* arm.c (adjacent_mem_locations): Reject volatile memory refs.
[official-gcc.git] / gcc / combine.c
blob297b58e914e050d132db5b3a12fb6d0118eb8475
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 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, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, 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 - a LOG_LINKS entry that refers to an insn with multiple SETs may be
57 removed because there is no way to know which register it was
58 linking
60 To simplify substitution, we combine only when the earlier insn(s)
61 consist of only a single assignment. To simplify updating afterward,
62 we never combine when a subroutine call appears in the middle.
64 Since we do not represent assignments to CC0 explicitly except when that
65 is all an insn does, there is no LOG_LINKS entry in an insn that uses
66 the condition code for the insn that set the condition code.
67 Fortunately, these two insns must be consecutive.
68 Therefore, every JUMP_INSN is taken to have an implicit logical link
69 to the preceding insn. This is not quite right, since non-jumps can
70 also use the condition code; but in practice such insns would not
71 combine anyway. */
73 #include "config.h"
74 #include "system.h"
75 #include "coretypes.h"
76 #include "tm.h"
77 #include "rtl.h"
78 #include "tree.h"
79 #include "tm_p.h"
80 #include "flags.h"
81 #include "regs.h"
82 #include "hard-reg-set.h"
83 #include "basic-block.h"
84 #include "insn-config.h"
85 #include "function.h"
86 /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */
87 #include "expr.h"
88 #include "insn-attr.h"
89 #include "recog.h"
90 #include "real.h"
91 #include "toplev.h"
92 #include "target.h"
93 #include "optabs.h"
94 #include "insn-codes.h"
95 #include "rtlhooks-def.h"
96 /* Include output.h for dump_file. */
97 #include "output.h"
98 #include "params.h"
100 /* Number of attempts to combine instructions in this function. */
102 static int combine_attempts;
104 /* Number of attempts that got as far as substitution in this function. */
106 static int combine_merges;
108 /* Number of instructions combined with added SETs in this function. */
110 static int combine_extras;
112 /* Number of instructions combined in this function. */
114 static int combine_successes;
116 /* Totals over entire compilation. */
118 static int total_attempts, total_merges, total_extras, total_successes;
121 /* Vector mapping INSN_UIDs to cuids.
122 The cuids are like uids but increase monotonically always.
123 Combine always uses cuids so that it can compare them.
124 But actually renumbering the uids, which we used to do,
125 proves to be a bad idea because it makes it hard to compare
126 the dumps produced by earlier passes with those from later passes. */
128 static int *uid_cuid;
129 static int max_uid_cuid;
131 /* Get the cuid of an insn. */
133 #define INSN_CUID(INSN) \
134 (INSN_UID (INSN) > max_uid_cuid ? insn_cuid (INSN) : uid_cuid[INSN_UID (INSN)])
136 /* In case BITS_PER_WORD == HOST_BITS_PER_WIDE_INT, shifting by
137 BITS_PER_WORD would invoke undefined behavior. Work around it. */
139 #define UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD(val) \
140 (((unsigned HOST_WIDE_INT) (val) << (BITS_PER_WORD - 1)) << 1)
142 /* Maximum register number, which is the size of the tables below. */
144 static unsigned int combine_max_regno;
146 struct reg_stat {
147 /* Record last point of death of (hard or pseudo) register n. */
148 rtx last_death;
150 /* Record last point of modification of (hard or pseudo) register n. */
151 rtx last_set;
153 /* The next group of fields allows the recording of the last value assigned
154 to (hard or pseudo) register n. We use this information to see if an
155 operation being processed is redundant given a prior operation performed
156 on the register. For example, an `and' with a constant is redundant if
157 all the zero bits are already known to be turned off.
159 We use an approach similar to that used by cse, but change it in the
160 following ways:
162 (1) We do not want to reinitialize at each label.
163 (2) It is useful, but not critical, to know the actual value assigned
164 to a register. Often just its form is helpful.
166 Therefore, we maintain the following fields:
168 last_set_value the last value assigned
169 last_set_label records the value of label_tick when the
170 register was assigned
171 last_set_table_tick records the value of label_tick when a
172 value using the register is assigned
173 last_set_invalid set to nonzero when it is not valid
174 to use the value of this register in some
175 register's value
177 To understand the usage of these tables, it is important to understand
178 the distinction between the value in last_set_value being valid and
179 the register being validly contained in some other expression in the
180 table.
182 (The next two parameters are out of date).
184 reg_stat[i].last_set_value is valid if it is nonzero, and either
185 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
187 Register I may validly appear in any expression returned for the value
188 of another register if reg_n_sets[i] is 1. It may also appear in the
189 value for register J if reg_stat[j].last_set_invalid is zero, or
190 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
192 If an expression is found in the table containing a register which may
193 not validly appear in an expression, the register is replaced by
194 something that won't match, (clobber (const_int 0)). */
196 /* Record last value assigned to (hard or pseudo) register n. */
198 rtx last_set_value;
200 /* Record the value of label_tick when an expression involving register n
201 is placed in last_set_value. */
203 int last_set_table_tick;
205 /* Record the value of label_tick when the value for register n is placed in
206 last_set_value. */
208 int last_set_label;
210 /* These fields are maintained in parallel with last_set_value and are
211 used to store the mode in which the register was last set, the bits
212 that were known to be zero when it was last set, and the number of
213 sign bits copies it was known to have when it was last set. */
215 unsigned HOST_WIDE_INT last_set_nonzero_bits;
216 char last_set_sign_bit_copies;
217 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
219 /* Set nonzero if references to register n in expressions should not be
220 used. last_set_invalid is set nonzero when this register is being
221 assigned to and last_set_table_tick == label_tick. */
223 char last_set_invalid;
225 /* Some registers that are set more than once and used in more than one
226 basic block are nevertheless always set in similar ways. For example,
227 a QImode register may be loaded from memory in two places on a machine
228 where byte loads zero extend.
230 We record in the following fields if a register has some leading bits
231 that are always equal to the sign bit, and what we know about the
232 nonzero bits of a register, specifically which bits are known to be
233 zero.
235 If an entry is zero, it means that we don't know anything special. */
237 unsigned char sign_bit_copies;
239 unsigned HOST_WIDE_INT nonzero_bits;
242 static struct reg_stat *reg_stat;
244 /* Record the cuid of the last insn that invalidated memory
245 (anything that writes memory, and subroutine calls, but not pushes). */
247 static int mem_last_set;
249 /* Record the cuid of the last CALL_INSN
250 so we can tell whether a potential combination crosses any calls. */
252 static int last_call_cuid;
254 /* When `subst' is called, this is the insn that is being modified
255 (by combining in a previous insn). The PATTERN of this insn
256 is still the old pattern partially modified and it should not be
257 looked at, but this may be used to examine the successors of the insn
258 to judge whether a simplification is valid. */
260 static rtx subst_insn;
262 /* This is the lowest CUID that `subst' is currently dealing with.
263 get_last_value will not return a value if the register was set at or
264 after this CUID. If not for this mechanism, we could get confused if
265 I2 or I1 in try_combine were an insn that used the old value of a register
266 to obtain a new value. In that case, we might erroneously get the
267 new value of the register when we wanted the old one. */
269 static int subst_low_cuid;
271 /* This contains any hard registers that are used in newpat; reg_dead_at_p
272 must consider all these registers to be always live. */
274 static HARD_REG_SET newpat_used_regs;
276 /* This is an insn to which a LOG_LINKS entry has been added. If this
277 insn is the earlier than I2 or I3, combine should rescan starting at
278 that location. */
280 static rtx added_links_insn;
282 /* Basic block in which we are performing combines. */
283 static basic_block this_basic_block;
285 /* A bitmap indicating which blocks had registers go dead at entry.
286 After combine, we'll need to re-do global life analysis with
287 those blocks as starting points. */
288 static sbitmap refresh_blocks;
290 /* The following array records the insn_rtx_cost for every insn
291 in the instruction stream. */
293 static int *uid_insn_cost;
295 /* Length of the currently allocated uid_insn_cost array. */
297 static int last_insn_cost;
299 /* Incremented for each label. */
301 static int label_tick;
303 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
304 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
306 static enum machine_mode nonzero_bits_mode;
308 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
309 be safely used. It is zero while computing them and after combine has
310 completed. This former test prevents propagating values based on
311 previously set values, which can be incorrect if a variable is modified
312 in a loop. */
314 static int nonzero_sign_valid;
317 /* Record one modification to rtl structure
318 to be undone by storing old_contents into *where.
319 is_int is 1 if the contents are an int. */
321 struct undo
323 struct undo *next;
324 int is_int;
325 union {rtx r; int i;} old_contents;
326 union {rtx *r; int *i;} where;
329 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
330 num_undo says how many are currently recorded.
332 other_insn is nonzero if we have modified some other insn in the process
333 of working on subst_insn. It must be verified too. */
335 struct undobuf
337 struct undo *undos;
338 struct undo *frees;
339 rtx other_insn;
342 static struct undobuf undobuf;
344 /* Number of times the pseudo being substituted for
345 was found and replaced. */
347 static int n_occurrences;
349 static rtx reg_nonzero_bits_for_combine (rtx, enum machine_mode, rtx,
350 enum machine_mode,
351 unsigned HOST_WIDE_INT,
352 unsigned HOST_WIDE_INT *);
353 static rtx reg_num_sign_bit_copies_for_combine (rtx, enum machine_mode, rtx,
354 enum machine_mode,
355 unsigned int, unsigned int *);
356 static void do_SUBST (rtx *, rtx);
357 static void do_SUBST_INT (int *, int);
358 static void init_reg_last (void);
359 static void setup_incoming_promotions (void);
360 static void set_nonzero_bits_and_sign_copies (rtx, rtx, void *);
361 static int cant_combine_insn_p (rtx);
362 static int can_combine_p (rtx, rtx, rtx, rtx, rtx *, rtx *);
363 static int combinable_i3pat (rtx, rtx *, rtx, rtx, int, rtx *);
364 static int contains_muldiv (rtx);
365 static rtx try_combine (rtx, rtx, rtx, int *);
366 static void undo_all (void);
367 static void undo_commit (void);
368 static rtx *find_split_point (rtx *, rtx);
369 static rtx subst (rtx, rtx, rtx, int, int);
370 static rtx combine_simplify_rtx (rtx, enum machine_mode, int);
371 static rtx simplify_if_then_else (rtx);
372 static rtx simplify_set (rtx);
373 static rtx simplify_logical (rtx);
374 static rtx expand_compound_operation (rtx);
375 static rtx expand_field_assignment (rtx);
376 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
377 rtx, unsigned HOST_WIDE_INT, int, int, int);
378 static rtx extract_left_shift (rtx, int);
379 static rtx make_compound_operation (rtx, enum rtx_code);
380 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
381 unsigned HOST_WIDE_INT *);
382 static rtx force_to_mode (rtx, enum machine_mode,
383 unsigned HOST_WIDE_INT, rtx, int);
384 static rtx if_then_else_cond (rtx, rtx *, rtx *);
385 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
386 static int rtx_equal_for_field_assignment_p (rtx, rtx);
387 static rtx make_field_assignment (rtx);
388 static rtx apply_distributive_law (rtx);
389 static rtx distribute_and_simplify_rtx (rtx, int);
390 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
391 unsigned HOST_WIDE_INT);
392 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
393 HOST_WIDE_INT, enum machine_mode, int *);
394 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
395 int);
396 static int recog_for_combine (rtx *, rtx, rtx *);
397 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
398 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
399 static void update_table_tick (rtx);
400 static void record_value_for_reg (rtx, rtx, rtx);
401 static void check_promoted_subreg (rtx, rtx);
402 static void record_dead_and_set_regs_1 (rtx, rtx, void *);
403 static void record_dead_and_set_regs (rtx);
404 static int get_last_value_validate (rtx *, rtx, int, int);
405 static rtx get_last_value (rtx);
406 static int use_crosses_set_p (rtx, int);
407 static void reg_dead_at_p_1 (rtx, rtx, void *);
408 static int reg_dead_at_p (rtx, rtx);
409 static void move_deaths (rtx, rtx, int, rtx, rtx *);
410 static int reg_bitfield_target_p (rtx, rtx);
411 static void distribute_notes (rtx, rtx, rtx, rtx);
412 static void distribute_links (rtx);
413 static void mark_used_regs_combine (rtx);
414 static int insn_cuid (rtx);
415 static void record_promoted_value (rtx, rtx);
416 static int unmentioned_reg_p_1 (rtx *, void *);
417 static bool unmentioned_reg_p (rtx, rtx);
420 /* It is not safe to use ordinary gen_lowpart in combine.
421 See comments in gen_lowpart_for_combine. */
422 #undef RTL_HOOKS_GEN_LOWPART
423 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine
425 #undef RTL_HOOKS_REG_NONZERO_REG_BITS
426 #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine
428 #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES
429 #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine
431 static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER;
434 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
435 insn. The substitution can be undone by undo_all. If INTO is already
436 set to NEWVAL, do not record this change. Because computing NEWVAL might
437 also call SUBST, we have to compute it before we put anything into
438 the undo table. */
440 static void
441 do_SUBST (rtx *into, rtx newval)
443 struct undo *buf;
444 rtx oldval = *into;
446 if (oldval == newval)
447 return;
449 /* We'd like to catch as many invalid transformations here as
450 possible. Unfortunately, there are way too many mode changes
451 that are perfectly valid, so we'd waste too much effort for
452 little gain doing the checks here. Focus on catching invalid
453 transformations involving integer constants. */
454 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
455 && GET_CODE (newval) == CONST_INT)
457 /* Sanity check that we're replacing oldval with a CONST_INT
458 that is a valid sign-extension for the original mode. */
459 gcc_assert (INTVAL (newval)
460 == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval)));
462 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
463 CONST_INT is not valid, because after the replacement, the
464 original mode would be gone. Unfortunately, we can't tell
465 when do_SUBST is called to replace the operand thereof, so we
466 perform this test on oldval instead, checking whether an
467 invalid replacement took place before we got here. */
468 gcc_assert (!(GET_CODE (oldval) == SUBREG
469 && GET_CODE (SUBREG_REG (oldval)) == CONST_INT));
470 gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
471 && GET_CODE (XEXP (oldval, 0)) == CONST_INT));
474 if (undobuf.frees)
475 buf = undobuf.frees, undobuf.frees = buf->next;
476 else
477 buf = xmalloc (sizeof (struct undo));
479 buf->is_int = 0;
480 buf->where.r = into;
481 buf->old_contents.r = oldval;
482 *into = newval;
484 buf->next = undobuf.undos, undobuf.undos = buf;
487 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
489 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
490 for the value of a HOST_WIDE_INT value (including CONST_INT) is
491 not safe. */
493 static void
494 do_SUBST_INT (int *into, int newval)
496 struct undo *buf;
497 int oldval = *into;
499 if (oldval == newval)
500 return;
502 if (undobuf.frees)
503 buf = undobuf.frees, undobuf.frees = buf->next;
504 else
505 buf = xmalloc (sizeof (struct undo));
507 buf->is_int = 1;
508 buf->where.i = into;
509 buf->old_contents.i = oldval;
510 *into = newval;
512 buf->next = undobuf.undos, undobuf.undos = buf;
515 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
517 /* Subroutine of try_combine. Determine whether the combine replacement
518 patterns NEWPAT and NEWI2PAT are cheaper according to insn_rtx_cost
519 that the original instruction sequence I1, I2 and I3. Note that I1
520 and/or NEWI2PAT may be NULL_RTX. This function returns false, if the
521 costs of all instructions can be estimated, and the replacements are
522 more expensive than the original sequence. */
524 static bool
525 combine_validate_cost (rtx i1, rtx i2, rtx i3, rtx newpat, rtx newi2pat)
527 int i1_cost, i2_cost, i3_cost;
528 int new_i2_cost, new_i3_cost;
529 int old_cost, new_cost;
531 /* Lookup the original insn_rtx_costs. */
532 i2_cost = INSN_UID (i2) <= last_insn_cost
533 ? uid_insn_cost[INSN_UID (i2)] : 0;
534 i3_cost = INSN_UID (i3) <= last_insn_cost
535 ? uid_insn_cost[INSN_UID (i3)] : 0;
537 if (i1)
539 i1_cost = INSN_UID (i1) <= last_insn_cost
540 ? uid_insn_cost[INSN_UID (i1)] : 0;
541 old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0)
542 ? i1_cost + i2_cost + i3_cost : 0;
544 else
546 old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0;
547 i1_cost = 0;
550 /* Calculate the replacement insn_rtx_costs. */
551 new_i3_cost = insn_rtx_cost (newpat);
552 if (newi2pat)
554 new_i2_cost = insn_rtx_cost (newi2pat);
555 new_cost = (new_i2_cost > 0 && new_i3_cost > 0)
556 ? new_i2_cost + new_i3_cost : 0;
558 else
560 new_cost = new_i3_cost;
561 new_i2_cost = 0;
564 if (undobuf.other_insn)
566 int old_other_cost, new_other_cost;
568 old_other_cost = (INSN_UID (undobuf.other_insn) <= last_insn_cost
569 ? uid_insn_cost[INSN_UID (undobuf.other_insn)] : 0);
570 new_other_cost = insn_rtx_cost (PATTERN (undobuf.other_insn));
571 if (old_other_cost > 0 && new_other_cost > 0)
573 old_cost += old_other_cost;
574 new_cost += new_other_cost;
576 else
577 old_cost = 0;
580 /* Disallow this recombination if both new_cost and old_cost are
581 greater than zero, and new_cost is greater than old cost. */
582 if (old_cost > 0
583 && new_cost > old_cost)
585 if (dump_file)
587 if (i1)
589 fprintf (dump_file,
590 "rejecting combination of insns %d, %d and %d\n",
591 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
592 fprintf (dump_file, "original costs %d + %d + %d = %d\n",
593 i1_cost, i2_cost, i3_cost, old_cost);
595 else
597 fprintf (dump_file,
598 "rejecting combination of insns %d and %d\n",
599 INSN_UID (i2), INSN_UID (i3));
600 fprintf (dump_file, "original costs %d + %d = %d\n",
601 i2_cost, i3_cost, old_cost);
604 if (newi2pat)
606 fprintf (dump_file, "replacement costs %d + %d = %d\n",
607 new_i2_cost, new_i3_cost, new_cost);
609 else
610 fprintf (dump_file, "replacement cost %d\n", new_cost);
613 return false;
616 /* Update the uid_insn_cost array with the replacement costs. */
617 uid_insn_cost[INSN_UID (i2)] = new_i2_cost;
618 uid_insn_cost[INSN_UID (i3)] = new_i3_cost;
619 if (i1)
620 uid_insn_cost[INSN_UID (i1)] = 0;
622 return true;
625 /* Main entry point for combiner. F is the first insn of the function.
626 NREGS is the first unused pseudo-reg number.
628 Return nonzero if the combiner has turned an indirect jump
629 instruction into a direct jump. */
631 combine_instructions (rtx f, unsigned int nregs)
633 rtx insn, next;
634 #ifdef HAVE_cc0
635 rtx prev;
636 #endif
637 int i;
638 rtx links, nextlinks;
640 int new_direct_jump_p = 0;
642 combine_attempts = 0;
643 combine_merges = 0;
644 combine_extras = 0;
645 combine_successes = 0;
647 combine_max_regno = nregs;
649 rtl_hooks = combine_rtl_hooks;
651 reg_stat = xcalloc (nregs, sizeof (struct reg_stat));
653 init_recog_no_volatile ();
655 /* Compute maximum uid value so uid_cuid can be allocated. */
657 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
658 if (INSN_UID (insn) > i)
659 i = INSN_UID (insn);
661 uid_cuid = xmalloc ((i + 1) * sizeof (int));
662 max_uid_cuid = i;
664 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
666 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
667 problems when, for example, we have j <<= 1 in a loop. */
669 nonzero_sign_valid = 0;
671 /* Compute the mapping from uids to cuids.
672 Cuids are numbers assigned to insns, like uids,
673 except that cuids increase monotonically through the code.
675 Scan all SETs and see if we can deduce anything about what
676 bits are known to be zero for some registers and how many copies
677 of the sign bit are known to exist for those registers.
679 Also set any known values so that we can use it while searching
680 for what bits are known to be set. */
682 label_tick = 1;
684 setup_incoming_promotions ();
686 refresh_blocks = sbitmap_alloc (last_basic_block);
687 sbitmap_zero (refresh_blocks);
689 /* Allocate array of current insn_rtx_costs. */
690 uid_insn_cost = xcalloc (max_uid_cuid + 1, sizeof (int));
691 last_insn_cost = max_uid_cuid;
693 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
695 uid_cuid[INSN_UID (insn)] = ++i;
696 subst_low_cuid = i;
697 subst_insn = insn;
699 if (INSN_P (insn))
701 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
702 NULL);
703 record_dead_and_set_regs (insn);
705 #ifdef AUTO_INC_DEC
706 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
707 if (REG_NOTE_KIND (links) == REG_INC)
708 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
709 NULL);
710 #endif
712 /* Record the current insn_rtx_cost of this instruction. */
713 if (NONJUMP_INSN_P (insn))
714 uid_insn_cost[INSN_UID (insn)] = insn_rtx_cost (PATTERN (insn));
715 if (dump_file)
716 fprintf(dump_file, "insn_cost %d: %d\n",
717 INSN_UID (insn), uid_insn_cost[INSN_UID (insn)]);
720 if (LABEL_P (insn))
721 label_tick++;
724 nonzero_sign_valid = 1;
726 /* Now scan all the insns in forward order. */
728 label_tick = 1;
729 last_call_cuid = 0;
730 mem_last_set = 0;
731 init_reg_last ();
732 setup_incoming_promotions ();
734 FOR_EACH_BB (this_basic_block)
736 for (insn = BB_HEAD (this_basic_block);
737 insn != NEXT_INSN (BB_END (this_basic_block));
738 insn = next ? next : NEXT_INSN (insn))
740 next = 0;
742 if (LABEL_P (insn))
743 label_tick++;
745 else if (INSN_P (insn))
747 /* See if we know about function return values before this
748 insn based upon SUBREG flags. */
749 check_promoted_subreg (insn, PATTERN (insn));
751 /* Try this insn with each insn it links back to. */
753 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
754 if ((next = try_combine (insn, XEXP (links, 0),
755 NULL_RTX, &new_direct_jump_p)) != 0)
756 goto retry;
758 /* Try each sequence of three linked insns ending with this one. */
760 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
762 rtx link = XEXP (links, 0);
764 /* If the linked insn has been replaced by a note, then there
765 is no point in pursuing this chain any further. */
766 if (NOTE_P (link))
767 continue;
769 for (nextlinks = LOG_LINKS (link);
770 nextlinks;
771 nextlinks = XEXP (nextlinks, 1))
772 if ((next = try_combine (insn, link,
773 XEXP (nextlinks, 0),
774 &new_direct_jump_p)) != 0)
775 goto retry;
778 #ifdef HAVE_cc0
779 /* Try to combine a jump insn that uses CC0
780 with a preceding insn that sets CC0, and maybe with its
781 logical predecessor as well.
782 This is how we make decrement-and-branch insns.
783 We need this special code because data flow connections
784 via CC0 do not get entered in LOG_LINKS. */
786 if (JUMP_P (insn)
787 && (prev = prev_nonnote_insn (insn)) != 0
788 && NONJUMP_INSN_P (prev)
789 && sets_cc0_p (PATTERN (prev)))
791 if ((next = try_combine (insn, prev,
792 NULL_RTX, &new_direct_jump_p)) != 0)
793 goto retry;
795 for (nextlinks = LOG_LINKS (prev); nextlinks;
796 nextlinks = XEXP (nextlinks, 1))
797 if ((next = try_combine (insn, prev,
798 XEXP (nextlinks, 0),
799 &new_direct_jump_p)) != 0)
800 goto retry;
803 /* Do the same for an insn that explicitly references CC0. */
804 if (NONJUMP_INSN_P (insn)
805 && (prev = prev_nonnote_insn (insn)) != 0
806 && NONJUMP_INSN_P (prev)
807 && sets_cc0_p (PATTERN (prev))
808 && GET_CODE (PATTERN (insn)) == SET
809 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
811 if ((next = try_combine (insn, prev,
812 NULL_RTX, &new_direct_jump_p)) != 0)
813 goto retry;
815 for (nextlinks = LOG_LINKS (prev); nextlinks;
816 nextlinks = XEXP (nextlinks, 1))
817 if ((next = try_combine (insn, prev,
818 XEXP (nextlinks, 0),
819 &new_direct_jump_p)) != 0)
820 goto retry;
823 /* Finally, see if any of the insns that this insn links to
824 explicitly references CC0. If so, try this insn, that insn,
825 and its predecessor if it sets CC0. */
826 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
827 if (NONJUMP_INSN_P (XEXP (links, 0))
828 && GET_CODE (PATTERN (XEXP (links, 0))) == SET
829 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0))))
830 && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0
831 && NONJUMP_INSN_P (prev)
832 && sets_cc0_p (PATTERN (prev))
833 && (next = try_combine (insn, XEXP (links, 0),
834 prev, &new_direct_jump_p)) != 0)
835 goto retry;
836 #endif
838 /* Try combining an insn with two different insns whose results it
839 uses. */
840 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
841 for (nextlinks = XEXP (links, 1); nextlinks;
842 nextlinks = XEXP (nextlinks, 1))
843 if ((next = try_combine (insn, XEXP (links, 0),
844 XEXP (nextlinks, 0),
845 &new_direct_jump_p)) != 0)
846 goto retry;
848 /* Try this insn with each REG_EQUAL note it links back to. */
849 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
851 rtx set, note;
852 rtx temp = XEXP (links, 0);
853 if ((set = single_set (temp)) != 0
854 && (note = find_reg_equal_equiv_note (temp)) != 0
855 && GET_CODE (XEXP (note, 0)) != EXPR_LIST
856 /* Avoid using a register that may already been marked
857 dead by an earlier instruction. */
858 && ! unmentioned_reg_p (XEXP (note, 0), SET_SRC (set)))
860 /* Temporarily replace the set's source with the
861 contents of the REG_EQUAL note. The insn will
862 be deleted or recognized by try_combine. */
863 rtx orig = SET_SRC (set);
864 SET_SRC (set) = XEXP (note, 0);
865 next = try_combine (insn, temp, NULL_RTX,
866 &new_direct_jump_p);
867 if (next)
868 goto retry;
869 SET_SRC (set) = orig;
873 if (!NOTE_P (insn))
874 record_dead_and_set_regs (insn);
876 retry:
881 clear_bb_flags ();
883 EXECUTE_IF_SET_IN_SBITMAP (refresh_blocks, 0, i,
884 BASIC_BLOCK (i)->flags |= BB_DIRTY);
885 new_direct_jump_p |= purge_all_dead_edges (0);
886 delete_noop_moves ();
888 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
889 PROP_DEATH_NOTES | PROP_SCAN_DEAD_CODE
890 | PROP_KILL_DEAD_CODE);
892 /* Clean up. */
893 sbitmap_free (refresh_blocks);
894 free (uid_insn_cost);
895 free (reg_stat);
896 free (uid_cuid);
899 struct undo *undo, *next;
900 for (undo = undobuf.frees; undo; undo = next)
902 next = undo->next;
903 free (undo);
905 undobuf.frees = 0;
908 total_attempts += combine_attempts;
909 total_merges += combine_merges;
910 total_extras += combine_extras;
911 total_successes += combine_successes;
913 nonzero_sign_valid = 0;
914 rtl_hooks = general_rtl_hooks;
916 /* Make recognizer allow volatile MEMs again. */
917 init_recog ();
919 return new_direct_jump_p;
922 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
924 static void
925 init_reg_last (void)
927 unsigned int i;
928 for (i = 0; i < combine_max_regno; i++)
929 memset (reg_stat + i, 0, offsetof (struct reg_stat, sign_bit_copies));
932 /* Set up any promoted values for incoming argument registers. */
934 static void
935 setup_incoming_promotions (void)
937 unsigned int regno;
938 rtx reg;
939 enum machine_mode mode;
940 int unsignedp;
941 rtx first = get_insns ();
943 if (targetm.calls.promote_function_args (TREE_TYPE (cfun->decl)))
945 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
946 /* Check whether this register can hold an incoming pointer
947 argument. FUNCTION_ARG_REGNO_P tests outgoing register
948 numbers, so translate if necessary due to register windows. */
949 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (regno))
950 && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0)
952 record_value_for_reg
953 (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND
954 : SIGN_EXTEND),
955 GET_MODE (reg),
956 gen_rtx_CLOBBER (mode, const0_rtx)));
961 /* Called via note_stores. If X is a pseudo that is narrower than
962 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
964 If we are setting only a portion of X and we can't figure out what
965 portion, assume all bits will be used since we don't know what will
966 be happening.
968 Similarly, set how many bits of X are known to be copies of the sign bit
969 at all locations in the function. This is the smallest number implied
970 by any set of X. */
972 static void
973 set_nonzero_bits_and_sign_copies (rtx x, rtx set,
974 void *data ATTRIBUTE_UNUSED)
976 unsigned int num;
978 if (REG_P (x)
979 && REGNO (x) >= FIRST_PSEUDO_REGISTER
980 /* If this register is undefined at the start of the file, we can't
981 say what its contents were. */
982 && ! REGNO_REG_SET_P (ENTRY_BLOCK_PTR->next_bb->global_live_at_start, REGNO (x))
983 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
985 if (set == 0 || GET_CODE (set) == CLOBBER)
987 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
988 reg_stat[REGNO (x)].sign_bit_copies = 1;
989 return;
992 /* If this is a complex assignment, see if we can convert it into a
993 simple assignment. */
994 set = expand_field_assignment (set);
996 /* If this is a simple assignment, or we have a paradoxical SUBREG,
997 set what we know about X. */
999 if (SET_DEST (set) == x
1000 || (GET_CODE (SET_DEST (set)) == SUBREG
1001 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
1002 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set)))))
1003 && SUBREG_REG (SET_DEST (set)) == x))
1005 rtx src = SET_SRC (set);
1007 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
1008 /* If X is narrower than a word and SRC is a non-negative
1009 constant that would appear negative in the mode of X,
1010 sign-extend it for use in reg_stat[].nonzero_bits because some
1011 machines (maybe most) will actually do the sign-extension
1012 and this is the conservative approach.
1014 ??? For 2.5, try to tighten up the MD files in this regard
1015 instead of this kludge. */
1017 if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD
1018 && GET_CODE (src) == CONST_INT
1019 && INTVAL (src) > 0
1020 && 0 != (INTVAL (src)
1021 & ((HOST_WIDE_INT) 1
1022 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
1023 src = GEN_INT (INTVAL (src)
1024 | ((HOST_WIDE_INT) (-1)
1025 << GET_MODE_BITSIZE (GET_MODE (x))));
1026 #endif
1028 /* Don't call nonzero_bits if it cannot change anything. */
1029 if (reg_stat[REGNO (x)].nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
1030 reg_stat[REGNO (x)].nonzero_bits
1031 |= nonzero_bits (src, nonzero_bits_mode);
1032 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
1033 if (reg_stat[REGNO (x)].sign_bit_copies == 0
1034 || reg_stat[REGNO (x)].sign_bit_copies > num)
1035 reg_stat[REGNO (x)].sign_bit_copies = num;
1037 else
1039 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1040 reg_stat[REGNO (x)].sign_bit_copies = 1;
1045 /* See if INSN can be combined into I3. PRED and SUCC are optionally
1046 insns that were previously combined into I3 or that will be combined
1047 into the merger of INSN and I3.
1049 Return 0 if the combination is not allowed for any reason.
1051 If the combination is allowed, *PDEST will be set to the single
1052 destination of INSN and *PSRC to the single source, and this function
1053 will return 1. */
1055 static int
1056 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED, rtx succ,
1057 rtx *pdest, rtx *psrc)
1059 int i;
1060 rtx set = 0, src, dest;
1061 rtx p;
1062 #ifdef AUTO_INC_DEC
1063 rtx link;
1064 #endif
1065 int all_adjacent = (succ ? (next_active_insn (insn) == succ
1066 && next_active_insn (succ) == i3)
1067 : next_active_insn (insn) == i3);
1069 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
1070 or a PARALLEL consisting of such a SET and CLOBBERs.
1072 If INSN has CLOBBER parallel parts, ignore them for our processing.
1073 By definition, these happen during the execution of the insn. When it
1074 is merged with another insn, all bets are off. If they are, in fact,
1075 needed and aren't also supplied in I3, they may be added by
1076 recog_for_combine. Otherwise, it won't match.
1078 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
1079 note.
1081 Get the source and destination of INSN. If more than one, can't
1082 combine. */
1084 if (GET_CODE (PATTERN (insn)) == SET)
1085 set = PATTERN (insn);
1086 else if (GET_CODE (PATTERN (insn)) == PARALLEL
1087 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
1089 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1091 rtx elt = XVECEXP (PATTERN (insn), 0, i);
1092 rtx note;
1094 switch (GET_CODE (elt))
1096 /* This is important to combine floating point insns
1097 for the SH4 port. */
1098 case USE:
1099 /* Combining an isolated USE doesn't make sense.
1100 We depend here on combinable_i3pat to reject them. */
1101 /* The code below this loop only verifies that the inputs of
1102 the SET in INSN do not change. We call reg_set_between_p
1103 to verify that the REG in the USE does not change between
1104 I3 and INSN.
1105 If the USE in INSN was for a pseudo register, the matching
1106 insn pattern will likely match any register; combining this
1107 with any other USE would only be safe if we knew that the
1108 used registers have identical values, or if there was
1109 something to tell them apart, e.g. different modes. For
1110 now, we forgo such complicated tests and simply disallow
1111 combining of USES of pseudo registers with any other USE. */
1112 if (REG_P (XEXP (elt, 0))
1113 && GET_CODE (PATTERN (i3)) == PARALLEL)
1115 rtx i3pat = PATTERN (i3);
1116 int i = XVECLEN (i3pat, 0) - 1;
1117 unsigned int regno = REGNO (XEXP (elt, 0));
1121 rtx i3elt = XVECEXP (i3pat, 0, i);
1123 if (GET_CODE (i3elt) == USE
1124 && REG_P (XEXP (i3elt, 0))
1125 && (REGNO (XEXP (i3elt, 0)) == regno
1126 ? reg_set_between_p (XEXP (elt, 0),
1127 PREV_INSN (insn), i3)
1128 : regno >= FIRST_PSEUDO_REGISTER))
1129 return 0;
1131 while (--i >= 0);
1133 break;
1135 /* We can ignore CLOBBERs. */
1136 case CLOBBER:
1137 break;
1139 case SET:
1140 /* Ignore SETs whose result isn't used but not those that
1141 have side-effects. */
1142 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1143 && (!(note = find_reg_note (insn, REG_EH_REGION, NULL_RTX))
1144 || INTVAL (XEXP (note, 0)) <= 0)
1145 && ! side_effects_p (elt))
1146 break;
1148 /* If we have already found a SET, this is a second one and
1149 so we cannot combine with this insn. */
1150 if (set)
1151 return 0;
1153 set = elt;
1154 break;
1156 default:
1157 /* Anything else means we can't combine. */
1158 return 0;
1162 if (set == 0
1163 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1164 so don't do anything with it. */
1165 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1166 return 0;
1168 else
1169 return 0;
1171 if (set == 0)
1172 return 0;
1174 set = expand_field_assignment (set);
1175 src = SET_SRC (set), dest = SET_DEST (set);
1177 /* Don't eliminate a store in the stack pointer. */
1178 if (dest == stack_pointer_rtx
1179 /* Don't combine with an insn that sets a register to itself if it has
1180 a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */
1181 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1182 /* Can't merge an ASM_OPERANDS. */
1183 || GET_CODE (src) == ASM_OPERANDS
1184 /* Can't merge a function call. */
1185 || GET_CODE (src) == CALL
1186 /* Don't eliminate a function call argument. */
1187 || (CALL_P (i3)
1188 && (find_reg_fusage (i3, USE, dest)
1189 || (REG_P (dest)
1190 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1191 && global_regs[REGNO (dest)])))
1192 /* Don't substitute into an incremented register. */
1193 || FIND_REG_INC_NOTE (i3, dest)
1194 || (succ && FIND_REG_INC_NOTE (succ, dest))
1195 /* Don't substitute into a non-local goto, this confuses CFG. */
1196 || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX))
1197 #if 0
1198 /* Don't combine the end of a libcall into anything. */
1199 /* ??? This gives worse code, and appears to be unnecessary, since no
1200 pass after flow uses REG_LIBCALL/REG_RETVAL notes. Local-alloc does
1201 use REG_RETVAL notes for noconflict blocks, but other code here
1202 makes sure that those insns don't disappear. */
1203 || find_reg_note (insn, REG_RETVAL, NULL_RTX)
1204 #endif
1205 /* Make sure that DEST is not used after SUCC but before I3. */
1206 || (succ && ! all_adjacent
1207 && reg_used_between_p (dest, succ, i3))
1208 /* Make sure that the value that is to be substituted for the register
1209 does not use any registers whose values alter in between. However,
1210 If the insns are adjacent, a use can't cross a set even though we
1211 think it might (this can happen for a sequence of insns each setting
1212 the same destination; last_set of that register might point to
1213 a NOTE). If INSN has a REG_EQUIV note, the register is always
1214 equivalent to the memory so the substitution is valid even if there
1215 are intervening stores. Also, don't move a volatile asm or
1216 UNSPEC_VOLATILE across any other insns. */
1217 || (! all_adjacent
1218 && (((!MEM_P (src)
1219 || ! find_reg_note (insn, REG_EQUIV, src))
1220 && use_crosses_set_p (src, INSN_CUID (insn)))
1221 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1222 || GET_CODE (src) == UNSPEC_VOLATILE))
1223 /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get
1224 better register allocation by not doing the combine. */
1225 || find_reg_note (i3, REG_NO_CONFLICT, dest)
1226 || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest))
1227 /* Don't combine across a CALL_INSN, because that would possibly
1228 change whether the life span of some REGs crosses calls or not,
1229 and it is a pain to update that information.
1230 Exception: if source is a constant, moving it later can't hurt.
1231 Accept that special case, because it helps -fforce-addr a lot. */
1232 || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src)))
1233 return 0;
1235 /* DEST must either be a REG or CC0. */
1236 if (REG_P (dest))
1238 /* If register alignment is being enforced for multi-word items in all
1239 cases except for parameters, it is possible to have a register copy
1240 insn referencing a hard register that is not allowed to contain the
1241 mode being copied and which would not be valid as an operand of most
1242 insns. Eliminate this problem by not combining with such an insn.
1244 Also, on some machines we don't want to extend the life of a hard
1245 register. */
1247 if (REG_P (src)
1248 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1249 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1250 /* Don't extend the life of a hard register unless it is
1251 user variable (if we have few registers) or it can't
1252 fit into the desired register (meaning something special
1253 is going on).
