* config/i386/i386.md (paritydi2, paritysi2): New expanders.
[official-gcc.git] / gcc / combine.c
blob6605b7a35584ad29bf2bd2a086c6d8ba3d5f432b
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;
126 /* combine_instructions may try to replace the right hand side of the
127 second instruction with the value of an associated REG_EQUAL note
128 before throwing it at try_combine. That is problematic when there
129 is a REG_DEAD note for a register used in the old right hand side
130 and can cause distribute_notes to do wrong things. This is the
131 second instruction if it has been so modified, null otherwise. */
133 static rtx i2mod;
135 /* When I2MOD is nonnull, this is a copy of the old right hand side. */
137 static rtx i2mod_old_rhs;
139 /* When I2MOD is nonnull, this is a copy of the new right hand side. */
141 static rtx i2mod_new_rhs;
143 /* Vector mapping INSN_UIDs to cuids.
144 The cuids are like uids but increase monotonically always.
145 Combine always uses cuids so that it can compare them.
146 But actually renumbering the uids, which we used to do,
147 proves to be a bad idea because it makes it hard to compare
148 the dumps produced by earlier passes with those from later passes. */
150 static int *uid_cuid;
151 static int max_uid_cuid;
153 /* Get the cuid of an insn. */
155 #define INSN_CUID(INSN) \
156 (INSN_UID (INSN) > max_uid_cuid ? insn_cuid (INSN) : uid_cuid[INSN_UID (INSN)])
158 /* Maximum register number, which is the size of the tables below. */
160 static unsigned int combine_max_regno;
162 struct reg_stat {
163 /* Record last point of death of (hard or pseudo) register n. */
164 rtx last_death;
166 /* Record last point of modification of (hard or pseudo) register n. */
167 rtx last_set;
169 /* The next group of fields allows the recording of the last value assigned
170 to (hard or pseudo) register n. We use this information to see if an
171 operation being processed is redundant given a prior operation performed
172 on the register. For example, an `and' with a constant is redundant if
173 all the zero bits are already known to be turned off.
175 We use an approach similar to that used by cse, but change it in the
176 following ways:
178 (1) We do not want to reinitialize at each label.
179 (2) It is useful, but not critical, to know the actual value assigned
180 to a register. Often just its form is helpful.
182 Therefore, we maintain the following fields:
184 last_set_value the last value assigned
185 last_set_label records the value of label_tick when the
186 register was assigned
187 last_set_table_tick records the value of label_tick when a
188 value using the register is assigned
189 last_set_invalid set to nonzero when it is not valid
190 to use the value of this register in some
191 register's value
193 To understand the usage of these tables, it is important to understand
194 the distinction between the value in last_set_value being valid and
195 the register being validly contained in some other expression in the
196 table.
198 (The next two parameters are out of date).
200 reg_stat[i].last_set_value is valid if it is nonzero, and either
201 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
203 Register I may validly appear in any expression returned for the value
204 of another register if reg_n_sets[i] is 1. It may also appear in the
205 value for register J if reg_stat[j].last_set_invalid is zero, or
206 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
208 If an expression is found in the table containing a register which may
209 not validly appear in an expression, the register is replaced by
210 something that won't match, (clobber (const_int 0)). */
212 /* Record last value assigned to (hard or pseudo) register n. */
214 rtx last_set_value;
216 /* Record the value of label_tick when an expression involving register n
217 is placed in last_set_value. */
219 int last_set_table_tick;
221 /* Record the value of label_tick when the value for register n is placed in
222 last_set_value. */
224 int last_set_label;
226 /* These fields are maintained in parallel with last_set_value and are
227 used to store the mode in which the register was last set, the bits
228 that were known to be zero when it was last set, and the number of
229 sign bits copies it was known to have when it was last set. */
231 unsigned HOST_WIDE_INT last_set_nonzero_bits;
232 char last_set_sign_bit_copies;
233 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
235 /* Set nonzero if references to register n in expressions should not be
236 used. last_set_invalid is set nonzero when this register is being
237 assigned to and last_set_table_tick == label_tick. */
239 char last_set_invalid;
241 /* Some registers that are set more than once and used in more than one
242 basic block are nevertheless always set in similar ways. For example,
243 a QImode register may be loaded from memory in two places on a machine
244 where byte loads zero extend.
246 We record in the following fields if a register has some leading bits
247 that are always equal to the sign bit, and what we know about the
248 nonzero bits of a register, specifically which bits are known to be
249 zero.
251 If an entry is zero, it means that we don't know anything special. */
253 unsigned char sign_bit_copies;
255 unsigned HOST_WIDE_INT nonzero_bits;
257 /* Record the value of the label_tick when the last truncation
258 happened. The field truncated_to_mode is only valid if
259 truncation_label == label_tick. */
261 int truncation_label;
263 /* Record the last truncation seen for this register. If truncation
264 is not a nop to this mode we might be able to save an explicit
265 truncation if we know that value already contains a truncated
266 value. */
268 ENUM_BITFIELD(machine_mode) truncated_to_mode : 8;
271 static struct reg_stat *reg_stat;
273 /* Record the cuid of the last insn that invalidated memory
274 (anything that writes memory, and subroutine calls, but not pushes). */
276 static int mem_last_set;
278 /* Record the cuid of the last CALL_INSN
279 so we can tell whether a potential combination crosses any calls. */
281 static int last_call_cuid;
283 /* When `subst' is called, this is the insn that is being modified
284 (by combining in a previous insn). The PATTERN of this insn
285 is still the old pattern partially modified and it should not be
286 looked at, but this may be used to examine the successors of the insn
287 to judge whether a simplification is valid. */
289 static rtx subst_insn;
291 /* This is the lowest CUID that `subst' is currently dealing with.
292 get_last_value will not return a value if the register was set at or
293 after this CUID. If not for this mechanism, we could get confused if
294 I2 or I1 in try_combine were an insn that used the old value of a register
295 to obtain a new value. In that case, we might erroneously get the
296 new value of the register when we wanted the old one. */
298 static int subst_low_cuid;
300 /* This contains any hard registers that are used in newpat; reg_dead_at_p
301 must consider all these registers to be always live. */
303 static HARD_REG_SET newpat_used_regs;
305 /* This is an insn to which a LOG_LINKS entry has been added. If this
306 insn is the earlier than I2 or I3, combine should rescan starting at
307 that location. */
309 static rtx added_links_insn;
311 /* Basic block in which we are performing combines. */
312 static basic_block this_basic_block;
314 /* A bitmap indicating which blocks had registers go dead at entry.
315 After combine, we'll need to re-do global life analysis with
316 those blocks as starting points. */
317 static sbitmap refresh_blocks;
319 /* The following array records the insn_rtx_cost for every insn
320 in the instruction stream. */
322 static int *uid_insn_cost;
324 /* Length of the currently allocated uid_insn_cost array. */
326 static int last_insn_cost;
328 /* Incremented for each label. */
330 static int label_tick;
332 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
333 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
335 static enum machine_mode nonzero_bits_mode;
337 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
338 be safely used. It is zero while computing them and after combine has
339 completed. This former test prevents propagating values based on
340 previously set values, which can be incorrect if a variable is modified
341 in a loop. */
343 static int nonzero_sign_valid;
346 /* Record one modification to rtl structure
347 to be undone by storing old_contents into *where. */
349 struct undo
351 struct undo *next;
352 enum { UNDO_RTX, UNDO_INT, UNDO_MODE } kind;
353 union { rtx r; int i; enum machine_mode m; } old_contents;
354 union { rtx *r; int *i; } where;
357 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
358 num_undo says how many are currently recorded.
360 other_insn is nonzero if we have modified some other insn in the process
361 of working on subst_insn. It must be verified too. */
363 struct undobuf
365 struct undo *undos;
366 struct undo *frees;
367 rtx other_insn;
370 static struct undobuf undobuf;
372 /* Number of times the pseudo being substituted for
373 was found and replaced. */
375 static int n_occurrences;
377 static rtx reg_nonzero_bits_for_combine (rtx, enum machine_mode, rtx,
378 enum machine_mode,
379 unsigned HOST_WIDE_INT,
380 unsigned HOST_WIDE_INT *);
381 static rtx reg_num_sign_bit_copies_for_combine (rtx, enum machine_mode, rtx,
382 enum machine_mode,
383 unsigned int, unsigned int *);
384 static void do_SUBST (rtx *, rtx);
385 static void do_SUBST_INT (int *, int);
386 static void init_reg_last (void);
387 static void setup_incoming_promotions (void);
388 static void set_nonzero_bits_and_sign_copies (rtx, rtx, void *);
389 static int cant_combine_insn_p (rtx);
390 static int can_combine_p (rtx, rtx, rtx, rtx, rtx *, rtx *);
391 static int combinable_i3pat (rtx, rtx *, rtx, rtx, int, rtx *);
392 static int contains_muldiv (rtx);
393 static rtx try_combine (rtx, rtx, rtx, int *);
394 static void undo_all (void);
395 static void undo_commit (void);
396 static rtx *find_split_point (rtx *, rtx);
397 static rtx subst (rtx, rtx, rtx, int, int);
398 static rtx combine_simplify_rtx (rtx, enum machine_mode, int);
399 static rtx simplify_if_then_else (rtx);
400 static rtx simplify_set (rtx);
401 static rtx simplify_logical (rtx);
402 static rtx expand_compound_operation (rtx);
403 static rtx expand_field_assignment (rtx);
404 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
405 rtx, unsigned HOST_WIDE_INT, int, int, int);
406 static rtx extract_left_shift (rtx, int);
407 static rtx make_compound_operation (rtx, enum rtx_code);
408 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
409 unsigned HOST_WIDE_INT *);
410 static rtx canon_reg_for_combine (rtx, rtx);
411 static rtx force_to_mode (rtx, enum machine_mode,
412 unsigned HOST_WIDE_INT, int);
413 static rtx if_then_else_cond (rtx, rtx *, rtx *);
414 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
415 static int rtx_equal_for_field_assignment_p (rtx, rtx);
416 static rtx make_field_assignment (rtx);
417 static rtx apply_distributive_law (rtx);
418 static rtx distribute_and_simplify_rtx (rtx, int);
419 static rtx simplify_and_const_int_1 (enum machine_mode, rtx,
420 unsigned HOST_WIDE_INT);
421 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
422 unsigned HOST_WIDE_INT);
423 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
424 HOST_WIDE_INT, enum machine_mode, int *);
425 static rtx simplify_shift_const_1 (enum rtx_code, enum machine_mode, rtx, int);
426 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
427 int);
428 static int recog_for_combine (rtx *, rtx, rtx *);
429 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
430 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
431 static void update_table_tick (rtx);
432 static void record_value_for_reg (rtx, rtx, rtx);
433 static void check_conversions (rtx, rtx);
434 static void record_dead_and_set_regs_1 (rtx, rtx, void *);
435 static void record_dead_and_set_regs (rtx);
436 static int get_last_value_validate (rtx *, rtx, int, int);
437 static rtx get_last_value (rtx);
438 static int use_crosses_set_p (rtx, int);
439 static void reg_dead_at_p_1 (rtx, rtx, void *);
440 static int reg_dead_at_p (rtx, rtx);
441 static void move_deaths (rtx, rtx, int, rtx, rtx *);
442 static int reg_bitfield_target_p (rtx, rtx);
443 static void distribute_notes (rtx, rtx, rtx, rtx, rtx, rtx);
444 static void distribute_links (rtx);
445 static void mark_used_regs_combine (rtx);
446 static int insn_cuid (rtx);
447 static void record_promoted_value (rtx, rtx);
448 static int unmentioned_reg_p_1 (rtx *, void *);
449 static bool unmentioned_reg_p (rtx, rtx);
450 static void record_truncated_value (rtx);
451 static bool reg_truncated_to_mode (enum machine_mode, rtx);
452 static rtx gen_lowpart_or_truncate (enum machine_mode, rtx);
455 /* It is not safe to use ordinary gen_lowpart in combine.
456 See comments in gen_lowpart_for_combine. */
457 #undef RTL_HOOKS_GEN_LOWPART
458 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine
460 /* Our implementation of gen_lowpart never emits a new pseudo. */
461 #undef RTL_HOOKS_GEN_LOWPART_NO_EMIT
462 #define RTL_HOOKS_GEN_LOWPART_NO_EMIT gen_lowpart_for_combine
464 #undef RTL_HOOKS_REG_NONZERO_REG_BITS
465 #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine
467 #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES
468 #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine
470 #undef RTL_HOOKS_REG_TRUNCATED_TO_MODE
471 #define RTL_HOOKS_REG_TRUNCATED_TO_MODE reg_truncated_to_mode
473 static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER;
476 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
477 insn. The substitution can be undone by undo_all. If INTO is already
478 set to NEWVAL, do not record this change. Because computing NEWVAL might
479 also call SUBST, we have to compute it before we put anything into
480 the undo table. */
482 static void
483 do_SUBST (rtx *into, rtx newval)
485 struct undo *buf;
486 rtx oldval = *into;
488 if (oldval == newval)
489 return;
491 /* We'd like to catch as many invalid transformations here as
492 possible. Unfortunately, there are way too many mode changes
493 that are perfectly valid, so we'd waste too much effort for
494 little gain doing the checks here. Focus on catching invalid
495 transformations involving integer constants. */
496 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
497 && GET_CODE (newval) == CONST_INT)
499 /* Sanity check that we're replacing oldval with a CONST_INT
500 that is a valid sign-extension for the original mode. */
501 gcc_assert (INTVAL (newval)
502 == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval)));
504 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
505 CONST_INT is not valid, because after the replacement, the
506 original mode would be gone. Unfortunately, we can't tell
507 when do_SUBST is called to replace the operand thereof, so we
508 perform this test on oldval instead, checking whether an
509 invalid replacement took place before we got here. */
510 gcc_assert (!(GET_CODE (oldval) == SUBREG
511 && GET_CODE (SUBREG_REG (oldval)) == CONST_INT));
512 gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
513 && GET_CODE (XEXP (oldval, 0)) == CONST_INT));
516 if (undobuf.frees)
517 buf = undobuf.frees, undobuf.frees = buf->next;
518 else
519 buf = XNEW (struct undo);
521 buf->kind = UNDO_RTX;
522 buf->where.r = into;
523 buf->old_contents.r = oldval;
524 *into = newval;
526 buf->next = undobuf.undos, undobuf.undos = buf;
529 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
531 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
532 for the value of a HOST_WIDE_INT value (including CONST_INT) is
533 not safe. */
535 static void
536 do_SUBST_INT (int *into, int newval)
538 struct undo *buf;
539 int oldval = *into;
541 if (oldval == newval)
542 return;
544 if (undobuf.frees)
545 buf = undobuf.frees, undobuf.frees = buf->next;
546 else
547 buf = XNEW (struct undo);
549 buf->kind = UNDO_INT;
550 buf->where.i = into;
551 buf->old_contents.i = oldval;
552 *into = newval;
554 buf->next = undobuf.undos, undobuf.undos = buf;
557 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
559 /* Similar to SUBST, but just substitute the mode. This is used when
560 changing the mode of a pseudo-register, so that any other
561 references to the entry in the regno_reg_rtx array will change as
562 well. */
564 static void
565 do_SUBST_MODE (rtx *into, enum machine_mode newval)
567 struct undo *buf;
568 enum machine_mode oldval = GET_MODE (*into);
570 if (oldval == newval)
571 return;
573 if (undobuf.frees)
574 buf = undobuf.frees, undobuf.frees = buf->next;
575 else
576 buf = XNEW (struct undo);
578 buf->kind = UNDO_MODE;
579 buf->where.r = into;
580 buf->old_contents.m = oldval;
581 PUT_MODE (*into, newval);
583 buf->next = undobuf.undos, undobuf.undos = buf;
586 #define SUBST_MODE(INTO, NEWVAL) do_SUBST_MODE(&(INTO), (NEWVAL))
588 /* Subroutine of try_combine. Determine whether the combine replacement
589 patterns NEWPAT and NEWI2PAT are cheaper according to insn_rtx_cost
590 that the original instruction sequence I1, I2 and I3. Note that I1
591 and/or NEWI2PAT may be NULL_RTX. This function returns false, if the
592 costs of all instructions can be estimated, and the replacements are
593 more expensive than the original sequence. */
595 static bool
596 combine_validate_cost (rtx i1, rtx i2, rtx i3, rtx newpat, rtx newi2pat)
598 int i1_cost, i2_cost, i3_cost;
599 int new_i2_cost, new_i3_cost;
600 int old_cost, new_cost;
602 /* Lookup the original insn_rtx_costs. */
603 i2_cost = INSN_UID (i2) <= last_insn_cost
604 ? uid_insn_cost[INSN_UID (i2)] : 0;
605 i3_cost = INSN_UID (i3) <= last_insn_cost
606 ? uid_insn_cost[INSN_UID (i3)] : 0;
608 if (i1)
610 i1_cost = INSN_UID (i1) <= last_insn_cost
611 ? uid_insn_cost[INSN_UID (i1)] : 0;
612 old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0)
613 ? i1_cost + i2_cost + i3_cost : 0;
615 else
617 old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0;
618 i1_cost = 0;
621 /* Calculate the replacement insn_rtx_costs. */
622 new_i3_cost = insn_rtx_cost (newpat);
623 if (newi2pat)
625 new_i2_cost = insn_rtx_cost (newi2pat);
626 new_cost = (new_i2_cost > 0 && new_i3_cost > 0)
627 ? new_i2_cost + new_i3_cost : 0;
629 else
631 new_cost = new_i3_cost;
632 new_i2_cost = 0;
635 if (undobuf.other_insn)
637 int old_other_cost, new_other_cost;
639 old_other_cost = (INSN_UID (undobuf.other_insn) <= last_insn_cost
640 ? uid_insn_cost[INSN_UID (undobuf.other_insn)] : 0);
641 new_other_cost = insn_rtx_cost (PATTERN (undobuf.other_insn));
642 if (old_other_cost > 0 && new_other_cost > 0)
644 old_cost += old_other_cost;
645 new_cost += new_other_cost;
647 else
648 old_cost = 0;
651 /* Disallow this recombination if both new_cost and old_cost are
652 greater than zero, and new_cost is greater than old cost. */
653 if (old_cost > 0
654 && new_cost > old_cost)
656 if (dump_file)
658 if (i1)
660 fprintf (dump_file,
661 "rejecting combination of insns %d, %d and %d\n",
662 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
663 fprintf (dump_file, "original costs %d + %d + %d = %d\n",
664 i1_cost, i2_cost, i3_cost, old_cost);
666 else
668 fprintf (dump_file,
669 "rejecting combination of insns %d and %d\n",
670 INSN_UID (i2), INSN_UID (i3));
671 fprintf (dump_file, "original costs %d + %d = %d\n",
672 i2_cost, i3_cost, old_cost);
675 if (newi2pat)
677 fprintf (dump_file, "replacement costs %d + %d = %d\n",
678 new_i2_cost, new_i3_cost, new_cost);
680 else
681 fprintf (dump_file, "replacement cost %d\n", new_cost);
684 return false;
687 /* Update the uid_insn_cost array with the replacement costs. */
688 uid_insn_cost[INSN_UID (i2)] = new_i2_cost;
689 uid_insn_cost[INSN_UID (i3)] = new_i3_cost;
690 if (i1)
691 uid_insn_cost[INSN_UID (i1)] = 0;
693 return true;
696 /* Main entry point for combiner. F is the first insn of the function.
697 NREGS is the first unused pseudo-reg number.
699 Return nonzero if the combiner has turned an indirect jump
700 instruction into a direct jump. */
701 static int
702 combine_instructions (rtx f, unsigned int nregs)
704 rtx insn, next;
705 #ifdef HAVE_cc0
706 rtx prev;
707 #endif
708 int i;
709 unsigned int j = 0;
710 rtx links, nextlinks;
711 sbitmap_iterator sbi;
713 int new_direct_jump_p = 0;
715 combine_attempts = 0;
716 combine_merges = 0;
717 combine_extras = 0;
718 combine_successes = 0;
720 combine_max_regno = nregs;
722 rtl_hooks = combine_rtl_hooks;
724 reg_stat = XCNEWVEC (struct reg_stat, nregs);
726 init_recog_no_volatile ();
728 /* Compute maximum uid value so uid_cuid can be allocated. */
730 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
731 if (INSN_UID (insn) > i)
732 i = INSN_UID (insn);
734 uid_cuid = XNEWVEC (int, i + 1);
735 max_uid_cuid = i;
737 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
739 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
740 problems when, for example, we have j <<= 1 in a loop. */
742 nonzero_sign_valid = 0;
744 /* Compute the mapping from uids to cuids.
745 Cuids are numbers assigned to insns, like uids,
746 except that cuids increase monotonically through the code.
748 Scan all SETs and see if we can deduce anything about what
749 bits are known to be zero for some registers and how many copies
750 of the sign bit are known to exist for those registers.
752 Also set any known values so that we can use it while searching
753 for what bits are known to be set. */
755 label_tick = 1;
757 setup_incoming_promotions ();
759 refresh_blocks = sbitmap_alloc (last_basic_block);
760 sbitmap_zero (refresh_blocks);
762 /* Allocate array of current insn_rtx_costs. */
763 uid_insn_cost = XCNEWVEC (int, max_uid_cuid + 1);
764 last_insn_cost = max_uid_cuid;
766 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
768 uid_cuid[INSN_UID (insn)] = ++i;
769 subst_low_cuid = i;
770 subst_insn = insn;
772 if (INSN_P (insn))
774 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
775 NULL);
776 record_dead_and_set_regs (insn);
778 #ifdef AUTO_INC_DEC
779 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
780 if (REG_NOTE_KIND (links) == REG_INC)
781 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
782 NULL);
783 #endif
785 /* Record the current insn_rtx_cost of this instruction. */
786 if (NONJUMP_INSN_P (insn))
787 uid_insn_cost[INSN_UID (insn)] = insn_rtx_cost (PATTERN (insn));
788 if (dump_file)
789 fprintf(dump_file, "insn_cost %d: %d\n",
790 INSN_UID (insn), uid_insn_cost[INSN_UID (insn)]);
793 if (LABEL_P (insn))
794 label_tick++;
797 nonzero_sign_valid = 1;
799 /* Now scan all the insns in forward order. */
801 label_tick = 1;
802 last_call_cuid = 0;
803 mem_last_set = 0;
804 init_reg_last ();
805 setup_incoming_promotions ();
807 FOR_EACH_BB (this_basic_block)
809 for (insn = BB_HEAD (this_basic_block);
810 insn != NEXT_INSN (BB_END (this_basic_block));
811 insn = next ? next : NEXT_INSN (insn))
813 next = 0;
815 if (LABEL_P (insn))
816 label_tick++;
818 else if (INSN_P (insn))
820 /* See if we know about function return values before this
821 insn based upon SUBREG flags. */
822 check_conversions (insn, PATTERN (insn));
824 /* Try this insn with each insn it links back to. */
826 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
827 if ((next = try_combine (insn, XEXP (links, 0),
828 NULL_RTX, &new_direct_jump_p)) != 0)
829 goto retry;
831 /* Try each sequence of three linked insns ending with this one. */
833 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
835 rtx link = XEXP (links, 0);
837 /* If the linked insn has been replaced by a note, then there
838 is no point in pursuing this chain any further. */
839 if (NOTE_P (link))
840 continue;
842 for (nextlinks = LOG_LINKS (link);
843 nextlinks;
844 nextlinks = XEXP (nextlinks, 1))
845 if ((next = try_combine (insn, link,
846 XEXP (nextlinks, 0),
847 &new_direct_jump_p)) != 0)
848 goto retry;
851 #ifdef HAVE_cc0
852 /* Try to combine a jump insn that uses CC0
853 with a preceding insn that sets CC0, and maybe with its
854 logical predecessor as well.
855 This is how we make decrement-and-branch insns.
856 We need this special code because data flow connections
857 via CC0 do not get entered in LOG_LINKS. */
859 if (JUMP_P (insn)
860 && (prev = prev_nonnote_insn (insn)) != 0
861 && NONJUMP_INSN_P (prev)
862 && sets_cc0_p (PATTERN (prev)))
864 if ((next = try_combine (insn, prev,
865 NULL_RTX, &new_direct_jump_p)) != 0)
866 goto retry;
868 for (nextlinks = LOG_LINKS (prev); nextlinks;
869 nextlinks = XEXP (nextlinks, 1))
870 if ((next = try_combine (insn, prev,
871 XEXP (nextlinks, 0),
872 &new_direct_jump_p)) != 0)
873 goto retry;
876 /* Do the same for an insn that explicitly references CC0. */
877 if (NONJUMP_INSN_P (insn)
878 && (prev = prev_nonnote_insn (insn)) != 0
879 && NONJUMP_INSN_P (prev)
880 && sets_cc0_p (PATTERN (prev))
881 && GET_CODE (PATTERN (insn)) == SET
882 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
884 if ((next = try_combine (insn, prev,
885 NULL_RTX, &new_direct_jump_p)) != 0)
886 goto retry;
888 for (nextlinks = LOG_LINKS (prev); nextlinks;
889 nextlinks = XEXP (nextlinks, 1))
890 if ((next = try_combine (insn, prev,
891 XEXP (nextlinks, 0),
892 &new_direct_jump_p)) != 0)
893 goto retry;
896 /* Finally, see if any of the insns that this insn links to
897 explicitly references CC0. If so, try this insn, that insn,
898 and its predecessor if it sets CC0. */
899 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
900 if (NONJUMP_INSN_P (XEXP (links, 0))
901 && GET_CODE (PATTERN (XEXP (links, 0))) == SET
902 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0))))
903 && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0
904 && NONJUMP_INSN_P (prev)
905 && sets_cc0_p (PATTERN (prev))
906 && (next = try_combine (insn, XEXP (links, 0),
907 prev, &new_direct_jump_p)) != 0)
908 goto retry;
909 #endif
911 /* Try combining an insn with two different insns whose results it
912 uses. */
913 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
914 for (nextlinks = XEXP (links, 1); nextlinks;
915 nextlinks = XEXP (nextlinks, 1))
916 if ((next = try_combine (insn, XEXP (links, 0),
917 XEXP (nextlinks, 0),
918 &new_direct_jump_p)) != 0)
919 goto retry;
921 /* Try this insn with each REG_EQUAL note it links back to. */
922 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
924 rtx set, note;
925 rtx temp = XEXP (links, 0);
926 if ((set = single_set (temp)) != 0
927 && (note = find_reg_equal_equiv_note (temp)) != 0
928 && (note = XEXP (note, 0), GET_CODE (note)) != EXPR_LIST
929 /* Avoid using a register that may already been marked
930 dead by an earlier instruction. */
931 && ! unmentioned_reg_p (note, SET_SRC (set))
932 && (GET_MODE (note) == VOIDmode
933 ? SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))
934 : GET_MODE (SET_DEST (set)) == GET_MODE (note)))
936 /* Temporarily replace the set's source with the
937 contents of the REG_EQUAL note. The insn will
938 be deleted or recognized by try_combine. */
939 rtx orig = SET_SRC (set);
940 SET_SRC (set) = note;
941 i2mod = temp;
942 i2mod_old_rhs = copy_rtx (orig);
943 i2mod_new_rhs = copy_rtx (note);
944 next = try_combine (insn, i2mod, NULL_RTX,
945 &new_direct_jump_p);
946 i2mod = NULL_RTX;
947 if (next)
948 goto retry;
949 SET_SRC (set) = orig;
953 if (!NOTE_P (insn))
954 record_dead_and_set_regs (insn);
956 retry:
961 clear_bb_flags ();
963 EXECUTE_IF_SET_IN_SBITMAP (refresh_blocks, 0, j, sbi)
964 BASIC_BLOCK (j)->flags |= BB_DIRTY;
965 new_direct_jump_p |= purge_all_dead_edges ();
966 delete_noop_moves ();
968 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
969 PROP_DEATH_NOTES | PROP_SCAN_DEAD_CODE
970 | PROP_KILL_DEAD_CODE);
972 /* Clean up. */
973 sbitmap_free (refresh_blocks);
974 free (uid_insn_cost);
975 free (reg_stat);
976 free (uid_cuid);
979 struct undo *undo, *next;
980 for (undo = undobuf.frees; undo; undo = next)
982 next = undo->next;
983 free (undo);
985 undobuf.frees = 0;
988 total_attempts += combine_attempts;
989 total_merges += combine_merges;
990 total_extras += combine_extras;
991 total_successes += combine_successes;
993 nonzero_sign_valid = 0;
994 rtl_hooks = general_rtl_hooks;
996 /* Make recognizer allow volatile MEMs again. */
997 init_recog ();
999 return new_direct_jump_p;
1002 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
1004 static void
1005 init_reg_last (void)
1007 unsigned int i;
1008 for (i = 0; i < combine_max_regno; i++)
1009 memset (reg_stat + i, 0, offsetof (struct reg_stat, sign_bit_copies));
1012 /* Set up any promoted values for incoming argument registers. */
1014 static void
1015 setup_incoming_promotions (void)
1017 unsigned int regno;
1018 rtx reg;
1019 enum machine_mode mode;
1020 int unsignedp;
1021 rtx first = get_insns ();
1023 if (targetm.calls.promote_function_args (TREE_TYPE (cfun->decl)))
1025 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1026 /* Check whether this register can hold an incoming pointer
1027 argument. FUNCTION_ARG_REGNO_P tests outgoing register
1028 numbers, so translate if necessary due to register windows. */
1029 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (regno))
1030 && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0)
1032 record_value_for_reg
1033 (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND
1034 : SIGN_EXTEND),
1035 GET_MODE (reg),
1036 gen_rtx_CLOBBER (mode, const0_rtx)));
1041 /* Called via note_stores. If X is a pseudo that is narrower than
1042 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
1044 If we are setting only a portion of X and we can't figure out what
1045 portion, assume all bits will be used since we don't know what will
1046 be happening.
1048 Similarly, set how many bits of X are known to be copies of the sign bit
1049 at all locations in the function. This is the smallest number implied
1050 by any set of X. */
1052 static void
1053 set_nonzero_bits_and_sign_copies (rtx x, rtx set,
1054 void *data ATTRIBUTE_UNUSED)
1056 unsigned int num;
1058 if (REG_P (x)
1059 && REGNO (x) >= FIRST_PSEUDO_REGISTER
1060 /* If this register is undefined at the start of the file, we can't
1061 say what its contents were. */
1062 && ! REGNO_REG_SET_P
1063 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start, REGNO (x))
1064 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
1066 if (set == 0 || GET_CODE (set) == CLOBBER)
1068 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1069 reg_stat[REGNO (x)].sign_bit_copies = 1;
1070 return;
1073 /* If this is a complex assignment, see if we can convert it into a
1074 simple assignment. */
1075 set = expand_field_assignment (set);
1077 /* If this is a simple assignment, or we have a paradoxical SUBREG,
1078 set what we know about X. */
1080 if (SET_DEST (set) == x
1081 || (GET_CODE (SET_DEST (set)) == SUBREG
1082 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
1083 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set)))))
1084 && SUBREG_REG (SET_DEST (set)) == x))
1086 rtx src = SET_SRC (set);
1088 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
1089 /* If X is narrower than a word and SRC is a non-negative
1090 constant that would appear negative in the mode of X,
1091 sign-extend it for use in reg_stat[].nonzero_bits because some
1092 machines (maybe most) will actually do the sign-extension
1093 and this is the conservative approach.
1095 ??? For 2.5, try to tighten up the MD files in this regard
1096 instead of this kludge. */
1098 if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD
1099 && GET_CODE (src) == CONST_INT
1100 && INTVAL (src) > 0
1101 && 0 != (INTVAL (src)
1102 & ((HOST_WIDE_INT) 1
1103 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
1104 src = GEN_INT (INTVAL (src)
1105 | ((HOST_WIDE_INT) (-1)
1106 << GET_MODE_BITSIZE (GET_MODE (x))));
1107 #endif
1109 /* Don't call nonzero_bits if it cannot change anything. */
1110 if (reg_stat[REGNO (x)].nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
1111 reg_stat[REGNO (x)].nonzero_bits
1112 |= nonzero_bits (src, nonzero_bits_mode);
1113 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
1114 if (reg_stat[REGNO (x)].sign_bit_copies == 0
1115 || reg_stat[REGNO (x)].sign_bit_copies > num)
1116 reg_stat[REGNO (x)].sign_bit_copies = num;
1118 else
1120 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1121 reg_stat[REGNO (x)].sign_bit_copies = 1;
1126 /* See if INSN can be combined into I3. PRED and SUCC are optionally
1127 insns that were previously combined into I3 or that will be combined
1128 into the merger of INSN and I3.
1130 Return 0 if the combination is not allowed for any reason.
1132 If the combination is allowed, *PDEST will be set to the single
1133 destination of INSN and *PSRC to the single source, and this function
1134 will return 1. */
1136 static int
1137 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED, rtx succ,
1138 rtx *pdest, rtx *psrc)
1140 int i;
1141 rtx set = 0, src, dest;
1142 rtx p;
1143 #ifdef AUTO_INC_DEC
1144 rtx link;
1145 #endif
1146 int all_adjacent = (succ ? (next_active_insn (insn) == succ
1147 && next_active_insn (succ) == i3)
1148 : next_active_insn (insn) == i3);
1150 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
1151 or a PARALLEL consisting of such a SET and CLOBBERs.
1153 If INSN has CLOBBER parallel parts, ignore them for our processing.
1154 By definition, these happen during the execution of the insn. When it
1155 is merged with another insn, all bets are off. If they are, in fact,
1156 needed and aren't also supplied in I3, they may be added by
1157 recog_for_combine. Otherwise, it won't match.
1159 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
1160 note.
1162 Get the source and destination of INSN. If more than one, can't
1163 combine. */
1165 if (GET_CODE (PATTERN (insn)) == SET)
1166 set = PATTERN (insn);
1167 else if (GET_CODE (PATTERN (insn)) == PARALLEL
1168 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
1170 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1172 rtx elt = XVECEXP (PATTERN (insn), 0, i);
1173 rtx note;
1175 switch (GET_CODE (elt))
1177 /* This is important to combine floating point insns
1178 for the SH4 port. */
1179 case USE:
1180 /* Combining an isolated USE doesn't make sense.
1181 We depend here on combinable_i3pat to reject them. */
1182 /* The code below this loop only verifies that the inputs of
1183 the SET in INSN do not change. We call reg_set_between_p
1184 to verify that the REG in the USE does not change between
1185 I3 and INSN.
1186 If the USE in INSN was for a pseudo register, the matching
1187 insn pattern will likely match any register; combining this
1188 with any other USE would only be safe if we knew that the
1189 used registers have identical values, or if there was
1190 something to tell them apart, e.g. different modes. For
1191 now, we forgo such complicated tests and simply disallow
1192 combining of USES of pseudo registers with any other USE. */
1193 if (REG_P (XEXP (elt, 0))
1194 && GET_CODE (PATTERN (i3)) == PARALLEL)
1196 rtx i3pat = PATTERN (i3);
1197 int i = XVECLEN (i3pat, 0) - 1;
1198 unsigned int regno = REGNO (XEXP (elt, 0));
1202 rtx i3elt = XVECEXP (i3pat, 0, i);
1204 if (GET_CODE (i3elt) == USE
1205 && REG_P (XEXP (i3elt, 0))
1206 && (REGNO (XEXP (i3elt, 0)) == regno
1207 ? reg_set_between_p (XEXP (elt, 0),
1208 PREV_INSN (insn), i3)
1209 : regno >= FIRST_PSEUDO_REGISTER))
1210 return 0;
1212 while (--i >= 0);
1214 break;
1216 /* We can ignore CLOBBERs. */
1217 case CLOBBER:
1218 break;
1220 case SET:
1221 /* Ignore SETs whose result isn't used but not those that
1222 have side-effects. */
1223 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1224 && (!(note = find_reg_note (insn, REG_EH_REGION, NULL_RTX))
1225 || INTVAL (XEXP (note, 0)) <= 0)
1226 && ! side_effects_p (elt))
1227 break;
1229 /* If we have already found a SET, this is a second one and
1230 so we cannot combine with this insn. */
1231 if (set)
1232 return 0;
1234 set = elt;
1235 break;
1237 default:
1238 /* Anything else means we can't combine. */
1239 return 0;
1243 if (set == 0
1244 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1245 so don't do anything with it. */
1246 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1247 return 0;
1249 else
1250 return 0;
1252 if (set == 0)
1253 return 0;
1255 set = expand_field_assignment (set);
1256 src = SET_SRC (set), dest = SET_DEST (set);
1258 /* Don't eliminate a store in the stack pointer. */
1259 if (dest == stack_pointer_rtx
1260 /* Don't combine with an insn that sets a register to itself if it has
1261 a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */
1262 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1263 /* Can't merge an ASM_OPERANDS. */
1264 || GET_CODE (src) == ASM_OPERANDS
1265 /* Can't merge a function call. */
1266 || GET_CODE (src) == CALL
1267 /* Don't eliminate a function call argument. */
1268 || (CALL_P (i3)
1269 && (find_reg_fusage (i3, USE, dest)
1270 || (REG_P (dest)
1271 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1272 && global_regs[REGNO (dest)])))
1273 /* Don't substitute into an incremented register. */
1274 || FIND_REG_INC_NOTE (i3, dest)
1275 || (succ && FIND_REG_INC_NOTE (succ, dest))
1276 /* Don't substitute into a non-local goto, this confuses CFG. */
1277 || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX))
1278 #if 0
1279 /* Don't combine the end of a libcall into anything. */
1280 /* ??? This gives worse code, and appears to be unnecessary, since no
1281 pass after flow uses REG_LIBCALL/REG_RETVAL notes. Local-alloc does
1282 use REG_RETVAL notes for noconflict blocks, but other code here
1283 makes sure that those insns don't disappear. */
1284 || find_reg_note (insn, REG_RETVAL, NULL_RTX)
1285 #endif
1286 /* Make sure that DEST is not used after SUCC but before I3. */
1287 || (succ && ! all_adjacent
1288 && reg_used_between_p (dest, succ, i3))
1289 /* Make sure that the value that is to be substituted for the register
1290 does not use any registers whose values alter in between. However,
1291 If the insns are adjacent, a use can't cross a set even though we
1292 think it might (this can happen for a sequence of insns each setting
1293 the same destination; last_set of that register might point to
1294 a NOTE). If INSN has a REG_EQUIV note, the register is always
1295 equivalent to the memory so the substitution is valid even if there
1296 are intervening stores. Also, don't move a volatile asm or
1297 UNSPEC_VOLATILE across any other insns. */
1298 || (! all_adjacent
1299 && (((!MEM_P (src)
1300 || ! find_reg_note (insn, REG_EQUIV, src))
1301 && use_crosses_set_p (src, INSN_CUID (insn)))
1302 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1303 || GET_CODE (src) == UNSPEC_VOLATILE))
1304 /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get
1305 better register allocation by not doing the combine. */
1306 || find_reg_note (i3, REG_NO_CONFLICT, dest)
1307 || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest))
1308 /* Don't combine across a CALL_INSN, because that would possibly
1309 change whether the life span of some REGs crosses calls or not,
1310 and it is a pain to update that information.
1311 Exception: if source is a constant, moving it later can't hurt.
1312 Accept that special case, because it helps -fforce-addr a lot. */
1313 || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src)))
1314 return 0;
1316 /* DEST must either be a REG or CC0. */
1317 if (REG_P (dest))
1319 /* If register alignment is being enforced for multi-word items in all
1320 cases except for parameters, it is possible to have a register copy
1321 insn referencing a hard register that is not allowed to contain the
1322 mode being copied and which would not be valid as an operand of most
1323 insns. Eliminate this problem by not combining with such an insn.
1325 Also, on some machines we don't want to extend the life of a hard
1326 register. */
1328 if (REG_P (src)
1329 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1330 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1331 /* Don't extend the life of a hard register unless it is
1332 user variable (if we have few registers) or it can't
1333 fit into the desired register (meaning something special
1334 is going on).
1335 Also avoid substituting a return register into I3, because
1336 reload can't handle a conflict with constraints of other
1337 inputs. */
1338 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1339 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1340 return 0;
1342 else if (GET_CODE (dest) != CC0)
1343 return 0;
1346 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1347 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1348 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER)
1350 /* Don't substitute for a register intended as a clobberable
1351 operand. */
1352 rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0);
1353 if (rtx_equal_p (reg, dest))
1354 return 0;
1356 /* If the clobber represents an earlyclobber operand, we must not
1357 substitute an expression containing the clobbered register.
1358 As we do not analyze the constraint strings here, we have to
1359 make the conservative assumption. However, if the register is
1360 a fixed hard reg, the clobber cannot represent any operand;
1361 we leave it up to the machine description to either accept or
1362 reject use-and-clobber patterns. */
1363 if (!REG_P (reg)
1364 || REGNO (reg) >= FIRST_PSEUDO_REGISTER
1365 || !fixed_regs[REGNO (reg)])
1366 if (reg_overlap_mentioned_p (reg, src))
1367 return 0;
1370 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1371 or not), reject, unless nothing volatile comes between it and I3 */
1373 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1375 /* Make sure succ doesn't contain a volatile reference. */
1376 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1377 return 0;
1379 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1380 if (INSN_P (p) && p != succ && volatile_refs_p (PATTERN (p)))
1381 return 0;
1384 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1385 to be an explicit register variable, and was chosen for a reason. */
1387 if (GET_CODE (src) == ASM_OPERANDS
1388 && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1389 return 0;
1391 /* If there are any volatile insns between INSN and I3, reject, because
1392 they might affect machine state. */
1394 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1395 if (INSN_P (p) && p != succ && volatile_insn_p (PATTERN (p)))
1396 return 0;
1398 /* If INSN contains an autoincrement or autodecrement, make sure that
1399 register is not used between there and I3, and not already used in
1400 I3 either. Neither must it be used in PRED or SUCC, if they exist.
1401 Also insist that I3 not be a jump; if it were one
1402 and the incremented register were spilled, we would lose. */
1404 #ifdef AUTO_INC_DEC
1405 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1406 if (REG_NOTE_KIND (link) == REG_INC
1407 && (JUMP_P (i3)
1408 || reg_used_between_p (XEXP (link, 0), insn, i3)
1409 || (pred != NULL_RTX
1410 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred)))
1411 || (succ != NULL_RTX
1412 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ)))
1413 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1414 return 0;
1415 #endif
1417 #ifdef HAVE_cc0
1418 /* Don't combine an insn that follows a CC0-setting insn.
1419 An insn that uses CC0 must not be separated from the one that sets it.
1420 We do, however, allow I2 to follow a CC0-setting insn if that insn
1421 is passed as I1; in that case it will be deleted also.
1422 We also allow combining in this case if all the insns are adjacent
1423 because that would leave the two CC0 insns adjacent as well.
1424 It would be more logical to test whether CC0 occurs inside I1 or I2,
1425 but that would be much slower, and this ought to be equivalent. */
1427 p = prev_nonnote_insn (insn);
1428 if (p && p != pred && NONJUMP_INSN_P (p) && sets_cc0_p (PATTERN (p))
1429 && ! all_adjacent)
1430 return 0;
1431 #endif
1433 /* If we get here, we have passed all the tests and the combination is
1434 to be allowed. */
1436 *pdest = dest;
1437 *psrc = src;
1439 return 1;
1442 /* LOC is the location within I3 that contains its pattern or the component
1443 of a PARALLEL of the pattern. We validate that it is valid for combining.
1445 One problem is if I3 modifies its output, as opposed to replacing it
1446 entirely, we can't allow the output to contain I2DEST or I1DEST as doing
1447 so would produce an insn that is not equivalent to the original insns.
1449 Consider:
1451 (set (reg:DI 101) (reg:DI 100))
1452 (set (subreg:SI (reg:DI 101) 0) <foo>)
1454 This is NOT equivalent to:
1456 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
1457 (set (reg:DI 101) (reg:DI 100))])
1459 Not only does this modify 100 (in which case it might still be valid
1460 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
1462 We can also run into a problem if I2 sets a register that I1
1463 uses and I1 gets directly substituted into I3 (not via I2). In that
1464 case, we would be getting the wrong value of I2DEST into I3, so we
1465 must reject the combination. This case occurs when I2 and I1 both
1466 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
1467 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
1468 of a SET must prevent combination from occurring.
1470 Before doing the above check, we first try to expand a field assignment
1471 into a set of logical operations.
1473 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
1474 we place a register that is both set and used within I3. If more than one
1475 such register is detected, we fail.
1477 Return 1 if the combination is valid, zero otherwise. */
1479 static int
1480 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest,
1481 int i1_not_in_src, rtx *pi3dest_killed)
1483 rtx x = *loc;
1485 if (GET_CODE (x) == SET)
1487 rtx set = x ;
1488 rtx dest = SET_DEST (set);
1489 rtx src = SET_SRC (set);
1490 rtx inner_dest = dest;
1491 rtx subdest;
1493 while (GET_CODE (inner_dest) == STRICT_LOW_PART
1494 || GET_CODE (inner_dest) == SUBREG
1495 || GET_CODE (inner_dest) == ZERO_EXTRACT)
1496 inner_dest = XEXP (inner_dest, 0);
1498 /* Check for the case where I3 modifies its output, as discussed
1499 above. We don't want to prevent pseudos from being combined
1500 into the address of a MEM, so only prevent the combination if
1501 i1 or i2 set the same MEM. */
1502 if ((inner_dest != dest &&
1503 (!MEM_P (inner_dest)
1504 || rtx_equal_p (i2dest, inner_dest)
1505 || (i1dest && rtx_equal_p (i1dest, inner_dest)))
1506 && (reg_overlap_mentioned_p (i2dest, inner_dest)
1507 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))))
1509 /* This is the same test done in can_combine_p except we can't test
1510 all_adjacent; we don't have to, since this instruction will stay
1511 in place, thus we are not considering increasing the lifetime of
1512 INNER_DEST.
1514 Also, if this insn sets a function argument, combining it with
1515 something that might need a spill could clobber a previous
1516 function argument; the all_adjacent test in can_combine_p also
1517 checks this; here, we do a more specific test for this case. */
1519 || (REG_P (inner_dest)
1520 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
1521 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
1522 GET_MODE (inner_dest))))
1523 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)))
1524 return 0;
1526 /* If DEST is used in I3, it is being killed in this insn, so
1527 record that for later. We have to consider paradoxical
1528 subregs here, since they kill the whole register, but we
1529 ignore partial subregs, STRICT_LOW_PART, etc.
1530 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
1531 STACK_POINTER_REGNUM, since these are always considered to be
1532 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
1533 subdest = dest;
1534 if (GET_CODE (subdest) == SUBREG
1535 && (GET_MODE_SIZE (GET_MODE (subdest))
1536 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (subdest)))))
1537 subdest = SUBREG_REG (subdest);
1538 if (pi3dest_killed
1539 && REG_P (subdest)
1540 && reg_referenced_p (subdest, PATTERN (i3))
1541 && REGNO (subdest) != FRAME_POINTER_REGNUM
1542 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1543 && REGNO (subdest) != HARD_FRAME_POINTER_REGNUM
1544 #endif
1545 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
1546 && (REGNO (subdest) != ARG_POINTER_REGNUM
1547 || ! fixed_regs [REGNO (subdest)])
1548 #endif
1549 && REGNO (subdest) != STACK_POINTER_REGNUM)
1551 if (*pi3dest_killed)
1552 return 0;
1554 *pi3dest_killed = subdest;
1558 else if (GET_CODE (x) == PARALLEL)
1560 int i;
1562 for (i = 0; i < XVECLEN (x, 0); i++)
1563 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest,
1564 i1_not_in_src, pi3dest_killed))
1565 return 0;
1568 return 1;
1571 /* Return 1 if X is an arithmetic expression that contains a multiplication
1572 and division. We don't count multiplications by powers of two here. */
1574 static int
1575 contains_muldiv (rtx x)
1577 switch (GET_CODE (x))
1579 case MOD: case DIV: case UMOD: case UDIV:
1580 return 1;
1582 case MULT:
1583 return ! (GET_CODE (XEXP (x, 1)) == CONST_INT
1584 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0);
1585 default:
1586 if (BINARY_P (x))
1587 return contains_muldiv (XEXP (x, 0))
1588 || contains_muldiv (XEXP (x, 1));
1590 if (UNARY_P (x))
1591 return contains_muldiv (XEXP (x, 0));
1593 return 0;
1597 /* Determine whether INSN can be used in a combination. Return nonzero if
1598 not. This is used in try_combine to detect early some cases where we
1599 can't perform combinations. */
1601 static int
1602 cant_combine_insn_p (rtx insn)
1604 rtx set;
1605 rtx src, dest;
1607 /* If this isn't really an insn, we can't do anything.
1608 This can occur when flow deletes an insn that it has merged into an
1609 auto-increment address. */
1610 if (! INSN_P (insn))
1611 return 1;
1613 /* Never combine loads and stores involving hard regs that are likely
1614 to be spilled. The register allocator can usually handle such
1615 reg-reg moves by tying. If we allow the combiner to make
1616 substitutions of likely-spilled regs, reload might die.
1617 As an exception, we allow combinations involving fixed regs; these are
1618 not available to the register allocator so there's no risk involved. */
1620 set = single_set (insn);
1621 if (! set)
1622 return 0;
1623 src = SET_SRC (set);
1624 dest = SET_DEST (set);
1625 if (GET_CODE (src) == SUBREG)
1626 src = SUBREG_REG (src);
1627 if (GET_CODE (dest) == SUBREG)
1628 dest = SUBREG_REG (dest);
1629 if (REG_P (src) && REG_P (dest)
1630 && ((REGNO (src) < FIRST_PSEUDO_REGISTER
1631 && ! fixed_regs[REGNO (src)]
1632 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (src))))
1633 || (REGNO (dest) < FIRST_PSEUDO_REGISTER
1634 && ! fixed_regs[REGNO (dest)]
1635 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (dest))))))
1636 return 1;
1638 return 0;
1641 struct likely_spilled_retval_info
1643 unsigned regno, nregs;
1644 unsigned mask;
1647 /* Called via note_stores by likely_spilled_retval_p. Remove from info->mask
1648 hard registers that are known to be written to / clobbered in full. */
1649 static void
1650 likely_spilled_retval_1 (rtx x, rtx set, void *data)
1652 struct likely_spilled_retval_info *info = data;
1653 unsigned regno, nregs;
1654 unsigned new_mask;
1656 if (!REG_P (XEXP (set, 0)))
1657 return;
1658 regno = REGNO (x);
1659 if (regno >= info->regno + info->nregs)
1660 return;
1661 nregs = hard_regno_nregs[regno][GET_MODE (x)];
1662 if (regno + nregs <= info->regno)
1663 return;
1664 new_mask = (2U << (nregs - 1)) - 1;
1665 if (regno < info->regno)
1666 new_mask >>= info->regno - regno;
1667 else
1668 new_mask <<= regno - info->regno;
1669 info->mask &= ~new_mask;
1672 /* Return nonzero iff part of the return value is live during INSN, and
1673 it is likely spilled. This can happen when more than one insn is needed
1674 to copy the return value, e.g. when we consider to combine into the
1675 second copy insn for a complex value. */
1677 static int
1678 likely_spilled_retval_p (rtx insn)
1680 rtx use = BB_END (this_basic_block);
1681 rtx reg, p;
1682 unsigned regno, nregs;
1683 /* We assume here that no machine mode needs more than
1684 32 hard registers when the value overlaps with a register
1685 for which FUNCTION_VALUE_REGNO_P is true. */
1686 unsigned mask;
1687 struct likely_spilled_retval_info info;
1689 if (!NONJUMP_INSN_P (use) || GET_CODE (PATTERN (use)) != USE || insn == use)
1690 return 0;
1691 reg = XEXP (PATTERN (use), 0);
1692 if (!REG_P (reg) || !FUNCTION_VALUE_REGNO_P (REGNO (reg)))
1693 return 0;
1694 regno = REGNO (reg);
1695 nregs = hard_regno_nregs[regno][GET_MODE (reg)];
1696 if (nregs == 1)
1697 return 0;
1698 mask = (2U << (nregs - 1)) - 1;
1700 /* Disregard parts of the return value that are set later. */
1701 info.regno = regno;
1702 info.nregs = nregs;
1703 info.mask = mask;
1704 for (p = PREV_INSN (use); info.mask && p != insn; p = PREV_INSN (p))
1705 if (INSN_P (p))
1706 note_stores (PATTERN (p), likely_spilled_retval_1, &info);
1707 mask = info.mask;
1709 /* Check if any of the (probably) live return value registers is
1710 likely spilled. */
1711 nregs --;
1714 if ((mask & 1 << nregs)
1715 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno + nregs)))
1716 return 1;
1717 } while (nregs--);
1718 return 0;
1721 /* Adjust INSN after we made a change to its destination.