1254 Also avoid substituting a return register into I3, because
1255 reload can't handle a conflict with constraints of other
1256 inputs. */
1257 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1258 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1259 return 0;
1261 else if (GET_CODE (dest) != CC0)
1262 return 0;
1265 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1266 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1267 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER)
1269 /* Don't substitute for a register intended as a clobberable
1270 operand. */
1271 rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0);
1272 if (rtx_equal_p (reg, dest))
1273 return 0;
1275 /* If the clobber represents an earlyclobber operand, we must not
1276 substitute an expression containing the clobbered register.
1277 As we do not analyze the constraint strings here, we have to
1278 make the conservative assumption. However, if the register is
1279 a fixed hard reg, the clobber cannot represent any operand;
1280 we leave it up to the machine description to either accept or
1281 reject use-and-clobber patterns. */
1282 if (!REG_P (reg)
1283 || REGNO (reg) >= FIRST_PSEUDO_REGISTER
1284 || !fixed_regs[REGNO (reg)])
1285 if (reg_overlap_mentioned_p (reg, src))
1286 return 0;
1289 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1290 or not), reject, unless nothing volatile comes between it and I3 */
1292 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1294 /* Make sure succ doesn't contain a volatile reference. */
1295 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1296 return 0;
1298 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1299 if (INSN_P (p) && p != succ && volatile_refs_p (PATTERN (p)))
1300 return 0;
1303 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1304 to be an explicit register variable, and was chosen for a reason. */
1306 if (GET_CODE (src) == ASM_OPERANDS
1307 && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1308 return 0;
1310 /* If there are any volatile insns between INSN and I3, reject, because
1311 they might affect machine state. */
1313 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1314 if (INSN_P (p) && p != succ && volatile_insn_p (PATTERN (p)))
1315 return 0;
1317 /* If INSN contains an autoincrement or autodecrement, make sure that
1318 register is not used between there and I3, and not already used in
1319 I3 either. Neither must it be used in PRED or SUCC, if they exist.
1320 Also insist that I3 not be a jump; if it were one
1321 and the incremented register were spilled, we would lose. */
1323 #ifdef AUTO_INC_DEC
1324 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1325 if (REG_NOTE_KIND (link) == REG_INC
1326 && (JUMP_P (i3)
1327 || reg_used_between_p (XEXP (link, 0), insn, i3)
1328 || (pred != NULL_RTX
1329 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred)))
1330 || (succ != NULL_RTX
1331 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ)))
1332 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1333 return 0;
1334 #endif
1336 #ifdef HAVE_cc0
1337 /* Don't combine an insn that follows a CC0-setting insn.
1338 An insn that uses CC0 must not be separated from the one that sets it.
1339 We do, however, allow I2 to follow a CC0-setting insn if that insn
1340 is passed as I1; in that case it will be deleted also.
1341 We also allow combining in this case if all the insns are adjacent
1342 because that would leave the two CC0 insns adjacent as well.
1343 It would be more logical to test whether CC0 occurs inside I1 or I2,
1344 but that would be much slower, and this ought to be equivalent. */
1346 p = prev_nonnote_insn (insn);
1347 if (p && p != pred && NONJUMP_INSN_P (p) && sets_cc0_p (PATTERN (p))
1348 && ! all_adjacent)
1349 return 0;
1350 #endif
1352 /* If we get here, we have passed all the tests and the combination is
1353 to be allowed. */
1355 *pdest = dest;
1356 *psrc = src;
1358 return 1;
1361 /* LOC is the location within I3 that contains its pattern or the component
1362 of a PARALLEL of the pattern. We validate that it is valid for combining.
1364 One problem is if I3 modifies its output, as opposed to replacing it
1365 entirely, we can't allow the output to contain I2DEST or I1DEST as doing
1366 so would produce an insn that is not equivalent to the original insns.
1368 Consider:
1370 (set (reg:DI 101) (reg:DI 100))
1371 (set (subreg:SI (reg:DI 101) 0) <foo>)
1373 This is NOT equivalent to:
1375 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
1376 (set (reg:DI 101) (reg:DI 100))])
1378 Not only does this modify 100 (in which case it might still be valid
1379 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
1381 We can also run into a problem if I2 sets a register that I1
1382 uses and I1 gets directly substituted into I3 (not via I2). In that
1383 case, we would be getting the wrong value of I2DEST into I3, so we
1384 must reject the combination. This case occurs when I2 and I1 both
1385 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
1386 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
1387 of a SET must prevent combination from occurring.
1389 Before doing the above check, we first try to expand a field assignment
1390 into a set of logical operations.
1392 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
1393 we place a register that is both set and used within I3. If more than one
1394 such register is detected, we fail.
1396 Return 1 if the combination is valid, zero otherwise. */
1398 static int
1399 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest,
1400 int i1_not_in_src, rtx *pi3dest_killed)
1402 rtx x = *loc;
1404 if (GET_CODE (x) == SET)
1406 rtx set = x ;
1407 rtx dest = SET_DEST (set);
1408 rtx src = SET_SRC (set);
1409 rtx inner_dest = dest;
1411 while (GET_CODE (inner_dest) == STRICT_LOW_PART
1412 || GET_CODE (inner_dest) == SUBREG
1413 || GET_CODE (inner_dest) == ZERO_EXTRACT)
1414 inner_dest = XEXP (inner_dest, 0);
1416 /* Check for the case where I3 modifies its output, as discussed
1417 above. We don't want to prevent pseudos from being combined
1418 into the address of a MEM, so only prevent the combination if
1419 i1 or i2 set the same MEM. */
1420 if ((inner_dest != dest &&
1421 (!MEM_P (inner_dest)
1422 || rtx_equal_p (i2dest, inner_dest)
1423 || (i1dest && rtx_equal_p (i1dest, inner_dest)))
1424 && (reg_overlap_mentioned_p (i2dest, inner_dest)
1425 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))))
1427 /* This is the same test done in can_combine_p except we can't test
1428 all_adjacent; we don't have to, since this instruction will stay
1429 in place, thus we are not considering increasing the lifetime of
1430 INNER_DEST.
1432 Also, if this insn sets a function argument, combining it with
1433 something that might need a spill could clobber a previous
1434 function argument; the all_adjacent test in can_combine_p also
1435 checks this; here, we do a more specific test for this case. */
1437 || (REG_P (inner_dest)
1438 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
1439 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
1440 GET_MODE (inner_dest))))
1441 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)))
1442 return 0;
1444 /* If DEST is used in I3, it is being killed in this insn,
1445 so record that for later.
1446 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
1447 STACK_POINTER_REGNUM, since these are always considered to be
1448 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
1449 if (pi3dest_killed && REG_P (dest)
1450 && reg_referenced_p (dest, PATTERN (i3))
1451 && REGNO (dest) != FRAME_POINTER_REGNUM
1452 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1453 && REGNO (dest) != HARD_FRAME_POINTER_REGNUM
1454 #endif
1455 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
1456 && (REGNO (dest) != ARG_POINTER_REGNUM
1457 || ! fixed_regs [REGNO (dest)])
1458 #endif
1459 && REGNO (dest) != STACK_POINTER_REGNUM)
1461 if (*pi3dest_killed)
1462 return 0;
1464 *pi3dest_killed = dest;
1468 else if (GET_CODE (x) == PARALLEL)
1470 int i;
1472 for (i = 0; i < XVECLEN (x, 0); i++)
1473 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest,
1474 i1_not_in_src, pi3dest_killed))
1475 return 0;
1478 return 1;
1481 /* Return 1 if X is an arithmetic expression that contains a multiplication
1482 and division. We don't count multiplications by powers of two here. */
1484 static int
1485 contains_muldiv (rtx x)
1487 switch (GET_CODE (x))
1489 case MOD: case DIV: case UMOD: case UDIV:
1490 return 1;
1492 case MULT:
1493 return ! (GET_CODE (XEXP (x, 1)) == CONST_INT
1494 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0);
1495 default:
1496 if (BINARY_P (x))
1497 return contains_muldiv (XEXP (x, 0))
1498 || contains_muldiv (XEXP (x, 1));
1500 if (UNARY_P (x))
1501 return contains_muldiv (XEXP (x, 0));
1503 return 0;
1507 /* Determine whether INSN can be used in a combination. Return nonzero if
1508 not. This is used in try_combine to detect early some cases where we
1509 can't perform combinations. */
1511 static int
1512 cant_combine_insn_p (rtx insn)
1514 rtx set;
1515 rtx src, dest;
1517 /* If this isn't really an insn, we can't do anything.
1518 This can occur when flow deletes an insn that it has merged into an
1519 auto-increment address. */
1520 if (! INSN_P (insn))
1521 return 1;
1523 /* Never combine loads and stores involving hard regs that are likely
1524 to be spilled. The register allocator can usually handle such
1525 reg-reg moves by tying. If we allow the combiner to make
1526 substitutions of likely-spilled regs, we may abort in reload.
1527 As an exception, we allow combinations involving fixed regs; these are
1528 not available to the register allocator so there's no risk involved. */
1530 set = single_set (insn);
1531 if (! set)
1532 return 0;
1533 src = SET_SRC (set);
1534 dest = SET_DEST (set);
1535 if (GET_CODE (src) == SUBREG)
1536 src = SUBREG_REG (src);
1537 if (GET_CODE (dest) == SUBREG)
1538 dest = SUBREG_REG (dest);
1539 if (REG_P (src) && REG_P (dest)
1540 && ((REGNO (src) < FIRST_PSEUDO_REGISTER
1541 && ! fixed_regs[REGNO (src)]
1542 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (src))))
1543 || (REGNO (dest) < FIRST_PSEUDO_REGISTER
1544 && ! fixed_regs[REGNO (dest)]
1545 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (dest))))))
1546 return 1;
1548 return 0;
1551 /* Adjust INSN after we made a change to its destination.
1553 Changing the destination can invalidate notes that say something about
1554 the results of the insn and a LOG_LINK pointing to the insn. */
1556 static void
1557 adjust_for_new_dest (rtx insn)
1559 rtx *loc;
1561 /* For notes, be conservative and simply remove them. */
1562 loc = &REG_NOTES (insn);
1563 while (*loc)
1565 enum reg_note kind = REG_NOTE_KIND (*loc);
1566 if (kind == REG_EQUAL || kind == REG_EQUIV)
1567 *loc = XEXP (*loc, 1);
1568 else
1569 loc = &XEXP (*loc, 1);
1572 /* The new insn will have a destination that was previously the destination
1573 of an insn just above it. Call distribute_links to make a LOG_LINK from
1574 the next use of that destination. */
1575 distribute_links (gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX));
1578 /* Try to combine the insns I1 and I2 into I3.
1579 Here I1 and I2 appear earlier than I3.
1580 I1 can be zero; then we combine just I2 into I3.
1582 If we are combining three insns and the resulting insn is not recognized,
1583 try splitting it into two insns. If that happens, I2 and I3 are retained
1584 and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2
1585 are pseudo-deleted.
1587 Return 0 if the combination does not work. Then nothing is changed.
1588 If we did the combination, return the insn at which combine should
1589 resume scanning.
1591 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
1592 new direct jump instruction. */
1594 static rtx
1595 try_combine (rtx i3, rtx i2, rtx i1, int *new_direct_jump_p)
1597 /* New patterns for I3 and I2, respectively. */
1598 rtx newpat, newi2pat = 0;
1599 rtvec newpat_vec_with_clobbers = 0;
1600 int substed_i2 = 0, substed_i1 = 0;
1601 /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */
1602 int added_sets_1, added_sets_2;
1603 /* Total number of SETs to put into I3. */
1604 int total_sets;
1605 /* Nonzero if I2's body now appears in I3. */
1606 int i2_is_used;
1607 /* INSN_CODEs for new I3, new I2, and user of condition code. */
1608 int insn_code_number, i2_code_number = 0, other_code_number = 0;
1609 /* Contains I3 if the destination of I3 is used in its source, which means
1610 that the old life of I3 is being killed. If that usage is placed into
1611 I2 and not in I3, a REG_DEAD note must be made. */
1612 rtx i3dest_killed = 0;
1613 /* SET_DEST and SET_SRC of I2 and I1. */
1614 rtx i2dest, i2src, i1dest = 0, i1src = 0;
1615 /* PATTERN (I2), or a copy of it in certain cases. */
1616 rtx i2pat;
1617 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
1618 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
1619 int i1_feeds_i3 = 0;
1620 /* Notes that must be added to REG_NOTES in I3 and I2. */
1621 rtx new_i3_notes, new_i2_notes;
1622 /* Notes that we substituted I3 into I2 instead of the normal case. */
1623 int i3_subst_into_i2 = 0;
1624 /* Notes that I1, I2 or I3 is a MULT operation. */
1625 int have_mult = 0;
1626 int swap_i2i3 = 0;
1628 int maxreg;
1629 rtx temp;
1630 rtx link;
1631 int i;
1633 /* Exit early if one of the insns involved can't be used for
1634 combinations. */
1635 if (cant_combine_insn_p (i3)
1636 || cant_combine_insn_p (i2)
1637 || (i1 && cant_combine_insn_p (i1))
1638 /* We also can't do anything if I3 has a
1639 REG_LIBCALL note since we don't want to disrupt the contiguity of a
1640 libcall. */
1641 #if 0
1642 /* ??? This gives worse code, and appears to be unnecessary, since no
1643 pass after flow uses REG_LIBCALL/REG_RETVAL notes. */
1644 || find_reg_note (i3, REG_LIBCALL, NULL_RTX)
1645 #endif
1647 return 0;
1649 combine_attempts++;
1650 undobuf.other_insn = 0;
1652 /* Reset the hard register usage information. */
1653 CLEAR_HARD_REG_SET (newpat_used_regs);
1655 /* If I1 and I2 both feed I3, they can be in any order. To simplify the
1656 code below, set I1 to be the earlier of the two insns. */
1657 if (i1 && INSN_CUID (i1) > INSN_CUID (i2))
1658 temp = i1, i1 = i2, i2 = temp;
1660 added_links_insn = 0;
1662 /* First check for one important special-case that the code below will
1663 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
1664 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
1665 we may be able to replace that destination with the destination of I3.
1666 This occurs in the common code where we compute both a quotient and
1667 remainder into a structure, in which case we want to do the computation
1668 directly into the structure to avoid register-register copies.
1670 Note that this case handles both multiple sets in I2 and also
1671 cases where I2 has a number of CLOBBER or PARALLELs.
1673 We make very conservative checks below and only try to handle the
1674 most common cases of this. For example, we only handle the case
1675 where I2 and I3 are adjacent to avoid making difficult register
1676 usage tests. */
1678 if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET
1679 && REG_P (SET_SRC (PATTERN (i3)))
1680 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
1681 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
1682 && GET_CODE (PATTERN (i2)) == PARALLEL
1683 && ! side_effects_p (SET_DEST (PATTERN (i3)))
1684 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
1685 below would need to check what is inside (and reg_overlap_mentioned_p
1686 doesn't support those codes anyway). Don't allow those destinations;
1687 the resulting insn isn't likely to be recognized anyway. */
1688 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
1689 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
1690 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
1691 SET_DEST (PATTERN (i3)))
1692 && next_real_insn (i2) == i3)
1694 rtx p2 = PATTERN (i2);
1696 /* Make sure that the destination of I3,
1697 which we are going to substitute into one output of I2,
1698 is not used within another output of I2. We must avoid making this:
1699 (parallel [(set (mem (reg 69)) ...)
1700 (set (reg 69) ...)])
1701 which is not well-defined as to order of actions.
1702 (Besides, reload can't handle output reloads for this.)
1704 The problem can also happen if the dest of I3 is a memory ref,
1705 if another dest in I2 is an indirect memory ref. */
1706 for (i = 0; i < XVECLEN (p2, 0); i++)
1707 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1708 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1709 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
1710 SET_DEST (XVECEXP (p2, 0, i))))
1711 break;
1713 if (i == XVECLEN (p2, 0))
1714 for (i = 0; i < XVECLEN (p2, 0); i++)
1715 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1716 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1717 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
1719 combine_merges++;
1721 subst_insn = i3;
1722 subst_low_cuid = INSN_CUID (i2);
1724 added_sets_2 = added_sets_1 = 0;
1725 i2dest = SET_SRC (PATTERN (i3));
1727 /* Replace the dest in I2 with our dest and make the resulting
1728 insn the new pattern for I3. Then skip to where we
1729 validate the pattern. Everything was set up above. */
1730 SUBST (SET_DEST (XVECEXP (p2, 0, i)),
1731 SET_DEST (PATTERN (i3)));
1733 newpat = p2;
1734 i3_subst_into_i2 = 1;
1735 goto validate_replacement;
1739 /* If I2 is setting a double-word pseudo to a constant and I3 is setting
1740 one of those words to another constant, merge them by making a new
1741 constant. */
1742 if (i1 == 0
1743 && (temp = single_set (i2)) != 0
1744 && (GET_CODE (SET_SRC (temp)) == CONST_INT
1745 || GET_CODE (SET_SRC (temp)) == CONST_DOUBLE)
1746 && REG_P (SET_DEST (temp))
1747 && GET_MODE_CLASS (GET_MODE (SET_DEST (temp))) == MODE_INT
1748 && GET_MODE_SIZE (GET_MODE (SET_DEST (temp))) == 2 * UNITS_PER_WORD
1749 && GET_CODE (PATTERN (i3)) == SET
1750 && GET_CODE (SET_DEST (PATTERN (i3))) == SUBREG
1751 && SUBREG_REG (SET_DEST (PATTERN (i3))) == SET_DEST (temp)
1752 && GET_MODE_CLASS (GET_MODE (SET_DEST (PATTERN (i3)))) == MODE_INT
1753 && GET_MODE_SIZE (GET_MODE (SET_DEST (PATTERN (i3)))) == UNITS_PER_WORD
1754 && GET_CODE (SET_SRC (PATTERN (i3))) == CONST_INT)
1756 HOST_WIDE_INT lo, hi;
1758 if (GET_CODE (SET_SRC (temp)) == CONST_INT)
1759 lo = INTVAL (SET_SRC (temp)), hi = lo < 0 ? -1 : 0;
1760 else
1762 lo = CONST_DOUBLE_LOW (SET_SRC (temp));
1763 hi = CONST_DOUBLE_HIGH (SET_SRC (temp));
1766 if (subreg_lowpart_p (SET_DEST (PATTERN (i3))))
1768 /* We don't handle the case of the target word being wider
1769 than a host wide int. */
1770 gcc_assert (HOST_BITS_PER_WIDE_INT >= BITS_PER_WORD);
1772 lo &= ~(UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1);
1773 lo |= (INTVAL (SET_SRC (PATTERN (i3)))
1774 & (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1));
1776 else if (HOST_BITS_PER_WIDE_INT == BITS_PER_WORD)
1777 hi = INTVAL (SET_SRC (PATTERN (i3)));
1778 else if (HOST_BITS_PER_WIDE_INT >= 2 * BITS_PER_WORD)
1780 int sign = -(int) ((unsigned HOST_WIDE_INT) lo
1781 >> (HOST_BITS_PER_WIDE_INT - 1));
1783 lo &= ~ (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD
1784 (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1));
1785 lo |= (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD
1786 (INTVAL (SET_SRC (PATTERN (i3)))));
1787 if (hi == sign)
1788 hi = lo < 0 ? -1 : 0;
1790 else
1791 /* We don't handle the case of the higher word not fitting
1792 entirely in either hi or lo. */
1793 gcc_unreachable ();
1795 combine_merges++;
1796 subst_insn = i3;
1797 subst_low_cuid = INSN_CUID (i2);
1798 added_sets_2 = added_sets_1 = 0;
1799 i2dest = SET_DEST (temp);
1801 SUBST (SET_SRC (temp),
1802 immed_double_const (lo, hi, GET_MODE (SET_DEST (temp))));
1804 newpat = PATTERN (i2);
1805 goto validate_replacement;
1808 #ifndef HAVE_cc0
1809 /* If we have no I1 and I2 looks like:
1810 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
1811 (set Y OP)])
1812 make up a dummy I1 that is
1813 (set Y OP)
1814 and change I2 to be
1815 (set (reg:CC X) (compare:CC Y (const_int 0)))
1817 (We can ignore any trailing CLOBBERs.)
1819 This undoes a previous combination and allows us to match a branch-and-
1820 decrement insn. */
1822 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
1823 && XVECLEN (PATTERN (i2), 0) >= 2
1824 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
1825 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
1826 == MODE_CC)
1827 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
1828 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
1829 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
1830 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)))
1831 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
1832 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
1834 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
1835 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
1836 break;
1838 if (i == 1)
1840 /* We make I1 with the same INSN_UID as I2. This gives it
1841 the same INSN_CUID for value tracking. Our fake I1 will
1842 never appear in the insn stream so giving it the same INSN_UID
1843 as I2 will not cause a problem. */
1845 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
1846 BLOCK_FOR_INSN (i2), INSN_LOCATOR (i2),
1847 XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX,
1848 NULL_RTX);
1850 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
1851 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
1852 SET_DEST (PATTERN (i1)));
1855 #endif
1857 /* Verify that I2 and I1 are valid for combining. */
1858 if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src)
1859 || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src)))
1861 undo_all ();
1862 return 0;
1865 /* Record whether I2DEST is used in I2SRC and similarly for the other
1866 cases. Knowing this will help in register status updating below. */
1867 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
1868 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
1869 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
1871 /* See if I1 directly feeds into I3. It does if I1DEST is not used
1872 in I2SRC. */
1873 i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src);
1875 /* Ensure that I3's pattern can be the destination of combines. */
1876 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest,
1877 i1 && i2dest_in_i1src && i1_feeds_i3,
1878 &i3dest_killed))
1880 undo_all ();
1881 return 0;
1884 /* See if any of the insns is a MULT operation. Unless one is, we will
1885 reject a combination that is, since it must be slower. Be conservative
1886 here. */
1887 if (GET_CODE (i2src) == MULT
1888 || (i1 != 0 && GET_CODE (i1src) == MULT)
1889 || (GET_CODE (PATTERN (i3)) == SET
1890 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
1891 have_mult = 1;
1893 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
1894 We used to do this EXCEPT in one case: I3 has a post-inc in an
1895 output operand. However, that exception can give rise to insns like
1896 mov r3,(r3)+
1897 which is a famous insn on the PDP-11 where the value of r3 used as the
1898 source was model-dependent. Avoid this sort of thing. */
1900 #if 0
1901 if (!(GET_CODE (PATTERN (i3)) == SET
1902 && REG_P (SET_SRC (PATTERN (i3)))
1903 && MEM_P (SET_DEST (PATTERN (i3)))
1904 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
1905 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
1906 /* It's not the exception. */
1907 #endif
1908 #ifdef AUTO_INC_DEC
1909 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
1910 if (REG_NOTE_KIND (link) == REG_INC
1911 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
1912 || (i1 != 0
1913 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
1915 undo_all ();
1916 return 0;
1918 #endif
1920 /* See if the SETs in I1 or I2 need to be kept around in the merged
1921 instruction: whenever the value set there is still needed past I3.
1922 For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3.
1924 For the SET in I1, we have two cases: If I1 and I2 independently
1925 feed into I3, the set in I1 needs to be kept around if I1DEST dies
1926 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
1927 in I1 needs to be kept around unless I1DEST dies or is set in either
1928 I2 or I3. We can distinguish these cases by seeing if I2SRC mentions
1929 I1DEST. If so, we know I1 feeds into I2. */
1931 added_sets_2 = ! dead_or_set_p (i3, i2dest);
1933 added_sets_1
1934 = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest)
1935 : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest)));
1937 /* If the set in I2 needs to be kept around, we must make a copy of
1938 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
1939 PATTERN (I2), we are only substituting for the original I1DEST, not into
1940 an already-substituted copy. This also prevents making self-referential
1941 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
1942 I2DEST. */
1944 i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL
1945 ? gen_rtx_SET (VOIDmode, i2dest, i2src)
1946 : PATTERN (i2));
1948 if (added_sets_2)
1949 i2pat = copy_rtx (i2pat);
1951 combine_merges++;
1953 /* Substitute in the latest insn for the regs set by the earlier ones. */
1955 maxreg = max_reg_num ();
1957 subst_insn = i3;
1959 /* It is possible that the source of I2 or I1 may be performing an
1960 unneeded operation, such as a ZERO_EXTEND of something that is known
1961 to have the high part zero. Handle that case by letting subst look at
1962 the innermost one of them.
1964 Another way to do this would be to have a function that tries to
1965 simplify a single insn instead of merging two or more insns. We don't
1966 do this because of the potential of infinite loops and because
1967 of the potential extra memory required. However, doing it the way
1968 we are is a bit of a kludge and doesn't catch all cases.
1970 But only do this if -fexpensive-optimizations since it slows things down
1971 and doesn't usually win. */
1973 if (flag_expensive_optimizations)
1975 /* Pass pc_rtx so no substitutions are done, just simplifications. */
1976 if (i1)
1978 subst_low_cuid = INSN_CUID (i1);
1979 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0);
1981 else
1983 subst_low_cuid = INSN_CUID (i2);
1984 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0);
1988 #ifndef HAVE_cc0
1989 /* Many machines that don't use CC0 have insns that can both perform an
1990 arithmetic operation and set the condition code. These operations will
1991 be represented as a PARALLEL with the first element of the vector
1992 being a COMPARE of an arithmetic operation with the constant zero.
1993 The second element of the vector will set some pseudo to the result
1994 of the same arithmetic operation. If we simplify the COMPARE, we won't
1995 match such a pattern and so will generate an extra insn. Here we test
1996 for this case, where both the comparison and the operation result are
1997 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
1998 I2SRC. Later we will make the PARALLEL that contains I2. */
2000 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
2001 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
2002 && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx
2003 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
2005 #ifdef SELECT_CC_MODE
2006 rtx *cc_use;
2007 enum machine_mode compare_mode;
2008 #endif
2010 newpat = PATTERN (i3);
2011 SUBST (XEXP (SET_SRC (newpat), 0), i2src);
2013 i2_is_used = 1;
2015 #ifdef SELECT_CC_MODE
2016 /* See if a COMPARE with the operand we substituted in should be done
2017 with the mode that is currently being used. If not, do the same
2018 processing we do in `subst' for a SET; namely, if the destination
2019 is used only once, try to replace it with a register of the proper
2020 mode and also replace the COMPARE. */
2021 if (undobuf.other_insn == 0
2022 && (cc_use = find_single_use (SET_DEST (newpat), i3,
2023 &undobuf.other_insn))
2024 && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use),
2025 i2src, const0_rtx))
2026 != GET_MODE (SET_DEST (newpat))))
2028 unsigned int regno = REGNO (SET_DEST (newpat));
2029 rtx new_dest = gen_rtx_REG (compare_mode, regno);
2031 if (regno < FIRST_PSEUDO_REGISTER
2032 || (REG_N_SETS (regno) == 1 && ! added_sets_2
2033 && ! REG_USERVAR_P (SET_DEST (newpat))))
2035 if (regno >= FIRST_PSEUDO_REGISTER)
2036 SUBST (regno_reg_rtx[regno], new_dest);
2038 SUBST (SET_DEST (newpat), new_dest);
2039 SUBST (XEXP (*cc_use, 0), new_dest);
2040 SUBST (SET_SRC (newpat),
2041 gen_rtx_COMPARE (compare_mode, i2src, const0_rtx));
2043 else
2044 undobuf.other_insn = 0;
2046 #endif
2048 else
2049 #endif
2051 n_occurrences = 0; /* `subst' counts here */
2053 /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we
2054 need to make a unique copy of I2SRC each time we substitute it
2055 to avoid self-referential rtl. */
2057 subst_low_cuid = INSN_CUID (i2);
2058 newpat = subst (PATTERN (i3), i2dest, i2src, 0,
2059 ! i1_feeds_i3 && i1dest_in_i1src);
2060 substed_i2 = 1;
2062 /* Record whether i2's body now appears within i3's body. */
2063 i2_is_used = n_occurrences;
2066 /* If we already got a failure, don't try to do more. Otherwise,
2067 try to substitute in I1 if we have it. */
2069 if (i1 && GET_CODE (newpat) != CLOBBER)
2071 /* Before we can do this substitution, we must redo the test done
2072 above (see detailed comments there) that ensures that I1DEST
2073 isn't mentioned in any SETs in NEWPAT that are field assignments. */
2075 if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX,
2076 0, (rtx*) 0))
2078 undo_all ();
2079 return 0;
2082 n_occurrences = 0;
2083 subst_low_cuid = INSN_CUID (i1);
2084 newpat = subst (newpat, i1dest, i1src, 0, 0);
2085 substed_i1 = 1;
2088 /* Fail if an autoincrement side-effect has been duplicated. Be careful
2089 to count all the ways that I2SRC and I1SRC can be used. */
2090 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
2091 && i2_is_used + added_sets_2 > 1)
2092 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
2093 && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3)
2094 > 1))
2095 /* Fail if we tried to make a new register (we used to abort, but there's
2096 really no reason to). */
2097 || max_reg_num () != maxreg
2098 /* Fail if we couldn't do something and have a CLOBBER. */
2099 || GET_CODE (newpat) == CLOBBER
2100 /* Fail if this new pattern is a MULT and we didn't have one before
2101 at the outer level. */
2102 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
2103 && ! have_mult))
2105 undo_all ();
2106 return 0;
2109 /* If the actions of the earlier insns must be kept
2110 in addition to substituting them into the latest one,
2111 we must make a new PARALLEL for the latest insn
2112 to hold additional the SETs. */
2114 if (added_sets_1 || added_sets_2)
2116 combine_extras++;
2118 if (GET_CODE (newpat) == PARALLEL)
2120 rtvec old = XVEC (newpat, 0);
2121 total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2;
2122 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2123 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
2124 sizeof (old->elem[0]) * old->num_elem);
2126 else
2128 rtx old = newpat;
2129 total_sets = 1 + added_sets_1 + added_sets_2;
2130 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2131 XVECEXP (newpat, 0, 0) = old;
2134 if (added_sets_1)
2135 XVECEXP (newpat, 0, --total_sets)
2136 = (GET_CODE (PATTERN (i1)) == PARALLEL
2137 ? gen_rtx_SET (VOIDmode, i1dest, i1src) : PATTERN (i1));
2139 if (added_sets_2)
2141 /* If there is no I1, use I2's body as is. We used to also not do
2142 the subst call below if I2 was substituted into I3,
2143 but that could lose a simplification. */
2144 if (i1 == 0)
2145 XVECEXP (newpat, 0, --total_sets) = i2pat;
2146 else
2147 /* See comment where i2pat is assigned. */
2148 XVECEXP (newpat, 0, --total_sets)
2149 = subst (i2pat, i1dest, i1src, 0, 0);
2153 /* We come here when we are replacing a destination in I2 with the
2154 destination of I3. */
2155 validate_replacement:
2157 /* Note which hard regs this insn has as inputs. */
2158 mark_used_regs_combine (newpat);
2160 /* If recog_for_combine fails, it strips existing clobbers. If we'll
2161 consider splitting this pattern, we might need these clobbers. */
2162 if (i1 && GET_CODE (newpat) == PARALLEL
2163 && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER)
2165 int len = XVECLEN (newpat, 0);
2167 newpat_vec_with_clobbers = rtvec_alloc (len);
2168 for (i = 0; i < len; i++)
2169 RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i);
2172 /* Is the result of combination a valid instruction? */
2173 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2175 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
2176 the second SET's destination is a register that is unused and isn't
2177 marked as an instruction that might trap in an EH region. In that case,
2178 we just need the first SET. This can occur when simplifying a divmod
2179 insn. We *must* test for this case here because the code below that
2180 splits two independent SETs doesn't handle this case correctly when it
2181 updates the register status.
2183 It's pointless doing this if we originally had two sets, one from
2184 i3, and one from i2. Combining then splitting the parallel results
2185 in the original i2 again plus an invalid insn (which we delete).
2186 The net effect is only to move instructions around, which makes
2187 debug info less accurate.
2189 Also check the case where the first SET's destination is unused.
2190 That would not cause incorrect code, but does cause an unneeded
2191 insn to remain. */
2193 if (insn_code_number < 0
2194 && !(added_sets_2 && i1 == 0)
2195 && GET_CODE (newpat) == PARALLEL
2196 && XVECLEN (newpat, 0) == 2
2197 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2198 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2199 && asm_noperands (newpat) < 0)
2201 rtx set0 = XVECEXP (newpat, 0, 0);
2202 rtx set1 = XVECEXP (newpat, 0, 1);
2203 rtx note;
2205 if (((REG_P (SET_DEST (set1))
2206 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
2207 || (GET_CODE (SET_DEST (set1)) == SUBREG
2208 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
2209 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2210 || INTVAL (XEXP (note, 0)) <= 0)
2211 && ! side_effects_p (SET_SRC (set1)))
2213 newpat = set0;
2214 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2217 else if (((REG_P (SET_DEST (set0))
2218 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
2219 || (GET_CODE (SET_DEST (set0)) == SUBREG
2220 && find_reg_note (i3, REG_UNUSED,
2221 SUBREG_REG (SET_DEST (set0)))))
2222 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2223 || INTVAL (XEXP (note, 0)) <= 0)
2224 && ! side_effects_p (SET_SRC (set0)))
2226 newpat = set1;
2227 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2229 if (insn_code_number >= 0)
2231 /* If we will be able to accept this, we have made a
2232 change to the destination of I3. This requires us to
2233 do a few adjustments. */
2235 PATTERN (i3) = newpat;
2236 adjust_for_new_dest (i3);
2241 /* If we were combining three insns and the result is a simple SET
2242 with no ASM_OPERANDS that wasn't recognized, try to split it into two
2243 insns. There are two ways to do this. It can be split using a
2244 machine-specific method (like when you have an addition of a large
2245 constant) or by combine in the function find_split_point. */
2247 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
2248 && asm_noperands (newpat) < 0)
2250 rtx m_split, *split;
2251 rtx ni2dest = i2dest;
2253 /* See if the MD file can split NEWPAT. If it can't, see if letting it
2254 use I2DEST as a scratch register will help. In the latter case,
2255 convert I2DEST to the mode of the source of NEWPAT if we can. */
2257 m_split = split_insns (newpat, i3);
2259 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
2260 inputs of NEWPAT. */
2262 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
2263 possible to try that as a scratch reg. This would require adding
2264 more code to make it work though. */
2266 if (m_split == 0 && ! reg_overlap_mentioned_p (ni2dest, newpat))
2268 /* If I2DEST is a hard register or the only use of a pseudo,
2269 we can change its mode. */
2270 if (GET_MODE (SET_DEST (newpat)) != GET_MODE (i2dest)
2271 && GET_MODE (SET_DEST (newpat)) != VOIDmode
2272 && REG_P (i2dest)
2273 && (REGNO (i2dest) < FIRST_PSEUDO_REGISTER
2274 || (REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2
2275 && ! REG_USERVAR_P (i2dest))))
2276 ni2dest = gen_rtx_REG (GET_MODE (SET_DEST (newpat)),
2277 REGNO (i2dest));
2279 m_split = split_insns (gen_rtx_PARALLEL
2280 (VOIDmode,
2281 gen_rtvec (2, newpat,
2282 gen_rtx_CLOBBER (VOIDmode,
2283 ni2dest))),
2284 i3);
2285 /* If the split with the mode-changed register didn't work, try
2286 the original register. */
2287 if (! m_split && ni2dest != i2dest)
2289 ni2dest = i2dest;
2290 m_split = split_insns (gen_rtx_PARALLEL
2291 (VOIDmode,
2292 gen_rtvec (2, newpat,
2293 gen_rtx_CLOBBER (VOIDmode,
2294 i2dest))),
2295 i3);
2299 /* If recog_for_combine has discarded clobbers, try to use them
2300 again for the split. */
2301 if (m_split == 0 && newpat_vec_with_clobbers)
2302 m_split
2303 = split_insns (gen_rtx_PARALLEL (VOIDmode,
2304 newpat_vec_with_clobbers), i3);
2306 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
2308 m_split = PATTERN (m_split);
2309 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
2310 if (insn_code_number >= 0)
2311 newpat = m_split;
2313 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
2314 && (next_real_insn (i2) == i3
2315 || ! use_crosses_set_p (PATTERN (m_split), INSN_CUID (i2))))
2317 rtx i2set, i3set;
2318 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
2319 newi2pat = PATTERN (m_split);
2321 i3set = single_set (NEXT_INSN (m_split));
2322 i2set = single_set (m_split);
2324 /* In case we changed the mode of I2DEST, replace it in the
2325 pseudo-register table here. We can't do it above in case this
2326 code doesn't get executed and we do a split the other way. */
2328 if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2329 SUBST (regno_reg_rtx[REGNO (i2dest)], ni2dest);
2331 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2333 /* If I2 or I3 has multiple SETs, we won't know how to track
2334 register status, so don't use these insns. If I2's destination
2335 is used between I2 and I3, we also can't use these insns. */
2337 if (i2_code_number >= 0 && i2set && i3set
2338 && (next_real_insn (i2) == i3
2339 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
2340 insn_code_number = recog_for_combine (&newi3pat, i3,
2341 &new_i3_notes);
2342 if (insn_code_number >= 0)
2343 newpat = newi3pat;
2345 /* It is possible that both insns now set the destination of I3.
2346 If so, we must show an extra use of it. */
2348 if (insn_code_number >= 0)
2350 rtx new_i3_dest = SET_DEST (i3set);
2351 rtx new_i2_dest = SET_DEST (i2set);
2353 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
2354 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
2355 || GET_CODE (new_i3_dest) == SUBREG)
2356 new_i3_dest = XEXP (new_i3_dest, 0);
2358 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
2359 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
2360 || GET_CODE (new_i2_dest) == SUBREG)
2361 new_i2_dest = XEXP (new_i2_dest, 0);
2363 if (REG_P (new_i3_dest)
2364 && REG_P (new_i2_dest)
2365 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
2366 REG_N_SETS (REGNO (new_i2_dest))++;
2370 /* If we can split it and use I2DEST, go ahead and see if that
2371 helps things be recognized. Verify that none of the registers
2372 are set between I2 and I3. */
2373 if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0
2374 #ifdef HAVE_cc0
2375 && REG_P (i2dest)
2376 #endif
2377 /* We need I2DEST in the proper mode. If it is a hard register
2378 or the only use of a pseudo, we can change its mode.