1723 Changing the destination can invalidate notes that say something about
1724 the results of the insn and a LOG_LINK pointing to the insn. */
1726 static void
1727 adjust_for_new_dest (rtx insn)
1729 rtx *loc;
1731 /* For notes, be conservative and simply remove them. */
1732 loc = &REG_NOTES (insn);
1733 while (*loc)
1735 enum reg_note kind = REG_NOTE_KIND (*loc);
1736 if (kind == REG_EQUAL || kind == REG_EQUIV)
1737 *loc = XEXP (*loc, 1);
1738 else
1739 loc = &XEXP (*loc, 1);
1742 /* The new insn will have a destination that was previously the destination
1743 of an insn just above it. Call distribute_links to make a LOG_LINK from
1744 the next use of that destination. */
1745 distribute_links (gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX));
1748 /* Return TRUE if combine can reuse reg X in mode MODE.
1749 ADDED_SETS is nonzero if the original set is still required. */
1750 static bool
1751 can_change_dest_mode (rtx x, int added_sets, enum machine_mode mode)
1753 unsigned int regno;
1755 if (!REG_P(x))
1756 return false;
1758 regno = REGNO (x);
1759 /* Allow hard registers if the new mode is legal, and occupies no more
1760 registers than the old mode. */
1761 if (regno < FIRST_PSEUDO_REGISTER)
1762 return (HARD_REGNO_MODE_OK (regno, mode)
1763 && (hard_regno_nregs[regno][GET_MODE (x)]
1764 >= hard_regno_nregs[regno][mode]));
1766 /* Or a pseudo that is only used once. */
1767 return (REG_N_SETS (regno) == 1 && !added_sets
1768 && !REG_USERVAR_P (x));
1772 /* Check whether X, the destination of a set, refers to part of
1773 the register specified by REG. */
1775 static bool
1776 reg_subword_p (rtx x, rtx reg)
1778 /* Check that reg is an integer mode register. */
1779 if (!REG_P (reg) || GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT)
1780 return false;
1782 if (GET_CODE (x) == STRICT_LOW_PART
1783 || GET_CODE (x) == ZERO_EXTRACT)
1784 x = XEXP (x, 0);
1786 return GET_CODE (x) == SUBREG
1787 && SUBREG_REG (x) == reg
1788 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT;
1792 /* Try to combine the insns I1 and I2 into I3.
1793 Here I1 and I2 appear earlier than I3.
1794 I1 can be zero; then we combine just I2 into I3.
1796 If we are combining three insns and the resulting insn is not recognized,
1797 try splitting it into two insns. If that happens, I2 and I3 are retained
1798 and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2
1799 are pseudo-deleted.
1801 Return 0 if the combination does not work. Then nothing is changed.
1802 If we did the combination, return the insn at which combine should
1803 resume scanning.
1805 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
1806 new direct jump instruction. */
1808 static rtx
1809 try_combine (rtx i3, rtx i2, rtx i1, int *new_direct_jump_p)
1811 /* New patterns for I3 and I2, respectively. */
1812 rtx newpat, newi2pat = 0;
1813 rtvec newpat_vec_with_clobbers = 0;
1814 int substed_i2 = 0, substed_i1 = 0;
1815 /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */
1816 int added_sets_1, added_sets_2;
1817 /* Total number of SETs to put into I3. */
1818 int total_sets;
1819 /* Nonzero if I2's body now appears in I3. */
1820 int i2_is_used;
1821 /* INSN_CODEs for new I3, new I2, and user of condition code. */
1822 int insn_code_number, i2_code_number = 0, other_code_number = 0;
1823 /* Contains I3 if the destination of I3 is used in its source, which means
1824 that the old life of I3 is being killed. If that usage is placed into
1825 I2 and not in I3, a REG_DEAD note must be made. */
1826 rtx i3dest_killed = 0;
1827 /* SET_DEST and SET_SRC of I2 and I1. */
1828 rtx i2dest, i2src, i1dest = 0, i1src = 0;
1829 /* PATTERN (I1) and PATTERN (I2), or a copy of it in certain cases. */
1830 rtx i1pat = 0, i2pat = 0;
1831 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
1832 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
1833 int i2dest_killed = 0, i1dest_killed = 0;
1834 int i1_feeds_i3 = 0;
1835 /* Notes that must be added to REG_NOTES in I3 and I2. */
1836 rtx new_i3_notes, new_i2_notes;
1837 /* Notes that we substituted I3 into I2 instead of the normal case. */
1838 int i3_subst_into_i2 = 0;
1839 /* Notes that I1, I2 or I3 is a MULT operation. */
1840 int have_mult = 0;
1841 int swap_i2i3 = 0;
1843 int maxreg;
1844 rtx temp;
1845 rtx link;
1846 int i;
1848 /* Exit early if one of the insns involved can't be used for
1849 combinations. */
1850 if (cant_combine_insn_p (i3)
1851 || cant_combine_insn_p (i2)
1852 || (i1 && cant_combine_insn_p (i1))
1853 || likely_spilled_retval_p (i3)
1854 /* We also can't do anything if I3 has a
1855 REG_LIBCALL note since we don't want to disrupt the contiguity of a
1856 libcall. */
1857 #if 0
1858 /* ??? This gives worse code, and appears to be unnecessary, since no
1859 pass after flow uses REG_LIBCALL/REG_RETVAL notes. */
1860 || find_reg_note (i3, REG_LIBCALL, NULL_RTX)
1861 #endif
1863 return 0;
1865 combine_attempts++;
1866 undobuf.other_insn = 0;
1868 /* Reset the hard register usage information. */
1869 CLEAR_HARD_REG_SET (newpat_used_regs);
1871 /* If I1 and I2 both feed I3, they can be in any order. To simplify the
1872 code below, set I1 to be the earlier of the two insns. */
1873 if (i1 && INSN_CUID (i1) > INSN_CUID (i2))
1874 temp = i1, i1 = i2, i2 = temp;
1876 added_links_insn = 0;
1878 /* First check for one important special-case that the code below will
1879 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
1880 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
1881 we may be able to replace that destination with the destination of I3.
1882 This occurs in the common code where we compute both a quotient and
1883 remainder into a structure, in which case we want to do the computation
1884 directly into the structure to avoid register-register copies.
1886 Note that this case handles both multiple sets in I2 and also
1887 cases where I2 has a number of CLOBBER or PARALLELs.
1889 We make very conservative checks below and only try to handle the
1890 most common cases of this. For example, we only handle the case
1891 where I2 and I3 are adjacent to avoid making difficult register
1892 usage tests. */
1894 if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET
1895 && REG_P (SET_SRC (PATTERN (i3)))
1896 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
1897 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
1898 && GET_CODE (PATTERN (i2)) == PARALLEL
1899 && ! side_effects_p (SET_DEST (PATTERN (i3)))
1900 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
1901 below would need to check what is inside (and reg_overlap_mentioned_p
1902 doesn't support those codes anyway). Don't allow those destinations;
1903 the resulting insn isn't likely to be recognized anyway. */
1904 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
1905 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
1906 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
1907 SET_DEST (PATTERN (i3)))
1908 && next_real_insn (i2) == i3)
1910 rtx p2 = PATTERN (i2);
1912 /* Make sure that the destination of I3,
1913 which we are going to substitute into one output of I2,
1914 is not used within another output of I2. We must avoid making this:
1915 (parallel [(set (mem (reg 69)) ...)
1916 (set (reg 69) ...)])
1917 which is not well-defined as to order of actions.
1918 (Besides, reload can't handle output reloads for this.)
1920 The problem can also happen if the dest of I3 is a memory ref,
1921 if another dest in I2 is an indirect memory ref. */
1922 for (i = 0; i < XVECLEN (p2, 0); i++)
1923 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1924 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1925 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
1926 SET_DEST (XVECEXP (p2, 0, i))))
1927 break;
1929 if (i == XVECLEN (p2, 0))
1930 for (i = 0; i < XVECLEN (p2, 0); i++)
1931 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1932 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1933 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
1935 combine_merges++;
1937 subst_insn = i3;
1938 subst_low_cuid = INSN_CUID (i2);
1940 added_sets_2 = added_sets_1 = 0;
1941 i2dest = SET_SRC (PATTERN (i3));
1942 i2dest_killed = dead_or_set_p (i2, i2dest);
1944 /* Replace the dest in I2 with our dest and make the resulting
1945 insn the new pattern for I3. Then skip to where we
1946 validate the pattern. Everything was set up above. */
1947 SUBST (SET_DEST (XVECEXP (p2, 0, i)),
1948 SET_DEST (PATTERN (i3)));
1950 newpat = p2;
1951 i3_subst_into_i2 = 1;
1952 goto validate_replacement;
1956 /* If I2 is setting a pseudo to a constant and I3 is setting some
1957 sub-part of it to another constant, merge them by making a new
1958 constant. */
1959 if (i1 == 0
1960 && (temp = single_set (i2)) != 0
1961 && (GET_CODE (SET_SRC (temp)) == CONST_INT
1962 || GET_CODE (SET_SRC (temp)) == CONST_DOUBLE)
1963 && GET_CODE (PATTERN (i3)) == SET
1964 && (GET_CODE (SET_SRC (PATTERN (i3))) == CONST_INT
1965 || GET_CODE (SET_SRC (PATTERN (i3))) == CONST_DOUBLE)
1966 && reg_subword_p (SET_DEST (PATTERN (i3)), SET_DEST (temp)))
1968 rtx dest = SET_DEST (PATTERN (i3));
1969 int offset = -1;
1970 int width = 0;
1972 if (GET_CODE (dest) == ZERO_EXTRACT)
1974 if (GET_CODE (XEXP (dest, 1)) == CONST_INT
1975 && GET_CODE (XEXP (dest, 2)) == CONST_INT)
1977 width = INTVAL (XEXP (dest, 1));
1978 offset = INTVAL (XEXP (dest, 2));
1979 dest = XEXP (dest, 0);
1980 if (BITS_BIG_ENDIAN)
1981 offset = GET_MODE_BITSIZE (GET_MODE (dest)) - width - offset;
1984 else
1986 if (GET_CODE (dest) == STRICT_LOW_PART)
1987 dest = XEXP (dest, 0);
1988 width = GET_MODE_BITSIZE (GET_MODE (dest));
1989 offset = 0;
1992 if (offset >= 0)
1994 /* If this is the low part, we're done. */
1995 if (subreg_lowpart_p (dest))
1997 /* Handle the case where inner is twice the size of outer. */
1998 else if (GET_MODE_BITSIZE (GET_MODE (SET_DEST (temp)))
1999 == 2 * GET_MODE_BITSIZE (GET_MODE (dest)))
2000 offset += GET_MODE_BITSIZE (GET_MODE (dest));
2001 /* Otherwise give up for now. */
2002 else
2003 offset = -1;
2006 if (offset >= 0
2007 && (GET_MODE_BITSIZE (GET_MODE (SET_DEST (temp)))
2008 <= HOST_BITS_PER_WIDE_INT * 2))
2010 HOST_WIDE_INT mhi, ohi, ihi;
2011 HOST_WIDE_INT mlo, olo, ilo;
2012 rtx inner = SET_SRC (PATTERN (i3));
2013 rtx outer = SET_SRC (temp);
2015 if (GET_CODE (outer) == CONST_INT)
2017 olo = INTVAL (outer);
2018 ohi = olo < 0 ? -1 : 0;
2020 else
2022 olo = CONST_DOUBLE_LOW (outer);
2023 ohi = CONST_DOUBLE_HIGH (outer);
2026 if (GET_CODE (inner) == CONST_INT)
2028 ilo = INTVAL (inner);
2029 ihi = ilo < 0 ? -1 : 0;
2031 else
2033 ilo = CONST_DOUBLE_LOW (inner);
2034 ihi = CONST_DOUBLE_HIGH (inner);
2037 if (width < HOST_BITS_PER_WIDE_INT)
2039 mlo = ((unsigned HOST_WIDE_INT) 1 << width) - 1;
2040 mhi = 0;
2042 else if (width < HOST_BITS_PER_WIDE_INT * 2)
2044 mhi = ((unsigned HOST_WIDE_INT) 1
2045 << (width - HOST_BITS_PER_WIDE_INT)) - 1;
2046 mlo = -1;
2048 else
2050 mlo = -1;
2051 mhi = -1;
2054 ilo &= mlo;
2055 ihi &= mhi;
2057 if (offset >= HOST_BITS_PER_WIDE_INT)
2059 mhi = mlo << (offset - HOST_BITS_PER_WIDE_INT);
2060 mlo = 0;
2061 ihi = ilo << (offset - HOST_BITS_PER_WIDE_INT);
2062 ilo = 0;
2064 else if (offset > 0)
2066 mhi = (mhi << offset) | ((unsigned HOST_WIDE_INT) mlo
2067 >> (HOST_BITS_PER_WIDE_INT - offset));
2068 mlo = mlo << offset;
2069 ihi = (ihi << offset) | ((unsigned HOST_WIDE_INT) ilo
2070 >> (HOST_BITS_PER_WIDE_INT - offset));
2071 ilo = ilo << offset;
2074 olo = (olo & ~mlo) | ilo;
2075 ohi = (ohi & ~mhi) | ihi;
2077 combine_merges++;
2078 subst_insn = i3;
2079 subst_low_cuid = INSN_CUID (i2);
2080 added_sets_2 = added_sets_1 = 0;
2081 i2dest = SET_DEST (temp);
2082 i2dest_killed = dead_or_set_p (i2, i2dest);
2084 SUBST (SET_SRC (temp),
2085 immed_double_const (olo, ohi, GET_MODE (SET_DEST (temp))));
2087 newpat = PATTERN (i2);
2088 goto validate_replacement;
2092 #ifndef HAVE_cc0
2093 /* If we have no I1 and I2 looks like:
2094 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
2095 (set Y OP)])
2096 make up a dummy I1 that is
2097 (set Y OP)
2098 and change I2 to be
2099 (set (reg:CC X) (compare:CC Y (const_int 0)))
2101 (We can ignore any trailing CLOBBERs.)
2103 This undoes a previous combination and allows us to match a branch-and-
2104 decrement insn. */
2106 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
2107 && XVECLEN (PATTERN (i2), 0) >= 2
2108 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
2109 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
2110 == MODE_CC)
2111 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
2112 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
2113 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
2114 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)))
2115 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
2116 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
2118 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
2119 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
2120 break;
2122 if (i == 1)
2124 /* We make I1 with the same INSN_UID as I2. This gives it
2125 the same INSN_CUID for value tracking. Our fake I1 will
2126 never appear in the insn stream so giving it the same INSN_UID
2127 as I2 will not cause a problem. */
2129 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
2130 BLOCK_FOR_INSN (i2), INSN_LOCATOR (i2),
2131 XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX,
2132 NULL_RTX);
2134 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
2135 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
2136 SET_DEST (PATTERN (i1)));
2139 #endif
2141 /* Verify that I2 and I1 are valid for combining. */
2142 if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src)
2143 || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src)))
2145 undo_all ();
2146 return 0;
2149 /* Record whether I2DEST is used in I2SRC and similarly for the other
2150 cases. Knowing this will help in register status updating below. */
2151 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
2152 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
2153 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
2154 i2dest_killed = dead_or_set_p (i2, i2dest);
2155 i1dest_killed = i1 && dead_or_set_p (i1, i1dest);
2157 /* See if I1 directly feeds into I3. It does if I1DEST is not used
2158 in I2SRC. */
2159 i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src);
2161 /* Ensure that I3's pattern can be the destination of combines. */
2162 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest,
2163 i1 && i2dest_in_i1src && i1_feeds_i3,
2164 &i3dest_killed))
2166 undo_all ();
2167 return 0;
2170 /* See if any of the insns is a MULT operation. Unless one is, we will
2171 reject a combination that is, since it must be slower. Be conservative
2172 here. */
2173 if (GET_CODE (i2src) == MULT
2174 || (i1 != 0 && GET_CODE (i1src) == MULT)
2175 || (GET_CODE (PATTERN (i3)) == SET
2176 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
2177 have_mult = 1;
2179 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
2180 We used to do this EXCEPT in one case: I3 has a post-inc in an
2181 output operand. However, that exception can give rise to insns like
2182 mov r3,(r3)+
2183 which is a famous insn on the PDP-11 where the value of r3 used as the
2184 source was model-dependent. Avoid this sort of thing. */
2186 #if 0
2187 if (!(GET_CODE (PATTERN (i3)) == SET
2188 && REG_P (SET_SRC (PATTERN (i3)))
2189 && MEM_P (SET_DEST (PATTERN (i3)))
2190 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
2191 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
2192 /* It's not the exception. */
2193 #endif
2194 #ifdef AUTO_INC_DEC
2195 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
2196 if (REG_NOTE_KIND (link) == REG_INC
2197 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
2198 || (i1 != 0
2199 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
2201 undo_all ();
2202 return 0;
2204 #endif
2206 /* See if the SETs in I1 or I2 need to be kept around in the merged
2207 instruction: whenever the value set there is still needed past I3.
2208 For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3.
2210 For the SET in I1, we have two cases: If I1 and I2 independently
2211 feed into I3, the set in I1 needs to be kept around if I1DEST dies
2212 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
2213 in I1 needs to be kept around unless I1DEST dies or is set in either
2214 I2 or I3. We can distinguish these cases by seeing if I2SRC mentions
2215 I1DEST. If so, we know I1 feeds into I2. */
2217 added_sets_2 = ! dead_or_set_p (i3, i2dest);
2219 added_sets_1
2220 = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest)
2221 : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest)));
2223 /* If the set in I2 needs to be kept around, we must make a copy of
2224 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
2225 PATTERN (I2), we are only substituting for the original I1DEST, not into
2226 an already-substituted copy. This also prevents making self-referential
2227 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
2228 I2DEST. */
2230 if (added_sets_2)
2232 if (GET_CODE (PATTERN (i2)) == PARALLEL)
2233 i2pat = gen_rtx_SET (VOIDmode, i2dest, copy_rtx (i2src));
2234 else
2235 i2pat = copy_rtx (PATTERN (i2));
2238 if (added_sets_1)
2240 if (GET_CODE (PATTERN (i1)) == PARALLEL)
2241 i1pat = gen_rtx_SET (VOIDmode, i1dest, copy_rtx (i1src));
2242 else
2243 i1pat = copy_rtx (PATTERN (i1));
2246 combine_merges++;
2248 /* Substitute in the latest insn for the regs set by the earlier ones. */
2250 maxreg = max_reg_num ();
2252 subst_insn = i3;
2254 #ifndef HAVE_cc0
2255 /* Many machines that don't use CC0 have insns that can both perform an
2256 arithmetic operation and set the condition code. These operations will
2257 be represented as a PARALLEL with the first element of the vector
2258 being a COMPARE of an arithmetic operation with the constant zero.
2259 The second element of the vector will set some pseudo to the result
2260 of the same arithmetic operation. If we simplify the COMPARE, we won't
2261 match such a pattern and so will generate an extra insn. Here we test
2262 for this case, where both the comparison and the operation result are
2263 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
2264 I2SRC. Later we will make the PARALLEL that contains I2. */
2266 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
2267 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
2268 && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx
2269 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
2271 #ifdef SELECT_CC_MODE
2272 rtx *cc_use;
2273 enum machine_mode compare_mode;
2274 #endif
2276 newpat = PATTERN (i3);
2277 SUBST (XEXP (SET_SRC (newpat), 0), i2src);
2279 i2_is_used = 1;
2281 #ifdef SELECT_CC_MODE
2282 /* See if a COMPARE with the operand we substituted in should be done
2283 with the mode that is currently being used. If not, do the same
2284 processing we do in `subst' for a SET; namely, if the destination
2285 is used only once, try to replace it with a register of the proper
2286 mode and also replace the COMPARE. */
2287 if (undobuf.other_insn == 0
2288 && (cc_use = find_single_use (SET_DEST (newpat), i3,
2289 &undobuf.other_insn))
2290 && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use),
2291 i2src, const0_rtx))
2292 != GET_MODE (SET_DEST (newpat))))
2294 if (can_change_dest_mode(SET_DEST (newpat), added_sets_2,
2295 compare_mode))
2297 unsigned int regno = REGNO (SET_DEST (newpat));
2298 rtx new_dest;
2300 if (regno < FIRST_PSEUDO_REGISTER)
2301 new_dest = gen_rtx_REG (compare_mode, regno);
2302 else
2304 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
2305 new_dest = regno_reg_rtx[regno];
2308 SUBST (SET_DEST (newpat), new_dest);
2309 SUBST (XEXP (*cc_use, 0), new_dest);
2310 SUBST (SET_SRC (newpat),
2311 gen_rtx_COMPARE (compare_mode, i2src, const0_rtx));
2313 else
2314 undobuf.other_insn = 0;
2316 #endif
2318 else
2319 #endif
2321 /* It is possible that the source of I2 or I1 may be performing
2322 an unneeded operation, such as a ZERO_EXTEND of something
2323 that is known to have the high part zero. Handle that case
2324 by letting subst look at the innermost one of them.
2326 Another way to do this would be to have a function that tries
2327 to simplify a single insn instead of merging two or more
2328 insns. We don't do this because of the potential of infinite
2329 loops and because of the potential extra memory required.
2330 However, doing it the way we are is a bit of a kludge and
2331 doesn't catch all cases.
2333 But only do this if -fexpensive-optimizations since it slows
2334 things down and doesn't usually win.
2336 This is not done in the COMPARE case above because the
2337 unmodified I2PAT is used in the PARALLEL and so a pattern
2338 with a modified I2SRC would not match. */
2340 if (flag_expensive_optimizations)
2342 /* Pass pc_rtx so no substitutions are done, just
2343 simplifications. */
2344 if (i1)
2346 subst_low_cuid = INSN_CUID (i1);
2347 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0);
2349 else
2351 subst_low_cuid = INSN_CUID (i2);
2352 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0);
2356 n_occurrences = 0; /* `subst' counts here */
2358 /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we
2359 need to make a unique copy of I2SRC each time we substitute it
2360 to avoid self-referential rtl. */
2362 subst_low_cuid = INSN_CUID (i2);
2363 newpat = subst (PATTERN (i3), i2dest, i2src, 0,
2364 ! i1_feeds_i3 && i1dest_in_i1src);
2365 substed_i2 = 1;
2367 /* Record whether i2's body now appears within i3's body. */
2368 i2_is_used = n_occurrences;
2371 /* If we already got a failure, don't try to do more. Otherwise,
2372 try to substitute in I1 if we have it. */
2374 if (i1 && GET_CODE (newpat) != CLOBBER)
2376 /* Before we can do this substitution, we must redo the test done
2377 above (see detailed comments there) that ensures that I1DEST
2378 isn't mentioned in any SETs in NEWPAT that are field assignments. */
2380 if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX,
2381 0, (rtx*) 0))
2383 undo_all ();
2384 return 0;
2387 n_occurrences = 0;
2388 subst_low_cuid = INSN_CUID (i1);
2389 newpat = subst (newpat, i1dest, i1src, 0, 0);
2390 substed_i1 = 1;
2393 /* Fail if an autoincrement side-effect has been duplicated. Be careful
2394 to count all the ways that I2SRC and I1SRC can be used. */
2395 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
2396 && i2_is_used + added_sets_2 > 1)
2397 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
2398 && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3)
2399 > 1))
2400 /* Fail if we tried to make a new register. */
2401 || max_reg_num () != maxreg
2402 /* Fail if we couldn't do something and have a CLOBBER. */
2403 || GET_CODE (newpat) == CLOBBER
2404 /* Fail if this new pattern is a MULT and we didn't have one before
2405 at the outer level. */
2406 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
2407 && ! have_mult))
2409 undo_all ();
2410 return 0;
2413 /* If the actions of the earlier insns must be kept
2414 in addition to substituting them into the latest one,
2415 we must make a new PARALLEL for the latest insn
2416 to hold additional the SETs. */
2418 if (added_sets_1 || added_sets_2)
2420 combine_extras++;
2422 if (GET_CODE (newpat) == PARALLEL)
2424 rtvec old = XVEC (newpat, 0);
2425 total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2;
2426 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2427 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
2428 sizeof (old->elem[0]) * old->num_elem);
2430 else
2432 rtx old = newpat;
2433 total_sets = 1 + added_sets_1 + added_sets_2;
2434 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2435 XVECEXP (newpat, 0, 0) = old;
2438 if (added_sets_1)
2439 XVECEXP (newpat, 0, --total_sets) = i1pat;
2441 if (added_sets_2)
2443 /* If there is no I1, use I2's body as is. We used to also not do
2444 the subst call below if I2 was substituted into I3,
2445 but that could lose a simplification. */
2446 if (i1 == 0)
2447 XVECEXP (newpat, 0, --total_sets) = i2pat;
2448 else
2449 /* See comment where i2pat is assigned. */
2450 XVECEXP (newpat, 0, --total_sets)
2451 = subst (i2pat, i1dest, i1src, 0, 0);
2455 /* We come here when we are replacing a destination in I2 with the
2456 destination of I3. */
2457 validate_replacement:
2459 /* Note which hard regs this insn has as inputs. */
2460 mark_used_regs_combine (newpat);
2462 /* If recog_for_combine fails, it strips existing clobbers. If we'll
2463 consider splitting this pattern, we might need these clobbers. */
2464 if (i1 && GET_CODE (newpat) == PARALLEL
2465 && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER)
2467 int len = XVECLEN (newpat, 0);
2469 newpat_vec_with_clobbers = rtvec_alloc (len);
2470 for (i = 0; i < len; i++)
2471 RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i);
2474 /* Is the result of combination a valid instruction? */
2475 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2477 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
2478 the second SET's destination is a register that is unused and isn't
2479 marked as an instruction that might trap in an EH region. In that case,
2480 we just need the first SET. This can occur when simplifying a divmod
2481 insn. We *must* test for this case here because the code below that
2482 splits two independent SETs doesn't handle this case correctly when it
2483 updates the register status.
2485 It's pointless doing this if we originally had two sets, one from
2486 i3, and one from i2. Combining then splitting the parallel results
2487 in the original i2 again plus an invalid insn (which we delete).
2488 The net effect is only to move instructions around, which makes
2489 debug info less accurate.
2491 Also check the case where the first SET's destination is unused.
2492 That would not cause incorrect code, but does cause an unneeded
2493 insn to remain. */
2495 if (insn_code_number < 0
2496 && !(added_sets_2 && i1 == 0)
2497 && GET_CODE (newpat) == PARALLEL
2498 && XVECLEN (newpat, 0) == 2
2499 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2500 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2501 && asm_noperands (newpat) < 0)
2503 rtx set0 = XVECEXP (newpat, 0, 0);
2504 rtx set1 = XVECEXP (newpat, 0, 1);
2505 rtx note;
2507 if (((REG_P (SET_DEST (set1))
2508 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
2509 || (GET_CODE (SET_DEST (set1)) == SUBREG
2510 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
2511 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2512 || INTVAL (XEXP (note, 0)) <= 0)
2513 && ! side_effects_p (SET_SRC (set1)))
2515 newpat = set0;
2516 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2519 else if (((REG_P (SET_DEST (set0))
2520 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
2521 || (GET_CODE (SET_DEST (set0)) == SUBREG
2522 && find_reg_note (i3, REG_UNUSED,
2523 SUBREG_REG (SET_DEST (set0)))))
2524 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2525 || INTVAL (XEXP (note, 0)) <= 0)
2526 && ! side_effects_p (SET_SRC (set0)))
2528 newpat = set1;
2529 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2531 if (insn_code_number >= 0)
2533 /* If we will be able to accept this, we have made a
2534 change to the destination of I3. This requires us to
2535 do a few adjustments. */
2537 PATTERN (i3) = newpat;
2538 adjust_for_new_dest (i3);
2543 /* If we were combining three insns and the result is a simple SET
2544 with no ASM_OPERANDS that wasn't recognized, try to split it into two
2545 insns. There are two ways to do this. It can be split using a
2546 machine-specific method (like when you have an addition of a large
2547 constant) or by combine in the function find_split_point. */
2549 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
2550 && asm_noperands (newpat) < 0)
2552 rtx m_split, *split;
2554 /* See if the MD file can split NEWPAT. If it can't, see if letting it
2555 use I2DEST as a scratch register will help. In the latter case,
2556 convert I2DEST to the mode of the source of NEWPAT if we can. */
2558 m_split = split_insns (newpat, i3);
2560 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
2561 inputs of NEWPAT. */
2563 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
2564 possible to try that as a scratch reg. This would require adding
2565 more code to make it work though. */
2567 if (m_split == 0 && ! reg_overlap_mentioned_p (i2dest, newpat))
2569 enum machine_mode new_mode = GET_MODE (SET_DEST (newpat));
2571 /* First try to split using the original register as a
2572 scratch register. */
2573 m_split = split_insns (gen_rtx_PARALLEL
2574 (VOIDmode,
2575 gen_rtvec (2, newpat,
2576 gen_rtx_CLOBBER (VOIDmode,
2577 i2dest))),
2578 i3);
2580 /* If that didn't work, try changing the mode of I2DEST if
2581 we can. */
2582 if (m_split == 0
2583 && new_mode != GET_MODE (i2dest)
2584 && new_mode != VOIDmode
2585 && can_change_dest_mode (i2dest, added_sets_2, new_mode))
2587 enum machine_mode old_mode = GET_MODE (i2dest);
2588 rtx ni2dest;
2590 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2591 ni2dest = gen_rtx_REG (new_mode, REGNO (i2dest));
2592 else
2594 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], new_mode);
2595 ni2dest = regno_reg_rtx[REGNO (i2dest)];
2598 m_split = split_insns (gen_rtx_PARALLEL
2599 (VOIDmode,
2600 gen_rtvec (2, newpat,
2601 gen_rtx_CLOBBER (VOIDmode,
2602 ni2dest))),
2603 i3);
2605 if (m_split == 0
2606 && REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2608 struct undo *buf;
2610 PUT_MODE (regno_reg_rtx[REGNO (i2dest)], old_mode);
2611 buf = undobuf.undos;
2612 undobuf.undos = buf->next;
2613 buf->next = undobuf.frees;
2614 undobuf.frees = buf;
2619 /* If recog_for_combine has discarded clobbers, try to use them
2620 again for the split. */
2621 if (m_split == 0 && newpat_vec_with_clobbers)
2622 m_split
2623 = split_insns (gen_rtx_PARALLEL (VOIDmode,
2624 newpat_vec_with_clobbers), i3);
2626 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
2628 m_split = PATTERN (m_split);
2629 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
2630 if (insn_code_number >= 0)
2631 newpat = m_split;
2633 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
2634 && (next_real_insn (i2) == i3
2635 || ! use_crosses_set_p (PATTERN (m_split), INSN_CUID (i2))))
2637 rtx i2set, i3set;
2638 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
2639 newi2pat = PATTERN (m_split);
2641 i3set = single_set (NEXT_INSN (m_split));
2642 i2set = single_set (m_split);
2644 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2646 /* If I2 or I3 has multiple SETs, we won't know how to track
2647 register status, so don't use these insns. If I2's destination
2648 is used between I2 and I3, we also can't use these insns. */
2650 if (i2_code_number >= 0 && i2set && i3set
2651 && (next_real_insn (i2) == i3
2652 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
2653 insn_code_number = recog_for_combine (&newi3pat, i3,
2654 &new_i3_notes);
2655 if (insn_code_number >= 0)
2656 newpat = newi3pat;
2658 /* It is possible that both insns now set the destination of I3.
2659 If so, we must show an extra use of it. */
2661 if (insn_code_number >= 0)
2663 rtx new_i3_dest = SET_DEST (i3set);
2664 rtx new_i2_dest = SET_DEST (i2set);
2666 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
2667 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
2668 || GET_CODE (new_i3_dest) == SUBREG)
2669 new_i3_dest = XEXP (new_i3_dest, 0);
2671 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
2672 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
2673 || GET_CODE (new_i2_dest) == SUBREG)
2674 new_i2_dest = XEXP (new_i2_dest, 0);
2676 if (REG_P (new_i3_dest)
2677 && REG_P (new_i2_dest)
2678 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
2679 REG_N_SETS (REGNO (new_i2_dest))++;
2683 /* If we can split it and use I2DEST, go ahead and see if that
2684 helps things be recognized. Verify that none of the registers
2685 are set between I2 and I3. */
2686 if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0
2687 #ifdef HAVE_cc0
2688 && REG_P (i2dest)
2689 #endif
2690 /* We need I2DEST in the proper mode. If it is a hard register
2691 or the only use of a pseudo, we can change its mode.
2692 Make sure we don't change a hard register to have a mode that
2693 isn't valid for it, or change the number of registers. */
2694 && (GET_MODE (*split) == GET_MODE (i2dest)
2695 || GET_MODE (*split) == VOIDmode
2696 || can_change_dest_mode (i2dest, added_sets_2,
2697 GET_MODE (*split)))
2698 && (next_real_insn (i2) == i3
2699 || ! use_crosses_set_p (*split, INSN_CUID (i2)))
2700 /* We can't overwrite I2DEST if its value is still used by
2701 NEWPAT. */
2702 && ! reg_referenced_p (i2dest, newpat))
2704 rtx newdest = i2dest;
2705 enum rtx_code split_code = GET_CODE (*split);
2706 enum machine_mode split_mode = GET_MODE (*split);
2707 bool subst_done = false;
2708 newi2pat = NULL_RTX;
2710 /* Get NEWDEST as a register in the proper mode. We have already
2711 validated that we can do this. */
2712 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
2714 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2715 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
2716 else
2718 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], split_mode);
2719 newdest = regno_reg_rtx[REGNO (i2dest)];
2723 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
2724 an ASHIFT. This can occur if it was inside a PLUS and hence
2725 appeared to be a memory address. This is a kludge. */
2726 if (split_code == MULT
2727 && GET_CODE (XEXP (*split, 1)) == CONST_INT
2728 && INTVAL (XEXP (*split, 1)) > 0
2729 && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0)
2731 SUBST (*split, gen_rtx_ASHIFT (split_mode,
2732 XEXP (*split, 0), GEN_INT (i)));
2733 /* Update split_code because we may not have a multiply
2734 anymore. */
2735 split_code = GET_CODE (*split);
2738 #ifdef INSN_SCHEDULING
2739 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
2740 be written as a ZERO_EXTEND. */
2741 if (split_code == SUBREG && MEM_P (SUBREG_REG (*split)))
2743 #ifdef LOAD_EXTEND_OP
2744 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
2745 what it really is. */
2746 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
2747 == SIGN_EXTEND)
2748 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
2749 SUBREG_REG (*split)));
2750 else
2751 #endif
2752 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
2753 SUBREG_REG (*split)));
2755 #endif
2757 /* Attempt to split binary operators using arithmetic identities. */
2758 if (BINARY_P (SET_SRC (newpat))
2759 && split_mode == GET_MODE (SET_SRC (newpat))
2760 && ! side_effects_p (SET_SRC (newpat)))
2762 rtx setsrc = SET_SRC (newpat);
2763 enum machine_mode mode = GET_MODE (setsrc);
2764 enum rtx_code code = GET_CODE (setsrc);
2765 rtx src_op0 = XEXP (setsrc, 0);
2766 rtx src_op1 = XEXP (setsrc, 1);
2768 /* Split "X = Y op Y" as "Z = Y; X = Z op Z". */
2769 if (rtx_equal_p (src_op0, src_op1))
2771 newi2pat = gen_rtx_SET (VOIDmode, newdest, src_op0);
2772 SUBST (XEXP (setsrc, 0), newdest);
2773 SUBST (XEXP (setsrc, 1), newdest);
2774 subst_done = true;
2776 /* Split "((P op Q) op R) op S" where op is PLUS or MULT. */
2777 else if ((code == PLUS || code == MULT)
2778 && GET_CODE (src_op0) == code
2779 && GET_CODE (XEXP (src_op0, 0)) == code
2780 && (INTEGRAL_MODE_P (mode)
2781 || (FLOAT_MODE_P (mode)
2782 && flag_unsafe_math_optimizations)))
2784 rtx p = XEXP (XEXP (src_op0, 0), 0);
2785 rtx q = XEXP (XEXP (src_op0, 0), 1);
2786 rtx r = XEXP (src_op0, 1);
2787 rtx s = src_op1;
2789 /* Split both "((X op Y) op X) op Y" and
2790 "((X op Y) op Y) op X" as "T op T" where T is
2791 "X op Y". */
2792 if ((rtx_equal_p (p,r) && rtx_equal_p (q,s))
2793 || (rtx_equal_p (p,s) && rtx_equal_p (q,r)))
2795 newi2pat = gen_rtx_SET (VOIDmode, newdest,
2796 XEXP (src_op0, 0));
2797 SUBST (XEXP (setsrc, 0), newdest);
2798 SUBST (XEXP (setsrc, 1), newdest);
2799 subst_done = true;
2801 /* Split "((X op X) op Y) op Y)" as "T op T" where
2802 T is "X op Y". */
2803 else if (rtx_equal_p (p,q) && rtx_equal_p (r,s))
2805 rtx tmp = simplify_gen_binary (code, mode, p, r);
2806 newi2pat = gen_rtx_SET (VOIDmode, newdest, tmp);
2807 SUBST (XEXP (setsrc, 0), newdest);
2808 SUBST (XEXP (setsrc, 1), newdest);
2809 subst_done = true;
2814 if (!subst_done)
2816 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
2817 SUBST (*split, newdest);
2820 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2822 /* recog_for_combine might have added CLOBBERs to newi2pat.
2823 Make sure NEWPAT does not depend on the clobbered regs. */
2824 if (GET_CODE (newi2pat) == PARALLEL)
2825 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
2826 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
2828 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
2829 if (reg_overlap_mentioned_p (reg, newpat))
2831 undo_all ();
2832 return 0;
2836 /* If the split point was a MULT and we didn't have one before,
2837 don't use one now. */
2838 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
2839 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2843 /* Check for a case where we loaded from memory in a narrow mode and
2844 then sign extended it, but we need both registers. In that case,
2845 we have a PARALLEL with both loads from the same memory location.
2846 We can split this into a load from memory followed by a register-register
2847 copy. This saves at least one insn, more if register allocation can
2848 eliminate the copy.
2850 We cannot do this if the destination of the first assignment is a
2851 condition code register or cc0. We eliminate this case by making sure
2852 the SET_DEST and SET_SRC have the same mode.
2854 We cannot do this if the destination of the second assignment is
2855 a register that we have already assumed is zero-extended. Similarly
2856 for a SUBREG of such a register. */
2858 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2859 && GET_CODE (newpat) == PARALLEL
2860 && XVECLEN (newpat, 0) == 2
2861 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2862 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
2863 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
2864 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
2865 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2866 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2867 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
2868 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2869 INSN_CUID (i2))
2870 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2871 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2872 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
2873 (REG_P (temp)
2874 && reg_stat[REGNO (temp)].nonzero_bits != 0
2875 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2876 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2877 && (reg_stat[REGNO (temp)].nonzero_bits
2878 != GET_MODE_MASK (word_mode))))
2879 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
2880 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
2881 (REG_P (temp)
2882 && reg_stat[REGNO (temp)].nonzero_bits != 0
2883 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2884 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2885 && (reg_stat[REGNO (temp)].nonzero_bits
2886 != GET_MODE_MASK (word_mode)))))
2887 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2888 SET_SRC (XVECEXP (newpat, 0, 1)))
2889 && ! find_reg_note (i3, REG_UNUSED,
2890 SET_DEST (XVECEXP (newpat, 0, 0))))
2892 rtx ni2dest;
2894 newi2pat = XVECEXP (newpat, 0, 0);
2895 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
2896 newpat = XVECEXP (newpat, 0, 1);
2897 SUBST (SET_SRC (newpat),
2898 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
2899 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2901 if (i2_code_number >= 0)
2902 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2904 if (insn_code_number >= 0)
2905 swap_i2i3 = 1;
2908 /* Similarly, check for a case where we have a PARALLEL of two independent
2909 SETs but we started with three insns. In this case, we can do the sets
2910 as two separate insns. This case occurs when some SET allows two
2911 other insns to combine, but the destination of that SET is still live. */
2913 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2914 && GET_CODE (newpat) == PARALLEL
2915 && XVECLEN (newpat, 0) == 2
2916 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2917 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
2918 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
2919 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2920 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2921 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2922 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2923 INSN_CUID (i2))
2924 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2925 XVECEXP (newpat, 0, 0))
2926 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
2927 XVECEXP (newpat, 0, 1))
2928 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
2929 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1))))
2930 #ifdef HAVE_cc0
2931 /* We cannot split the parallel into two sets if both sets
2932 reference cc0. */
2933 && ! (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0))
2934 && reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 1)))
2935 #endif
2938 /* Normally, it doesn't matter which of the two is done first,
2939 but it does if one references cc0. In that case, it has to
2940 be first. */
2941 #ifdef HAVE_cc0
2942 if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0)))
2944 newi2pat = XVECEXP (newpat, 0, 0);
2945 newpat = XVECEXP (newpat, 0, 1);
2947 else
2948 #endif
2950 newi2pat = XVECEXP (newpat, 0, 1);
2951 newpat = XVECEXP (newpat, 0, 0);
2954 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2956 if (i2_code_number >= 0)
2957 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2960 /* If it still isn't recognized, fail and change things back the way they
2961 were. */
2962 if ((insn_code_number < 0
2963 /* Is the result a reasonable ASM_OPERANDS? */
2964 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
2966 undo_all ();
2967 return 0;
2970 /* If we had to change another insn, make sure it is valid also. */
2971 if (undobuf.other_insn)
2973 rtx other_pat = PATTERN (undobuf.other_insn);
2974 rtx new_other_notes;
2975 rtx note, next;
2977 CLEAR_HARD_REG_SET (newpat_used_regs);
2979 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
2980 &new_other_notes);
2982 if (other_code_number < 0 && ! check_asm_operands (other_pat))
2984 undo_all ();
2985 return 0;
2988 PATTERN (undobuf.other_insn) = other_pat;
2990 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
2991 are still valid. Then add any non-duplicate notes added by
2992 recog_for_combine. */
2993 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
2995 next = XEXP (note, 1);
2997 if (REG_NOTE_KIND (note) == REG_UNUSED
2998 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
3000 if (REG_P (XEXP (note, 0)))
3001 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
3003 remove_note (undobuf.other_insn, note);
3007 for (note = new_other_notes; note; note = XEXP (note, 1))
3008 if (REG_P (XEXP (note, 0)))
3009 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
3011 distribute_notes (new_other_notes, undobuf.other_insn,
3012 undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX);
3014 #ifdef HAVE_cc0
3015 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
3016 they are adjacent to each other or not. */
3018 rtx p = prev_nonnote_insn (i3);
3019 if (p && p != i2 && NONJUMP_INSN_P (p) && newi2pat
3020 && sets_cc0_p (newi2pat))
3022 undo_all ();
3023 return 0;
3026 #endif
3028 /* Only allow this combination if insn_rtx_costs reports that the
3029 replacement instructions are cheaper than the originals. */
3030 if (!combine_validate_cost (i1, i2, i3, newpat, newi2pat))
3032 undo_all ();
3033 return 0;
3036 /* We now know that we can do this combination. Merge the insns and
3037 update the status of registers and LOG_LINKS. */
3039 if (swap_i2i3)
3041 rtx insn;
3042 rtx link;
3043 rtx ni2dest;
3045 /* I3 now uses what used to be its destination and which is now
3046 I2's destination. This requires us to do a few adjustments. */
3047 PATTERN (i3) = newpat;
3048 adjust_for_new_dest (i3);
3050 /* We need a LOG_LINK from I3 to I2. But we used to have one,
3051 so we still will.
3053 However, some later insn might be using I2's dest and have
3054 a LOG_LINK pointing at I3. We must remove this link.
3055 The simplest way to remove the link is to point it at I1,
3056 which we know will be a NOTE. */
3058 /* newi2pat is usually a SET here; however, recog_for_combine might
3059 have added some clobbers. */
3060 if (GET_CODE (newi2pat) == PARALLEL)
3061 ni2dest = SET_DEST (XVECEXP (newi2pat, 0, 0));
3062 else
3063 ni2dest = SET_DEST (newi2pat);
3065 for (insn = NEXT_INSN (i3);
3066 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3067 || insn != BB_HEAD (this_basic_block->next_bb));
3068 insn = NEXT_INSN (insn))
3070 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
3072 for (link = LOG_LINKS (insn); link;
3073 link = XEXP (link, 1))
3074 if (XEXP (link, 0) == i3)
3075 XEXP (link, 0) = i1;
3077 break;
3083 rtx i3notes, i2notes, i1notes = 0;
3084 rtx i3links, i2links, i1links = 0;
3085 rtx midnotes = 0;
3086 unsigned int regno;
3087 /* Compute which registers we expect to eliminate. newi2pat may be setting
3088 either i3dest or i2dest, so we must check it. Also, i1dest may be the
3089 same as i3dest, in which case newi2pat may be setting i1dest. */
3090 rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat))
3091 || i2dest_in_i2src || i2dest_in_i1src
3092 || !i2dest_killed
3093 ? 0 : i2dest);
3094 rtx elim_i1 = (i1 == 0 || i1dest_in_i1src
3095 || (newi2pat && reg_set_p (i1dest, newi2pat))
3096 || !i1dest_killed
3097 ? 0 : i1dest);
3099 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
3100 clear them. */
3101 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
3102 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
3103 if (i1)
3104 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
3106 /* Ensure that we do not have something that should not be shared but
3107 occurs multiple times in the new insns. Check this by first
3108 resetting all the `used' flags and then copying anything is shared. */
3110 reset_used_flags (i3notes);
3111 reset_used_flags (i2notes);
3112 reset_used_flags (i1notes);
3113 reset_used_flags (newpat);
3114 reset_used_flags (newi2pat);
3115 if (undobuf.other_insn)
3116 reset_used_flags (PATTERN (undobuf.other_insn));
3118 i3notes = copy_rtx_if_shared (i3notes);
3119 i2notes = copy_rtx_if_shared (i2notes);
3120 i1notes = copy_rtx_if_shared (i1notes);
3121 newpat = copy_rtx_if_shared (newpat);
3122 newi2pat = copy_rtx_if_shared (newi2pat);
3123 if (undobuf.other_insn)
3124 reset_used_flags (PATTERN (undobuf.other_insn));
3126 INSN_CODE (i3) = insn_code_number;
3127 PATTERN (i3) = newpat;
3129 if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3))
3131 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
3133 reset_used_flags (call_usage);
3134 call_usage = copy_rtx (call_usage);
3136 if (substed_i2)
3137 replace_rtx (call_usage, i2dest, i2src);
3139 if (substed_i1)
3140 replace_rtx (call_usage, i1dest, i1src);
3142 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
3145 if (undobuf.other_insn)
3146 INSN_CODE (undobuf.other_insn) = other_code_number;
3148 /* We had one special case above where I2 had more than one set and
3149 we replaced a destination of one of those sets with the destination
3150 of I3. In that case, we have to update LOG_LINKS of insns later
3151 in this basic block. Note that this (expensive) case is rare.
3153 Also, in this case, we must pretend that all REG_NOTEs for I2
3154 actually came from I3, so that REG_UNUSED notes from I2 will be
3155 properly handled. */
3157 if (i3_subst_into_i2)
3159 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
3160 if ((GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == SET
3161 || GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == CLOBBER)
3162 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i)))
3163 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
3164 && ! find_reg_note (i2, REG_UNUSED,
3165 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
3166 for (temp = NEXT_INSN (i2);
3167 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3168 || BB_HEAD (this_basic_block) != temp);
3169 temp = NEXT_INSN (temp))
3170 if (temp != i3 && INSN_P (temp))
3171 for (link = LOG_LINKS (temp); link; link = XEXP (link, 1))
3172 if (XEXP (link, 0) == i2)
3173 XEXP (link, 0) = i3;
3175 if (i3notes)
3177 rtx link = i3notes;
3178 while (XEXP (link, 1))
3179 link = XEXP (link, 1);
3180 XEXP (link, 1) = i2notes;
3182 else
3183 i3notes = i2notes;
3184 i2notes = 0;
3187 LOG_LINKS (i3) = 0;
3188 REG_NOTES (i3) = 0;
3189 LOG_LINKS (i2) = 0;
3190 REG_NOTES (i2) = 0;
3192 if (newi2pat)
3194 INSN_CODE (i2) = i2_code_number;
3195 PATTERN (i2) = newi2pat;
3197 else
3198 SET_INSN_DELETED (i2);
3200 if (i1)
3202 LOG_LINKS (i1) = 0;
3203 REG_NOTES (i1) = 0;
3204 SET_INSN_DELETED (i1);
3207 /* Get death notes for everything that is now used in either I3 or
3208 I2 and used to die in a previous insn. If we built two new
3209 patterns, move from I1 to I2 then I2 to I3 so that we get the
3210 proper movement on registers that I2 modifies. */
3212 if (newi2pat)
3214 move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes);
3215 move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes);
3217 else
3218 move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2),
3219 i3, &midnotes);
3221 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
3222 if (i3notes)
3223 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX,
3224 elim_i2, elim_i1);
3225 if (i2notes)
3226 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX,
3227 elim_i2, elim_i1);
3228 if (i1notes)
3229 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX,
3230 elim_i2, elim_i1);
3231 if (midnotes)
3232 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3233 elim_i2, elim_i1);
3235 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
3236 know these are REG_UNUSED and want them to go to the desired insn,
3237 so we always pass it as i3. We have not counted the notes in
3238 reg_n_deaths yet, so we need to do so now. */
3240 if (newi2pat && new_i2_notes)
3242 for (temp = new_i2_notes; temp; temp = XEXP (temp, 1))
3243 if (REG_P (XEXP (temp, 0)))
3244 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3246 distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3249 if (new_i3_notes)
3251 for (temp = new_i3_notes; temp; temp = XEXP (temp, 1))
3252 if (REG_P (XEXP (temp, 0)))
3253 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3255 distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX);
3258 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
3259 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
3260 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
3261 in that case, it might delete I2. Similarly for I2 and I1.
3262 Show an additional death due to the REG_DEAD note we make here. If
3263 we discard it in distribute_notes, we will decrement it again. */
3265 if (i3dest_killed)
3267 if (REG_P (i3dest_killed))
3268 REG_N_DEATHS (REGNO (i3dest_killed))++;
3270 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
3271 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3272 NULL_RTX),
3273 NULL_RTX, i2, NULL_RTX, elim_i2, elim_i1);
3274 else
3275 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3276 NULL_RTX),
3277 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3278 elim_i2, elim_i1);
3281 if (i2dest_in_i2src)
3283 if (REG_P (i2dest))
3284 REG_N_DEATHS (REGNO (i2dest))++;
3286 if (newi2pat && reg_set_p (i2dest, newi2pat))
3287 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3288 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3289 else
3290 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3291 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3292 NULL_RTX, NULL_RTX);
3295 if (i1dest_in_i1src)
3297 if (REG_P (i1dest))
3298 REG_N_DEATHS (REGNO (i1dest))++;
3300 if (newi2pat && reg_set_p (i1dest, newi2pat))
3301 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3302 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3303 else
3304 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3305 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3306 NULL_RTX, NULL_RTX);
3309 distribute_links (i3links);
3310 distribute_links (i2links);
3311 distribute_links (i1links);
3313 if (REG_P (i2dest))
3315 rtx link;
3316 rtx i2_insn = 0, i2_val = 0, set;
3318 /* The insn that used to set this register doesn't exist, and
3319 this life of the register may not exist either. See if one of
3320 I3's links points to an insn that sets I2DEST. If it does,
3321 that is now the last known value for I2DEST. If we don't update
3322 this and I2 set the register to a value that depended on its old
3323 contents, we will get confused. If this insn is used, thing
3324 will be set correctly in combine_instructions. */
3326 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3327 if ((set = single_set (XEXP (link, 0))) != 0
3328 && rtx_equal_p (i2dest, SET_DEST (set)))
3329 i2_insn = XEXP (link, 0), i2_val = SET_SRC (set);
3331 record_value_for_reg (i2dest, i2_insn, i2_val);
3333 /* If the reg formerly set in I2 died only once and that was in I3,
3334 zero its use count so it won't make `reload' do any work. */
3335 if (! added_sets_2
3336 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
3337 && ! i2dest_in_i2src)
3339 regno = REGNO (i2dest);
3340 REG_N_SETS (regno)--;
3344 if (i1 && REG_P (i1dest))
3346 rtx link;
3347 rtx i1_insn = 0, i1_val = 0, set;
3349 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3350 if ((set = single_set (XEXP (link, 0))) != 0
3351 && rtx_equal_p (i1dest, SET_DEST (set)))
3352 i1_insn = XEXP (link, 0), i1_val = SET_SRC (set);
3354 record_value_for_reg (i1dest, i1_insn, i1_val);
3356 regno = REGNO (i1dest);
3357 if (! added_sets_1 && ! i1dest_in_i1src)
3358 REG_N_SETS (regno)--;
3361 /* Update reg_stat[].nonzero_bits et al for any changes that may have
3362 been made to this insn. The order of
3363 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
3364 can affect nonzero_bits of newpat */
3365 if (newi2pat)
3366 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
3367 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
3369 /* Set new_direct_jump_p if a new return or simple jump instruction
3370 has been created.
3372 If I3 is now an unconditional jump, ensure that it has a
3373 BARRIER following it since it may have initially been a
3374 conditional jump. It may also be the last nonnote insn. */
3376 if (returnjump_p (i3) || any_uncondjump_p (i3))
3378 *new_direct_jump_p = 1;
3379 mark_jump_label (PATTERN (i3), i3, 0);
3381 if ((temp = next_nonnote_insn (i3)) == NULL_RTX
3382 || !BARRIER_P (temp))
3383 emit_barrier_after (i3);
3386 if (undobuf.other_insn != NULL_RTX
3387 && (returnjump_p (undobuf.other_insn)
3388 || any_uncondjump_p (undobuf.other_insn)))
3390 *new_direct_jump_p = 1;
3392 if ((temp = next_nonnote_insn (undobuf.other_insn)) == NULL_RTX
3393 || !BARRIER_P (temp))
3394 emit_barrier_after (undobuf.other_insn);
3397 /* An NOOP jump does not need barrier, but it does need cleaning up
3398 of CFG. */
3399 if (GET_CODE (newpat) == SET
3400 && SET_SRC (newpat) == pc_rtx
3401 && SET_DEST (newpat) == pc_rtx)
3402 *new_direct_jump_p = 1;
3405 combine_successes++;
3406 undo_commit ();
3408 if (added_links_insn
3409 && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2))
3410 && INSN_CUID (added_links_insn) < INSN_CUID (i3))
3411 return added_links_insn;
3412 else
3413 return newi2pat ? i2 : i3;
3416 /* Undo all the modifications recorded in undobuf. */
3418 static void
3419 undo_all (void)
3421 struct undo *undo, *next;
3423 for (undo = undobuf.undos; undo; undo = next)
3425 next = undo->next;
3426 switch (undo->kind)
3428 case UNDO_RTX:
3429 *undo->where.r = undo->old_contents.r;
3430 break;
3431 case UNDO_INT:
3432 *undo->where.i = undo->old_contents.i;
3433 break;
3434 case UNDO_MODE:
3435 PUT_MODE (*undo->where.r, undo->old_contents.m);
3436 break;
3437 default:
3438 gcc_unreachable ();
3441 undo->next = undobuf.frees;
3442 undobuf.frees = undo;
3445 undobuf.undos = 0;
3448 /* We've committed to accepting the changes we made. Move all
3449 of the undos to the free list. */
3451 static void
3452 undo_commit (void)
3454 struct undo *undo, *next;
3456 for (undo = undobuf.undos; undo; undo = next)
3458 next = undo->next;
3459 undo->next = undobuf.frees;
3460 undobuf.frees = undo;
3462 undobuf.undos = 0;
3465 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
3466 where we have an arithmetic expression and return that point. LOC will
3467 be inside INSN.
3469 try_combine will call this function to see if an insn can be split into
3470 two insns. */
3472 static rtx *
3473 find_split_point (rtx *loc, rtx insn)
3475 rtx x = *loc;
3476 enum rtx_code code = GET_CODE (x);
3477 rtx *split;
3478 unsigned HOST_WIDE_INT len = 0;
3479 HOST_WIDE_INT pos = 0;
3480 int unsignedp = 0;
3481 rtx inner = NULL_RTX;
3483 /* First special-case some codes. */
3484 switch (code)
3486 case SUBREG:
3487 #ifdef INSN_SCHEDULING
3488 /* If we are making a paradoxical SUBREG invalid, it becomes a split
3489 point. */
3490 if (MEM_P (SUBREG_REG (x)))
3491 return loc;
3492 #endif
3493 return find_split_point (&SUBREG_REG (x), insn);
3495 case MEM:
3496 #ifdef HAVE_lo_sum
3497 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
3498 using LO_SUM and HIGH. */
3499 if (GET_CODE (XEXP (x, 0)) == CONST
3500 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
3502 SUBST (XEXP (x, 0),
3503 gen_rtx_LO_SUM (Pmode,
3504 gen_rtx_HIGH (Pmode, XEXP (x, 0)),
3505 XEXP (x, 0)));
3506 return &XEXP (XEXP (x, 0), 0);
3508 #endif
3510 /* If we have a PLUS whose second operand is a constant and the
3511 address is not valid, perhaps will can split it up using
3512 the machine-specific way to split large constants. We use
3513 the first pseudo-reg (one of the virtual regs) as a placeholder;
3514 it will not remain in the result. */
3515 if (GET_CODE (XEXP (x, 0)) == PLUS
3516 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3517 && ! memory_address_p (GET_MODE (x), XEXP (x, 0)))
3519 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
3520 rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)),
3521 subst_insn);
3523 /* This should have produced two insns, each of which sets our
3524 placeholder. If the source of the second is a valid address,
3525 we can make put both sources together and make a split point
3526 in the middle. */
3528 if (seq
3529 && NEXT_INSN (seq) != NULL_RTX
3530 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
3531 && NONJUMP_INSN_P (seq)
3532 && GET_CODE (PATTERN (seq)) == SET
3533 && SET_DEST (PATTERN (seq)) == reg
3534 && ! reg_mentioned_p (reg,
3535 SET_SRC (PATTERN (seq)))
3536 && NONJUMP_INSN_P (NEXT_INSN (seq))
3537 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
3538 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
3539 && memory_address_p (GET_MODE (x),
3540 SET_SRC (PATTERN (NEXT_INSN (seq)))))
3542 rtx src1 = SET_SRC (PATTERN (seq));
3543 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
3545 /* Replace the placeholder in SRC2 with SRC1. If we can
3546 find where in SRC2 it was placed, that can become our
3547 split point and we can replace this address with SRC2.
3548 Just try two obvious places. */
3550 src2 = replace_rtx (src2, reg, src1);
3551 split = 0;
3552 if (XEXP (src2, 0) == src1)
3553 split = &XEXP (src2, 0);
3554 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
3555 && XEXP (XEXP (src2, 0), 0) == src1)
3556 split = &XEXP (XEXP (src2, 0), 0);
3558 if (split)
3560 SUBST (XEXP (x, 0), src2);
3561 return split;
3565 /* If that didn't work, perhaps the first operand is complex and
3566 needs to be computed separately, so make a split point there.
3567 This will occur on machines that just support REG + CONST
3568 and have a constant moved through some previous computation. */
3570 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
3571 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
3572 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
3573 return &XEXP (XEXP (x, 0), 0);
3575 break;
3577 case SET:
3578 #ifdef HAVE_cc0
3579 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
3580 ZERO_EXTRACT, the most likely reason why this doesn't match is that
3581 we need to put the operand into a register. So split at that
3582 point. */
3584 if (SET_DEST (x) == cc0_rtx
3585 && GET_CODE (SET_SRC (x)) != COMPARE
3586 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
3587 && !OBJECT_P (SET_SRC (x))
3588 && ! (GET_CODE (SET_SRC (x)) == SUBREG
3589 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
3590 return &SET_SRC (x);
3591 #endif
3593 /* See if we can split SET_SRC as it stands. */
3594 split = find_split_point (&SET_SRC (x), insn);
3595 if (split && split != &SET_SRC (x))
3596 return split;
3598 /* See if we can split SET_DEST as it stands. */
3599 split = find_split_point (&SET_DEST (x), insn);
3600 if (split && split != &SET_DEST (x))
3601 return split;
3603 /* See if this is a bitfield assignment with everything constant. If
3604 so, this is an IOR of an AND, so split it into that. */
3605 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
3606 && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))
3607 <= HOST_BITS_PER_WIDE_INT)
3608 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT
3609 && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT
3610 && GET_CODE (SET_SRC (x)) == CONST_INT
3611 && ((INTVAL (XEXP (SET_DEST (x), 1))
3612 + INTVAL (XEXP (SET_DEST (x), 2)))
3613 <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))))
3614 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
3616 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
3617 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
3618 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
3619 rtx dest = XEXP (SET_DEST (x), 0);
3620 enum machine_mode mode = GET_MODE (dest);
3621 unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1;
3622 rtx or_mask;
3624 if (BITS_BIG_ENDIAN)
3625 pos = GET_MODE_BITSIZE (mode) - len - pos;
3627 or_mask = gen_int_mode (src << pos, mode);
3628 if (src == mask)
3629 SUBST (SET_SRC (x),
3630 simplify_gen_binary (IOR, mode, dest, or_mask));
3631 else
3633 rtx negmask = gen_int_mode (~(mask << pos), mode);
3634 SUBST (SET_SRC (x),
3635 simplify_gen_binary (IOR, mode,
3636 simplify_gen_binary (AND, mode,
3637 dest, negmask),
3638 or_mask));
3641 SUBST (SET_DEST (x), dest);
3643 split = find_split_point (&SET_SRC (x), insn);
3644 if (split && split != &SET_SRC (x))
3645 return split;
3648 /* Otherwise, see if this is an operation that we can split into two.
3649 If so, try to split that. */
3650 code = GET_CODE (SET_SRC (x));
3652 switch (code)
3654 case AND:
3655 /* If we are AND'ing with a large constant that is only a single
3656 bit and the result is only being used in a context where we
3657 need to know if it is zero or nonzero, replace it with a bit
3658 extraction. This will avoid the large constant, which might
3659 have taken more than one insn to make. If the constant were
3660 not a valid argument to the AND but took only one insn to make,
3661 this is no worse, but if it took more than one insn, it will
3662 be better. */
3664 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3665 && REG_P (XEXP (SET_SRC (x), 0))
3666 && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7
3667 && REG_P (SET_DEST (x))
3668 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
3669 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
3670 && XEXP (*split, 0) == SET_DEST (x)
3671 && XEXP (*split, 1) == const0_rtx)
3673 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
3674 XEXP (SET_SRC (x), 0),
3675 pos, NULL_RTX, 1, 1, 0, 0);
3676 if (extraction != 0)
3678 SUBST (SET_SRC (x), extraction);
3679 return find_split_point (loc, insn);
3682 break;
3684 case NE:
3685 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
3686 is known to be on, this can be converted into a NEG of a shift. */
3687 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
3688 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
3689 && 1 <= (pos = exact_log2
3690 (nonzero_bits (XEXP (SET_SRC (x), 0),
3691 GET_MODE (XEXP (SET_SRC (x), 0))))))
3693 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
3695 SUBST (SET_SRC (x),
3696 gen_rtx_NEG (mode,
3697 gen_rtx_LSHIFTRT (mode,
3698 XEXP (SET_SRC (x), 0),
3699 GEN_INT (pos))));
3701 split = find_split_point (&SET_SRC (x), insn);
3702 if (split && split != &SET_SRC (x))
3703 return split;
3705 break;
3707 case SIGN_EXTEND:
3708 inner = XEXP (SET_SRC (x), 0);
3710 /* We can't optimize if either mode is a partial integer
3711 mode as we don't know how many bits are significant
3712 in those modes. */
3713 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
3714 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
3715 break;
3717 pos = 0;
3718 len = GET_MODE_BITSIZE (GET_MODE (inner));
3719 unsignedp = 0;
3720 break;
3722 case SIGN_EXTRACT:
3723 case ZERO_EXTRACT:
3724 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3725 && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT)
3727 inner = XEXP (SET_SRC (x), 0);
3728 len = INTVAL (XEXP (SET_SRC (x), 1));
3729 pos = INTVAL (XEXP (SET_SRC (x), 2));
3731 if (BITS_BIG_ENDIAN)
3732 pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos;
3733 unsignedp = (code == ZERO_EXTRACT);
3735 break;
3737 default:
3738 break;
3741 if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner)))
3743 enum machine_mode mode = GET_MODE (SET_SRC (x));
3745 /* For unsigned, we have a choice of a shift followed by an
3746 AND or two shifts. Use two shifts for field sizes where the
3747 constant might be too large. We assume here that we can
3748 always at least get 8-bit constants in an AND insn, which is
3749 true for every current RISC. */
3751 if (unsignedp && len <= 8)
3753 SUBST (SET_SRC (x),
3754 gen_rtx_AND (mode,
3755 gen_rtx_LSHIFTRT
3756 (mode, gen_lowpart (mode, inner),
3757 GEN_INT (pos)),
3758 GEN_INT (((HOST_WIDE_INT) 1 << len) - 1)));
3760 split = find_split_point (&SET_SRC (x), insn);
3761 if (split && split != &SET_SRC (x))
3762 return split;
3764 else
3766 SUBST (SET_SRC (x),
3767 gen_rtx_fmt_ee
3768 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
3769 gen_rtx_ASHIFT (mode,
3770 gen_lowpart (mode, inner),
3771 GEN_INT (GET_MODE_BITSIZE (mode)
3772 - len - pos)),
3773 GEN_INT (GET_MODE_BITSIZE (mode) - len)));
3775 split = find_split_point (&SET_SRC (x), insn);
3776 if (split && split != &SET_SRC (x))
3777 return split;
3781 /* See if this is a simple operation with a constant as the second
3782 operand. It might be that this constant is out of range and hence
3783 could be used as a split point. */
3784 if (BINARY_P (SET_SRC (x))
3785 && CONSTANT_P (XEXP (SET_SRC (x), 1))
3786 && (OBJECT_P (XEXP (SET_SRC (x), 0))
3787 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
3788 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
3789 return &XEXP (SET_SRC (x), 1);
3791 /* Finally, see if this is a simple operation with its first operand
3792 not in a register. The operation might require this operand in a
3793 register, so return it as a split point. We can always do this
3794 because if the first operand were another operation, we would have
3795 already found it as a split point. */
3796 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
3797 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
3798 return &XEXP (SET_SRC (x), 0);
3800 return 0;
3802 case AND:
3803 case IOR:
3804 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
3805 it is better to write this as (not (ior A B)) so we can split it.
3806 Similarly for IOR. */
3807 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
3809 SUBST (*loc,
3810 gen_rtx_NOT (GET_MODE (x),
3811 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
3812 GET_MODE (x),
3813 XEXP (XEXP (x, 0), 0),
3814 XEXP (XEXP (x, 1), 0))));
3815 return find_split_point (loc, insn);
3818 /* Many RISC machines have a large set of logical insns. If the
3819 second operand is a NOT, put it first so we will try to split the
3820 other operand first. */
3821 if (GET_CODE (XEXP (x, 1)) == NOT)
3823 rtx tem = XEXP (x, 0);
3824 SUBST (XEXP (x, 0), XEXP (x, 1));
3825 SUBST (XEXP (x, 1), tem);
3827 break;
3829 default:
3830 break;
3833 /* Otherwise, select our actions depending on our rtx class. */
3834 switch (GET_RTX_CLASS (code))
3836 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
3837 case RTX_TERNARY:
3838 split = find_split_point (&XEXP (x, 2), insn);
3839 if (split)
3840 return split;
3841 /* ... fall through ... */
3842 case RTX_BIN_ARITH:
3843 case RTX_COMM_ARITH:
3844 case RTX_COMPARE:
3845 case RTX_COMM_COMPARE:
3846 split = find_split_point (&XEXP (x, 1), insn);
3847 if (split)
3848 return split;
3849 /* ... fall through ... */
3850 case RTX_UNARY:
3851 /* Some machines have (and (shift ...) ...) insns. If X is not
3852 an AND, but XEXP (X, 0) is, use it as our split point. */
3853 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
3854 return &XEXP (x, 0);
3856 split = find_split_point (&XEXP (x, 0), insn);
3857 if (split)
3858 return split;
3859 return loc;
3861 default:
3862 /* Otherwise, we don't have a split point. */
3863 return 0;
3867 /* Throughout X, replace FROM with TO, and return the result.
3868 The result is TO if X is FROM;
3869 otherwise the result is X, but its contents may have been modified.
3870 If they were modified, a record was made in undobuf so that
3871 undo_all will (among other things) return X to its original state.
3873 If the number of changes necessary is too much to record to undo,
3874 the excess changes are not made, so the result is invalid.
3875 The changes already made can still be undone.
3876 undobuf.num_undo is incremented for such changes, so by testing that
3877 the caller can tell whether the result is valid.
3879 `n_occurrences' is incremented each time FROM is replaced.
3881 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
3883 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
3884 by copying if `n_occurrences' is nonzero. */
3886 static rtx
3887 subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy)
3889 enum rtx_code code = GET_CODE (x);
3890 enum machine_mode op0_mode = VOIDmode;
3891 const char *fmt;
3892 int len, i;
3893 rtx new;
3895 /* Two expressions are equal if they are identical copies of a shared
3896 RTX or if they are both registers with the same register number
3897 and mode. */
3899 #define COMBINE_RTX_EQUAL_P(X,Y) \
3900 ((X) == (Y) \
3901 || (REG_P (X) && REG_P (Y) \
3902 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
3904 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
3906 n_occurrences++;
3907 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
3910 /* If X and FROM are the same register but different modes, they will
3911 not have been seen as equal above. However, flow.c will make a
3912 LOG_LINKS entry for that case. If we do nothing, we will try to
3913 rerecognize our original insn and, when it succeeds, we will
3914 delete the feeding insn, which is incorrect.
3916 So force this insn not to match in this (rare) case. */
3917 if (! in_dest && code == REG && REG_P (from)
3918 && REGNO (x) == REGNO (from))
3919 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
3921 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
3922 of which may contain things that can be combined. */
3923 if (code != MEM && code != LO_SUM && OBJECT_P (x))
3924 return x;
3926 /* It is possible to have a subexpression appear twice in the insn.
3927 Suppose that FROM is a register that appears within TO.
3928 Then, after that subexpression has been scanned once by `subst',
3929 the second time it is scanned, TO may be found. If we were
3930 to scan TO here, we would find FROM within it and create a
3931 self-referent rtl structure which is completely wrong. */
3932 if (COMBINE_RTX_EQUAL_P (x, to))
3933 return to;
3935 /* Parallel asm_operands need special attention because all of the
3936 inputs are shared across the arms. Furthermore, unsharing the
3937 rtl results in recognition failures. Failure to handle this case
3938 specially can result in circular rtl.
3940 Solve this by doing a normal pass across the first entry of the
3941 parallel, and only processing the SET_DESTs of the subsequent
3942 entries. Ug. */
3944 if (code == PARALLEL
3945 && GET_CODE (XVECEXP (x, 0, 0)) == SET
3946 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
3948 new = subst (XVECEXP (x, 0, 0), from, to, 0, unique_copy);
3950 /* If this substitution failed, this whole thing fails. */
3951 if (GET_CODE (new) == CLOBBER
3952 && XEXP (new, 0) == const0_rtx)
3953 return new;
3955 SUBST (XVECEXP (x, 0, 0), new);
3957 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
3959 rtx dest = SET_DEST (XVECEXP (x, 0, i));
3961 if (!REG_P (dest)
3962 && GET_CODE (dest) != CC0
3963 && GET_CODE (dest) != PC)
3965 new = subst (dest, from, to, 0, unique_copy);
3967 /* If this substitution failed, this whole thing fails. */
3968 if (GET_CODE (new) == CLOBBER
3969 && XEXP (new, 0) == const0_rtx)
3970 return new;
3972 SUBST (SET_DEST (XVECEXP (x, 0, i)), new);
3976 else
3978 len = GET_RTX_LENGTH (code);
3979 fmt = GET_RTX_FORMAT (code);
3981 /* We don't need to process a SET_DEST that is a register, CC0,
3982 or PC, so set up to skip this common case. All other cases
3983 where we want to suppress replacing something inside a
3984 SET_SRC are handled via the IN_DEST operand. */
3985 if (code == SET
3986 && (REG_P (SET_DEST (x))
3987 || GET_CODE (SET_DEST (x)) == CC0
3988 || GET_CODE (SET_DEST (x)) == PC))
3989 fmt = "ie";
3991 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
3992 constant. */
3993 if (fmt[0] == 'e')
3994 op0_mode = GET_MODE (XEXP (x, 0));
3996 for (i = 0; i < len; i++)
3998 if (fmt[i] == 'E')
4000 int j;
4001 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4003 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
4005 new = (unique_copy && n_occurrences
4006 ? copy_rtx (to) : to);
4007 n_occurrences++;
4009 else
4011 new = subst (XVECEXP (x, i, j), from, to, 0,
4012 unique_copy);
4014 /* If this substitution failed, this whole thing
4015 fails. */
4016 if (GET_CODE (new) == CLOBBER
4017 && XEXP (new, 0) == const0_rtx)
4018 return new;
4021 SUBST (XVECEXP (x, i, j), new);
4024 else if (fmt[i] == 'e')
4026 /* If this is a register being set, ignore it. */
4027 new = XEXP (x, i);
4028 if (in_dest
4029 && i == 0
4030 && (((code == SUBREG || code == ZERO_EXTRACT)
4031 && REG_P (new))
4032 || code == STRICT_LOW_PART))
4035 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
4037 /* In general, don't install a subreg involving two
4038 modes not tieable. It can worsen register
4039 allocation, and can even make invalid reload
4040 insns, since the reg inside may need to be copied
4041 from in the outside mode, and that may be invalid
4042 if it is an fp reg copied in integer mode.
4044 We allow two exceptions to this: It is valid if
4045 it is inside another SUBREG and the mode of that
4046 SUBREG and the mode of the inside of TO is
4047 tieable and it is valid if X is a SET that copies
4048 FROM to CC0. */
4050 if (GET_CODE (to) == SUBREG
4051 && ! MODES_TIEABLE_P (GET_MODE (to),
4052 GET_MODE (SUBREG_REG (to)))
4053 && ! (code == SUBREG
4054 && MODES_TIEABLE_P (GET_MODE (x),
4055 GET_MODE (SUBREG_REG (to))))
4056 #ifdef HAVE_cc0
4057 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
4058 #endif
4060 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4062 #ifdef CANNOT_CHANGE_MODE_CLASS
4063 if (code == SUBREG
4064 && REG_P (to)
4065 && REGNO (to) < FIRST_PSEUDO_REGISTER
4066 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
4067 GET_MODE (to),
4068 GET_MODE (x)))
4069 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4070 #endif
4072 new = (unique_copy && n_occurrences ? copy_rtx (to) : to);
4073 n_occurrences++;
4075 else
4076 /* If we are in a SET_DEST, suppress most cases unless we
4077 have gone inside a MEM, in which case we want to
4078 simplify the address. We assume here that things that
4079 are actually part of the destination have their inner
4080 parts in the first expression. This is true for SUBREG,
4081 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
4082 things aside from REG and MEM that should appear in a
4083 SET_DEST. */
4084 new = subst (XEXP (x, i), from, to,
4085 (((in_dest
4086 && (code == SUBREG || code == STRICT_LOW_PART
4087 || code == ZERO_EXTRACT))
4088 || code == SET)
4089 && i == 0), unique_copy);
4091 /* If we found that we will have to reject this combination,
4092 indicate that by returning the CLOBBER ourselves, rather than
4093 an expression containing it. This will speed things up as
4094 well as prevent accidents where two CLOBBERs are considered
4095 to be equal, thus producing an incorrect simplification. */
4097 if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx)
4098 return new;
4100 if (GET_CODE (x) == SUBREG
4101 && (GET_CODE (new) == CONST_INT
4102 || GET_CODE (new) == CONST_DOUBLE))
4104 enum machine_mode mode = GET_MODE (x);
4106 x = simplify_subreg (GET_MODE (x), new,
4107 GET_MODE (SUBREG_REG (x)),
4108 SUBREG_BYTE (x));
4109 if (! x)
4110 x = gen_rtx_CLOBBER (mode, const0_rtx);
4112 else if (GET_CODE (new) == CONST_INT
4113 && GET_CODE (x) == ZERO_EXTEND)
4115 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
4116 new, GET_MODE (XEXP (x, 0)));
4117 gcc_assert (x);
4119 else
4120 SUBST (XEXP (x, i), new);
4125 /* Try to simplify X. If the simplification changed the code, it is likely
4126 that further simplification will help, so loop, but limit the number
4127 of repetitions that will be performed. */
4129 for (i = 0; i < 4; i++)
4131 /* If X is sufficiently simple, don't bother trying to do anything
4132 with it. */
4133 if (code != CONST_INT && code != REG && code != CLOBBER)
4134 x = combine_simplify_rtx (x, op0_mode, in_dest);
4136 if (GET_CODE (x) == code)
4137 break;
4139 code = GET_CODE (x);
4141 /* We no longer know the original mode of operand 0 since we
4142 have changed the form of X) */
4143 op0_mode = VOIDmode;
4146 return x;
4149 /* Simplify X, a piece of RTL. We just operate on the expression at the
4150 outer level; call `subst' to simplify recursively. Return the new
4151 expression.
4153 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
4154 if we are inside a SET_DEST. */
4156 static rtx
4157 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest)
4159 enum rtx_code code = GET_CODE (x);
4160 enum machine_mode mode = GET_MODE (x);
4161 rtx temp;
4162 int i;
4164 /* If this is a commutative operation, put a constant last and a complex
4165 expression first. We don't need to do this for comparisons here. */
4166 if (COMMUTATIVE_ARITH_P (x)
4167 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
4169 temp = XEXP (x, 0);
4170 SUBST (XEXP (x, 0), XEXP (x, 1));
4171 SUBST (XEXP (x, 1), temp);
4174 /* If this is a simple operation applied to an IF_THEN_ELSE, try
4175 applying it to the arms of the IF_THEN_ELSE. This often simplifies
4176 things. Check for cases where both arms are testing the same
4177 condition.
4179 Don't do anything if all operands are very simple. */
4181 if ((BINARY_P (x)
4182 && ((!OBJECT_P (XEXP (x, 0))
4183 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4184 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
4185 || (!OBJECT_P (XEXP (x, 1))
4186 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
4187 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
4188 || (UNARY_P (x)
4189 && (!OBJECT_P (XEXP (x, 0))
4190 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4191 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
4193 rtx cond, true_rtx, false_rtx;
4195 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
4196 if (cond != 0
4197 /* If everything is a comparison, what we have is highly unlikely
4198 to be simpler, so don't use it. */
4199 && ! (COMPARISON_P (x)
4200 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
4202 rtx cop1 = const0_rtx;
4203 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
4205 if (cond_code == NE && COMPARISON_P (cond))
4206 return x;
4208 /* Simplify the alternative arms; this may collapse the true and
4209 false arms to store-flag values. Be careful to use copy_rtx
4210 here since true_rtx or false_rtx might share RTL with x as a
4211 result of the if_then_else_cond call above. */
4212 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0);
4213 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0);
4215 /* If true_rtx and false_rtx are not general_operands, an if_then_else
4216 is unlikely to be simpler. */
4217 if (general_operand (true_rtx, VOIDmode)
4218 && general_operand (false_rtx, VOIDmode))
4220 enum rtx_code reversed;
4222 /* Restarting if we generate a store-flag expression will cause
4223 us to loop. Just drop through in this case. */
4225 /* If the result values are STORE_FLAG_VALUE and zero, we can
4226 just make the comparison operation. */
4227 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
4228 x = simplify_gen_relational (cond_code, mode, VOIDmode,
4229 cond, cop1);
4230 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
4231 && ((reversed = reversed_comparison_code_parts
4232 (cond_code, cond, cop1, NULL))
4233 != UNKNOWN))
4234 x = simplify_gen_relational (reversed, mode, VOIDmode,
4235 cond, cop1);
4237 /* Likewise, we can make the negate of a comparison operation
4238 if the result values are - STORE_FLAG_VALUE and zero. */
4239 else if (GET_CODE (true_rtx) == CONST_INT
4240 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
4241 && false_rtx == const0_rtx)
4242 x = simplify_gen_unary (NEG, mode,
4243 simplify_gen_relational (cond_code,
4244 mode, VOIDmode,
4245 cond, cop1),
4246 mode);
4247 else if (GET_CODE (false_rtx) == CONST_INT
4248 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
4249 && true_rtx == const0_rtx
4250 && ((reversed = reversed_comparison_code_parts
4251 (cond_code, cond, cop1, NULL))
4252 != UNKNOWN))
4253 x = simplify_gen_unary (NEG, mode,
4254 simplify_gen_relational (reversed,
4255 mode, VOIDmode,
4256 cond, cop1),
4257 mode);
4258 else
4259 return gen_rtx_IF_THEN_ELSE (mode,
4260 simplify_gen_relational (cond_code,
4261 mode,
4262 VOIDmode,
4263 cond,
4264 cop1),
4265 true_rtx, false_rtx);
4267 code = GET_CODE (x);
4268 op0_mode = VOIDmode;
4273 /* Try to fold this expression in case we have constants that weren't
4274 present before. */
4275 temp = 0;
4276 switch (GET_RTX_CLASS (code))
4278 case RTX_UNARY:
4279 if (op0_mode == VOIDmode)
4280 op0_mode = GET_MODE (XEXP (x, 0));
4281 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
4282 break;
4283 case RTX_COMPARE:
4284 case RTX_COMM_COMPARE:
4286 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
4287 if (cmp_mode == VOIDmode)
4289 cmp_mode = GET_MODE (XEXP (x, 1));
4290 if (cmp_mode == VOIDmode)
4291 cmp_mode = op0_mode;
4293 temp = simplify_relational_operation (code, mode, cmp_mode,
4294 XEXP (x, 0), XEXP (x, 1));
4296 break;
4297 case RTX_COMM_ARITH:
4298 case RTX_BIN_ARITH:
4299 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
4300 break;
4301 case RTX_BITFIELD_OPS:
4302 case RTX_TERNARY:
4303 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
4304 XEXP (x, 1), XEXP (x, 2));
4305 break;
4306 default:
4307 break;
4310 if (temp)
4312 x = temp;
4313 code = GET_CODE (temp);
4314 op0_mode = VOIDmode;
4315 mode = GET_MODE (temp);
4318 /* First see if we can apply the inverse distributive law. */
4319 if (code == PLUS || code == MINUS
4320 || code == AND || code == IOR || code == XOR)
4322 x = apply_distributive_law (x);
4323 code = GET_CODE (x);
4324 op0_mode = VOIDmode;
4327 /* If CODE is an associative operation not otherwise handled, see if we
4328 can associate some operands. This can win if they are constants or
4329 if they are logically related (i.e. (a & b) & a). */
4330 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
4331 || code == AND || code == IOR || code == XOR
4332 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
4333 && ((INTEGRAL_MODE_P (mode) && code != DIV)
4334 || (flag_unsafe_math_optimizations && FLOAT_MODE_P (mode))))
4336 if (GET_CODE (XEXP (x, 0)) == code)
4338 rtx other = XEXP (XEXP (x, 0), 0);
4339 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
4340 rtx inner_op1 = XEXP (x, 1);
4341 rtx inner;
4343 /* Make sure we pass the constant operand if any as the second
4344 one if this is a commutative operation. */
4345 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
4347 rtx tem = inner_op0;
4348 inner_op0 = inner_op1;
4349 inner_op1 = tem;
4351 inner = simplify_binary_operation (code == MINUS ? PLUS
4352 : code == DIV ? MULT
4353 : code,
4354 mode, inner_op0, inner_op1);
4356 /* For commutative operations, try the other pair if that one
4357 didn't simplify. */
4358 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
4360 other = XEXP (XEXP (x, 0), 1);
4361 inner = simplify_binary_operation (code, mode,
4362 XEXP (XEXP (x, 0), 0),
4363 XEXP (x, 1));
4366 if (inner)
4367 return simplify_gen_binary (code, mode, other, inner);
4371 /* A little bit of algebraic simplification here. */
4372 switch (code)
4374 case MEM:
4375 /* Ensure that our address has any ASHIFTs converted to MULT in case
4376 address-recognizing predicates are called later. */
4377 temp = make_compound_operation (XEXP (x, 0), MEM);
4378 SUBST (XEXP (x, 0), temp);
4379 break;
4381 case SUBREG:
4382 if (op0_mode == VOIDmode)
4383 op0_mode = GET_MODE (SUBREG_REG (x));
4385 /* See if this can be moved to simplify_subreg. */
4386 if (CONSTANT_P (SUBREG_REG (x))
4387 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
4388 /* Don't call gen_lowpart if the inner mode
4389 is VOIDmode and we cannot simplify it, as SUBREG without
4390 inner mode is invalid. */
4391 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
4392 || gen_lowpart_common (mode, SUBREG_REG (x))))
4393 return gen_lowpart (mode, SUBREG_REG (x));
4395 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
4396 break;
4398 rtx temp;
4399 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,
4400 SUBREG_BYTE (x));
4401 if (temp)
4402 return temp;
4405 /* Don't change the mode of the MEM if that would change the meaning
4406 of the address. */
4407 if (MEM_P (SUBREG_REG (x))
4408 && (MEM_VOLATILE_P (SUBREG_REG (x))
4409 || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0))))
4410 return gen_rtx_CLOBBER (mode, const0_rtx);
4412 /* Note that we cannot do any narrowing for non-constants since
4413 we might have been counting on using the fact that some bits were
4414 zero. We now do this in the SET. */
4416 break;
4418 case NEG:
4419 temp = expand_compound_operation (XEXP (x, 0));
4421 /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be
4422 replaced by (lshiftrt X C). This will convert
4423 (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */
4425 if (GET_CODE (temp) == ASHIFTRT
4426 && GET_CODE (XEXP (temp, 1)) == CONST_INT
4427 && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1)
4428 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (temp, 0),
4429 INTVAL (XEXP (temp, 1)));
4431 /* If X has only a single bit that might be nonzero, say, bit I, convert
4432 (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of
4433 MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to
4434 (sign_extract X 1 Y). But only do this if TEMP isn't a register
4435 or a SUBREG of one since we'd be making the expression more
4436 complex if it was just a register. */
4438 if (!REG_P (temp)
4439 && ! (GET_CODE (temp) == SUBREG
4440 && REG_P (SUBREG_REG (temp)))
4441 && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0)
4443 rtx temp1 = simplify_shift_const
4444 (NULL_RTX, ASHIFTRT, mode,
4445 simplify_shift_const (NULL_RTX, ASHIFT, mode, temp,
4446 GET_MODE_BITSIZE (mode) - 1 - i),
4447 GET_MODE_BITSIZE (mode) - 1 - i);
4449 /* If all we did was surround TEMP with the two shifts, we
4450 haven't improved anything, so don't use it. Otherwise,
4451 we are better off with TEMP1. */
4452 if (GET_CODE (temp1) != ASHIFTRT
4453 || GET_CODE (XEXP (temp1, 0)) != ASHIFT
4454 || XEXP (XEXP (temp1, 0), 0) != temp)
4455 return temp1;
4457 break;
4459 case TRUNCATE:
4460 /* We can't handle truncation to a partial integer mode here
4461 because we don't know the real bitsize of the partial
4462 integer mode. */
4463 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
4464 break;
4466 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4467 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
4468 GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))))
4469 SUBST (XEXP (x, 0),
4470 force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
4471 GET_MODE_MASK (mode), 0));
4473 /* Similarly to what we do in simplify-rtx.c, a truncate of a register
4474 whose value is a comparison can be replaced with a subreg if
4475 STORE_FLAG_VALUE permits. */
4476 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4477 && ((HOST_WIDE_INT) STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0
4478 && (temp = get_last_value (XEXP (x, 0)))
4479 && COMPARISON_P (temp))
4480 return gen_lowpart (mode, XEXP (x, 0));
4481 break;
4483 #ifdef HAVE_cc0
4484 case COMPARE:
4485 /* Convert (compare FOO (const_int 0)) to FOO unless we aren't
4486 using cc0, in which case we want to leave it as a COMPARE
4487 so we can distinguish it from a register-register-copy. */
4488 if (XEXP (x, 1) == const0_rtx)
4489 return XEXP (x, 0);
4491 /* x - 0 is the same as x unless x's mode has signed zeros and
4492 allows rounding towards -infinity. Under those conditions,
4493 0 - 0 is -0. */
4494 if (!(HONOR_SIGNED_ZEROS (GET_MODE (XEXP (x, 0)))
4495 && HONOR_SIGN_DEPENDENT_ROUNDING (GET_MODE (XEXP (x, 0))))
4496 && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0))))
4497 return XEXP (x, 0);
4498 break;
4499 #endif
4501 case CONST:
4502 /* (const (const X)) can become (const X). Do it this way rather than
4503 returning the inner CONST since CONST can be shared with a
4504 REG_EQUAL note. */
4505 if (GET_CODE (XEXP (x, 0)) == CONST)
4506 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4507 break;
4509 #ifdef HAVE_lo_sum
4510 case LO_SUM:
4511 /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we
4512 can add in an offset. find_split_point will split this address up
4513 again if it doesn't match. */
4514 if (GET_CODE (XEXP (x, 0)) == HIGH
4515 && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
4516 return XEXP (x, 1);
4517 break;
4518 #endif
4520 case PLUS:
4521 /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>)
4522 when c is (const_int (pow2 + 1) / 2) is a sign extension of a
4523 bit-field and can be replaced by either a sign_extend or a
4524 sign_extract. The `and' may be a zero_extend and the two
4525 <c>, -<c> constants may be reversed. */
4526 if (GET_CODE (XEXP (x, 0)) == XOR
4527 && GET_CODE (XEXP (x, 1)) == CONST_INT
4528 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
4529 && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1))
4530 && ((i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0
4531 || (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
4532 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4533 && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND
4534 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
4535 && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))
4536 == ((HOST_WIDE_INT) 1 << (i + 1)) - 1))
4537 || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND
4538 && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))
4539 == (unsigned int) i + 1))))
4540 return simplify_shift_const
4541 (NULL_RTX, ASHIFTRT, mode,
4542 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4543 XEXP (XEXP (XEXP (x, 0), 0), 0),
4544 GET_MODE_BITSIZE (mode) - (i + 1)),
4545 GET_MODE_BITSIZE (mode) - (i + 1));
4547 /* If only the low-order bit of X is possibly nonzero, (plus x -1)
4548 can become (ashiftrt (ashift (xor x 1) C) C) where C is
4549 the bitsize of the mode - 1. This allows simplification of
4550 "a = (b & 8) == 0;" */
4551 if (XEXP (x, 1) == constm1_rtx
4552 && !REG_P (XEXP (x, 0))
4553 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4554 && REG_P (SUBREG_REG (XEXP (x, 0))))
4555 && nonzero_bits (XEXP (x, 0), mode) == 1)
4556 return simplify_shift_const (NULL_RTX, ASHIFTRT, mode,
4557 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4558 gen_rtx_XOR (mode, XEXP (x, 0), const1_rtx),
4559 GET_MODE_BITSIZE (mode) - 1),
4560 GET_MODE_BITSIZE (mode) - 1);
4562 /* If we are adding two things that have no bits in common, convert
4563 the addition into an IOR. This will often be further simplified,
4564 for example in cases like ((a & 1) + (a & 2)), which can
4565 become a & 3. */
4567 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4568 && (nonzero_bits (XEXP (x, 0), mode)
4569 & nonzero_bits (XEXP (x, 1), mode)) == 0)
4571 /* Try to simplify the expression further. */
4572 rtx tor = simplify_gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1));
4573 temp = combine_simplify_rtx (tor, mode, in_dest);
4575 /* If we could, great. If not, do not go ahead with the IOR
4576 replacement, since PLUS appears in many special purpose
4577 address arithmetic instructions. */
4578 if (GET_CODE (temp) != CLOBBER && temp != tor)
4579 return temp;
4581 break;
4583 case MINUS:
4584 /* (minus <foo> (and <foo> (const_int -pow2))) becomes
4585 (and <foo> (const_int pow2-1)) */
4586 if (GET_CODE (XEXP (x, 1)) == AND
4587 && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
4588 && exact_log2 (-INTVAL (XEXP (XEXP (x, 1), 1))) >= 0
4589 && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
4590 return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0),
4591 -INTVAL (XEXP (XEXP (x, 1), 1)) - 1);
4592 break;
4594 case MULT:
4595 /* If we have (mult (plus A B) C), apply the distributive law and then
4596 the inverse distributive law to see if things simplify. This
4597 occurs mostly in addresses, often when unrolling loops. */
4599 if (GET_CODE (XEXP (x, 0)) == PLUS)
4601 rtx result = distribute_and_simplify_rtx (x, 0);
4602 if (result)
4603 return result;
4606 /* Try simplify a*(b/c) as (a*b)/c. */
4607 if (FLOAT_MODE_P (mode) && flag_unsafe_math_optimizations
4608 && GET_CODE (XEXP (x, 0)) == DIV)
4610 rtx tem = simplify_binary_operation (MULT, mode,
4611 XEXP (XEXP (x, 0), 0),
4612 XEXP (x, 1));
4613 if (tem)
4614 return simplify_gen_binary (DIV, mode, tem, XEXP (XEXP (x, 0), 1));
4616 break;
4618 case UDIV:
4619 /* If this is a divide by a power of two, treat it as a shift if
4620 its first operand is a shift. */
4621 if (GET_CODE (XEXP (x, 1)) == CONST_INT
4622 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0
4623 && (GET_CODE (XEXP (x, 0)) == ASHIFT
4624 || GET_CODE (XEXP (x, 0)) == LSHIFTRT
4625 || GET_CODE (XEXP (x, 0)) == ASHIFTRT
4626 || GET_CODE (XEXP (x, 0)) == ROTATE
4627 || GET_CODE (XEXP (x, 0)) == ROTATERT))
4628 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i);
4629 break;
4631 case EQ: case NE:
4632 case GT: case GTU: case GE: case GEU:
4633 case LT: case LTU: case LE: case LEU:
4634 case UNEQ: case LTGT:
4635 case UNGT: case UNGE:
4636 case UNLT: case UNLE:
4637 case UNORDERED: case ORDERED:
4638 /* If the first operand is a condition code, we can't do anything
4639 with it. */
4640 if (GET_CODE (XEXP (x, 0)) == COMPARE
4641 || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC
4642 && ! CC0_P (XEXP (x, 0))))
4644 rtx op0 = XEXP (x, 0);
4645 rtx op1 = XEXP (x, 1);
4646 enum rtx_code new_code;
4648 if (GET_CODE (op0) == COMPARE)
4649 op1 = XEXP (op0, 1), op0 = XEXP (op0, 0);
4651 /* Simplify our comparison, if possible. */
4652 new_code = simplify_comparison (code, &op0, &op1);
4654 /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X
4655 if only the low-order bit is possibly nonzero in X (such as when
4656 X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to
4657 (xor X 1) or (minus 1 X); we use the former. Finally, if X is
4658 known to be either 0 or -1, NE becomes a NEG and EQ becomes
4659 (plus X 1).
4661 Remove any ZERO_EXTRACT we made when thinking this was a
4662 comparison. It may now be simpler to use, e.g., an AND. If a
4663 ZERO_EXTRACT is indeed appropriate, it will be placed back by
4664 the call to make_compound_operation in the SET case. */
4666 if (STORE_FLAG_VALUE == 1
4667 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4668 && op1 == const0_rtx
4669 && mode == GET_MODE (op0)
4670 && nonzero_bits (op0, mode) == 1)
4671 return gen_lowpart (mode,
4672 expand_compound_operation (op0));
4674 else if (STORE_FLAG_VALUE == 1
4675 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4676 && op1 == const0_rtx
4677 && mode == GET_MODE (op0)
4678 && (num_sign_bit_copies (op0, mode)
4679 == GET_MODE_BITSIZE (mode)))
4681 op0 = expand_compound_operation (op0);
4682 return simplify_gen_unary (NEG, mode,
4683 gen_lowpart (mode, op0),
4684 mode);
4687 else if (STORE_FLAG_VALUE == 1
4688 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4689 && op1 == const0_rtx
4690 && mode == GET_MODE (op0)
4691 && nonzero_bits (op0, mode) == 1)
4693 op0 = expand_compound_operation (op0);
4694 return simplify_gen_binary (XOR, mode,
4695 gen_lowpart (mode, op0),
4696 const1_rtx);
4699 else if (STORE_FLAG_VALUE == 1
4700 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4701 && op1 == const0_rtx
4702 && mode == GET_MODE (op0)
4703 && (num_sign_bit_copies (op0, mode)
4704 == GET_MODE_BITSIZE (mode)))
4706 op0 = expand_compound_operation (op0);
4707 return plus_constant (gen_lowpart (mode, op0), 1);
4710 /* If STORE_FLAG_VALUE is -1, we have cases similar to
4711 those above. */
4712 if (STORE_FLAG_VALUE == -1
4713 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4714 && op1 == const0_rtx
4715 && (num_sign_bit_copies (op0, mode)
4716 == GET_MODE_BITSIZE (mode)))
4717 return gen_lowpart (mode,
4718 expand_compound_operation (op0));
4720 else if (STORE_FLAG_VALUE == -1
4721 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4722 && op1 == const0_rtx
4723 && mode == GET_MODE (op0)
4724 && nonzero_bits (op0, mode) == 1)
4726 op0 = expand_compound_operation (op0);
4727 return simplify_gen_unary (NEG, mode,
4728 gen_lowpart (mode, op0),
4729 mode);
4732 else if (STORE_FLAG_VALUE == -1
4733 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4734 && op1 == const0_rtx
4735 && mode == GET_MODE (op0)
4736 && (num_sign_bit_copies (op0, mode)
4737 == GET_MODE_BITSIZE (mode)))
4739 op0 = expand_compound_operation (op0);
4740 return simplify_gen_unary (NOT, mode,
4741 gen_lowpart (mode, op0),
4742 mode);
4745 /* If X is 0/1, (eq X 0) is X-1. */
4746 else if (STORE_FLAG_VALUE == -1
4747 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4748 && op1 == const0_rtx
4749 && mode == GET_MODE (op0)
4750 && nonzero_bits (op0, mode) == 1)
4752 op0 = expand_compound_operation (op0);
4753 return plus_constant (gen_lowpart (mode, op0), -1);
4756 /* If STORE_FLAG_VALUE says to just test the sign bit and X has just
4757 one bit that might be nonzero, we can convert (ne x 0) to
4758 (ashift x c) where C puts the bit in the sign bit. Remove any
4759 AND with STORE_FLAG_VALUE when we are done, since we are only
4760 going to test the sign bit. */
4761 if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4762 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4763 && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
4764 == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))
4765 && op1 == const0_rtx
4766 && mode == GET_MODE (op0)
4767 && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0)
4769 x = simplify_shift_const (NULL_RTX, ASHIFT, mode,
4770 expand_compound_operation (op0),
4771 GET_MODE_BITSIZE (mode) - 1 - i);
4772 if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx)
4773 return XEXP (x, 0);
4774 else
4775 return x;
4778 /* If the code changed, return a whole new comparison. */
4779 if (new_code != code)
4780 return gen_rtx_fmt_ee (new_code, mode, op0, op1);
4782 /* Otherwise, keep this operation, but maybe change its operands.