2379 Make sure we don't change a hard register to have a mode that
2380 isn't valid for it, or change the number of registers. */
2381 && (GET_MODE (*split) == GET_MODE (i2dest)
2382 || GET_MODE (*split) == VOIDmode
2383 || (REGNO (i2dest) < FIRST_PSEUDO_REGISTER
2384 && HARD_REGNO_MODE_OK (REGNO (i2dest), GET_MODE (*split))
2385 && (HARD_REGNO_NREGS (REGNO (i2dest), GET_MODE (i2dest))
2386 == HARD_REGNO_NREGS (REGNO (i2dest), GET_MODE (*split))))
2387 || (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER
2388 && REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2
2389 && ! REG_USERVAR_P (i2dest)))
2390 && (next_real_insn (i2) == i3
2391 || ! use_crosses_set_p (*split, INSN_CUID (i2)))
2392 /* We can't overwrite I2DEST if its value is still used by
2393 NEWPAT. */
2394 && ! reg_referenced_p (i2dest, newpat))
2396 rtx newdest = i2dest;
2397 enum rtx_code split_code = GET_CODE (*split);
2398 enum machine_mode split_mode = GET_MODE (*split);
2400 /* Get NEWDEST as a register in the proper mode. We have already
2401 validated that we can do this. */
2402 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
2404 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
2406 if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2407 SUBST (regno_reg_rtx[REGNO (i2dest)], newdest);
2410 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
2411 an ASHIFT. This can occur if it was inside a PLUS and hence
2412 appeared to be a memory address. This is a kludge. */
2413 if (split_code == MULT
2414 && GET_CODE (XEXP (*split, 1)) == CONST_INT
2415 && INTVAL (XEXP (*split, 1)) > 0
2416 && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0)
2418 SUBST (*split, gen_rtx_ASHIFT (split_mode,
2419 XEXP (*split, 0), GEN_INT (i)));
2420 /* Update split_code because we may not have a multiply
2421 anymore. */
2422 split_code = GET_CODE (*split);
2425 #ifdef INSN_SCHEDULING
2426 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
2427 be written as a ZERO_EXTEND. */
2428 if (split_code == SUBREG && MEM_P (SUBREG_REG (*split)))
2430 #ifdef LOAD_EXTEND_OP
2431 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
2432 what it really is. */
2433 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
2434 == SIGN_EXTEND)
2435 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
2436 SUBREG_REG (*split)));
2437 else
2438 #endif
2439 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
2440 SUBREG_REG (*split)));
2442 #endif
2444 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
2445 SUBST (*split, newdest);
2446 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2448 /* recog_for_combine might have added CLOBBERs to newi2pat.
2449 Make sure NEWPAT does not depend on the clobbered regs. */
2450 if (GET_CODE (newi2pat) == PARALLEL)
2451 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
2452 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
2454 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
2455 if (reg_overlap_mentioned_p (reg, newpat))
2457 undo_all ();
2458 return 0;
2462 /* If the split point was a MULT and we didn't have one before,
2463 don't use one now. */
2464 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
2465 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2469 /* Check for a case where we loaded from memory in a narrow mode and
2470 then sign extended it, but we need both registers. In that case,
2471 we have a PARALLEL with both loads from the same memory location.
2472 We can split this into a load from memory followed by a register-register
2473 copy. This saves at least one insn, more if register allocation can
2474 eliminate the copy.
2476 We cannot do this if the destination of the first assignment is a
2477 condition code register or cc0. We eliminate this case by making sure
2478 the SET_DEST and SET_SRC have the same mode.
2480 We cannot do this if the destination of the second assignment is
2481 a register that we have already assumed is zero-extended. Similarly
2482 for a SUBREG of such a register. */
2484 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2485 && GET_CODE (newpat) == PARALLEL
2486 && XVECLEN (newpat, 0) == 2
2487 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2488 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
2489 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
2490 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
2491 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2492 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2493 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
2494 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2495 INSN_CUID (i2))
2496 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2497 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2498 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
2499 (REG_P (temp)
2500 && reg_stat[REGNO (temp)].nonzero_bits != 0
2501 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2502 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2503 && (reg_stat[REGNO (temp)].nonzero_bits
2504 != GET_MODE_MASK (word_mode))))
2505 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
2506 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
2507 (REG_P (temp)
2508 && reg_stat[REGNO (temp)].nonzero_bits != 0
2509 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2510 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2511 && (reg_stat[REGNO (temp)].nonzero_bits
2512 != GET_MODE_MASK (word_mode)))))
2513 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2514 SET_SRC (XVECEXP (newpat, 0, 1)))
2515 && ! find_reg_note (i3, REG_UNUSED,
2516 SET_DEST (XVECEXP (newpat, 0, 0))))
2518 rtx ni2dest;
2520 newi2pat = XVECEXP (newpat, 0, 0);
2521 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
2522 newpat = XVECEXP (newpat, 0, 1);
2523 SUBST (SET_SRC (newpat),
2524 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
2525 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2527 if (i2_code_number >= 0)
2528 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2530 if (insn_code_number >= 0)
2531 swap_i2i3 = 1;
2534 /* Similarly, check for a case where we have a PARALLEL of two independent
2535 SETs but we started with three insns. In this case, we can do the sets
2536 as two separate insns. This case occurs when some SET allows two
2537 other insns to combine, but the destination of that SET is still live. */
2539 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2540 && GET_CODE (newpat) == PARALLEL
2541 && XVECLEN (newpat, 0) == 2
2542 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2543 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
2544 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
2545 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2546 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2547 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2548 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2549 INSN_CUID (i2))
2550 /* Don't pass sets with (USE (MEM ...)) dests to the following. */
2551 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != USE
2552 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != USE
2553 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2554 XVECEXP (newpat, 0, 0))
2555 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
2556 XVECEXP (newpat, 0, 1))
2557 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
2558 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1)))))
2560 /* Normally, it doesn't matter which of the two is done first,
2561 but it does if one references cc0. In that case, it has to
2562 be first. */
2563 #ifdef HAVE_cc0
2564 if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0)))
2566 newi2pat = XVECEXP (newpat, 0, 0);
2567 newpat = XVECEXP (newpat, 0, 1);
2569 else
2570 #endif
2572 newi2pat = XVECEXP (newpat, 0, 1);
2573 newpat = XVECEXP (newpat, 0, 0);
2576 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2578 if (i2_code_number >= 0)
2579 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2582 /* If it still isn't recognized, fail and change things back the way they
2583 were. */
2584 if ((insn_code_number < 0
2585 /* Is the result a reasonable ASM_OPERANDS? */
2586 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
2588 undo_all ();
2589 return 0;
2592 /* If we had to change another insn, make sure it is valid also. */
2593 if (undobuf.other_insn)
2595 rtx other_pat = PATTERN (undobuf.other_insn);
2596 rtx new_other_notes;
2597 rtx note, next;
2599 CLEAR_HARD_REG_SET (newpat_used_regs);
2601 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
2602 &new_other_notes);
2604 if (other_code_number < 0 && ! check_asm_operands (other_pat))
2606 undo_all ();
2607 return 0;
2610 PATTERN (undobuf.other_insn) = other_pat;
2612 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
2613 are still valid. Then add any non-duplicate notes added by
2614 recog_for_combine. */
2615 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
2617 next = XEXP (note, 1);
2619 if (REG_NOTE_KIND (note) == REG_UNUSED
2620 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
2622 if (REG_P (XEXP (note, 0)))
2623 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
2625 remove_note (undobuf.other_insn, note);
2629 for (note = new_other_notes; note; note = XEXP (note, 1))
2630 if (REG_P (XEXP (note, 0)))
2631 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
2633 distribute_notes (new_other_notes, undobuf.other_insn,
2634 undobuf.other_insn, NULL_RTX);
2636 #ifdef HAVE_cc0
2637 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
2638 they are adjacent to each other or not. */
2640 rtx p = prev_nonnote_insn (i3);
2641 if (p && p != i2 && NONJUMP_INSN_P (p) && newi2pat
2642 && sets_cc0_p (newi2pat))
2644 undo_all ();
2645 return 0;
2648 #endif
2650 /* Only allow this combination if insn_rtx_costs reports that the
2651 replacement instructions are cheaper than the originals. */
2652 if (!combine_validate_cost (i1, i2, i3, newpat, newi2pat))
2654 undo_all ();
2655 return 0;
2658 /* We now know that we can do this combination. Merge the insns and
2659 update the status of registers and LOG_LINKS. */
2661 if (swap_i2i3)
2663 rtx insn;
2664 rtx link;
2665 rtx ni2dest;
2667 /* I3 now uses what used to be its destination and which is now
2668 I2's destination. This requires us to do a few adjustments. */
2669 PATTERN (i3) = newpat;
2670 adjust_for_new_dest (i3);
2672 /* We need a LOG_LINK from I3 to I2. But we used to have one,
2673 so we still will.
2675 However, some later insn might be using I2's dest and have
2676 a LOG_LINK pointing at I3. We must remove this link.
2677 The simplest way to remove the link is to point it at I1,
2678 which we know will be a NOTE. */
2680 /* newi2pat is usually a SET here; however, recog_for_combine might
2681 have added some clobbers. */
2682 if (GET_CODE (newi2pat) == PARALLEL)
2683 ni2dest = SET_DEST (XVECEXP (newi2pat, 0, 0));
2684 else
2685 ni2dest = SET_DEST (newi2pat);
2687 for (insn = NEXT_INSN (i3);
2688 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
2689 || insn != BB_HEAD (this_basic_block->next_bb));
2690 insn = NEXT_INSN (insn))
2692 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
2694 for (link = LOG_LINKS (insn); link;
2695 link = XEXP (link, 1))
2696 if (XEXP (link, 0) == i3)
2697 XEXP (link, 0) = i1;
2699 break;
2705 rtx i3notes, i2notes, i1notes = 0;
2706 rtx i3links, i2links, i1links = 0;
2707 rtx midnotes = 0;
2708 unsigned int regno;
2710 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
2711 clear them. */
2712 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
2713 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
2714 if (i1)
2715 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
2717 /* Ensure that we do not have something that should not be shared but
2718 occurs multiple times in the new insns. Check this by first
2719 resetting all the `used' flags and then copying anything is shared. */
2721 reset_used_flags (i3notes);
2722 reset_used_flags (i2notes);
2723 reset_used_flags (i1notes);
2724 reset_used_flags (newpat);
2725 reset_used_flags (newi2pat);
2726 if (undobuf.other_insn)
2727 reset_used_flags (PATTERN (undobuf.other_insn));
2729 i3notes = copy_rtx_if_shared (i3notes);
2730 i2notes = copy_rtx_if_shared (i2notes);
2731 i1notes = copy_rtx_if_shared (i1notes);
2732 newpat = copy_rtx_if_shared (newpat);
2733 newi2pat = copy_rtx_if_shared (newi2pat);
2734 if (undobuf.other_insn)
2735 reset_used_flags (PATTERN (undobuf.other_insn));
2737 INSN_CODE (i3) = insn_code_number;
2738 PATTERN (i3) = newpat;
2740 if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3))
2742 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
2744 reset_used_flags (call_usage);
2745 call_usage = copy_rtx (call_usage);
2747 if (substed_i2)
2748 replace_rtx (call_usage, i2dest, i2src);
2750 if (substed_i1)
2751 replace_rtx (call_usage, i1dest, i1src);
2753 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
2756 if (undobuf.other_insn)
2757 INSN_CODE (undobuf.other_insn) = other_code_number;
2759 /* We had one special case above where I2 had more than one set and
2760 we replaced a destination of one of those sets with the destination
2761 of I3. In that case, we have to update LOG_LINKS of insns later
2762 in this basic block. Note that this (expensive) case is rare.
2764 Also, in this case, we must pretend that all REG_NOTEs for I2
2765 actually came from I3, so that REG_UNUSED notes from I2 will be
2766 properly handled. */
2768 if (i3_subst_into_i2)
2770 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
2771 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != USE
2772 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i)))
2773 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
2774 && ! find_reg_note (i2, REG_UNUSED,
2775 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
2776 for (temp = NEXT_INSN (i2);
2777 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
2778 || BB_HEAD (this_basic_block) != temp);
2779 temp = NEXT_INSN (temp))
2780 if (temp != i3 && INSN_P (temp))
2781 for (link = LOG_LINKS (temp); link; link = XEXP (link, 1))
2782 if (XEXP (link, 0) == i2)
2783 XEXP (link, 0) = i3;
2785 if (i3notes)
2787 rtx link = i3notes;
2788 while (XEXP (link, 1))
2789 link = XEXP (link, 1);
2790 XEXP (link, 1) = i2notes;
2792 else
2793 i3notes = i2notes;
2794 i2notes = 0;
2797 LOG_LINKS (i3) = 0;
2798 REG_NOTES (i3) = 0;
2799 LOG_LINKS (i2) = 0;
2800 REG_NOTES (i2) = 0;
2802 if (newi2pat)
2804 INSN_CODE (i2) = i2_code_number;
2805 PATTERN (i2) = newi2pat;
2807 else
2808 SET_INSN_DELETED (i2);
2810 if (i1)
2812 LOG_LINKS (i1) = 0;
2813 REG_NOTES (i1) = 0;
2814 SET_INSN_DELETED (i1);
2817 /* Get death notes for everything that is now used in either I3 or
2818 I2 and used to die in a previous insn. If we built two new
2819 patterns, move from I1 to I2 then I2 to I3 so that we get the
2820 proper movement on registers that I2 modifies. */
2822 if (newi2pat)
2824 move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes);
2825 move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes);
2827 else
2828 move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2),
2829 i3, &midnotes);
2831 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
2832 if (i3notes)
2833 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX);
2834 if (i2notes)
2835 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX);
2836 if (i1notes)
2837 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX);
2838 if (midnotes)
2839 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2841 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
2842 know these are REG_UNUSED and want them to go to the desired insn,
2843 so we always pass it as i3. We have not counted the notes in
2844 reg_n_deaths yet, so we need to do so now. */
2846 if (newi2pat && new_i2_notes)
2848 for (temp = new_i2_notes; temp; temp = XEXP (temp, 1))
2849 if (REG_P (XEXP (temp, 0)))
2850 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
2852 distribute_notes (new_i2_notes, i2, i2, NULL_RTX);
2855 if (new_i3_notes)
2857 for (temp = new_i3_notes; temp; temp = XEXP (temp, 1))
2858 if (REG_P (XEXP (temp, 0)))
2859 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
2861 distribute_notes (new_i3_notes, i3, i3, NULL_RTX);
2864 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
2865 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
2866 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
2867 in that case, it might delete I2. Similarly for I2 and I1.
2868 Show an additional death due to the REG_DEAD note we make here. If
2869 we discard it in distribute_notes, we will decrement it again. */
2871 if (i3dest_killed)
2873 if (REG_P (i3dest_killed))
2874 REG_N_DEATHS (REGNO (i3dest_killed))++;
2876 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
2877 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
2878 NULL_RTX),
2879 NULL_RTX, i2, NULL_RTX);
2880 else
2881 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
2882 NULL_RTX),
2883 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2886 if (i2dest_in_i2src)
2888 if (REG_P (i2dest))
2889 REG_N_DEATHS (REGNO (i2dest))++;
2891 if (newi2pat && reg_set_p (i2dest, newi2pat))
2892 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
2893 NULL_RTX, i2, NULL_RTX);
2894 else
2895 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
2896 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2899 if (i1dest_in_i1src)
2901 if (REG_P (i1dest))
2902 REG_N_DEATHS (REGNO (i1dest))++;
2904 if (newi2pat && reg_set_p (i1dest, newi2pat))
2905 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
2906 NULL_RTX, i2, NULL_RTX);
2907 else
2908 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
2909 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2912 distribute_links (i3links);
2913 distribute_links (i2links);
2914 distribute_links (i1links);
2916 if (REG_P (i2dest))
2918 rtx link;
2919 rtx i2_insn = 0, i2_val = 0, set;
2921 /* The insn that used to set this register doesn't exist, and
2922 this life of the register may not exist either. See if one of
2923 I3's links points to an insn that sets I2DEST. If it does,
2924 that is now the last known value for I2DEST. If we don't update
2925 this and I2 set the register to a value that depended on its old
2926 contents, we will get confused. If this insn is used, thing
2927 will be set correctly in combine_instructions. */
2929 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
2930 if ((set = single_set (XEXP (link, 0))) != 0
2931 && rtx_equal_p (i2dest, SET_DEST (set)))
2932 i2_insn = XEXP (link, 0), i2_val = SET_SRC (set);
2934 record_value_for_reg (i2dest, i2_insn, i2_val);
2936 /* If the reg formerly set in I2 died only once and that was in I3,
2937 zero its use count so it won't make `reload' do any work. */
2938 if (! added_sets_2
2939 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
2940 && ! i2dest_in_i2src)
2942 regno = REGNO (i2dest);
2943 REG_N_SETS (regno)--;
2947 if (i1 && REG_P (i1dest))
2949 rtx link;
2950 rtx i1_insn = 0, i1_val = 0, set;
2952 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
2953 if ((set = single_set (XEXP (link, 0))) != 0
2954 && rtx_equal_p (i1dest, SET_DEST (set)))
2955 i1_insn = XEXP (link, 0), i1_val = SET_SRC (set);
2957 record_value_for_reg (i1dest, i1_insn, i1_val);
2959 regno = REGNO (i1dest);
2960 if (! added_sets_1 && ! i1dest_in_i1src)
2961 REG_N_SETS (regno)--;
2964 /* Update reg_stat[].nonzero_bits et al for any changes that may have
2965 been made to this insn. The order of
2966 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
2967 can affect nonzero_bits of newpat */
2968 if (newi2pat)
2969 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
2970 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
2972 /* Set new_direct_jump_p if a new return or simple jump instruction
2973 has been created.
2975 If I3 is now an unconditional jump, ensure that it has a
2976 BARRIER following it since it may have initially been a
2977 conditional jump. It may also be the last nonnote insn. */
2979 if (returnjump_p (i3) || any_uncondjump_p (i3))
2981 *new_direct_jump_p = 1;
2982 mark_jump_label (PATTERN (i3), i3, 0);
2984 if ((temp = next_nonnote_insn (i3)) == NULL_RTX
2985 || !BARRIER_P (temp))
2986 emit_barrier_after (i3);
2989 if (undobuf.other_insn != NULL_RTX
2990 && (returnjump_p (undobuf.other_insn)
2991 || any_uncondjump_p (undobuf.other_insn)))
2993 *new_direct_jump_p = 1;
2995 if ((temp = next_nonnote_insn (undobuf.other_insn)) == NULL_RTX
2996 || !BARRIER_P (temp))
2997 emit_barrier_after (undobuf.other_insn);
3000 /* An NOOP jump does not need barrier, but it does need cleaning up
3001 of CFG. */
3002 if (GET_CODE (newpat) == SET
3003 && SET_SRC (newpat) == pc_rtx
3004 && SET_DEST (newpat) == pc_rtx)
3005 *new_direct_jump_p = 1;
3008 combine_successes++;
3009 undo_commit ();
3011 if (added_links_insn
3012 && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2))
3013 && INSN_CUID (added_links_insn) < INSN_CUID (i3))
3014 return added_links_insn;
3015 else
3016 return newi2pat ? i2 : i3;
3019 /* Undo all the modifications recorded in undobuf. */
3021 static void
3022 undo_all (void)
3024 struct undo *undo, *next;
3026 for (undo = undobuf.undos; undo; undo = next)
3028 next = undo->next;
3029 if (undo->is_int)
3030 *undo->where.i = undo->old_contents.i;
3031 else
3032 *undo->where.r = undo->old_contents.r;
3034 undo->next = undobuf.frees;
3035 undobuf.frees = undo;
3038 undobuf.undos = 0;
3041 /* We've committed to accepting the changes we made. Move all
3042 of the undos to the free list. */
3044 static void
3045 undo_commit (void)
3047 struct undo *undo, *next;
3049 for (undo = undobuf.undos; undo; undo = next)
3051 next = undo->next;
3052 undo->next = undobuf.frees;
3053 undobuf.frees = undo;
3055 undobuf.undos = 0;
3059 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
3060 where we have an arithmetic expression and return that point. LOC will
3061 be inside INSN.
3063 try_combine will call this function to see if an insn can be split into
3064 two insns. */
3066 static rtx *
3067 find_split_point (rtx *loc, rtx insn)
3069 rtx x = *loc;
3070 enum rtx_code code = GET_CODE (x);
3071 rtx *split;
3072 unsigned HOST_WIDE_INT len = 0;
3073 HOST_WIDE_INT pos = 0;
3074 int unsignedp = 0;
3075 rtx inner = NULL_RTX;
3077 /* First special-case some codes. */
3078 switch (code)
3080 case SUBREG:
3081 #ifdef INSN_SCHEDULING
3082 /* If we are making a paradoxical SUBREG invalid, it becomes a split
3083 point. */
3084 if (MEM_P (SUBREG_REG (x)))
3085 return loc;
3086 #endif
3087 return find_split_point (&SUBREG_REG (x), insn);
3089 case MEM:
3090 #ifdef HAVE_lo_sum
3091 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
3092 using LO_SUM and HIGH. */
3093 if (GET_CODE (XEXP (x, 0)) == CONST
3094 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
3096 SUBST (XEXP (x, 0),
3097 gen_rtx_LO_SUM (Pmode,
3098 gen_rtx_HIGH (Pmode, XEXP (x, 0)),
3099 XEXP (x, 0)));
3100 return &XEXP (XEXP (x, 0), 0);
3102 #endif
3104 /* If we have a PLUS whose second operand is a constant and the
3105 address is not valid, perhaps will can split it up using
3106 the machine-specific way to split large constants. We use
3107 the first pseudo-reg (one of the virtual regs) as a placeholder;
3108 it will not remain in the result. */
3109 if (GET_CODE (XEXP (x, 0)) == PLUS
3110 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3111 && ! memory_address_p (GET_MODE (x), XEXP (x, 0)))
3113 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
3114 rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)),
3115 subst_insn);
3117 /* This should have produced two insns, each of which sets our
3118 placeholder. If the source of the second is a valid address,
3119 we can make put both sources together and make a split point
3120 in the middle. */
3122 if (seq
3123 && NEXT_INSN (seq) != NULL_RTX
3124 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
3125 && NONJUMP_INSN_P (seq)
3126 && GET_CODE (PATTERN (seq)) == SET
3127 && SET_DEST (PATTERN (seq)) == reg
3128 && ! reg_mentioned_p (reg,
3129 SET_SRC (PATTERN (seq)))
3130 && NONJUMP_INSN_P (NEXT_INSN (seq))
3131 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
3132 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
3133 && memory_address_p (GET_MODE (x),
3134 SET_SRC (PATTERN (NEXT_INSN (seq)))))
3136 rtx src1 = SET_SRC (PATTERN (seq));
3137 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
3139 /* Replace the placeholder in SRC2 with SRC1. If we can
3140 find where in SRC2 it was placed, that can become our
3141 split point and we can replace this address with SRC2.
3142 Just try two obvious places. */
3144 src2 = replace_rtx (src2, reg, src1);
3145 split = 0;
3146 if (XEXP (src2, 0) == src1)
3147 split = &XEXP (src2, 0);
3148 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
3149 && XEXP (XEXP (src2, 0), 0) == src1)
3150 split = &XEXP (XEXP (src2, 0), 0);
3152 if (split)
3154 SUBST (XEXP (x, 0), src2);
3155 return split;
3159 /* If that didn't work, perhaps the first operand is complex and
3160 needs to be computed separately, so make a split point there.
3161 This will occur on machines that just support REG + CONST
3162 and have a constant moved through some previous computation. */
3164 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
3165 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
3166 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
3167 return &XEXP (XEXP (x, 0), 0);
3169 break;
3171 case SET:
3172 #ifdef HAVE_cc0
3173 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
3174 ZERO_EXTRACT, the most likely reason why this doesn't match is that
3175 we need to put the operand into a register. So split at that
3176 point. */
3178 if (SET_DEST (x) == cc0_rtx
3179 && GET_CODE (SET_SRC (x)) != COMPARE
3180 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
3181 && !OBJECT_P (SET_SRC (x))
3182 && ! (GET_CODE (SET_SRC (x)) == SUBREG
3183 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
3184 return &SET_SRC (x);
3185 #endif
3187 /* See if we can split SET_SRC as it stands. */
3188 split = find_split_point (&SET_SRC (x), insn);
3189 if (split && split != &SET_SRC (x))
3190 return split;
3192 /* See if we can split SET_DEST as it stands. */
3193 split = find_split_point (&SET_DEST (x), insn);
3194 if (split && split != &SET_DEST (x))
3195 return split;
3197 /* See if this is a bitfield assignment with everything constant. If
3198 so, this is an IOR of an AND, so split it into that. */
3199 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
3200 && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))
3201 <= HOST_BITS_PER_WIDE_INT)
3202 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT
3203 && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT
3204 && GET_CODE (SET_SRC (x)) == CONST_INT
3205 && ((INTVAL (XEXP (SET_DEST (x), 1))
3206 + INTVAL (XEXP (SET_DEST (x), 2)))
3207 <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))))
3208 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
3210 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
3211 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
3212 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
3213 rtx dest = XEXP (SET_DEST (x), 0);
3214 enum machine_mode mode = GET_MODE (dest);
3215 unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1;
3217 if (BITS_BIG_ENDIAN)
3218 pos = GET_MODE_BITSIZE (mode) - len - pos;
3220 if (src == mask)
3221 SUBST (SET_SRC (x),
3222 simplify_gen_binary (IOR, mode, dest, GEN_INT (src << pos)));
3223 else
3225 rtx negmask = gen_int_mode (~(mask << pos), mode);
3226 SUBST (SET_SRC (x),
3227 simplify_gen_binary (IOR, mode,
3228 simplify_gen_binary (AND, mode,
3229 dest, negmask),
3230 GEN_INT (src << pos)));
3233 SUBST (SET_DEST (x), dest);
3235 split = find_split_point (&SET_SRC (x), insn);
3236 if (split && split != &SET_SRC (x))
3237 return split;
3240 /* Otherwise, see if this is an operation that we can split into two.
3241 If so, try to split that. */
3242 code = GET_CODE (SET_SRC (x));
3244 switch (code)
3246 case AND:
3247 /* If we are AND'ing with a large constant that is only a single
3248 bit and the result is only being used in a context where we
3249 need to know if it is zero or nonzero, replace it with a bit
3250 extraction. This will avoid the large constant, which might
3251 have taken more than one insn to make. If the constant were
3252 not a valid argument to the AND but took only one insn to make,
3253 this is no worse, but if it took more than one insn, it will
3254 be better. */
3256 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3257 && REG_P (XEXP (SET_SRC (x), 0))
3258 && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7
3259 && REG_P (SET_DEST (x))
3260 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
3261 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
3262 && XEXP (*split, 0) == SET_DEST (x)
3263 && XEXP (*split, 1) == const0_rtx)
3265 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
3266 XEXP (SET_SRC (x), 0),
3267 pos, NULL_RTX, 1, 1, 0, 0);
3268 if (extraction != 0)
3270 SUBST (SET_SRC (x), extraction);
3271 return find_split_point (loc, insn);
3274 break;
3276 case NE:
3277 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
3278 is known to be on, this can be converted into a NEG of a shift. */
3279 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
3280 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
3281 && 1 <= (pos = exact_log2
3282 (nonzero_bits (XEXP (SET_SRC (x), 0),
3283 GET_MODE (XEXP (SET_SRC (x), 0))))))
3285 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
3287 SUBST (SET_SRC (x),
3288 gen_rtx_NEG (mode,
3289 gen_rtx_LSHIFTRT (mode,
3290 XEXP (SET_SRC (x), 0),
3291 GEN_INT (pos))));
3293 split = find_split_point (&SET_SRC (x), insn);
3294 if (split && split != &SET_SRC (x))
3295 return split;
3297 break;
3299 case SIGN_EXTEND:
3300 inner = XEXP (SET_SRC (x), 0);
3302 /* We can't optimize if either mode is a partial integer
3303 mode as we don't know how many bits are significant
3304 in those modes. */
3305 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
3306 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
3307 break;
3309 pos = 0;
3310 len = GET_MODE_BITSIZE (GET_MODE (inner));
3311 unsignedp = 0;
3312 break;
3314 case SIGN_EXTRACT:
3315 case ZERO_EXTRACT:
3316 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3317 && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT)
3319 inner = XEXP (SET_SRC (x), 0);
3320 len = INTVAL (XEXP (SET_SRC (x), 1));
3321 pos = INTVAL (XEXP (SET_SRC (x), 2));
3323 if (BITS_BIG_ENDIAN)
3324 pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos;
3325 unsignedp = (code == ZERO_EXTRACT);
3327 break;
3329 default:
3330 break;
3333 if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner)))
3335 enum machine_mode mode = GET_MODE (SET_SRC (x));
3337 /* For unsigned, we have a choice of a shift followed by an
3338 AND or two shifts. Use two shifts for field sizes where the
3339 constant might be too large. We assume here that we can
3340 always at least get 8-bit constants in an AND insn, which is
3341 true for every current RISC. */
3343 if (unsignedp && len <= 8)
3345 SUBST (SET_SRC (x),
3346 gen_rtx_AND (mode,
3347 gen_rtx_LSHIFTRT
3348 (mode, gen_lowpart (mode, inner),
3349 GEN_INT (pos)),
3350 GEN_INT (((HOST_WIDE_INT) 1 << len) - 1)));
3352 split = find_split_point (&SET_SRC (x), insn);
3353 if (split && split != &SET_SRC (x))
3354 return split;
3356 else
3358 SUBST (SET_SRC (x),
3359 gen_rtx_fmt_ee
3360 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
3361 gen_rtx_ASHIFT (mode,
3362 gen_lowpart (mode, inner),
3363 GEN_INT (GET_MODE_BITSIZE (mode)
3364 - len - pos)),
3365 GEN_INT (GET_MODE_BITSIZE (mode) - len)));
3367 split = find_split_point (&SET_SRC (x), insn);
3368 if (split && split != &SET_SRC (x))
3369 return split;
3373 /* See if this is a simple operation with a constant as the second
3374 operand. It might be that this constant is out of range and hence
3375 could be used as a split point. */
3376 if (BINARY_P (SET_SRC (x))
3377 && CONSTANT_P (XEXP (SET_SRC (x), 1))
3378 && (OBJECT_P (XEXP (SET_SRC (x), 0))
3379 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
3380 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
3381 return &XEXP (SET_SRC (x), 1);
3383 /* Finally, see if this is a simple operation with its first operand
3384 not in a register. The operation might require this operand in a
3385 register, so return it as a split point. We can always do this
3386 because if the first operand were another operation, we would have
3387 already found it as a split point. */
3388 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
3389 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
3390 return &XEXP (SET_SRC (x), 0);
3392 return 0;
3394 case AND:
3395 case IOR:
3396 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
3397 it is better to write this as (not (ior A B)) so we can split it.
3398 Similarly for IOR. */
3399 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
3401 SUBST (*loc,
3402 gen_rtx_NOT (GET_MODE (x),
3403 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
3404 GET_MODE (x),
3405 XEXP (XEXP (x, 0), 0),
3406 XEXP (XEXP (x, 1), 0))));
3407 return find_split_point (loc, insn);
3410 /* Many RISC machines have a large set of logical insns. If the
3411 second operand is a NOT, put it first so we will try to split the
3412 other operand first. */
3413 if (GET_CODE (XEXP (x, 1)) == NOT)
3415 rtx tem = XEXP (x, 0);
3416 SUBST (XEXP (x, 0), XEXP (x, 1));
3417 SUBST (XEXP (x, 1), tem);
3419 break;
3421 default:
3422 break;
3425 /* Otherwise, select our actions depending on our rtx class. */
3426 switch (GET_RTX_CLASS (code))
3428 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
3429 case RTX_TERNARY:
3430 split = find_split_point (&XEXP (x, 2), insn);
3431 if (split)
3432 return split;
3433 /* ... fall through ... */
3434 case RTX_BIN_ARITH:
3435 case RTX_COMM_ARITH:
3436 case RTX_COMPARE:
3437 case RTX_COMM_COMPARE:
3438 split = find_split_point (&XEXP (x, 1), insn);
3439 if (split)
3440 return split;
3441 /* ... fall through ... */
3442 case RTX_UNARY:
3443 /* Some machines have (and (shift ...) ...) insns. If X is not
3444 an AND, but XEXP (X, 0) is, use it as our split point. */
3445 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
3446 return &XEXP (x, 0);
3448 split = find_split_point (&XEXP (x, 0), insn);
3449 if (split)
3450 return split;
3451 return loc;
3453 default:
3454 /* Otherwise, we don't have a split point. */
3455 return 0;
3459 /* Throughout X, replace FROM with TO, and return the result.
3460 The result is TO if X is FROM;
3461 otherwise the result is X, but its contents may have been modified.
3462 If they were modified, a record was made in undobuf so that
3463 undo_all will (among other things) return X to its original state.
3465 If the number of changes necessary is too much to record to undo,
3466 the excess changes are not made, so the result is invalid.
3467 The changes already made can still be undone.
3468 undobuf.num_undo is incremented for such changes, so by testing that
3469 the caller can tell whether the result is valid.
3471 `n_occurrences' is incremented each time FROM is replaced.
3473 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
3475 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
3476 by copying if `n_occurrences' is nonzero. */
3478 static rtx
3479 subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy)
3481 enum rtx_code code = GET_CODE (x);
3482 enum machine_mode op0_mode = VOIDmode;
3483 const char *fmt;
3484 int len, i;
3485 rtx new;
3487 /* Two expressions are equal if they are identical copies of a shared
3488 RTX or if they are both registers with the same register number
3489 and mode. */
3491 #define COMBINE_RTX_EQUAL_P(X,Y) \
3492 ((X) == (Y) \
3493 || (REG_P (X) && REG_P (Y) \
3494 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
3496 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
3498 n_occurrences++;
3499 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
3502 /* If X and FROM are the same register but different modes, they will
3503 not have been seen as equal above. However, flow.c will make a
3504 LOG_LINKS entry for that case. If we do nothing, we will try to
3505 rerecognize our original insn and, when it succeeds, we will
3506 delete the feeding insn, which is incorrect.
3508 So force this insn not to match in this (rare) case. */
3509 if (! in_dest && code == REG && REG_P (from)
3510 && REGNO (x) == REGNO (from))
3511 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
3513 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
3514 of which may contain things that can be combined. */
3515 if (code != MEM && code != LO_SUM && OBJECT_P (x))
3516 return x;
3518 /* It is possible to have a subexpression appear twice in the insn.
3519 Suppose that FROM is a register that appears within TO.
3520 Then, after that subexpression has been scanned once by `subst',
3521 the second time it is scanned, TO may be found. If we were
3522 to scan TO here, we would find FROM within it and create a
3523 self-referent rtl structure which is completely wrong. */
3524 if (COMBINE_RTX_EQUAL_P (x, to))
3525 return to;
3527 /* Parallel asm_operands need special attention because all of the
3528 inputs are shared across the arms. Furthermore, unsharing the
3529 rtl results in recognition failures. Failure to handle this case
3530 specially can result in circular rtl.
3532 Solve this by doing a normal pass across the first entry of the
3533 parallel, and only processing the SET_DESTs of the subsequent
3534 entries. Ug. */
3536 if (code == PARALLEL
3537 && GET_CODE (XVECEXP (x, 0, 0)) == SET
3538 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
3540 new = subst (XVECEXP (x, 0, 0), from, to, 0, unique_copy);
3542 /* If this substitution failed, this whole thing fails. */
3543 if (GET_CODE (new) == CLOBBER
3544 && XEXP (new, 0) == const0_rtx)
3545 return new;
3547 SUBST (XVECEXP (x, 0, 0), new);
3549 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
3551 rtx dest = SET_DEST (XVECEXP (x, 0, i));
3553 if (!REG_P (dest)
3554 && GET_CODE (dest) != CC0
3555 && GET_CODE (dest) != PC)
3557 new = subst (dest, from, to, 0, unique_copy);
3559 /* If this substitution failed, this whole thing fails. */
3560 if (GET_CODE (new) == CLOBBER
3561 && XEXP (new, 0) == const0_rtx)
3562 return new;
3564 SUBST (SET_DEST (XVECEXP (x, 0, i)), new);
3568 else
3570 len = GET_RTX_LENGTH (code);
3571 fmt = GET_RTX_FORMAT (code);
3573 /* We don't need to process a SET_DEST that is a register, CC0,
3574 or PC, so set up to skip this common case. All other cases
3575 where we want to suppress replacing something inside a
3576 SET_SRC are handled via the IN_DEST operand. */
3577 if (code == SET
3578 && (REG_P (SET_DEST (x))
3579 || GET_CODE (SET_DEST (x)) == CC0
3580 || GET_CODE (SET_DEST (x)) == PC))
3581 fmt = "ie";
3583 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
3584 constant. */
3585 if (fmt[0] == 'e')
3586 op0_mode = GET_MODE (XEXP (x, 0));
3588 for (i = 0; i < len; i++)
3590 if (fmt[i] == 'E')
3592 int j;
3593 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3595 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
3597 new = (unique_copy && n_occurrences
3598 ? copy_rtx (to) : to);
3599 n_occurrences++;
3601 else
3603 new = subst (XVECEXP (x, i, j), from, to, 0,
3604 unique_copy);
3606 /* If this substitution failed, this whole thing
3607 fails. */
3608 if (GET_CODE (new) == CLOBBER
3609 && XEXP (new, 0) == const0_rtx)
3610 return new;
3613 SUBST (XVECEXP (x, i, j), new);
3616 else if (fmt[i] == 'e')
3618 /* If this is a register being set, ignore it. */
3619 new = XEXP (x, i);
3620 if (in_dest
3621 && i == 0
3622 && (((code == SUBREG || code == ZERO_EXTRACT)
3623 && REG_P (new))
3624 || code == STRICT_LOW_PART))
3627 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
3629 /* In general, don't install a subreg involving two
3630 modes not tieable. It can worsen register
3631 allocation, and can even make invalid reload
3632 insns, since the reg inside may need to be copied
3633 from in the outside mode, and that may be invalid
3634 if it is an fp reg copied in integer mode.