4783 This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */
4784 SUBST (XEXP (x, 0), op0);
4785 SUBST (XEXP (x, 1), op1);
4787 break;
4789 case IF_THEN_ELSE:
4790 return simplify_if_then_else (x);
4792 case ZERO_EXTRACT:
4793 case SIGN_EXTRACT:
4794 case ZERO_EXTEND:
4795 case SIGN_EXTEND:
4796 /* If we are processing SET_DEST, we are done. */
4797 if (in_dest)
4798 return x;
4800 return expand_compound_operation (x);
4802 case SET:
4803 return simplify_set (x);
4805 case AND:
4806 case IOR:
4807 return simplify_logical (x);
4809 case ASHIFT:
4810 case LSHIFTRT:
4811 case ASHIFTRT:
4812 case ROTATE:
4813 case ROTATERT:
4814 /* If this is a shift by a constant amount, simplify it. */
4815 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
4816 return simplify_shift_const (x, code, mode, XEXP (x, 0),
4817 INTVAL (XEXP (x, 1)));
4819 else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1)))
4820 SUBST (XEXP (x, 1),
4821 force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)),
4822 ((HOST_WIDE_INT) 1
4823 << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))))
4824 - 1,
4825 0));
4826 break;
4828 default:
4829 break;
4832 return x;
4835 /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */
4837 static rtx
4838 simplify_if_then_else (rtx x)
4840 enum machine_mode mode = GET_MODE (x);
4841 rtx cond = XEXP (x, 0);
4842 rtx true_rtx = XEXP (x, 1);
4843 rtx false_rtx = XEXP (x, 2);
4844 enum rtx_code true_code = GET_CODE (cond);
4845 int comparison_p = COMPARISON_P (cond);
4846 rtx temp;
4847 int i;
4848 enum rtx_code false_code;
4849 rtx reversed;
4851 /* Simplify storing of the truth value. */
4852 if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx)
4853 return simplify_gen_relational (true_code, mode, VOIDmode,
4854 XEXP (cond, 0), XEXP (cond, 1));
4856 /* Also when the truth value has to be reversed. */
4857 if (comparison_p
4858 && true_rtx == const0_rtx && false_rtx == const_true_rtx
4859 && (reversed = reversed_comparison (cond, mode)))
4860 return reversed;
4862 /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used
4863 in it is being compared against certain values. Get the true and false
4864 comparisons and see if that says anything about the value of each arm. */
4866 if (comparison_p
4867 && ((false_code = reversed_comparison_code (cond, NULL))
4868 != UNKNOWN)
4869 && REG_P (XEXP (cond, 0)))
4871 HOST_WIDE_INT nzb;
4872 rtx from = XEXP (cond, 0);
4873 rtx true_val = XEXP (cond, 1);
4874 rtx false_val = true_val;
4875 int swapped = 0;
4877 /* If FALSE_CODE is EQ, swap the codes and arms. */
4879 if (false_code == EQ)
4881 swapped = 1, true_code = EQ, false_code = NE;
4882 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4885 /* If we are comparing against zero and the expression being tested has
4886 only a single bit that might be nonzero, that is its value when it is
4887 not equal to zero. Similarly if it is known to be -1 or 0. */
4889 if (true_code == EQ && true_val == const0_rtx
4890 && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0)
4891 false_code = EQ, false_val = GEN_INT (nzb);
4892 else if (true_code == EQ && true_val == const0_rtx
4893 && (num_sign_bit_copies (from, GET_MODE (from))
4894 == GET_MODE_BITSIZE (GET_MODE (from))))
4895 false_code = EQ, false_val = constm1_rtx;
4897 /* Now simplify an arm if we know the value of the register in the
4898 branch and it is used in the arm. Be careful due to the potential
4899 of locally-shared RTL. */
4901 if (reg_mentioned_p (from, true_rtx))
4902 true_rtx = subst (known_cond (copy_rtx (true_rtx), true_code,
4903 from, true_val),
4904 pc_rtx, pc_rtx, 0, 0);
4905 if (reg_mentioned_p (from, false_rtx))
4906 false_rtx = subst (known_cond (copy_rtx (false_rtx), false_code,
4907 from, false_val),
4908 pc_rtx, pc_rtx, 0, 0);
4910 SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx);
4911 SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx);
4913 true_rtx = XEXP (x, 1);
4914 false_rtx = XEXP (x, 2);
4915 true_code = GET_CODE (cond);
4918 /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be
4919 reversed, do so to avoid needing two sets of patterns for
4920 subtract-and-branch insns. Similarly if we have a constant in the true
4921 arm, the false arm is the same as the first operand of the comparison, or
4922 the false arm is more complicated than the true arm. */
4924 if (comparison_p
4925 && reversed_comparison_code (cond, NULL) != UNKNOWN
4926 && (true_rtx == pc_rtx
4927 || (CONSTANT_P (true_rtx)
4928 && GET_CODE (false_rtx) != CONST_INT && false_rtx != pc_rtx)
4929 || true_rtx == const0_rtx
4930 || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx))
4931 || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx))
4932 && !OBJECT_P (false_rtx))
4933 || reg_mentioned_p (true_rtx, false_rtx)
4934 || rtx_equal_p (false_rtx, XEXP (cond, 0))))
4936 true_code = reversed_comparison_code (cond, NULL);
4937 SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond)));
4938 SUBST (XEXP (x, 1), false_rtx);
4939 SUBST (XEXP (x, 2), true_rtx);
4941 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4942 cond = XEXP (x, 0);
4944 /* It is possible that the conditional has been simplified out. */
4945 true_code = GET_CODE (cond);
4946 comparison_p = COMPARISON_P (cond);
4949 /* If the two arms are identical, we don't need the comparison. */
4951 if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond))
4952 return true_rtx;
4954 /* Convert a == b ? b : a to "a". */
4955 if (true_code == EQ && ! side_effects_p (cond)
4956 && !HONOR_NANS (mode)
4957 && rtx_equal_p (XEXP (cond, 0), false_rtx)
4958 && rtx_equal_p (XEXP (cond, 1), true_rtx))
4959 return false_rtx;
4960 else if (true_code == NE && ! side_effects_p (cond)
4961 && !HONOR_NANS (mode)
4962 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4963 && rtx_equal_p (XEXP (cond, 1), false_rtx))
4964 return true_rtx;
4966 /* Look for cases where we have (abs x) or (neg (abs X)). */
4968 if (GET_MODE_CLASS (mode) == MODE_INT
4969 && GET_CODE (false_rtx) == NEG
4970 && rtx_equal_p (true_rtx, XEXP (false_rtx, 0))
4971 && comparison_p
4972 && rtx_equal_p (true_rtx, XEXP (cond, 0))
4973 && ! side_effects_p (true_rtx))
4974 switch (true_code)
4976 case GT:
4977 case GE:
4978 return simplify_gen_unary (ABS, mode, true_rtx, mode);
4979 case LT:
4980 case LE:
4981 return
4982 simplify_gen_unary (NEG, mode,
4983 simplify_gen_unary (ABS, mode, true_rtx, mode),
4984 mode);
4985 default:
4986 break;
4989 /* Look for MIN or MAX. */
4991 if ((! FLOAT_MODE_P (mode) || flag_unsafe_math_optimizations)
4992 && comparison_p
4993 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4994 && rtx_equal_p (XEXP (cond, 1), false_rtx)
4995 && ! side_effects_p (cond))
4996 switch (true_code)
4998 case GE:
4999 case GT:
5000 return simplify_gen_binary (SMAX, mode, true_rtx, false_rtx);
5001 case LE:
5002 case LT:
5003 return simplify_gen_binary (SMIN, mode, true_rtx, false_rtx);
5004 case GEU:
5005 case GTU:
5006 return simplify_gen_binary (UMAX, mode, true_rtx, false_rtx);
5007 case LEU:
5008 case LTU:
5009 return simplify_gen_binary (UMIN, mode, true_rtx, false_rtx);
5010 default:
5011 break;
5014 /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its
5015 second operand is zero, this can be done as (OP Z (mult COND C2)) where
5016 C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or
5017 SIGN_EXTEND as long as Z is already extended (so we don't destroy it).
5018 We can do this kind of thing in some cases when STORE_FLAG_VALUE is
5019 neither 1 or -1, but it isn't worth checking for. */
5021 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
5022 && comparison_p
5023 && GET_MODE_CLASS (mode) == MODE_INT
5024 && ! side_effects_p (x))
5026 rtx t = make_compound_operation (true_rtx, SET);
5027 rtx f = make_compound_operation (false_rtx, SET);
5028 rtx cond_op0 = XEXP (cond, 0);
5029 rtx cond_op1 = XEXP (cond, 1);
5030 enum rtx_code op = UNKNOWN, extend_op = UNKNOWN;
5031 enum machine_mode m = mode;
5032 rtx z = 0, c1 = NULL_RTX;
5034 if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS
5035 || GET_CODE (t) == IOR || GET_CODE (t) == XOR
5036 || GET_CODE (t) == ASHIFT
5037 || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT)
5038 && rtx_equal_p (XEXP (t, 0), f))
5039 c1 = XEXP (t, 1), op = GET_CODE (t), z = f;
5041 /* If an identity-zero op is commutative, check whether there
5042 would be a match if we swapped the operands. */
5043 else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR
5044 || GET_CODE (t) == XOR)
5045 && rtx_equal_p (XEXP (t, 1), f))
5046 c1 = XEXP (t, 0), op = GET_CODE (t), z = f;
5047 else if (GET_CODE (t) == SIGN_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 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5057 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5058 && (num_sign_bit_copies (f, GET_MODE (f))
5059 > (unsigned int)
5060 (GET_MODE_BITSIZE (mode)
5061 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 0))))))
5063 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5064 extend_op = SIGN_EXTEND;
5065 m = GET_MODE (XEXP (t, 0));
5067 else if (GET_CODE (t) == SIGN_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 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5073 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5074 && (num_sign_bit_copies (f, GET_MODE (f))
5075 > (unsigned int)
5076 (GET_MODE_BITSIZE (mode)
5077 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 1))))))
5079 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5080 extend_op = SIGN_EXTEND;
5081 m = GET_MODE (XEXP (t, 0));
5083 else if (GET_CODE (t) == ZERO_EXTEND
5084 && (GET_CODE (XEXP (t, 0)) == PLUS
5085 || GET_CODE (XEXP (t, 0)) == MINUS
5086 || GET_CODE (XEXP (t, 0)) == IOR
5087 || GET_CODE (XEXP (t, 0)) == XOR
5088 || GET_CODE (XEXP (t, 0)) == ASHIFT
5089 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5090 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5091 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5092 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5093 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5094 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5095 && ((nonzero_bits (f, GET_MODE (f))
5096 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0))))
5097 == 0))
5099 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5100 extend_op = ZERO_EXTEND;
5101 m = GET_MODE (XEXP (t, 0));
5103 else if (GET_CODE (t) == ZERO_EXTEND
5104 && (GET_CODE (XEXP (t, 0)) == PLUS
5105 || GET_CODE (XEXP (t, 0)) == IOR
5106 || GET_CODE (XEXP (t, 0)) == XOR)
5107 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5108 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5109 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5110 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5111 && ((nonzero_bits (f, GET_MODE (f))
5112 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1))))
5113 == 0))
5115 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5116 extend_op = ZERO_EXTEND;
5117 m = GET_MODE (XEXP (t, 0));
5120 if (z)
5122 temp = subst (simplify_gen_relational (true_code, m, VOIDmode,
5123 cond_op0, cond_op1),
5124 pc_rtx, pc_rtx, 0, 0);
5125 temp = simplify_gen_binary (MULT, m, temp,
5126 simplify_gen_binary (MULT, m, c1,
5127 const_true_rtx));
5128 temp = subst (temp, pc_rtx, pc_rtx, 0, 0);
5129 temp = simplify_gen_binary (op, m, gen_lowpart (m, z), temp);
5131 if (extend_op != UNKNOWN)
5132 temp = simplify_gen_unary (extend_op, mode, temp, m);
5134 return temp;
5138 /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or
5139 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the
5140 negation of a single bit, we can convert this operation to a shift. We
5141 can actually do this more generally, but it doesn't seem worth it. */
5143 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5144 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5145 && ((1 == nonzero_bits (XEXP (cond, 0), mode)
5146 && (i = exact_log2 (INTVAL (true_rtx))) >= 0)
5147 || ((num_sign_bit_copies (XEXP (cond, 0), mode)
5148 == GET_MODE_BITSIZE (mode))
5149 && (i = exact_log2 (-INTVAL (true_rtx))) >= 0)))
5150 return
5151 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5152 gen_lowpart (mode, XEXP (cond, 0)), i);
5154 /* (IF_THEN_ELSE (NE REG 0) (0) (8)) is REG for nonzero_bits (REG) == 8. */
5155 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5156 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5157 && GET_MODE (XEXP (cond, 0)) == mode
5158 && (INTVAL (true_rtx) & GET_MODE_MASK (mode))
5159 == nonzero_bits (XEXP (cond, 0), mode)
5160 && (i = exact_log2 (INTVAL (true_rtx) & GET_MODE_MASK (mode))) >= 0)
5161 return XEXP (cond, 0);
5163 return x;
5166 /* Simplify X, a SET expression. Return the new expression. */
5168 static rtx
5169 simplify_set (rtx x)
5171 rtx src = SET_SRC (x);
5172 rtx dest = SET_DEST (x);
5173 enum machine_mode mode
5174 = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest);
5175 rtx other_insn;
5176 rtx *cc_use;
5178 /* (set (pc) (return)) gets written as (return). */
5179 if (GET_CODE (dest) == PC && GET_CODE (src) == RETURN)
5180 return src;
5182 /* Now that we know for sure which bits of SRC we are using, see if we can
5183 simplify the expression for the object knowing that we only need the
5184 low-order bits. */
5186 if (GET_MODE_CLASS (mode) == MODE_INT
5187 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
5189 src = force_to_mode (src, mode, ~(HOST_WIDE_INT) 0, 0);
5190 SUBST (SET_SRC (x), src);
5193 /* If we are setting CC0 or if the source is a COMPARE, look for the use of
5194 the comparison result and try to simplify it unless we already have used
5195 undobuf.other_insn. */
5196 if ((GET_MODE_CLASS (mode) == MODE_CC
5197 || GET_CODE (src) == COMPARE
5198 || CC0_P (dest))
5199 && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0
5200 && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn)
5201 && COMPARISON_P (*cc_use)
5202 && rtx_equal_p (XEXP (*cc_use, 0), dest))
5204 enum rtx_code old_code = GET_CODE (*cc_use);
5205 enum rtx_code new_code;
5206 rtx op0, op1, tmp;
5207 int other_changed = 0;
5208 enum machine_mode compare_mode = GET_MODE (dest);
5210 if (GET_CODE (src) == COMPARE)
5211 op0 = XEXP (src, 0), op1 = XEXP (src, 1);
5212 else
5213 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
5215 tmp = simplify_relational_operation (old_code, compare_mode, VOIDmode,
5216 op0, op1);
5217 if (!tmp)
5218 new_code = old_code;
5219 else if (!CONSTANT_P (tmp))
5221 new_code = GET_CODE (tmp);
5222 op0 = XEXP (tmp, 0);
5223 op1 = XEXP (tmp, 1);
5225 else
5227 rtx pat = PATTERN (other_insn);
5228 undobuf.other_insn = other_insn;
5229 SUBST (*cc_use, tmp);
5231 /* Attempt to simplify CC user. */
5232 if (GET_CODE (pat) == SET)
5234 rtx new = simplify_rtx (SET_SRC (pat));
5235 if (new != NULL_RTX)
5236 SUBST (SET_SRC (pat), new);
5239 /* Convert X into a no-op move. */
5240 SUBST (SET_DEST (x), pc_rtx);
5241 SUBST (SET_SRC (x), pc_rtx);
5242 return x;
5245 /* Simplify our comparison, if possible. */
5246 new_code = simplify_comparison (new_code, &op0, &op1);
5248 #ifdef SELECT_CC_MODE
5249 /* If this machine has CC modes other than CCmode, check to see if we
5250 need to use a different CC mode here. */
5251 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
5252 compare_mode = GET_MODE (op0);
5253 else
5254 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
5256 #ifndef HAVE_cc0
5257 /* If the mode changed, we have to change SET_DEST, the mode in the
5258 compare, and the mode in the place SET_DEST is used. If SET_DEST is
5259 a hard register, just build new versions with the proper mode. If it
5260 is a pseudo, we lose unless it is only time we set the pseudo, in
5261 which case we can safely change its mode. */
5262 if (compare_mode != GET_MODE (dest))
5264 if (can_change_dest_mode (dest, 0, compare_mode))
5266 unsigned int regno = REGNO (dest);
5267 rtx new_dest;
5269 if (regno < FIRST_PSEUDO_REGISTER)
5270 new_dest = gen_rtx_REG (compare_mode, regno);
5271 else
5273 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
5274 new_dest = regno_reg_rtx[regno];
5277 SUBST (SET_DEST (x), new_dest);
5278 SUBST (XEXP (*cc_use, 0), new_dest);
5279 other_changed = 1;
5281 dest = new_dest;
5284 #endif /* cc0 */
5285 #endif /* SELECT_CC_MODE */
5287 /* If the code changed, we have to build a new comparison in
5288 undobuf.other_insn. */
5289 if (new_code != old_code)
5291 int other_changed_previously = other_changed;
5292 unsigned HOST_WIDE_INT mask;
5294 SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use),
5295 dest, const0_rtx));
5296 other_changed = 1;
5298 /* If the only change we made was to change an EQ into an NE or
5299 vice versa, OP0 has only one bit that might be nonzero, and OP1
5300 is zero, check if changing the user of the condition code will
5301 produce a valid insn. If it won't, we can keep the original code
5302 in that insn by surrounding our operation with an XOR. */
5304 if (((old_code == NE && new_code == EQ)
5305 || (old_code == EQ && new_code == NE))
5306 && ! other_changed_previously && op1 == const0_rtx
5307 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
5308 && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0)
5310 rtx pat = PATTERN (other_insn), note = 0;
5312 if ((recog_for_combine (&pat, other_insn, &note) < 0
5313 && ! check_asm_operands (pat)))
5315 PUT_CODE (*cc_use, old_code);
5316 other_changed = 0;
5318 op0 = simplify_gen_binary (XOR, GET_MODE (op0),
5319 op0, GEN_INT (mask));
5324 if (other_changed)
5325 undobuf.other_insn = other_insn;
5327 #ifdef HAVE_cc0
5328 /* If we are now comparing against zero, change our source if
5329 needed. If we do not use cc0, we always have a COMPARE. */
5330 if (op1 == const0_rtx && dest == cc0_rtx)
5332 SUBST (SET_SRC (x), op0);
5333 src = op0;
5335 else
5336 #endif
5338 /* Otherwise, if we didn't previously have a COMPARE in the
5339 correct mode, we need one. */
5340 if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode)
5342 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
5343 src = SET_SRC (x);
5345 else if (GET_MODE (op0) == compare_mode && op1 == const0_rtx)
5347 SUBST(SET_SRC (x), op0);
5348 src = SET_SRC (x);
5350 else
5352 /* Otherwise, update the COMPARE if needed. */
5353 SUBST (XEXP (src, 0), op0);
5354 SUBST (XEXP (src, 1), op1);
5357 else
5359 /* Get SET_SRC in a form where we have placed back any
5360 compound expressions. Then do the checks below. */
5361 src = make_compound_operation (src, SET);
5362 SUBST (SET_SRC (x), src);
5365 /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation,
5366 and X being a REG or (subreg (reg)), we may be able to convert this to
5367 (set (subreg:m2 x) (op)).
5369 We can always do this if M1 is narrower than M2 because that means that
5370 we only care about the low bits of the result.
5372 However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot
5373 perform a narrower operation than requested since the high-order bits will
5374 be undefined. On machine where it is defined, this transformation is safe
5375 as long as M1 and M2 have the same number of words. */
5377 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5378 && !OBJECT_P (SUBREG_REG (src))
5379 && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1))
5380 / UNITS_PER_WORD)
5381 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5382 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))
5383 #ifndef WORD_REGISTER_OPERATIONS
5384 && (GET_MODE_SIZE (GET_MODE (src))
5385 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5386 #endif
5387 #ifdef CANNOT_CHANGE_MODE_CLASS
5388 && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER
5389 && REG_CANNOT_CHANGE_MODE_P (REGNO (dest),
5390 GET_MODE (SUBREG_REG (src)),
5391 GET_MODE (src)))
5392 #endif
5393 && (REG_P (dest)
5394 || (GET_CODE (dest) == SUBREG
5395 && REG_P (SUBREG_REG (dest)))))
5397 SUBST (SET_DEST (x),
5398 gen_lowpart (GET_MODE (SUBREG_REG (src)),
5399 dest));
5400 SUBST (SET_SRC (x), SUBREG_REG (src));
5402 src = SET_SRC (x), dest = SET_DEST (x);
5405 #ifdef HAVE_cc0
5406 /* If we have (set (cc0) (subreg ...)), we try to remove the subreg
5407 in SRC. */
5408 if (dest == cc0_rtx
5409 && GET_CODE (src) == SUBREG
5410 && subreg_lowpart_p (src)
5411 && (GET_MODE_BITSIZE (GET_MODE (src))
5412 < GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (src)))))
5414 rtx inner = SUBREG_REG (src);
5415 enum machine_mode inner_mode = GET_MODE (inner);
5417 /* Here we make sure that we don't have a sign bit on. */
5418 if (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT
5419 && (nonzero_bits (inner, inner_mode)
5420 < ((unsigned HOST_WIDE_INT) 1
5421 << (GET_MODE_BITSIZE (GET_MODE (src)) - 1))))
5423 SUBST (SET_SRC (x), inner);
5424 src = SET_SRC (x);
5427 #endif
5429 #ifdef LOAD_EXTEND_OP
5430 /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this
5431 would require a paradoxical subreg. Replace the subreg with a
5432 zero_extend to avoid the reload that would otherwise be required. */
5434 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5435 && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != UNKNOWN
5436 && SUBREG_BYTE (src) == 0
5437 && (GET_MODE_SIZE (GET_MODE (src))
5438 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5439 && MEM_P (SUBREG_REG (src)))
5441 SUBST (SET_SRC (x),
5442 gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))),
5443 GET_MODE (src), SUBREG_REG (src)));
5445 src = SET_SRC (x);
5447 #endif
5449 /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we
5450 are comparing an item known to be 0 or -1 against 0, use a logical
5451 operation instead. Check for one of the arms being an IOR of the other
5452 arm with some value. We compute three terms to be IOR'ed together. In
5453 practice, at most two will be nonzero. Then we do the IOR's. */
5455 if (GET_CODE (dest) != PC
5456 && GET_CODE (src) == IF_THEN_ELSE
5457 && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT
5458 && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE)
5459 && XEXP (XEXP (src, 0), 1) == const0_rtx
5460 && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0))
5461 #ifdef HAVE_conditional_move
5462 && ! can_conditionally_move_p (GET_MODE (src))
5463 #endif
5464 && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0),
5465 GET_MODE (XEXP (XEXP (src, 0), 0)))
5466 == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (src, 0), 0))))
5467 && ! side_effects_p (src))
5469 rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE
5470 ? XEXP (src, 1) : XEXP (src, 2));
5471 rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE
5472 ? XEXP (src, 2) : XEXP (src, 1));
5473 rtx term1 = const0_rtx, term2, term3;
5475 if (GET_CODE (true_rtx) == IOR
5476 && rtx_equal_p (XEXP (true_rtx, 0), false_rtx))
5477 term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx;
5478 else if (GET_CODE (true_rtx) == IOR
5479 && rtx_equal_p (XEXP (true_rtx, 1), false_rtx))
5480 term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx;
5481 else if (GET_CODE (false_rtx) == IOR
5482 && rtx_equal_p (XEXP (false_rtx, 0), true_rtx))
5483 term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx;
5484 else if (GET_CODE (false_rtx) == IOR
5485 && rtx_equal_p (XEXP (false_rtx, 1), true_rtx))
5486 term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx;
5488 term2 = simplify_gen_binary (AND, GET_MODE (src),
5489 XEXP (XEXP (src, 0), 0), true_rtx);
5490 term3 = simplify_gen_binary (AND, GET_MODE (src),
5491 simplify_gen_unary (NOT, GET_MODE (src),
5492 XEXP (XEXP (src, 0), 0),
5493 GET_MODE (src)),
5494 false_rtx);
5496 SUBST (SET_SRC (x),
5497 simplify_gen_binary (IOR, GET_MODE (src),
5498 simplify_gen_binary (IOR, GET_MODE (src),
5499 term1, term2),
5500 term3));
5502 src = SET_SRC (x);
5505 /* If either SRC or DEST is a CLOBBER of (const_int 0), make this
5506 whole thing fail. */
5507 if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx)
5508 return src;
5509 else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx)
5510 return dest;
5511 else
5512 /* Convert this into a field assignment operation, if possible. */
5513 return make_field_assignment (x);
5516 /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified
5517 result. */
5519 static rtx
5520 simplify_logical (rtx x)
5522 enum machine_mode mode = GET_MODE (x);
5523 rtx op0 = XEXP (x, 0);
5524 rtx op1 = XEXP (x, 1);
5526 switch (GET_CODE (x))
5528 case AND:
5529 /* We can call simplify_and_const_int only if we don't lose
5530 any (sign) bits when converting INTVAL (op1) to
5531 "unsigned HOST_WIDE_INT". */
5532 if (GET_CODE (op1) == CONST_INT
5533 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5534 || INTVAL (op1) > 0))
5536 x = simplify_and_const_int (x, mode, op0, INTVAL (op1));
5537 if (GET_CODE (x) != AND)
5538 return x;
5540 op0 = XEXP (x, 0);
5541 op1 = XEXP (x, 1);
5544 /* If we have any of (and (ior A B) C) or (and (xor A B) C),
5545 apply the distributive law and then the inverse distributive
5546 law to see if things simplify. */
5547 if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR)
5549 rtx result = distribute_and_simplify_rtx (x, 0);
5550 if (result)
5551 return result;
5553 if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR)
5555 rtx result = distribute_and_simplify_rtx (x, 1);
5556 if (result)
5557 return result;
5559 break;
5561 case IOR:
5562 /* If we have (ior (and A B) C), apply the distributive law and then
5563 the inverse distributive law to see if things simplify. */
5565 if (GET_CODE (op0) == AND)
5567 rtx result = distribute_and_simplify_rtx (x, 0);
5568 if (result)
5569 return result;
5572 if (GET_CODE (op1) == AND)
5574 rtx result = distribute_and_simplify_rtx (x, 1);
5575 if (result)
5576 return result;
5578 break;
5580 default:
5581 gcc_unreachable ();
5584 return x;
5587 /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound
5588 operations" because they can be replaced with two more basic operations.
5589 ZERO_EXTEND is also considered "compound" because it can be replaced with
5590 an AND operation, which is simpler, though only one operation.
5592 The function expand_compound_operation is called with an rtx expression
5593 and will convert it to the appropriate shifts and AND operations,
5594 simplifying at each stage.
5596 The function make_compound_operation is called to convert an expression
5597 consisting of shifts and ANDs into the equivalent compound expression.
5598 It is the inverse of this function, loosely speaking. */
5600 static rtx
5601 expand_compound_operation (rtx x)
5603 unsigned HOST_WIDE_INT pos = 0, len;
5604 int unsignedp = 0;
5605 unsigned int modewidth;
5606 rtx tem;
5608 switch (GET_CODE (x))
5610 case ZERO_EXTEND:
5611 unsignedp = 1;
5612 case SIGN_EXTEND:
5613 /* We can't necessarily use a const_int for a multiword mode;
5614 it depends on implicitly extending the value.
5615 Since we don't know the right way to extend it,
5616 we can't tell whether the implicit way is right.
5618 Even for a mode that is no wider than a const_int,
5619 we can't win, because we need to sign extend one of its bits through
5620 the rest of it, and we don't know which bit. */
5621 if (GET_CODE (XEXP (x, 0)) == CONST_INT)
5622 return x;
5624 /* Return if (subreg:MODE FROM 0) is not a safe replacement for
5625 (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM
5626 because (SUBREG (MEM...)) is guaranteed to cause the MEM to be
5627 reloaded. If not for that, MEM's would very rarely be safe.
5629 Reject MODEs bigger than a word, because we might not be able
5630 to reference a two-register group starting with an arbitrary register
5631 (and currently gen_lowpart might crash for a SUBREG). */
5633 if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD)
5634 return x;
5636 /* Reject MODEs that aren't scalar integers because turning vector
5637 or complex modes into shifts causes problems. */
5639 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5640 return x;
5642 len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)));
5643 /* If the inner object has VOIDmode (the only way this can happen
5644 is if it is an ASM_OPERANDS), we can't do anything since we don't
5645 know how much masking to do. */
5646 if (len == 0)
5647 return x;
5649 break;
5651 case ZERO_EXTRACT:
5652 unsignedp = 1;
5654 /* ... fall through ... */
5656 case SIGN_EXTRACT:
5657 /* If the operand is a CLOBBER, just return it. */
5658 if (GET_CODE (XEXP (x, 0)) == CLOBBER)
5659 return XEXP (x, 0);
5661 if (GET_CODE (XEXP (x, 1)) != CONST_INT
5662 || GET_CODE (XEXP (x, 2)) != CONST_INT
5663 || GET_MODE (XEXP (x, 0)) == VOIDmode)
5664 return x;
5666 /* Reject MODEs that aren't scalar integers because turning vector
5667 or complex modes into shifts causes problems. */
5669 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5670 return x;
5672 len = INTVAL (XEXP (x, 1));
5673 pos = INTVAL (XEXP (x, 2));
5675 /* This should stay within the object being extracted, fail otherwise. */
5676 if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))
5677 return x;
5679 if (BITS_BIG_ENDIAN)
5680 pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos;
5682 break;
5684 default:
5685 return x;
5687 /* Convert sign extension to zero extension, if we know that the high
5688 bit is not set, as this is easier to optimize. It will be converted
5689 back to cheaper alternative in make_extraction. */
5690 if (GET_CODE (x) == SIGN_EXTEND
5691 && (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5692 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
5693 & ~(((unsigned HOST_WIDE_INT)
5694 GET_MODE_MASK (GET_MODE (XEXP (x, 0))))
5695 >> 1))
5696 == 0)))
5698 rtx temp = gen_rtx_ZERO_EXTEND (GET_MODE (x), XEXP (x, 0));
5699 rtx temp2 = expand_compound_operation (temp);
5701 /* Make sure this is a profitable operation. */
5702 if (rtx_cost (x, SET) > rtx_cost (temp2, SET))
5703 return temp2;
5704 else if (rtx_cost (x, SET) > rtx_cost (temp, SET))
5705 return temp;
5706 else
5707 return x;
5710 /* We can optimize some special cases of ZERO_EXTEND. */
5711 if (GET_CODE (x) == ZERO_EXTEND)
5713 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we
5714 know that the last value didn't have any inappropriate bits
5715 set. */
5716 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5717 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5718 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5719 && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x))
5720 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5721 return XEXP (XEXP (x, 0), 0);
5723 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5724 if (GET_CODE (XEXP (x, 0)) == SUBREG
5725 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5726 && subreg_lowpart_p (XEXP (x, 0))
5727 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5728 && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x))
5729 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5730 return SUBREG_REG (XEXP (x, 0));
5732 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo
5733 is a comparison and STORE_FLAG_VALUE permits. This is like
5734 the first case, but it works even when GET_MODE (x) is larger
5735 than HOST_WIDE_INT. */
5736 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5737 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5738 && COMPARISON_P (XEXP (XEXP (x, 0), 0))
5739 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5740 <= HOST_BITS_PER_WIDE_INT)
5741 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5742 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5743 return XEXP (XEXP (x, 0), 0);
5745 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5746 if (GET_CODE (XEXP (x, 0)) == SUBREG
5747 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5748 && subreg_lowpart_p (XEXP (x, 0))
5749 && COMPARISON_P (SUBREG_REG (XEXP (x, 0)))
5750 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5751 <= HOST_BITS_PER_WIDE_INT)
5752 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5753 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5754 return SUBREG_REG (XEXP (x, 0));
5758 /* If we reach here, we want to return a pair of shifts. The inner
5759 shift is a left shift of BITSIZE - POS - LEN bits. The outer
5760 shift is a right shift of BITSIZE - LEN bits. It is arithmetic or
5761 logical depending on the value of UNSIGNEDP.
5763 If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be
5764 converted into an AND of a shift.
5766 We must check for the case where the left shift would have a negative
5767 count. This can happen in a case like (x >> 31) & 255 on machines
5768 that can't shift by a constant. On those machines, we would first
5769 combine the shift with the AND to produce a variable-position
5770 extraction. Then the constant of 31 would be substituted in to produce
5771 a such a position. */
5773 modewidth = GET_MODE_BITSIZE (GET_MODE (x));
5774 if (modewidth + len >= pos)
5776 enum machine_mode mode = GET_MODE (x);
5777 tem = gen_lowpart (mode, XEXP (x, 0));
5778 if (!tem || GET_CODE (tem) == CLOBBER)
5779 return x;
5780 tem = simplify_shift_const (NULL_RTX, ASHIFT, mode,
5781 tem, modewidth - pos - len);
5782 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT,
5783 mode, tem, modewidth - len);
5785 else if (unsignedp && len < HOST_BITS_PER_WIDE_INT)
5786 tem = simplify_and_const_int (NULL_RTX, GET_MODE (x),
5787 simplify_shift_const (NULL_RTX, LSHIFTRT,
5788 GET_MODE (x),
5789 XEXP (x, 0), pos),
5790 ((HOST_WIDE_INT) 1 << len) - 1);
5791 else
5792 /* Any other cases we can't handle. */
5793 return x;
5795 /* If we couldn't do this for some reason, return the original
5796 expression. */
5797 if (GET_CODE (tem) == CLOBBER)
5798 return x;
5800 return tem;
5803 /* X is a SET which contains an assignment of one object into
5804 a part of another (such as a bit-field assignment, STRICT_LOW_PART,
5805 or certain SUBREGS). If possible, convert it into a series of
5806 logical operations.
5808 We half-heartedly support variable positions, but do not at all
5809 support variable lengths. */
5811 static rtx
5812 expand_field_assignment (rtx x)
5814 rtx inner;
5815 rtx pos; /* Always counts from low bit. */
5816 int len;
5817 rtx mask, cleared, masked;
5818 enum machine_mode compute_mode;
5820 /* Loop until we find something we can't simplify. */
5821 while (1)
5823 if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
5824 && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG)
5826 inner = SUBREG_REG (XEXP (SET_DEST (x), 0));
5827 len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)));
5828 pos = GEN_INT (subreg_lsb (XEXP (SET_DEST (x), 0)));
5830 else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
5831 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT)
5833 inner = XEXP (SET_DEST (x), 0);
5834 len = INTVAL (XEXP (SET_DEST (x), 1));
5835 pos = XEXP (SET_DEST (x), 2);
5837 /* A constant position should stay within the width of INNER. */
5838 if (GET_CODE (pos) == CONST_INT
5839 && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner)))
5840 break;
5842 if (BITS_BIG_ENDIAN)
5844 if (GET_CODE (pos) == CONST_INT)
5845 pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len
5846 - INTVAL (pos));
5847 else if (GET_CODE (pos) == MINUS
5848 && GET_CODE (XEXP (pos, 1)) == CONST_INT
5849 && (INTVAL (XEXP (pos, 1))
5850 == GET_MODE_BITSIZE (GET_MODE (inner)) - len))
5851 /* If position is ADJUST - X, new position is X. */
5852 pos = XEXP (pos, 0);
5853 else
5854 pos = simplify_gen_binary (MINUS, GET_MODE (pos),
5855 GEN_INT (GET_MODE_BITSIZE (
5856 GET_MODE (inner))
5857 - len),
5858 pos);
5862 /* A SUBREG between two modes that occupy the same numbers of words
5863 can be done by moving the SUBREG to the source. */
5864 else if (GET_CODE (SET_DEST (x)) == SUBREG
5865 /* We need SUBREGs to compute nonzero_bits properly. */
5866 && nonzero_sign_valid
5867 && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
5868 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
5869 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
5870 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))
5872 x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)),
5873 gen_lowpart
5874 (GET_MODE (SUBREG_REG (SET_DEST (x))),
5875 SET_SRC (x)));
5876 continue;
5878 else
5879 break;
5881 while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5882 inner = SUBREG_REG (inner);
5884 compute_mode = GET_MODE (inner);
5886 /* Don't attempt bitwise arithmetic on non scalar integer modes. */
5887 if (! SCALAR_INT_MODE_P (compute_mode))
5889 enum machine_mode imode;
5891 /* Don't do anything for vector or complex integral types. */
5892 if (! FLOAT_MODE_P (compute_mode))
5893 break;
5895 /* Try to find an integral mode to pun with. */
5896 imode = mode_for_size (GET_MODE_BITSIZE (compute_mode), MODE_INT, 0);
5897 if (imode == BLKmode)
5898 break;
5900 compute_mode = imode;
5901 inner = gen_lowpart (imode, inner);
5904 /* Compute a mask of LEN bits, if we can do this on the host machine. */
5905 if (len >= HOST_BITS_PER_WIDE_INT)
5906 break;
5908 /* Now compute the equivalent expression. Make a copy of INNER
5909 for the SET_DEST in case it is a MEM into which we will substitute;
5910 we don't want shared RTL in that case. */
5911 mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1);
5912 cleared = simplify_gen_binary (AND, compute_mode,
5913 simplify_gen_unary (NOT, compute_mode,
5914 simplify_gen_binary (ASHIFT,
5915 compute_mode,
5916 mask, pos),
5917 compute_mode),
5918 inner);
5919 masked = simplify_gen_binary (ASHIFT, compute_mode,
5920 simplify_gen_binary (
5921 AND, compute_mode,
5922 gen_lowpart (compute_mode, SET_SRC (x)),
5923 mask),
5924 pos);
5926 x = gen_rtx_SET (VOIDmode, copy_rtx (inner),
5927 simplify_gen_binary (IOR, compute_mode,
5928 cleared, masked));
5931 return x;
5934 /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero,
5935 it is an RTX that represents a variable starting position; otherwise,
5936 POS is the (constant) starting bit position (counted from the LSB).
5938 UNSIGNEDP is nonzero for an unsigned reference and zero for a
5939 signed reference.
5941 IN_DEST is nonzero if this is a reference in the destination of a
5942 SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero,
5943 a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will
5944 be used.
5946 IN_COMPARE is nonzero if we are in a COMPARE. This means that a
5947 ZERO_EXTRACT should be built even for bits starting at bit 0.
5949 MODE is the desired mode of the result (if IN_DEST == 0).
5951 The result is an RTX for the extraction or NULL_RTX if the target
5952 can't handle it. */
5954 static rtx
5955 make_extraction (enum machine_mode mode, rtx inner, HOST_WIDE_INT pos,
5956 rtx pos_rtx, unsigned HOST_WIDE_INT len, int unsignedp,
5957 int in_dest, int in_compare)
5959 /* This mode describes the size of the storage area
5960 to fetch the overall value from. Within that, we
5961 ignore the POS lowest bits, etc. */
5962 enum machine_mode is_mode = GET_MODE (inner);
5963 enum machine_mode inner_mode;
5964 enum machine_mode wanted_inner_mode;
5965 enum machine_mode wanted_inner_reg_mode = word_mode;
5966 enum machine_mode pos_mode = word_mode;
5967 enum machine_mode extraction_mode = word_mode;
5968 enum machine_mode tmode = mode_for_size (len, MODE_INT, 1);
5969 rtx new = 0;
5970 rtx orig_pos_rtx = pos_rtx;
5971 HOST_WIDE_INT orig_pos;
5973 if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5975 /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...),
5976 consider just the QI as the memory to extract from.
5977 The subreg adds or removes high bits; its mode is
5978 irrelevant to the meaning of this extraction,
5979 since POS and LEN count from the lsb. */
5980 if (MEM_P (SUBREG_REG (inner)))
5981 is_mode = GET_MODE (SUBREG_REG (inner));
5982 inner = SUBREG_REG (inner);
5984 else if (GET_CODE (inner) == ASHIFT
5985 && GET_CODE (XEXP (inner, 1)) == CONST_INT
5986 && pos_rtx == 0 && pos == 0
5987 && len > (unsigned HOST_WIDE_INT) INTVAL (XEXP (inner, 1)))
5989 /* We're extracting the least significant bits of an rtx
5990 (ashift X (const_int C)), where LEN > C. Extract the
5991 least significant (LEN - C) bits of X, giving an rtx
5992 whose mode is MODE, then shift it left C times. */
5993 new = make_extraction (mode, XEXP (inner, 0),
5994 0, 0, len - INTVAL (XEXP (inner, 1)),
5995 unsignedp, in_dest, in_compare);
5996 if (new != 0)
5997 return gen_rtx_ASHIFT (mode, new, XEXP (inner, 1));
6000 inner_mode = GET_MODE (inner);
6002 if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT)
6003 pos = INTVAL (pos_rtx), pos_rtx = 0;
6005 /* See if this can be done without an extraction. We never can if the
6006 width of the field is not the same as that of some integer mode. For
6007 registers, we can only avoid the extraction if the position is at the
6008 low-order bit and this is either not in the destination or we have the
6009 appropriate STRICT_LOW_PART operation available.
6011 For MEM, we can avoid an extract if the field starts on an appropriate
6012 boundary and we can change the mode of the memory reference. */
6014 if (tmode != BLKmode
6015 && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
6016 && !MEM_P (inner)
6017 && (inner_mode == tmode
6018 || !REG_P (inner)
6019 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
6020 GET_MODE_BITSIZE (inner_mode))
6021 || reg_truncated_to_mode (tmode, inner))
6022 && (! in_dest
6023 || (REG_P (inner)
6024 && have_insn_for (STRICT_LOW_PART, tmode))))
6025 || (MEM_P (inner) && pos_rtx == 0
6026 && (pos
6027 % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode)
6028 : BITS_PER_UNIT)) == 0
6029 /* We can't do this if we are widening INNER_MODE (it
6030 may not be aligned, for one thing). */
6031 && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode)
6032 && (inner_mode == tmode
6033 || (! mode_dependent_address_p (XEXP (inner, 0))
6034 && ! MEM_VOLATILE_P (inner))))))
6036 /* If INNER is a MEM, make a new MEM that encompasses just the desired
6037 field. If the original and current mode are the same, we need not
6038 adjust the offset. Otherwise, we do if bytes big endian.
6040 If INNER is not a MEM, get a piece consisting of just the field
6041 of interest (in this case POS % BITS_PER_WORD must be 0). */
6043 if (MEM_P (inner))
6045 HOST_WIDE_INT offset;
6047 /* POS counts from lsb, but make OFFSET count in memory order. */
6048 if (BYTES_BIG_ENDIAN)
6049 offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT;
6050 else
6051 offset = pos / BITS_PER_UNIT;
6053 new = adjust_address_nv (inner, tmode, offset);
6055 else if (REG_P (inner))
6057 if (tmode != inner_mode)
6059 /* We can't call gen_lowpart in a DEST since we
6060 always want a SUBREG (see below) and it would sometimes
6061 return a new hard register. */
6062 if (pos || in_dest)
6064 HOST_WIDE_INT final_word = pos / BITS_PER_WORD;
6066 if (WORDS_BIG_ENDIAN
6067 && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD)
6068 final_word = ((GET_MODE_SIZE (inner_mode)
6069 - GET_MODE_SIZE (tmode))
6070 / UNITS_PER_WORD) - final_word;
6072 final_word *= UNITS_PER_WORD;
6073 if (BYTES_BIG_ENDIAN &&
6074 GET_MODE_SIZE (inner_mode) > GET_MODE_SIZE (tmode))
6075 final_word += (GET_MODE_SIZE (inner_mode)
6076 - GET_MODE_SIZE (tmode)) % UNITS_PER_WORD;
6078 /* Avoid creating invalid subregs, for example when
6079 simplifying (x>>32)&255. */
6080 if (!validate_subreg (tmode, inner_mode, inner, final_word))
6081 return NULL_RTX;
6083 new = gen_rtx_SUBREG (tmode, inner, final_word);
6085 else
6086 new = gen_lowpart (tmode, inner);
6088 else
6089 new = inner;
6091 else
6092 new = force_to_mode (inner, tmode,
6093 len >= HOST_BITS_PER_WIDE_INT
6094 ? ~(unsigned HOST_WIDE_INT) 0
6095 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
6098 /* If this extraction is going into the destination of a SET,
6099 make a STRICT_LOW_PART unless we made a MEM. */
6101 if (in_dest)
6102 return (MEM_P (new) ? new
6103 : (GET_CODE (new) != SUBREG
6104 ? gen_rtx_CLOBBER (tmode, const0_rtx)
6105 : gen_rtx_STRICT_LOW_PART (VOIDmode, new)));
6107 if (mode == tmode)
6108 return new;
6110 if (GET_CODE (new) == CONST_INT)
6111 return gen_int_mode (INTVAL (new), mode);
6113 /* If we know that no extraneous bits are set, and that the high
6114 bit is not set, convert the extraction to the cheaper of
6115 sign and zero extension, that are equivalent in these cases. */
6116 if (flag_expensive_optimizations
6117 && (GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
6118 && ((nonzero_bits (new, tmode)
6119 & ~(((unsigned HOST_WIDE_INT)
6120 GET_MODE_MASK (tmode))
6121 >> 1))
6122 == 0)))
6124 rtx temp = gen_rtx_ZERO_EXTEND (mode, new);
6125 rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new);
6127 /* Prefer ZERO_EXTENSION, since it gives more information to
6128 backends. */
6129 if (rtx_cost (temp, SET) <= rtx_cost (temp1, SET))
6130 return temp;
6131 return temp1;
6134 /* Otherwise, sign- or zero-extend unless we already are in the
6135 proper mode. */
6137 return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
6138 mode, new));
6141 /* Unless this is a COMPARE or we have a funny memory reference,
6142 don't do anything with zero-extending field extracts starting at
6143 the low-order bit since they are simple AND operations. */
6144 if (pos_rtx == 0 && pos == 0 && ! in_dest
6145 && ! in_compare && unsignedp)
6146 return 0;
6148 /* Unless INNER is not MEM, reject this if we would be spanning bytes or
6149 if the position is not a constant and the length is not 1. In all
6150 other cases, we would only be going outside our object in cases when
6151 an original shift would have been undefined. */
6152 if (MEM_P (inner)
6153 && ((pos_rtx == 0 && pos + len > GET_MODE_BITSIZE (is_mode))
6154 || (pos_rtx != 0 && len != 1)))
6155 return 0;
6157 /* Get the mode to use should INNER not be a MEM, the mode for the position,
6158 and the mode for the result. */
6159 if (in_dest && mode_for_extraction (EP_insv, -1) != MAX_MACHINE_MODE)
6161 wanted_inner_reg_mode = mode_for_extraction (EP_insv, 0);
6162 pos_mode = mode_for_extraction (EP_insv, 2);
6163 extraction_mode = mode_for_extraction (EP_insv, 3);
6166 if (! in_dest && unsignedp
6167 && mode_for_extraction (EP_extzv, -1) != MAX_MACHINE_MODE)
6169 wanted_inner_reg_mode = mode_for_extraction (EP_extzv, 1);
6170 pos_mode = mode_for_extraction (EP_extzv, 3);
6171 extraction_mode = mode_for_extraction (EP_extzv, 0);
6174 if (! in_dest && ! unsignedp
6175 && mode_for_extraction (EP_extv, -1) != MAX_MACHINE_MODE)
6177 wanted_inner_reg_mode = mode_for_extraction (EP_extv, 1);
6178 pos_mode = mode_for_extraction (EP_extv, 3);
6179 extraction_mode = mode_for_extraction (EP_extv, 0);
6182 /* Never narrow an object, since that might not be safe. */
6184 if (mode != VOIDmode
6185 && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode))
6186 extraction_mode = mode;
6188 if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode
6189 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6190 pos_mode = GET_MODE (pos_rtx);
6192 /* If this is not from memory, the desired mode is the preferred mode
6193 for an extraction pattern's first input operand, or word_mode if there
6194 is none. */
6195 if (!MEM_P (inner))
6196 wanted_inner_mode = wanted_inner_reg_mode;
6197 else
6199 /* Be careful not to go beyond the extracted object and maintain the
6200 natural alignment of the memory. */
6201 wanted_inner_mode = smallest_mode_for_size (len, MODE_INT);
6202 while (pos % GET_MODE_BITSIZE (wanted_inner_mode) + len
6203 > GET_MODE_BITSIZE (wanted_inner_mode))
6205 wanted_inner_mode = GET_MODE_WIDER_MODE (wanted_inner_mode);
6206 gcc_assert (wanted_inner_mode != VOIDmode);
6209 /* If we have to change the mode of memory and cannot, the desired mode
6210 is EXTRACTION_MODE. */
6211 if (inner_mode != wanted_inner_mode
6212 && (mode_dependent_address_p (XEXP (inner, 0))
6213 || MEM_VOLATILE_P (inner)
6214 || pos_rtx))
6215 wanted_inner_mode = extraction_mode;
6218 orig_pos = pos;
6220 if (BITS_BIG_ENDIAN)
6222 /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to
6223 BITS_BIG_ENDIAN style. If position is constant, compute new
6224 position. Otherwise, build subtraction.
6225 Note that POS is relative to the mode of the original argument.
6226 If it's a MEM we need to recompute POS relative to that.
6227 However, if we're extracting from (or inserting into) a register,
6228 we want to recompute POS relative to wanted_inner_mode. */
6229 int width = (MEM_P (inner)
6230 ? GET_MODE_BITSIZE (is_mode)
6231 : GET_MODE_BITSIZE (wanted_inner_mode));
6233 if (pos_rtx == 0)
6234 pos = width - len - pos;
6235 else
6236 pos_rtx
6237 = gen_rtx_MINUS (GET_MODE (pos_rtx), GEN_INT (width - len), pos_rtx);
6238 /* POS may be less than 0 now, but we check for that below.
6239 Note that it can only be less than 0 if !MEM_P (inner). */
6242 /* If INNER has a wider mode, and this is a constant extraction, try to
6243 make it smaller and adjust the byte to point to the byte containing
6244 the value. */
6245 if (wanted_inner_mode != VOIDmode
6246 && inner_mode != wanted_inner_mode
6247 && ! pos_rtx
6248 && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode)
6249 && MEM_P (inner)
6250 && ! mode_dependent_address_p (XEXP (inner, 0))
6251 && ! MEM_VOLATILE_P (inner))
6253 int offset = 0;
6255 /* The computations below will be correct if the machine is big
6256 endian in both bits and bytes or little endian in bits and bytes.