3636 We allow two exceptions to this: It is valid if
3637 it is inside another SUBREG and the mode of that
3638 SUBREG and the mode of the inside of TO is
3639 tieable and it is valid if X is a SET that copies
3640 FROM to CC0. */
3642 if (GET_CODE (to) == SUBREG
3643 && ! MODES_TIEABLE_P (GET_MODE (to),
3644 GET_MODE (SUBREG_REG (to)))
3645 && ! (code == SUBREG
3646 && MODES_TIEABLE_P (GET_MODE (x),
3647 GET_MODE (SUBREG_REG (to))))
3648 #ifdef HAVE_cc0
3649 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
3650 #endif
3652 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
3654 #ifdef CANNOT_CHANGE_MODE_CLASS
3655 if (code == SUBREG
3656 && REG_P (to)
3657 && REGNO (to) < FIRST_PSEUDO_REGISTER
3658 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
3659 GET_MODE (to),
3660 GET_MODE (x)))
3661 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
3662 #endif
3664 new = (unique_copy && n_occurrences ? copy_rtx (to) : to);
3665 n_occurrences++;
3667 else
3668 /* If we are in a SET_DEST, suppress most cases unless we
3669 have gone inside a MEM, in which case we want to
3670 simplify the address. We assume here that things that
3671 are actually part of the destination have their inner
3672 parts in the first expression. This is true for SUBREG,
3673 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
3674 things aside from REG and MEM that should appear in a
3675 SET_DEST. */
3676 new = subst (XEXP (x, i), from, to,
3677 (((in_dest
3678 && (code == SUBREG || code == STRICT_LOW_PART
3679 || code == ZERO_EXTRACT))
3680 || code == SET)
3681 && i == 0), unique_copy);
3683 /* If we found that we will have to reject this combination,
3684 indicate that by returning the CLOBBER ourselves, rather than
3685 an expression containing it. This will speed things up as
3686 well as prevent accidents where two CLOBBERs are considered
3687 to be equal, thus producing an incorrect simplification. */
3689 if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx)
3690 return new;
3692 if (GET_CODE (x) == SUBREG
3693 && (GET_CODE (new) == CONST_INT
3694 || GET_CODE (new) == CONST_DOUBLE))
3696 enum machine_mode mode = GET_MODE (x);
3698 x = simplify_subreg (GET_MODE (x), new,
3699 GET_MODE (SUBREG_REG (x)),
3700 SUBREG_BYTE (x));
3701 if (! x)
3702 x = gen_rtx_CLOBBER (mode, const0_rtx);
3704 else if (GET_CODE (new) == CONST_INT
3705 && GET_CODE (x) == ZERO_EXTEND)
3707 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
3708 new, GET_MODE (XEXP (x, 0)));
3709 gcc_assert (x);
3711 else
3712 SUBST (XEXP (x, i), new);
3717 /* Try to simplify X. If the simplification changed the code, it is likely
3718 that further simplification will help, so loop, but limit the number
3719 of repetitions that will be performed. */
3721 for (i = 0; i < 4; i++)
3723 /* If X is sufficiently simple, don't bother trying to do anything
3724 with it. */
3725 if (code != CONST_INT && code != REG && code != CLOBBER)
3726 x = combine_simplify_rtx (x, op0_mode, in_dest);
3728 if (GET_CODE (x) == code)
3729 break;
3731 code = GET_CODE (x);
3733 /* We no longer know the original mode of operand 0 since we
3734 have changed the form of X) */
3735 op0_mode = VOIDmode;
3738 return x;
3741 /* Simplify X, a piece of RTL. We just operate on the expression at the
3742 outer level; call `subst' to simplify recursively. Return the new
3743 expression.
3745 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
3746 if we are inside a SET_DEST. */
3748 static rtx
3749 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest)
3751 enum rtx_code code = GET_CODE (x);
3752 enum machine_mode mode = GET_MODE (x);
3753 rtx temp;
3754 rtx reversed;
3755 int i;
3757 /* If this is a commutative operation, put a constant last and a complex
3758 expression first. We don't need to do this for comparisons here. */
3759 if (COMMUTATIVE_ARITH_P (x)
3760 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
3762 temp = XEXP (x, 0);
3763 SUBST (XEXP (x, 0), XEXP (x, 1));
3764 SUBST (XEXP (x, 1), temp);
3767 /* If this is a PLUS, MINUS, or MULT, and the first operand is the
3768 sign extension of a PLUS with a constant, reverse the order of the sign
3769 extension and the addition. Note that this not the same as the original
3770 code, but overflow is undefined for signed values. Also note that the
3771 PLUS will have been partially moved "inside" the sign-extension, so that
3772 the first operand of X will really look like:
3773 (ashiftrt (plus (ashift A C4) C5) C4).
3774 We convert this to
3775 (plus (ashiftrt (ashift A C4) C2) C4)
3776 and replace the first operand of X with that expression. Later parts
3777 of this function may simplify the expression further.
3779 For example, if we start with (mult (sign_extend (plus A C1)) C2),
3780 we swap the SIGN_EXTEND and PLUS. Later code will apply the
3781 distributive law to produce (plus (mult (sign_extend X) C1) C3).
3783 We do this to simplify address expressions. */
3785 if ((code == PLUS || code == MINUS || code == MULT)
3786 && GET_CODE (XEXP (x, 0)) == ASHIFTRT
3787 && GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS
3788 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ASHIFT
3789 && GET_CODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1)) == CONST_INT
3790 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3791 && XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1) == XEXP (XEXP (x, 0), 1)
3792 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
3793 && (temp = simplify_binary_operation (ASHIFTRT, mode,
3794 XEXP (XEXP (XEXP (x, 0), 0), 1),
3795 XEXP (XEXP (x, 0), 1))) != 0)
3797 rtx new
3798 = simplify_shift_const (NULL_RTX, ASHIFT, mode,
3799 XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0),
3800 INTVAL (XEXP (XEXP (x, 0), 1)));
3802 new = simplify_shift_const (NULL_RTX, ASHIFTRT, mode, new,
3803 INTVAL (XEXP (XEXP (x, 0), 1)));
3805 SUBST (XEXP (x, 0), simplify_gen_binary (PLUS, mode, new, temp));
3808 /* If this is a simple operation applied to an IF_THEN_ELSE, try
3809 applying it to the arms of the IF_THEN_ELSE. This often simplifies
3810 things. Check for cases where both arms are testing the same
3811 condition.
3813 Don't do anything if all operands are very simple. */
3815 if ((BINARY_P (x)
3816 && ((!OBJECT_P (XEXP (x, 0))
3817 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
3818 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
3819 || (!OBJECT_P (XEXP (x, 1))
3820 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
3821 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
3822 || (UNARY_P (x)
3823 && (!OBJECT_P (XEXP (x, 0))
3824 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
3825 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
3827 rtx cond, true_rtx, false_rtx;
3829 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
3830 if (cond != 0
3831 /* If everything is a comparison, what we have is highly unlikely
3832 to be simpler, so don't use it. */
3833 && ! (COMPARISON_P (x)
3834 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
3836 rtx cop1 = const0_rtx;
3837 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
3839 if (cond_code == NE && COMPARISON_P (cond))
3840 return x;
3842 /* Simplify the alternative arms; this may collapse the true and
3843 false arms to store-flag values. Be careful to use copy_rtx
3844 here since true_rtx or false_rtx might share RTL with x as a
3845 result of the if_then_else_cond call above. */
3846 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0);
3847 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0);
3849 /* If true_rtx and false_rtx are not general_operands, an if_then_else
3850 is unlikely to be simpler. */
3851 if (general_operand (true_rtx, VOIDmode)
3852 && general_operand (false_rtx, VOIDmode))
3854 enum rtx_code reversed;
3856 /* Restarting if we generate a store-flag expression will cause
3857 us to loop. Just drop through in this case. */
3859 /* If the result values are STORE_FLAG_VALUE and zero, we can
3860 just make the comparison operation. */
3861 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
3862 x = simplify_gen_relational (cond_code, mode, VOIDmode,
3863 cond, cop1);
3864 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
3865 && ((reversed = reversed_comparison_code_parts
3866 (cond_code, cond, cop1, NULL))
3867 != UNKNOWN))
3868 x = simplify_gen_relational (reversed, mode, VOIDmode,
3869 cond, cop1);
3871 /* Likewise, we can make the negate of a comparison operation
3872 if the result values are - STORE_FLAG_VALUE and zero. */
3873 else if (GET_CODE (true_rtx) == CONST_INT
3874 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
3875 && false_rtx == const0_rtx)
3876 x = simplify_gen_unary (NEG, mode,
3877 simplify_gen_relational (cond_code,
3878 mode, VOIDmode,
3879 cond, cop1),
3880 mode);
3881 else if (GET_CODE (false_rtx) == CONST_INT
3882 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
3883 && true_rtx == const0_rtx
3884 && ((reversed = reversed_comparison_code_parts
3885 (cond_code, cond, cop1, NULL))
3886 != UNKNOWN))
3887 x = simplify_gen_unary (NEG, mode,
3888 simplify_gen_relational (reversed,
3889 mode, VOIDmode,
3890 cond, cop1),
3891 mode);
3892 else
3893 return gen_rtx_IF_THEN_ELSE (mode,
3894 simplify_gen_relational (cond_code,
3895 mode,
3896 VOIDmode,
3897 cond,
3898 cop1),
3899 true_rtx, false_rtx);
3901 code = GET_CODE (x);
3902 op0_mode = VOIDmode;
3907 /* Try to fold this expression in case we have constants that weren't
3908 present before. */
3909 temp = 0;
3910 switch (GET_RTX_CLASS (code))
3912 case RTX_UNARY:
3913 if (op0_mode == VOIDmode)
3914 op0_mode = GET_MODE (XEXP (x, 0));
3915 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
3916 break;
3917 case RTX_COMPARE:
3918 case RTX_COMM_COMPARE:
3920 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
3921 if (cmp_mode == VOIDmode)
3923 cmp_mode = GET_MODE (XEXP (x, 1));
3924 if (cmp_mode == VOIDmode)
3925 cmp_mode = op0_mode;
3927 temp = simplify_relational_operation (code, mode, cmp_mode,
3928 XEXP (x, 0), XEXP (x, 1));
3930 break;
3931 case RTX_COMM_ARITH:
3932 case RTX_BIN_ARITH:
3933 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
3934 break;
3935 case RTX_BITFIELD_OPS:
3936 case RTX_TERNARY:
3937 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
3938 XEXP (x, 1), XEXP (x, 2));
3939 break;
3940 default:
3941 break;
3944 if (temp)
3946 x = temp;
3947 code = GET_CODE (temp);
3948 op0_mode = VOIDmode;
3949 mode = GET_MODE (temp);
3952 /* First see if we can apply the inverse distributive law. */
3953 if (code == PLUS || code == MINUS
3954 || code == AND || code == IOR || code == XOR)
3956 x = apply_distributive_law (x);
3957 code = GET_CODE (x);
3958 op0_mode = VOIDmode;
3961 /* If CODE is an associative operation not otherwise handled, see if we
3962 can associate some operands. This can win if they are constants or
3963 if they are logically related (i.e. (a & b) & a). */
3964 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
3965 || code == AND || code == IOR || code == XOR
3966 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
3967 && ((INTEGRAL_MODE_P (mode) && code != DIV)
3968 || (flag_unsafe_math_optimizations && FLOAT_MODE_P (mode))))
3970 if (GET_CODE (XEXP (x, 0)) == code)
3972 rtx other = XEXP (XEXP (x, 0), 0);
3973 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
3974 rtx inner_op1 = XEXP (x, 1);
3975 rtx inner;
3977 /* Make sure we pass the constant operand if any as the second
3978 one if this is a commutative operation. */
3979 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
3981 rtx tem = inner_op0;
3982 inner_op0 = inner_op1;
3983 inner_op1 = tem;
3985 inner = simplify_binary_operation (code == MINUS ? PLUS
3986 : code == DIV ? MULT
3987 : code,
3988 mode, inner_op0, inner_op1);
3990 /* For commutative operations, try the other pair if that one
3991 didn't simplify. */
3992 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
3994 other = XEXP (XEXP (x, 0), 1);
3995 inner = simplify_binary_operation (code, mode,
3996 XEXP (XEXP (x, 0), 0),
3997 XEXP (x, 1));
4000 if (inner)
4001 return simplify_gen_binary (code, mode, other, inner);
4005 /* A little bit of algebraic simplification here. */
4006 switch (code)
4008 case MEM:
4009 /* Ensure that our address has any ASHIFTs converted to MULT in case
4010 address-recognizing predicates are called later. */
4011 temp = make_compound_operation (XEXP (x, 0), MEM);
4012 SUBST (XEXP (x, 0), temp);
4013 break;
4015 case SUBREG:
4016 if (op0_mode == VOIDmode)
4017 op0_mode = GET_MODE (SUBREG_REG (x));
4019 /* See if this can be moved to simplify_subreg. */
4020 if (CONSTANT_P (SUBREG_REG (x))
4021 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
4022 /* Don't call gen_lowpart if the inner mode
4023 is VOIDmode and we cannot simplify it, as SUBREG without
4024 inner mode is invalid. */
4025 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
4026 || gen_lowpart_common (mode, SUBREG_REG (x))))
4027 return gen_lowpart (mode, SUBREG_REG (x));
4029 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
4030 break;
4032 rtx temp;
4033 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,
4034 SUBREG_BYTE (x));
4035 if (temp)
4036 return temp;
4039 /* Don't change the mode of the MEM if that would change the meaning
4040 of the address. */
4041 if (MEM_P (SUBREG_REG (x))
4042 && (MEM_VOLATILE_P (SUBREG_REG (x))
4043 || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0))))
4044 return gen_rtx_CLOBBER (mode, const0_rtx);
4046 /* Note that we cannot do any narrowing for non-constants since
4047 we might have been counting on using the fact that some bits were
4048 zero. We now do this in the SET. */
4050 break;
4052 case NOT:
4053 if (GET_CODE (XEXP (x, 0)) == SUBREG
4054 && subreg_lowpart_p (XEXP (x, 0))
4055 && (GET_MODE_SIZE (GET_MODE (XEXP (x, 0)))
4056 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (x, 0)))))
4057 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == ASHIFT
4058 && XEXP (SUBREG_REG (XEXP (x, 0)), 0) == const1_rtx)
4060 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (XEXP (x, 0)));
4062 x = gen_rtx_ROTATE (inner_mode,
4063 simplify_gen_unary (NOT, inner_mode, const1_rtx,
4064 inner_mode),
4065 XEXP (SUBREG_REG (XEXP (x, 0)), 1));
4066 return gen_lowpart (mode, x);
4069 /* Apply De Morgan's laws to reduce number of patterns for machines
4070 with negating logical insns (and-not, nand, etc.). If result has
4071 only one NOT, put it first, since that is how the patterns are
4072 coded. */
4074 if (GET_CODE (XEXP (x, 0)) == IOR || GET_CODE (XEXP (x, 0)) == AND)
4076 rtx in1 = XEXP (XEXP (x, 0), 0), in2 = XEXP (XEXP (x, 0), 1);
4077 enum machine_mode op_mode;
4079 op_mode = GET_MODE (in1);
4080 in1 = simplify_gen_unary (NOT, op_mode, in1, op_mode);
4082 op_mode = GET_MODE (in2);
4083 if (op_mode == VOIDmode)
4084 op_mode = mode;
4085 in2 = simplify_gen_unary (NOT, op_mode, in2, op_mode);
4087 if (GET_CODE (in2) == NOT && GET_CODE (in1) != NOT)
4089 rtx tem = in2;
4090 in2 = in1; in1 = tem;
4093 return gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)) == IOR ? AND : IOR,
4094 mode, in1, in2);
4096 break;
4098 case NEG:
4099 /* (neg (xor A 1)) is (plus A -1) if A is known to be either 0 or 1. */
4100 if (GET_CODE (XEXP (x, 0)) == XOR
4101 && XEXP (XEXP (x, 0), 1) == const1_rtx
4102 && nonzero_bits (XEXP (XEXP (x, 0), 0), mode) == 1)
4103 return simplify_gen_binary (PLUS, mode, XEXP (XEXP (x, 0), 0),
4104 constm1_rtx);
4106 temp = expand_compound_operation (XEXP (x, 0));
4108 /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be
4109 replaced by (lshiftrt X C). This will convert
4110 (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */
4112 if (GET_CODE (temp) == ASHIFTRT
4113 && GET_CODE (XEXP (temp, 1)) == CONST_INT
4114 && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1)
4115 return simplify_shift_const (temp, LSHIFTRT, mode, XEXP (temp, 0),
4116 INTVAL (XEXP (temp, 1)));
4118 /* If X has only a single bit that might be nonzero, say, bit I, convert
4119 (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of
4120 MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to
4121 (sign_extract X 1 Y). But only do this if TEMP isn't a register
4122 or a SUBREG of one since we'd be making the expression more
4123 complex if it was just a register. */
4125 if (!REG_P (temp)
4126 && ! (GET_CODE (temp) == SUBREG
4127 && REG_P (SUBREG_REG (temp)))
4128 && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0)
4130 rtx temp1 = simplify_shift_const
4131 (NULL_RTX, ASHIFTRT, mode,
4132 simplify_shift_const (NULL_RTX, ASHIFT, mode, temp,
4133 GET_MODE_BITSIZE (mode) - 1 - i),
4134 GET_MODE_BITSIZE (mode) - 1 - i);
4136 /* If all we did was surround TEMP with the two shifts, we
4137 haven't improved anything, so don't use it. Otherwise,
4138 we are better off with TEMP1. */
4139 if (GET_CODE (temp1) != ASHIFTRT
4140 || GET_CODE (XEXP (temp1, 0)) != ASHIFT
4141 || XEXP (XEXP (temp1, 0), 0) != temp)
4142 return temp1;
4144 break;
4146 case TRUNCATE:
4147 /* We can't handle truncation to a partial integer mode here
4148 because we don't know the real bitsize of the partial
4149 integer mode. */
4150 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
4151 break;
4153 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4154 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
4155 GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))))
4156 SUBST (XEXP (x, 0),
4157 force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
4158 GET_MODE_MASK (mode), NULL_RTX, 0));
4160 /* (truncate:SI ({sign,zero}_extend:DI foo:SI)) == foo:SI. */
4161 if ((GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
4162 || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND)
4163 && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode)
4164 return XEXP (XEXP (x, 0), 0);
4166 /* (truncate:SI (OP:DI ({sign,zero}_extend:DI foo:SI))) is
4167 (OP:SI foo:SI) if OP is NEG or ABS. */
4168 if ((GET_CODE (XEXP (x, 0)) == ABS
4169 || GET_CODE (XEXP (x, 0)) == NEG)
4170 && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SIGN_EXTEND
4171 || GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND)
4172 && GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == mode)
4173 return simplify_gen_unary (GET_CODE (XEXP (x, 0)), mode,
4174 XEXP (XEXP (XEXP (x, 0), 0), 0), mode);
4176 /* (truncate:SI (subreg:DI (truncate:SI X) 0)) is
4177 (truncate:SI x). */
4178 if (GET_CODE (XEXP (x, 0)) == SUBREG
4179 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == TRUNCATE
4180 && subreg_lowpart_p (XEXP (x, 0)))
4181 return SUBREG_REG (XEXP (x, 0));
4183 /* If we know that the value is already truncated, we can
4184 replace the TRUNCATE with a SUBREG if TRULY_NOOP_TRUNCATION
4185 is nonzero for the corresponding modes. But don't do this
4186 for an (LSHIFTRT (MULT ...)) since this will cause problems
4187 with the umulXi3_highpart patterns. */
4188 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
4189 GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))
4190 && num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
4191 >= (unsigned int) (GET_MODE_BITSIZE (mode) + 1)
4192 && ! (GET_CODE (XEXP (x, 0)) == LSHIFTRT
4193 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT))
4194 return gen_lowpart (mode, XEXP (x, 0));
4196 /* A truncate of a comparison can be replaced with a subreg if
4197 STORE_FLAG_VALUE permits. This is like the previous test,
4198 but it works even if the comparison is done in a mode larger
4199 than HOST_BITS_PER_WIDE_INT. */
4200 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4201 && COMPARISON_P (XEXP (x, 0))
4202 && ((HOST_WIDE_INT) STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0)
4203 return gen_lowpart (mode, XEXP (x, 0));
4205 /* Similarly, a truncate of a register whose value is a
4206 comparison can be replaced with a subreg if STORE_FLAG_VALUE
4207 permits. */
4208 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4209 && ((HOST_WIDE_INT) STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0
4210 && (temp = get_last_value (XEXP (x, 0)))
4211 && COMPARISON_P (temp))
4212 return gen_lowpart (mode, XEXP (x, 0));
4214 break;
4216 case FLOAT_TRUNCATE:
4217 /* (float_truncate:SF (float_extend:DF foo:SF)) = foo:SF. */
4218 if (GET_CODE (XEXP (x, 0)) == FLOAT_EXTEND
4219 && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode)
4220 return XEXP (XEXP (x, 0), 0);
4222 /* (float_truncate:SF (float_truncate:DF foo:XF))
4223 = (float_truncate:SF foo:XF).
4224 This may eliminate double rounding, so it is unsafe.
4226 (float_truncate:SF (float_extend:XF foo:DF))
4227 = (float_truncate:SF foo:DF).
4229 (float_truncate:DF (float_extend:XF foo:SF))
4230 = (float_extend:SF foo:DF). */
4231 if ((GET_CODE (XEXP (x, 0)) == FLOAT_TRUNCATE
4232 && flag_unsafe_math_optimizations)
4233 || GET_CODE (XEXP (x, 0)) == FLOAT_EXTEND)
4234 return simplify_gen_unary (GET_MODE_SIZE (GET_MODE (XEXP (XEXP (x, 0),
4235 0)))
4236 > GET_MODE_SIZE (mode)
4237 ? FLOAT_TRUNCATE : FLOAT_EXTEND,
4238 mode,
4239 XEXP (XEXP (x, 0), 0), mode);
4241 /* (float_truncate (float x)) is (float x) */
4242 if (GET_CODE (XEXP (x, 0)) == FLOAT
4243 && (flag_unsafe_math_optimizations
4244 || ((unsigned)significand_size (GET_MODE (XEXP (x, 0)))
4245 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (x, 0), 0)))
4246 - num_sign_bit_copies (XEXP (XEXP (x, 0), 0),
4247 GET_MODE (XEXP (XEXP (x, 0), 0)))))))
4248 return simplify_gen_unary (FLOAT, mode,
4249 XEXP (XEXP (x, 0), 0),
4250 GET_MODE (XEXP (XEXP (x, 0), 0)));
4252 /* (float_truncate:SF (OP:DF (float_extend:DF foo:sf))) is
4253 (OP:SF foo:SF) if OP is NEG or ABS. */
4254 if ((GET_CODE (XEXP (x, 0)) == ABS
4255 || GET_CODE (XEXP (x, 0)) == NEG)
4256 && GET_CODE (XEXP (XEXP (x, 0), 0)) == FLOAT_EXTEND
4257 && GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == mode)
4258 return simplify_gen_unary (GET_CODE (XEXP (x, 0)), mode,
4259 XEXP (XEXP (XEXP (x, 0), 0), 0), mode);
4261 /* (float_truncate:SF (subreg:DF (float_truncate:SF X) 0))
4262 is (float_truncate:SF x). */
4263 if (GET_CODE (XEXP (x, 0)) == SUBREG
4264 && subreg_lowpart_p (XEXP (x, 0))
4265 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == FLOAT_TRUNCATE)
4266 return SUBREG_REG (XEXP (x, 0));
4267 break;
4268 case FLOAT_EXTEND:
4269 /* (float_extend (float_extend x)) is (float_extend x)
4271 (float_extend (float x)) is (float x) assuming that double
4272 rounding can't happen.
4274 if (GET_CODE (XEXP (x, 0)) == FLOAT_EXTEND
4275 || (GET_CODE (XEXP (x, 0)) == FLOAT
4276 && ((unsigned)significand_size (GET_MODE (XEXP (x, 0)))
4277 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (x, 0), 0)))
4278 - num_sign_bit_copies (XEXP (XEXP (x, 0), 0),
4279 GET_MODE (XEXP (XEXP (x, 0), 0)))))))
4280 return simplify_gen_unary (GET_CODE (XEXP (x, 0)), mode,
4281 XEXP (XEXP (x, 0), 0),
4282 GET_MODE (XEXP (XEXP (x, 0), 0)));
4284 break;
4285 #ifdef HAVE_cc0
4286 case COMPARE:
4287 /* Convert (compare FOO (const_int 0)) to FOO unless we aren't
4288 using cc0, in which case we want to leave it as a COMPARE
4289 so we can distinguish it from a register-register-copy. */
4290 if (XEXP (x, 1) == const0_rtx)
4291 return XEXP (x, 0);
4293 /* x - 0 is the same as x unless x's mode has signed zeros and
4294 allows rounding towards -infinity. Under those conditions,
4295 0 - 0 is -0. */
4296 if (!(HONOR_SIGNED_ZEROS (GET_MODE (XEXP (x, 0)))
4297 && HONOR_SIGN_DEPENDENT_ROUNDING (GET_MODE (XEXP (x, 0))))
4298 && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0))))
4299 return XEXP (x, 0);
4300 break;
4301 #endif
4303 case CONST:
4304 /* (const (const X)) can become (const X). Do it this way rather than
4305 returning the inner CONST since CONST can be shared with a
4306 REG_EQUAL note. */
4307 if (GET_CODE (XEXP (x, 0)) == CONST)
4308 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4309 break;
4311 #ifdef HAVE_lo_sum
4312 case LO_SUM:
4313 /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we
4314 can add in an offset. find_split_point will split this address up
4315 again if it doesn't match. */
4316 if (GET_CODE (XEXP (x, 0)) == HIGH
4317 && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
4318 return XEXP (x, 1);
4319 break;
4320 #endif
4322 case PLUS:
4323 /* Canonicalize (plus (mult (neg B) C) A) to (minus A (mult B C)).
4325 if (GET_CODE (XEXP (x, 0)) == MULT
4326 && GET_CODE (XEXP (XEXP (x, 0), 0)) == NEG)
4328 rtx in1, in2;
4330 in1 = XEXP (XEXP (XEXP (x, 0), 0), 0);
4331 in2 = XEXP (XEXP (x, 0), 1);
4332 return simplify_gen_binary (MINUS, mode, XEXP (x, 1),
4333 simplify_gen_binary (MULT, mode,
4334 in1, in2));
4337 /* If we have (plus (plus (A const) B)), associate it so that CONST is
4338 outermost. That's because that's the way indexed addresses are
4339 supposed to appear. This code used to check many more cases, but
4340 they are now checked elsewhere. */
4341 if (GET_CODE (XEXP (x, 0)) == PLUS
4342 && CONSTANT_ADDRESS_P (XEXP (XEXP (x, 0), 1)))
4343 return simplify_gen_binary (PLUS, mode,
4344 simplify_gen_binary (PLUS, mode,
4345 XEXP (XEXP (x, 0), 0),
4346 XEXP (x, 1)),
4347 XEXP (XEXP (x, 0), 1));
4349 /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>)
4350 when c is (const_int (pow2 + 1) / 2) is a sign extension of a
4351 bit-field and can be replaced by either a sign_extend or a
4352 sign_extract. The `and' may be a zero_extend and the two
4353 <c>, -<c> constants may be reversed. */
4354 if (GET_CODE (XEXP (x, 0)) == XOR
4355 && GET_CODE (XEXP (x, 1)) == CONST_INT
4356 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
4357 && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1))
4358 && ((i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0
4359 || (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
4360 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4361 && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND
4362 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
4363 && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))
4364 == ((HOST_WIDE_INT) 1 << (i + 1)) - 1))
4365 || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND
4366 && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))
4367 == (unsigned int) i + 1))))
4368 return simplify_shift_const
4369 (NULL_RTX, ASHIFTRT, mode,
4370 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4371 XEXP (XEXP (XEXP (x, 0), 0), 0),
4372 GET_MODE_BITSIZE (mode) - (i + 1)),
4373 GET_MODE_BITSIZE (mode) - (i + 1));
4375 /* (plus (comparison A B) C) can become (neg (rev-comp A B)) if
4376 C is 1 and STORE_FLAG_VALUE is -1 or if C is -1 and STORE_FLAG_VALUE
4377 is 1. This produces better code than the alternative immediately
4378 below. */
4379 if (COMPARISON_P (XEXP (x, 0))
4380 && ((STORE_FLAG_VALUE == -1 && XEXP (x, 1) == const1_rtx)
4381 || (STORE_FLAG_VALUE == 1 && XEXP (x, 1) == constm1_rtx))
4382 && (reversed = reversed_comparison (XEXP (x, 0), mode)))
4383 return
4384 simplify_gen_unary (NEG, mode, reversed, mode);
4386 /* If only the low-order bit of X is possibly nonzero, (plus x -1)
4387 can become (ashiftrt (ashift (xor x 1) C) C) where C is
4388 the bitsize of the mode - 1. This allows simplification of
4389 "a = (b & 8) == 0;" */
4390 if (XEXP (x, 1) == constm1_rtx
4391 && !REG_P (XEXP (x, 0))
4392 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4393 && REG_P (SUBREG_REG (XEXP (x, 0))))
4394 && nonzero_bits (XEXP (x, 0), mode) == 1)
4395 return simplify_shift_const (NULL_RTX, ASHIFTRT, mode,
4396 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4397 gen_rtx_XOR (mode, XEXP (x, 0), const1_rtx),
4398 GET_MODE_BITSIZE (mode) - 1),
4399 GET_MODE_BITSIZE (mode) - 1);
4401 /* If we are adding two things that have no bits in common, convert
4402 the addition into an IOR. This will often be further simplified,
4403 for example in cases like ((a & 1) + (a & 2)), which can
4404 become a & 3. */
4406 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4407 && (nonzero_bits (XEXP (x, 0), mode)
4408 & nonzero_bits (XEXP (x, 1), mode)) == 0)
4410 /* Try to simplify the expression further. */
4411 rtx tor = simplify_gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1));
4412 temp = combine_simplify_rtx (tor, mode, in_dest);
4414 /* If we could, great. If not, do not go ahead with the IOR
4415 replacement, since PLUS appears in many special purpose
4416 address arithmetic instructions. */
4417 if (GET_CODE (temp) != CLOBBER && temp != tor)
4418 return temp;
4420 break;
4422 case MINUS:
4423 /* If STORE_FLAG_VALUE is 1, (minus 1 (comparison foo bar)) can be done
4424 by reversing the comparison code if valid. */
4425 if (STORE_FLAG_VALUE == 1
4426 && XEXP (x, 0) == const1_rtx
4427 && COMPARISON_P (XEXP (x, 1))
4428 && (reversed = reversed_comparison (XEXP (x, 1), mode)))
4429 return reversed;
4431 /* (minus <foo> (and <foo> (const_int -pow2))) becomes
4432 (and <foo> (const_int pow2-1)) */
4433 if (GET_CODE (XEXP (x, 1)) == AND
4434 && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
4435 && exact_log2 (-INTVAL (XEXP (XEXP (x, 1), 1))) >= 0
4436 && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
4437 return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0),
4438 -INTVAL (XEXP (XEXP (x, 1), 1)) - 1);
4440 /* Canonicalize (minus A (mult (neg B) C)) to (plus (mult B C) A).
4442 if (GET_CODE (XEXP (x, 1)) == MULT
4443 && GET_CODE (XEXP (XEXP (x, 1), 0)) == NEG)
4445 rtx in1, in2;
4447 in1 = XEXP (XEXP (XEXP (x, 1), 0), 0);
4448 in2 = XEXP (XEXP (x, 1), 1);
4449 return simplify_gen_binary (PLUS, mode,
4450 simplify_gen_binary (MULT, mode,
4451 in1, in2),
4452 XEXP (x, 0));
4455 /* Canonicalize (minus (neg A) (mult B C)) to
4456 (minus (mult (neg B) C) A). */
4457 if (GET_CODE (XEXP (x, 1)) == MULT
4458 && GET_CODE (XEXP (x, 0)) == NEG)
4460 rtx in1, in2;
4462 in1 = simplify_gen_unary (NEG, mode, XEXP (XEXP (x, 1), 0), mode);
4463 in2 = XEXP (XEXP (x, 1), 1);
4464 return simplify_gen_binary (MINUS, mode,
4465 simplify_gen_binary (MULT, mode,
4466 in1, in2),
4467 XEXP (XEXP (x, 0), 0));
4470 /* Canonicalize (minus A (plus B C)) to (minus (minus A B) C) for
4471 integers. */
4472 if (GET_CODE (XEXP (x, 1)) == PLUS && INTEGRAL_MODE_P (mode))
4473 return simplify_gen_binary (MINUS, mode,
4474 simplify_gen_binary (MINUS, mode,
4475 XEXP (x, 0),
4476 XEXP (XEXP (x, 1), 0)),
4477 XEXP (XEXP (x, 1), 1));
4478 break;
4480 case MULT:
4481 /* If we have (mult (plus A B) C), apply the distributive law and then
4482 the inverse distributive law to see if things simplify. This
4483 occurs mostly in addresses, often when unrolling loops. */
4485 if (GET_CODE (XEXP (x, 0)) == PLUS)
4487 rtx result = distribute_and_simplify_rtx (x, 0);
4488 if (result)
4489 return result;
4492 /* Try simplify a*(b/c) as (a*b)/c. */
4493 if (FLOAT_MODE_P (mode) && flag_unsafe_math_optimizations
4494 && GET_CODE (XEXP (x, 0)) == DIV)
4496 rtx tem = simplify_binary_operation (MULT, mode,
4497 XEXP (XEXP (x, 0), 0),
4498 XEXP (x, 1));
4499 if (tem)
4500 return simplify_gen_binary (DIV, mode, tem, XEXP (XEXP (x, 0), 1));
4502 break;
4504 case UDIV:
4505 /* If this is a divide by a power of two, treat it as a shift if
4506 its first operand is a shift. */
4507 if (GET_CODE (XEXP (x, 1)) == CONST_INT
4508 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0
4509 && (GET_CODE (XEXP (x, 0)) == ASHIFT
4510 || GET_CODE (XEXP (x, 0)) == LSHIFTRT
4511 || GET_CODE (XEXP (x, 0)) == ASHIFTRT
4512 || GET_CODE (XEXP (x, 0)) == ROTATE
4513 || GET_CODE (XEXP (x, 0)) == ROTATERT))
4514 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i);
4515 break;
4517 case EQ: case NE:
4518 case GT: case GTU: case GE: case GEU:
4519 case LT: case LTU: case LE: case LEU:
4520 case UNEQ: case LTGT:
4521 case UNGT: case UNGE:
4522 case UNLT: case UNLE:
4523 case UNORDERED: case ORDERED:
4524 /* If the first operand is a condition code, we can't do anything
4525 with it. */
4526 if (GET_CODE (XEXP (x, 0)) == COMPARE
4527 || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC
4528 && ! CC0_P (XEXP (x, 0))))
4530 rtx op0 = XEXP (x, 0);
4531 rtx op1 = XEXP (x, 1);
4532 enum rtx_code new_code;
4534 if (GET_CODE (op0) == COMPARE)
4535 op1 = XEXP (op0, 1), op0 = XEXP (op0, 0);
4537 /* Simplify our comparison, if possible. */
4538 new_code = simplify_comparison (code, &op0, &op1);
4540 /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X
4541 if only the low-order bit is possibly nonzero in X (such as when
4542 X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to
4543 (xor X 1) or (minus 1 X); we use the former. Finally, if X is
4544 known to be either 0 or -1, NE becomes a NEG and EQ becomes
4545 (plus X 1).
4547 Remove any ZERO_EXTRACT we made when thinking this was a
4548 comparison. It may now be simpler to use, e.g., an AND. If a
4549 ZERO_EXTRACT is indeed appropriate, it will be placed back by
4550 the call to make_compound_operation in the SET case. */
4552 if (STORE_FLAG_VALUE == 1
4553 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4554 && op1 == const0_rtx
4555 && mode == GET_MODE (op0)
4556 && nonzero_bits (op0, mode) == 1)
4557 return gen_lowpart (mode,
4558 expand_compound_operation (op0));
4560 else if (STORE_FLAG_VALUE == 1
4561 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4562 && op1 == const0_rtx
4563 && mode == GET_MODE (op0)
4564 && (num_sign_bit_copies (op0, mode)
4565 == GET_MODE_BITSIZE (mode)))
4567 op0 = expand_compound_operation (op0);
4568 return simplify_gen_unary (NEG, mode,
4569 gen_lowpart (mode, op0),
4570 mode);
4573 else if (STORE_FLAG_VALUE == 1
4574 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4575 && op1 == const0_rtx
4576 && mode == GET_MODE (op0)
4577 && nonzero_bits (op0, mode) == 1)
4579 op0 = expand_compound_operation (op0);
4580 return simplify_gen_binary (XOR, mode,
4581 gen_lowpart (mode, op0),
4582 const1_rtx);
4585 else if (STORE_FLAG_VALUE == 1
4586 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4587 && op1 == const0_rtx
4588 && mode == GET_MODE (op0)
4589 && (num_sign_bit_copies (op0, mode)
4590 == GET_MODE_BITSIZE (mode)))
4592 op0 = expand_compound_operation (op0);
4593 return plus_constant (gen_lowpart (mode, op0), 1);
4596 /* If STORE_FLAG_VALUE is -1, we have cases similar to
4597 those above. */
4598 if (STORE_FLAG_VALUE == -1
4599 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4600 && op1 == const0_rtx
4601 && (num_sign_bit_copies (op0, mode)
4602 == GET_MODE_BITSIZE (mode)))
4603 return gen_lowpart (mode,
4604 expand_compound_operation (op0));
4606 else if (STORE_FLAG_VALUE == -1
4607 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4608 && op1 == const0_rtx
4609 && mode == GET_MODE (op0)
4610 && nonzero_bits (op0, mode) == 1)
4612 op0 = expand_compound_operation (op0);
4613 return simplify_gen_unary (NEG, mode,
4614 gen_lowpart (mode, op0),
4615 mode);
4618 else if (STORE_FLAG_VALUE == -1
4619 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4620 && op1 == const0_rtx
4621 && mode == GET_MODE (op0)
4622 && (num_sign_bit_copies (op0, mode)
4623 == GET_MODE_BITSIZE (mode)))
4625 op0 = expand_compound_operation (op0);
4626 return simplify_gen_unary (NOT, mode,
4627 gen_lowpart (mode, op0),
4628 mode);
4631 /* If X is 0/1, (eq X 0) is X-1. */
4632 else if (STORE_FLAG_VALUE == -1
4633 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4634 && op1 == const0_rtx
4635 && mode == GET_MODE (op0)
4636 && nonzero_bits (op0, mode) == 1)
4638 op0 = expand_compound_operation (op0);
4639 return plus_constant (gen_lowpart (mode, op0), -1);
4642 /* If STORE_FLAG_VALUE says to just test the sign bit and X has just
4643 one bit that might be nonzero, we can convert (ne x 0) to
4644 (ashift x c) where C puts the bit in the sign bit. Remove any
4645 AND with STORE_FLAG_VALUE when we are done, since we are only
4646 going to test the sign bit. */
4647 if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4648 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4649 && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
4650 == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))
4651 && op1 == const0_rtx
4652 && mode == GET_MODE (op0)
4653 && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0)
4655 x = simplify_shift_const (NULL_RTX, ASHIFT, mode,
4656 expand_compound_operation (op0),
4657 GET_MODE_BITSIZE (mode) - 1 - i);
4658 if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx)
4659 return XEXP (x, 0);
4660 else
4661 return x;
4664 /* If the code changed, return a whole new comparison. */
4665 if (new_code != code)
4666 return gen_rtx_fmt_ee (new_code, mode, op0, op1);
4668 /* Otherwise, keep this operation, but maybe change its operands.