6257 If it is mixed, we must adjust. */
6259 /* If bytes are big endian and we had a paradoxical SUBREG, we must
6260 adjust OFFSET to compensate. */
6261 if (BYTES_BIG_ENDIAN
6262 && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode))
6263 offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode);
6265 /* We can now move to the desired byte. */
6266 offset += (pos / GET_MODE_BITSIZE (wanted_inner_mode))
6267 * GET_MODE_SIZE (wanted_inner_mode);
6268 pos %= GET_MODE_BITSIZE (wanted_inner_mode);
6270 if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN
6271 && is_mode != wanted_inner_mode)
6272 offset = (GET_MODE_SIZE (is_mode)
6273 - GET_MODE_SIZE (wanted_inner_mode) - offset);
6275 inner = adjust_address_nv (inner, wanted_inner_mode, offset);
6278 /* If INNER is not memory, we can always get it into the proper mode. If we
6279 are changing its mode, POS must be a constant and smaller than the size
6280 of the new mode. */
6281 else if (!MEM_P (inner))
6283 if (GET_MODE (inner) != wanted_inner_mode
6284 && (pos_rtx != 0
6285 || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode)))
6286 return 0;
6288 if (orig_pos < 0)
6289 return 0;
6291 inner = force_to_mode (inner, wanted_inner_mode,
6292 pos_rtx
6293 || len + orig_pos >= HOST_BITS_PER_WIDE_INT
6294 ? ~(unsigned HOST_WIDE_INT) 0
6295 : ((((unsigned HOST_WIDE_INT) 1 << len) - 1)
6296 << orig_pos),
6300 /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we
6301 have to zero extend. Otherwise, we can just use a SUBREG. */
6302 if (pos_rtx != 0
6303 && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx)))
6305 rtx temp = gen_rtx_ZERO_EXTEND (pos_mode, pos_rtx);
6307 /* If we know that no extraneous bits are set, and that the high
6308 bit is not set, convert extraction to cheaper one - either
6309 SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these
6310 cases. */
6311 if (flag_expensive_optimizations
6312 && (GET_MODE_BITSIZE (GET_MODE (pos_rtx)) <= HOST_BITS_PER_WIDE_INT
6313 && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx))
6314 & ~(((unsigned HOST_WIDE_INT)
6315 GET_MODE_MASK (GET_MODE (pos_rtx)))
6316 >> 1))
6317 == 0)))
6319 rtx temp1 = gen_rtx_SIGN_EXTEND (pos_mode, pos_rtx);
6321 /* Prefer ZERO_EXTENSION, since it gives more information to
6322 backends. */
6323 if (rtx_cost (temp1, SET) < rtx_cost (temp, SET))
6324 temp = temp1;
6326 pos_rtx = temp;
6328 else if (pos_rtx != 0
6329 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6330 pos_rtx = gen_lowpart (pos_mode, pos_rtx);
6332 /* Make POS_RTX unless we already have it and it is correct. If we don't
6333 have a POS_RTX but we do have an ORIG_POS_RTX, the latter must
6334 be a CONST_INT. */
6335 if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos)
6336 pos_rtx = orig_pos_rtx;
6338 else if (pos_rtx == 0)
6339 pos_rtx = GEN_INT (pos);
6341 /* Make the required operation. See if we can use existing rtx. */
6342 new = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT,
6343 extraction_mode, inner, GEN_INT (len), pos_rtx);
6344 if (! in_dest)
6345 new = gen_lowpart (mode, new);
6347 return new;
6350 /* See if X contains an ASHIFT of COUNT or more bits that can be commuted
6351 with any other operations in X. Return X without that shift if so. */
6353 static rtx
6354 extract_left_shift (rtx x, int count)
6356 enum rtx_code code = GET_CODE (x);
6357 enum machine_mode mode = GET_MODE (x);
6358 rtx tem;
6360 switch (code)
6362 case ASHIFT:
6363 /* This is the shift itself. If it is wide enough, we will return
6364 either the value being shifted if the shift count is equal to
6365 COUNT or a shift for the difference. */
6366 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6367 && INTVAL (XEXP (x, 1)) >= count)
6368 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0),
6369 INTVAL (XEXP (x, 1)) - count);
6370 break;
6372 case NEG: case NOT:
6373 if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6374 return simplify_gen_unary (code, mode, tem, mode);
6376 break;
6378 case PLUS: case IOR: case XOR: case AND:
6379 /* If we can safely shift this constant and we find the inner shift,
6380 make a new operation. */
6381 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6382 && (INTVAL (XEXP (x, 1)) & ((((HOST_WIDE_INT) 1 << count)) - 1)) == 0
6383 && (tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6384 return simplify_gen_binary (code, mode, tem,
6385 GEN_INT (INTVAL (XEXP (x, 1)) >> count));
6387 break;
6389 default:
6390 break;
6393 return 0;
6396 /* Look at the expression rooted at X. Look for expressions
6397 equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND.
6398 Form these expressions.
6400 Return the new rtx, usually just X.
6402 Also, for machines like the VAX that don't have logical shift insns,
6403 try to convert logical to arithmetic shift operations in cases where
6404 they are equivalent. This undoes the canonicalizations to logical
6405 shifts done elsewhere.
6407 We try, as much as possible, to re-use rtl expressions to save memory.
6409 IN_CODE says what kind of expression we are processing. Normally, it is
6410 SET. In a memory address (inside a MEM, PLUS or minus, the latter two
6411 being kludges), it is MEM. When processing the arguments of a comparison
6412 or a COMPARE against zero, it is COMPARE. */
6414 static rtx
6415 make_compound_operation (rtx x, enum rtx_code in_code)
6417 enum rtx_code code = GET_CODE (x);
6418 enum machine_mode mode = GET_MODE (x);
6419 int mode_width = GET_MODE_BITSIZE (mode);
6420 rtx rhs, lhs;
6421 enum rtx_code next_code;
6422 int i;
6423 rtx new = 0;
6424 rtx tem;
6425 const char *fmt;
6427 /* Select the code to be used in recursive calls. Once we are inside an
6428 address, we stay there. If we have a comparison, set to COMPARE,
6429 but once inside, go back to our default of SET. */
6431 next_code = (code == MEM || code == PLUS || code == MINUS ? MEM
6432 : ((code == COMPARE || COMPARISON_P (x))
6433 && XEXP (x, 1) == const0_rtx) ? COMPARE
6434 : in_code == COMPARE ? SET : in_code);
6436 /* Process depending on the code of this operation. If NEW is set
6437 nonzero, it will be returned. */
6439 switch (code)
6441 case ASHIFT:
6442 /* Convert shifts by constants into multiplications if inside
6443 an address. */
6444 if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT
6445 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
6446 && INTVAL (XEXP (x, 1)) >= 0)
6448 new = make_compound_operation (XEXP (x, 0), next_code);
6449 new = gen_rtx_MULT (mode, new,
6450 GEN_INT ((HOST_WIDE_INT) 1
6451 << INTVAL (XEXP (x, 1))));
6453 break;
6455 case AND:
6456 /* If the second operand is not a constant, we can't do anything
6457 with it. */
6458 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6459 break;
6461 /* If the constant is a power of two minus one and the first operand
6462 is a logical right shift, make an extraction. */
6463 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6464 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6466 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6467 new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1,
6468 0, in_code == COMPARE);
6471 /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */
6472 else if (GET_CODE (XEXP (x, 0)) == SUBREG
6473 && subreg_lowpart_p (XEXP (x, 0))
6474 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT
6475 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6477 new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0),
6478 next_code);
6479 new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0,
6480 XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1,
6481 0, in_code == COMPARE);
6483 /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */
6484 else if ((GET_CODE (XEXP (x, 0)) == XOR
6485 || GET_CODE (XEXP (x, 0)) == IOR)
6486 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT
6487 && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT
6488 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6490 /* Apply the distributive law, and then try to make extractions. */
6491 new = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode,
6492 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0),
6493 XEXP (x, 1)),
6494 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1),
6495 XEXP (x, 1)));
6496 new = make_compound_operation (new, in_code);
6499 /* If we are have (and (rotate X C) M) and C is larger than the number
6500 of bits in M, this is an extraction. */
6502 else if (GET_CODE (XEXP (x, 0)) == ROTATE
6503 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6504 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0
6505 && i <= INTVAL (XEXP (XEXP (x, 0), 1)))
6507 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6508 new = make_extraction (mode, new,
6509 (GET_MODE_BITSIZE (mode)
6510 - INTVAL (XEXP (XEXP (x, 0), 1))),
6511 NULL_RTX, i, 1, 0, in_code == COMPARE);
6514 /* On machines without logical shifts, if the operand of the AND is
6515 a logical shift and our mask turns off all the propagated sign
6516 bits, we can replace the logical shift with an arithmetic shift. */
6517 else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6518 && !have_insn_for (LSHIFTRT, mode)
6519 && have_insn_for (ASHIFTRT, mode)
6520 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6521 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
6522 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
6523 && mode_width <= HOST_BITS_PER_WIDE_INT)
6525 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
6527 mask >>= INTVAL (XEXP (XEXP (x, 0), 1));
6528 if ((INTVAL (XEXP (x, 1)) & ~mask) == 0)
6529 SUBST (XEXP (x, 0),
6530 gen_rtx_ASHIFTRT (mode,
6531 make_compound_operation
6532 (XEXP (XEXP (x, 0), 0), next_code),
6533 XEXP (XEXP (x, 0), 1)));
6536 /* If the constant is one less than a power of two, this might be
6537 representable by an extraction even if no shift is present.
6538 If it doesn't end up being a ZERO_EXTEND, we will ignore it unless
6539 we are in a COMPARE. */
6540 else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6541 new = make_extraction (mode,
6542 make_compound_operation (XEXP (x, 0),
6543 next_code),
6544 0, NULL_RTX, i, 1, 0, in_code == COMPARE);
6546 /* If we are in a comparison and this is an AND with a power of two,
6547 convert this into the appropriate bit extract. */
6548 else if (in_code == COMPARE
6549 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
6550 new = make_extraction (mode,
6551 make_compound_operation (XEXP (x, 0),
6552 next_code),
6553 i, NULL_RTX, 1, 1, 0, 1);
6555 break;
6557 case LSHIFTRT:
6558 /* If the sign bit is known to be zero, replace this with an
6559 arithmetic shift. */
6560 if (have_insn_for (ASHIFTRT, mode)
6561 && ! have_insn_for (LSHIFTRT, mode)
6562 && mode_width <= HOST_BITS_PER_WIDE_INT
6563 && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0)
6565 new = gen_rtx_ASHIFTRT (mode,
6566 make_compound_operation (XEXP (x, 0),
6567 next_code),
6568 XEXP (x, 1));
6569 break;
6572 /* ... fall through ... */
6574 case ASHIFTRT:
6575 lhs = XEXP (x, 0);
6576 rhs = XEXP (x, 1);
6578 /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1,
6579 this is a SIGN_EXTRACT. */
6580 if (GET_CODE (rhs) == CONST_INT
6581 && GET_CODE (lhs) == ASHIFT
6582 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
6583 && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1)))
6585 new = make_compound_operation (XEXP (lhs, 0), next_code);
6586 new = make_extraction (mode, new,
6587 INTVAL (rhs) - INTVAL (XEXP (lhs, 1)),
6588 NULL_RTX, mode_width - INTVAL (rhs),
6589 code == LSHIFTRT, 0, in_code == COMPARE);
6590 break;
6593 /* See if we have operations between an ASHIFTRT and an ASHIFT.
6594 If so, try to merge the shifts into a SIGN_EXTEND. We could
6595 also do this for some cases of SIGN_EXTRACT, but it doesn't
6596 seem worth the effort; the case checked for occurs on Alpha. */
6598 if (!OBJECT_P (lhs)
6599 && ! (GET_CODE (lhs) == SUBREG
6600 && (OBJECT_P (SUBREG_REG (lhs))))
6601 && GET_CODE (rhs) == CONST_INT
6602 && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT
6603 && (new = extract_left_shift (lhs, INTVAL (rhs))) != 0)
6604 new = make_extraction (mode, make_compound_operation (new, next_code),
6605 0, NULL_RTX, mode_width - INTVAL (rhs),
6606 code == LSHIFTRT, 0, in_code == COMPARE);
6608 break;
6610 case SUBREG:
6611 /* Call ourselves recursively on the inner expression. If we are
6612 narrowing the object and it has a different RTL code from
6613 what it originally did, do this SUBREG as a force_to_mode. */
6615 tem = make_compound_operation (SUBREG_REG (x), in_code);
6618 rtx simplified;
6619 simplified = simplify_subreg (GET_MODE (x), tem, GET_MODE (tem),
6620 SUBREG_BYTE (x));
6622 if (simplified)
6623 tem = simplified;
6625 if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x))
6626 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem))
6627 && subreg_lowpart_p (x))
6629 rtx newer = force_to_mode (tem, mode, ~(HOST_WIDE_INT) 0,
6632 /* If we have something other than a SUBREG, we might have
6633 done an expansion, so rerun ourselves. */
6634 if (GET_CODE (newer) != SUBREG)
6635 newer = make_compound_operation (newer, in_code);
6637 return newer;
6640 if (simplified)
6641 return tem;
6643 break;
6645 default:
6646 break;
6649 if (new)
6651 x = gen_lowpart (mode, new);
6652 code = GET_CODE (x);
6655 /* Now recursively process each operand of this operation. */
6656 fmt = GET_RTX_FORMAT (code);
6657 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6658 if (fmt[i] == 'e')
6660 new = make_compound_operation (XEXP (x, i), next_code);
6661 SUBST (XEXP (x, i), new);
6664 /* If this is a commutative operation, the changes to the operands
6665 may have made it noncanonical. */
6666 if (COMMUTATIVE_ARITH_P (x)
6667 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
6669 tem = XEXP (x, 0);
6670 SUBST (XEXP (x, 0), XEXP (x, 1));
6671 SUBST (XEXP (x, 1), tem);
6674 return x;
6677 /* Given M see if it is a value that would select a field of bits
6678 within an item, but not the entire word. Return -1 if not.
6679 Otherwise, return the starting position of the field, where 0 is the
6680 low-order bit.
6682 *PLEN is set to the length of the field. */
6684 static int
6685 get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen)
6687 /* Get the bit number of the first 1 bit from the right, -1 if none. */
6688 int pos = exact_log2 (m & -m);
6689 int len = 0;
6691 if (pos >= 0)
6692 /* Now shift off the low-order zero bits and see if we have a
6693 power of two minus 1. */
6694 len = exact_log2 ((m >> pos) + 1);
6696 if (len <= 0)
6697 pos = -1;
6699 *plen = len;
6700 return pos;
6703 /* If X refers to a register that equals REG in value, replace these
6704 references with REG. */
6705 static rtx
6706 canon_reg_for_combine (rtx x, rtx reg)
6708 rtx op0, op1, op2;
6709 const char *fmt;
6710 int i;
6711 bool copied;
6713 enum rtx_code code = GET_CODE (x);
6714 switch (GET_RTX_CLASS (code))
6716 case RTX_UNARY:
6717 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6718 if (op0 != XEXP (x, 0))
6719 return simplify_gen_unary (GET_CODE (x), GET_MODE (x), op0,
6720 GET_MODE (reg));
6721 break;
6723 case RTX_BIN_ARITH:
6724 case RTX_COMM_ARITH:
6725 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6726 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6727 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6728 return simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1);
6729 break;
6731 case RTX_COMPARE:
6732 case RTX_COMM_COMPARE:
6733 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6734 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6735 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6736 return simplify_gen_relational (GET_CODE (x), GET_MODE (x),
6737 GET_MODE (op0), op0, op1);
6738 break;
6740 case RTX_TERNARY:
6741 case RTX_BITFIELD_OPS:
6742 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6743 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6744 op2 = canon_reg_for_combine (XEXP (x, 2), reg);
6745 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1) || op2 != XEXP (x, 2))
6746 return simplify_gen_ternary (GET_CODE (x), GET_MODE (x),
6747 GET_MODE (op0), op0, op1, op2);
6749 case RTX_OBJ:
6750 if (REG_P (x))
6752 if (rtx_equal_p (get_last_value (reg), x)
6753 || rtx_equal_p (reg, get_last_value (x)))
6754 return reg;
6755 else
6756 break;
6759 /* fall through */
6761 default:
6762 fmt = GET_RTX_FORMAT (code);
6763 copied = false;
6764 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6765 if (fmt[i] == 'e')
6767 rtx op = canon_reg_for_combine (XEXP (x, i), reg);
6768 if (op != XEXP (x, i))
6770 if (!copied)
6772 copied = true;
6773 x = copy_rtx (x);
6775 XEXP (x, i) = op;
6778 else if (fmt[i] == 'E')
6780 int j;
6781 for (j = 0; j < XVECLEN (x, i); j++)
6783 rtx op = canon_reg_for_combine (XVECEXP (x, i, j), reg);
6784 if (op != XVECEXP (x, i, j))
6786 if (!copied)
6788 copied = true;
6789 x = copy_rtx (x);
6791 XVECEXP (x, i, j) = op;
6796 break;
6799 return x;
6802 /* Return X converted to MODE. If the value is already truncated to
6803 MODE we can just return a subreg even though in the general case we
6804 would need an explicit truncation. */
6806 static rtx
6807 gen_lowpart_or_truncate (enum machine_mode mode, rtx x)
6809 if (GET_MODE_SIZE (GET_MODE (x)) <= GET_MODE_SIZE (mode)
6810 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
6811 GET_MODE_BITSIZE (GET_MODE (x)))
6812 || (REG_P (x) && reg_truncated_to_mode (mode, x)))
6813 return gen_lowpart (mode, x);
6814 else
6815 return simplify_gen_unary (TRUNCATE, mode, x, GET_MODE (x));
6818 /* See if X can be simplified knowing that we will only refer to it in
6819 MODE and will only refer to those bits that are nonzero in MASK.
6820 If other bits are being computed or if masking operations are done
6821 that select a superset of the bits in MASK, they can sometimes be
6822 ignored.
6824 Return a possibly simplified expression, but always convert X to
6825 MODE. If X is a CONST_INT, AND the CONST_INT with MASK.
6827 If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK
6828 are all off in X. This is used when X will be complemented, by either
6829 NOT, NEG, or XOR. */
6831 static rtx
6832 force_to_mode (rtx x, enum machine_mode mode, unsigned HOST_WIDE_INT mask,
6833 int just_select)
6835 enum rtx_code code = GET_CODE (x);
6836 int next_select = just_select || code == XOR || code == NOT || code == NEG;
6837 enum machine_mode op_mode;
6838 unsigned HOST_WIDE_INT fuller_mask, nonzero;
6839 rtx op0, op1, temp;
6841 /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the
6842 code below will do the wrong thing since the mode of such an
6843 expression is VOIDmode.
6845 Also do nothing if X is a CLOBBER; this can happen if X was
6846 the return value from a call to gen_lowpart. */
6847 if (code == CALL || code == ASM_OPERANDS || code == CLOBBER)
6848 return x;
6850 /* We want to perform the operation is its present mode unless we know
6851 that the operation is valid in MODE, in which case we do the operation
6852 in MODE. */
6853 op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x))
6854 && have_insn_for (code, mode))
6855 ? mode : GET_MODE (x));
6857 /* It is not valid to do a right-shift in a narrower mode
6858 than the one it came in with. */
6859 if ((code == LSHIFTRT || code == ASHIFTRT)
6860 && GET_MODE_BITSIZE (mode) < GET_MODE_BITSIZE (GET_MODE (x)))
6861 op_mode = GET_MODE (x);
6863 /* Truncate MASK to fit OP_MODE. */
6864 if (op_mode)
6865 mask &= GET_MODE_MASK (op_mode);
6867 /* When we have an arithmetic operation, or a shift whose count we
6868 do not know, we need to assume that all bits up to the highest-order
6869 bit in MASK will be needed. This is how we form such a mask. */
6870 if (mask & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)))
6871 fuller_mask = ~(unsigned HOST_WIDE_INT) 0;
6872 else
6873 fuller_mask = (((unsigned HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1))
6874 - 1);
6876 /* Determine what bits of X are guaranteed to be (non)zero. */
6877 nonzero = nonzero_bits (x, mode);
6879 /* If none of the bits in X are needed, return a zero. */
6880 if (!just_select && (nonzero & mask) == 0 && !side_effects_p (x))
6881 x = const0_rtx;
6883 /* If X is a CONST_INT, return a new one. Do this here since the
6884 test below will fail. */
6885 if (GET_CODE (x) == CONST_INT)
6887 if (SCALAR_INT_MODE_P (mode))
6888 return gen_int_mode (INTVAL (x) & mask, mode);
6889 else
6891 x = GEN_INT (INTVAL (x) & mask);
6892 return gen_lowpart_common (mode, x);
6896 /* If X is narrower than MODE and we want all the bits in X's mode, just
6897 get X in the proper mode. */
6898 if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)
6899 && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0)
6900 return gen_lowpart (mode, x);
6902 switch (code)
6904 case CLOBBER:
6905 /* If X is a (clobber (const_int)), return it since we know we are
6906 generating something that won't match. */
6907 return x;
6909 case SIGN_EXTEND:
6910 case ZERO_EXTEND:
6911 case ZERO_EXTRACT:
6912 case SIGN_EXTRACT:
6913 x = expand_compound_operation (x);
6914 if (GET_CODE (x) != code)
6915 return force_to_mode (x, mode, mask, next_select);
6916 break;
6918 case SUBREG:
6919 if (subreg_lowpart_p (x)
6920 /* We can ignore the effect of this SUBREG if it narrows the mode or
6921 if the constant masks to zero all the bits the mode doesn't
6922 have. */
6923 && ((GET_MODE_SIZE (GET_MODE (x))
6924 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
6925 || (0 == (mask
6926 & GET_MODE_MASK (GET_MODE (x))
6927 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))))))
6928 return force_to_mode (SUBREG_REG (x), mode, mask, next_select);
6929 break;
6931 case AND:
6932 /* If this is an AND with a constant, convert it into an AND
6933 whose constant is the AND of that constant with MASK. If it
6934 remains an AND of MASK, delete it since it is redundant. */
6936 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6938 x = simplify_and_const_int (x, op_mode, XEXP (x, 0),
6939 mask & INTVAL (XEXP (x, 1)));
6941 /* If X is still an AND, see if it is an AND with a mask that
6942 is just some low-order bits. If so, and it is MASK, we don't
6943 need it. */
6945 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6946 && ((INTVAL (XEXP (x, 1)) & GET_MODE_MASK (GET_MODE (x)))
6947 == mask))
6948 x = XEXP (x, 0);
6950 /* If it remains an AND, try making another AND with the bits
6951 in the mode mask that aren't in MASK turned on. If the
6952 constant in the AND is wide enough, this might make a
6953 cheaper constant. */
6955 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6956 && GET_MODE_MASK (GET_MODE (x)) != mask
6957 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
6959 HOST_WIDE_INT cval = (INTVAL (XEXP (x, 1))
6960 | (GET_MODE_MASK (GET_MODE (x)) & ~mask));
6961 int width = GET_MODE_BITSIZE (GET_MODE (x));
6962 rtx y;
6964 /* If MODE is narrower than HOST_WIDE_INT and CVAL is a negative
6965 number, sign extend it. */
6966 if (width > 0 && width < HOST_BITS_PER_WIDE_INT
6967 && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6968 cval |= (HOST_WIDE_INT) -1 << width;
6970 y = simplify_gen_binary (AND, GET_MODE (x),
6971 XEXP (x, 0), GEN_INT (cval));
6972 if (rtx_cost (y, SET) < rtx_cost (x, SET))
6973 x = y;
6976 break;
6979 goto binop;
6981 case PLUS:
6982 /* In (and (plus FOO C1) M), if M is a mask that just turns off
6983 low-order bits (as in an alignment operation) and FOO is already
6984 aligned to that boundary, mask C1 to that boundary as well.
6985 This may eliminate that PLUS and, later, the AND. */
6988 unsigned int width = GET_MODE_BITSIZE (mode);
6989 unsigned HOST_WIDE_INT smask = mask;
6991 /* If MODE is narrower than HOST_WIDE_INT and mask is a negative
6992 number, sign extend it. */
6994 if (width < HOST_BITS_PER_WIDE_INT
6995 && (smask & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6996 smask |= (HOST_WIDE_INT) -1 << width;
6998 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6999 && exact_log2 (- smask) >= 0
7000 && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0
7001 && (INTVAL (XEXP (x, 1)) & ~smask) != 0)
7002 return force_to_mode (plus_constant (XEXP (x, 0),
7003 (INTVAL (XEXP (x, 1)) & smask)),
7004 mode, smask, next_select);
7007 /* ... fall through ... */
7009 case MULT:
7010 /* For PLUS, MINUS and MULT, we need any bits less significant than the
7011 most significant bit in MASK since carries from those bits will
7012 affect the bits we are interested in. */
7013 mask = fuller_mask;
7014 goto binop;
7016 case MINUS:
7017 /* If X is (minus C Y) where C's least set bit is larger than any bit
7018 in the mask, then we may replace with (neg Y). */
7019 if (GET_CODE (XEXP (x, 0)) == CONST_INT
7020 && (((unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 0))
7021 & -INTVAL (XEXP (x, 0))))
7022 > mask))
7024 x = simplify_gen_unary (NEG, GET_MODE (x), XEXP (x, 1),
7025 GET_MODE (x));
7026 return force_to_mode (x, mode, mask, next_select);
7029 /* Similarly, if C contains every bit in the fuller_mask, then we may
7030 replace with (not Y). */
7031 if (GET_CODE (XEXP (x, 0)) == CONST_INT
7032 && ((INTVAL (XEXP (x, 0)) | (HOST_WIDE_INT) fuller_mask)
7033 == INTVAL (XEXP (x, 0))))
7035 x = simplify_gen_unary (NOT, GET_MODE (x),
7036 XEXP (x, 1), GET_MODE (x));
7037 return force_to_mode (x, mode, mask, next_select);
7040 mask = fuller_mask;
7041 goto binop;
7043 case IOR:
7044 case XOR:
7045 /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and
7046 LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...)
7047 operation which may be a bitfield extraction. Ensure that the
7048 constant we form is not wider than the mode of X. */
7050 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7051 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7052 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7053 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
7054 && GET_CODE (XEXP (x, 1)) == CONST_INT
7055 && ((INTVAL (XEXP (XEXP (x, 0), 1))
7056 + floor_log2 (INTVAL (XEXP (x, 1))))
7057 < GET_MODE_BITSIZE (GET_MODE (x)))
7058 && (INTVAL (XEXP (x, 1))
7059 & ~nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0)
7061 temp = GEN_INT ((INTVAL (XEXP (x, 1)) & mask)
7062 << INTVAL (XEXP (XEXP (x, 0), 1)));
7063 temp = simplify_gen_binary (GET_CODE (x), GET_MODE (x),
7064 XEXP (XEXP (x, 0), 0), temp);
7065 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), temp,
7066 XEXP (XEXP (x, 0), 1));
7067 return force_to_mode (x, mode, mask, next_select);
7070 binop:
7071 /* For most binary operations, just propagate into the operation and
7072 change the mode if we have an operation of that mode. */
7074 op0 = gen_lowpart_or_truncate (op_mode,
7075 force_to_mode (XEXP (x, 0), mode, mask,
7076 next_select));
7077 op1 = gen_lowpart_or_truncate (op_mode,
7078 force_to_mode (XEXP (x, 1), mode, mask,
7079 next_select));
7081 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7082 x = simplify_gen_binary (code, op_mode, op0, op1);
7083 break;
7085 case ASHIFT:
7086 /* For left shifts, do the same, but just for the first operand.
7087 However, we cannot do anything with shifts where we cannot
7088 guarantee that the counts are smaller than the size of the mode
7089 because such a count will have a different meaning in a
7090 wider mode. */
7092 if (! (GET_CODE (XEXP (x, 1)) == CONST_INT
7093 && INTVAL (XEXP (x, 1)) >= 0
7094 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode))
7095 && ! (GET_MODE (XEXP (x, 1)) != VOIDmode
7096 && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1)))
7097 < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode))))
7098 break;
7100 /* If the shift count is a constant and we can do arithmetic in
7101 the mode of the shift, refine which bits we need. Otherwise, use the
7102 conservative form of the mask. */
7103 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7104 && INTVAL (XEXP (x, 1)) >= 0
7105 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (op_mode)
7106 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7107 mask >>= INTVAL (XEXP (x, 1));
7108 else
7109 mask = fuller_mask;
7111 op0 = gen_lowpart_or_truncate (op_mode,
7112 force_to_mode (XEXP (x, 0), op_mode,
7113 mask, next_select));
7115 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7116 x = simplify_gen_binary (code, op_mode, op0, XEXP (x, 1));
7117 break;
7119 case LSHIFTRT:
7120 /* Here we can only do something if the shift count is a constant,
7121 this shift constant is valid for the host, and we can do arithmetic
7122 in OP_MODE. */
7124 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7125 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
7126 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7128 rtx inner = XEXP (x, 0);
7129 unsigned HOST_WIDE_INT inner_mask;
7131 /* Select the mask of the bits we need for the shift operand. */
7132 inner_mask = mask << INTVAL (XEXP (x, 1));
7134 /* We can only change the mode of the shift if we can do arithmetic
7135 in the mode of the shift and INNER_MASK is no wider than the
7136 width of X's mode. */
7137 if ((inner_mask & ~GET_MODE_MASK (GET_MODE (x))) != 0)
7138 op_mode = GET_MODE (x);
7140 inner = force_to_mode (inner, op_mode, inner_mask, next_select);
7142 if (GET_MODE (x) != op_mode || inner != XEXP (x, 0))
7143 x = simplify_gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1));
7146 /* If we have (and (lshiftrt FOO C1) C2) where the combination of the
7147 shift and AND produces only copies of the sign bit (C2 is one less
7148 than a power of two), we can do this with just a shift. */
7150 if (GET_CODE (x) == LSHIFTRT
7151 && GET_CODE (XEXP (x, 1)) == CONST_INT
7152 /* The shift puts one of the sign bit copies in the least significant
7153 bit. */
7154 && ((INTVAL (XEXP (x, 1))
7155 + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))))
7156 >= GET_MODE_BITSIZE (GET_MODE (x)))
7157 && exact_log2 (mask + 1) >= 0
7158 /* Number of bits left after the shift must be more than the mask
7159 needs. */
7160 && ((INTVAL (XEXP (x, 1)) + exact_log2 (mask + 1))
7161 <= GET_MODE_BITSIZE (GET_MODE (x)))
7162 /* Must be more sign bit copies than the mask needs. */
7163 && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
7164 >= exact_log2 (mask + 1)))
7165 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7166 GEN_INT (GET_MODE_BITSIZE (GET_MODE (x))
7167 - exact_log2 (mask + 1)));
7169 goto shiftrt;
7171 case ASHIFTRT:
7172 /* If we are just looking for the sign bit, we don't need this shift at
7173 all, even if it has a variable count. */
7174 if (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
7175 && (mask == ((unsigned HOST_WIDE_INT) 1
7176 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
7177 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7179 /* If this is a shift by a constant, get a mask that contains those bits
7180 that are not copies of the sign bit. We then have two cases: If
7181 MASK only includes those bits, this can be a logical shift, which may
7182 allow simplifications. If MASK is a single-bit field not within
7183 those bits, we are requesting a copy of the sign bit and hence can
7184 shift the sign bit to the appropriate location. */
7186 if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0
7187 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT)
7189 int i;
7191 /* If the considered data is wider than HOST_WIDE_INT, we can't
7192 represent a mask for all its bits in a single scalar.
7193 But we only care about the lower bits, so calculate these. */
7195 if (GET_MODE_BITSIZE (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT)
7197 nonzero = ~(HOST_WIDE_INT) 0;
7199 /* GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7200 is the number of bits a full-width mask would have set.
7201 We need only shift if these are fewer than nonzero can
7202 hold. If not, we must keep all bits set in nonzero. */
7204 if (GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7205 < HOST_BITS_PER_WIDE_INT)
7206 nonzero >>= INTVAL (XEXP (x, 1))
7207 + HOST_BITS_PER_WIDE_INT
7208 - GET_MODE_BITSIZE (GET_MODE (x)) ;
7210 else
7212 nonzero = GET_MODE_MASK (GET_MODE (x));
7213 nonzero >>= INTVAL (XEXP (x, 1));
7216 if ((mask & ~nonzero) == 0)
7218 x = simplify_shift_const (NULL_RTX, LSHIFTRT, GET_MODE (x),
7219 XEXP (x, 0), INTVAL (XEXP (x, 1)));
7220 if (GET_CODE (x) != ASHIFTRT)
7221 return force_to_mode (x, mode, mask, next_select);
7224 else if ((i = exact_log2 (mask)) >= 0)
7226 x = simplify_shift_const
7227 (NULL_RTX, LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7228 GET_MODE_BITSIZE (GET_MODE (x)) - 1 - i);
7230 if (GET_CODE (x) != ASHIFTRT)
7231 return force_to_mode (x, mode, mask, next_select);
7235 /* If MASK is 1, convert this to an LSHIFTRT. This can be done
7236 even if the shift count isn't a constant. */
7237 if (mask == 1)
7238 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7239 XEXP (x, 0), XEXP (x, 1));
7241 shiftrt:
7243 /* If this is a zero- or sign-extension operation that just affects bits
7244 we don't care about, remove it. Be sure the call above returned
7245 something that is still a shift. */
7247 if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT)
7248 && GET_CODE (XEXP (x, 1)) == CONST_INT
7249 && INTVAL (XEXP (x, 1)) >= 0
7250 && (INTVAL (XEXP (x, 1))
7251 <= GET_MODE_BITSIZE (GET_MODE (x)) - (floor_log2 (mask) + 1))
7252 && GET_CODE (XEXP (x, 0)) == ASHIFT
7253 && XEXP (XEXP (x, 0), 1) == XEXP (x, 1))
7254 return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask,
7255 next_select);
7257 break;
7259 case ROTATE:
7260 case ROTATERT:
7261 /* If the shift count is constant and we can do computations
7262 in the mode of X, compute where the bits we care about are.
7263 Otherwise, we can't do anything. Don't change the mode of
7264 the shift or propagate MODE into the shift, though. */
7265 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7266 && INTVAL (XEXP (x, 1)) >= 0)
7268 temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE,
7269 GET_MODE (x), GEN_INT (mask),
7270 XEXP (x, 1));
7271 if (temp && GET_CODE (temp) == CONST_INT)
7272 SUBST (XEXP (x, 0),
7273 force_to_mode (XEXP (x, 0), GET_MODE (x),
7274 INTVAL (temp), next_select));
7276 break;
7278 case NEG:
7279 /* If we just want the low-order bit, the NEG isn't needed since it
7280 won't change the low-order bit. */
7281 if (mask == 1)
7282 return force_to_mode (XEXP (x, 0), mode, mask, just_select);
7284 /* We need any bits less significant than the most significant bit in
7285 MASK since carries from those bits will affect the bits we are
7286 interested in. */
7287 mask = fuller_mask;
7288 goto unop;
7290 case NOT:
7291 /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the
7292 same as the XOR case above. Ensure that the constant we form is not
7293 wider than the mode of X. */
7295 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7296 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7297 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7298 && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask)
7299 < GET_MODE_BITSIZE (GET_MODE (x)))
7300 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT)
7302 temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)),
7303 GET_MODE (x));
7304 temp = simplify_gen_binary (XOR, GET_MODE (x),
7305 XEXP (XEXP (x, 0), 0), temp);
7306 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7307 temp, XEXP (XEXP (x, 0), 1));
7309 return force_to_mode (x, mode, mask, next_select);
7312 /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must
7313 use the full mask inside the NOT. */
7314 mask = fuller_mask;
7316 unop:
7317 op0 = gen_lowpart_or_truncate (op_mode,
7318 force_to_mode (XEXP (x, 0), mode, mask,
7319 next_select));
7320 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7321 x = simplify_gen_unary (code, op_mode, op0, op_mode);
7322 break;
7324 case NE:
7325 /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included
7326 in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero,
7327 which is equal to STORE_FLAG_VALUE. */
7328 if ((mask & ~STORE_FLAG_VALUE) == 0 && XEXP (x, 1) == const0_rtx
7329 && GET_MODE (XEXP (x, 0)) == mode
7330 && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0
7331 && (nonzero_bits (XEXP (x, 0), mode)
7332 == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE))
7333 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7335 break;
7337 case IF_THEN_ELSE:
7338 /* We have no way of knowing if the IF_THEN_ELSE can itself be
7339 written in a narrower mode. We play it safe and do not do so. */
7341 SUBST (XEXP (x, 1),
7342 gen_lowpart_or_truncate (GET_MODE (x),
7343 force_to_mode (XEXP (x, 1), mode,
7344 mask, next_select)));
7345 SUBST (XEXP (x, 2),
7346 gen_lowpart_or_truncate (GET_MODE (x),
7347 force_to_mode (XEXP (x, 2), mode,
7348 mask, next_select)));
7349 break;
7351 default:
7352 break;
7355 /* Ensure we return a value of the proper mode. */
7356 return gen_lowpart_or_truncate (mode, x);
7359 /* Return nonzero if X is an expression that has one of two values depending on
7360 whether some other value is zero or nonzero. In that case, we return the
7361 value that is being tested, *PTRUE is set to the value if the rtx being
7362 returned has a nonzero value, and *PFALSE is set to the other alternative.
7364 If we return zero, we set *PTRUE and *PFALSE to X. */
7366 static rtx
7367 if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse)
7369 enum machine_mode mode = GET_MODE (x);
7370 enum rtx_code code = GET_CODE (x);
7371 rtx cond0, cond1, true0, true1, false0, false1;
7372 unsigned HOST_WIDE_INT nz;
7374 /* If we are comparing a value against zero, we are done. */
7375 if ((code == NE || code == EQ)
7376 && XEXP (x, 1) == const0_rtx)
7378 *ptrue = (code == NE) ? const_true_rtx : const0_rtx;
7379 *pfalse = (code == NE) ? const0_rtx : const_true_rtx;
7380 return XEXP (x, 0);
7383 /* If this is a unary operation whose operand has one of two values, apply
7384 our opcode to compute those values. */
7385 else if (UNARY_P (x)
7386 && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0)
7388 *ptrue = simplify_gen_unary (code, mode, true0, GET_MODE (XEXP (x, 0)));
7389 *pfalse = simplify_gen_unary (code, mode, false0,
7390 GET_MODE (XEXP (x, 0)));
7391 return cond0;
7394 /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would
7395 make can't possibly match and would suppress other optimizations. */
7396 else if (code == COMPARE)
7399 /* If this is a binary operation, see if either side has only one of two
7400 values. If either one does or if both do and they are conditional on
7401 the same value, compute the new true and false values. */
7402 else if (BINARY_P (x))
7404 cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0);
7405 cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1);
7407 if ((cond0 != 0 || cond1 != 0)
7408 && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1)))
7410 /* If if_then_else_cond returned zero, then true/false are the
7411 same rtl. We must copy one of them to prevent invalid rtl
7412 sharing. */
7413 if (cond0 == 0)
7414 true0 = copy_rtx (true0);
7415 else if (cond1 == 0)
7416 true1 = copy_rtx (true1);
7418 if (COMPARISON_P (x))
7420 *ptrue = simplify_gen_relational (code, mode, VOIDmode,
7421 true0, true1);
7422 *pfalse = simplify_gen_relational (code, mode, VOIDmode,
7423 false0, false1);
7425 else
7427 *ptrue = simplify_gen_binary (code, mode, true0, true1);
7428 *pfalse = simplify_gen_binary (code, mode, false0, false1);
7431 return cond0 ? cond0 : cond1;
7434 /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the
7435 operands is zero when the other is nonzero, and vice-versa,
7436 and STORE_FLAG_VALUE is 1 or -1. */
7438 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7439 && (code == PLUS || code == IOR || code == XOR || code == MINUS
7440 || code == UMAX)
7441 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7443 rtx op0 = XEXP (XEXP (x, 0), 1);
7444 rtx op1 = XEXP (XEXP (x, 1), 1);
7446 cond0 = XEXP (XEXP (x, 0), 0);
7447 cond1 = XEXP (XEXP (x, 1), 0);
7449 if (COMPARISON_P (cond0)
7450 && COMPARISON_P (cond1)
7451 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7452 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7453 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7454 || ((swap_condition (GET_CODE (cond0))
7455 == reversed_comparison_code (cond1, NULL))
7456 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7457 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7458 && ! side_effects_p (x))
7460 *ptrue = simplify_gen_binary (MULT, mode, op0, const_true_rtx);
7461 *pfalse = simplify_gen_binary (MULT, mode,
7462 (code == MINUS
7463 ? simplify_gen_unary (NEG, mode,
7464 op1, mode)
7465 : op1),
7466 const_true_rtx);
7467 return cond0;
7471 /* Similarly for MULT, AND and UMIN, except that for these the result
7472 is always zero. */
7473 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7474 && (code == MULT || code == AND || code == UMIN)
7475 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7477 cond0 = XEXP (XEXP (x, 0), 0);
7478 cond1 = XEXP (XEXP (x, 1), 0);
7480 if (COMPARISON_P (cond0)
7481 && COMPARISON_P (cond1)
7482 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7483 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7484 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7485 || ((swap_condition (GET_CODE (cond0))
7486 == reversed_comparison_code (cond1, NULL))
7487 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7488 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7489 && ! side_effects_p (x))
7491 *ptrue = *pfalse = const0_rtx;
7492 return cond0;
7497 else if (code == IF_THEN_ELSE)
7499 /* If we have IF_THEN_ELSE already, extract the condition and
7500 canonicalize it if it is NE or EQ. */
7501 cond0 = XEXP (x, 0);
7502 *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2);
7503 if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx)
7504 return XEXP (cond0, 0);
7505 else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx)
7507 *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1);
7508 return XEXP (cond0, 0);
7510 else
7511 return cond0;
7514 /* If X is a SUBREG, we can narrow both the true and false values
7515 if the inner expression, if there is a condition. */
7516 else if (code == SUBREG
7517 && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x),
7518 &true0, &false0)))
7520 true0 = simplify_gen_subreg (mode, true0,
7521 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7522 false0 = simplify_gen_subreg (mode, false0,
7523 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7524 if (true0 && false0)
7526 *ptrue = true0;
7527 *pfalse = false0;
7528 return cond0;
7532 /* If X is a constant, this isn't special and will cause confusions
7533 if we treat it as such. Likewise if it is equivalent to a constant. */
7534 else if (CONSTANT_P (x)
7535 || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0)))
7538 /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that
7539 will be least confusing to the rest of the compiler. */
7540 else if (mode == BImode)
7542 *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx;
7543 return x;
7546 /* If X is known to be either 0 or -1, those are the true and
7547 false values when testing X. */
7548 else if (x == constm1_rtx || x == const0_rtx
7549 || (mode != VOIDmode
7550 && num_sign_bit_copies (x, mode) == GET_MODE_BITSIZE (mode)))
7552 *ptrue = constm1_rtx, *pfalse = const0_rtx;
7553 return x;
7556 /* Likewise for 0 or a single bit. */
7557 else if (SCALAR_INT_MODE_P (mode)
7558 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
7559 && exact_log2 (nz = nonzero_bits (x, mode)) >= 0)
7561 *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx;
7562 return x;
7565 /* Otherwise fail; show no condition with true and false values the same. */
7566 *ptrue = *pfalse = x;
7567 return 0;
7570 /* Return the value of expression X given the fact that condition COND
7571 is known to be true when applied to REG as its first operand and VAL
7572 as its second. X is known to not be shared and so can be modified in
7573 place.
7575 We only handle the simplest cases, and specifically those cases that
7576 arise with IF_THEN_ELSE expressions. */
7578 static rtx
7579 known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val)
7581 enum rtx_code code = GET_CODE (x);
7582 rtx temp;
7583 const char *fmt;
7584 int i, j;
7586 if (side_effects_p (x))
7587 return x;
7589 /* If either operand of the condition is a floating point value,
7590 then we have to avoid collapsing an EQ comparison. */
7591 if (cond == EQ
7592 && rtx_equal_p (x, reg)
7593 && ! FLOAT_MODE_P (GET_MODE (x))
7594 && ! FLOAT_MODE_P (GET_MODE (val)))
7595 return val;
7597 if (cond == UNEQ && rtx_equal_p (x, reg))
7598 return val;
7600 /* If X is (abs REG) and we know something about REG's relationship
7601 with zero, we may be able to simplify this. */
7603 if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx)
7604 switch (cond)
7606 case GE: case GT: case EQ:
7607 return XEXP (x, 0);
7608 case LT: case LE:
7609 return simplify_gen_unary (NEG, GET_MODE (XEXP (x, 0)),
7610 XEXP (x, 0),
7611 GET_MODE (XEXP (x, 0)));
7612 default:
7613 break;
7616 /* The only other cases we handle are MIN, MAX, and comparisons if the
7617 operands are the same as REG and VAL. */
7619 else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x))
7621 if (rtx_equal_p (XEXP (x, 0), val))
7622 cond = swap_condition (cond), temp = val, val = reg, reg = temp;
7624 if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val))
7626 if (COMPARISON_P (x))
7628 if (comparison_dominates_p (cond, code))
7629 return const_true_rtx;
7631 code = reversed_comparison_code (x, NULL);
7632 if (code != UNKNOWN
7633 && comparison_dominates_p (cond, code))
7634 return const0_rtx;
7635 else
7636 return x;
7638 else if (code == SMAX || code == SMIN
7639 || code == UMIN || code == UMAX)
7641 int unsignedp = (code == UMIN || code == UMAX);
7643 /* Do not reverse the condition when it is NE or EQ.
7644 This is because we cannot conclude anything about
7645 the value of 'SMAX (x, y)' when x is not equal to y,
7646 but we can when x equals y. */
7647 if ((code == SMAX || code == UMAX)
7648 && ! (cond == EQ || cond == NE))
7649 cond = reverse_condition (cond);
7651 switch (cond)
7653 case GE: case GT:
7654 return unsignedp ? x : XEXP (x, 1);
7655 case LE: case LT:
7656 return unsignedp ? x : XEXP (x, 0);
7657 case GEU: case GTU:
7658 return unsignedp ? XEXP (x, 1) : x;
7659 case LEU: case LTU:
7660 return unsignedp ? XEXP (x, 0) : x;
7661 default:
7662 break;
7667 else if (code == SUBREG)
7669 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (x));
7670 rtx new, r = known_cond (SUBREG_REG (x), cond, reg, val);
7672 if (SUBREG_REG (x) != r)
7674 /* We must simplify subreg here, before we lose track of the
7675 original inner_mode. */
7676 new = simplify_subreg (GET_MODE (x), r,
7677 inner_mode, SUBREG_BYTE (x));
7678 if (new)
7679 return new;
7680 else
7681 SUBST (SUBREG_REG (x), r);
7684 return x;
7686 /* We don't have to handle SIGN_EXTEND here, because even in the
7687 case of replacing something with a modeless CONST_INT, a
7688 CONST_INT is already (supposed to be) a valid sign extension for
7689 its narrower mode, which implies it's already properly
7690 sign-extended for the wider mode. Now, for ZERO_EXTEND, the
7691 story is different. */
7692 else if (code == ZERO_EXTEND)
7694 enum machine_mode inner_mode = GET_MODE (XEXP (x, 0));
7695 rtx new, r = known_cond (XEXP (x, 0), cond, reg, val);
7697 if (XEXP (x, 0) != r)
7699 /* We must simplify the zero_extend here, before we lose
7700 track of the original inner_mode. */
7701 new = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
7702 r, inner_mode);
7703 if (new)
7704 return new;
7705 else
7706 SUBST (XEXP (x, 0), r);
7709 return x;
7712 fmt = GET_RTX_FORMAT (code);
7713 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7715 if (fmt[i] == 'e')
7716 SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val));
7717 else if (fmt[i] == 'E')
7718 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7719 SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j),
7720 cond, reg, val));
7723 return x;
7726 /* See if X and Y are equal for the purposes of seeing if we can rewrite an
7727 assignment as a field assignment. */
7729 static int
7730 rtx_equal_for_field_assignment_p (rtx x, rtx y)
7732 if (x == y || rtx_equal_p (x, y))
7733 return 1;
7735 if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y))
7736 return 0;
7738 /* Check for a paradoxical SUBREG of a MEM compared with the MEM.
7739 Note that all SUBREGs of MEM are paradoxical; otherwise they
7740 would have been rewritten. */
7741 if (MEM_P (x) && GET_CODE (y) == SUBREG
7742 && MEM_P (SUBREG_REG (y))
7743 && rtx_equal_p (SUBREG_REG (y),
7744 gen_lowpart (GET_MODE (SUBREG_REG (y)), x)))
7745 return 1;
7747 if (MEM_P (y) && GET_CODE (x) == SUBREG
7748 && MEM_P (SUBREG_REG (x))
7749 && rtx_equal_p (SUBREG_REG (x),
7750 gen_lowpart (GET_MODE (SUBREG_REG (x)), y)))
7751 return 1;
7753 /* We used to see if get_last_value of X and Y were the same but that's
7754 not correct. In one direction, we'll cause the assignment to have
7755 the wrong destination and in the case, we'll import a register into this
7756 insn that might have already have been dead. So fail if none of the
7757 above cases are true. */
7758 return 0;
7761 /* See if X, a SET operation, can be rewritten as a bit-field assignment.
7762 Return that assignment if so.
7764 We only handle the most common cases. */
7766 static rtx
7767 make_field_assignment (rtx x)
7769 rtx dest = SET_DEST (x);
7770 rtx src = SET_SRC (x);
7771 rtx assign;
7772 rtx rhs, lhs;
7773 HOST_WIDE_INT c1;
7774 HOST_WIDE_INT pos;
7775 unsigned HOST_WIDE_INT len;
7776 rtx other;
7777 enum machine_mode mode;
7779 /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is
7780 a clear of a one-bit field. We will have changed it to
7781 (and (rotate (const_int -2) POS) DEST), so check for that. Also check
7782 for a SUBREG. */
7784 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE
7785 && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT
7786 && INTVAL (XEXP (XEXP (src, 0), 0)) == -2
7787 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7789 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7790 1, 1, 1, 0);
7791 if (assign != 0)
7792 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7793 return x;
7796 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG
7797 && subreg_lowpart_p (XEXP (src, 0))
7798 && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0)))
7799 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0)))))
7800 && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE
7801 && GET_CODE (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == CONST_INT
7802 && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2
7803 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7805 assign = make_extraction (VOIDmode, dest, 0,
7806 XEXP (SUBREG_REG (XEXP (src, 0)), 1),
7807 1, 1, 1, 0);
7808 if (assign != 0)
7809 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7810 return x;
7813 /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a
7814 one-bit field. */
7815 if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT
7816 && XEXP (XEXP (src, 0), 0) == const1_rtx
7817 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7819 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7820 1, 1, 1, 0);
7821 if (assign != 0)
7822 return gen_rtx_SET (VOIDmode, assign, const1_rtx);
7823 return x;
7826 /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the
7827 SRC is an AND with all bits of that field set, then we can discard
7828 the AND. */
7829 if (GET_CODE (dest) == ZERO_EXTRACT
7830 && GET_CODE (XEXP (dest, 1)) == CONST_INT
7831 && GET_CODE (src) == AND
7832 && GET_CODE (XEXP (src, 1)) == CONST_INT)
7834 HOST_WIDE_INT width = INTVAL (XEXP (dest, 1));
7835 unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1));
7836 unsigned HOST_WIDE_INT ze_mask;
7838 if (width >= HOST_BITS_PER_WIDE_INT)
7839 ze_mask = -1;
7840 else
7841 ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1;
7843 /* Complete overlap. We can remove the source AND. */
7844 if ((and_mask & ze_mask) == ze_mask)
7845 return gen_rtx_SET (VOIDmode, dest, XEXP (src, 0));
7847 /* Partial overlap. We can reduce the source AND. */
7848 if ((and_mask & ze_mask) != and_mask)
7850 mode = GET_MODE (src);
7851 src = gen_rtx_AND (mode, XEXP (src, 0),
7852 gen_int_mode (and_mask & ze_mask, mode));
7853 return gen_rtx_SET (VOIDmode, dest, src);
7857 /* The other case we handle is assignments into a constant-position
7858 field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents
7859 a mask that has all one bits except for a group of zero bits and
7860 OTHER is known to have zeros where C1 has ones, this is such an
7861 assignment. Compute the position and length from C1. Shift OTHER
7862 to the appropriate position, force it to the required mode, and
7863 make the extraction. Check for the AND in both operands. */
7865 if (GET_CODE (src) != IOR && GET_CODE (src) != XOR)
7866 return x;
7868 rhs = expand_compound_operation (XEXP (src, 0));
7869 lhs = expand_compound_operation (XEXP (src, 1));
7871 if (GET_CODE (rhs) == AND
7872 && GET_CODE (XEXP (rhs, 1)) == CONST_INT
7873 && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest))
7874 c1 = INTVAL (XEXP (rhs, 1)), other = lhs;
7875 else if (GET_CODE (lhs) == AND
7876 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
7877 && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest))
7878 c1 = INTVAL (XEXP (lhs, 1)), other = rhs;
7879 else
7880 return x;
7882 pos = get_pos_from_mask ((~c1) & GET_MODE_MASK (GET_MODE (dest)), &len);
7883 if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest))
7884 || GET_MODE_BITSIZE (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT
7885 || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0)
7886 return x;
7888 assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0);
7889 if (assign == 0)
7890 return x;
7892 /* The mode to use for the source is the mode of the assignment, or of
7893 what is inside a possible STRICT_LOW_PART. */
7894 mode = (GET_CODE (assign) == STRICT_LOW_PART
7895 ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign));
7897 /* Shift OTHER right POS places and make it the source, restricting it
7898 to the proper length and mode. */
7900 src = canon_reg_for_combine (simplify_shift_const (NULL_RTX, LSHIFTRT,
7901 GET_MODE (src),
7902 other, pos),
7903 dest);
7904 src = force_to_mode (src, mode,
7905 GET_MODE_BITSIZE (mode) >= HOST_BITS_PER_WIDE_INT
7906 ? ~(unsigned HOST_WIDE_INT) 0
7907 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
7910 /* If SRC is masked by an AND that does not make a difference in
7911 the value being stored, strip it. */
7912 if (GET_CODE (assign) == ZERO_EXTRACT
7913 && GET_CODE (XEXP (assign, 1)) == CONST_INT
7914 && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT
7915 && GET_CODE (src) == AND
7916 && GET_CODE (XEXP (src, 1)) == CONST_INT
7917 && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (src, 1))
7918 == ((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (assign, 1))) - 1))
7919 src = XEXP (src, 0);
7921 return gen_rtx_SET (VOIDmode, assign, src);
7924 /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c)
7925 if so. */
7927 static rtx
7928 apply_distributive_law (rtx x)
7930 enum rtx_code code = GET_CODE (x);
7931 enum rtx_code inner_code;
7932 rtx lhs, rhs, other;
7933 rtx tem;
7935 /* Distributivity is not true for floating point as it can change the
7936 value. So we don't do it unless -funsafe-math-optimizations. */
7937 if (FLOAT_MODE_P (GET_MODE (x))
7938 && ! flag_unsafe_math_optimizations)
7939 return x;
7941 /* The outer operation can only be one of the following: */
7942 if (code != IOR && code != AND && code != XOR
7943 && code != PLUS && code != MINUS)
7944 return x;
7946 lhs = XEXP (x, 0);
7947 rhs = XEXP (x, 1);
7949 /* If either operand is a primitive we can't do anything, so get out
7950 fast. */
7951 if (OBJECT_P (lhs) || OBJECT_P (rhs))
7952 return x;
7954 lhs = expand_compound_operation (lhs);
7955 rhs = expand_compound_operation (rhs);
7956 inner_code = GET_CODE (lhs);
7957 if (inner_code != GET_CODE (rhs))
7958 return x;
7960 /* See if the inner and outer operations distribute. */
7961 switch (inner_code)
7963 case LSHIFTRT:
7964 case ASHIFTRT:
7965 case AND:
7966 case IOR:
7967 /* These all distribute except over PLUS. */
7968 if (code == PLUS || code == MINUS)
7969 return x;
7970 break;
7972 case MULT:
7973 if (code != PLUS && code != MINUS)
7974 return x;
7975 break;
7977 case ASHIFT:
7978 /* This is also a multiply, so it distributes over everything. */
7979 break;
7981 case SUBREG:
7982 /* Non-paradoxical SUBREGs distributes over all operations,
7983 provided the inner modes and byte offsets are the same, this
7984 is an extraction of a low-order part, we don't convert an fp
7985 operation to int or vice versa, this is not a vector mode,
7986 and we would not be converting a single-word operation into a
7987 multi-word operation. The latter test is not required, but
7988 it prevents generating unneeded multi-word operations. Some
7989 of the previous tests are redundant given the latter test,
7990 but are retained because they are required for correctness.
7992 We produce the result slightly differently in this case. */
7994 if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs))
7995 || SUBREG_BYTE (lhs) != SUBREG_BYTE (rhs)
7996 || ! subreg_lowpart_p (lhs)
7997 || (GET_MODE_CLASS (GET_MODE (lhs))
7998 != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs))))
7999 || (GET_MODE_SIZE (GET_MODE (lhs))
8000 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))))
8001 || VECTOR_MODE_P (GET_MODE (lhs))
8002 || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD
8003 /* Result might need to be truncated. Don't change mode if
8004 explicit truncation is needed. */
8005 || !TRULY_NOOP_TRUNCATION
8006 (GET_MODE_BITSIZE (GET_MODE (x)),
8007 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (lhs)))))
8008 return x;
8010 tem = simplify_gen_binary (code, GET_MODE (SUBREG_REG (lhs)),
8011 SUBREG_REG (lhs), SUBREG_REG (rhs));
8012 return gen_lowpart (GET_MODE (x), tem);
8014 default:
8015 return x;
8018 /* Set LHS and RHS to the inner operands (A and B in the example
8019 above) and set OTHER to the common operand (C in the example).
8020 There is only one way to do this unless the inner operation is
8021 commutative. */
8022 if (COMMUTATIVE_ARITH_P (lhs)
8023 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0)))
8024 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1);
8025 else if (COMMUTATIVE_ARITH_P (lhs)
8026 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1)))
8027 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0);
8028 else if (COMMUTATIVE_ARITH_P (lhs)
8029 && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0)))
8030 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1);
8031 else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1)))
8032 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0);
8033 else
8034 return x;
8036 /* Form the new inner operation, seeing if it simplifies first. */
8037 tem = simplify_gen_binary (code, GET_MODE (x), lhs, rhs);
8039 /* There is one exception to the general way of distributing:
8040 (a | c) ^ (b | c) -> (a ^ b) & ~c */
8041 if (code == XOR && inner_code == IOR)
8043 inner_code = AND;
8044 other = simplify_gen_unary (NOT, GET_MODE (x), other, GET_MODE (x));
8047 /* We may be able to continuing distributing the result, so call
8048 ourselves recursively on the inner operation before forming the
8049 outer operation, which we return. */
8050 return simplify_gen_binary (inner_code, GET_MODE (x),
8051 apply_distributive_law (tem), other);
8054 /* See if X is of the form (* (+ A B) C), and if so convert to
8055 (+ (* A C) (* B C)) and try to simplify.
8057 Most of the time, this results in no change. However, if some of
8058 the operands are the same or inverses of each other, simplifications
8059 will result.
8061 For example, (and (ior A B) (not B)) can occur as the result of
8062 expanding a bit field assignment. When we apply the distributive
8063 law to this, we get (ior (and (A (not B))) (and (B (not B)))),
8064 which then simplifies to (and (A (not B))).
8066 Note that no checks happen on the validity of applying the inverse
8067 distributive law. This is pointless since we can do it in the
8068 few places where this routine is called.
8070 N is the index of the term that is decomposed (the arithmetic operation,
8071 i.e. (+ A B) in the first example above). !N is the index of the term that
8072 is distributed, i.e. of C in the first example above. */
8073 static rtx
8074 distribute_and_simplify_rtx (rtx x, int n)
8076 enum machine_mode mode;
8077 enum rtx_code outer_code, inner_code;
8078 rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp;
8080 decomposed = XEXP (x, n);
8081 if (!ARITHMETIC_P (decomposed))
8082 return NULL_RTX;
8084 mode = GET_MODE (x);
8085 outer_code = GET_CODE (x);
8086 distributed = XEXP (x, !n);
8088 inner_code = GET_CODE (decomposed);
8089 inner_op0 = XEXP (decomposed, 0);
8090 inner_op1 = XEXP (decomposed, 1);
8092 /* Special case (and (xor B C) (not A)), which is equivalent to
8093 (xor (ior A B) (ior A C)) */
8094 if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT)
8096 distributed = XEXP (distributed, 0);
8097 outer_code = IOR;
8100 if (n == 0)
8102 /* Distribute the second term. */
8103 new_op0 = simplify_gen_binary (outer_code, mode, inner_op0, distributed);
8104 new_op1 = simplify_gen_binary (outer_code, mode, inner_op1, distributed);
8106 else
8108 /* Distribute the first term. */
8109 new_op0 = simplify_gen_binary (outer_code, mode, distributed, inner_op0);
8110 new_op1 = simplify_gen_binary (outer_code, mode, distributed, inner_op1);
8113 tmp = apply_distributive_law (simplify_gen_binary (inner_code, mode,
8114 new_op0, new_op1));
8115 if (GET_CODE (tmp) != outer_code
8116 && rtx_cost (tmp, SET) < rtx_cost (x, SET))
8117 return tmp;
8119 return NULL_RTX;
8122 /* Simplify a logical `and' of VAROP with the constant CONSTOP, to be done
8123 in MODE. Return an equivalent form, if different from (and VAROP
8124 (const_int CONSTOP)). Otherwise, return NULL_RTX. */
8126 static rtx
8127 simplify_and_const_int_1 (enum machine_mode mode, rtx varop,
8128 unsigned HOST_WIDE_INT constop)
8130 unsigned HOST_WIDE_INT nonzero;
8131 unsigned HOST_WIDE_INT orig_constop;
8132 rtx orig_varop;
8133 int i;
8135 orig_varop = varop;
8136 orig_constop = constop;
8137 if (GET_CODE (varop) == CLOBBER)
8138 return NULL_RTX;
8140 /* Simplify VAROP knowing that we will be only looking at some of the
8141 bits in it.
8143 Note by passing in CONSTOP, we guarantee that the bits not set in
8144 CONSTOP are not significant and will never be examined. We must
8145 ensure that is the case by explicitly masking out those bits
8146 before returning. */
8147 varop = force_to_mode (varop, mode, constop, 0);
8149 /* If VAROP is a CLOBBER, we will fail so return it. */
8150 if (GET_CODE (varop) == CLOBBER)
8151 return varop;
8153 /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP
8154 to VAROP and return the new constant. */
8155 if (GET_CODE (varop) == CONST_INT)
8156 return gen_int_mode (INTVAL (varop) & constop, mode);
8158 /* See what bits may be nonzero in VAROP. Unlike the general case of
8159 a call to nonzero_bits, here we don't care about bits outside
8160 MODE. */
8162 nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
8164 /* Turn off all bits in the constant that are known to already be zero.
8165 Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
8166 which is tested below. */
8168 constop &= nonzero;
8170 /* If we don't have any bits left, return zero. */
8171 if (constop == 0)
8172 return const0_rtx;
8174 /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is
8175 a power of two, we can replace this with an ASHIFT. */
8176 if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1
8177 && (i = exact_log2 (constop)) >= 0)
8178 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i);
8180 /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR
8181 or XOR, then try to apply the distributive law. This may eliminate
8182 operations if either branch can be simplified because of the AND.
8183 It may also make some cases more complex, but those cases probably
8184 won't match a pattern either with or without this. */
8186 if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR)
8187 return
8188 gen_lowpart
8189 (mode,
8190 apply_distributive_law
8191 (simplify_gen_binary (GET_CODE (varop), GET_MODE (varop),
8192 simplify_and_const_int (NULL_RTX,
8193 GET_MODE (varop),
8194 XEXP (varop, 0),
8195 constop),
8196 simplify_and_const_int (NULL_RTX,
8197 GET_MODE (varop),
8198 XEXP (varop, 1),
8199 constop))));
8201 /* If VAROP is PLUS, and the constant is a mask of low bits, distribute
8202 the AND and see if one of the operands simplifies to zero. If so, we
8203 may eliminate it. */
8205 if (GET_CODE (varop) == PLUS
8206 && exact_log2 (constop + 1) >= 0)
8208 rtx o0, o1;
8210 o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop);
8211 o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop);
8212 if (o0 == const0_rtx)
8213 return o1;
8214 if (o1 == const0_rtx)
8215 return o0;
8218 /* Make a SUBREG if necessary. If we can't make it, fail. */
8219 varop = gen_lowpart (mode, varop);
8220 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
8221 return NULL_RTX;
8223 /* If we are only masking insignificant bits, return VAROP. */
8224 if (constop == nonzero)
8225 return varop;
8227 if (varop == orig_varop && constop == orig_constop)
8228 return NULL_RTX;
8230 /* Otherwise, return an AND. */
8231 return simplify_gen_binary (AND, mode, varop, gen_int_mode (constop, mode));
8235 /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done
8236 in MODE.
8238 Return an equivalent form, if different from X. Otherwise, return X. If
8239 X is zero, we are to always construct the equivalent form. */
8241 static rtx
8242 simplify_and_const_int (rtx x, enum machine_mode mode, rtx varop,
8243 unsigned HOST_WIDE_INT constop)
8245 rtx tem = simplify_and_const_int_1 (mode, varop, constop);
8246 if (tem)
8247 return tem;
8249 if (!x)
8250 x = simplify_gen_binary (AND, GET_MODE (varop), varop,
8251 gen_int_mode (constop, mode));
8252 if (GET_MODE (x) != mode)
8253 x = gen_lowpart (mode, x);
8254 return x;
8257 /* Given a REG, X, compute which bits in X can be nonzero.
8258 We don't care about bits outside of those defined in MODE.
8260 For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is
8261 a shift, AND, or zero_extract, we can do better. */
8263 static rtx
8264 reg_nonzero_bits_for_combine (rtx x, enum machine_mode mode,
8265 rtx known_x ATTRIBUTE_UNUSED,
8266 enum machine_mode known_mode ATTRIBUTE_UNUSED,
8267 unsigned HOST_WIDE_INT known_ret ATTRIBUTE_UNUSED,
8268 unsigned HOST_WIDE_INT *nonzero)
8270 rtx tem;
8272 /* If X is a register whose nonzero bits value is current, use it.
8273 Otherwise, if X is a register whose value we can find, use that
8274 value. Otherwise, use the previously-computed global nonzero bits
8275 for this register. */
8277 if (reg_stat[REGNO (x)].last_set_value != 0
8278 && (reg_stat[REGNO (x)].last_set_mode == mode
8279 || (GET_MODE_CLASS (reg_stat[REGNO (x)].last_set_mode) == MODE_INT
8280 && GET_MODE_CLASS (mode) == MODE_INT))
8281 && (reg_stat[REGNO (x)].last_set_label == label_tick
8282 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8283 && REG_N_SETS (REGNO (x)) == 1
8284 && ! REGNO_REG_SET_P
8285 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8286 REGNO (x))))
8287 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8289 *nonzero &= reg_stat[REGNO (x)].last_set_nonzero_bits;
8290 return NULL;
8293 tem = get_last_value (x);
8295 if (tem)
8297 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
8298 /* If X is narrower than MODE and TEM is a non-negative
8299 constant that would appear negative in the mode of X,
8300 sign-extend it for use in reg_nonzero_bits because some
8301 machines (maybe most) will actually do the sign-extension
8302 and this is the conservative approach.
8304 ??? For 2.5, try to tighten up the MD files in this regard
8305 instead of this kludge. */
8307 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode)
8308 && GET_CODE (tem) == CONST_INT
8309 && INTVAL (tem) > 0
8310 && 0 != (INTVAL (tem)
8311 & ((HOST_WIDE_INT) 1
8312 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
8313 tem = GEN_INT (INTVAL (tem)
8314 | ((HOST_WIDE_INT) (-1)
8315 << GET_MODE_BITSIZE (GET_MODE (x))));
8316 #endif
8317 return tem;
8319 else if (nonzero_sign_valid && reg_stat[REGNO (x)].nonzero_bits)
8321 unsigned HOST_WIDE_INT mask = reg_stat[REGNO (x)].nonzero_bits;
8323 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode))
8324 /* We don't know anything about the upper bits. */
8325 mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (GET_MODE (x));
8326 *nonzero &= mask;
8329 return NULL;
8332 /* Return the number of bits at the high-order end of X that are known to
8333 be equal to the sign bit. X will be used in mode MODE; if MODE is
8334 VOIDmode, X will be used in its own mode. The returned value will always
8335 be between 1 and the number of bits in MODE. */
8337 static rtx
8338 reg_num_sign_bit_copies_for_combine (rtx x, enum machine_mode mode,
8339 rtx known_x ATTRIBUTE_UNUSED,
8340 enum machine_mode known_mode
8341 ATTRIBUTE_UNUSED,
8342 unsigned int known_ret ATTRIBUTE_UNUSED,
8343 unsigned int *result)
8345 rtx tem;
8347 if (reg_stat[REGNO (x)].last_set_value != 0
8348 && reg_stat[REGNO (x)].last_set_mode == mode
8349 && (reg_stat[REGNO (x)].last_set_label == label_tick
8350 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8351 && REG_N_SETS (REGNO (x)) == 1
8352 && ! REGNO_REG_SET_P
8353 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8354 REGNO (x))))
8355 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8357 *result = reg_stat[REGNO (x)].last_set_sign_bit_copies;
8358 return NULL;
8361 tem = get_last_value (x);
8362 if (tem != 0)
8363 return tem;
8365 if (nonzero_sign_valid && reg_stat[REGNO (x)].sign_bit_copies != 0
8366 && GET_MODE_BITSIZE (GET_MODE (x)) == GET_MODE_BITSIZE (mode))
8367 *result = reg_stat[REGNO (x)].sign_bit_copies;
8369 return NULL;
8372 /* Return the number of "extended" bits there are in X, when interpreted
8373 as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For
8374 unsigned quantities, this is the number of high-order zero bits.
8375 For signed quantities, this is the number of copies of the sign bit
8376 minus 1. In both case, this function returns the number of "spare"
8377 bits. For example, if two quantities for which this function returns
8378 at least 1 are added, the addition is known not to overflow.
8380 This function will always return 0 unless called during combine, which
8381 implies that it must be called from a define_split. */
8383 unsigned int
8384 extended_count (rtx x, enum machine_mode mode, int unsignedp)
8386 if (nonzero_sign_valid == 0)
8387 return 0;
8389 return (unsignedp
8390 ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8391 ? (unsigned int) (GET_MODE_BITSIZE (mode) - 1
8392 - floor_log2 (nonzero_bits (x, mode)))
8393 : 0)
8394 : num_sign_bit_copies (x, mode) - 1);
8397 /* This function is called from `simplify_shift_const' to merge two
8398 outer operations. Specifically, we have already found that we need
8399 to perform operation *POP0 with constant *PCONST0 at the outermost
8400 position. We would now like to also perform OP1 with constant CONST1
8401 (with *POP0 being done last).
8403 Return 1 if we can do the operation and update *POP0 and *PCONST0 with
8404 the resulting operation. *PCOMP_P is set to 1 if we would need to
8405 complement the innermost operand, otherwise it is unchanged.
8407 MODE is the mode in which the operation will be done. No bits outside
8408 the width of this mode matter. It is assumed that the width of this mode
8409 is smaller than or equal to HOST_BITS_PER_WIDE_INT.
8411 If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS,
8412 IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper
8413 result is simply *PCONST0.
8415 If the resulting operation cannot be expressed as one operation, we
8416 return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */
8418 static int
8419 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)
8421 enum rtx_code op0 = *pop0;
8422 HOST_WIDE_INT const0 = *pconst0;
8424 const0 &= GET_MODE_MASK (mode);
8425 const1 &= GET_MODE_MASK (mode);
8427 /* If OP0 is an AND, clear unimportant bits in CONST1. */
8428 if (op0 == AND)
8429 const1 &= const0;
8431 /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or
8432 if OP0 is SET. */
8434 if (op1 == UNKNOWN || op0 == SET)
8435 return 1;
8437 else if (op0 == UNKNOWN)
8438 op0 = op1, const0 = const1;
8440 else if (op0 == op1)
8442 switch (op0)
8444 case AND:
8445 const0 &= const1;
8446 break;
8447 case IOR:
8448 const0 |= const1;
8449 break;
8450 case XOR:
8451 const0 ^= const1;
8452 break;
8453 case PLUS:
8454 const0 += const1;
8455 break;
8456 case NEG:
8457 op0 = UNKNOWN;
8458 break;
8459 default:
8460 break;
8464 /* Otherwise, if either is a PLUS or NEG, we can't do anything. */
8465 else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG)
8466 return 0;
8468 /* If the two constants aren't the same, we can't do anything. The
8469 remaining six cases can all be done. */
8470 else if (const0 != const1)
8471 return 0;
8473 else
8474 switch (op0)
8476 case IOR:
8477 if (op1 == AND)
8478 /* (a & b) | b == b */
8479 op0 = SET;
8480 else /* op1 == XOR */
8481 /* (a ^ b) | b == a | b */
8483 break;
8485 case XOR:
8486 if (op1 == AND)
8487 /* (a & b) ^ b == (~a) & b */
8488 op0 = AND, *pcomp_p = 1;
8489 else /* op1 == IOR */
8490 /* (a | b) ^ b == a & ~b */
8491 op0 = AND, const0 = ~const0;
8492 break;
8494 case AND:
8495 if (op1 == IOR)
8496 /* (a | b) & b == b */
8497 op0 = SET;
8498 else /* op1 == XOR */
8499 /* (a ^ b) & b) == (~a) & b */
8500 *pcomp_p = 1;
8501 break;
8502 default:
8503 break;
8506 /* Check for NO-OP cases. */
8507 const0 &= GET_MODE_MASK (mode);
8508 if (const0 == 0
8509 && (op0 == IOR || op0 == XOR || op0 == PLUS))
8510 op0 = UNKNOWN;
8511 else if (const0 == 0 && op0 == AND)
8512 op0 = SET;
8513 else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode)
8514 && op0 == AND)
8515 op0 = UNKNOWN;
8517 /* ??? Slightly redundant with the above mask, but not entirely.
8518 Moving this above means we'd have to sign-extend the mode mask
8519 for the final test. */
8520 const0 = trunc_int_for_mode (const0, mode);
8522 *pop0 = op0;
8523 *pconst0 = const0;
8525 return 1;
8528 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
8529 The result of the shift is RESULT_MODE. Return NULL_RTX if we cannot
8530 simplify it. Otherwise, return a simplified value.
8532 The shift is normally computed in the widest mode we find in VAROP, as
8533 long as it isn't a different number of words than RESULT_MODE. Exceptions
8534 are ASHIFTRT and ROTATE, which are always done in their original mode. */
8536 static rtx
8537 simplify_shift_const_1 (enum rtx_code code, enum machine_mode result_mode,
8538 rtx varop, int orig_count)
8540 enum rtx_code orig_code = code;
8541 rtx orig_varop = varop;
8542 int count;
8543 enum machine_mode mode = result_mode;
8544 enum machine_mode shift_mode, tmode;
8545 unsigned int mode_words
8546 = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
8547 /* We form (outer_op (code varop count) (outer_const)). */
8548 enum rtx_code outer_op = UNKNOWN;
8549 HOST_WIDE_INT outer_const = 0;
8550 int complement_p = 0;
8551 rtx new, x;
8553 /* Make sure and truncate the "natural" shift on the way in. We don't
8554 want to do this inside the loop as it makes it more difficult to
8555 combine shifts. */
8556 if (SHIFT_COUNT_TRUNCATED)
8557 orig_count &= GET_MODE_BITSIZE (mode) - 1;
8559 /* If we were given an invalid count, don't do anything except exactly
8560 what was requested. */
8562 if (orig_count < 0 || orig_count >= (int) GET_MODE_BITSIZE (mode))
8563 return NULL_RTX;
8565 count = orig_count;
8567 /* Unless one of the branches of the `if' in this loop does a `continue',
8568 we will `break' the loop after the `if'. */
8570 while (count != 0)
8572 /* If we have an operand of (clobber (const_int 0)), fail. */
8573 if (GET_CODE (varop) == CLOBBER)
8574 return NULL_RTX;
8576 /* If we discovered we had to complement VAROP, leave. Making a NOT
8577 here would cause an infinite loop. */
8578 if (complement_p)
8579 break;
8581 /* Convert ROTATERT to ROTATE. */
8582 if (code == ROTATERT)
8584 unsigned int bitsize = GET_MODE_BITSIZE (result_mode);;
8585 code = ROTATE;
8586 if (VECTOR_MODE_P (result_mode))
8587 count = bitsize / GET_MODE_NUNITS (result_mode) - count;
8588 else
8589 count = bitsize - count;
8592 /* We need to determine what mode we will do the shift in. If the
8593 shift is a right shift or a ROTATE, we must always do it in the mode
8594 it was originally done in. Otherwise, we can do it in MODE, the
8595 widest mode encountered. */
8596 shift_mode
8597 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
8598 ? result_mode : mode);
8600 /* Handle cases where the count is greater than the size of the mode
8601 minus 1. For ASHIFT, use the size minus one as the count (this can
8602 occur when simplifying (lshiftrt (ashiftrt ..))). For rotates,
8603 take the count modulo the size. For other shifts, the result is
8604 zero.
8606 Since these shifts are being produced by the compiler by combining
8607 multiple operations, each of which are defined, we know what the
8608 result is supposed to be. */
8610 if (count > (GET_MODE_BITSIZE (shift_mode) - 1))
8612 if (code == ASHIFTRT)
8613 count = GET_MODE_BITSIZE (shift_mode) - 1;
8614 else if (code == ROTATE || code == ROTATERT)
8615 count %= GET_MODE_BITSIZE (shift_mode);
8616 else
8618 /* We can't simply return zero because there may be an
8619 outer op. */
8620 varop = const0_rtx;
8621 count = 0;
8622 break;
8626 /* An arithmetic right shift of a quantity known to be -1 or 0
8627 is a no-op. */
8628 if (code == ASHIFTRT
8629 && (num_sign_bit_copies (varop, shift_mode)
8630 == GET_MODE_BITSIZE (shift_mode)))
8632 count = 0;
8633 break;
8636 /* If we are doing an arithmetic right shift and discarding all but
8637 the sign bit copies, this is equivalent to doing a shift by the
8638 bitsize minus one. Convert it into that shift because it will often
8639 allow other simplifications. */
8641 if (code == ASHIFTRT
8642 && (count + num_sign_bit_copies (varop, shift_mode)
8643 >= GET_MODE_BITSIZE (shift_mode)))
8644 count = GET_MODE_BITSIZE (shift_mode) - 1;
8646 /* We simplify the tests below and elsewhere by converting
8647 ASHIFTRT to LSHIFTRT if we know the sign bit is clear.
8648 `make_compound_operation' will convert it to an ASHIFTRT for
8649 those machines (such as VAX) that don't have an LSHIFTRT. */
8650 if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8651 && code == ASHIFTRT
8652 && ((nonzero_bits (varop, shift_mode)
8653 & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1)))
8654 == 0))
8655 code = LSHIFTRT;
8657 if (((code == LSHIFTRT
8658 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8659 && !(nonzero_bits (varop, shift_mode) >> count))
8660 || (code == ASHIFT
8661 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8662 && !((nonzero_bits (varop, shift_mode) << count)
8663 & GET_MODE_MASK (shift_mode))))
8664 && !side_effects_p (varop))
8665 varop = const0_rtx;
8667 switch (GET_CODE (varop))
8669 case SIGN_EXTEND:
8670 case ZERO_EXTEND:
8671 case SIGN_EXTRACT:
8672 case ZERO_EXTRACT:
8673 new = expand_compound_operation (varop);
8674 if (new != varop)
8676 varop = new;
8677 continue;
8679 break;
8681 case MEM:
8682 /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH
8683 minus the width of a smaller mode, we can do this with a
8684 SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */
8685 if ((code == ASHIFTRT || code == LSHIFTRT)
8686 && ! mode_dependent_address_p (XEXP (varop, 0))
8687 && ! MEM_VOLATILE_P (varop)
8688 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
8689 MODE_INT, 1)) != BLKmode)
8691 new = adjust_address_nv (varop, tmode,
8692 BYTES_BIG_ENDIAN ? 0
8693 : count / BITS_PER_UNIT);
8695 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
8696 : ZERO_EXTEND, mode, new);
8697 count = 0;
8698 continue;
8700 break;
8702 case SUBREG:
8703 /* If VAROP is a SUBREG, strip it as long as the inner operand has
8704 the same number of words as what we've seen so far. Then store
8705 the widest mode in MODE. */
8706 if (subreg_lowpart_p (varop)
8707 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8708 > GET_MODE_SIZE (GET_MODE (varop)))
8709 && (unsigned int) ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8710 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
8711 == mode_words)
8713 varop = SUBREG_REG (varop);
8714 if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode))
8715 mode = GET_MODE (varop);
8716 continue;
8718 break;
8720 case MULT:
8721 /* Some machines use MULT instead of ASHIFT because MULT
8722 is cheaper. But it is still better on those machines to
8723 merge two shifts into one. */
8724 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8725 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8727 varop
8728 = simplify_gen_binary (ASHIFT, GET_MODE (varop),
8729 XEXP (varop, 0),
8730 GEN_INT (exact_log2 (
8731 INTVAL (XEXP (varop, 1)))));
8732 continue;
8734 break;
8736 case UDIV:
8737 /* Similar, for when divides are cheaper. */
8738 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8739 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8741 varop
8742 = simplify_gen_binary (LSHIFTRT, GET_MODE (varop),
8743 XEXP (varop, 0),
8744 GEN_INT (exact_log2 (
8745 INTVAL (XEXP (varop, 1)))));
8746 continue;
8748 break;
8750 case ASHIFTRT:
8751 /* If we are extracting just the sign bit of an arithmetic
8752 right shift, that shift is not needed. However, the sign
8753 bit of a wider mode may be different from what would be
8754 interpreted as the sign bit in a narrower mode, so, if
8755 the result is narrower, don't discard the shift. */
8756 if (code == LSHIFTRT
8757 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8758 && (GET_MODE_BITSIZE (result_mode)
8759 >= GET_MODE_BITSIZE (GET_MODE (varop))))
8761 varop = XEXP (varop, 0);
8762 continue;
8765 /* ... fall through ... */
8767 case LSHIFTRT:
8768 case ASHIFT:
8769 case ROTATE:
8770 /* Here we have two nested shifts. The result is usually the
8771 AND of a new shift with a mask. We compute the result below. */
8772 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8773 && INTVAL (XEXP (varop, 1)) >= 0
8774 && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop))
8775 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8776 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8777 && !VECTOR_MODE_P (result_mode))
8779 enum rtx_code first_code = GET_CODE (varop);
8780 unsigned int first_count = INTVAL (XEXP (varop, 1));
8781 unsigned HOST_WIDE_INT mask;
8782 rtx mask_rtx;
8784 /* We have one common special case. We can't do any merging if
8785 the inner code is an ASHIFTRT of a smaller mode. However, if
8786 we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2)
8787 with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2),
8788 we can convert it to
8789 (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1).
8790 This simplifies certain SIGN_EXTEND operations. */
8791 if (code == ASHIFT && first_code == ASHIFTRT
8792 && count == (GET_MODE_BITSIZE (result_mode)
8793 - GET_MODE_BITSIZE (GET_MODE (varop))))
8795 /* C3 has the low-order C1 bits zero. */
8797 mask = (GET_MODE_MASK (mode)
8798 & ~(((HOST_WIDE_INT) 1 << first_count) - 1));
8800 varop = simplify_and_const_int (NULL_RTX, result_mode,
8801 XEXP (varop, 0), mask);
8802 varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode,
8803 varop, count);
8804 count = first_count;
8805 code = ASHIFTRT;
8806 continue;
8809 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more
8810 than C1 high-order bits equal to the sign bit, we can convert
8811 this to either an ASHIFT or an ASHIFTRT depending on the
8812 two counts.
8814 We cannot do this if VAROP's mode is not SHIFT_MODE. */
8816 if (code == ASHIFTRT && first_code == ASHIFT
8817 && GET_MODE (varop) == shift_mode
8818 && (num_sign_bit_copies (XEXP (varop, 0), shift_mode)
8819 > first_count))
8821 varop = XEXP (varop, 0);
8822 count -= first_count;
8823 if (count < 0)
8825 count = -count;
8826 code = ASHIFT;
8829 continue;
8832 /* There are some cases we can't do. If CODE is ASHIFTRT,
8833 we can only do this if FIRST_CODE is also ASHIFTRT.
8835 We can't do the case when CODE is ROTATE and FIRST_CODE is
8836 ASHIFTRT.
8838 If the mode of this shift is not the mode of the outer shift,
8839 we can't do this if either shift is a right shift or ROTATE.
8841 Finally, we can't do any of these if the mode is too wide
8842 unless the codes are the same.
8844 Handle the case where the shift codes are the same
8845 first. */
8847 if (code == first_code)
8849 if (GET_MODE (varop) != result_mode
8850 && (code == ASHIFTRT || code == LSHIFTRT
8851 || code == ROTATE))
8852 break;
8854 count += first_count;
8855 varop = XEXP (varop, 0);
8856 continue;
8859 if (code == ASHIFTRT
8860 || (code == ROTATE && first_code == ASHIFTRT)
8861 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT
8862 || (GET_MODE (varop) != result_mode
8863 && (first_code == ASHIFTRT || first_code == LSHIFTRT
8864 || first_code == ROTATE
8865 || code == ROTATE)))
8866 break;
8868 /* To compute the mask to apply after the shift, shift the
8869 nonzero bits of the inner shift the same way the
8870 outer shift will. */
8872 mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop)));
8874 mask_rtx
8875 = simplify_const_binary_operation (code, result_mode, mask_rtx,
8876 GEN_INT (count));
8878 /* Give up if we can't compute an outer operation to use. */
8879 if (mask_rtx == 0
8880 || GET_CODE (mask_rtx) != CONST_INT
8881 || ! merge_outer_ops (&outer_op, &outer_const, AND,
8882 INTVAL (mask_rtx),
8883 result_mode, &complement_p))
8884 break;
8886 /* If the shifts are in the same direction, we add the
8887 counts. Otherwise, we subtract them. */
8888 if ((code == ASHIFTRT || code == LSHIFTRT)
8889 == (first_code == ASHIFTRT || first_code == LSHIFTRT))
8890 count += first_count;
8891 else
8892 count -= first_count;
8894 /* If COUNT is positive, the new shift is usually CODE,
8895 except for the two exceptions below, in which case it is
8896 FIRST_CODE. If the count is negative, FIRST_CODE should
8897 always be used */
8898 if (count > 0
8899 && ((first_code == ROTATE && code == ASHIFT)
8900 || (first_code == ASHIFTRT && code == LSHIFTRT)))
8901 code = first_code;
8902 else if (count < 0)
8903 code = first_code, count = -count;
8905 varop = XEXP (varop, 0);
8906 continue;
8909 /* If we have (A << B << C) for any shift, we can convert this to
8910 (A << C << B). This wins if A is a constant. Only try this if
8911 B is not a constant. */
8913 else if (GET_CODE (varop) == code
8914 && GET_CODE (XEXP (varop, 0)) == CONST_INT
8915 && GET_CODE (XEXP (varop, 1)) != CONST_INT)
8917 rtx new = simplify_const_binary_operation (code, mode,
8918 XEXP (varop, 0),
8919 GEN_INT (count));
8920 varop = gen_rtx_fmt_ee (code, mode, new, XEXP (varop, 1));
8921 count = 0;
8922 continue;
8924 break;
8926 case NOT:
8927 /* Make this fit the case below. */
8928 varop = gen_rtx_XOR (mode, XEXP (varop, 0),
8929 GEN_INT (GET_MODE_MASK (mode)));
8930 continue;
8932 case IOR:
8933 case AND:
8934 case XOR:
8935 /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C)
8936 with C the size of VAROP - 1 and the shift is logical if
8937 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
8938 we have an (le X 0) operation. If we have an arithmetic shift
8939 and STORE_FLAG_VALUE is 1 or we have a logical shift with
8940 STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */
8942 if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS
8943 && XEXP (XEXP (varop, 0), 1) == constm1_rtx
8944 && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
8945 && (code == LSHIFTRT || code == ASHIFTRT)
8946 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
8947 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
8949 count = 0;
8950 varop = gen_rtx_LE (GET_MODE (varop), XEXP (varop, 1),
8951 const0_rtx);
8953 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
8954 varop = gen_rtx_NEG (GET_MODE (varop), varop);
8956 continue;
8959 /* If we have (shift (logical)), move the logical to the outside
8960 to allow it to possibly combine with another logical and the
8961 shift to combine with another shift. This also canonicalizes to
8962 what a ZERO_EXTRACT looks like. Also, some machines have
8963 (and (shift)) insns. */
8965 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8966 /* We can't do this if we have (ashiftrt (xor)) and the
8967 constant has its sign bit set in shift_mode. */
8968 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8969 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8970 shift_mode))
8971 && (new = simplify_const_binary_operation (code, result_mode,
8972 XEXP (varop, 1),
8973 GEN_INT (count))) != 0
8974 && GET_CODE (new) == CONST_INT
8975 && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop),
8976 INTVAL (new), result_mode, &complement_p))
8978 varop = XEXP (varop, 0);
8979 continue;
8982 /* If we can't do that, try to simplify the shift in each arm of the
8983 logical expression, make a new logical expression, and apply
8984 the inverse distributive law. This also can't be done
8985 for some (ashiftrt (xor)). */
8986 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8987 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8988 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8989 shift_mode)))
8991 rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8992 XEXP (varop, 0), count);
8993 rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8994 XEXP (varop, 1), count);
8996 varop = simplify_gen_binary (GET_CODE (varop), shift_mode,
8997 lhs, rhs);
8998 varop = apply_distributive_law (varop);
9000 count = 0;
9001 continue;
9003 break;
9005 case EQ:
9006 /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE
9007 says that the sign bit can be tested, FOO has mode MODE, C is
9008 GET_MODE_BITSIZE (MODE) - 1, and FOO has only its low-order bit
9009 that may be nonzero. */
9010 if (code == LSHIFTRT
9011 && XEXP (varop, 1) == const0_rtx
9012 && GET_MODE (XEXP (varop, 0)) == result_mode
9013 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9014 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9015 && STORE_FLAG_VALUE == -1
9016 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9017 && merge_outer_ops (&outer_op, &outer_const, XOR,
9018 (HOST_WIDE_INT) 1, result_mode,
9019 &complement_p))
9021 varop = XEXP (varop, 0);
9022 count = 0;
9023 continue;
9025 break;
9027 case NEG:
9028 /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less
9029 than the number of bits in the mode is equivalent to A. */
9030 if (code == LSHIFTRT
9031 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9032 && nonzero_bits (XEXP (varop, 0), result_mode) == 1)
9034 varop = XEXP (varop, 0);
9035 count = 0;
9036 continue;
9039 /* NEG commutes with ASHIFT since it is multiplication. Move the
9040 NEG outside to allow shifts to combine. */
9041 if (code == ASHIFT
9042 && merge_outer_ops (&outer_op, &outer_const, NEG,
9043 (HOST_WIDE_INT) 0, result_mode,
9044 &complement_p))
9046 varop = XEXP (varop, 0);
9047 continue;
9049 break;
9051 case PLUS:
9052 /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C
9053 is one less than the number of bits in the mode is
9054 equivalent to (xor A 1). */
9055 if (code == LSHIFTRT
9056 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9057 && XEXP (varop, 1) == constm1_rtx
9058 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9059 && merge_outer_ops (&outer_op, &outer_const, XOR,
9060 (HOST_WIDE_INT) 1, result_mode,
9061 &complement_p))
9063 count = 0;
9064 varop = XEXP (varop, 0);
9065 continue;
9068 /* If we have (xshiftrt (plus FOO BAR) C), and the only bits
9069 that might be nonzero in BAR are those being shifted out and those
9070 bits are known zero in FOO, we can replace the PLUS with FOO.
9071 Similarly in the other operand order. This code occurs when
9072 we are computing the size of a variable-size array. */
9074 if ((code == ASHIFTRT || code == LSHIFTRT)
9075 && count < HOST_BITS_PER_WIDE_INT
9076 && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0
9077 && (nonzero_bits (XEXP (varop, 1), result_mode)
9078 & nonzero_bits (XEXP (varop, 0), result_mode)) == 0)
9080 varop = XEXP (varop, 0);
9081 continue;
9083 else if ((code == ASHIFTRT || code == LSHIFTRT)
9084 && count < HOST_BITS_PER_WIDE_INT
9085 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9086 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9087 >> count)
9088 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9089 & nonzero_bits (XEXP (varop, 1),
9090 result_mode)))
9092 varop = XEXP (varop, 1);
9093 continue;
9096 /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */
9097 if (code == ASHIFT
9098 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9099 && (new = simplify_const_binary_operation (ASHIFT, result_mode,
9100 XEXP (varop, 1),
9101 GEN_INT (count))) != 0
9102 && GET_CODE (new) == CONST_INT
9103 && merge_outer_ops (&outer_op, &outer_const, PLUS,
9104 INTVAL (new), result_mode, &complement_p))
9106 varop = XEXP (varop, 0);
9107 continue;
9110 /* Check for 'PLUS signbit', which is the canonical form of 'XOR
9111 signbit', and attempt to change the PLUS to an XOR and move it to
9112 the outer operation as is done above in the AND/IOR/XOR case
9113 leg for shift(logical). See details in logical handling above
9114 for reasoning in doing so. */
9115 if (code == LSHIFTRT
9116 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9117 && mode_signbit_p (result_mode, XEXP (varop, 1))
9118 && (new = simplify_const_binary_operation (code, result_mode,
9119 XEXP (varop, 1),
9120 GEN_INT (count))) != 0
9121 && GET_CODE (new) == CONST_INT
9122 && merge_outer_ops (&outer_op, &outer_const, XOR,
9123 INTVAL (new), result_mode, &complement_p))
9125 varop = XEXP (varop, 0);
9126 continue;
9129 break;
9131 case MINUS:
9132 /* If we have (xshiftrt (minus (ashiftrt X C)) X) C)
9133 with C the size of VAROP - 1 and the shift is logical if
9134 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
9135 we have a (gt X 0) operation. If the shift is arithmetic with
9136 STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1,
9137 we have a (neg (gt X 0)) operation. */
9139 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
9140 && GET_CODE (XEXP (varop, 0)) == ASHIFTRT
9141 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
9142 && (code == LSHIFTRT || code == ASHIFTRT)
9143 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9144 && INTVAL (XEXP (XEXP (varop, 0), 1)) == count
9145 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
9147 count = 0;
9148 varop = gen_rtx_GT (GET_MODE (varop), XEXP (varop, 1),
9149 const0_rtx);
9151 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
9152 varop = gen_rtx_NEG (GET_MODE (varop), varop);
9154 continue;
9156 break;
9158 case TRUNCATE:
9159 /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt))
9160 if the truncate does not affect the value. */
9161 if (code == LSHIFTRT
9162 && GET_CODE (XEXP (varop, 0)) == LSHIFTRT
9163 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9164 && (INTVAL (XEXP (XEXP (varop, 0), 1))
9165 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (varop, 0)))
9166 - GET_MODE_BITSIZE (GET_MODE (varop)))))
9168 rtx varop_inner = XEXP (varop, 0);
9170 varop_inner
9171 = gen_rtx_LSHIFTRT (GET_MODE (varop_inner),
9172 XEXP (varop_inner, 0),
9173 GEN_INT
9174 (count + INTVAL (XEXP (varop_inner, 1))));
9175 varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner);
9176 count = 0;
9177 continue;
9179 break;
9181 default:
9182 break;
9185 break;
9188 /* We need to determine what mode to do the shift in. If the shift is
9189 a right shift or ROTATE, we must always do it in the mode it was
9190 originally done in. Otherwise, we can do it in MODE, the widest mode
9191 encountered. The code we care about is that of the shift that will
9192 actually be done, not the shift that was originally requested. */
9193 shift_mode
9194 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
9195 ? result_mode : mode);
9197 /* We have now finished analyzing the shift. The result should be
9198 a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If
9199 OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied
9200 to the result of the shift. OUTER_CONST is the relevant constant,
9201 but we must turn off all bits turned off in the shift. */
9203 if (outer_op == UNKNOWN
9204 && orig_code == code && orig_count == count
9205 && varop == orig_varop
9206 && shift_mode == GET_MODE (varop))
9207 return NULL_RTX;
9209 /* Make a SUBREG if necessary. If we can't make it, fail. */
9210 varop = gen_lowpart (shift_mode, varop);
9211 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
9212 return NULL_RTX;
9214 /* If we have an outer operation and we just made a shift, it is
9215 possible that we could have simplified the shift were it not
9216 for the outer operation. So try to do the simplification
9217 recursively. */
9219 if (outer_op != UNKNOWN)
9220 x = simplify_shift_const_1 (code, shift_mode, varop, count);
9221 else
9222 x = NULL_RTX;
9224 if (x == NULL_RTX)
9225 x = simplify_gen_binary (code, shift_mode, varop, GEN_INT (count));
9227 /* If we were doing an LSHIFTRT in a wider mode than it was originally,
9228 turn off all the bits that the shift would have turned off. */
9229 if (orig_code == LSHIFTRT && result_mode != shift_mode)
9230 x = simplify_and_const_int (NULL_RTX, shift_mode, x,
9231 GET_MODE_MASK (result_mode) >> orig_count);
9233 /* Do the remainder of the processing in RESULT_MODE. */
9234 x = gen_lowpart_or_truncate (result_mode, x);
9236 /* If COMPLEMENT_P is set, we have to complement X before doing the outer
9237 operation. */
9238 if (complement_p)
9239 x = simplify_gen_unary (NOT, result_mode, x, result_mode);
9241 if (outer_op != UNKNOWN)
9243 if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT)
9244 outer_const = trunc_int_for_mode (outer_const, result_mode);
9246 if (outer_op == AND)
9247 x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const);
9248 else if (outer_op == SET)
9250 /* This means that we have determined that the result is
9251 equivalent to a constant. This should be rare. */
9252 if (!side_effects_p (x))
9253 x = GEN_INT (outer_const);
9255 else if (GET_RTX_CLASS (outer_op) == RTX_UNARY)
9256 x = simplify_gen_unary (outer_op, result_mode, x, result_mode);
9257 else
9258 x = simplify_gen_binary (outer_op, result_mode, x,
9259 GEN_INT (outer_const));
9262 return x;
9265 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
9266 The result of the shift is RESULT_MODE. If we cannot simplify it,
9267 return X or, if it is NULL, synthesize the expression with
9268 simplify_gen_binary. Otherwise, return a simplified value.
9270 The shift is normally computed in the widest mode we find in VAROP, as
9271 long as it isn't a different number of words than RESULT_MODE. Exceptions
9272 are ASHIFTRT and ROTATE, which are always done in their original mode. */
9274 static rtx
9275 simplify_shift_const (rtx x, enum rtx_code code, enum machine_mode result_mode,
9276 rtx varop, int count)
9278 rtx tem = simplify_shift_const_1 (code, result_mode, varop, count);
9279 if (tem)
9280 return tem;
9282 if (!x)
9283 x = simplify_gen_binary (code, GET_MODE (varop), varop, GEN_INT (count));
9284 if (GET_MODE (x) != result_mode)
9285 x = gen_lowpart (result_mode, x);
9286 return x;
9290 /* Like recog, but we receive the address of a pointer to a new pattern.
9291 We try to match the rtx that the pointer points to.
9292 If that fails, we may try to modify or replace the pattern,
9293 storing the replacement into the same pointer object.
9295 Modifications include deletion or addition of CLOBBERs.
9297 PNOTES is a pointer to a location where any REG_UNUSED notes added for
9298 the CLOBBERs are placed.
9300 The value is the final insn code from the pattern ultimately matched,
9301 or -1. */
9303 static int
9304 recog_for_combine (rtx *pnewpat, rtx insn, rtx *pnotes)
9306 rtx pat = *pnewpat;
9307 int insn_code_number;
9308 int num_clobbers_to_add = 0;
9309 int i;
9310 rtx notes = 0;
9311 rtx old_notes, old_pat;
9313 /* If PAT is a PARALLEL, check to see if it contains the CLOBBER
9314 we use to indicate that something didn't match. If we find such a
9315 thing, force rejection. */
9316 if (GET_CODE (pat) == PARALLEL)
9317 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
9318 if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER
9319 && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx)
9320 return -1;
9322 old_pat = PATTERN (insn);
9323 old_notes = REG_NOTES (insn);
9324 PATTERN (insn) = pat;
9325 REG_NOTES (insn) = 0;
9327 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9329 /* If it isn't, there is the possibility that we previously had an insn
9330 that clobbered some register as a side effect, but the combined
9331 insn doesn't need to do that. So try once more without the clobbers
9332 unless this represents an ASM insn. */
9334 if (insn_code_number < 0 && ! check_asm_operands (pat)
9335 && GET_CODE (pat) == PARALLEL)
9337 int pos;
9339 for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++)
9340 if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER)
9342 if (i != pos)
9343 SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i));
9344 pos++;
9347 SUBST_INT (XVECLEN (pat, 0), pos);
9349 if (pos == 1)
9350 pat = XVECEXP (pat, 0, 0);
9352 PATTERN (insn) = pat;
9353 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9355 PATTERN (insn) = old_pat;
9356 REG_NOTES (insn) = old_notes;
9358 /* Recognize all noop sets, these will be killed by followup pass. */
9359 if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat))
9360 insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0;
9362 /* If we had any clobbers to add, make a new pattern than contains
9363 them. Then check to make sure that all of them are dead. */
9364 if (num_clobbers_to_add)
9366 rtx newpat = gen_rtx_PARALLEL (VOIDmode,
9367 rtvec_alloc (GET_CODE (pat) == PARALLEL
9368 ? (XVECLEN (pat, 0)
9369 + num_clobbers_to_add)
9370 : num_clobbers_to_add + 1));
9372 if (GET_CODE (pat) == PARALLEL)
9373 for (i = 0; i < XVECLEN (pat, 0); i++)
9374 XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i);
9375 else
9376 XVECEXP (newpat, 0, 0) = pat;
9378 add_clobbers (newpat, insn_code_number);
9380 for (i = XVECLEN (newpat, 0) - num_clobbers_to_add;
9381 i < XVECLEN (newpat, 0); i++)
9383 if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))
9384 && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn))
9385 return -1;
9386 notes = gen_rtx_EXPR_LIST (REG_UNUSED,
9387 XEXP (XVECEXP (newpat, 0, i), 0), notes);
9389 pat = newpat;
9392 *pnewpat = pat;
9393 *pnotes = notes;
9395 return insn_code_number;
9398 /* Like gen_lowpart_general but for use by combine. In combine it
9399 is not possible to create any new pseudoregs. However, it is
9400 safe to create invalid memory addresses, because combine will
9401 try to recognize them and all they will do is make the combine
9402 attempt fail.
9404 If for some reason this cannot do its job, an rtx
9405 (clobber (const_int 0)) is returned.