4669 This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */
4670 SUBST (XEXP (x, 0), op0);
4671 SUBST (XEXP (x, 1), op1);
4673 break;
4675 case IF_THEN_ELSE:
4676 return simplify_if_then_else (x);
4678 case ZERO_EXTRACT:
4679 case SIGN_EXTRACT:
4680 case ZERO_EXTEND:
4681 case SIGN_EXTEND:
4682 /* If we are processing SET_DEST, we are done. */
4683 if (in_dest)
4684 return x;
4686 return expand_compound_operation (x);
4688 case SET:
4689 return simplify_set (x);
4691 case AND:
4692 case IOR:
4693 case XOR:
4694 return simplify_logical (x);
4696 case ABS:
4697 /* (abs (neg <foo>)) -> (abs <foo>) */
4698 if (GET_CODE (XEXP (x, 0)) == NEG)
4699 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4701 /* If the mode of the operand is VOIDmode (i.e. if it is ASM_OPERANDS),
4702 do nothing. */
4703 if (GET_MODE (XEXP (x, 0)) == VOIDmode)
4704 break;
4706 /* If operand is something known to be positive, ignore the ABS. */
4707 if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS
4708 || ((GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
4709 <= HOST_BITS_PER_WIDE_INT)
4710 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
4711 & ((HOST_WIDE_INT) 1
4712 << (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - 1)))
4713 == 0)))
4714 return XEXP (x, 0);
4716 /* If operand is known to be only -1 or 0, convert ABS to NEG. */
4717 if (num_sign_bit_copies (XEXP (x, 0), mode) == GET_MODE_BITSIZE (mode))
4718 return gen_rtx_NEG (mode, XEXP (x, 0));
4720 break;
4722 case FFS:
4723 /* (ffs (*_extend <X>)) = (ffs <X>) */
4724 if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
4725 || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND)
4726 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4727 break;
4729 case POPCOUNT:
4730 case PARITY:
4731 /* (pop* (zero_extend <X>)) = (pop* <X>) */
4732 if (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND)
4733 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4734 break;
4736 case FLOAT:
4737 /* (float (sign_extend <X>)) = (float <X>). */
4738 if (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)
4739 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4740 break;
4742 case ASHIFT:
4743 case LSHIFTRT:
4744 case ASHIFTRT:
4745 case ROTATE:
4746 case ROTATERT:
4747 /* If this is a shift by a constant amount, simplify it. */
4748 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
4749 return simplify_shift_const (x, code, mode, XEXP (x, 0),
4750 INTVAL (XEXP (x, 1)));
4752 else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1)))
4753 SUBST (XEXP (x, 1),
4754 force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)),
4755 ((HOST_WIDE_INT) 1
4756 << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))))
4757 - 1,
4758 NULL_RTX, 0));
4759 break;
4761 case VEC_SELECT:
4763 rtx op0 = XEXP (x, 0);
4764 rtx op1 = XEXP (x, 1);
4765 int len;
4767 gcc_assert (GET_CODE (op1) == PARALLEL);
4768 len = XVECLEN (op1, 0);
4769 if (len == 1
4770 && GET_CODE (XVECEXP (op1, 0, 0)) == CONST_INT
4771 && GET_CODE (op0) == VEC_CONCAT)
4773 int offset = INTVAL (XVECEXP (op1, 0, 0)) * GET_MODE_SIZE (GET_MODE (x));
4775 /* Try to find the element in the VEC_CONCAT. */
4776 for (;;)
4778 if (GET_MODE (op0) == GET_MODE (x))
4779 return op0;
4780 if (GET_CODE (op0) == VEC_CONCAT)
4782 HOST_WIDE_INT op0_size = GET_MODE_SIZE (GET_MODE (XEXP (op0, 0)));
4783 if (op0_size < offset)
4784 op0 = XEXP (op0, 0);
4785 else
4787 offset -= op0_size;
4788 op0 = XEXP (op0, 1);
4791 else
4792 break;
4797 break;
4799 default:
4800 break;
4803 return x;
4806 /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */
4808 static rtx
4809 simplify_if_then_else (rtx x)
4811 enum machine_mode mode = GET_MODE (x);
4812 rtx cond = XEXP (x, 0);
4813 rtx true_rtx = XEXP (x, 1);
4814 rtx false_rtx = XEXP (x, 2);
4815 enum rtx_code true_code = GET_CODE (cond);
4816 int comparison_p = COMPARISON_P (cond);
4817 rtx temp;
4818 int i;
4819 enum rtx_code false_code;
4820 rtx reversed;
4822 /* Simplify storing of the truth value. */
4823 if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx)
4824 return simplify_gen_relational (true_code, mode, VOIDmode,
4825 XEXP (cond, 0), XEXP (cond, 1));
4827 /* Also when the truth value has to be reversed. */
4828 if (comparison_p
4829 && true_rtx == const0_rtx && false_rtx == const_true_rtx
4830 && (reversed = reversed_comparison (cond, mode)))
4831 return reversed;
4833 /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used
4834 in it is being compared against certain values. Get the true and false
4835 comparisons and see if that says anything about the value of each arm. */
4837 if (comparison_p
4838 && ((false_code = reversed_comparison_code (cond, NULL))
4839 != UNKNOWN)
4840 && REG_P (XEXP (cond, 0)))
4842 HOST_WIDE_INT nzb;
4843 rtx from = XEXP (cond, 0);
4844 rtx true_val = XEXP (cond, 1);
4845 rtx false_val = true_val;
4846 int swapped = 0;
4848 /* If FALSE_CODE is EQ, swap the codes and arms. */
4850 if (false_code == EQ)
4852 swapped = 1, true_code = EQ, false_code = NE;
4853 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4856 /* If we are comparing against zero and the expression being tested has
4857 only a single bit that might be nonzero, that is its value when it is
4858 not equal to zero. Similarly if it is known to be -1 or 0. */
4860 if (true_code == EQ && true_val == const0_rtx
4861 && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0)
4862 false_code = EQ, false_val = GEN_INT (nzb);
4863 else if (true_code == EQ && true_val == const0_rtx
4864 && (num_sign_bit_copies (from, GET_MODE (from))
4865 == GET_MODE_BITSIZE (GET_MODE (from))))
4866 false_code = EQ, false_val = constm1_rtx;
4868 /* Now simplify an arm if we know the value of the register in the
4869 branch and it is used in the arm. Be careful due to the potential
4870 of locally-shared RTL. */
4872 if (reg_mentioned_p (from, true_rtx))
4873 true_rtx = subst (known_cond (copy_rtx (true_rtx), true_code,
4874 from, true_val),
4875 pc_rtx, pc_rtx, 0, 0);
4876 if (reg_mentioned_p (from, false_rtx))
4877 false_rtx = subst (known_cond (copy_rtx (false_rtx), false_code,
4878 from, false_val),
4879 pc_rtx, pc_rtx, 0, 0);
4881 SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx);
4882 SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx);
4884 true_rtx = XEXP (x, 1);
4885 false_rtx = XEXP (x, 2);
4886 true_code = GET_CODE (cond);
4889 /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be
4890 reversed, do so to avoid needing two sets of patterns for
4891 subtract-and-branch insns. Similarly if we have a constant in the true
4892 arm, the false arm is the same as the first operand of the comparison, or
4893 the false arm is more complicated than the true arm. */
4895 if (comparison_p
4896 && reversed_comparison_code (cond, NULL) != UNKNOWN
4897 && (true_rtx == pc_rtx
4898 || (CONSTANT_P (true_rtx)
4899 && GET_CODE (false_rtx) != CONST_INT && false_rtx != pc_rtx)
4900 || true_rtx == const0_rtx
4901 || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx))
4902 || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx))
4903 && !OBJECT_P (false_rtx))
4904 || reg_mentioned_p (true_rtx, false_rtx)
4905 || rtx_equal_p (false_rtx, XEXP (cond, 0))))
4907 true_code = reversed_comparison_code (cond, NULL);
4908 SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond)));
4909 SUBST (XEXP (x, 1), false_rtx);
4910 SUBST (XEXP (x, 2), true_rtx);
4912 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4913 cond = XEXP (x, 0);
4915 /* It is possible that the conditional has been simplified out. */
4916 true_code = GET_CODE (cond);
4917 comparison_p = COMPARISON_P (cond);
4920 /* If the two arms are identical, we don't need the comparison. */
4922 if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond))
4923 return true_rtx;
4925 /* Convert a == b ? b : a to "a". */
4926 if (true_code == EQ && ! side_effects_p (cond)
4927 && !HONOR_NANS (mode)
4928 && rtx_equal_p (XEXP (cond, 0), false_rtx)
4929 && rtx_equal_p (XEXP (cond, 1), true_rtx))
4930 return false_rtx;
4931 else if (true_code == NE && ! side_effects_p (cond)
4932 && !HONOR_NANS (mode)
4933 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4934 && rtx_equal_p (XEXP (cond, 1), false_rtx))
4935 return true_rtx;
4937 /* Look for cases where we have (abs x) or (neg (abs X)). */
4939 if (GET_MODE_CLASS (mode) == MODE_INT
4940 && GET_CODE (false_rtx) == NEG
4941 && rtx_equal_p (true_rtx, XEXP (false_rtx, 0))
4942 && comparison_p
4943 && rtx_equal_p (true_rtx, XEXP (cond, 0))
4944 && ! side_effects_p (true_rtx))
4945 switch (true_code)
4947 case GT:
4948 case GE:
4949 return simplify_gen_unary (ABS, mode, true_rtx, mode);
4950 case LT:
4951 case LE:
4952 return
4953 simplify_gen_unary (NEG, mode,
4954 simplify_gen_unary (ABS, mode, true_rtx, mode),
4955 mode);
4956 default:
4957 break;
4960 /* Look for MIN or MAX. */
4962 if ((! FLOAT_MODE_P (mode) || flag_unsafe_math_optimizations)
4963 && comparison_p
4964 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4965 && rtx_equal_p (XEXP (cond, 1), false_rtx)
4966 && ! side_effects_p (cond))
4967 switch (true_code)
4969 case GE:
4970 case GT:
4971 return simplify_gen_binary (SMAX, mode, true_rtx, false_rtx);
4972 case LE:
4973 case LT:
4974 return simplify_gen_binary (SMIN, mode, true_rtx, false_rtx);
4975 case GEU:
4976 case GTU:
4977 return simplify_gen_binary (UMAX, mode, true_rtx, false_rtx);
4978 case LEU:
4979 case LTU:
4980 return simplify_gen_binary (UMIN, mode, true_rtx, false_rtx);
4981 default:
4982 break;
4985 /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its
4986 second operand is zero, this can be done as (OP Z (mult COND C2)) where
4987 C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or
4988 SIGN_EXTEND as long as Z is already extended (so we don't destroy it).
4989 We can do this kind of thing in some cases when STORE_FLAG_VALUE is
4990 neither 1 or -1, but it isn't worth checking for. */
4992 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
4993 && comparison_p
4994 && GET_MODE_CLASS (mode) == MODE_INT
4995 && ! side_effects_p (x))
4997 rtx t = make_compound_operation (true_rtx, SET);
4998 rtx f = make_compound_operation (false_rtx, SET);
4999 rtx cond_op0 = XEXP (cond, 0);
5000 rtx cond_op1 = XEXP (cond, 1);
5001 enum rtx_code op = UNKNOWN, extend_op = UNKNOWN;
5002 enum machine_mode m = mode;
5003 rtx z = 0, c1 = NULL_RTX;
5005 if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS
5006 || GET_CODE (t) == IOR || GET_CODE (t) == XOR
5007 || GET_CODE (t) == ASHIFT
5008 || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT)
5009 && rtx_equal_p (XEXP (t, 0), f))
5010 c1 = XEXP (t, 1), op = GET_CODE (t), z = f;
5012 /* If an identity-zero op is commutative, check whether there
5013 would be a match if we swapped the operands. */
5014 else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR
5015 || GET_CODE (t) == XOR)
5016 && rtx_equal_p (XEXP (t, 1), f))
5017 c1 = XEXP (t, 0), op = GET_CODE (t), z = f;
5018 else if (GET_CODE (t) == SIGN_EXTEND
5019 && (GET_CODE (XEXP (t, 0)) == PLUS
5020 || GET_CODE (XEXP (t, 0)) == MINUS
5021 || GET_CODE (XEXP (t, 0)) == IOR
5022 || GET_CODE (XEXP (t, 0)) == XOR
5023 || GET_CODE (XEXP (t, 0)) == ASHIFT
5024 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5025 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5026 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5027 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5028 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5029 && (num_sign_bit_copies (f, GET_MODE (f))
5030 > (unsigned int)
5031 (GET_MODE_BITSIZE (mode)
5032 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 0))))))
5034 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5035 extend_op = SIGN_EXTEND;
5036 m = GET_MODE (XEXP (t, 0));
5038 else if (GET_CODE (t) == SIGN_EXTEND
5039 && (GET_CODE (XEXP (t, 0)) == PLUS
5040 || GET_CODE (XEXP (t, 0)) == IOR
5041 || GET_CODE (XEXP (t, 0)) == XOR)
5042 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5043 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5044 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5045 && (num_sign_bit_copies (f, GET_MODE (f))
5046 > (unsigned int)
5047 (GET_MODE_BITSIZE (mode)
5048 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 1))))))
5050 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5051 extend_op = SIGN_EXTEND;
5052 m = GET_MODE (XEXP (t, 0));
5054 else if (GET_CODE (t) == ZERO_EXTEND
5055 && (GET_CODE (XEXP (t, 0)) == PLUS
5056 || GET_CODE (XEXP (t, 0)) == MINUS
5057 || GET_CODE (XEXP (t, 0)) == IOR
5058 || GET_CODE (XEXP (t, 0)) == XOR
5059 || GET_CODE (XEXP (t, 0)) == ASHIFT
5060 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5061 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5062 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5063 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5064 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5065 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5066 && ((nonzero_bits (f, GET_MODE (f))
5067 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0))))
5068 == 0))
5070 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5071 extend_op = ZERO_EXTEND;
5072 m = GET_MODE (XEXP (t, 0));
5074 else if (GET_CODE (t) == ZERO_EXTEND
5075 && (GET_CODE (XEXP (t, 0)) == PLUS
5076 || GET_CODE (XEXP (t, 0)) == IOR
5077 || GET_CODE (XEXP (t, 0)) == XOR)
5078 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5079 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5080 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5081 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5082 && ((nonzero_bits (f, GET_MODE (f))
5083 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1))))
5084 == 0))
5086 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5087 extend_op = ZERO_EXTEND;
5088 m = GET_MODE (XEXP (t, 0));
5091 if (z)
5093 temp = subst (simplify_gen_relational (true_code, m, VOIDmode,
5094 cond_op0, cond_op1),
5095 pc_rtx, pc_rtx, 0, 0);
5096 temp = simplify_gen_binary (MULT, m, temp,
5097 simplify_gen_binary (MULT, m, c1,
5098 const_true_rtx));
5099 temp = subst (temp, pc_rtx, pc_rtx, 0, 0);
5100 temp = simplify_gen_binary (op, m, gen_lowpart (m, z), temp);
5102 if (extend_op != UNKNOWN)
5103 temp = simplify_gen_unary (extend_op, mode, temp, m);
5105 return temp;
5109 /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or
5110 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the
5111 negation of a single bit, we can convert this operation to a shift. We
5112 can actually do this more generally, but it doesn't seem worth it. */
5114 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5115 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5116 && ((1 == nonzero_bits (XEXP (cond, 0), mode)
5117 && (i = exact_log2 (INTVAL (true_rtx))) >= 0)
5118 || ((num_sign_bit_copies (XEXP (cond, 0), mode)
5119 == GET_MODE_BITSIZE (mode))
5120 && (i = exact_log2 (-INTVAL (true_rtx))) >= 0)))
5121 return
5122 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5123 gen_lowpart (mode, XEXP (cond, 0)), i);
5125 /* (IF_THEN_ELSE (NE REG 0) (0) (8)) is REG for nonzero_bits (REG) == 8. */
5126 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5127 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5128 && GET_MODE (XEXP (cond, 0)) == mode
5129 && (INTVAL (true_rtx) & GET_MODE_MASK (mode))
5130 == nonzero_bits (XEXP (cond, 0), mode)
5131 && (i = exact_log2 (INTVAL (true_rtx) & GET_MODE_MASK (mode))) >= 0)
5132 return XEXP (cond, 0);
5134 return x;
5137 /* Simplify X, a SET expression. Return the new expression. */
5139 static rtx
5140 simplify_set (rtx x)
5142 rtx src = SET_SRC (x);
5143 rtx dest = SET_DEST (x);
5144 enum machine_mode mode
5145 = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest);
5146 rtx other_insn;
5147 rtx *cc_use;
5149 /* (set (pc) (return)) gets written as (return). */
5150 if (GET_CODE (dest) == PC && GET_CODE (src) == RETURN)
5151 return src;
5153 /* Now that we know for sure which bits of SRC we are using, see if we can
5154 simplify the expression for the object knowing that we only need the
5155 low-order bits. */
5157 if (GET_MODE_CLASS (mode) == MODE_INT
5158 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
5160 src = force_to_mode (src, mode, ~(HOST_WIDE_INT) 0, NULL_RTX, 0);
5161 SUBST (SET_SRC (x), src);
5164 /* If we are setting CC0 or if the source is a COMPARE, look for the use of
5165 the comparison result and try to simplify it unless we already have used
5166 undobuf.other_insn. */
5167 if ((GET_MODE_CLASS (mode) == MODE_CC
5168 || GET_CODE (src) == COMPARE
5169 || CC0_P (dest))
5170 && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0
5171 && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn)
5172 && COMPARISON_P (*cc_use)
5173 && rtx_equal_p (XEXP (*cc_use, 0), dest))
5175 enum rtx_code old_code = GET_CODE (*cc_use);
5176 enum rtx_code new_code;
5177 rtx op0, op1, tmp;
5178 int other_changed = 0;
5179 enum machine_mode compare_mode = GET_MODE (dest);
5181 if (GET_CODE (src) == COMPARE)
5182 op0 = XEXP (src, 0), op1 = XEXP (src, 1);
5183 else
5184 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
5186 tmp = simplify_relational_operation (old_code, compare_mode, VOIDmode,
5187 op0, op1);
5188 if (!tmp)
5189 new_code = old_code;
5190 else if (!CONSTANT_P (tmp))
5192 new_code = GET_CODE (tmp);
5193 op0 = XEXP (tmp, 0);
5194 op1 = XEXP (tmp, 1);
5196 else
5198 rtx pat = PATTERN (other_insn);
5199 undobuf.other_insn = other_insn;
5200 SUBST (*cc_use, tmp);
5202 /* Attempt to simplify CC user. */
5203 if (GET_CODE (pat) == SET)
5205 rtx new = simplify_rtx (SET_SRC (pat));
5206 if (new != NULL_RTX)
5207 SUBST (SET_SRC (pat), new);
5210 /* Convert X into a no-op move. */
5211 SUBST (SET_DEST (x), pc_rtx);
5212 SUBST (SET_SRC (x), pc_rtx);
5213 return x;
5216 /* Simplify our comparison, if possible. */
5217 new_code = simplify_comparison (new_code, &op0, &op1);
5219 #ifdef SELECT_CC_MODE
5220 /* If this machine has CC modes other than CCmode, check to see if we
5221 need to use a different CC mode here. */
5222 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
5223 compare_mode = GET_MODE (op0);
5224 else
5225 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
5227 #ifndef HAVE_cc0
5228 /* If the mode changed, we have to change SET_DEST, the mode in the
5229 compare, and the mode in the place SET_DEST is used. If SET_DEST is
5230 a hard register, just build new versions with the proper mode. If it
5231 is a pseudo, we lose unless it is only time we set the pseudo, in
5232 which case we can safely change its mode. */
5233 if (compare_mode != GET_MODE (dest))
5235 unsigned int regno = REGNO (dest);
5236 rtx new_dest = gen_rtx_REG (compare_mode, regno);
5238 if (regno < FIRST_PSEUDO_REGISTER
5239 || (REG_N_SETS (regno) == 1 && ! REG_USERVAR_P (dest)))
5241 if (regno >= FIRST_PSEUDO_REGISTER)
5242 SUBST (regno_reg_rtx[regno], new_dest);
5244 SUBST (SET_DEST (x), new_dest);
5245 SUBST (XEXP (*cc_use, 0), new_dest);
5246 other_changed = 1;
5248 dest = new_dest;
5251 #endif /* cc0 */
5252 #endif /* SELECT_CC_MODE */
5254 /* If the code changed, we have to build a new comparison in
5255 undobuf.other_insn. */
5256 if (new_code != old_code)
5258 int other_changed_previously = other_changed;
5259 unsigned HOST_WIDE_INT mask;
5261 SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use),
5262 dest, const0_rtx));
5263 other_changed = 1;
5265 /* If the only change we made was to change an EQ into an NE or
5266 vice versa, OP0 has only one bit that might be nonzero, and OP1
5267 is zero, check if changing the user of the condition code will
5268 produce a valid insn. If it won't, we can keep the original code
5269 in that insn by surrounding our operation with an XOR. */
5271 if (((old_code == NE && new_code == EQ)
5272 || (old_code == EQ && new_code == NE))
5273 && ! other_changed_previously && op1 == const0_rtx
5274 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
5275 && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0)
5277 rtx pat = PATTERN (other_insn), note = 0;
5279 if ((recog_for_combine (&pat, other_insn, &note) < 0
5280 && ! check_asm_operands (pat)))
5282 PUT_CODE (*cc_use, old_code);
5283 other_changed = 0;
5285 op0 = simplify_gen_binary (XOR, GET_MODE (op0),
5286 op0, GEN_INT (mask));
5291 if (other_changed)
5292 undobuf.other_insn = other_insn;
5294 #ifdef HAVE_cc0
5295 /* If we are now comparing against zero, change our source if
5296 needed. If we do not use cc0, we always have a COMPARE. */
5297 if (op1 == const0_rtx && dest == cc0_rtx)
5299 SUBST (SET_SRC (x), op0);
5300 src = op0;
5302 else
5303 #endif
5305 /* Otherwise, if we didn't previously have a COMPARE in the
5306 correct mode, we need one. */
5307 if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode)
5309 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
5310 src = SET_SRC (x);
5312 else
5314 /* Otherwise, update the COMPARE if needed. */
5315 SUBST (XEXP (src, 0), op0);
5316 SUBST (XEXP (src, 1), op1);
5319 else
5321 /* Get SET_SRC in a form where we have placed back any
5322 compound expressions. Then do the checks below. */
5323 src = make_compound_operation (src, SET);
5324 SUBST (SET_SRC (x), src);
5327 /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation,
5328 and X being a REG or (subreg (reg)), we may be able to convert this to
5329 (set (subreg:m2 x) (op)).
5331 We can always do this if M1 is narrower than M2 because that means that
5332 we only care about the low bits of the result.
5334 However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot
5335 perform a narrower operation than requested since the high-order bits will
5336 be undefined. On machine where it is defined, this transformation is safe
5337 as long as M1 and M2 have the same number of words. */
5339 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5340 && !OBJECT_P (SUBREG_REG (src))
5341 && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1))
5342 / UNITS_PER_WORD)
5343 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5344 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))
5345 #ifndef WORD_REGISTER_OPERATIONS
5346 && (GET_MODE_SIZE (GET_MODE (src))
5347 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5348 #endif
5349 #ifdef CANNOT_CHANGE_MODE_CLASS
5350 && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER
5351 && REG_CANNOT_CHANGE_MODE_P (REGNO (dest),
5352 GET_MODE (SUBREG_REG (src)),
5353 GET_MODE (src)))
5354 #endif
5355 && (REG_P (dest)
5356 || (GET_CODE (dest) == SUBREG
5357 && REG_P (SUBREG_REG (dest)))))
5359 SUBST (SET_DEST (x),
5360 gen_lowpart (GET_MODE (SUBREG_REG (src)),
5361 dest));
5362 SUBST (SET_SRC (x), SUBREG_REG (src));
5364 src = SET_SRC (x), dest = SET_DEST (x);
5367 #ifdef HAVE_cc0
5368 /* If we have (set (cc0) (subreg ...)), we try to remove the subreg
5369 in SRC. */
5370 if (dest == cc0_rtx
5371 && GET_CODE (src) == SUBREG
5372 && subreg_lowpart_p (src)
5373 && (GET_MODE_BITSIZE (GET_MODE (src))
5374 < GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (src)))))
5376 rtx inner = SUBREG_REG (src);
5377 enum machine_mode inner_mode = GET_MODE (inner);
5379 /* Here we make sure that we don't have a sign bit on. */
5380 if (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT
5381 && (nonzero_bits (inner, inner_mode)
5382 < ((unsigned HOST_WIDE_INT) 1
5383 << (GET_MODE_BITSIZE (GET_MODE (src)) - 1))))
5385 SUBST (SET_SRC (x), inner);
5386 src = SET_SRC (x);
5389 #endif
5391 #ifdef LOAD_EXTEND_OP
5392 /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this
5393 would require a paradoxical subreg. Replace the subreg with a
5394 zero_extend to avoid the reload that would otherwise be required. */
5396 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5397 && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != UNKNOWN
5398 && SUBREG_BYTE (src) == 0
5399 && (GET_MODE_SIZE (GET_MODE (src))
5400 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5401 && MEM_P (SUBREG_REG (src)))
5403 SUBST (SET_SRC (x),
5404 gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))),
5405 GET_MODE (src), SUBREG_REG (src)));
5407 src = SET_SRC (x);
5409 #endif
5411 /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we
5412 are comparing an item known to be 0 or -1 against 0, use a logical
5413 operation instead. Check for one of the arms being an IOR of the other
5414 arm with some value. We compute three terms to be IOR'ed together. In
5415 practice, at most two will be nonzero. Then we do the IOR's. */
5417 if (GET_CODE (dest) != PC
5418 && GET_CODE (src) == IF_THEN_ELSE
5419 && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT
5420 && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE)
5421 && XEXP (XEXP (src, 0), 1) == const0_rtx
5422 && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0))
5423 #ifdef HAVE_conditional_move
5424 && ! can_conditionally_move_p (GET_MODE (src))
5425 #endif
5426 && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0),
5427 GET_MODE (XEXP (XEXP (src, 0), 0)))
5428 == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (src, 0), 0))))
5429 && ! side_effects_p (src))
5431 rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE
5432 ? XEXP (src, 1) : XEXP (src, 2));
5433 rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE
5434 ? XEXP (src, 2) : XEXP (src, 1));
5435 rtx term1 = const0_rtx, term2, term3;
5437 if (GET_CODE (true_rtx) == IOR
5438 && rtx_equal_p (XEXP (true_rtx, 0), false_rtx))
5439 term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx;
5440 else if (GET_CODE (true_rtx) == IOR
5441 && rtx_equal_p (XEXP (true_rtx, 1), false_rtx))
5442 term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx;
5443 else if (GET_CODE (false_rtx) == IOR
5444 && rtx_equal_p (XEXP (false_rtx, 0), true_rtx))
5445 term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx;
5446 else if (GET_CODE (false_rtx) == IOR
5447 && rtx_equal_p (XEXP (false_rtx, 1), true_rtx))
5448 term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx;
5450 term2 = simplify_gen_binary (AND, GET_MODE (src),
5451 XEXP (XEXP (src, 0), 0), true_rtx);
5452 term3 = simplify_gen_binary (AND, GET_MODE (src),
5453 simplify_gen_unary (NOT, GET_MODE (src),
5454 XEXP (XEXP (src, 0), 0),
5455 GET_MODE (src)),
5456 false_rtx);
5458 SUBST (SET_SRC (x),
5459 simplify_gen_binary (IOR, GET_MODE (src),
5460 simplify_gen_binary (IOR, GET_MODE (src),
5461 term1, term2),
5462 term3));
5464 src = SET_SRC (x);
5467 /* If either SRC or DEST is a CLOBBER of (const_int 0), make this
5468 whole thing fail. */
5469 if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx)
5470 return src;
5471 else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx)
5472 return dest;
5473 else
5474 /* Convert this into a field assignment operation, if possible. */
5475 return make_field_assignment (x);
5478 /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified
5479 result. */
5481 static rtx
5482 simplify_logical (rtx x)
5484 enum machine_mode mode = GET_MODE (x);
5485 rtx op0 = XEXP (x, 0);
5486 rtx op1 = XEXP (x, 1);
5487 rtx reversed;
5489 switch (GET_CODE (x))
5491 case AND:
5492 /* Convert (A ^ B) & A to A & (~B) since the latter is often a single
5493 insn (and may simplify more). */
5494 if (GET_CODE (op0) == XOR
5495 && rtx_equal_p (XEXP (op0, 0), op1)
5496 && ! side_effects_p (op1))
5497 x = simplify_gen_binary (AND, mode,
5498 simplify_gen_unary (NOT, mode,
5499 XEXP (op0, 1), mode),
5500 op1);
5502 if (GET_CODE (op0) == XOR
5503 && rtx_equal_p (XEXP (op0, 1), op1)
5504 && ! side_effects_p (op1))
5505 x = simplify_gen_binary (AND, mode,
5506 simplify_gen_unary (NOT, mode,
5507 XEXP (op0, 0), mode),
5508 op1);
5510 /* Similarly for (~(A ^ B)) & A. */
5511 if (GET_CODE (op0) == NOT
5512 && GET_CODE (XEXP (op0, 0)) == XOR
5513 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), op1)
5514 && ! side_effects_p (op1))
5515 x = simplify_gen_binary (AND, mode, XEXP (XEXP (op0, 0), 1), op1);
5517 if (GET_CODE (op0) == NOT
5518 && GET_CODE (XEXP (op0, 0)) == XOR
5519 && rtx_equal_p (XEXP (XEXP (op0, 0), 1), op1)
5520 && ! side_effects_p (op1))
5521 x = simplify_gen_binary (AND, mode, XEXP (XEXP (op0, 0), 0), op1);
5523 /* We can call simplify_and_const_int only if we don't lose
5524 any (sign) bits when converting INTVAL (op1) to
5525 "unsigned HOST_WIDE_INT". */
5526 if (GET_CODE (op1) == CONST_INT
5527 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5528 || INTVAL (op1) > 0))
5530 x = simplify_and_const_int (x, mode, op0, INTVAL (op1));
5532 /* If we have (ior (and (X C1) C2)) and the next restart would be
5533 the last, simplify this by making C1 as small as possible
5534 and then exit. Only do this if C1 actually changes: for now
5535 this only saves memory but, should this transformation be
5536 moved to simplify-rtx.c, we'd risk unbounded recursion there. */
5537 if (GET_CODE (x) == IOR && GET_CODE (op0) == AND
5538 && GET_CODE (XEXP (op0, 1)) == CONST_INT
5539 && GET_CODE (op1) == CONST_INT
5540 && (INTVAL (XEXP (op0, 1)) & INTVAL (op1)) != 0)
5541 return simplify_gen_binary (IOR, mode,
5542 simplify_gen_binary
5543 (AND, mode, XEXP (op0, 0),
5544 GEN_INT (INTVAL (XEXP (op0, 1))
5545 & ~INTVAL (op1))), op1);
5547 if (GET_CODE (x) != AND)
5548 return x;
5550 op0 = XEXP (x, 0);
5551 op1 = XEXP (x, 1);
5554 /* Convert (A | B) & A to A. */
5555 if (GET_CODE (op0) == IOR
5556 && (rtx_equal_p (XEXP (op0, 0), op1)
5557 || rtx_equal_p (XEXP (op0, 1), op1))
5558 && ! side_effects_p (XEXP (op0, 0))
5559 && ! side_effects_p (XEXP (op0, 1)))
5560 return op1;
5562 /* If we have any of (and (ior A B) C) or (and (xor A B) C),
5563 apply the distributive law and then the inverse distributive
5564 law to see if things simplify. */
5565 if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR)
5567 rtx result = distribute_and_simplify_rtx (x, 0);
5568 if (result)
5569 return result;
5571 if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR)
5573 rtx result = distribute_and_simplify_rtx (x, 1);
5574 if (result)
5575 return result;
5577 break;
5579 case IOR:
5580 /* (ior A C) is C if all bits of A that might be nonzero are on in C. */
5581 if (GET_CODE (op1) == CONST_INT
5582 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5583 && (nonzero_bits (op0, mode) & ~INTVAL (op1)) == 0)
5584 return op1;
5586 /* Convert (A & B) | A to A. */
5587 if (GET_CODE (op0) == AND
5588 && (rtx_equal_p (XEXP (op0, 0), op1)
5589 || rtx_equal_p (XEXP (op0, 1), op1))
5590 && ! side_effects_p (XEXP (op0, 0))
5591 && ! side_effects_p (XEXP (op0, 1)))
5592 return op1;
5594 /* If we have (ior (and A B) C), apply the distributive law and then
5595 the inverse distributive law to see if things simplify. */
5597 if (GET_CODE (op0) == AND)
5599 rtx result = distribute_and_simplify_rtx (x, 0);
5600 if (result)
5601 return result;
5604 if (GET_CODE (op1) == AND)
5606 rtx result = distribute_and_simplify_rtx (x, 1);
5607 if (result)
5608 return result;
5611 /* Convert (ior (ashift A CX) (lshiftrt A CY)) where CX+CY equals the
5612 mode size to (rotate A CX). */
5614 if (((GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT)
5615 || (GET_CODE (op1) == ASHIFT && GET_CODE (op0) == LSHIFTRT))
5616 && rtx_equal_p (XEXP (op0, 0), XEXP (op1, 0))
5617 && GET_CODE (XEXP (op0, 1)) == CONST_INT
5618 && GET_CODE (XEXP (op1, 1)) == CONST_INT
5619 && (INTVAL (XEXP (op0, 1)) + INTVAL (XEXP (op1, 1))
5620 == GET_MODE_BITSIZE (mode)))
5621 return gen_rtx_ROTATE (mode, XEXP (op0, 0),
5622 (GET_CODE (op0) == ASHIFT
5623 ? XEXP (op0, 1) : XEXP (op1, 1)));
5625 /* If OP0 is (ashiftrt (plus ...) C), it might actually be
5626 a (sign_extend (plus ...)). If so, OP1 is a CONST_INT, and the PLUS
5627 does not affect any of the bits in OP1, it can really be done
5628 as a PLUS and we can associate. We do this by seeing if OP1
5629 can be safely shifted left C bits. */
5630 if (GET_CODE (op1) == CONST_INT && GET_CODE (op0) == ASHIFTRT
5631 && GET_CODE (XEXP (op0, 0)) == PLUS
5632 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
5633 && GET_CODE (XEXP (op0, 1)) == CONST_INT
5634 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT)
5636 int count = INTVAL (XEXP (op0, 1));
5637 HOST_WIDE_INT mask = INTVAL (op1) << count;
5639 if (mask >> count == INTVAL (op1)
5640 && (mask & nonzero_bits (XEXP (op0, 0), mode)) == 0)
5642 SUBST (XEXP (XEXP (op0, 0), 1),
5643 GEN_INT (INTVAL (XEXP (XEXP (op0, 0), 1)) | mask));
5644 return op0;
5647 break;
5649 case XOR:
5650 /* If we are XORing two things that have no bits in common,
5651 convert them into an IOR. This helps to detect rotation encoded
5652 using those methods and possibly other simplifications. */
5654 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5655 && (nonzero_bits (op0, mode)
5656 & nonzero_bits (op1, mode)) == 0)
5657 return (simplify_gen_binary (IOR, mode, op0, op1));
5659 /* Convert (XOR (NOT x) (NOT y)) to (XOR x y).
5660 Also convert (XOR (NOT x) y) to (NOT (XOR x y)), similarly for
5661 (NOT y). */
5663 int num_negated = 0;
5665 if (GET_CODE (op0) == NOT)
5666 num_negated++, op0 = XEXP (op0, 0);
5667 if (GET_CODE (op1) == NOT)
5668 num_negated++, op1 = XEXP (op1, 0);
5670 if (num_negated == 2)
5672 SUBST (XEXP (x, 0), op0);
5673 SUBST (XEXP (x, 1), op1);
5675 else if (num_negated == 1)
5676 return
5677 simplify_gen_unary (NOT, mode,
5678 simplify_gen_binary (XOR, mode, op0, op1),
5679 mode);
5682 /* Convert (xor (and A B) B) to (and (not A) B). The latter may
5683 correspond to a machine insn or result in further simplifications
5684 if B is a constant. */
5686 if (GET_CODE (op0) == AND
5687 && rtx_equal_p (XEXP (op0, 1), op1)
5688 && ! side_effects_p (op1))
5689 return simplify_gen_binary (AND, mode,
5690 simplify_gen_unary (NOT, mode,
5691 XEXP (op0, 0), mode),
5692 op1);
5694 else if (GET_CODE (op0) == AND
5695 && rtx_equal_p (XEXP (op0, 0), op1)
5696 && ! side_effects_p (op1))
5697 return simplify_gen_binary (AND, mode,
5698 simplify_gen_unary (NOT, mode,
5699 XEXP (op0, 1), mode),
5700 op1);
5702 /* (xor (comparison foo bar) (const_int 1)) can become the reversed
5703 comparison if STORE_FLAG_VALUE is 1. */
5704 if (STORE_FLAG_VALUE == 1
5705 && op1 == const1_rtx
5706 && COMPARISON_P (op0)
5707 && (reversed = reversed_comparison (op0, mode)))
5708 return reversed;
5710 /* (lshiftrt foo C) where C is the number of bits in FOO minus 1
5711 is (lt foo (const_int 0)), so we can perform the above
5712 simplification if STORE_FLAG_VALUE is 1. */
5714 if (STORE_FLAG_VALUE == 1
5715 && op1 == const1_rtx
5716 && GET_CODE (op0) == LSHIFTRT
5717 && GET_CODE (XEXP (op0, 1)) == CONST_INT
5718 && INTVAL (XEXP (op0, 1)) == GET_MODE_BITSIZE (mode) - 1)
5719 return gen_rtx_GE (mode, XEXP (op0, 0), const0_rtx);
5721 /* (xor (comparison foo bar) (const_int sign-bit))
5722 when STORE_FLAG_VALUE is the sign bit. */
5723 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5724 && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
5725 == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))
5726 && op1 == const_true_rtx
5727 && COMPARISON_P (op0)
5728 && (reversed = reversed_comparison (op0, mode)))
5729 return reversed;
5731 break;
5733 default:
5734 gcc_unreachable ();
5737 return x;
5740 /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound
5741 operations" because they can be replaced with two more basic operations.