9406 An insn containing that will not be recognized. */
9408 static rtx
9409 gen_lowpart_for_combine (enum machine_mode omode, rtx x)
9411 enum machine_mode imode = GET_MODE (x);
9412 unsigned int osize = GET_MODE_SIZE (omode);
9413 unsigned int isize = GET_MODE_SIZE (imode);
9414 rtx result;
9416 if (omode == imode)
9417 return x;
9419 /* Return identity if this is a CONST or symbolic reference. */
9420 if (omode == Pmode
9421 && (GET_CODE (x) == CONST
9422 || GET_CODE (x) == SYMBOL_REF
9423 || GET_CODE (x) == LABEL_REF))
9424 return x;
9426 /* We can only support MODE being wider than a word if X is a
9427 constant integer or has a mode the same size. */
9428 if (GET_MODE_SIZE (omode) > UNITS_PER_WORD
9429 && ! ((imode == VOIDmode
9430 && (GET_CODE (x) == CONST_INT
9431 || GET_CODE (x) == CONST_DOUBLE))
9432 || isize == osize))
9433 goto fail;
9435 /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart
9436 won't know what to do. So we will strip off the SUBREG here and
9437 process normally. */
9438 if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x)))
9440 x = SUBREG_REG (x);
9442 /* For use in case we fall down into the address adjustments
9443 further below, we need to adjust the known mode and size of
9444 x; imode and isize, since we just adjusted x. */
9445 imode = GET_MODE (x);
9447 if (imode == omode)
9448 return x;
9450 isize = GET_MODE_SIZE (imode);
9453 result = gen_lowpart_common (omode, x);
9455 #ifdef CANNOT_CHANGE_MODE_CLASS
9456 if (result != 0 && GET_CODE (result) == SUBREG)
9457 record_subregs_of_mode (result);
9458 #endif
9460 if (result)
9461 return result;
9463 if (MEM_P (x))
9465 int offset = 0;
9467 /* Refuse to work on a volatile memory ref or one with a mode-dependent
9468 address. */
9469 if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0)))
9470 goto fail;
9472 /* If we want to refer to something bigger than the original memref,
9473 generate a paradoxical subreg instead. That will force a reload
9474 of the original memref X. */
9475 if (isize < osize)
9476 return gen_rtx_SUBREG (omode, x, 0);
9478 if (WORDS_BIG_ENDIAN)
9479 offset = MAX (isize, UNITS_PER_WORD) - MAX (osize, UNITS_PER_WORD);
9481 /* Adjust the address so that the address-after-the-data is
9482 unchanged. */
9483 if (BYTES_BIG_ENDIAN)
9484 offset -= MIN (UNITS_PER_WORD, osize) - MIN (UNITS_PER_WORD, isize);
9486 return adjust_address_nv (x, omode, offset);
9489 /* If X is a comparison operator, rewrite it in a new mode. This
9490 probably won't match, but may allow further simplifications. */
9491 else if (COMPARISON_P (x))
9492 return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1));
9494 /* If we couldn't simplify X any other way, just enclose it in a
9495 SUBREG. Normally, this SUBREG won't match, but some patterns may
9496 include an explicit SUBREG or we may simplify it further in combine. */
9497 else
9499 int offset = 0;
9500 rtx res;
9502 offset = subreg_lowpart_offset (omode, imode);
9503 if (imode == VOIDmode)
9505 imode = int_mode_for_mode (omode);
9506 x = gen_lowpart_common (imode, x);
9507 if (x == NULL)
9508 goto fail;
9510 res = simplify_gen_subreg (omode, x, imode, offset);
9511 if (res)
9512 return res;
9515 fail:
9516 return gen_rtx_CLOBBER (imode, const0_rtx);
9519 /* Simplify a comparison between *POP0 and *POP1 where CODE is the
9520 comparison code that will be tested.
9522 The result is a possibly different comparison code to use. *POP0 and
9523 *POP1 may be updated.
9525 It is possible that we might detect that a comparison is either always
9526 true or always false. However, we do not perform general constant
9527 folding in combine, so this knowledge isn't useful. Such tautologies
9528 should have been detected earlier. Hence we ignore all such cases. */
9530 static enum rtx_code
9531 simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1)
9533 rtx op0 = *pop0;
9534 rtx op1 = *pop1;
9535 rtx tem, tem1;
9536 int i;
9537 enum machine_mode mode, tmode;
9539 /* Try a few ways of applying the same transformation to both operands. */
9540 while (1)
9542 #ifndef WORD_REGISTER_OPERATIONS
9543 /* The test below this one won't handle SIGN_EXTENDs on these machines,
9544 so check specially. */
9545 if (code != GTU && code != GEU && code != LTU && code != LEU
9546 && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT
9547 && GET_CODE (XEXP (op0, 0)) == ASHIFT
9548 && GET_CODE (XEXP (op1, 0)) == ASHIFT
9549 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG
9550 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG
9551 && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))
9552 == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))))
9553 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9554 && XEXP (op0, 1) == XEXP (op1, 1)
9555 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
9556 && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1)
9557 && (INTVAL (XEXP (op0, 1))
9558 == (GET_MODE_BITSIZE (GET_MODE (op0))
9559 - (GET_MODE_BITSIZE
9560 (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))))))))
9562 op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0));
9563 op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0));
9565 #endif
9567 /* If both operands are the same constant shift, see if we can ignore the
9568 shift. We can if the shift is a rotate or if the bits shifted out of
9569 this shift are known to be zero for both inputs and if the type of
9570 comparison is compatible with the shift. */
9571 if (GET_CODE (op0) == GET_CODE (op1)
9572 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
9573 && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ))
9574 || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT)
9575 && (code != GT && code != LT && code != GE && code != LE))
9576 || (GET_CODE (op0) == ASHIFTRT
9577 && (code != GTU && code != LTU
9578 && code != GEU && code != LEU)))
9579 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9580 && INTVAL (XEXP (op0, 1)) >= 0
9581 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
9582 && XEXP (op0, 1) == XEXP (op1, 1))
9584 enum machine_mode mode = GET_MODE (op0);
9585 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9586 int shift_count = INTVAL (XEXP (op0, 1));
9588 if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT)
9589 mask &= (mask >> shift_count) << shift_count;
9590 else if (GET_CODE (op0) == ASHIFT)
9591 mask = (mask & (mask << shift_count)) >> shift_count;
9593 if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0
9594 && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0)
9595 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0);
9596 else
9597 break;
9600 /* If both operands are AND's of a paradoxical SUBREG by constant, the
9601 SUBREGs are of the same mode, and, in both cases, the AND would
9602 be redundant if the comparison was done in the narrower mode,
9603 do the comparison in the narrower mode (e.g., we are AND'ing with 1
9604 and the operand's possibly nonzero bits are 0xffffff01; in that case
9605 if we only care about QImode, we don't need the AND). This case
9606 occurs if the output mode of an scc insn is not SImode and
9607 STORE_FLAG_VALUE == 1 (e.g., the 386).
9609 Similarly, check for a case where the AND's are ZERO_EXTEND
9610 operations from some narrower mode even though a SUBREG is not
9611 present. */
9613 else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND
9614 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9615 && GET_CODE (XEXP (op1, 1)) == CONST_INT)
9617 rtx inner_op0 = XEXP (op0, 0);
9618 rtx inner_op1 = XEXP (op1, 0);
9619 HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1));
9620 HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1));
9621 int changed = 0;
9623 if (GET_CODE (inner_op0) == SUBREG && GET_CODE (inner_op1) == SUBREG
9624 && (GET_MODE_SIZE (GET_MODE (inner_op0))
9625 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner_op0))))
9626 && (GET_MODE (SUBREG_REG (inner_op0))
9627 == GET_MODE (SUBREG_REG (inner_op1)))
9628 && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (inner_op0)))
9629 <= HOST_BITS_PER_WIDE_INT)
9630 && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0),
9631 GET_MODE (SUBREG_REG (inner_op0)))))
9632 && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1),
9633 GET_MODE (SUBREG_REG (inner_op1))))))
9635 op0 = SUBREG_REG (inner_op0);
9636 op1 = SUBREG_REG (inner_op1);
9638 /* The resulting comparison is always unsigned since we masked
9639 off the original sign bit. */
9640 code = unsigned_condition (code);
9642 changed = 1;
9645 else if (c0 == c1)
9646 for (tmode = GET_CLASS_NARROWEST_MODE
9647 (GET_MODE_CLASS (GET_MODE (op0)));
9648 tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode))
9649 if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode))
9651 op0 = gen_lowpart (tmode, inner_op0);
9652 op1 = gen_lowpart (tmode, inner_op1);
9653 code = unsigned_condition (code);
9654 changed = 1;
9655 break;
9658 if (! changed)
9659 break;
9662 /* If both operands are NOT, we can strip off the outer operation
9663 and adjust the comparison code for swapped operands; similarly for
9664 NEG, except that this must be an equality comparison. */
9665 else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT)
9666 || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG
9667 && (code == EQ || code == NE)))
9668 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code);
9670 else
9671 break;
9674 /* If the first operand is a constant, swap the operands and adjust the
9675 comparison code appropriately, but don't do this if the second operand
9676 is already a constant integer. */
9677 if (swap_commutative_operands_p (op0, op1))
9679 tem = op0, op0 = op1, op1 = tem;
9680 code = swap_condition (code);
9683 /* We now enter a loop during which we will try to simplify the comparison.
9684 For the most part, we only are concerned with comparisons with zero,
9685 but some things may really be comparisons with zero but not start
9686 out looking that way. */
9688 while (GET_CODE (op1) == CONST_INT)
9690 enum machine_mode mode = GET_MODE (op0);
9691 unsigned int mode_width = GET_MODE_BITSIZE (mode);
9692 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9693 int equality_comparison_p;
9694 int sign_bit_comparison_p;
9695 int unsigned_comparison_p;
9696 HOST_WIDE_INT const_op;
9698 /* We only want to handle integral modes. This catches VOIDmode,
9699 CCmode, and the floating-point modes. An exception is that we
9700 can handle VOIDmode if OP0 is a COMPARE or a comparison
9701 operation. */
9703 if (GET_MODE_CLASS (mode) != MODE_INT
9704 && ! (mode == VOIDmode
9705 && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0))))
9706 break;
9708 /* Get the constant we are comparing against and turn off all bits
9709 not on in our mode. */
9710 const_op = INTVAL (op1);
9711 if (mode != VOIDmode)
9712 const_op = trunc_int_for_mode (const_op, mode);
9713 op1 = GEN_INT (const_op);
9715 /* If we are comparing against a constant power of two and the value
9716 being compared can only have that single bit nonzero (e.g., it was
9717 `and'ed with that bit), we can replace this with a comparison
9718 with zero. */
9719 if (const_op
9720 && (code == EQ || code == NE || code == GE || code == GEU
9721 || code == LT || code == LTU)
9722 && mode_width <= HOST_BITS_PER_WIDE_INT
9723 && exact_log2 (const_op) >= 0
9724 && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
9726 code = (code == EQ || code == GE || code == GEU ? NE : EQ);
9727 op1 = const0_rtx, const_op = 0;
9730 /* Similarly, if we are comparing a value known to be either -1 or
9731 0 with -1, change it to the opposite comparison against zero. */
9733 if (const_op == -1
9734 && (code == EQ || code == NE || code == GT || code == LE
9735 || code == GEU || code == LTU)
9736 && num_sign_bit_copies (op0, mode) == mode_width)
9738 code = (code == EQ || code == LE || code == GEU ? NE : EQ);
9739 op1 = const0_rtx, const_op = 0;
9742 /* Do some canonicalizations based on the comparison code. We prefer
9743 comparisons against zero and then prefer equality comparisons.
9744 If we can reduce the size of a constant, we will do that too. */
9746 switch (code)
9748 case LT:
9749 /* < C is equivalent to <= (C - 1) */
9750 if (const_op > 0)
9752 const_op -= 1;
9753 op1 = GEN_INT (const_op);
9754 code = LE;
9755 /* ... fall through to LE case below. */
9757 else
9758 break;
9760 case LE:
9761 /* <= C is equivalent to < (C + 1); we do this for C < 0 */
9762 if (const_op < 0)
9764 const_op += 1;
9765 op1 = GEN_INT (const_op);
9766 code = LT;
9769 /* If we are doing a <= 0 comparison on a value known to have
9770 a zero sign bit, we can replace this with == 0. */
9771 else if (const_op == 0
9772 && mode_width <= HOST_BITS_PER_WIDE_INT
9773 && (nonzero_bits (op0, mode)
9774 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9775 code = EQ;
9776 break;
9778 case GE:
9779 /* >= C is equivalent to > (C - 1). */
9780 if (const_op > 0)
9782 const_op -= 1;
9783 op1 = GEN_INT (const_op);
9784 code = GT;
9785 /* ... fall through to GT below. */
9787 else
9788 break;
9790 case GT:
9791 /* > C is equivalent to >= (C + 1); we do this for C < 0. */
9792 if (const_op < 0)
9794 const_op += 1;
9795 op1 = GEN_INT (const_op);
9796 code = GE;
9799 /* If we are doing a > 0 comparison on a value known to have
9800 a zero sign bit, we can replace this with != 0. */
9801 else if (const_op == 0
9802 && mode_width <= HOST_BITS_PER_WIDE_INT
9803 && (nonzero_bits (op0, mode)
9804 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9805 code = NE;
9806 break;
9808 case LTU:
9809 /* < C is equivalent to <= (C - 1). */
9810 if (const_op > 0)
9812 const_op -= 1;
9813 op1 = GEN_INT (const_op);
9814 code = LEU;
9815 /* ... fall through ... */
9818 /* (unsigned) < 0x80000000 is equivalent to >= 0. */
9819 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9820 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9822 const_op = 0, op1 = const0_rtx;
9823 code = GE;
9824 break;
9826 else
9827 break;
9829 case LEU:
9830 /* unsigned <= 0 is equivalent to == 0 */
9831 if (const_op == 0)
9832 code = EQ;
9834 /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */
9835 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9836 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9838 const_op = 0, op1 = const0_rtx;
9839 code = GE;
9841 break;
9843 case GEU:
9844 /* >= C is equivalent to > (C - 1). */
9845 if (const_op > 1)
9847 const_op -= 1;
9848 op1 = GEN_INT (const_op);
9849 code = GTU;
9850 /* ... fall through ... */
9853 /* (unsigned) >= 0x80000000 is equivalent to < 0. */
9854 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9855 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9857 const_op = 0, op1 = const0_rtx;
9858 code = LT;
9859 break;
9861 else
9862 break;
9864 case GTU:
9865 /* unsigned > 0 is equivalent to != 0 */
9866 if (const_op == 0)
9867 code = NE;
9869 /* (unsigned) > 0x7fffffff is equivalent to < 0. */
9870 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9871 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9873 const_op = 0, op1 = const0_rtx;
9874 code = LT;
9876 break;
9878 default:
9879 break;
9882 /* Compute some predicates to simplify code below. */
9884 equality_comparison_p = (code == EQ || code == NE);
9885 sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0);
9886 unsigned_comparison_p = (code == LTU || code == LEU || code == GTU
9887 || code == GEU);
9889 /* If this is a sign bit comparison and we can do arithmetic in
9890 MODE, say that we will only be needing the sign bit of OP0. */
9891 if (sign_bit_comparison_p
9892 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
9893 op0 = force_to_mode (op0, mode,
9894 ((HOST_WIDE_INT) 1
9895 << (GET_MODE_BITSIZE (mode) - 1)),
9898 /* Now try cases based on the opcode of OP0. If none of the cases
9899 does a "continue", we exit this loop immediately after the
9900 switch. */
9902 switch (GET_CODE (op0))
9904 case ZERO_EXTRACT:
9905 /* If we are extracting a single bit from a variable position in
9906 a constant that has only a single bit set and are comparing it
9907 with zero, we can convert this into an equality comparison
9908 between the position and the location of the single bit. */
9909 /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might
9910 have already reduced the shift count modulo the word size. */
9911 if (!SHIFT_COUNT_TRUNCATED
9912 && GET_CODE (XEXP (op0, 0)) == CONST_INT
9913 && XEXP (op0, 1) == const1_rtx
9914 && equality_comparison_p && const_op == 0
9915 && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0)
9917 if (BITS_BIG_ENDIAN)
9919 enum machine_mode new_mode
9920 = mode_for_extraction (EP_extzv, 1);
9921 if (new_mode == MAX_MACHINE_MODE)
9922 i = BITS_PER_WORD - 1 - i;
9923 else
9925 mode = new_mode;
9926 i = (GET_MODE_BITSIZE (mode) - 1 - i);
9930 op0 = XEXP (op0, 2);
9931 op1 = GEN_INT (i);
9932 const_op = i;
9934 /* Result is nonzero iff shift count is equal to I. */
9935 code = reverse_condition (code);
9936 continue;
9939 /* ... fall through ... */
9941 case SIGN_EXTRACT:
9942 tem = expand_compound_operation (op0);
9943 if (tem != op0)
9945 op0 = tem;
9946 continue;
9948 break;
9950 case NOT:
9951 /* If testing for equality, we can take the NOT of the constant. */
9952 if (equality_comparison_p
9953 && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0)
9955 op0 = XEXP (op0, 0);
9956 op1 = tem;
9957 continue;
9960 /* If just looking at the sign bit, reverse the sense of the
9961 comparison. */
9962 if (sign_bit_comparison_p)
9964 op0 = XEXP (op0, 0);
9965 code = (code == GE ? LT : GE);
9966 continue;
9968 break;
9970 case NEG:
9971 /* If testing for equality, we can take the NEG of the constant. */
9972 if (equality_comparison_p
9973 && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0)
9975 op0 = XEXP (op0, 0);
9976 op1 = tem;
9977 continue;
9980 /* The remaining cases only apply to comparisons with zero. */
9981 if (const_op != 0)
9982 break;
9984 /* When X is ABS or is known positive,
9985 (neg X) is < 0 if and only if X != 0. */
9987 if (sign_bit_comparison_p
9988 && (GET_CODE (XEXP (op0, 0)) == ABS
9989 || (mode_width <= HOST_BITS_PER_WIDE_INT
9990 && (nonzero_bits (XEXP (op0, 0), mode)
9991 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)))
9993 op0 = XEXP (op0, 0);
9994 code = (code == LT ? NE : EQ);
9995 continue;
9998 /* If we have NEG of something whose two high-order bits are the
9999 same, we know that "(-a) < 0" is equivalent to "a > 0". */
10000 if (num_sign_bit_copies (op0, mode) >= 2)
10002 op0 = XEXP (op0, 0);
10003 code = swap_condition (code);
10004 continue;
10006 break;
10008 case ROTATE:
10009 /* If we are testing equality and our count is a constant, we
10010 can perform the inverse operation on our RHS. */
10011 if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10012 && (tem = simplify_binary_operation (ROTATERT, mode,
10013 op1, XEXP (op0, 1))) != 0)
10015 op0 = XEXP (op0, 0);
10016 op1 = tem;
10017 continue;
10020 /* If we are doing a < 0 or >= 0 comparison, it means we are testing
10021 a particular bit. Convert it to an AND of a constant of that
10022 bit. This will be converted into a ZERO_EXTRACT. */
10023 if (const_op == 0 && sign_bit_comparison_p
10024 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10025 && mode_width <= HOST_BITS_PER_WIDE_INT)
10027 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10028 ((HOST_WIDE_INT) 1
10029 << (mode_width - 1
10030 - INTVAL (XEXP (op0, 1)))));
10031 code = (code == LT ? NE : EQ);
10032 continue;
10035 /* Fall through. */
10037 case ABS:
10038 /* ABS is ignorable inside an equality comparison with zero. */
10039 if (const_op == 0 && equality_comparison_p)
10041 op0 = XEXP (op0, 0);
10042 continue;
10044 break;
10046 case SIGN_EXTEND:
10047 /* Can simplify (compare (zero/sign_extend FOO) CONST) to
10048 (compare FOO CONST) if CONST fits in FOO's mode and we
10049 are either testing inequality or have an unsigned
10050 comparison with ZERO_EXTEND or a signed comparison with
10051 SIGN_EXTEND. But don't do it if we don't have a compare
10052 insn of the given mode, since we'd have to revert it
10053 later on, and then we wouldn't know whether to sign- or
10054 zero-extend. */
10055 mode = GET_MODE (XEXP (op0, 0));
10056 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10057 && ! unsigned_comparison_p
10058 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10059 && ((unsigned HOST_WIDE_INT) const_op
10060 < (((unsigned HOST_WIDE_INT) 1
10061 << (GET_MODE_BITSIZE (mode) - 1))))
10062 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10064 op0 = XEXP (op0, 0);
10065 continue;
10067 break;
10069 case SUBREG:
10070 /* Check for the case where we are comparing A - C1 with C2, that is
10072 (subreg:MODE (plus (A) (-C1))) op (C2)
10074 with C1 a constant, and try to lift the SUBREG, i.e. to do the
10075 comparison in the wider mode. One of the following two conditions
10076 must be true in order for this to be valid:
10078 1. The mode extension results in the same bit pattern being added
10079 on both sides and the comparison is equality or unsigned. As
10080 C2 has been truncated to fit in MODE, the pattern can only be
10081 all 0s or all 1s.
10083 2. The mode extension results in the sign bit being copied on
10084 each side.
10086 The difficulty here is that we have predicates for A but not for
10087 (A - C1) so we need to check that C1 is within proper bounds so
10088 as to perturbate A as little as possible. */
10090 if (mode_width <= HOST_BITS_PER_WIDE_INT
10091 && subreg_lowpart_p (op0)
10092 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) > mode_width
10093 && GET_CODE (SUBREG_REG (op0)) == PLUS
10094 && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT)
10096 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
10097 rtx a = XEXP (SUBREG_REG (op0), 0);
10098 HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1));
10100 if ((c1 > 0
10101 && (unsigned HOST_WIDE_INT) c1
10102 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)
10103 && (equality_comparison_p || unsigned_comparison_p)
10104 /* (A - C1) zero-extends if it is positive and sign-extends
10105 if it is negative, C2 both zero- and sign-extends. */
10106 && ((0 == (nonzero_bits (a, inner_mode)
10107 & ~GET_MODE_MASK (mode))
10108 && const_op >= 0)
10109 /* (A - C1) sign-extends if it is positive and 1-extends
10110 if it is negative, C2 both sign- and 1-extends. */
10111 || (num_sign_bit_copies (a, inner_mode)
10112 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10113 - mode_width)
10114 && const_op < 0)))
10115 || ((unsigned HOST_WIDE_INT) c1
10116 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 2)
10117 /* (A - C1) always sign-extends, like C2. */
10118 && num_sign_bit_copies (a, inner_mode)
10119 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10120 - (mode_width - 1))))
10122 op0 = SUBREG_REG (op0);
10123 continue;
10127 /* If the inner mode is narrower and we are extracting the low part,
10128 we can treat the SUBREG as if it were a ZERO_EXTEND. */
10129 if (subreg_lowpart_p (op0)
10130 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width)
10131 /* Fall through */ ;
10132 else
10133 break;
10135 /* ... fall through ... */
10137 case ZERO_EXTEND:
10138 mode = GET_MODE (XEXP (op0, 0));
10139 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10140 && (unsigned_comparison_p || equality_comparison_p)
10141 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10142 && ((unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode))
10143 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10145 op0 = XEXP (op0, 0);
10146 continue;
10148 break;
10150 case PLUS:
10151 /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do
10152 this for equality comparisons due to pathological cases involving
10153 overflows. */
10154 if (equality_comparison_p
10155 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10156 op1, XEXP (op0, 1))))
10158 op0 = XEXP (op0, 0);
10159 op1 = tem;
10160 continue;
10163 /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */
10164 if (const_op == 0 && XEXP (op0, 1) == constm1_rtx
10165 && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p)
10167 op0 = XEXP (XEXP (op0, 0), 0);
10168 code = (code == LT ? EQ : NE);
10169 continue;
10171 break;
10173 case MINUS:
10174 /* We used to optimize signed comparisons against zero, but that
10175 was incorrect. Unsigned comparisons against zero (GTU, LEU)
10176 arrive here as equality comparisons, or (GEU, LTU) are
10177 optimized away. No need to special-case them. */
10179 /* (eq (minus A B) C) -> (eq A (plus B C)) or
10180 (eq B (minus A C)), whichever simplifies. We can only do
10181 this for equality comparisons due to pathological cases involving
10182 overflows. */
10183 if (equality_comparison_p
10184 && 0 != (tem = simplify_binary_operation (PLUS, mode,
10185 XEXP (op0, 1), op1)))
10187 op0 = XEXP (op0, 0);
10188 op1 = tem;
10189 continue;
10192 if (equality_comparison_p
10193 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10194 XEXP (op0, 0), op1)))
10196 op0 = XEXP (op0, 1);
10197 op1 = tem;
10198 continue;
10201 /* The sign bit of (minus (ashiftrt X C) X), where C is the number
10202 of bits in X minus 1, is one iff X > 0. */
10203 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT
10204 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10205 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (op0, 0), 1))
10206 == mode_width - 1
10207 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10209 op0 = XEXP (op0, 1);
10210 code = (code == GE ? LE : GT);
10211 continue;
10213 break;
10215 case XOR:
10216 /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification
10217 if C is zero or B is a constant. */
10218 if (equality_comparison_p
10219 && 0 != (tem = simplify_binary_operation (XOR, mode,
10220 XEXP (op0, 1), op1)))
10222 op0 = XEXP (op0, 0);
10223 op1 = tem;
10224 continue;
10226 break;
10228 case EQ: case NE:
10229 case UNEQ: case LTGT:
10230 case LT: case LTU: case UNLT: case LE: case LEU: case UNLE:
10231 case GT: case GTU: case UNGT: case GE: case GEU: case UNGE:
10232 case UNORDERED: case ORDERED:
10233 /* We can't do anything if OP0 is a condition code value, rather
10234 than an actual data value. */
10235 if (const_op != 0
10236 || CC0_P (XEXP (op0, 0))
10237 || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC)
10238 break;
10240 /* Get the two operands being compared. */
10241 if (GET_CODE (XEXP (op0, 0)) == COMPARE)
10242 tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1);
10243 else
10244 tem = XEXP (op0, 0), tem1 = XEXP (op0, 1);
10246 /* Check for the cases where we simply want the result of the
10247 earlier test or the opposite of that result. */
10248 if (code == NE || code == EQ
10249 || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
10250 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10251 && (STORE_FLAG_VALUE
10252 & (((HOST_WIDE_INT) 1
10253 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
10254 && (code == LT || code == GE)))
10256 enum rtx_code new_code;
10257 if (code == LT || code == NE)
10258 new_code = GET_CODE (op0);
10259 else
10260 new_code = reversed_comparison_code (op0, NULL);
10262 if (new_code != UNKNOWN)
10264 code = new_code;
10265 op0 = tem;
10266 op1 = tem1;
10267 continue;
10270 break;
10272 case IOR:
10273 /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero
10274 iff X <= 0. */
10275 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS
10276 && XEXP (XEXP (op0, 0), 1) == constm1_rtx
10277 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10279 op0 = XEXP (op0, 1);
10280 code = (code == GE ? GT : LE);
10281 continue;
10283 break;
10285 case AND:
10286 /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This
10287 will be converted to a ZERO_EXTRACT later. */
10288 if (const_op == 0 && equality_comparison_p
10289 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10290 && XEXP (XEXP (op0, 0), 0) == const1_rtx)
10292 op0 = simplify_and_const_int
10293 (NULL_RTX, mode, gen_rtx_LSHIFTRT (mode,
10294 XEXP (op0, 1),
10295 XEXP (XEXP (op0, 0), 1)),
10296 (HOST_WIDE_INT) 1);
10297 continue;
10300 /* If we are comparing (and (lshiftrt X C1) C2) for equality with
10301 zero and X is a comparison and C1 and C2 describe only bits set
10302 in STORE_FLAG_VALUE, we can compare with X. */
10303 if (const_op == 0 && equality_comparison_p
10304 && mode_width <= HOST_BITS_PER_WIDE_INT
10305 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10306 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT
10307 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10308 && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0
10309 && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT)
10311 mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10312 << INTVAL (XEXP (XEXP (op0, 0), 1)));
10313 if ((~STORE_FLAG_VALUE & mask) == 0
10314 && (COMPARISON_P (XEXP (XEXP (op0, 0), 0))
10315 || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0
10316 && COMPARISON_P (tem))))
10318 op0 = XEXP (XEXP (op0, 0), 0);
10319 continue;
10323 /* If we are doing an equality comparison of an AND of a bit equal
10324 to the sign bit, replace this with a LT or GE comparison of
10325 the underlying value. */
10326 if (equality_comparison_p
10327 && const_op == 0
10328 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10329 && mode_width <= HOST_BITS_PER_WIDE_INT
10330 && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10331 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10333 op0 = XEXP (op0, 0);
10334 code = (code == EQ ? GE : LT);
10335 continue;
10338 /* If this AND operation is really a ZERO_EXTEND from a narrower
10339 mode, the constant fits within that mode, and this is either an
10340 equality or unsigned comparison, try to do this comparison in
10341 the narrower mode.
10343 Note that in:
10345 (ne:DI (and:DI (reg:DI 4) (const_int 0xffffffff)) (const_int 0))
10346 -> (ne:DI (reg:SI 4) (const_int 0))
10348 unless TRULY_NOOP_TRUNCATION allows it or the register is
10349 known to hold a value of the required mode the
10350 transformation is invalid. */
10351 if ((equality_comparison_p || unsigned_comparison_p)
10352 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10353 && (i = exact_log2 ((INTVAL (XEXP (op0, 1))
10354 & GET_MODE_MASK (mode))
10355 + 1)) >= 0
10356 && const_op >> i == 0
10357 && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode
10358 && (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
10359 GET_MODE_BITSIZE (GET_MODE (op0)))
10360 || (REG_P (XEXP (op0, 0))
10361 && reg_truncated_to_mode (tmode, XEXP (op0, 0)))))
10363 op0 = gen_lowpart (tmode, XEXP (op0, 0));
10364 continue;
10367 /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1
10368 fits in both M1 and M2 and the SUBREG is either paradoxical
10369 or represents the low part, permute the SUBREG and the AND
10370 and try again. */
10371 if (GET_CODE (XEXP (op0, 0)) == SUBREG)
10373 unsigned HOST_WIDE_INT c1;
10374 tmode = GET_MODE (SUBREG_REG (XEXP (op0, 0)));
10375 /* Require an integral mode, to avoid creating something like
10376 (AND:SF ...). */
10377 if (SCALAR_INT_MODE_P (tmode)
10378 /* It is unsafe to commute the AND into the SUBREG if the
10379 SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is
10380 not defined. As originally written the upper bits
10381 have a defined value due to the AND operation.
10382 However, if we commute the AND inside the SUBREG then
10383 they no longer have defined values and the meaning of
10384 the code has been changed. */
10385 && (0
10386 #ifdef WORD_REGISTER_OPERATIONS
10387 || (mode_width > GET_MODE_BITSIZE (tmode)
10388 && mode_width <= BITS_PER_WORD)
10389 #endif
10390 || (mode_width <= GET_MODE_BITSIZE (tmode)
10391 && subreg_lowpart_p (XEXP (op0, 0))))
10392 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10393 && mode_width <= HOST_BITS_PER_WIDE_INT
10394 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
10395 && ((c1 = INTVAL (XEXP (op0, 1))) & ~mask) == 0
10396 && (c1 & ~GET_MODE_MASK (tmode)) == 0
10397 && c1 != mask
10398 && c1 != GET_MODE_MASK (tmode))
10400 op0 = simplify_gen_binary (AND, tmode,
10401 SUBREG_REG (XEXP (op0, 0)),
10402 gen_int_mode (c1, tmode));
10403 op0 = gen_lowpart (mode, op0);
10404 continue;
10408 /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */
10409 if (const_op == 0 && equality_comparison_p
10410 && XEXP (op0, 1) == const1_rtx
10411 && GET_CODE (XEXP (op0, 0)) == NOT)
10413 op0 = simplify_and_const_int
10414 (NULL_RTX, mode, XEXP (XEXP (op0, 0), 0), (HOST_WIDE_INT) 1);
10415 code = (code == NE ? EQ : NE);
10416 continue;
10419 /* Convert (ne (and (lshiftrt (not X)) 1) 0) to
10420 (eq (and (lshiftrt X) 1) 0).
10421 Also handle the case where (not X) is expressed using xor. */
10422 if (const_op == 0 && equality_comparison_p
10423 && XEXP (op0, 1) == const1_rtx
10424 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT)
10426 rtx shift_op = XEXP (XEXP (op0, 0), 0);
10427 rtx shift_count = XEXP (XEXP (op0, 0), 1);
10429 if (GET_CODE (shift_op) == NOT
10430 || (GET_CODE (shift_op) == XOR
10431 && GET_CODE (XEXP (shift_op, 1)) == CONST_INT
10432 && GET_CODE (shift_count) == CONST_INT
10433 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
10434 && (INTVAL (XEXP (shift_op, 1))
10435 == (HOST_WIDE_INT) 1 << INTVAL (shift_count))))
10437 op0 = simplify_and_const_int
10438 (NULL_RTX, mode,
10439 gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count),
10440 (HOST_WIDE_INT) 1);
10441 code = (code == NE ? EQ : NE);
10442 continue;
10445 break;
10447 case ASHIFT:
10448 /* If we have (compare (ashift FOO N) (const_int C)) and
10449 the high order N bits of FOO (N+1 if an inequality comparison)
10450 are known to be zero, we can do this by comparing FOO with C
10451 shifted right N bits so long as the low-order N bits of C are
10452 zero. */
10453 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10454 && INTVAL (XEXP (op0, 1)) >= 0
10455 && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p)
10456 < HOST_BITS_PER_WIDE_INT)
10457 && ((const_op
10458 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0)
10459 && mode_width <= HOST_BITS_PER_WIDE_INT
10460 && (nonzero_bits (XEXP (op0, 0), mode)
10461 & ~(mask >> (INTVAL (XEXP (op0, 1))
10462 + ! equality_comparison_p))) == 0)
10464 /* We must perform a logical shift, not an arithmetic one,
10465 as we want the top N bits of C to be zero. */
10466 unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode);
10468 temp >>= INTVAL (XEXP (op0, 1));
10469 op1 = gen_int_mode (temp, mode);
10470 op0 = XEXP (op0, 0);
10471 continue;
10474 /* If we are doing a sign bit comparison, it means we are testing
10475 a particular bit. Convert it to the appropriate AND. */
10476 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10477 && mode_width <= HOST_BITS_PER_WIDE_INT)
10479 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10480 ((HOST_WIDE_INT) 1
10481 << (mode_width - 1
10482 - INTVAL (XEXP (op0, 1)))));
10483 code = (code == LT ? NE : EQ);
10484 continue;
10487 /* If this an equality comparison with zero and we are shifting
10488 the low bit to the sign bit, we can convert this to an AND of the
10489 low-order bit. */
10490 if (const_op == 0 && equality_comparison_p
10491 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10492 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10493 == mode_width - 1)
10495 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10496 (HOST_WIDE_INT) 1);
10497 continue;
10499 break;
10501 case ASHIFTRT:
10502 /* If this is an equality comparison with zero, we can do this
10503 as a logical shift, which might be much simpler. */
10504 if (equality_comparison_p && const_op == 0
10505 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
10507 op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode,
10508 XEXP (op0, 0),
10509 INTVAL (XEXP (op0, 1)));
10510 continue;
10513 /* If OP0 is a sign extension and CODE is not an unsigned comparison,
10514 do the comparison in a narrower mode. */
10515 if (! unsigned_comparison_p
10516 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10517 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10518 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
10519 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10520 MODE_INT, 1)) != BLKmode
10521 && (((unsigned HOST_WIDE_INT) const_op
10522 + (GET_MODE_MASK (tmode) >> 1) + 1)
10523 <= GET_MODE_MASK (tmode)))
10525 op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0));
10526 continue;
10529 /* Likewise if OP0 is a PLUS of a sign extension with a
10530 constant, which is usually represented with the PLUS
10531 between the shifts. */
10532 if (! unsigned_comparison_p
10533 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10534 && GET_CODE (XEXP (op0, 0)) == PLUS
10535 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10536 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT
10537 && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1)
10538 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10539 MODE_INT, 1)) != BLKmode
10540 && (((unsigned HOST_WIDE_INT) const_op
10541 + (GET_MODE_MASK (tmode) >> 1) + 1)
10542 <= GET_MODE_MASK (tmode)))
10544 rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0);
10545 rtx add_const = XEXP (XEXP (op0, 0), 1);
10546 rtx new_const = simplify_gen_binary (ASHIFTRT, GET_MODE (op0),
10547 add_const, XEXP (op0, 1));
10549 op0 = simplify_gen_binary (PLUS, tmode,
10550 gen_lowpart (tmode, inner),
10551 new_const);
10552 continue;
10555 /* ... fall through ... */
10556 case LSHIFTRT:
10557 /* If we have (compare (xshiftrt FOO N) (const_int C)) and
10558 the low order N bits of FOO are known to be zero, we can do this
10559 by comparing FOO with C shifted left N bits so long as no
10560 overflow occurs. */
10561 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10562 && INTVAL (XEXP (op0, 1)) >= 0
10563 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
10564 && mode_width <= HOST_BITS_PER_WIDE_INT
10565 && (nonzero_bits (XEXP (op0, 0), mode)
10566 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0
10567 && (((unsigned HOST_WIDE_INT) const_op
10568 + (GET_CODE (op0) != LSHIFTRT
10569 ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1)
10570 + 1)
10571 : 0))
10572 <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1))))
10574 /* If the shift was logical, then we must make the condition
10575 unsigned. */
10576 if (GET_CODE (op0) == LSHIFTRT)
10577 code = unsigned_condition (code);
10579 const_op <<= INTVAL (XEXP (op0, 1));
10580 op1 = GEN_INT (const_op);
10581 op0 = XEXP (op0, 0);
10582 continue;
10585 /* If we are using this shift to extract just the sign bit, we
10586 can replace this with an LT or GE comparison. */
10587 if (const_op == 0
10588 && (equality_comparison_p || sign_bit_comparison_p)
10589 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10590 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10591 == mode_width - 1)
10593 op0 = XEXP (op0, 0);
10594 code = (code == NE || code == GT ? LT : GE);
10595 continue;
10597 break;
10599 default:
10600 break;
10603 break;
10606 /* Now make any compound operations involved in this comparison. Then,
10607 check for an outmost SUBREG on OP0 that is not doing anything or is
10608 paradoxical. The latter transformation must only be performed when
10609 it is known that the "extra" bits will be the same in op0 and op1 or
10610 that they don't matter. There are three cases to consider:
10612 1. SUBREG_REG (op0) is a register. In this case the bits are don't
10613 care bits and we can assume they have any convenient value. So
10614 making the transformation is safe.
10616 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not defined.
10617 In this case the upper bits of op0 are undefined. We should not make
10618 the simplification in that case as we do not know the contents of
10619 those bits.
10621 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined and not
10622 UNKNOWN. In that case we know those bits are zeros or ones. We must
10623 also be sure that they are the same as the upper bits of op1.
10625 We can never remove a SUBREG for a non-equality comparison because
10626 the sign bit is in a different place in the underlying object. */
10628 op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET);
10629 op1 = make_compound_operation (op1, SET);
10631 if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0)
10632 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10633 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op0))) == MODE_INT
10634 && (code == NE || code == EQ))
10636 if (GET_MODE_SIZE (GET_MODE (op0))
10637 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))
10639 /* For paradoxical subregs, allow case 1 as above. Case 3 isn't
10640 implemented. */
10641 if (REG_P (SUBREG_REG (op0)))
10643 op0 = SUBREG_REG (op0);
10644 op1 = gen_lowpart (GET_MODE (op0), op1);
10647 else if ((GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0)))
10648 <= HOST_BITS_PER_WIDE_INT)
10649 && (nonzero_bits (SUBREG_REG (op0),
10650 GET_MODE (SUBREG_REG (op0)))
10651 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10653 tem = gen_lowpart (GET_MODE (SUBREG_REG (op0)), op1);
10655 if ((nonzero_bits (tem, GET_MODE (SUBREG_REG (op0)))
10656 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10657 op0 = SUBREG_REG (op0), op1 = tem;
10661 /* We now do the opposite procedure: Some machines don't have compare
10662 insns in all modes. If OP0's mode is an integer mode smaller than a
10663 word and we can't do a compare in that mode, see if there is a larger
10664 mode for which we can do the compare. There are a number of cases in
10665 which we can use the wider mode. */
10667 mode = GET_MODE (op0);
10668 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10669 && GET_MODE_SIZE (mode) < UNITS_PER_WORD
10670 && ! have_insn_for (COMPARE, mode))
10671 for (tmode = GET_MODE_WIDER_MODE (mode);
10672 (tmode != VOIDmode
10673 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT);
10674 tmode = GET_MODE_WIDER_MODE (tmode))
10675 if (have_insn_for (COMPARE, tmode))
10677 int zero_extended;
10679 /* If the only nonzero bits in OP0 and OP1 are those in the
10680 narrower mode and this is an equality or unsigned comparison,
10681 we can use the wider mode. Similarly for sign-extended
10682 values, in which case it is true for all comparisons. */
10683 zero_extended = ((code == EQ || code == NE
10684 || code == GEU || code == GTU
10685 || code == LEU || code == LTU)
10686 && (nonzero_bits (op0, tmode)
10687 & ~GET_MODE_MASK (mode)) == 0
10688 && ((GET_CODE (op1) == CONST_INT
10689 || (nonzero_bits (op1, tmode)
10690 & ~GET_MODE_MASK (mode)) == 0)));
10692 if (zero_extended
10693 || ((num_sign_bit_copies (op0, tmode)
10694 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10695 - GET_MODE_BITSIZE (mode)))
10696 && (num_sign_bit_copies (op1, tmode)
10697 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10698 - GET_MODE_BITSIZE (mode)))))
10700 /* If OP0 is an AND and we don't have an AND in MODE either,
10701 make a new AND in the proper mode. */
10702 if (GET_CODE (op0) == AND
10703 && !have_insn_for (AND, mode))
10704 op0 = simplify_gen_binary (AND, tmode,
10705 gen_lowpart (tmode,
10706 XEXP (op0, 0)),
10707 gen_lowpart (tmode,
10708 XEXP (op0, 1)));
10710 op0 = gen_lowpart (tmode, op0);
10711 if (zero_extended && GET_CODE (op1) == CONST_INT)
10712 op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (mode));
10713 op1 = gen_lowpart (tmode, op1);
10714 break;
10717 /* If this is a test for negative, we can make an explicit
10718 test of the sign bit. */
10720 if (op1 == const0_rtx && (code == LT || code == GE)
10721 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10723 op0 = simplify_gen_binary (AND, tmode,
10724 gen_lowpart (tmode, op0),
10725 GEN_INT ((HOST_WIDE_INT) 1
10726 << (GET_MODE_BITSIZE (mode)
10727 - 1)));
10728 code = (code == LT) ? NE : EQ;
10729 break;
10733 #ifdef CANONICALIZE_COMPARISON
10734 /* If this machine only supports a subset of valid comparisons, see if we
10735 can convert an unsupported one into a supported one. */
10736 CANONICALIZE_COMPARISON (code, op0, op1);
10737 #endif
10739 *pop0 = op0;
10740 *pop1 = op1;
10742 return code;
10745 /* Utility function for record_value_for_reg. Count number of
10746 rtxs in X. */
10747 static int
10748 count_rtxs (rtx x)
10750 enum rtx_code code = GET_CODE (x);
10751 const char *fmt;
10752 int i, ret = 1;
10754 if (GET_RTX_CLASS (code) == '2'
10755 || GET_RTX_CLASS (code) == 'c')
10757 rtx x0 = XEXP (x, 0);
10758 rtx x1 = XEXP (x, 1);
10760 if (x0 == x1)
10761 return 1 + 2 * count_rtxs (x0);
10763 if ((GET_RTX_CLASS (GET_CODE (x1)) == '2'
10764 || GET_RTX_CLASS (GET_CODE (x1)) == 'c')
10765 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10766 return 2 + 2 * count_rtxs (x0)
10767 + count_rtxs (x == XEXP (x1, 0)
10768 ? XEXP (x1, 1) : XEXP (x1, 0));
10770 if ((GET_RTX_CLASS (GET_CODE (x0)) == '2'
10771 || GET_RTX_CLASS (GET_CODE (x0)) == 'c')
10772 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10773 return 2 + 2 * count_rtxs (x1)
10774 + count_rtxs (x == XEXP (x0, 0)
10775 ? XEXP (x0, 1) : XEXP (x0, 0));
10778 fmt = GET_RTX_FORMAT (code);
10779 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10780 if (fmt[i] == 'e')
10781 ret += count_rtxs (XEXP (x, i));
10783 return ret;
10786 /* Utility function for following routine. Called when X is part of a value
10787 being stored into last_set_value. Sets last_set_table_tick
10788 for each register mentioned. Similar to mention_regs in cse.c */
10790 static void
10791 update_table_tick (rtx x)
10793 enum rtx_code code = GET_CODE (x);
10794 const char *fmt = GET_RTX_FORMAT (code);
10795 int i;
10797 if (code == REG)
10799 unsigned int regno = REGNO (x);
10800 unsigned int endregno
10801 = regno + (regno < FIRST_PSEUDO_REGISTER
10802 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
10803 unsigned int r;
10805 for (r = regno; r < endregno; r++)
10806 reg_stat[r].last_set_table_tick = label_tick;
10808 return;
10811 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10812 /* Note that we can't have an "E" in values stored; see
10813 get_last_value_validate. */
10814 if (fmt[i] == 'e')
10816 /* Check for identical subexpressions. If x contains
10817 identical subexpression we only have to traverse one of
10818 them. */
10819 if (i == 0 && ARITHMETIC_P (x))
10821 /* Note that at this point x1 has already been
10822 processed. */
10823 rtx x0 = XEXP (x, 0);
10824 rtx x1 = XEXP (x, 1);
10826 /* If x0 and x1 are identical then there is no need to
10827 process x0. */
10828 if (x0 == x1)
10829 break;
10831 /* If x0 is identical to a subexpression of x1 then while
10832 processing x1, x0 has already been processed. Thus we
10833 are done with x. */
10834 if (ARITHMETIC_P (x1)
10835 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10836 break;
10838 /* If x1 is identical to a subexpression of x0 then we
10839 still have to process the rest of x0. */
10840 if (ARITHMETIC_P (x0)
10841 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10843 update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0));
10844 break;
10848 update_table_tick (XEXP (x, i));
10852 /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we
10853 are saying that the register is clobbered and we no longer know its
10854 value. If INSN is zero, don't update reg_stat[].last_set; this is
10855 only permitted with VALUE also zero and is used to invalidate the
10856 register. */
10858 static void
10859 record_value_for_reg (rtx reg, rtx insn, rtx value)
10861 unsigned int regno = REGNO (reg);
10862 unsigned int endregno
10863 = regno + (regno < FIRST_PSEUDO_REGISTER
10864 ? hard_regno_nregs[regno][GET_MODE (reg)] : 1);
10865 unsigned int i;
10867 /* If VALUE contains REG and we have a previous value for REG, substitute
10868 the previous value. */
10869 if (value && insn && reg_overlap_mentioned_p (reg, value))
10871 rtx tem;
10873 /* Set things up so get_last_value is allowed to see anything set up to
10874 our insn. */
10875 subst_low_cuid = INSN_CUID (insn);
10876 tem = get_last_value (reg);
10878 /* If TEM is simply a binary operation with two CLOBBERs as operands,
10879 it isn't going to be useful and will take a lot of time to process,
10880 so just use the CLOBBER. */
10882 if (tem)
10884 if (ARITHMETIC_P (tem)
10885 && GET_CODE (XEXP (tem, 0)) == CLOBBER
10886 && GET_CODE (XEXP (tem, 1)) == CLOBBER)
10887 tem = XEXP (tem, 0);
10888 else if (count_occurrences (value, reg, 1) >= 2)
10890 /* If there are two or more occurrences of REG in VALUE,
10891 prevent the value from growing too much. */
10892 if (count_rtxs (tem) > MAX_LAST_VALUE_RTL)
10893 tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx);
10896 value = replace_rtx (copy_rtx (value), reg, tem);
10900 /* For each register modified, show we don't know its value, that
10901 we don't know about its bitwise content, that its value has been
10902 updated, and that we don't know the location of the death of the
10903 register. */
10904 for (i = regno; i < endregno; i++)
10906 if (insn)
10907 reg_stat[i].last_set = insn;
10909 reg_stat[i].last_set_value = 0;
10910 reg_stat[i].last_set_mode = 0;
10911 reg_stat[i].last_set_nonzero_bits = 0;
10912 reg_stat[i].last_set_sign_bit_copies = 0;
10913 reg_stat[i].last_death = 0;
10914 reg_stat[i].truncated_to_mode = 0;
10917 /* Mark registers that are being referenced in this value. */
10918 if (value)
10919 update_table_tick (value);
10921 /* Now update the status of each register being set.
10922 If someone is using this register in this block, set this register
10923 to invalid since we will get confused between the two lives in this
10924 basic block. This makes using this register always invalid. In cse, we
10925 scan the table to invalidate all entries using this register, but this
10926 is too much work for us. */
10928 for (i = regno; i < endregno; i++)
10930 reg_stat[i].last_set_label = label_tick;
10931 if (!insn || (value && reg_stat[i].last_set_table_tick == label_tick))
10932 reg_stat[i].last_set_invalid = 1;
10933 else
10934 reg_stat[i].last_set_invalid = 0;
10937 /* The value being assigned might refer to X (like in "x++;"). In that
10938 case, we must replace it with (clobber (const_int 0)) to prevent
10939 infinite loops. */
10940 if (value && ! get_last_value_validate (&value, insn,
10941 reg_stat[regno].last_set_label, 0))
10943 value = copy_rtx (value);
10944 if (! get_last_value_validate (&value, insn,
10945 reg_stat[regno].last_set_label, 1))
10946 value = 0;
10949 /* For the main register being modified, update the value, the mode, the
10950 nonzero bits, and the number of sign bit copies. */
10952 reg_stat[regno].last_set_value = value;
10954 if (value)
10956 enum machine_mode mode = GET_MODE (reg);
10957 subst_low_cuid = INSN_CUID (insn);
10958 reg_stat[regno].last_set_mode = mode;
10959 if (GET_MODE_CLASS (mode) == MODE_INT
10960 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10961 mode = nonzero_bits_mode;
10962 reg_stat[regno].last_set_nonzero_bits = nonzero_bits (value, mode);
10963 reg_stat[regno].last_set_sign_bit_copies
10964 = num_sign_bit_copies (value, GET_MODE (reg));
10968 /* Called via note_stores from record_dead_and_set_regs to handle one
10969 SET or CLOBBER in an insn. DATA is the instruction in which the
10970 set is occurring. */
10972 static void
10973 record_dead_and_set_regs_1 (rtx dest, rtx setter, void *data)
10975 rtx record_dead_insn = (rtx) data;
10977 if (GET_CODE (dest) == SUBREG)
10978 dest = SUBREG_REG (dest);
10980 if (!record_dead_insn)
10982 if (REG_P (dest))
10983 record_value_for_reg (dest, NULL_RTX, NULL_RTX);
10984 return;
10987 if (REG_P (dest))
10989 /* If we are setting the whole register, we know its value. Otherwise
10990 show that we don't know the value. We can handle SUBREG in
10991 some cases. */
10992 if (GET_CODE (setter) == SET && dest == SET_DEST (setter))
10993 record_value_for_reg (dest, record_dead_insn, SET_SRC (setter));
10994 else if (GET_CODE (setter) == SET
10995 && GET_CODE (SET_DEST (setter)) == SUBREG
10996 && SUBREG_REG (SET_DEST (setter)) == dest
10997 && GET_MODE_BITSIZE (GET_MODE (dest)) <= BITS_PER_WORD
10998 && subreg_lowpart_p (SET_DEST (setter)))
10999 record_value_for_reg (dest, record_dead_insn,
11000 gen_lowpart (GET_MODE (dest),
11001 SET_SRC (setter)));
11002 else
11003 record_value_for_reg (dest, record_dead_insn, NULL_RTX);
11005 else if (MEM_P (dest)
11006 /* Ignore pushes, they clobber nothing. */
11007 && ! push_operand (dest, GET_MODE (dest)))
11008 mem_last_set = INSN_CUID (record_dead_insn);
11011 /* Update the records of when each REG was most recently set or killed
11012 for the things done by INSN. This is the last thing done in processing
11013 INSN in the combiner loop.
11015 We update reg_stat[], in particular fields last_set, last_set_value,
11016 last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies,
11017 last_death, and also the similar information mem_last_set (which insn
11018 most recently modified memory) and last_call_cuid (which insn was the
11019 most recent subroutine call). */
11021 static void
11022 record_dead_and_set_regs (rtx insn)
11024 rtx link;
11025 unsigned int i;
11027 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
11029 if (REG_NOTE_KIND (link) == REG_DEAD
11030 && REG_P (XEXP (link, 0)))
11032 unsigned int regno = REGNO (XEXP (link, 0));
11033 unsigned int endregno
11034 = regno + (regno < FIRST_PSEUDO_REGISTER
11035 ? hard_regno_nregs[regno][GET_MODE (XEXP (link, 0))]
11036 : 1);
11038 for (i = regno; i < endregno; i++)
11039 reg_stat[i].last_death = insn;
11041 else if (REG_NOTE_KIND (link) == REG_INC)
11042 record_value_for_reg (XEXP (link, 0), insn, NULL_RTX);
11045 if (CALL_P (insn))
11047 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
11048 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
11050 reg_stat[i].last_set_value = 0;
11051 reg_stat[i].last_set_mode = 0;
11052 reg_stat[i].last_set_nonzero_bits = 0;
11053 reg_stat[i].last_set_sign_bit_copies = 0;
11054 reg_stat[i].last_death = 0;
11055 reg_stat[i].truncated_to_mode = 0;
11058 last_call_cuid = mem_last_set = INSN_CUID (insn);
11060 /* We can't combine into a call pattern. Remember, though, that
11061 the return value register is set at this CUID. We could
11062 still replace a register with the return value from the
11063 wrong subroutine call! */
11064 note_stores (PATTERN (insn), record_dead_and_set_regs_1, NULL_RTX);
11066 else
11067 note_stores (PATTERN (insn), record_dead_and_set_regs_1, insn);
11070 /* If a SUBREG has the promoted bit set, it is in fact a property of the
11071 register present in the SUBREG, so for each such SUBREG go back and
11072 adjust nonzero and sign bit information of the registers that are
11073 known to have some zero/sign bits set.
11075 This is needed because when combine blows the SUBREGs away, the
11076 information on zero/sign bits is lost and further combines can be
11077 missed because of that. */
11079 static void
11080 record_promoted_value (rtx insn, rtx subreg)
11082 rtx links, set;
11083 unsigned int regno = REGNO (SUBREG_REG (subreg));
11084 enum machine_mode mode = GET_MODE (subreg);
11086 if (GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
11087 return;
11089 for (links = LOG_LINKS (insn); links;)
11091 insn = XEXP (links, 0);
11092 set = single_set (insn);
11094 if (! set || !REG_P (SET_DEST (set))
11095 || REGNO (SET_DEST (set)) != regno
11096 || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg)))
11098 links = XEXP (links, 1);
11099 continue;
11102 if (reg_stat[regno].last_set == insn)
11104 if (SUBREG_PROMOTED_UNSIGNED_P (subreg) > 0)
11105 reg_stat[regno].last_set_nonzero_bits &= GET_MODE_MASK (mode);
11108 if (REG_P (SET_SRC (set)))
11110 regno = REGNO (SET_SRC (set));
11111 links = LOG_LINKS (insn);
11113 else
11114 break;
11118 /* Check if X, a register, is known to contain a value already
11119 truncated to MODE. In this case we can use a subreg to refer to
11120 the truncated value even though in the generic case we would need
11121 an explicit truncation. */
11123 static bool
11124 reg_truncated_to_mode (enum machine_mode mode, rtx x)
11126 enum machine_mode truncated = reg_stat[REGNO (x)].truncated_to_mode;
11128 if (truncated == 0 || reg_stat[REGNO (x)].truncation_label != label_tick)
11129 return false;
11130 if (GET_MODE_SIZE (truncated) <= GET_MODE_SIZE (mode))
11131 return true;
11132 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
11133 GET_MODE_BITSIZE (truncated)))
11134 return true;
11135 return false;
11138 /* X is a REG or a SUBREG. If X is some sort of a truncation record
11139 it. For non-TRULY_NOOP_TRUNCATION targets we might be able to turn
11140 a truncate into a subreg using this information. */
11142 static void
11143 record_truncated_value (rtx x)
11145 enum machine_mode truncated_mode;
11147 if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)))
11149 enum machine_mode original_mode = GET_MODE (SUBREG_REG (x));
11150 truncated_mode = GET_MODE (x);
11152 if (GET_MODE_SIZE (original_mode) <= GET_MODE_SIZE (truncated_mode))
11153 return;
11155 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (truncated_mode),
11156 GET_MODE_BITSIZE (original_mode)))
11157 return;
11159 x = SUBREG_REG (x);
11161 /* ??? For hard-regs we now record everything. We might be able to
11162 optimize this using last_set_mode. */
11163 else if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
11164 truncated_mode = GET_MODE (x);
11165 else
11166 return;
11168 if (reg_stat[REGNO (x)].truncated_to_mode == 0
11169 || reg_stat[REGNO (x)].truncation_label < label_tick
11170 || (GET_MODE_SIZE (truncated_mode)
11171 < GET_MODE_SIZE (reg_stat[REGNO (x)].truncated_to_mode)))
11173 reg_stat[REGNO (x)].truncated_to_mode = truncated_mode;
11174 reg_stat[REGNO (x)].truncation_label = label_tick;
11178 /* Scan X for promoted SUBREGs and truncated REGs. For each one
11179 found, note what it implies to the registers used in it. */
11181 static void
11182 check_conversions (rtx insn, rtx x)
11184 if (GET_CODE (x) == SUBREG || REG_P (x))
11186 if (GET_CODE (x) == SUBREG
11187 && SUBREG_PROMOTED_VAR_P (x)
11188 && REG_P (SUBREG_REG (x)))
11189 record_promoted_value (insn, x);
11191 record_truncated_value (x);
11193 else
11195 const char *format = GET_RTX_FORMAT (GET_CODE (x));
11196 int i, j;
11198 for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++)
11199 switch (format[i])
11201 case 'e':
11202 check_conversions (insn, XEXP (x, i));
11203 break;
11204 case 'V':
11205 case 'E':
11206 if (XVEC (x, i) != 0)
11207 for (j = 0; j < XVECLEN (x, i); j++)
11208 check_conversions (insn, XVECEXP (x, i, j));
11209 break;
11214 /* Utility routine for the following function. Verify that all the registers
11215 mentioned in *LOC are valid when *LOC was part of a value set when
11216 label_tick == TICK. Return 0 if some are not.
11218 If REPLACE is nonzero, replace the invalid reference with
11219 (clobber (const_int 0)) and return 1. This replacement is useful because
11220 we often can get useful information about the form of a value (e.g., if
11221 it was produced by a shift that always produces -1 or 0) even though
11222 we don't know exactly what registers it was produced from. */
11224 static int
11225 get_last_value_validate (rtx *loc, rtx insn, int tick, int replace)
11227 rtx x = *loc;
11228 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
11229 int len = GET_RTX_LENGTH (GET_CODE (x));
11230 int i;
11232 if (REG_P (x))
11234 unsigned int regno = REGNO (x);
11235 unsigned int endregno
11236 = regno + (regno < FIRST_PSEUDO_REGISTER
11237 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11238 unsigned int j;
11240 for (j = regno; j < endregno; j++)
11241 if (reg_stat[j].last_set_invalid
11242 /* If this is a pseudo-register that was only set once and not
11243 live at the beginning of the function, it is always valid. */
11244 || (! (regno >= FIRST_PSEUDO_REGISTER
11245 && REG_N_SETS (regno) == 1
11246 && (! REGNO_REG_SET_P
11247 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11248 regno)))
11249 && reg_stat[j].last_set_label > tick))
11251 if (replace)
11252 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11253 return replace;
11256 return 1;
11258 /* If this is a memory reference, make sure that there were
11259 no stores after it that might have clobbered the value. We don't
11260 have alias info, so we assume any store invalidates it. */
11261 else if (MEM_P (x) && !MEM_READONLY_P (x)
11262 && INSN_CUID (insn) <= mem_last_set)
11264 if (replace)
11265 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11266 return replace;
11269 for (i = 0; i < len; i++)
11271 if (fmt[i] == 'e')
11273 /* Check for identical subexpressions. If x contains
11274 identical subexpression we only have to traverse one of
11275 them. */
11276 if (i == 1 && ARITHMETIC_P (x))
11278 /* Note that at this point x0 has already been checked
11279 and found valid. */
11280 rtx x0 = XEXP (x, 0);
11281 rtx x1 = XEXP (x, 1);
11283 /* If x0 and x1 are identical then x is also valid. */
11284 if (x0 == x1)
11285 return 1;
11287 /* If x1 is identical to a subexpression of x0 then
11288 while checking x0, x1 has already been checked. Thus
11289 it is valid and so as x. */
11290 if (ARITHMETIC_P (x0)
11291 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
11292 return 1;
11294 /* If x0 is identical to a subexpression of x1 then x is
11295 valid iff the rest of x1 is valid. */
11296 if (ARITHMETIC_P (x1)
11297 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
11298 return
11299 get_last_value_validate (&XEXP (x1,
11300 x0 == XEXP (x1, 0) ? 1 : 0),
11301 insn, tick, replace);
11304 if (get_last_value_validate (&XEXP (x, i), insn, tick,
11305 replace) == 0)
11306 return 0;
11308 /* Don't bother with these. They shouldn't occur anyway. */
11309 else if (fmt[i] == 'E')
11310 return 0;
11313 /* If we haven't found a reason for it to be invalid, it is valid. */
11314 return 1;
11317 /* Get the last value assigned to X, if known. Some registers
11318 in the value may be replaced with (clobber (const_int 0)) if their value
11319 is known longer known reliably. */
11321 static rtx
11322 get_last_value (rtx x)
11324 unsigned int regno;
11325 rtx value;
11327 /* If this is a non-paradoxical SUBREG, get the value of its operand and
11328 then convert it to the desired mode. If this is a paradoxical SUBREG,
11329 we cannot predict what values the "extra" bits might have. */
11330 if (GET_CODE (x) == SUBREG
11331 && subreg_lowpart_p (x)
11332 && (GET_MODE_SIZE (GET_MODE (x))
11333 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
11334 && (value = get_last_value (SUBREG_REG (x))) != 0)
11335 return gen_lowpart (GET_MODE (x), value);
11337 if (!REG_P (x))
11338 return 0;
11340 regno = REGNO (x);
11341 value = reg_stat[regno].last_set_value;
11343 /* If we don't have a value, or if it isn't for this basic block and
11344 it's either a hard register, set more than once, or it's a live
11345 at the beginning of the function, return 0.
11347 Because if it's not live at the beginning of the function then the reg
11348 is always set before being used (is never used without being set).
11349 And, if it's set only once, and it's always set before use, then all
11350 uses must have the same last value, even if it's not from this basic
11351 block. */
11353 if (value == 0
11354 || (reg_stat[regno].last_set_label != label_tick
11355 && (regno < FIRST_PSEUDO_REGISTER
11356 || REG_N_SETS (regno) != 1
11357 || (REGNO_REG_SET_P
11358 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11359 regno)))))
11360 return 0;
11362 /* If the value was set in a later insn than the ones we are processing,
11363 we can't use it even if the register was only set once. */
11364 if (INSN_CUID (reg_stat[regno].last_set) >= subst_low_cuid)
11365 return 0;
11367 /* If the value has all its registers valid, return it. */
11368 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11369 reg_stat[regno].last_set_label, 0))
11370 return value;
11372 /* Otherwise, make a copy and replace any invalid register with
11373 (clobber (const_int 0)). If that fails for some reason, return 0. */
11375 value = copy_rtx (value);
11376 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11377 reg_stat[regno].last_set_label, 1))
11378 return value;
11380 return 0;
11383 /* Return nonzero if expression X refers to a REG or to memory
11384 that is set in an instruction more recent than FROM_CUID. */
11386 static int
11387 use_crosses_set_p (rtx x, int from_cuid)
11389 const char *fmt;
11390 int i;
11391 enum rtx_code code = GET_CODE (x);
11393 if (code == REG)
11395 unsigned int regno = REGNO (x);
11396 unsigned endreg = regno + (regno < FIRST_PSEUDO_REGISTER
11397 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11399 #ifdef PUSH_ROUNDING
11400 /* Don't allow uses of the stack pointer to be moved,
11401 because we don't know whether the move crosses a push insn. */
11402 if (regno == STACK_POINTER_REGNUM && PUSH_ARGS)
11403 return 1;
11404 #endif
11405 for (; regno < endreg; regno++)
11406 if (reg_stat[regno].last_set
11407 && INSN_CUID (reg_stat[regno].last_set) > from_cuid)
11408 return 1;
11409 return 0;
11412 if (code == MEM && mem_last_set > from_cuid)
11413 return 1;
11415 fmt = GET_RTX_FORMAT (code);
11417 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11419 if (fmt[i] == 'E')
11421 int j;
11422 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11423 if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid))
11424 return 1;
11426 else if (fmt[i] == 'e'
11427 && use_crosses_set_p (XEXP (x, i), from_cuid))
11428 return 1;
11430 return 0;
11433 /* Define three variables used for communication between the following
11434 routines. */
11436 static unsigned int reg_dead_regno, reg_dead_endregno;
11437 static int reg_dead_flag;
11439 /* Function called via note_stores from reg_dead_at_p.
11441 If DEST is within [reg_dead_regno, reg_dead_endregno), set
11442 reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */
11444 static void
11445 reg_dead_at_p_1 (rtx dest, rtx x, void *data ATTRIBUTE_UNUSED)
11447 unsigned int regno, endregno;
11449 if (!REG_P (dest))
11450 return;
11452 regno = REGNO (dest);
11453 endregno = regno + (regno < FIRST_PSEUDO_REGISTER
11454 ? hard_regno_nregs[regno][GET_MODE (dest)] : 1);
11456 if (reg_dead_endregno > regno && reg_dead_regno < endregno)
11457 reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1;
11460 /* Return nonzero if REG is known to be dead at INSN.
11462 We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER
11463 referencing REG, it is dead. If we hit a SET referencing REG, it is
11464 live. Otherwise, see if it is live or dead at the start of the basic
11465 block we are in. Hard regs marked as being live in NEWPAT_USED_REGS
11466 must be assumed to be always live. */
11468 static int
11469 reg_dead_at_p (rtx reg, rtx insn)
11471 basic_block block;
11472 unsigned int i;
11474 /* Set variables for reg_dead_at_p_1. */
11475 reg_dead_regno = REGNO (reg);
11476 reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER
11477 ? hard_regno_nregs[reg_dead_regno]
11478 [GET_MODE (reg)]
11479 : 1);
11481 reg_dead_flag = 0;
11483 /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers
11484 we allow the machine description to decide whether use-and-clobber
11485 patterns are OK. */
11486 if (reg_dead_regno < FIRST_PSEUDO_REGISTER)
11488 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11489 if (!fixed_regs[i] && TEST_HARD_REG_BIT (newpat_used_regs, i))
11490 return 0;
11493 /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or
11494 beginning of function. */
11495 for (; insn && !LABEL_P (insn) && !BARRIER_P (insn);
11496 insn = prev_nonnote_insn (insn))
11498 note_stores (PATTERN (insn), reg_dead_at_p_1, NULL);
11499 if (reg_dead_flag)
11500 return reg_dead_flag == 1 ? 1 : 0;
11502 if (find_regno_note (insn, REG_DEAD, reg_dead_regno))
11503 return 1;
11506 /* Get the basic block that we were in. */
11507 if (insn == 0)
11508 block = ENTRY_BLOCK_PTR->next_bb;
11509 else
11511 FOR_EACH_BB (block)
11512 if (insn == BB_HEAD (block))
11513 break;
11515 if (block == EXIT_BLOCK_PTR)
11516 return 0;
11519 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11520 if (REGNO_REG_SET_P (block->il.rtl->global_live_at_start, i))
11521 return 0;
11523 return 1;
11526 /* Note hard registers in X that are used. This code is similar to
11527 that in flow.c, but much simpler since we don't care about pseudos. */
11529 static void
11530 mark_used_regs_combine (rtx x)
11532 RTX_CODE code = GET_CODE (x);
11533 unsigned int regno;
11534 int i;
11536 switch (code)
11538 case LABEL_REF:
11539 case SYMBOL_REF:
11540 case CONST_INT:
11541 case CONST:
11542 case CONST_DOUBLE:
11543 case CONST_VECTOR:
11544 case PC:
11545 case ADDR_VEC:
11546 case ADDR_DIFF_VEC:
11547 case ASM_INPUT:
11548 #ifdef HAVE_cc0
11549 /* CC0 must die in the insn after it is set, so we don't need to take
11550 special note of it here. */
11551 case CC0:
11552 #endif
11553 return;
11555 case CLOBBER:
11556 /* If we are clobbering a MEM, mark any hard registers inside the
11557 address as used. */
11558 if (MEM_P (XEXP (x, 0)))
11559 mark_used_regs_combine (XEXP (XEXP (x, 0), 0));
11560 return;
11562 case REG:
11563 regno = REGNO (x);
11564 /* A hard reg in a wide mode may really be multiple registers.
11565 If so, mark all of them just like the first. */
11566 if (regno < FIRST_PSEUDO_REGISTER)
11568 unsigned int endregno, r;
11570 /* None of this applies to the stack, frame or arg pointers. */
11571 if (regno == STACK_POINTER_REGNUM
11572 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
11573 || regno == HARD_FRAME_POINTER_REGNUM
11574 #endif
11575 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
11576 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
11577 #endif
11578 || regno == FRAME_POINTER_REGNUM)
11579 return;
11581 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11582 for (r = regno; r < endregno; r++)
11583 SET_HARD_REG_BIT (newpat_used_regs, r);
11585 return;
11587 case SET:
11589 /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in
11590 the address. */
11591 rtx testreg = SET_DEST (x);
11593 while (GET_CODE (testreg) == SUBREG
11594 || GET_CODE (testreg) == ZERO_EXTRACT
11595 || GET_CODE (testreg) == STRICT_LOW_PART)
11596 testreg = XEXP (testreg, 0);
11598 if (MEM_P (testreg))
11599 mark_used_regs_combine (XEXP (testreg, 0));
11601 mark_used_regs_combine (SET_SRC (x));
11603 return;
11605 default:
11606 break;
11609 /* Recursively scan the operands of this expression. */
11612 const char *fmt = GET_RTX_FORMAT (code);
11614 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11616 if (fmt[i] == 'e')
11617 mark_used_regs_combine (XEXP (x, i));
11618 else if (fmt[i] == 'E')
11620 int j;
11622 for (j = 0; j < XVECLEN (x, i); j++)
11623 mark_used_regs_combine (XVECEXP (x, i, j));
11629 /* Remove register number REGNO from the dead registers list of INSN.
11631 Return the note used to record the death, if there was one. */
11634 remove_death (unsigned int regno, rtx insn)
11636 rtx note = find_regno_note (insn, REG_DEAD, regno);
11638 if (note)
11640 REG_N_DEATHS (regno)--;
11641 remove_note (insn, note);
11644 return note;
11647 /* For each register (hardware or pseudo) used within expression X, if its
11648 death is in an instruction with cuid between FROM_CUID (inclusive) and
11649 TO_INSN (exclusive), put a REG_DEAD note for that register in the
11650 list headed by PNOTES.
11652 That said, don't move registers killed by maybe_kill_insn.
11654 This is done when X is being merged by combination into TO_INSN. These
11655 notes will then be distributed as needed. */
11657 static void
11658 move_deaths (rtx x, rtx maybe_kill_insn, int from_cuid, rtx to_insn,
11659 rtx *pnotes)
11661 const char *fmt;
11662 int len, i;
11663 enum rtx_code code = GET_CODE (x);
11665 if (code == REG)
11667 unsigned int regno = REGNO (x);
11668 rtx where_dead = reg_stat[regno].last_death;
11669 rtx before_dead, after_dead;
11671 /* Don't move the register if it gets killed in between from and to. */
11672 if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn)
11673 && ! reg_referenced_p (x, maybe_kill_insn))
11674 return;
11676 /* WHERE_DEAD could be a USE insn made by combine, so first we
11677 make sure that we have insns with valid INSN_CUID values. */
11678 before_dead = where_dead;
11679 while (before_dead && INSN_UID (before_dead) > max_uid_cuid)
11680 before_dead = PREV_INSN (before_dead);
11682 after_dead = where_dead;
11683 while (after_dead && INSN_UID (after_dead) > max_uid_cuid)
11684 after_dead = NEXT_INSN (after_dead);
11686 if (before_dead && after_dead
11687 && INSN_CUID (before_dead) >= from_cuid
11688 && (INSN_CUID (after_dead) < INSN_CUID (to_insn)
11689 || (where_dead != after_dead
11690 && INSN_CUID (after_dead) == INSN_CUID (to_insn))))
11692 rtx note = remove_death (regno, where_dead);
11694 /* It is possible for the call above to return 0. This can occur
11695 when last_death points to I2 or I1 that we combined with.
11696 In that case make a new note.
11698 We must also check for the case where X is a hard register
11699 and NOTE is a death note for a range of hard registers
11700 including X. In that case, we must put REG_DEAD notes for
11701 the remaining registers in place of NOTE. */
11703 if (note != 0 && regno < FIRST_PSEUDO_REGISTER
11704 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11705 > GET_MODE_SIZE (GET_MODE (x))))
11707 unsigned int deadregno = REGNO (XEXP (note, 0));
11708 unsigned int deadend
11709 = (deadregno + hard_regno_nregs[deadregno]
11710 [GET_MODE (XEXP (note, 0))]);
11711 unsigned int ourend
11712 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11713 unsigned int i;
11715 for (i = deadregno; i < deadend; i++)
11716 if (i < regno || i >= ourend)
11717 REG_NOTES (where_dead)
11718 = gen_rtx_EXPR_LIST (REG_DEAD,
11719 regno_reg_rtx[i],
11720 REG_NOTES (where_dead));
11723 /* If we didn't find any note, or if we found a REG_DEAD note that
11724 covers only part of the given reg, and we have a multi-reg hard
11725 register, then to be safe we must check for REG_DEAD notes
11726 for each register other than the first. They could have
11727 their own REG_DEAD notes lying around. */
11728 else if ((note == 0
11729 || (note != 0
11730 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11731 < GET_MODE_SIZE (GET_MODE (x)))))
11732 && regno < FIRST_PSEUDO_REGISTER
11733 && hard_regno_nregs[regno][GET_MODE (x)] > 1)
11735 unsigned int ourend
11736 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11737 unsigned int i, offset;
11738 rtx oldnotes = 0;
11740 if (note)
11741 offset = hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))];
11742 else
11743 offset = 1;
11745 for (i = regno + offset; i < ourend; i++)
11746 move_deaths (regno_reg_rtx[i],
11747 maybe_kill_insn, from_cuid, to_insn, &oldnotes);
11750 if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x))
11752 XEXP (note, 1) = *pnotes;
11753 *pnotes = note;
11755 else
11756 *pnotes = gen_rtx_EXPR_LIST (REG_DEAD, x, *pnotes);
11758 REG_N_DEATHS (regno)++;
11761 return;
11764 else if (GET_CODE (x) == SET)
11766 rtx dest = SET_DEST (x);
11768 move_deaths (SET_SRC (x), maybe_kill_insn, from_cuid, to_insn, pnotes);
11770 /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG
11771 that accesses one word of a multi-word item, some
11772 piece of everything register in the expression is used by
11773 this insn, so remove any old death. */
11774 /* ??? So why do we test for equality of the sizes? */
11776 if (GET_CODE (dest) == ZERO_EXTRACT
11777 || GET_CODE (dest) == STRICT_LOW_PART
11778 || (GET_CODE (dest) == SUBREG
11779 && (((GET_MODE_SIZE (GET_MODE (dest))
11780 + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
11781 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))
11782 + UNITS_PER_WORD - 1) / UNITS_PER_WORD))))
11784 move_deaths (dest, maybe_kill_insn, from_cuid, to_insn, pnotes);
11785 return;
11788 /* If this is some other SUBREG, we know it replaces the entire
11789 value, so use that as the destination. */
11790 if (GET_CODE (dest) == SUBREG)
11791 dest = SUBREG_REG (dest);
11793 /* If this is a MEM, adjust deaths of anything used in the address.
11794 For a REG (the only other possibility), the entire value is
11795 being replaced so the old value is not used in this insn. */
11797 if (MEM_P (dest))
11798 move_deaths (XEXP (dest, 0), maybe_kill_insn, from_cuid,
11799 to_insn, pnotes);
11800 return;
11803 else if (GET_CODE (x) == CLOBBER)
11804 return;
11806 len = GET_RTX_LENGTH (code);
11807 fmt = GET_RTX_FORMAT (code);
11809 for (i = 0; i < len; i++)
11811 if (fmt[i] == 'E')
11813 int j;
11814 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11815 move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_cuid,
11816 to_insn, pnotes);
11818 else if (fmt[i] == 'e')
11819 move_deaths (XEXP (x, i), maybe_kill_insn, from_cuid, to_insn, pnotes);
11823 /* Return 1 if X is the target of a bit-field assignment in BODY, the
11824 pattern of an insn. X must be a REG. */
11826 static int
11827 reg_bitfield_target_p (rtx x, rtx body)
11829 int i;
11831 if (GET_CODE (body) == SET)
11833 rtx dest = SET_DEST (body);
11834 rtx target;
11835 unsigned int regno, tregno, endregno, endtregno;
11837 if (GET_CODE (dest) == ZERO_EXTRACT)
11838 target = XEXP (dest, 0);
11839 else if (GET_CODE (dest) == STRICT_LOW_PART)
11840 target = SUBREG_REG (XEXP (dest, 0));
11841 else
11842 return 0;
11844 if (GET_CODE (target) == SUBREG)
11845 target = SUBREG_REG (target);
11847 if (!REG_P (target))
11848 return 0;
11850 tregno = REGNO (target), regno = REGNO (x);
11851 if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER)
11852 return target == x;
11854 endtregno = tregno + hard_regno_nregs[tregno][GET_MODE (target)];
11855 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11857 return endregno > tregno && regno < endtregno;
11860 else if (GET_CODE (body) == PARALLEL)
11861 for (i = XVECLEN (body, 0) - 1; i >= 0; i--)
11862 if (reg_bitfield_target_p (x, XVECEXP (body, 0, i)))
11863 return 1;
11865 return 0;
11868 /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them
11869 as appropriate. I3 and I2 are the insns resulting from the combination
11870 insns including FROM (I2 may be zero).
11872 ELIM_I2 and ELIM_I1 are either zero or registers that we know will
11873 not need REG_DEAD notes because they are being substituted for. This
11874 saves searching in the most common cases.
11876 Each note in the list is either ignored or placed on some insns, depending
11877 on the type of note. */
11879 static void
11880 distribute_notes (rtx notes, rtx from_insn, rtx i3, rtx i2, rtx elim_i2,
11881 rtx elim_i1)
11883 rtx note, next_note;
11884 rtx tem;
11886 for (note = notes; note; note = next_note)
11888 rtx place = 0, place2 = 0;
11890 next_note = XEXP (note, 1);
11891 switch (REG_NOTE_KIND (note))
11893 case REG_BR_PROB:
11894 case REG_BR_PRED:
11895 /* Doesn't matter much where we put this, as long as it's somewhere.
11896 It is preferable to keep these notes on branches, which is most
11897 likely to be i3. */
11898 place = i3;
11899 break;
11901 case REG_VALUE_PROFILE:
11902 /* Just get rid of this note, as it is unused later anyway. */
11903 break;
11905 case REG_NON_LOCAL_GOTO:
11906 if (JUMP_P (i3))
11907 place = i3;
11908 else
11910 gcc_assert (i2 && JUMP_P (i2));
11911 place = i2;
11913 break;
11915 case REG_EH_REGION:
11916 /* These notes must remain with the call or trapping instruction. */
11917 if (CALL_P (i3))
11918 place = i3;
11919 else if (i2 && CALL_P (i2))
11920 place = i2;
11921 else
11923 gcc_assert (flag_non_call_exceptions);
11924 if (may_trap_p (i3))
11925 place = i3;
11926 else if (i2 && may_trap_p (i2))
11927 place = i2;
11928 /* ??? Otherwise assume we've combined things such that we
11929 can now prove that the instructions can't trap. Drop the
11930 note in this case. */
11932 break;
11934 case REG_NORETURN:
11935 case REG_SETJMP:
11936 /* These notes must remain with the call. It should not be
11937 possible for both I2 and I3 to be a call. */
11938 if (CALL_P (i3))
11939 place = i3;
11940 else
11942 gcc_assert (i2 && CALL_P (i2));
11943 place = i2;
11945 break;
11947 case REG_UNUSED:
11948 /* Any clobbers for i3 may still exist, and so we must process
11949 REG_UNUSED notes from that insn.
11951 Any clobbers from i2 or i1 can only exist if they were added by
11952 recog_for_combine. In that case, recog_for_combine created the
11953 necessary REG_UNUSED notes. Trying to keep any original
11954 REG_UNUSED notes from these insns can cause incorrect output
11955 if it is for the same register as the original i3 dest.
11956 In that case, we will notice that the register is set in i3,
11957 and then add a REG_UNUSED note for the destination of i3, which
11958 is wrong. However, it is possible to have REG_UNUSED notes from
11959 i2 or i1 for register which were both used and clobbered, so
11960 we keep notes from i2 or i1 if they will turn into REG_DEAD
11961 notes. */
11963 /* If this register is set or clobbered in I3, put the note there
11964 unless there is one already. */
11965 if (reg_set_p (XEXP (note, 0), PATTERN (i3)))
11967 if (from_insn != i3)
11968 break;
11970 if (! (REG_P (XEXP (note, 0))
11971 ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0)))
11972 : find_reg_note (i3, REG_UNUSED, XEXP (note, 0))))
11973 place = i3;
11975 /* Otherwise, if this register is used by I3, then this register
11976 now dies here, so we must put a REG_DEAD note here unless there
11977 is one already. */
11978 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))
11979 && ! (REG_P (XEXP (note, 0))
11980 ? find_regno_note (i3, REG_DEAD,
11981 REGNO (XEXP (note, 0)))
11982 : find_reg_note (i3, REG_DEAD, XEXP (note, 0))))
11984 PUT_REG_NOTE_KIND (note, REG_DEAD);
11985 place = i3;
11987 break;
11989 case REG_EQUAL:
11990 case REG_EQUIV:
11991 case REG_NOALIAS:
11992 /* These notes say something about results of an insn. We can
11993 only support them if they used to be on I3 in which case they
11994 remain on I3. Otherwise they are ignored.
11996 If the note refers to an expression that is not a constant, we
11997 must also ignore the note since we cannot tell whether the
11998 equivalence is still true. It might be possible to do
11999 slightly better than this (we only have a problem if I2DEST
12000 or I1DEST is present in the expression), but it doesn't
12001 seem worth the trouble. */
12003 if (from_insn == i3
12004 && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0))))
12005 place = i3;
12006 break;
12008 case REG_INC:
12009 case REG_NO_CONFLICT:
12010 /* These notes say something about how a register is used. They must
12011 be present on any use of the register in I2 or I3. */
12012 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3)))
12013 place = i3;
12015 if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2)))
12017 if (place)
12018 place2 = i2;
12019 else
12020 place = i2;
12022 break;
12024 case REG_LABEL:
12025 /* This can show up in several ways -- either directly in the
12026 pattern, or hidden off in the constant pool with (or without?)
12027 a REG_EQUAL note. */
12028 /* ??? Ignore the without-reg_equal-note problem for now. */
12029 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))
12030 || ((tem = find_reg_note (i3, REG_EQUAL, NULL_RTX))
12031 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
12032 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0)))
12033 place = i3;
12035 if (i2
12036 && (reg_mentioned_p (XEXP (note, 0), PATTERN (i2))
12037 || ((tem = find_reg_note (i2, REG_EQUAL, NULL_RTX))
12038 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
12039 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0))))
12041 if (place)
12042 place2 = i2;
12043 else
12044 place = i2;
12047 /* Don't attach REG_LABEL note to a JUMP_INSN. Add
12048 a JUMP_LABEL instead or decrement LABEL_NUSES. */
12049 if (place && JUMP_P (place))
12051 rtx label = JUMP_LABEL (place);
12053 if (!label)
12054 JUMP_LABEL (place) = XEXP (note, 0);
12055 else
12057 gcc_assert (label == XEXP (note, 0));
12058 if (LABEL_P (label))
12059 LABEL_NUSES (label)--;
12061 place = 0;
12063 if (place2 && JUMP_P (place2))
12065 rtx label = JUMP_LABEL (place2);
12067 if (!label)
12068 JUMP_LABEL (place2) = XEXP (note, 0);
12069 else
12071 gcc_assert (label == XEXP (note, 0));
12072 if (LABEL_P (label))
12073 LABEL_NUSES (label)--;
12075 place2 = 0;
12077 break;
12079 case REG_NONNEG:
12080 /* This note says something about the value of a register prior
12081 to the execution of an insn. It is too much trouble to see
12082 if the note is still correct in all situations. It is better
12083 to simply delete it. */
12084 break;
12086 case REG_RETVAL:
12087 /* If the insn previously containing this note still exists,
12088 put it back where it was. Otherwise move it to the previous
12089 insn. Adjust the corresponding REG_LIBCALL note. */
12090 if (!NOTE_P (from_insn))
12091 place = from_insn;
12092 else
12094 tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX);
12095 place = prev_real_insn (from_insn);
12096 if (tem && place)
12097 XEXP (tem, 0) = place;
12098 /* If we're deleting the last remaining instruction of a
12099 libcall sequence, don't add the notes. */
12100 else if (XEXP (note, 0) == from_insn)
12101 tem = place = 0;
12102 /* Don't add the dangling REG_RETVAL note. */
12103 else if (! tem)
12104 place = 0;
12106 break;
12108 case REG_LIBCALL:
12109 /* This is handled similarly to REG_RETVAL. */
12110 if (!NOTE_P (from_insn))
12111 place = from_insn;
12112 else
12114 tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX);
12115 place = next_real_insn (from_insn);
12116 if (tem && place)
12117 XEXP (tem, 0) = place;
12118 /* If we're deleting the last remaining instruction of a
12119 libcall sequence, don't add the notes. */
12120 else if (XEXP (note, 0) == from_insn)
12121 tem = place = 0;
12122 /* Don't add the dangling REG_LIBCALL note. */
12123 else if (! tem)
12124 place = 0;
12126 break;
12128 case REG_DEAD:
12129 /* If we replaced the right hand side of FROM_INSN with a
12130 REG_EQUAL note, the original use of the dying register
12131 will not have been combined into I3 and I2. In such cases,
12132 FROM_INSN is guaranteed to be the first of the combined
12133 instructions, so we simply need to search back before
12134 FROM_INSN for the previous use or set of this register,
12135 then alter the notes there appropriately.
12137 If the register is used as an input in I3, it dies there.
12138 Similarly for I2, if it is nonzero and adjacent to I3.
12140 If the register is not used as an input in either I3 or I2
12141 and it is not one of the registers we were supposed to eliminate,
12142 there are two possibilities. We might have a non-adjacent I2
12143 or we might have somehow eliminated an additional register
12144 from a computation. For example, we might have had A & B where
12145 we discover that B will always be zero. In this case we will
12146 eliminate the reference to A.
12148 In both cases, we must search to see if we can find a previous
12149 use of A and put the death note there. */
12151 if (from_insn
12152 && from_insn == i2mod
12153 && !reg_overlap_mentioned_p (XEXP (note, 0), i2mod_new_rhs))
12154 tem = from_insn;
12155 else
12157 if (from_insn
12158 && CALL_P (from_insn)
12159 && find_reg_fusage (from_insn, USE, XEXP (note, 0)))
12160 place = from_insn;
12161 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)))
12162 place = i3;
12163 else if (i2 != 0 && next_nonnote_insn (i2) == i3
12164 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12165 place = i2;
12166 else if ((rtx_equal_p (XEXP (note, 0), elim_i2)
12167 && !(i2mod
12168 && reg_overlap_mentioned_p (XEXP (note, 0),
12169 i2mod_old_rhs)))
12170 || rtx_equal_p (XEXP (note, 0), elim_i1))
12171 break;
12172 tem = i3;
12175 if (place == 0)
12177 basic_block bb = this_basic_block;
12179 for (tem = PREV_INSN (tem); place == 0; tem = PREV_INSN (tem))
12181 if (! INSN_P (tem))
12183 if (tem == BB_HEAD (bb))
12184 break;
12185 continue;
12188 /* If the register is being set at TEM, see if that is all
12189 TEM is doing. If so, delete TEM. Otherwise, make this
12190 into a REG_UNUSED note instead. Don't delete sets to
12191 global register vars. */
12192 if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER
12193 || !global_regs[REGNO (XEXP (note, 0))])
12194 && reg_set_p (XEXP (note, 0), PATTERN (tem)))
12196 rtx set = single_set (tem);
12197 rtx inner_dest = 0;
12198 #ifdef HAVE_cc0
12199 rtx cc0_setter = NULL_RTX;
12200 #endif
12202 if (set != 0)
12203 for (inner_dest = SET_DEST (set);
12204 (GET_CODE (inner_dest) == STRICT_LOW_PART
12205 || GET_CODE (inner_dest) == SUBREG
12206 || GET_CODE (inner_dest) == ZERO_EXTRACT);
12207 inner_dest = XEXP (inner_dest, 0))
12210 /* Verify that it was the set, and not a clobber that
12211 modified the register.
12213 CC0 targets must be careful to maintain setter/user
12214 pairs. If we cannot delete the setter due to side
12215 effects, mark the user with an UNUSED note instead
12216 of deleting it. */
12218 if (set != 0 && ! side_effects_p (SET_SRC (set))
12219 && rtx_equal_p (XEXP (note, 0), inner_dest)
12220 #ifdef HAVE_cc0
12221 && (! reg_mentioned_p (cc0_rtx, SET_SRC (set))
12222 || ((cc0_setter = prev_cc0_setter (tem)) != NULL
12223 && sets_cc0_p (PATTERN (cc0_setter)) > 0))
12224 #endif
12227 /* Move the notes and links of TEM elsewhere.
12228 This might delete other dead insns recursively.
12229 First set the pattern to something that won't use
12230 any register. */
12231 rtx old_notes = REG_NOTES (tem);
12233 PATTERN (tem) = pc_rtx;
12234 REG_NOTES (tem) = NULL;
12236 distribute_notes (old_notes, tem, tem, NULL_RTX,
12237 NULL_RTX, NULL_RTX);
12238 distribute_links (LOG_LINKS (tem));
12240 SET_INSN_DELETED (tem);
12242 #ifdef HAVE_cc0
12243 /* Delete the setter too. */
12244 if (cc0_setter)
12246 PATTERN (cc0_setter) = pc_rtx;
12247 old_notes = REG_NOTES (cc0_setter);
12248 REG_NOTES (cc0_setter) = NULL;
12250 distribute_notes (old_notes, cc0_setter,
12251 cc0_setter, NULL_RTX,
12252 NULL_RTX, NULL_RTX);
12253 distribute_links (LOG_LINKS (cc0_setter));
12255 SET_INSN_DELETED (cc0_setter);
12257 #endif
12259 else
12261 PUT_REG_NOTE_KIND (note, REG_UNUSED);
12263 /* If there isn't already a REG_UNUSED note, put one
12264 here. Do not place a REG_DEAD note, even if
12265 the register is also used here; that would not
12266 match the algorithm used in lifetime analysis
12267 and can cause the consistency check in the
12268 scheduler to fail. */
12269 if (! find_regno_note (tem, REG_UNUSED,
12270 REGNO (XEXP (note, 0))))
12271 place = tem;
12272 break;
12275 else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))
12276 || (CALL_P (tem)
12277 && find_reg_fusage (tem, USE, XEXP (note, 0))))
12279 place = tem;
12281 /* If we are doing a 3->2 combination, and we have a
12282 register which formerly died in i3 and was not used
12283 by i2, which now no longer dies in i3 and is used in
12284 i2 but does not die in i2, and place is between i2
12285 and i3, then we may need to move a link from place to
12286 i2. */
12287 if (i2 && INSN_UID (place) <= max_uid_cuid
12288 && INSN_CUID (place) > INSN_CUID (i2)
12289 && from_insn
12290 && INSN_CUID (from_insn) > INSN_CUID (i2)
12291 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12293 rtx links = LOG_LINKS (place);
12294 LOG_LINKS (place) = 0;
12295 distribute_links (links);
12297 break;
12300 if (tem == BB_HEAD (bb))
12301 break;
12304 /* We haven't found an insn for the death note and it
12305 is still a REG_DEAD note, but we have hit the beginning
12306 of the block. If the existing life info says the reg
12307 was dead, there's nothing left to do. Otherwise, we'll
12308 need to do a global life update after combine. */
12309 if (REG_NOTE_KIND (note) == REG_DEAD && place == 0
12310 && REGNO_REG_SET_P (bb->il.rtl->global_live_at_start,
12311 REGNO (XEXP (note, 0))))
12312 SET_BIT (refresh_blocks, this_basic_block->index);
12315 /* If the register is set or already dead at PLACE, we needn't do
12316 anything with this note if it is still a REG_DEAD note.
12317 We check here if it is set at all, not if is it totally replaced,
12318 which is what `dead_or_set_p' checks, so also check for it being
12319 set partially. */
12321 if (place && REG_NOTE_KIND (note) == REG_DEAD)
12323 unsigned int regno = REGNO (XEXP (note, 0));
12325 /* Similarly, if the instruction on which we want to place
12326 the note is a noop, we'll need do a global live update
12327 after we remove them in delete_noop_moves. */
12328 if (noop_move_p (place))
12329 SET_BIT (refresh_blocks, this_basic_block->index);
12331 if (dead_or_set_p (place, XEXP (note, 0))
12332 || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place)))
12334 /* Unless the register previously died in PLACE, clear
12335 last_death. [I no longer understand why this is
12336 being done.] */
12337 if (reg_stat[regno].last_death != place)
12338 reg_stat[regno].last_death = 0;
12339 place = 0;
12341 else
12342 reg_stat[regno].last_death = place;
12344 /* If this is a death note for a hard reg that is occupying
12345 multiple registers, ensure that we are still using all
12346 parts of the object. If we find a piece of the object
12347 that is unused, we must arrange for an appropriate REG_DEAD
12348 note to be added for it. However, we can't just emit a USE
12349 and tag the note to it, since the register might actually
12350 be dead; so we recourse, and the recursive call then finds
12351 the previous insn that used this register. */
12353 if (place && regno < FIRST_PSEUDO_REGISTER
12354 && hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))] > 1)
12356 unsigned int endregno
12357 = regno + hard_regno_nregs[regno]
12358 [GET_MODE (XEXP (note, 0))];
12359 int all_used = 1;
12360 unsigned int i;
12362 for (i = regno; i < endregno; i++)
12363 if ((! refers_to_regno_p (i, i + 1, PATTERN (place), 0)
12364 && ! find_regno_fusage (place, USE, i))
12365 || dead_or_set_regno_p (place, i))
12366 all_used = 0;
12368 if (! all_used)
12370 /* Put only REG_DEAD notes for pieces that are
12371 not already dead or set. */
12373 for (i = regno; i < endregno;
12374 i += hard_regno_nregs[i][reg_raw_mode[i]])
12376 rtx piece = regno_reg_rtx[i];
12377 basic_block bb = this_basic_block;
12379 if (! dead_or_set_p (place, piece)
12380 && ! reg_bitfield_target_p (piece,
12381 PATTERN (place)))
12383 rtx new_note
12384 = gen_rtx_EXPR_LIST (REG_DEAD, piece, NULL_RTX);
12386 distribute_notes (new_note, place, place,
12387 NULL_RTX, NULL_RTX, NULL_RTX);
12389 else if (! refers_to_regno_p (i, i + 1,
12390 PATTERN (place), 0)
12391 && ! find_regno_fusage (place, USE, i))
12392 for (tem = PREV_INSN (place); ;
12393 tem = PREV_INSN (tem))
12395 if (! INSN_P (tem))
12397 if (tem == BB_HEAD (bb))
12399 SET_BIT (refresh_blocks,
12400 this_basic_block->index);
12401 break;
12403 continue;
12405 if (dead_or_set_p (tem, piece)
12406 || reg_bitfield_target_p (piece,
12407 PATTERN (tem)))
12409 REG_NOTES (tem)
12410 = gen_rtx_EXPR_LIST (REG_UNUSED, piece,
12411 REG_NOTES (tem));
12412 break;
12418 place = 0;
12422 break;
12424 default:
12425 /* Any other notes should not be present at this point in the
12426 compilation. */
12427 gcc_unreachable ();
12430 if (place)
12432 XEXP (note, 1) = REG_NOTES (place);
12433 REG_NOTES (place) = note;
12435 else if ((REG_NOTE_KIND (note) == REG_DEAD
12436 || REG_NOTE_KIND (note) == REG_UNUSED)
12437 && REG_P (XEXP (note, 0)))
12438 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
12440 if (place2)
12442 if ((REG_NOTE_KIND (note) == REG_DEAD
12443 || REG_NOTE_KIND (note) == REG_UNUSED)
12444 && REG_P (XEXP (note, 0)))
12445 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
12447 REG_NOTES (place2) = gen_rtx_fmt_ee (GET_CODE (note),
12448 REG_NOTE_KIND (note),
12449 XEXP (note, 0),
12450 REG_NOTES (place2));
12455 /* Similarly to above, distribute the LOG_LINKS that used to be present on
12456 I3, I2, and I1 to new locations. This is also called to add a link
12457 pointing at I3 when I3's destination is changed. */
12459 static void
12460 distribute_links (rtx links)
12462 rtx link, next_link;
12464 for (link = links; link; link = next_link)
12466 rtx place = 0;
12467 rtx insn;
12468 rtx set, reg;
12470 next_link = XEXP (link, 1);
12472 /* If the insn that this link points to is a NOTE or isn't a single
12473 set, ignore it. In the latter case, it isn't clear what we
12474 can do other than ignore the link, since we can't tell which
12475 register it was for. Such links wouldn't be used by combine
12476 anyway.
12478 It is not possible for the destination of the target of the link to
12479 have been changed by combine. The only potential of this is if we
12480 replace I3, I2, and I1 by I3 and I2. But in that case the
12481 destination of I2 also remains unchanged. */
12483 if (NOTE_P (XEXP (link, 0))
12484 || (set = single_set (XEXP (link, 0))) == 0)
12485 continue;
12487 reg = SET_DEST (set);
12488 while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
12489 || GET_CODE (reg) == STRICT_LOW_PART)
12490 reg = XEXP (reg, 0);
12492 /* A LOG_LINK is defined as being placed on the first insn that uses
12493 a register and points to the insn that sets the register. Start
12494 searching at the next insn after the target of the link and stop
12495 when we reach a set of the register or the end of the basic block.
12497 Note that this correctly handles the link that used to point from
12498 I3 to I2. Also note that not much searching is typically done here
12499 since most links don't point very far away. */
12501 for (insn = NEXT_INSN (XEXP (link, 0));
12502 (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
12503 || BB_HEAD (this_basic_block->next_bb) != insn));
12504 insn = NEXT_INSN (insn))
12505 if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
12507 if (reg_referenced_p (reg, PATTERN (insn)))
12508 place = insn;
12509 break;
12511 else if (CALL_P (insn)
12512 && find_reg_fusage (insn, USE, reg))
12514 place = insn;
12515 break;
12517 else if (INSN_P (insn) && reg_set_p (reg, insn))
12518 break;
12520 /* If we found a place to put the link, place it there unless there
12521 is already a link to the same insn as LINK at that point. */
12523 if (place)
12525 rtx link2;
12527 for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1))
12528 if (XEXP (link2, 0) == XEXP (link, 0))
12529 break;
12531 if (link2 == 0)
12533 XEXP (link, 1) = LOG_LINKS (place);
12534 LOG_LINKS (place) = link;
12536 /* Set added_links_insn to the earliest insn we added a
12537 link to. */
12538 if (added_links_insn == 0
12539 || INSN_CUID (added_links_insn) > INSN_CUID (place))
12540 added_links_insn = place;
12546 /* Subroutine of unmentioned_reg_p and callback from for_each_rtx.
12547 Check whether the expression pointer to by LOC is a register or
12548 memory, and if so return 1 if it isn't mentioned in the rtx EXPR.
12549 Otherwise return zero. */
12551 static int
12552 unmentioned_reg_p_1 (rtx *loc, void *expr)
12554 rtx x = *loc;
12556 if (x != NULL_RTX
12557 && (REG_P (x) || MEM_P (x))
12558 && ! reg_mentioned_p (x, (rtx) expr))
12559 return 1;
12560 return 0;
12563 /* Check for any register or memory mentioned in EQUIV that is not
12564 mentioned in EXPR. This is used to restrict EQUIV to "specializations"
12565 of EXPR where some registers may have been replaced by constants. */
12567 static bool
12568 unmentioned_reg_p (rtx equiv, rtx expr)
12570 return for_each_rtx (&equiv, unmentioned_reg_p_1, expr);
12573 /* Compute INSN_CUID for INSN, which is an insn made by combine. */
12575 static int
12576 insn_cuid (rtx insn)
12578 while (insn != 0 && INSN_UID (insn) > max_uid_cuid
12579 && NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE)
12580 insn = NEXT_INSN (insn);
12582 gcc_assert (INSN_UID (insn) <= max_uid_cuid);
12584 return INSN_CUID (insn);
12587 void
12588 dump_combine_stats (FILE *file)
12590 fprintf
12591 (file,
12592 ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n",
12593 combine_attempts, combine_merges, combine_extras, combine_successes);
12596 void
12597 dump_combine_total_stats (FILE *file)
12599 fprintf
12600 (file,
12601 "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n",
12602 total_attempts, total_merges, total_extras, total_successes);
12606 static bool
12607 gate_handle_combine (void)
12609 return (optimize > 0);
12612 /* Try combining insns through substitution. */
12613 static unsigned int
12614 rest_of_handle_combine (void)
12616 int rebuild_jump_labels_after_combine
12617 = combine_instructions (get_insns (), max_reg_num ());
12619 /* Combining insns may have turned an indirect jump into a
12620 direct jump. Rebuild the JUMP_LABEL fields of jumping
12621 instructions. */
12622 if (rebuild_jump_labels_after_combine)
12624 timevar_push (TV_JUMP);
12625 rebuild_jump_labels (get_insns ());
12626 timevar_pop (TV_JUMP);
12628 delete_dead_jumptables ();
12629 cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_UPDATE_LIFE);
12631 return 0;
12634 struct tree_opt_pass pass_combine =
12636 "combine", /* name */
12637 gate_handle_combine, /* gate */
12638 rest_of_handle_combine, /* execute */
12639 NULL, /* sub */
12640 NULL, /* next */
12641 0, /* static_pass_number */
12642 TV_COMBINE, /* tv_id */
12643 0, /* properties_required */
12644 0, /* properties_provided */
12645 0, /* properties_destroyed */
12646 0, /* todo_flags_start */
12647 TODO_dump_func |
12648 TODO_ggc_collect, /* todo_flags_finish */
12649 'c' /* letter */