5742 ZERO_EXTEND is also considered "compound" because it can be replaced with
5743 an AND operation, which is simpler, though only one operation.
5745 The function expand_compound_operation is called with an rtx expression
5746 and will convert it to the appropriate shifts and AND operations,
5747 simplifying at each stage.
5749 The function make_compound_operation is called to convert an expression
5750 consisting of shifts and ANDs into the equivalent compound expression.
5751 It is the inverse of this function, loosely speaking. */
5753 static rtx
5754 expand_compound_operation (rtx x)
5756 unsigned HOST_WIDE_INT pos = 0, len;
5757 int unsignedp = 0;
5758 unsigned int modewidth;
5759 rtx tem;
5761 switch (GET_CODE (x))
5763 case ZERO_EXTEND:
5764 unsignedp = 1;
5765 case SIGN_EXTEND:
5766 /* We can't necessarily use a const_int for a multiword mode;
5767 it depends on implicitly extending the value.
5768 Since we don't know the right way to extend it,
5769 we can't tell whether the implicit way is right.
5771 Even for a mode that is no wider than a const_int,
5772 we can't win, because we need to sign extend one of its bits through
5773 the rest of it, and we don't know which bit. */
5774 if (GET_CODE (XEXP (x, 0)) == CONST_INT)
5775 return x;
5777 /* Return if (subreg:MODE FROM 0) is not a safe replacement for
5778 (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM
5779 because (SUBREG (MEM...)) is guaranteed to cause the MEM to be
5780 reloaded. If not for that, MEM's would very rarely be safe.
5782 Reject MODEs bigger than a word, because we might not be able
5783 to reference a two-register group starting with an arbitrary register
5784 (and currently gen_lowpart might crash for a SUBREG). */
5786 if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD)
5787 return x;
5789 /* Reject MODEs that aren't scalar integers because turning vector
5790 or complex modes into shifts causes problems. */
5792 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5793 return x;
5795 len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)));
5796 /* If the inner object has VOIDmode (the only way this can happen
5797 is if it is an ASM_OPERANDS), we can't do anything since we don't
5798 know how much masking to do. */
5799 if (len == 0)
5800 return x;
5802 break;
5804 case ZERO_EXTRACT:
5805 unsignedp = 1;
5807 /* ... fall through ... */
5809 case SIGN_EXTRACT:
5810 /* If the operand is a CLOBBER, just return it. */
5811 if (GET_CODE (XEXP (x, 0)) == CLOBBER)
5812 return XEXP (x, 0);
5814 if (GET_CODE (XEXP (x, 1)) != CONST_INT
5815 || GET_CODE (XEXP (x, 2)) != CONST_INT
5816 || GET_MODE (XEXP (x, 0)) == VOIDmode)
5817 return x;
5819 /* Reject MODEs that aren't scalar integers because turning vector
5820 or complex modes into shifts causes problems. */
5822 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5823 return x;
5825 len = INTVAL (XEXP (x, 1));
5826 pos = INTVAL (XEXP (x, 2));
5828 /* If this goes outside the object being extracted, replace the object
5829 with a (use (mem ...)) construct that only combine understands
5830 and is used only for this purpose. */
5831 if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))
5832 SUBST (XEXP (x, 0), gen_rtx_USE (GET_MODE (x), XEXP (x, 0)));
5834 if (BITS_BIG_ENDIAN)
5835 pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos;
5837 break;
5839 default:
5840 return x;
5842 /* Convert sign extension to zero extension, if we know that the high
5843 bit is not set, as this is easier to optimize. It will be converted
5844 back to cheaper alternative in make_extraction. */
5845 if (GET_CODE (x) == SIGN_EXTEND
5846 && (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5847 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
5848 & ~(((unsigned HOST_WIDE_INT)
5849 GET_MODE_MASK (GET_MODE (XEXP (x, 0))))
5850 >> 1))
5851 == 0)))
5853 rtx temp = gen_rtx_ZERO_EXTEND (GET_MODE (x), XEXP (x, 0));
5854 rtx temp2 = expand_compound_operation (temp);
5856 /* Make sure this is a profitable operation. */
5857 if (rtx_cost (x, SET) > rtx_cost (temp2, SET))
5858 return temp2;
5859 else if (rtx_cost (x, SET) > rtx_cost (temp, SET))
5860 return temp;
5861 else
5862 return x;
5865 /* We can optimize some special cases of ZERO_EXTEND. */
5866 if (GET_CODE (x) == ZERO_EXTEND)
5868 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we
5869 know that the last value didn't have any inappropriate bits
5870 set. */
5871 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5872 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5873 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5874 && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x))
5875 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5876 return XEXP (XEXP (x, 0), 0);
5878 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5879 if (GET_CODE (XEXP (x, 0)) == SUBREG
5880 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5881 && subreg_lowpart_p (XEXP (x, 0))
5882 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5883 && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x))
5884 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5885 return SUBREG_REG (XEXP (x, 0));
5887 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo
5888 is a comparison and STORE_FLAG_VALUE permits. This is like
5889 the first case, but it works even when GET_MODE (x) is larger
5890 than HOST_WIDE_INT. */
5891 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5892 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5893 && COMPARISON_P (XEXP (XEXP (x, 0), 0))
5894 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5895 <= HOST_BITS_PER_WIDE_INT)
5896 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5897 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5898 return XEXP (XEXP (x, 0), 0);
5900 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5901 if (GET_CODE (XEXP (x, 0)) == SUBREG
5902 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5903 && subreg_lowpart_p (XEXP (x, 0))
5904 && COMPARISON_P (SUBREG_REG (XEXP (x, 0)))
5905 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5906 <= HOST_BITS_PER_WIDE_INT)
5907 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5908 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5909 return SUBREG_REG (XEXP (x, 0));
5913 /* If we reach here, we want to return a pair of shifts. The inner
5914 shift is a left shift of BITSIZE - POS - LEN bits. The outer
5915 shift is a right shift of BITSIZE - LEN bits. It is arithmetic or
5916 logical depending on the value of UNSIGNEDP.
5918 If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be
5919 converted into an AND of a shift.
5921 We must check for the case where the left shift would have a negative
5922 count. This can happen in a case like (x >> 31) & 255 on machines
5923 that can't shift by a constant. On those machines, we would first
5924 combine the shift with the AND to produce a variable-position
5925 extraction. Then the constant of 31 would be substituted in to produce
5926 a such a position. */
5928 modewidth = GET_MODE_BITSIZE (GET_MODE (x));
5929 if (modewidth + len >= pos)
5930 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT,
5931 GET_MODE (x),
5932 simplify_shift_const (NULL_RTX, ASHIFT,
5933 GET_MODE (x),
5934 XEXP (x, 0),
5935 modewidth - pos - len),
5936 modewidth - len);
5938 else if (unsignedp && len < HOST_BITS_PER_WIDE_INT)
5939 tem = simplify_and_const_int (NULL_RTX, GET_MODE (x),
5940 simplify_shift_const (NULL_RTX, LSHIFTRT,
5941 GET_MODE (x),
5942 XEXP (x, 0), pos),
5943 ((HOST_WIDE_INT) 1 << len) - 1);
5944 else
5945 /* Any other cases we can't handle. */
5946 return x;
5948 /* If we couldn't do this for some reason, return the original
5949 expression. */
5950 if (GET_CODE (tem) == CLOBBER)
5951 return x;
5953 return tem;
5956 /* X is a SET which contains an assignment of one object into
5957 a part of another (such as a bit-field assignment, STRICT_LOW_PART,
5958 or certain SUBREGS). If possible, convert it into a series of
5959 logical operations.
5961 We half-heartedly support variable positions, but do not at all
5962 support variable lengths. */
5964 static rtx
5965 expand_field_assignment (rtx x)
5967 rtx inner;
5968 rtx pos; /* Always counts from low bit. */
5969 int len;
5970 rtx mask, cleared, masked;
5971 enum machine_mode compute_mode;
5973 /* Loop until we find something we can't simplify. */
5974 while (1)
5976 if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
5977 && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG)
5979 inner = SUBREG_REG (XEXP (SET_DEST (x), 0));
5980 len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)));
5981 pos = GEN_INT (subreg_lsb (XEXP (SET_DEST (x), 0)));
5983 else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
5984 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT)
5986 inner = XEXP (SET_DEST (x), 0);
5987 len = INTVAL (XEXP (SET_DEST (x), 1));
5988 pos = XEXP (SET_DEST (x), 2);
5990 /* If the position is constant and spans the width of INNER,
5991 surround INNER with a USE to indicate this. */
5992 if (GET_CODE (pos) == CONST_INT
5993 && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner)))
5994 inner = gen_rtx_USE (GET_MODE (SET_DEST (x)), inner);
5996 if (BITS_BIG_ENDIAN)
5998 if (GET_CODE (pos) == CONST_INT)
5999 pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len
6000 - INTVAL (pos));
6001 else if (GET_CODE (pos) == MINUS
6002 && GET_CODE (XEXP (pos, 1)) == CONST_INT
6003 && (INTVAL (XEXP (pos, 1))
6004 == GET_MODE_BITSIZE (GET_MODE (inner)) - len))
6005 /* If position is ADJUST - X, new position is X. */
6006 pos = XEXP (pos, 0);
6007 else
6008 pos = simplify_gen_binary (MINUS, GET_MODE (pos),
6009 GEN_INT (GET_MODE_BITSIZE (
6010 GET_MODE (inner))
6011 - len),
6012 pos);
6016 /* A SUBREG between two modes that occupy the same numbers of words
6017 can be done by moving the SUBREG to the source. */
6018 else if (GET_CODE (SET_DEST (x)) == SUBREG
6019 /* We need SUBREGs to compute nonzero_bits properly. */
6020 && nonzero_sign_valid
6021 && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
6022 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
6023 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
6024 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))
6026 x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)),
6027 gen_lowpart
6028 (GET_MODE (SUBREG_REG (SET_DEST (x))),
6029 SET_SRC (x)));
6030 continue;
6032 else
6033 break;
6035 while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
6036 inner = SUBREG_REG (inner);
6038 compute_mode = GET_MODE (inner);
6040 /* Don't attempt bitwise arithmetic on non scalar integer modes. */
6041 if (! SCALAR_INT_MODE_P (compute_mode))
6043 enum machine_mode imode;
6045 /* Don't do anything for vector or complex integral types. */
6046 if (! FLOAT_MODE_P (compute_mode))
6047 break;
6049 /* Try to find an integral mode to pun with. */
6050 imode = mode_for_size (GET_MODE_BITSIZE (compute_mode), MODE_INT, 0);
6051 if (imode == BLKmode)
6052 break;
6054 compute_mode = imode;
6055 inner = gen_lowpart (imode, inner);
6058 /* Compute a mask of LEN bits, if we can do this on the host machine. */
6059 if (len >= HOST_BITS_PER_WIDE_INT)
6060 break;
6062 /* Now compute the equivalent expression. Make a copy of INNER
6063 for the SET_DEST in case it is a MEM into which we will substitute;
6064 we don't want shared RTL in that case. */
6065 mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1);
6066 cleared = simplify_gen_binary (AND, compute_mode,
6067 simplify_gen_unary (NOT, compute_mode,
6068 simplify_gen_binary (ASHIFT,
6069 compute_mode,
6070 mask, pos),
6071 compute_mode),
6072 inner);
6073 masked = simplify_gen_binary (ASHIFT, compute_mode,
6074 simplify_gen_binary (
6075 AND, compute_mode,
6076 gen_lowpart (compute_mode, SET_SRC (x)),
6077 mask),
6078 pos);
6080 x = gen_rtx_SET (VOIDmode, copy_rtx (inner),
6081 simplify_gen_binary (IOR, compute_mode,
6082 cleared, masked));
6085 return x;
6088 /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero,
6089 it is an RTX that represents a variable starting position; otherwise,
6090 POS is the (constant) starting bit position (counted from the LSB).
6092 INNER may be a USE. This will occur when we started with a bitfield
6093 that went outside the boundary of the object in memory, which is
6094 allowed on most machines. To isolate this case, we produce a USE
6095 whose mode is wide enough and surround the MEM with it. The only
6096 code that understands the USE is this routine. If it is not removed,
6097 it will cause the resulting insn not to match.
6099 UNSIGNEDP is nonzero for an unsigned reference and zero for a
6100 signed reference.
6102 IN_DEST is nonzero if this is a reference in the destination of a
6103 SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero,
6104 a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will
6105 be used.
6107 IN_COMPARE is nonzero if we are in a COMPARE. This means that a
6108 ZERO_EXTRACT should be built even for bits starting at bit 0.
6110 MODE is the desired mode of the result (if IN_DEST == 0).
6112 The result is an RTX for the extraction or NULL_RTX if the target
6113 can't handle it. */
6115 static rtx
6116 make_extraction (enum machine_mode mode, rtx inner, HOST_WIDE_INT pos,
6117 rtx pos_rtx, unsigned HOST_WIDE_INT len, int unsignedp,
6118 int in_dest, int in_compare)
6120 /* This mode describes the size of the storage area
6121 to fetch the overall value from. Within that, we
6122 ignore the POS lowest bits, etc. */
6123 enum machine_mode is_mode = GET_MODE (inner);
6124 enum machine_mode inner_mode;
6125 enum machine_mode wanted_inner_mode = byte_mode;
6126 enum machine_mode wanted_inner_reg_mode = word_mode;
6127 enum machine_mode pos_mode = word_mode;
6128 enum machine_mode extraction_mode = word_mode;
6129 enum machine_mode tmode = mode_for_size (len, MODE_INT, 1);
6130 int spans_byte = 0;
6131 rtx new = 0;
6132 rtx orig_pos_rtx = pos_rtx;
6133 HOST_WIDE_INT orig_pos;
6135 /* Get some information about INNER and get the innermost object. */
6136 if (GET_CODE (inner) == USE)
6137 /* (use:SI (mem:QI foo)) stands for (mem:SI foo). */
6138 /* We don't need to adjust the position because we set up the USE
6139 to pretend that it was a full-word object. */
6140 spans_byte = 1, inner = XEXP (inner, 0);
6141 else if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
6143 /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...),
6144 consider just the QI as the memory to extract from.
6145 The subreg adds or removes high bits; its mode is
6146 irrelevant to the meaning of this extraction,
6147 since POS and LEN count from the lsb. */
6148 if (MEM_P (SUBREG_REG (inner)))
6149 is_mode = GET_MODE (SUBREG_REG (inner));
6150 inner = SUBREG_REG (inner);
6152 else if (GET_CODE (inner) == ASHIFT
6153 && GET_CODE (XEXP (inner, 1)) == CONST_INT
6154 && pos_rtx == 0 && pos == 0
6155 && len > (unsigned HOST_WIDE_INT) INTVAL (XEXP (inner, 1)))
6157 /* We're extracting the least significant bits of an rtx
6158 (ashift X (const_int C)), where LEN > C. Extract the
6159 least significant (LEN - C) bits of X, giving an rtx
6160 whose mode is MODE, then shift it left C times. */
6161 new = make_extraction (mode, XEXP (inner, 0),
6162 0, 0, len - INTVAL (XEXP (inner, 1)),
6163 unsignedp, in_dest, in_compare);
6164 if (new != 0)
6165 return gen_rtx_ASHIFT (mode, new, XEXP (inner, 1));
6168 inner_mode = GET_MODE (inner);
6170 if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT)
6171 pos = INTVAL (pos_rtx), pos_rtx = 0;
6173 /* See if this can be done without an extraction. We never can if the
6174 width of the field is not the same as that of some integer mode. For
6175 registers, we can only avoid the extraction if the position is at the
6176 low-order bit and this is either not in the destination or we have the
6177 appropriate STRICT_LOW_PART operation available.
6179 For MEM, we can avoid an extract if the field starts on an appropriate
6180 boundary and we can change the mode of the memory reference. However,
6181 we cannot directly access the MEM if we have a USE and the underlying
6182 MEM is not TMODE. This combination means that MEM was being used in a
6183 context where bits outside its mode were being referenced; that is only
6184 valid in bit-field insns. */
6186 if (tmode != BLKmode
6187 && ! (spans_byte && inner_mode != tmode)
6188 && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
6189 && !MEM_P (inner)
6190 && (! in_dest
6191 || (REG_P (inner)
6192 && have_insn_for (STRICT_LOW_PART, tmode))))
6193 || (MEM_P (inner) && pos_rtx == 0
6194 && (pos
6195 % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode)
6196 : BITS_PER_UNIT)) == 0
6197 /* We can't do this if we are widening INNER_MODE (it
6198 may not be aligned, for one thing). */
6199 && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode)
6200 && (inner_mode == tmode
6201 || (! mode_dependent_address_p (XEXP (inner, 0))
6202 && ! MEM_VOLATILE_P (inner))))))
6204 /* If INNER is a MEM, make a new MEM that encompasses just the desired
6205 field. If the original and current mode are the same, we need not
6206 adjust the offset. Otherwise, we do if bytes big endian.
6208 If INNER is not a MEM, get a piece consisting of just the field
6209 of interest (in this case POS % BITS_PER_WORD must be 0). */
6211 if (MEM_P (inner))
6213 HOST_WIDE_INT offset;
6215 /* POS counts from lsb, but make OFFSET count in memory order. */
6216 if (BYTES_BIG_ENDIAN)
6217 offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT;
6218 else
6219 offset = pos / BITS_PER_UNIT;
6221 new = adjust_address_nv (inner, tmode, offset);
6223 else if (REG_P (inner))
6225 if (tmode != inner_mode)
6227 /* We can't call gen_lowpart in a DEST since we
6228 always want a SUBREG (see below) and it would sometimes
6229 return a new hard register. */
6230 if (pos || in_dest)
6232 HOST_WIDE_INT final_word = pos / BITS_PER_WORD;
6234 if (WORDS_BIG_ENDIAN
6235 && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD)
6236 final_word = ((GET_MODE_SIZE (inner_mode)
6237 - GET_MODE_SIZE (tmode))
6238 / UNITS_PER_WORD) - final_word;
6240 final_word *= UNITS_PER_WORD;
6241 if (BYTES_BIG_ENDIAN &&
6242 GET_MODE_SIZE (inner_mode) > GET_MODE_SIZE (tmode))
6243 final_word += (GET_MODE_SIZE (inner_mode)
6244 - GET_MODE_SIZE (tmode)) % UNITS_PER_WORD;
6246 /* Avoid creating invalid subregs, for example when
6247 simplifying (x>>32)&255. */
6248 if (final_word >= GET_MODE_SIZE (inner_mode))
6249 return NULL_RTX;
6251 new = gen_rtx_SUBREG (tmode, inner, final_word);
6253 else
6254 new = gen_lowpart (tmode, inner);
6256 else
6257 new = inner;
6259 else
6260 new = force_to_mode (inner, tmode,
6261 len >= HOST_BITS_PER_WIDE_INT
6262 ? ~(unsigned HOST_WIDE_INT) 0
6263 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
6264 NULL_RTX, 0);
6266 /* If this extraction is going into the destination of a SET,
6267 make a STRICT_LOW_PART unless we made a MEM. */
6269 if (in_dest)
6270 return (MEM_P (new) ? new
6271 : (GET_CODE (new) != SUBREG
6272 ? gen_rtx_CLOBBER (tmode, const0_rtx)
6273 : gen_rtx_STRICT_LOW_PART (VOIDmode, new)));
6275 if (mode == tmode)
6276 return new;
6278 if (GET_CODE (new) == CONST_INT)
6279 return gen_int_mode (INTVAL (new), mode);
6281 /* If we know that no extraneous bits are set, and that the high
6282 bit is not set, convert the extraction to the cheaper of
6283 sign and zero extension, that are equivalent in these cases. */
6284 if (flag_expensive_optimizations
6285 && (GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
6286 && ((nonzero_bits (new, tmode)
6287 & ~(((unsigned HOST_WIDE_INT)
6288 GET_MODE_MASK (tmode))
6289 >> 1))
6290 == 0)))
6292 rtx temp = gen_rtx_ZERO_EXTEND (mode, new);
6293 rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new);
6295 /* Prefer ZERO_EXTENSION, since it gives more information to
6296 backends. */
6297 if (rtx_cost (temp, SET) <= rtx_cost (temp1, SET))
6298 return temp;
6299 return temp1;
6302 /* Otherwise, sign- or zero-extend unless we already are in the
6303 proper mode. */
6305 return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
6306 mode, new));
6309 /* Unless this is a COMPARE or we have a funny memory reference,
6310 don't do anything with zero-extending field extracts starting at
6311 the low-order bit since they are simple AND operations. */
6312 if (pos_rtx == 0 && pos == 0 && ! in_dest
6313 && ! in_compare && ! spans_byte && unsignedp)
6314 return 0;
6316 /* Unless we are allowed to span bytes or INNER is not MEM, reject this if
6317 we would be spanning bytes or if the position is not a constant and the
6318 length is not 1. In all other cases, we would only be going outside
6319 our object in cases when an original shift would have been
6320 undefined. */
6321 if (! spans_byte && MEM_P (inner)
6322 && ((pos_rtx == 0 && pos + len > GET_MODE_BITSIZE (is_mode))
6323 || (pos_rtx != 0 && len != 1)))
6324 return 0;
6326 /* Get the mode to use should INNER not be a MEM, the mode for the position,
6327 and the mode for the result. */
6328 if (in_dest && mode_for_extraction (EP_insv, -1) != MAX_MACHINE_MODE)
6330 wanted_inner_reg_mode = mode_for_extraction (EP_insv, 0);
6331 pos_mode = mode_for_extraction (EP_insv, 2);
6332 extraction_mode = mode_for_extraction (EP_insv, 3);
6335 if (! in_dest && unsignedp
6336 && mode_for_extraction (EP_extzv, -1) != MAX_MACHINE_MODE)
6338 wanted_inner_reg_mode = mode_for_extraction (EP_extzv, 1);
6339 pos_mode = mode_for_extraction (EP_extzv, 3);
6340 extraction_mode = mode_for_extraction (EP_extzv, 0);
6343 if (! in_dest && ! unsignedp
6344 && mode_for_extraction (EP_extv, -1) != MAX_MACHINE_MODE)
6346 wanted_inner_reg_mode = mode_for_extraction (EP_extv, 1);
6347 pos_mode = mode_for_extraction (EP_extv, 3);
6348 extraction_mode = mode_for_extraction (EP_extv, 0);
6351 /* Never narrow an object, since that might not be safe. */
6353 if (mode != VOIDmode
6354 && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode))
6355 extraction_mode = mode;
6357 if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode
6358 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6359 pos_mode = GET_MODE (pos_rtx);
6361 /* If this is not from memory, the desired mode is wanted_inner_reg_mode;
6362 if we have to change the mode of memory and cannot, the desired mode is
6363 EXTRACTION_MODE. */
6364 if (!MEM_P (inner))
6365 wanted_inner_mode = wanted_inner_reg_mode;
6366 else if (inner_mode != wanted_inner_mode
6367 && (mode_dependent_address_p (XEXP (inner, 0))
6368 || MEM_VOLATILE_P (inner)))
6369 wanted_inner_mode = extraction_mode;
6371 orig_pos = pos;
6373 if (BITS_BIG_ENDIAN)
6375 /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to
6376 BITS_BIG_ENDIAN style. If position is constant, compute new
6377 position. Otherwise, build subtraction.
6378 Note that POS is relative to the mode of the original argument.
6379 If it's a MEM we need to recompute POS relative to that.
6380 However, if we're extracting from (or inserting into) a register,
6381 we want to recompute POS relative to wanted_inner_mode. */
6382 int width = (MEM_P (inner)
6383 ? GET_MODE_BITSIZE (is_mode)
6384 : GET_MODE_BITSIZE (wanted_inner_mode));
6386 if (pos_rtx == 0)
6387 pos = width - len - pos;
6388 else
6389 pos_rtx
6390 = gen_rtx_MINUS (GET_MODE (pos_rtx), GEN_INT (width - len), pos_rtx);
6391 /* POS may be less than 0 now, but we check for that below.
6392 Note that it can only be less than 0 if !MEM_P (inner). */
6395 /* If INNER has a wider mode, make it smaller. If this is a constant
6396 extract, try to adjust the byte to point to the byte containing
6397 the value. */
6398 if (wanted_inner_mode != VOIDmode
6399 && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode)
6400 && ((MEM_P (inner)
6401 && (inner_mode == wanted_inner_mode
6402 || (! mode_dependent_address_p (XEXP (inner, 0))
6403 && ! MEM_VOLATILE_P (inner))))))
6405 int offset = 0;
6407 /* The computations below will be correct if the machine is big
6408 endian in both bits and bytes or little endian in bits and bytes.
6409 If it is mixed, we must adjust. */
6411 /* If bytes are big endian and we had a paradoxical SUBREG, we must
6412 adjust OFFSET to compensate. */
6413 if (BYTES_BIG_ENDIAN
6414 && ! spans_byte
6415 && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode))
6416 offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode);
6418 /* If this is a constant position, we can move to the desired byte. */
6419 if (pos_rtx == 0)
6421 offset += pos / BITS_PER_UNIT;
6422 pos %= GET_MODE_BITSIZE (wanted_inner_mode);
6425 if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN
6426 && ! spans_byte
6427 && is_mode != wanted_inner_mode)
6428 offset = (GET_MODE_SIZE (is_mode)
6429 - GET_MODE_SIZE (wanted_inner_mode) - offset);
6431 if (offset != 0 || inner_mode != wanted_inner_mode)
6432 inner = adjust_address_nv (inner, wanted_inner_mode, offset);
6435 /* If INNER is not memory, we can always get it into the proper mode. If we
6436 are changing its mode, POS must be a constant and smaller than the size
6437 of the new mode. */
6438 else if (!MEM_P (inner))
6440 if (GET_MODE (inner) != wanted_inner_mode
6441 && (pos_rtx != 0
6442 || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode)))
6443 return 0;
6445 inner = force_to_mode (inner, wanted_inner_mode,
6446 pos_rtx
6447 || len + orig_pos >= HOST_BITS_PER_WIDE_INT
6448 ? ~(unsigned HOST_WIDE_INT) 0
6449 : ((((unsigned HOST_WIDE_INT) 1 << len) - 1)
6450 << orig_pos),
6451 NULL_RTX, 0);
6454 /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we
6455 have to zero extend. Otherwise, we can just use a SUBREG. */
6456 if (pos_rtx != 0
6457 && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx)))
6459 rtx temp = gen_rtx_ZERO_EXTEND (pos_mode, pos_rtx);
6461 /* If we know that no extraneous bits are set, and that the high
6462 bit is not set, convert extraction to cheaper one - either
6463 SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these
6464 cases. */
6465 if (flag_expensive_optimizations
6466 && (GET_MODE_BITSIZE (GET_MODE (pos_rtx)) <= HOST_BITS_PER_WIDE_INT
6467 && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx))
6468 & ~(((unsigned HOST_WIDE_INT)
6469 GET_MODE_MASK (GET_MODE (pos_rtx)))
6470 >> 1))
6471 == 0)))
6473 rtx temp1 = gen_rtx_SIGN_EXTEND (pos_mode, pos_rtx);
6475 /* Prefer ZERO_EXTENSION, since it gives more information to
6476 backends. */
6477 if (rtx_cost (temp1, SET) < rtx_cost (temp, SET))
6478 temp = temp1;
6480 pos_rtx = temp;
6482 else if (pos_rtx != 0
6483 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6484 pos_rtx = gen_lowpart (pos_mode, pos_rtx);
6486 /* Make POS_RTX unless we already have it and it is correct. If we don't
6487 have a POS_RTX but we do have an ORIG_POS_RTX, the latter must
6488 be a CONST_INT. */
6489 if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos)
6490 pos_rtx = orig_pos_rtx;
6492 else if (pos_rtx == 0)
6493 pos_rtx = GEN_INT (pos);
6495 /* Make the required operation. See if we can use existing rtx. */
6496 new = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT,
6497 extraction_mode, inner, GEN_INT (len), pos_rtx);
6498 if (! in_dest)
6499 new = gen_lowpart (mode, new);
6501 return new;
6504 /* See if X contains an ASHIFT of COUNT or more bits that can be commuted
6505 with any other operations in X. Return X without that shift if so. */
6507 static rtx
6508 extract_left_shift (rtx x, int count)
6510 enum rtx_code code = GET_CODE (x);
6511 enum machine_mode mode = GET_MODE (x);
6512 rtx tem;
6514 switch (code)
6516 case ASHIFT:
6517 /* This is the shift itself. If it is wide enough, we will return
6518 either the value being shifted if the shift count is equal to
6519 COUNT or a shift for the difference. */
6520 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6521 && INTVAL (XEXP (x, 1)) >= count)
6522 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0),
6523 INTVAL (XEXP (x, 1)) - count);
6524 break;
6526 case NEG: case NOT:
6527 if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6528 return simplify_gen_unary (code, mode, tem, mode);
6530 break;
6532 case PLUS: case IOR: case XOR: case AND:
6533 /* If we can safely shift this constant and we find the inner shift,
6534 make a new operation. */
6535 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6536 && (INTVAL (XEXP (x, 1)) & ((((HOST_WIDE_INT) 1 << count)) - 1)) == 0
6537 && (tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6538 return simplify_gen_binary (code, mode, tem,
6539 GEN_INT (INTVAL (XEXP (x, 1)) >> count));
6541 break;
6543 default:
6544 break;
6547 return 0;
6550 /* Look at the expression rooted at X. Look for expressions
6551 equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND.
6552 Form these expressions.
6554 Return the new rtx, usually just X.
6556 Also, for machines like the VAX that don't have logical shift insns,
6557 try to convert logical to arithmetic shift operations in cases where
6558 they are equivalent. This undoes the canonicalizations to logical
6559 shifts done elsewhere.
6561 We try, as much as possible, to re-use rtl expressions to save memory.
6563 IN_CODE says what kind of expression we are processing. Normally, it is
6564 SET. In a memory address (inside a MEM, PLUS or minus, the latter two
6565 being kludges), it is MEM. When processing the arguments of a comparison
6566 or a COMPARE against zero, it is COMPARE. */
6568 static rtx
6569 make_compound_operation (rtx x, enum rtx_code in_code)
6571 enum rtx_code code = GET_CODE (x);
6572 enum machine_mode mode = GET_MODE (x);
6573 int mode_width = GET_MODE_BITSIZE (mode);
6574 rtx rhs, lhs;
6575 enum rtx_code next_code;
6576 int i;
6577 rtx new = 0;
6578 rtx tem;
6579 const char *fmt;
6581 /* Select the code to be used in recursive calls. Once we are inside an
6582 address, we stay there. If we have a comparison, set to COMPARE,
6583 but once inside, go back to our default of SET. */
6585 next_code = (code == MEM || code == PLUS || code == MINUS ? MEM
6586 : ((code == COMPARE || COMPARISON_P (x))
6587 && XEXP (x, 1) == const0_rtx) ? COMPARE
6588 : in_code == COMPARE ? SET : in_code);
6590 /* Process depending on the code of this operation. If NEW is set
6591 nonzero, it will be returned. */
6593 switch (code)
6595 case ASHIFT:
6596 /* Convert shifts by constants into multiplications if inside
6597 an address. */
6598 if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT
6599 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
6600 && INTVAL (XEXP (x, 1)) >= 0)
6602 new = make_compound_operation (XEXP (x, 0), next_code);
6603 new = gen_rtx_MULT (mode, new,
6604 GEN_INT ((HOST_WIDE_INT) 1
6605 << INTVAL (XEXP (x, 1))));
6607 break;
6609 case AND:
6610 /* If the second operand is not a constant, we can't do anything
6611 with it. */
6612 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6613 break;
6615 /* If the constant is a power of two minus one and the first operand
6616 is a logical right shift, make an extraction. */
6617 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6618 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6620 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6621 new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1,
6622 0, in_code == COMPARE);
6625 /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */
6626 else if (GET_CODE (XEXP (x, 0)) == SUBREG
6627 && subreg_lowpart_p (XEXP (x, 0))
6628 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT
6629 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6631 new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0),
6632 next_code);
6633 new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0,
6634 XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1,
6635 0, in_code == COMPARE);
6637 /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */
6638 else if ((GET_CODE (XEXP (x, 0)) == XOR
6639 || GET_CODE (XEXP (x, 0)) == IOR)
6640 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT
6641 && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT
6642 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6644 /* Apply the distributive law, and then try to make extractions. */
6645 new = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode,
6646 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0),
6647 XEXP (x, 1)),
6648 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1),
6649 XEXP (x, 1)));
6650 new = make_compound_operation (new, in_code);
6653 /* If we are have (and (rotate X C) M) and C is larger than the number
6654 of bits in M, this is an extraction. */
6656 else if (GET_CODE (XEXP (x, 0)) == ROTATE
6657 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6658 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0
6659 && i <= INTVAL (XEXP (XEXP (x, 0), 1)))
6661 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6662 new = make_extraction (mode, new,
6663 (GET_MODE_BITSIZE (mode)
6664 - INTVAL (XEXP (XEXP (x, 0), 1))),
6665 NULL_RTX, i, 1, 0, in_code == COMPARE);
6668 /* On machines without logical shifts, if the operand of the AND is
6669 a logical shift and our mask turns off all the propagated sign
6670 bits, we can replace the logical shift with an arithmetic shift. */
6671 else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6672 && !have_insn_for (LSHIFTRT, mode)
6673 && have_insn_for (ASHIFTRT, mode)
6674 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6675 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
6676 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
6677 && mode_width <= HOST_BITS_PER_WIDE_INT)
6679 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
6681 mask >>= INTVAL (XEXP (XEXP (x, 0), 1));
6682 if ((INTVAL (XEXP (x, 1)) & ~mask) == 0)
6683 SUBST (XEXP (x, 0),
6684 gen_rtx_ASHIFTRT (mode,
6685 make_compound_operation
6686 (XEXP (XEXP (x, 0), 0), next_code),
6687 XEXP (XEXP (x, 0), 1)));
6690 /* If the constant is one less than a power of two, this might be
6691 representable by an extraction even if no shift is present.
6692 If it doesn't end up being a ZERO_EXTEND, we will ignore it unless
6693 we are in a COMPARE. */
6694 else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6695 new = make_extraction (mode,
6696 make_compound_operation (XEXP (x, 0),
6697 next_code),
6698 0, NULL_RTX, i, 1, 0, in_code == COMPARE);
6700 /* If we are in a comparison and this is an AND with a power of two,
6701 convert this into the appropriate bit extract. */
6702 else if (in_code == COMPARE
6703 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
6704 new = make_extraction (mode,
6705 make_compound_operation (XEXP (x, 0),
6706 next_code),
6707 i, NULL_RTX, 1, 1, 0, 1);
6709 break;
6711 case LSHIFTRT:
6712 /* If the sign bit is known to be zero, replace this with an
6713 arithmetic shift. */
6714 if (have_insn_for (ASHIFTRT, mode)
6715 && ! have_insn_for (LSHIFTRT, mode)
6716 && mode_width <= HOST_BITS_PER_WIDE_INT
6717 && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0)
6719 new = gen_rtx_ASHIFTRT (mode,
6720 make_compound_operation (XEXP (x, 0),
6721 next_code),
6722 XEXP (x, 1));
6723 break;
6726 /* ... fall through ... */
6728 case ASHIFTRT:
6729 lhs = XEXP (x, 0);
6730 rhs = XEXP (x, 1);
6732 /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1,
6733 this is a SIGN_EXTRACT. */
6734 if (GET_CODE (rhs) == CONST_INT
6735 && GET_CODE (lhs) == ASHIFT
6736 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
6737 && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1)))
6739 new = make_compound_operation (XEXP (lhs, 0), next_code);
6740 new = make_extraction (mode, new,
6741 INTVAL (rhs) - INTVAL (XEXP (lhs, 1)),
6742 NULL_RTX, mode_width - INTVAL (rhs),
6743 code == LSHIFTRT, 0, in_code == COMPARE);
6744 break;
6747 /* See if we have operations between an ASHIFTRT and an ASHIFT.
6748 If so, try to merge the shifts into a SIGN_EXTEND. We could
6749 also do this for some cases of SIGN_EXTRACT, but it doesn't
6750 seem worth the effort; the case checked for occurs on Alpha. */
6752 if (!OBJECT_P (lhs)
6753 && ! (GET_CODE (lhs) == SUBREG
6754 && (OBJECT_P (SUBREG_REG (lhs))))
6755 && GET_CODE (rhs) == CONST_INT
6756 && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT
6757 && (new = extract_left_shift (lhs, INTVAL (rhs))) != 0)
6758 new = make_extraction (mode, make_compound_operation (new, next_code),
6759 0, NULL_RTX, mode_width - INTVAL (rhs),
6760 code == LSHIFTRT, 0, in_code == COMPARE);
6762 break;
6764 case SUBREG:
6765 /* Call ourselves recursively on the inner expression. If we are
6766 narrowing the object and it has a different RTL code from
6767 what it originally did, do this SUBREG as a force_to_mode. */
6769 tem = make_compound_operation (SUBREG_REG (x), in_code);
6770 if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x))
6771 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem))
6772 && subreg_lowpart_p (x))
6774 rtx newer = force_to_mode (tem, mode, ~(HOST_WIDE_INT) 0,
6775 NULL_RTX, 0);
6777 /* If we have something other than a SUBREG, we might have
6778 done an expansion, so rerun ourselves. */
6779 if (GET_CODE (newer) != SUBREG)
6780 newer = make_compound_operation (newer, in_code);
6782 return newer;
6785 /* If this is a paradoxical subreg, and the new code is a sign or
6786 zero extension, omit the subreg and widen the extension. If it
6787 is a regular subreg, we can still get rid of the subreg by not
6788 widening so much, or in fact removing the extension entirely. */
6789 if ((GET_CODE (tem) == SIGN_EXTEND
6790 || GET_CODE (tem) == ZERO_EXTEND)
6791 && subreg_lowpart_p (x))
6793 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (tem))
6794 || (GET_MODE_SIZE (mode) >
6795 GET_MODE_SIZE (GET_MODE (XEXP (tem, 0)))))
6797 if (! SCALAR_INT_MODE_P (mode))
6798 break;
6799 tem = gen_rtx_fmt_e (GET_CODE (tem), mode, XEXP (tem, 0));
6801 else
6802 tem = gen_lowpart (mode, XEXP (tem, 0));
6803 return tem;
6805 break;
6807 default:
6808 break;
6811 if (new)
6813 x = gen_lowpart (mode, new);
6814 code = GET_CODE (x);
6817 /* Now recursively process each operand of this operation. */
6818 fmt = GET_RTX_FORMAT (code);
6819 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6820 if (fmt[i] == 'e')
6822 new = make_compound_operation (XEXP (x, i), next_code);
6823 SUBST (XEXP (x, i), new);
6826 return x;
6829 /* Given M see if it is a value that would select a field of bits
6830 within an item, but not the entire word. Return -1 if not.
6831 Otherwise, return the starting position of the field, where 0 is the
6832 low-order bit.
6834 *PLEN is set to the length of the field. */
6836 static int
6837 get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen)
6839 /* Get the bit number of the first 1 bit from the right, -1 if none. */
6840 int pos = exact_log2 (m & -m);
6841 int len = 0;
6843 if (pos >= 0)
6844 /* Now shift off the low-order zero bits and see if we have a
6845 power of two minus 1. */
6846 len = exact_log2 ((m >> pos) + 1);
6848 if (len <= 0)
6849 pos = -1;
6851 *plen = len;
6852 return pos;
6855 /* See if X can be simplified knowing that we will only refer to it in
6856 MODE and will only refer to those bits that are nonzero in MASK.
6857 If other bits are being computed or if masking operations are done
6858 that select a superset of the bits in MASK, they can sometimes be
6859 ignored.
6861 Return a possibly simplified expression, but always convert X to
6862 MODE. If X is a CONST_INT, AND the CONST_INT with MASK.
6864 Also, if REG is nonzero and X is a register equal in value to REG,
6865 replace X with REG.
6867 If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK
6868 are all off in X. This is used when X will be complemented, by either
6869 NOT, NEG, or XOR. */
6871 static rtx
6872 force_to_mode (rtx x, enum machine_mode mode, unsigned HOST_WIDE_INT mask,
6873 rtx reg, int just_select)
6875 enum rtx_code code = GET_CODE (x);
6876 int next_select = just_select || code == XOR || code == NOT || code == NEG;
6877 enum machine_mode op_mode;
6878 unsigned HOST_WIDE_INT fuller_mask, nonzero;
6879 rtx op0, op1, temp;
6881 /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the
6882 code below will do the wrong thing since the mode of such an
6883 expression is VOIDmode.
6885 Also do nothing if X is a CLOBBER; this can happen if X was
6886 the return value from a call to gen_lowpart. */
6887 if (code == CALL || code == ASM_OPERANDS || code == CLOBBER)
6888 return x;
6890 /* We want to perform the operation is its present mode unless we know
6891 that the operation is valid in MODE, in which case we do the operation
6892 in MODE. */
6893 op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x))
6894 && have_insn_for (code, mode))
6895 ? mode : GET_MODE (x));
6897 /* It is not valid to do a right-shift in a narrower mode
6898 than the one it came in with. */
6899 if ((code == LSHIFTRT || code == ASHIFTRT)
6900 && GET_MODE_BITSIZE (mode) < GET_MODE_BITSIZE (GET_MODE (x)))
6901 op_mode = GET_MODE (x);
6903 /* Truncate MASK to fit OP_MODE. */
6904 if (op_mode)
6905 mask &= GET_MODE_MASK (op_mode);
6907 /* When we have an arithmetic operation, or a shift whose count we
6908 do not know, we need to assume that all bits up to the highest-order
6909 bit in MASK will be needed. This is how we form such a mask. */
6910 if (mask & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)))
6911 fuller_mask = ~(unsigned HOST_WIDE_INT) 0;
6912 else
6913 fuller_mask = (((unsigned HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1))
6914 - 1);
6916 /* Determine what bits of X are guaranteed to be (non)zero. */
6917 nonzero = nonzero_bits (x, mode);
6919 /* If none of the bits in X are needed, return a zero. */
6920 if (! just_select && (nonzero & mask) == 0)
6921 x = const0_rtx;
6923 /* If X is a CONST_INT, return a new one. Do this here since the
6924 test below will fail. */
6925 if (GET_CODE (x) == CONST_INT)
6927 if (SCALAR_INT_MODE_P (mode))
6928 return gen_int_mode (INTVAL (x) & mask, mode);
6929 else
6931 x = GEN_INT (INTVAL (x) & mask);
6932 return gen_lowpart_common (mode, x);
6936 /* If X is narrower than MODE and we want all the bits in X's mode, just
6937 get X in the proper mode. */
6938 if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)
6939 && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0)
6940 return gen_lowpart (mode, x);
6942 switch (code)
6944 case CLOBBER:
6945 /* If X is a (clobber (const_int)), return it since we know we are
6946 generating something that won't match. */
6947 return x;
6949 case USE:
6950 /* X is a (use (mem ..)) that was made from a bit-field extraction that
6951 spanned the boundary of the MEM. If we are now masking so it is
6952 within that boundary, we don't need the USE any more. */
6953 if (! BITS_BIG_ENDIAN
6954 && (mask & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
6955 return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select);
6956 break;
6958 case SIGN_EXTEND:
6959 case ZERO_EXTEND:
6960 case ZERO_EXTRACT:
6961 case SIGN_EXTRACT:
6962 x = expand_compound_operation (x);
6963 if (GET_CODE (x) != code)
6964 return force_to_mode (x, mode, mask, reg, next_select);
6965 break;
6967 case REG:
6968 if (reg != 0 && (rtx_equal_p (get_last_value (reg), x)
6969 || rtx_equal_p (reg, get_last_value (x))))
6970 x = reg;
6971 break;
6973 case SUBREG:
6974 if (subreg_lowpart_p (x)
6975 /* We can ignore the effect of this SUBREG if it narrows the mode or
6976 if the constant masks to zero all the bits the mode doesn't
6977 have. */
6978 && ((GET_MODE_SIZE (GET_MODE (x))
6979 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
6980 || (0 == (mask
6981 & GET_MODE_MASK (GET_MODE (x))
6982 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))))))
6983 return force_to_mode (SUBREG_REG (x), mode, mask, reg, next_select);
6984 break;
6986 case AND:
6987 /* If this is an AND with a constant, convert it into an AND
6988 whose constant is the AND of that constant with MASK. If it
6989 remains an AND of MASK, delete it since it is redundant. */
6991 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6993 x = simplify_and_const_int (x, op_mode, XEXP (x, 0),
6994 mask & INTVAL (XEXP (x, 1)));
6996 /* If X is still an AND, see if it is an AND with a mask that
6997 is just some low-order bits. If so, and it is MASK, we don't
6998 need it. */
7000 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
7001 && ((INTVAL (XEXP (x, 1)) & GET_MODE_MASK (GET_MODE (x)))
7002 == mask))
7003 x = XEXP (x, 0);
7005 /* If it remains an AND, try making another AND with the bits
7006 in the mode mask that aren't in MASK turned on. If the
7007 constant in the AND is wide enough, this might make a
7008 cheaper constant. */
7010 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
7011 && GET_MODE_MASK (GET_MODE (x)) != mask
7012 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
7014 HOST_WIDE_INT cval = (INTVAL (XEXP (x, 1))
7015 | (GET_MODE_MASK (GET_MODE (x)) & ~mask));
7016 int width = GET_MODE_BITSIZE (GET_MODE (x));
7017 rtx y;
7019 /* If MODE is narrower than HOST_WIDE_INT and CVAL is a negative
7020 number, sign extend it. */
7021 if (width > 0 && width < HOST_BITS_PER_WIDE_INT
7022 && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
7023 cval |= (HOST_WIDE_INT) -1 << width;
7025 y = simplify_gen_binary (AND, GET_MODE (x),
7026 XEXP (x, 0), GEN_INT (cval));
7027 if (rtx_cost (y, SET) < rtx_cost (x, SET))
7028 x = y;
7031 break;
7034 goto binop;
7036 case PLUS:
7037 /* In (and (plus FOO C1) M), if M is a mask that just turns off
7038 low-order bits (as in an alignment operation) and FOO is already
7039 aligned to that boundary, mask C1 to that boundary as well.
7040 This may eliminate that PLUS and, later, the AND. */
7043 unsigned int width = GET_MODE_BITSIZE (mode);
7044 unsigned HOST_WIDE_INT smask = mask;
7046 /* If MODE is narrower than HOST_WIDE_INT and mask is a negative
7047 number, sign extend it. */
7049 if (width < HOST_BITS_PER_WIDE_INT
7050 && (smask & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
7051 smask |= (HOST_WIDE_INT) -1 << width;
7053 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7054 && exact_log2 (- smask) >= 0
7055 && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0
7056 && (INTVAL (XEXP (x, 1)) & ~smask) != 0)
7057 return force_to_mode (plus_constant (XEXP (x, 0),
7058 (INTVAL (XEXP (x, 1)) & smask)),
7059 mode, smask, reg, next_select);
7062 /* ... fall through ... */
7064 case MULT:
7065 /* For PLUS, MINUS and MULT, we need any bits less significant than the
7066 most significant bit in MASK since carries from those bits will
7067 affect the bits we are interested in. */
7068 mask = fuller_mask;
7069 goto binop;
7071 case MINUS:
7072 /* If X is (minus C Y) where C's least set bit is larger than any bit
7073 in the mask, then we may replace with (neg Y). */
7074 if (GET_CODE (XEXP (x, 0)) == CONST_INT
7075 && (((unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 0))
7076 & -INTVAL (XEXP (x, 0))))
7077 > mask))
7079 x = simplify_gen_unary (NEG, GET_MODE (x), XEXP (x, 1),
7080 GET_MODE (x));
7081 return force_to_mode (x, mode, mask, reg, next_select);
7084 /* Similarly, if C contains every bit in the fuller_mask, then we may
7085 replace with (not Y). */
7086 if (GET_CODE (XEXP (x, 0)) == CONST_INT
7087 && ((INTVAL (XEXP (x, 0)) | (HOST_WIDE_INT) fuller_mask)
7088 == INTVAL (XEXP (x, 0))))
7090 x = simplify_gen_unary (NOT, GET_MODE (x),
7091 XEXP (x, 1), GET_MODE (x));
7092 return force_to_mode (x, mode, mask, reg, next_select);
7095 mask = fuller_mask;
7096 goto binop;
7098 case IOR:
7099 case XOR:
7100 /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and
7101 LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...)
7102 operation which may be a bitfield extraction. Ensure that the
7103 constant we form is not wider than the mode of X. */
7105 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7106 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7107 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7108 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
7109 && GET_CODE (XEXP (x, 1)) == CONST_INT
7110 && ((INTVAL (XEXP (XEXP (x, 0), 1))
7111 + floor_log2 (INTVAL (XEXP (x, 1))))
7112 < GET_MODE_BITSIZE (GET_MODE (x)))
7113 && (INTVAL (XEXP (x, 1))
7114 & ~nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0)
7116 temp = GEN_INT ((INTVAL (XEXP (x, 1)) & mask)
7117 << INTVAL (XEXP (XEXP (x, 0), 1)));
7118 temp = simplify_gen_binary (GET_CODE (x), GET_MODE (x),
7119 XEXP (XEXP (x, 0), 0), temp);
7120 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), temp,
7121 XEXP (XEXP (x, 0), 1));
7122 return force_to_mode (x, mode, mask, reg, next_select);
7125 binop:
7126 /* For most binary operations, just propagate into the operation and
7127 change the mode if we have an operation of that mode. */
7129 op0 = gen_lowpart (op_mode,
7130 force_to_mode (XEXP (x, 0), mode, mask,
7131 reg, next_select));
7132 op1 = gen_lowpart (op_mode,
7133 force_to_mode (XEXP (x, 1), mode, mask,
7134 reg, next_select));
7136 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7137 x = simplify_gen_binary (code, op_mode, op0, op1);
7138 break;
7140 case ASHIFT:
7141 /* For left shifts, do the same, but just for the first operand.
7142 However, we cannot do anything with shifts where we cannot
7143 guarantee that the counts are smaller than the size of the mode
7144 because such a count will have a different meaning in a
7145 wider mode. */
7147 if (! (GET_CODE (XEXP (x, 1)) == CONST_INT
7148 && INTVAL (XEXP (x, 1)) >= 0
7149 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode))
7150 && ! (GET_MODE (XEXP (x, 1)) != VOIDmode
7151 && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1)))
7152 < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode))))
7153 break;
7155 /* If the shift count is a constant and we can do arithmetic in
7156 the mode of the shift, refine which bits we need. Otherwise, use the
7157 conservative form of the mask. */
7158 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7159 && INTVAL (XEXP (x, 1)) >= 0
7160 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (op_mode)
7161 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7162 mask >>= INTVAL (XEXP (x, 1));
7163 else
7164 mask = fuller_mask;
7166 op0 = gen_lowpart (op_mode,
7167 force_to_mode (XEXP (x, 0), op_mode,
7168 mask, reg, next_select));
7170 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7171 x = simplify_gen_binary (code, op_mode, op0, XEXP (x, 1));
7172 break;
7174 case LSHIFTRT:
7175 /* Here we can only do something if the shift count is a constant,
7176 this shift constant is valid for the host, and we can do arithmetic
7177 in OP_MODE. */
7179 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7180 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
7181 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7183 rtx inner = XEXP (x, 0);
7184 unsigned HOST_WIDE_INT inner_mask;
7186 /* Select the mask of the bits we need for the shift operand. */
7187 inner_mask = mask << INTVAL (XEXP (x, 1));
7189 /* We can only change the mode of the shift if we can do arithmetic
7190 in the mode of the shift and INNER_MASK is no wider than the
7191 width of X's mode. */
7192 if ((inner_mask & ~GET_MODE_MASK (GET_MODE (x))) != 0)
7193 op_mode = GET_MODE (x);
7195 inner = force_to_mode (inner, op_mode, inner_mask, reg, next_select);
7197 if (GET_MODE (x) != op_mode || inner != XEXP (x, 0))
7198 x = simplify_gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1));
7201 /* If we have (and (lshiftrt FOO C1) C2) where the combination of the
7202 shift and AND produces only copies of the sign bit (C2 is one less
7203 than a power of two), we can do this with just a shift. */
7205 if (GET_CODE (x) == LSHIFTRT
7206 && GET_CODE (XEXP (x, 1)) == CONST_INT
7207 /* The shift puts one of the sign bit copies in the least significant
7208 bit. */
7209 && ((INTVAL (XEXP (x, 1))
7210 + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))))
7211 >= GET_MODE_BITSIZE (GET_MODE (x)))
7212 && exact_log2 (mask + 1) >= 0
7213 /* Number of bits left after the shift must be more than the mask
7214 needs. */
7215 && ((INTVAL (XEXP (x, 1)) + exact_log2 (mask + 1))
7216 <= GET_MODE_BITSIZE (GET_MODE (x)))
7217 /* Must be more sign bit copies than the mask needs. */
7218 && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
7219 >= exact_log2 (mask + 1)))
7220 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7221 GEN_INT (GET_MODE_BITSIZE (GET_MODE (x))
7222 - exact_log2 (mask + 1)));
7224 goto shiftrt;
7226 case ASHIFTRT:
7227 /* If we are just looking for the sign bit, we don't need this shift at
7228 all, even if it has a variable count. */
7229 if (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
7230 && (mask == ((unsigned HOST_WIDE_INT) 1
7231 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
7232 return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select);
7234 /* If this is a shift by a constant, get a mask that contains those bits
7235 that are not copies of the sign bit. We then have two cases: If
7236 MASK only includes those bits, this can be a logical shift, which may
7237 allow simplifications. If MASK is a single-bit field not within
7238 those bits, we are requesting a copy of the sign bit and hence can
7239 shift the sign bit to the appropriate location. */
7241 if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0
7242 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT)
7244 int i = -1;
7246 /* If the considered data is wider than HOST_WIDE_INT, we can't
7247 represent a mask for all its bits in a single scalar.
7248 But we only care about the lower bits, so calculate these. */
7250 if (GET_MODE_BITSIZE (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT)
7252 nonzero = ~(HOST_WIDE_INT) 0;
7254 /* GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7255 is the number of bits a full-width mask would have set.
7256 We need only shift if these are fewer than nonzero can
7257 hold. If not, we must keep all bits set in nonzero. */
7259 if (GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7260 < HOST_BITS_PER_WIDE_INT)
7261 nonzero >>= INTVAL (XEXP (x, 1))
7262 + HOST_BITS_PER_WIDE_INT
7263 - GET_MODE_BITSIZE (GET_MODE (x)) ;
7265 else
7267 nonzero = GET_MODE_MASK (GET_MODE (x));
7268 nonzero >>= INTVAL (XEXP (x, 1));
7271 if ((mask & ~nonzero) == 0
7272 || (i = exact_log2 (mask)) >= 0)
7274 x = simplify_shift_const
7275 (x, LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7276 i < 0 ? INTVAL (XEXP (x, 1))
7277 : GET_MODE_BITSIZE (GET_MODE (x)) - 1 - i);
7279 if (GET_CODE (x) != ASHIFTRT)
7280 return force_to_mode (x, mode, mask, reg, next_select);
7284 /* If MASK is 1, convert this to an LSHIFTRT. This can be done
7285 even if the shift count isn't a constant. */
7286 if (mask == 1)
7287 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7288 XEXP (x, 0), XEXP (x, 1));
7290 shiftrt:
7292 /* If this is a zero- or sign-extension operation that just affects bits
7293 we don't care about, remove it. Be sure the call above returned
7294 something that is still a shift. */
7296 if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT)
7297 && GET_CODE (XEXP (x, 1)) == CONST_INT
7298 && INTVAL (XEXP (x, 1)) >= 0
7299 && (INTVAL (XEXP (x, 1))
7300 <= GET_MODE_BITSIZE (GET_MODE (x)) - (floor_log2 (mask) + 1))
7301 && GET_CODE (XEXP (x, 0)) == ASHIFT
7302 && XEXP (XEXP (x, 0), 1) == XEXP (x, 1))
7303 return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask,
7304 reg, next_select);
7306 break;
7308 case ROTATE:
7309 case ROTATERT:
7310 /* If the shift count is constant and we can do computations
7311 in the mode of X, compute where the bits we care about are.
7312 Otherwise, we can't do anything. Don't change the mode of
7313 the shift or propagate MODE into the shift, though. */
7314 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7315 && INTVAL (XEXP (x, 1)) >= 0)
7317 temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE,
7318 GET_MODE (x), GEN_INT (mask),
7319 XEXP (x, 1));
7320 if (temp && GET_CODE (temp) == CONST_INT)
7321 SUBST (XEXP (x, 0),
7322 force_to_mode (XEXP (x, 0), GET_MODE (x),
7323 INTVAL (temp), reg, next_select));
7325 break;
7327 case NEG:
7328 /* If we just want the low-order bit, the NEG isn't needed since it
7329 won't change the low-order bit. */
7330 if (mask == 1)
7331 return force_to_mode (XEXP (x, 0), mode, mask, reg, just_select);
7333 /* We need any bits less significant than the most significant bit in
7334 MASK since carries from those bits will affect the bits we are
7335 interested in. */
7336 mask = fuller_mask;
7337 goto unop;
7339 case NOT:
7340 /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the
7341 same as the XOR case above. Ensure that the constant we form is not
7342 wider than the mode of X. */
7344 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7345 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7346 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7347 && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask)
7348 < GET_MODE_BITSIZE (GET_MODE (x)))
7349 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT)
7351 temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)),
7352 GET_MODE (x));
7353 temp = simplify_gen_binary (XOR, GET_MODE (x),
7354 XEXP (XEXP (x, 0), 0), temp);
7355 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7356 temp, XEXP (XEXP (x, 0), 1));
7358 return force_to_mode (x, mode, mask, reg, next_select);
7361 /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must
7362 use the full mask inside the NOT. */
7363 mask = fuller_mask;
7365 unop:
7366 op0 = gen_lowpart (op_mode,
7367 force_to_mode (XEXP (x, 0), mode, mask,
7368 reg, next_select));
7369 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7370 x = simplify_gen_unary (code, op_mode, op0, op_mode);
7371 break;
7373 case NE:
7374 /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included
7375 in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero,
7376 which is equal to STORE_FLAG_VALUE. */
7377 if ((mask & ~STORE_FLAG_VALUE) == 0 && XEXP (x, 1) == const0_rtx
7378 && GET_MODE (XEXP (x, 0)) == mode
7379 && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0
7380 && (nonzero_bits (XEXP (x, 0), mode)
7381 == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE))
7382 return force_to_mode (XEXP (x, 0), mode, mask, reg, next_select);
7384 break;
7386 case IF_THEN_ELSE:
7387 /* We have no way of knowing if the IF_THEN_ELSE can itself be
7388 written in a narrower mode. We play it safe and do not do so. */
7390 SUBST (XEXP (x, 1),
7391 gen_lowpart (GET_MODE (x),
7392 force_to_mode (XEXP (x, 1), mode,
7393 mask, reg, next_select)));
7394 SUBST (XEXP (x, 2),
7395 gen_lowpart (GET_MODE (x),
7396 force_to_mode (XEXP (x, 2), mode,
7397 mask, reg, next_select)));
7398 break;
7400 default:
7401 break;
7404 /* Ensure we return a value of the proper mode. */
7405 return gen_lowpart (mode, x);
7408 /* Return nonzero if X is an expression that has one of two values depending on
7409 whether some other value is zero or nonzero. In that case, we return the
7410 value that is being tested, *PTRUE is set to the value if the rtx being
7411 returned has a nonzero value, and *PFALSE is set to the other alternative.
7413 If we return zero, we set *PTRUE and *PFALSE to X. */
7415 static rtx
7416 if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse)
7418 enum machine_mode mode = GET_MODE (x);
7419 enum rtx_code code = GET_CODE (x);
7420 rtx cond0, cond1, true0, true1, false0, false1;
7421 unsigned HOST_WIDE_INT nz;
7423 /* If we are comparing a value against zero, we are done. */
7424 if ((code == NE || code == EQ)
7425 && XEXP (x, 1) == const0_rtx)
7427 *ptrue = (code == NE) ? const_true_rtx : const0_rtx;
7428 *pfalse = (code == NE) ? const0_rtx : const_true_rtx;
7429 return XEXP (x, 0);
7432 /* If this is a unary operation whose operand has one of two values, apply
7433 our opcode to compute those values. */
7434 else if (UNARY_P (x)
7435 && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0)
7437 *ptrue = simplify_gen_unary (code, mode, true0, GET_MODE (XEXP (x, 0)));
7438 *pfalse = simplify_gen_unary (code, mode, false0,
7439 GET_MODE (XEXP (x, 0)));
7440 return cond0;
7443 /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would
7444 make can't possibly match and would suppress other optimizations. */
7445 else if (code == COMPARE)
7448 /* If this is a binary operation, see if either side has only one of two
7449 values. If either one does or if both do and they are conditional on
7450 the same value, compute the new true and false values. */
7451 else if (BINARY_P (x))
7453 cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0);
7454 cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1);
7456 if ((cond0 != 0 || cond1 != 0)
7457 && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1)))
7459 /* If if_then_else_cond returned zero, then true/false are the
7460 same rtl. We must copy one of them to prevent invalid rtl
7461 sharing. */
7462 if (cond0 == 0)
7463 true0 = copy_rtx (true0);
7464 else if (cond1 == 0)
7465 true1 = copy_rtx (true1);
7467 if (COMPARISON_P (x))
7469 *ptrue = simplify_gen_relational (code, mode, VOIDmode,
7470 true0, true1);
7471 *pfalse = simplify_gen_relational (code, mode, VOIDmode,
7472 false0, false1);
7474 else
7476 *ptrue = simplify_gen_binary (code, mode, true0, true1);
7477 *pfalse = simplify_gen_binary (code, mode, false0, false1);
7480 return cond0 ? cond0 : cond1;
7483 /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the
7484 operands is zero when the other is nonzero, and vice-versa,
7485 and STORE_FLAG_VALUE is 1 or -1. */
7487 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7488 && (code == PLUS || code == IOR || code == XOR || code == MINUS
7489 || code == UMAX)
7490 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7492 rtx op0 = XEXP (XEXP (x, 0), 1);
7493 rtx op1 = XEXP (XEXP (x, 1), 1);
7495 cond0 = XEXP (XEXP (x, 0), 0);
7496 cond1 = XEXP (XEXP (x, 1), 0);
7498 if (COMPARISON_P (cond0)
7499 && COMPARISON_P (cond1)
7500 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7501 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7502 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7503 || ((swap_condition (GET_CODE (cond0))
7504 == reversed_comparison_code (cond1, NULL))
7505 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7506 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7507 && ! side_effects_p (x))
7509 *ptrue = simplify_gen_binary (MULT, mode, op0, const_true_rtx);
7510 *pfalse = simplify_gen_binary (MULT, mode,
7511 (code == MINUS
7512 ? simplify_gen_unary (NEG, mode,
7513 op1, mode)
7514 : op1),
7515 const_true_rtx);
7516 return cond0;
7520 /* Similarly for MULT, AND and UMIN, except that for these the result
7521 is always zero. */
7522 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7523 && (code == MULT || code == AND || code == UMIN)
7524 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7526 cond0 = XEXP (XEXP (x, 0), 0);
7527 cond1 = XEXP (XEXP (x, 1), 0);
7529 if (COMPARISON_P (cond0)
7530 && COMPARISON_P (cond1)
7531 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7532 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7533 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7534 || ((swap_condition (GET_CODE (cond0))
7535 == reversed_comparison_code (cond1, NULL))
7536 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7537 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7538 && ! side_effects_p (x))
7540 *ptrue = *pfalse = const0_rtx;
7541 return cond0;
7546 else if (code == IF_THEN_ELSE)
7548 /* If we have IF_THEN_ELSE already, extract the condition and
7549 canonicalize it if it is NE or EQ. */
7550 cond0 = XEXP (x, 0);
7551 *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2);
7552 if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx)
7553 return XEXP (cond0, 0);
7554 else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx)
7556 *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1);
7557 return XEXP (cond0, 0);
7559 else
7560 return cond0;
7563 /* If X is a SUBREG, we can narrow both the true and false values
7564 if the inner expression, if there is a condition. */
7565 else if (code == SUBREG
7566 && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x),
7567 &true0, &false0)))
7569 true0 = simplify_gen_subreg (mode, true0,
7570 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7571 false0 = simplify_gen_subreg (mode, false0,
7572 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7573 if (true0 && false0)
7575 *ptrue = true0;
7576 *pfalse = false0;
7577 return cond0;
7581 /* If X is a constant, this isn't special and will cause confusions
7582 if we treat it as such. Likewise if it is equivalent to a constant. */
7583 else if (CONSTANT_P (x)
7584 || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0)))
7587 /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that
7588 will be least confusing to the rest of the compiler. */
7589 else if (mode == BImode)
7591 *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx;
7592 return x;
7595 /* If X is known to be either 0 or -1, those are the true and
7596 false values when testing X. */
7597 else if (x == constm1_rtx || x == const0_rtx
7598 || (mode != VOIDmode
7599 && num_sign_bit_copies (x, mode) == GET_MODE_BITSIZE (mode)))
7601 *ptrue = constm1_rtx, *pfalse = const0_rtx;
7602 return x;
7605 /* Likewise for 0 or a single bit. */
7606 else if (SCALAR_INT_MODE_P (mode)
7607 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
7608 && exact_log2 (nz = nonzero_bits (x, mode)) >= 0)
7610 *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx;
7611 return x;
7614 /* Otherwise fail; show no condition with true and false values the same. */
7615 *ptrue = *pfalse = x;
7616 return 0;
7619 /* Return the value of expression X given the fact that condition COND
7620 is known to be true when applied to REG as its first operand and VAL
7621 as its second. X is known to not be shared and so can be modified in
7622 place.
7624 We only handle the simplest cases, and specifically those cases that
7625 arise with IF_THEN_ELSE expressions. */
7627 static rtx
7628 known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val)
7630 enum rtx_code code = GET_CODE (x);
7631 rtx temp;
7632 const char *fmt;
7633 int i, j;
7635 if (side_effects_p (x))
7636 return x;
7638 /* If either operand of the condition is a floating point value,
7639 then we have to avoid collapsing an EQ comparison. */
7640 if (cond == EQ
7641 && rtx_equal_p (x, reg)
7642 && ! FLOAT_MODE_P (GET_MODE (x))
7643 && ! FLOAT_MODE_P (GET_MODE (val)))
7644 return val;
7646 if (cond == UNEQ && rtx_equal_p (x, reg))
7647 return val;
7649 /* If X is (abs REG) and we know something about REG's relationship
7650 with zero, we may be able to simplify this. */
7652 if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx)
7653 switch (cond)
7655 case GE: case GT: case EQ:
7656 return XEXP (x, 0);
7657 case LT: case LE:
7658 return simplify_gen_unary (NEG, GET_MODE (XEXP (x, 0)),
7659 XEXP (x, 0),
7660 GET_MODE (XEXP (x, 0)));
7661 default:
7662 break;
7665 /* The only other cases we handle are MIN, MAX, and comparisons if the
7666 operands are the same as REG and VAL. */
7668 else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x))
7670 if (rtx_equal_p (XEXP (x, 0), val))
7671 cond = swap_condition (cond), temp = val, val = reg, reg = temp;
7673 if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val))
7675 if (COMPARISON_P (x))
7677 if (comparison_dominates_p (cond, code))
7678 return const_true_rtx;
7680 code = reversed_comparison_code (x, NULL);
7681 if (code != UNKNOWN
7682 && comparison_dominates_p (cond, code))
7683 return const0_rtx;
7684 else
7685 return x;
7687 else if (code == SMAX || code == SMIN
7688 || code == UMIN || code == UMAX)
7690 int unsignedp = (code == UMIN || code == UMAX);
7692 /* Do not reverse the condition when it is NE or EQ.
7693 This is because we cannot conclude anything about
7694 the value of 'SMAX (x, y)' when x is not equal to y,
7695 but we can when x equals y. */
7696 if ((code == SMAX || code == UMAX)
7697 && ! (cond == EQ || cond == NE))
7698 cond = reverse_condition (cond);
7700 switch (cond)
7702 case GE: case GT:
7703 return unsignedp ? x : XEXP (x, 1);
7704 case LE: case LT:
7705 return unsignedp ? x : XEXP (x, 0);
7706 case GEU: case GTU:
7707 return unsignedp ? XEXP (x, 1) : x;
7708 case LEU: case LTU:
7709 return unsignedp ? XEXP (x, 0) : x;
7710 default:
7711 break;
7716 else if (code == SUBREG)
7718 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (x));
7719 rtx new, r = known_cond (SUBREG_REG (x), cond, reg, val);
7721 if (SUBREG_REG (x) != r)
7723 /* We must simplify subreg here, before we lose track of the
7724 original inner_mode. */
7725 new = simplify_subreg (GET_MODE (x), r,
7726 inner_mode, SUBREG_BYTE (x));
7727 if (new)
7728 return new;
7729 else
7730 SUBST (SUBREG_REG (x), r);
7733 return x;
7735 /* We don't have to handle SIGN_EXTEND here, because even in the
7736 case of replacing something with a modeless CONST_INT, a
7737 CONST_INT is already (supposed to be) a valid sign extension for
7738 its narrower mode, which implies it's already properly
7739 sign-extended for the wider mode. Now, for ZERO_EXTEND, the
7740 story is different. */
7741 else if (code == ZERO_EXTEND)
7743 enum machine_mode inner_mode = GET_MODE (XEXP (x, 0));
7744 rtx new, r = known_cond (XEXP (x, 0), cond, reg, val);
7746 if (XEXP (x, 0) != r)
7748 /* We must simplify the zero_extend here, before we lose
7749 track of the original inner_mode. */
7750 new = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
7751 r, inner_mode);
7752 if (new)
7753 return new;
7754 else
7755 SUBST (XEXP (x, 0), r);
7758 return x;
7761 fmt = GET_RTX_FORMAT (code);
7762 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7764 if (fmt[i] == 'e')
7765 SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val));
7766 else if (fmt[i] == 'E')
7767 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7768 SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j),
7769 cond, reg, val));
7772 return x;
7775 /* See if X and Y are equal for the purposes of seeing if we can rewrite an
7776 assignment as a field assignment. */
7778 static int
7779 rtx_equal_for_field_assignment_p (rtx x, rtx y)
7781 if (x == y || rtx_equal_p (x, y))
7782 return 1;
7784 if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y))
7785 return 0;
7787 /* Check for a paradoxical SUBREG of a MEM compared with the MEM.
7788 Note that all SUBREGs of MEM are paradoxical; otherwise they
7789 would have been rewritten. */
7790 if (MEM_P (x) && GET_CODE (y) == SUBREG
7791 && MEM_P (SUBREG_REG (y))
7792 && rtx_equal_p (SUBREG_REG (y),
7793 gen_lowpart (GET_MODE (SUBREG_REG (y)), x)))
7794 return 1;
7796 if (MEM_P (y) && GET_CODE (x) == SUBREG
7797 && MEM_P (SUBREG_REG (x))
7798 && rtx_equal_p (SUBREG_REG (x),
7799 gen_lowpart (GET_MODE (SUBREG_REG (x)), y)))
7800 return 1;
7802 /* We used to see if get_last_value of X and Y were the same but that's
7803 not correct. In one direction, we'll cause the assignment to have
7804 the wrong destination and in the case, we'll import a register into this
7805 insn that might have already have been dead. So fail if none of the
7806 above cases are true. */
7807 return 0;
7810 /* See if X, a SET operation, can be rewritten as a bit-field assignment.
7811 Return that assignment if so.
7813 We only handle the most common cases. */
7815 static rtx
7816 make_field_assignment (rtx x)
7818 rtx dest = SET_DEST (x);
7819 rtx src = SET_SRC (x);
7820 rtx assign;
7821 rtx rhs, lhs;
7822 HOST_WIDE_INT c1;
7823 HOST_WIDE_INT pos;
7824 unsigned HOST_WIDE_INT len;
7825 rtx other;
7826 enum machine_mode mode;
7828 /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is
7829 a clear of a one-bit field. We will have changed it to
7830 (and (rotate (const_int -2) POS) DEST), so check for that. Also check
7831 for a SUBREG. */
7833 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE
7834 && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT
7835 && INTVAL (XEXP (XEXP (src, 0), 0)) == -2
7836 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7838 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7839 1, 1, 1, 0);
7840 if (assign != 0)
7841 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7842 return x;
7845 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG
7846 && subreg_lowpart_p (XEXP (src, 0))
7847 && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0)))
7848 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0)))))
7849 && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE
7850 && GET_CODE (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == CONST_INT
7851 && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2
7852 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7854 assign = make_extraction (VOIDmode, dest, 0,
7855 XEXP (SUBREG_REG (XEXP (src, 0)), 1),
7856 1, 1, 1, 0);
7857 if (assign != 0)
7858 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7859 return x;
7862 /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a
7863 one-bit field. */
7864 if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT
7865 && XEXP (XEXP (src, 0), 0) == const1_rtx
7866 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7868 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7869 1, 1, 1, 0);
7870 if (assign != 0)
7871 return gen_rtx_SET (VOIDmode, assign, const1_rtx);
7872 return x;
7875 /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the
7876 SRC is an AND with all bits of that field set, then we can discard
7877 the AND. */
7878 if (GET_CODE (dest) == ZERO_EXTRACT
7879 && GET_CODE (XEXP (dest, 1)) == CONST_INT
7880 && GET_CODE (src) == AND
7881 && GET_CODE (XEXP (src, 1)) == CONST_INT)
7883 HOST_WIDE_INT width = INTVAL (XEXP (dest, 1));
7884 unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1));
7885 unsigned HOST_WIDE_INT ze_mask;
7887 if (width >= HOST_BITS_PER_WIDE_INT)
7888 ze_mask = -1;
7889 else
7890 ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1;
7892 /* Complete overlap. We can remove the source AND. */
7893 if ((and_mask & ze_mask) == ze_mask)
7894 return gen_rtx_SET (VOIDmode, dest, XEXP (src, 0));
7896 /* Partial overlap. We can reduce the source AND. */
7897 if ((and_mask & ze_mask) != and_mask)
7899 mode = GET_MODE (src);
7900 src = gen_rtx_AND (mode, XEXP (src, 0),
7901 gen_int_mode (and_mask & ze_mask, mode));
7902 return gen_rtx_SET (VOIDmode, dest, src);
7906 /* The other case we handle is assignments into a constant-position
7907 field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents
7908 a mask that has all one bits except for a group of zero bits and
7909 OTHER is known to have zeros where C1 has ones, this is such an
7910 assignment. Compute the position and length from C1. Shift OTHER
7911 to the appropriate position, force it to the required mode, and
7912 make the extraction. Check for the AND in both operands. */
7914 if (GET_CODE (src) != IOR && GET_CODE (src) != XOR)
7915 return x;
7917 rhs = expand_compound_operation (XEXP (src, 0));
7918 lhs = expand_compound_operation (XEXP (src, 1));
7920 if (GET_CODE (rhs) == AND
7921 && GET_CODE (XEXP (rhs, 1)) == CONST_INT
7922 && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest))
7923 c1 = INTVAL (XEXP (rhs, 1)), other = lhs;
7924 else if (GET_CODE (lhs) == AND
7925 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
7926 && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest))
7927 c1 = INTVAL (XEXP (lhs, 1)), other = rhs;
7928 else
7929 return x;
7931 pos = get_pos_from_mask ((~c1) & GET_MODE_MASK (GET_MODE (dest)), &len);
7932 if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest))
7933 || GET_MODE_BITSIZE (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT
7934 || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0)
7935 return x;
7937 assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0);
7938 if (assign == 0)
7939 return x;
7941 /* The mode to use for the source is the mode of the assignment, or of
7942 what is inside a possible STRICT_LOW_PART. */
7943 mode = (GET_CODE (assign) == STRICT_LOW_PART
7944 ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign));
7946 /* Shift OTHER right POS places and make it the source, restricting it
7947 to the proper length and mode. */
7949 src = force_to_mode (simplify_shift_const (NULL_RTX, LSHIFTRT,
7950 GET_MODE (src), other, pos),
7951 mode,
7952 GET_MODE_BITSIZE (mode) >= HOST_BITS_PER_WIDE_INT
7953 ? ~(unsigned HOST_WIDE_INT) 0
7954 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
7955 dest, 0);
7957 /* If SRC is masked by an AND that does not make a difference in
7958 the value being stored, strip it. */
7959 if (GET_CODE (assign) == ZERO_EXTRACT
7960 && GET_CODE (XEXP (assign, 1)) == CONST_INT
7961 && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT
7962 && GET_CODE (src) == AND
7963 && GET_CODE (XEXP (src, 1)) == CONST_INT
7964 && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (src, 1))
7965 == ((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (assign, 1))) - 1))
7966 src = XEXP (src, 0);
7968 return gen_rtx_SET (VOIDmode, assign, src);
7971 /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c)
7972 if so. */
7974 static rtx
7975 apply_distributive_law (rtx x)
7977 enum rtx_code code = GET_CODE (x);
7978 enum rtx_code inner_code;
7979 rtx lhs, rhs, other;
7980 rtx tem;
7982 /* Distributivity is not true for floating point as it can change the
7983 value. So we don't do it unless -funsafe-math-optimizations. */
7984 if (FLOAT_MODE_P (GET_MODE (x))
7985 && ! flag_unsafe_math_optimizations)
7986 return x;
7988 /* The outer operation can only be one of the following: */
7989 if (code != IOR && code != AND && code != XOR
7990 && code != PLUS && code != MINUS)
7991 return x;
7993 lhs = XEXP (x, 0);
7994 rhs = XEXP (x, 1);
7996 /* If either operand is a primitive we can't do anything, so get out
7997 fast. */
7998 if (OBJECT_P (lhs) || OBJECT_P (rhs))
7999 return x;
8001 lhs = expand_compound_operation (lhs);
8002 rhs = expand_compound_operation (rhs);
8003 inner_code = GET_CODE (lhs);
8004 if (inner_code != GET_CODE (rhs))
8005 return x;
8007 /* See if the inner and outer operations distribute. */
8008 switch (inner_code)
8010 case LSHIFTRT:
8011 case ASHIFTRT:
8012 case AND:
8013 case IOR:
8014 /* These all distribute except over PLUS. */
8015 if (code == PLUS || code == MINUS)
8016 return x;
8017 break;
8019 case MULT:
8020 if (code != PLUS && code != MINUS)
8021 return x;
8022 break;
8024 case ASHIFT:
8025 /* This is also a multiply, so it distributes over everything. */
8026 break;
8028 case SUBREG:
8029 /* Non-paradoxical SUBREGs distributes over all operations, provided
8030 the inner modes and byte offsets are the same, this is an extraction
8031 of a low-order part, we don't convert an fp operation to int or
8032 vice versa, and we would not be converting a single-word
8033 operation into a multi-word operation. The latter test is not
8034 required, but it prevents generating unneeded multi-word operations.
8035 Some of the previous tests are redundant given the latter test, but
8036 are retained because they are required for correctness.
8038 We produce the result slightly differently in this case. */
8040 if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs))
8041 || SUBREG_BYTE (lhs) != SUBREG_BYTE (rhs)
8042 || ! subreg_lowpart_p (lhs)
8043 || (GET_MODE_CLASS (GET_MODE (lhs))
8044 != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs))))
8045 || (GET_MODE_SIZE (GET_MODE (lhs))
8046 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))))
8047 || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD)
8048 return x;
8050 tem = simplify_gen_binary (code, GET_MODE (SUBREG_REG (lhs)),
8051 SUBREG_REG (lhs), SUBREG_REG (rhs));
8052 return gen_lowpart (GET_MODE (x), tem);
8054 default:
8055 return x;
8058 /* Set LHS and RHS to the inner operands (A and B in the example
8059 above) and set OTHER to the common operand (C in the example).
8060 There is only one way to do this unless the inner operation is
8061 commutative. */
8062 if (COMMUTATIVE_ARITH_P (lhs)
8063 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0)))
8064 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1);
8065 else if (COMMUTATIVE_ARITH_P (lhs)
8066 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1)))
8067 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0);
8068 else if (COMMUTATIVE_ARITH_P (lhs)
8069 && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0)))
8070 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1);
8071 else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1)))
8072 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0);
8073 else
8074 return x;
8076 /* Form the new inner operation, seeing if it simplifies first. */
8077 tem = simplify_gen_binary (code, GET_MODE (x), lhs, rhs);
8079 /* There is one exception to the general way of distributing:
8080 (a | c) ^ (b | c) -> (a ^ b) & ~c */
8081 if (code == XOR && inner_code == IOR)
8083 inner_code = AND;
8084 other = simplify_gen_unary (NOT, GET_MODE (x), other, GET_MODE (x));
8087 /* We may be able to continuing distributing the result, so call
8088 ourselves recursively on the inner operation before forming the
8089 outer operation, which we return. */
8090 return simplify_gen_binary (inner_code, GET_MODE (x),
8091 apply_distributive_law (tem), other);
8094 /* See if X is of the form (* (+ A B) C), and if so convert to
8095 (+ (* A C) (* B C)) and try to simplify.
8097 Most of the time, this results in no change. However, if some of
8098 the operands are the same or inverses of each other, simplifications
8099 will result.
8101 For example, (and (ior A B) (not B)) can occur as the result of
8102 expanding a bit field assignment. When we apply the distributive
8103 law to this, we get (ior (and (A (not B))) (and (B (not B)))),
8104 which then simplifies to (and (A (not B))).
8106 Note that no checks happen on the validity of applying the inverse
8107 distributive law. This is pointless since we can do it in the
8108 few places where this routine is called.
8110 N is the index of the term that is decomposed (the arithmetic operation,
8111 i.e. (+ A B) in the first example above). !N is the index of the term that
8112 is distributed, i.e. of C in the first example above. */
8113 static rtx
8114 distribute_and_simplify_rtx (rtx x, int n)
8116 enum machine_mode mode;
8117 enum rtx_code outer_code, inner_code;
8118 rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp;
8120 decomposed = XEXP (x, n);
8121 if (!ARITHMETIC_P (decomposed))
8122 return NULL_RTX;
8124 mode = GET_MODE (x);
8125 outer_code = GET_CODE (x);
8126 distributed = XEXP (x, !n);
8128 inner_code = GET_CODE (decomposed);
8129 inner_op0 = XEXP (decomposed, 0);
8130 inner_op1 = XEXP (decomposed, 1);
8132 /* Special case (and (xor B C) (not A)), which is equivalent to
8133 (xor (ior A B) (ior A C)) */
8134 if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT)
8136 distributed = XEXP (distributed, 0);
8137 outer_code = IOR;
8140 if (n == 0)
8142 /* Distribute the second term. */
8143 new_op0 = simplify_gen_binary (outer_code, mode, inner_op0, distributed);
8144 new_op1 = simplify_gen_binary (outer_code, mode, inner_op1, distributed);
8146 else
8148 /* Distribute the first term. */
8149 new_op0 = simplify_gen_binary (outer_code, mode, distributed, inner_op0);
8150 new_op1 = simplify_gen_binary (outer_code, mode, distributed, inner_op1);
8153 tmp = apply_distributive_law (simplify_gen_binary (inner_code, mode,
8154 new_op0, new_op1));
8155 if (GET_CODE (tmp) != outer_code
8156 && rtx_cost (tmp, SET) < rtx_cost (x, SET))
8157 return tmp;
8159 return NULL_RTX;
8162 /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done
8163 in MODE.
8165 Return an equivalent form, if different from X. Otherwise, return X. If
8166 X is zero, we are to always construct the equivalent form. */
8168 static rtx
8169 simplify_and_const_int (rtx x, enum machine_mode mode, rtx varop,
8170 unsigned HOST_WIDE_INT constop)
8172 unsigned HOST_WIDE_INT nonzero;
8173 int i;
8175 /* Simplify VAROP knowing that we will be only looking at some of the
8176 bits in it.
8178 Note by passing in CONSTOP, we guarantee that the bits not set in
8179 CONSTOP are not significant and will never be examined. We must
8180 ensure that is the case by explicitly masking out those bits
8181 before returning. */
8182 varop = force_to_mode (varop, mode, constop, NULL_RTX, 0);
8184 /* If VAROP is a CLOBBER, we will fail so return it. */
8185 if (GET_CODE (varop) == CLOBBER)
8186 return varop;
8188 /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP
8189 to VAROP and return the new constant. */
8190 if (GET_CODE (varop) == CONST_INT)
8191 return gen_int_mode (INTVAL (varop) & constop, mode);
8193 /* See what bits may be nonzero in VAROP. Unlike the general case of
8194 a call to nonzero_bits, here we don't care about bits outside
8195 MODE. */
8197 nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
8199 /* Turn off all bits in the constant that are known to already be zero.
8200 Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
8201 which is tested below. */
8203 constop &= nonzero;
8205 /* If we don't have any bits left, return zero. */
8206 if (constop == 0)
8207 return const0_rtx;
8209 /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is
8210 a power of two, we can replace this with an ASHIFT. */
8211 if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1
8212 && (i = exact_log2 (constop)) >= 0)
8213 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i);
8215 /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR
8216 or XOR, then try to apply the distributive law. This may eliminate
8217 operations if either branch can be simplified because of the AND.
8218 It may also make some cases more complex, but those cases probably
8219 won't match a pattern either with or without this. */
8221 if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR)
8222 return
8223 gen_lowpart
8224 (mode,
8225 apply_distributive_law
8226 (simplify_gen_binary (GET_CODE (varop), GET_MODE (varop),
8227 simplify_and_const_int (NULL_RTX,
8228 GET_MODE (varop),
8229 XEXP (varop, 0),
8230 constop),
8231 simplify_and_const_int (NULL_RTX,
8232 GET_MODE (varop),
8233 XEXP (varop, 1),
8234 constop))));
8236 /* If VAROP is PLUS, and the constant is a mask of low bite, distribute
8237 the AND and see if one of the operands simplifies to zero. If so, we
8238 may eliminate it. */
8240 if (GET_CODE (varop) == PLUS
8241 && exact_log2 (constop + 1) >= 0)
8243 rtx o0, o1;
8245 o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop);
8246 o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop);
8247 if (o0 == const0_rtx)
8248 return o1;
8249 if (o1 == const0_rtx)
8250 return o0;
8253 /* Get VAROP in MODE. Try to get a SUBREG if not. Don't make a new SUBREG
8254 if we already had one (just check for the simplest cases). */
8255 if (x && GET_CODE (XEXP (x, 0)) == SUBREG
8256 && GET_MODE (XEXP (x, 0)) == mode
8257 && SUBREG_REG (XEXP (x, 0)) == varop)
8258 varop = XEXP (x, 0);
8259 else
8260 varop = gen_lowpart (mode, varop);
8262 /* If we can't make the SUBREG, try to return what we were given. */
8263 if (GET_CODE (varop) == CLOBBER)
8264 return x ? x : varop;
8266 /* If we are only masking insignificant bits, return VAROP. */
8267 if (constop == nonzero)
8268 x = varop;
8269 else
8271 /* Otherwise, return an AND. */
8272 constop = trunc_int_for_mode (constop, mode);
8273 /* See how much, if any, of X we can use. */
8274 if (x == 0 || GET_CODE (x) != AND || GET_MODE (x) != mode)
8275 x = simplify_gen_binary (AND, mode, varop, GEN_INT (constop));
8277 else
8279 if (GET_CODE (XEXP (x, 1)) != CONST_INT
8280 || (unsigned HOST_WIDE_INT) INTVAL (XEXP (x, 1)) != constop)
8281 SUBST (XEXP (x, 1), GEN_INT (constop));
8283 SUBST (XEXP (x, 0), varop);
8287 return x;
8290 /* Given a REG, X, compute which bits in X can be nonzero.
8291 We don't care about bits outside of those defined in MODE.
8293 For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is
8294 a shift, AND, or zero_extract, we can do better. */
8296 static rtx
8297 reg_nonzero_bits_for_combine (rtx x, enum machine_mode mode,
8298 rtx known_x ATTRIBUTE_UNUSED,
8299 enum machine_mode known_mode ATTRIBUTE_UNUSED,
8300 unsigned HOST_WIDE_INT known_ret ATTRIBUTE_UNUSED,
8301 unsigned HOST_WIDE_INT *nonzero)
8303 rtx tem;
8305 /* If X is a register whose nonzero bits value is current, use it.
8306 Otherwise, if X is a register whose value we can find, use that
8307 value. Otherwise, use the previously-computed global nonzero bits
8308 for this register. */
8310 if (reg_stat[REGNO (x)].last_set_value != 0
8311 && (reg_stat[REGNO (x)].last_set_mode == mode
8312 || (GET_MODE_CLASS (reg_stat[REGNO (x)].last_set_mode) == MODE_INT
8313 && GET_MODE_CLASS (mode) == MODE_INT))
8314 && (reg_stat[REGNO (x)].last_set_label == label_tick
8315 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8316 && REG_N_SETS (REGNO (x)) == 1
8317 && ! REGNO_REG_SET_P (ENTRY_BLOCK_PTR->next_bb->global_live_at_start,
8318 REGNO (x))))
8319 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8321 *nonzero &= reg_stat[REGNO (x)].last_set_nonzero_bits;
8322 return NULL;
8325 tem = get_last_value (x);
8327 if (tem)
8329 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
8330 /* If X is narrower than MODE and TEM is a non-negative
8331 constant that would appear negative in the mode of X,
8332 sign-extend it for use in reg_nonzero_bits because some
8333 machines (maybe most) will actually do the sign-extension
8334 and this is the conservative approach.
8336 ??? For 2.5, try to tighten up the MD files in this regard
8337 instead of this kludge. */
8339 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode)
8340 && GET_CODE (tem) == CONST_INT
8341 && INTVAL (tem) > 0
8342 && 0 != (INTVAL (tem)
8343 & ((HOST_WIDE_INT) 1
8344 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
8345 tem = GEN_INT (INTVAL (tem)
8346 | ((HOST_WIDE_INT) (-1)
8347 << GET_MODE_BITSIZE (GET_MODE (x))));
8348 #endif
8349 return tem;
8351 else if (nonzero_sign_valid && reg_stat[REGNO (x)].nonzero_bits)
8353 unsigned HOST_WIDE_INT mask = reg_stat[REGNO (x)].nonzero_bits;
8355 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode))
8356 /* We don't know anything about the upper bits. */
8357 mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (GET_MODE (x));
8358 *nonzero &= mask;
8361 return NULL;
8364 /* Return the number of bits at the high-order end of X that are known to
8365 be equal to the sign bit. X will be used in mode MODE; if MODE is
8366 VOIDmode, X will be used in its own mode. The returned value will always
8367 be between 1 and the number of bits in MODE. */
8369 static rtx
8370 reg_num_sign_bit_copies_for_combine (rtx x, enum machine_mode mode,
8371 rtx known_x ATTRIBUTE_UNUSED,
8372 enum machine_mode known_mode
8373 ATTRIBUTE_UNUSED,
8374 unsigned int known_ret ATTRIBUTE_UNUSED,
8375 unsigned int *result)
8377 rtx tem;
8379 if (reg_stat[REGNO (x)].last_set_value != 0
8380 && reg_stat[REGNO (x)].last_set_mode == mode
8381 && (reg_stat[REGNO (x)].last_set_label == label_tick
8382 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8383 && REG_N_SETS (REGNO (x)) == 1
8384 && ! REGNO_REG_SET_P (ENTRY_BLOCK_PTR->next_bb->global_live_at_start,
8385 REGNO (x))))
8386 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8388 *result = reg_stat[REGNO (x)].last_set_sign_bit_copies;
8389 return NULL;
8392 tem = get_last_value (x);
8393 if (tem != 0)
8394 return tem;
8396 if (nonzero_sign_valid && reg_stat[REGNO (x)].sign_bit_copies != 0
8397 && GET_MODE_BITSIZE (GET_MODE (x)) == GET_MODE_BITSIZE (mode))
8398 *result = reg_stat[REGNO (x)].sign_bit_copies;
8400 return NULL;
8403 /* Return the number of "extended" bits there are in X, when interpreted
8404 as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For
8405 unsigned quantities, this is the number of high-order zero bits.
8406 For signed quantities, this is the number of copies of the sign bit
8407 minus 1. In both case, this function returns the number of "spare"
8408 bits. For example, if two quantities for which this function returns
8409 at least 1 are added, the addition is known not to overflow.
8411 This function will always return 0 unless called during combine, which
8412 implies that it must be called from a define_split. */
8414 unsigned int
8415 extended_count (rtx x, enum machine_mode mode, int unsignedp)
8417 if (nonzero_sign_valid == 0)
8418 return 0;
8420 return (unsignedp
8421 ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8422 ? (unsigned int) (GET_MODE_BITSIZE (mode) - 1
8423 - floor_log2 (nonzero_bits (x, mode)))
8424 : 0)
8425 : num_sign_bit_copies (x, mode) - 1);
8428 /* This function is called from `simplify_shift_const' to merge two
8429 outer operations. Specifically, we have already found that we need
8430 to perform operation *POP0 with constant *PCONST0 at the outermost
8431 position. We would now like to also perform OP1 with constant CONST1
8432 (with *POP0 being done last).
8434 Return 1 if we can do the operation and update *POP0 and *PCONST0 with
8435 the resulting operation. *PCOMP_P is set to 1 if we would need to
8436 complement the innermost operand, otherwise it is unchanged.
8438 MODE is the mode in which the operation will be done. No bits outside
8439 the width of this mode matter. It is assumed that the width of this mode
8440 is smaller than or equal to HOST_BITS_PER_WIDE_INT.
8442 If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS,
8443 IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper
8444 result is simply *PCONST0.
8446 If the resulting operation cannot be expressed as one operation, we
8447 return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */
8449 static int
8450 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)
8452 enum rtx_code op0 = *pop0;
8453 HOST_WIDE_INT const0 = *pconst0;
8455 const0 &= GET_MODE_MASK (mode);
8456 const1 &= GET_MODE_MASK (mode);
8458 /* If OP0 is an AND, clear unimportant bits in CONST1. */
8459 if (op0 == AND)
8460 const1 &= const0;
8462 /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or
8463 if OP0 is SET. */
8465 if (op1 == UNKNOWN || op0 == SET)
8466 return 1;
8468 else if (op0 == UNKNOWN)
8469 op0 = op1, const0 = const1;
8471 else if (op0 == op1)
8473 switch (op0)
8475 case AND:
8476 const0 &= const1;
8477 break;
8478 case IOR:
8479 const0 |= const1;
8480 break;
8481 case XOR:
8482 const0 ^= const1;
8483 break;
8484 case PLUS:
8485 const0 += const1;
8486 break;
8487 case NEG:
8488 op0 = UNKNOWN;
8489 break;
8490 default:
8491 break;
8495 /* Otherwise, if either is a PLUS or NEG, we can't do anything. */
8496 else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG)
8497 return 0;
8499 /* If the two constants aren't the same, we can't do anything. The
8500 remaining six cases can all be done. */
8501 else if (const0 != const1)
8502 return 0;
8504 else
8505 switch (op0)
8507 case IOR:
8508 if (op1 == AND)
8509 /* (a & b) | b == b */
8510 op0 = SET;
8511 else /* op1 == XOR */
8512 /* (a ^ b) | b == a | b */
8514 break;
8516 case XOR:
8517 if (op1 == AND)
8518 /* (a & b) ^ b == (~a) & b */
8519 op0 = AND, *pcomp_p = 1;
8520 else /* op1 == IOR */
8521 /* (a | b) ^ b == a & ~b */
8522 op0 = AND, const0 = ~const0;
8523 break;
8525 case AND:
8526 if (op1 == IOR)
8527 /* (a | b) & b == b */
8528 op0 = SET;
8529 else /* op1 == XOR */
8530 /* (a ^ b) & b) == (~a) & b */
8531 *pcomp_p = 1;
8532 break;
8533 default:
8534 break;
8537 /* Check for NO-OP cases. */
8538 const0 &= GET_MODE_MASK (mode);
8539 if (const0 == 0
8540 && (op0 == IOR || op0 == XOR || op0 == PLUS))
8541 op0 = UNKNOWN;
8542 else if (const0 == 0 && op0 == AND)
8543 op0 = SET;
8544 else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode)
8545 && op0 == AND)
8546 op0 = UNKNOWN;
8548 /* ??? Slightly redundant with the above mask, but not entirely.
8549 Moving this above means we'd have to sign-extend the mode mask
8550 for the final test. */
8551 const0 = trunc_int_for_mode (const0, mode);
8553 *pop0 = op0;
8554 *pconst0 = const0;
8556 return 1;
8559 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
8560 The result of the shift is RESULT_MODE. X, if nonzero, is an expression
8561 that we started with.
8563 The shift is normally computed in the widest mode we find in VAROP, as
8564 long as it isn't a different number of words than RESULT_MODE. Exceptions
8565 are ASHIFTRT and ROTATE, which are always done in their original mode, */
8567 static rtx
8568 simplify_shift_const (rtx x, enum rtx_code code,
8569 enum machine_mode result_mode, rtx varop,
8570 int orig_count)
8572 enum rtx_code orig_code = code;
8573 unsigned int count;
8574 int signed_count;
8575 enum machine_mode mode = result_mode;
8576 enum machine_mode shift_mode, tmode;
8577 unsigned int mode_words
8578 = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
8579 /* We form (outer_op (code varop count) (outer_const)). */
8580 enum rtx_code outer_op = UNKNOWN;
8581 HOST_WIDE_INT outer_const = 0;
8582 rtx const_rtx;
8583 int complement_p = 0;
8584 rtx new;
8586 /* Make sure and truncate the "natural" shift on the way in. We don't
8587 want to do this inside the loop as it makes it more difficult to
8588 combine shifts. */
8589 if (SHIFT_COUNT_TRUNCATED)
8590 orig_count &= GET_MODE_BITSIZE (mode) - 1;
8592 /* If we were given an invalid count, don't do anything except exactly
8593 what was requested. */
8595 if (orig_count < 0 || orig_count >= (int) GET_MODE_BITSIZE (mode))
8597 if (x)
8598 return x;
8600 return gen_rtx_fmt_ee (code, mode, varop, GEN_INT (orig_count));
8603 count = orig_count;
8605 /* Unless one of the branches of the `if' in this loop does a `continue',
8606 we will `break' the loop after the `if'. */
8608 while (count != 0)
8610 /* If we have an operand of (clobber (const_int 0)), just return that
8611 value. */
8612 if (GET_CODE (varop) == CLOBBER)
8613 return varop;
8615 /* If we discovered we had to complement VAROP, leave. Making a NOT
8616 here would cause an infinite loop. */
8617 if (complement_p)
8618 break;
8620 /* Convert ROTATERT to ROTATE. */
8621 if (code == ROTATERT)
8623 unsigned int bitsize = GET_MODE_BITSIZE (result_mode);;
8624 code = ROTATE;
8625 if (VECTOR_MODE_P (result_mode))
8626 count = bitsize / GET_MODE_NUNITS (result_mode) - count;
8627 else
8628 count = bitsize - count;
8631 /* We need to determine what mode we will do the shift in. If the
8632 shift is a right shift or a ROTATE, we must always do it in the mode
8633 it was originally done in. Otherwise, we can do it in MODE, the
8634 widest mode encountered. */
8635 shift_mode
8636 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
8637 ? result_mode : mode);
8639 /* Handle cases where the count is greater than the size of the mode
8640 minus 1. For ASHIFT, use the size minus one as the count (this can
8641 occur when simplifying (lshiftrt (ashiftrt ..))). For rotates,
8642 take the count modulo the size. For other shifts, the result is
8643 zero.
8645 Since these shifts are being produced by the compiler by combining
8646 multiple operations, each of which are defined, we know what the
8647 result is supposed to be. */
8649 if (count > (unsigned int) (GET_MODE_BITSIZE (shift_mode) - 1))
8651 if (code == ASHIFTRT)
8652 count = GET_MODE_BITSIZE (shift_mode) - 1;
8653 else if (code == ROTATE || code == ROTATERT)
8654 count %= GET_MODE_BITSIZE (shift_mode);
8655 else
8657 /* We can't simply return zero because there may be an
8658 outer op. */
8659 varop = const0_rtx;
8660 count = 0;
8661 break;
8665 /* An arithmetic right shift of a quantity known to be -1 or 0
8666 is a no-op. */
8667 if (code == ASHIFTRT
8668 && (num_sign_bit_copies (varop, shift_mode)
8669 == GET_MODE_BITSIZE (shift_mode)))
8671 count = 0;
8672 break;
8675 /* If we are doing an arithmetic right shift and discarding all but
8676 the sign bit copies, this is equivalent to doing a shift by the
8677 bitsize minus one. Convert it into that shift because it will often
8678 allow other simplifications. */
8680 if (code == ASHIFTRT
8681 && (count + num_sign_bit_copies (varop, shift_mode)
8682 >= GET_MODE_BITSIZE (shift_mode)))
8683 count = GET_MODE_BITSIZE (shift_mode) - 1;
8685 /* We simplify the tests below and elsewhere by converting
8686 ASHIFTRT to LSHIFTRT if we know the sign bit is clear.
8687 `make_compound_operation' will convert it to an ASHIFTRT for
8688 those machines (such as VAX) that don't have an LSHIFTRT. */
8689 if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8690 && code == ASHIFTRT
8691 && ((nonzero_bits (varop, shift_mode)
8692 & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1)))
8693 == 0))
8694 code = LSHIFTRT;
8696 if (code == LSHIFTRT
8697 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8698 && !(nonzero_bits (varop, shift_mode) >> count))
8699 varop = const0_rtx;
8700 if (code == ASHIFT
8701 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8702 && !((nonzero_bits (varop, shift_mode) << count)
8703 & GET_MODE_MASK (shift_mode)))
8704 varop = const0_rtx;
8706 switch (GET_CODE (varop))
8708 case SIGN_EXTEND:
8709 case ZERO_EXTEND:
8710 case SIGN_EXTRACT:
8711 case ZERO_EXTRACT:
8712 new = expand_compound_operation (varop);
8713 if (new != varop)
8715 varop = new;
8716 continue;
8718 break;
8720 case MEM:
8721 /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH
8722 minus the width of a smaller mode, we can do this with a
8723 SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */
8724 if ((code == ASHIFTRT || code == LSHIFTRT)
8725 && ! mode_dependent_address_p (XEXP (varop, 0))
8726 && ! MEM_VOLATILE_P (varop)
8727 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
8728 MODE_INT, 1)) != BLKmode)
8730 new = adjust_address_nv (varop, tmode,
8731 BYTES_BIG_ENDIAN ? 0
8732 : count / BITS_PER_UNIT);
8734 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
8735 : ZERO_EXTEND, mode, new);
8736 count = 0;
8737 continue;
8739 break;
8741 case USE:
8742 /* Similar to the case above, except that we can only do this if
8743 the resulting mode is the same as that of the underlying
8744 MEM and adjust the address depending on the *bits* endianness
8745 because of the way that bit-field extract insns are defined. */
8746 if ((code == ASHIFTRT || code == LSHIFTRT)
8747 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
8748 MODE_INT, 1)) != BLKmode
8749 && tmode == GET_MODE (XEXP (varop, 0)))
8751 if (BITS_BIG_ENDIAN)
8752 new = XEXP (varop, 0);
8753 else
8755 new = copy_rtx (XEXP (varop, 0));
8756 SUBST (XEXP (new, 0),
8757 plus_constant (XEXP (new, 0),
8758 count / BITS_PER_UNIT));
8761 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
8762 : ZERO_EXTEND, mode, new);
8763 count = 0;
8764 continue;
8766 break;
8768 case SUBREG:
8769 /* If VAROP is a SUBREG, strip it as long as the inner operand has
8770 the same number of words as what we've seen so far. Then store
8771 the widest mode in MODE. */
8772 if (subreg_lowpart_p (varop)
8773 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8774 > GET_MODE_SIZE (GET_MODE (varop)))
8775 && (unsigned int) ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8776 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
8777 == mode_words)
8779 varop = SUBREG_REG (varop);
8780 if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode))
8781 mode = GET_MODE (varop);
8782 continue;
8784 break;
8786 case MULT:
8787 /* Some machines use MULT instead of ASHIFT because MULT
8788 is cheaper. But it is still better on those machines to
8789 merge two shifts into one. */
8790 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8791 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8793 varop
8794 = simplify_gen_binary (ASHIFT, GET_MODE (varop),
8795 XEXP (varop, 0),
8796 GEN_INT (exact_log2 (
8797 INTVAL (XEXP (varop, 1)))));
8798 continue;
8800 break;
8802 case UDIV:
8803 /* Similar, for when divides are cheaper. */
8804 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8805 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8807 varop
8808 = simplify_gen_binary (LSHIFTRT, GET_MODE (varop),
8809 XEXP (varop, 0),
8810 GEN_INT (exact_log2 (
8811 INTVAL (XEXP (varop, 1)))));
8812 continue;
8814 break;
8816 case ASHIFTRT:
8817 /* If we are extracting just the sign bit of an arithmetic
8818 right shift, that shift is not needed. However, the sign
8819 bit of a wider mode may be different from what would be
8820 interpreted as the sign bit in a narrower mode, so, if
8821 the result is narrower, don't discard the shift. */
8822 if (code == LSHIFTRT
8823 && count == (unsigned int) (GET_MODE_BITSIZE (result_mode) - 1)
8824 && (GET_MODE_BITSIZE (result_mode)
8825 >= GET_MODE_BITSIZE (GET_MODE (varop))))
8827 varop = XEXP (varop, 0);
8828 continue;
8831 /* ... fall through ... */
8833 case LSHIFTRT:
8834 case ASHIFT:
8835 case ROTATE:
8836 /* Here we have two nested shifts. The result is usually the
8837 AND of a new shift with a mask. We compute the result below. */
8838 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8839 && INTVAL (XEXP (varop, 1)) >= 0
8840 && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop))
8841 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8842 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
8844 enum rtx_code first_code = GET_CODE (varop);
8845 unsigned int first_count = INTVAL (XEXP (varop, 1));
8846 unsigned HOST_WIDE_INT mask;
8847 rtx mask_rtx;
8849 /* We have one common special case. We can't do any merging if
8850 the inner code is an ASHIFTRT of a smaller mode. However, if
8851 we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2)
8852 with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2),
8853 we can convert it to
8854 (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1).
8855 This simplifies certain SIGN_EXTEND operations. */
8856 if (code == ASHIFT && first_code == ASHIFTRT
8857 && count == (unsigned int)
8858 (GET_MODE_BITSIZE (result_mode)
8859 - GET_MODE_BITSIZE (GET_MODE (varop))))
8861 /* C3 has the low-order C1 bits zero. */
8863 mask = (GET_MODE_MASK (mode)
8864 & ~(((HOST_WIDE_INT) 1 << first_count) - 1));
8866 varop = simplify_and_const_int (NULL_RTX, result_mode,
8867 XEXP (varop, 0), mask);
8868 varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode,
8869 varop, count);
8870 count = first_count;
8871 code = ASHIFTRT;
8872 continue;
8875 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more
8876 than C1 high-order bits equal to the sign bit, we can convert
8877 this to either an ASHIFT or an ASHIFTRT depending on the
8878 two counts.
8880 We cannot do this if VAROP's mode is not SHIFT_MODE. */
8882 if (code == ASHIFTRT && first_code == ASHIFT
8883 && GET_MODE (varop) == shift_mode
8884 && (num_sign_bit_copies (XEXP (varop, 0), shift_mode)
8885 > first_count))
8887 varop = XEXP (varop, 0);
8889 signed_count = count - first_count;
8890 if (signed_count < 0)
8891 count = -signed_count, code = ASHIFT;
8892 else
8893 count = signed_count;
8895 continue;
8898 /* There are some cases we can't do. If CODE is ASHIFTRT,
8899 we can only do this if FIRST_CODE is also ASHIFTRT.
8901 We can't do the case when CODE is ROTATE and FIRST_CODE is
8902 ASHIFTRT.
8904 If the mode of this shift is not the mode of the outer shift,
8905 we can't do this if either shift is a right shift or ROTATE.
8907 Finally, we can't do any of these if the mode is too wide
8908 unless the codes are the same.
8910 Handle the case where the shift codes are the same
8911 first. */
8913 if (code == first_code)
8915 if (GET_MODE (varop) != result_mode
8916 && (code == ASHIFTRT || code == LSHIFTRT
8917 || code == ROTATE))
8918 break;
8920 count += first_count;
8921 varop = XEXP (varop, 0);
8922 continue;
8925 if (code == ASHIFTRT
8926 || (code == ROTATE && first_code == ASHIFTRT)
8927 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT
8928 || (GET_MODE (varop) != result_mode
8929 && (first_code == ASHIFTRT || first_code == LSHIFTRT
8930 || first_code == ROTATE
8931 || code == ROTATE)))
8932 break;
8934 /* To compute the mask to apply after the shift, shift the
8935 nonzero bits of the inner shift the same way the
8936 outer shift will. */
8938 mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop)));
8940 mask_rtx
8941 = simplify_binary_operation (code, result_mode, mask_rtx,
8942 GEN_INT (count));
8944 /* Give up if we can't compute an outer operation to use. */
8945 if (mask_rtx == 0
8946 || GET_CODE (mask_rtx) != CONST_INT
8947 || ! merge_outer_ops (&outer_op, &outer_const, AND,
8948 INTVAL (mask_rtx),
8949 result_mode, &complement_p))
8950 break;
8952 /* If the shifts are in the same direction, we add the
8953 counts. Otherwise, we subtract them. */
8954 signed_count = count;
8955 if ((code == ASHIFTRT || code == LSHIFTRT)
8956 == (first_code == ASHIFTRT || first_code == LSHIFTRT))
8957 signed_count += first_count;
8958 else
8959 signed_count -= first_count;
8961 /* If COUNT is positive, the new shift is usually CODE,
8962 except for the two exceptions below, in which case it is
8963 FIRST_CODE. If the count is negative, FIRST_CODE should
8964 always be used */
8965 if (signed_count > 0
8966 && ((first_code == ROTATE && code == ASHIFT)
8967 || (first_code == ASHIFTRT && code == LSHIFTRT)))
8968 code = first_code, count = signed_count;
8969 else if (signed_count < 0)
8970 code = first_code, count = -signed_count;
8971 else
8972 count = signed_count;
8974 varop = XEXP (varop, 0);
8975 continue;
8978 /* If we have (A << B << C) for any shift, we can convert this to
8979 (A << C << B). This wins if A is a constant. Only try this if
8980 B is not a constant. */
8982 else if (GET_CODE (varop) == code
8983 && GET_CODE (XEXP (varop, 1)) != CONST_INT
8984 && 0 != (new
8985 = simplify_binary_operation (code, mode,
8986 XEXP (varop, 0),
8987 GEN_INT (count))))
8989 varop = gen_rtx_fmt_ee (code, mode, new, XEXP (varop, 1));
8990 count = 0;
8991 continue;
8993 break;
8995 case NOT:
8996 /* Make this fit the case below. */
8997 varop = gen_rtx_XOR (mode, XEXP (varop, 0),
8998 GEN_INT (GET_MODE_MASK (mode)));
8999 continue;
9001 case IOR:
9002 case AND:
9003 case XOR:
9004 /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C)
9005 with C the size of VAROP - 1 and the shift is logical if
9006 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
9007 we have an (le X 0) operation. If we have an arithmetic shift
9008 and STORE_FLAG_VALUE is 1 or we have a logical shift with
9009 STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */
9011 if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS
9012 && XEXP (XEXP (varop, 0), 1) == constm1_rtx
9013 && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
9014 && (code == LSHIFTRT || code == ASHIFTRT)
9015 && count == (unsigned int)
9016 (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
9017 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
9019 count = 0;
9020 varop = gen_rtx_LE (GET_MODE (varop), XEXP (varop, 1),
9021 const0_rtx);
9023 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
9024 varop = gen_rtx_NEG (GET_MODE (varop), varop);
9026 continue;
9029 /* If we have (shift (logical)), move the logical to the outside
9030 to allow it to possibly combine with another logical and the
9031 shift to combine with another shift. This also canonicalizes to
9032 what a ZERO_EXTRACT looks like. Also, some machines have
9033 (and (shift)) insns. */
9035 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
9036 /* We can't do this if we have (ashiftrt (xor)) and the
9037 constant has its sign bit set in shift_mode. */
9038 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
9039 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
9040 shift_mode))
9041 && (new = simplify_binary_operation (code, result_mode,
9042 XEXP (varop, 1),
9043 GEN_INT (count))) != 0
9044 && GET_CODE (new) == CONST_INT
9045 && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop),
9046 INTVAL (new), result_mode, &complement_p))
9048 varop = XEXP (varop, 0);
9049 continue;
9052 /* If we can't do that, try to simplify the shift in each arm of the
9053 logical expression, make a new logical expression, and apply
9054 the inverse distributive law. This also can't be done
9055 for some (ashiftrt (xor)). */
9056 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
9057 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
9058 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
9059 shift_mode)))
9061 rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode,
9062 XEXP (varop, 0), count);
9063 rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode,
9064 XEXP (varop, 1), count);
9066 varop = simplify_gen_binary (GET_CODE (varop), shift_mode,
9067 lhs, rhs);
9068 varop = apply_distributive_law (varop);
9070 count = 0;
9071 continue;
9073 break;
9075 case EQ:
9076 /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE
9077 says that the sign bit can be tested, FOO has mode MODE, C is
9078 GET_MODE_BITSIZE (MODE) - 1, and FOO has only its low-order bit
9079 that may be nonzero. */
9080 if (code == LSHIFTRT
9081 && XEXP (varop, 1) == const0_rtx
9082 && GET_MODE (XEXP (varop, 0)) == result_mode
9083 && count == (unsigned int) (GET_MODE_BITSIZE (result_mode) - 1)
9084 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9085 && ((STORE_FLAG_VALUE
9086 & ((HOST_WIDE_INT) 1
9087 < (GET_MODE_BITSIZE (result_mode) - 1))))
9088 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9089 && merge_outer_ops (&outer_op, &outer_const, XOR,
9090 (HOST_WIDE_INT) 1, result_mode,
9091 &complement_p))
9093 varop = XEXP (varop, 0);
9094 count = 0;
9095 continue;
9097 break;
9099 case NEG:
9100 /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less
9101 than the number of bits in the mode is equivalent to A. */
9102 if (code == LSHIFTRT
9103 && count == (unsigned int) (GET_MODE_BITSIZE (result_mode) - 1)
9104 && nonzero_bits (XEXP (varop, 0), result_mode) == 1)
9106 varop = XEXP (varop, 0);
9107 count = 0;
9108 continue;
9111 /* NEG commutes with ASHIFT since it is multiplication. Move the
9112 NEG outside to allow shifts to combine. */
9113 if (code == ASHIFT
9114 && merge_outer_ops (&outer_op, &outer_const, NEG,
9115 (HOST_WIDE_INT) 0, result_mode,
9116 &complement_p))
9118 varop = XEXP (varop, 0);
9119 continue;
9121 break;
9123 case PLUS:
9124 /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C
9125 is one less than the number of bits in the mode is
9126 equivalent to (xor A 1). */
9127 if (code == LSHIFTRT
9128 && count == (unsigned int) (GET_MODE_BITSIZE (result_mode) - 1)
9129 && XEXP (varop, 1) == constm1_rtx
9130 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9131 && merge_outer_ops (&outer_op, &outer_const, XOR,
9132 (HOST_WIDE_INT) 1, result_mode,
9133 &complement_p))
9135 count = 0;
9136 varop = XEXP (varop, 0);
9137 continue;
9140 /* If we have (xshiftrt (plus FOO BAR) C), and the only bits
9141 that might be nonzero in BAR are those being shifted out and those
9142 bits are known zero in FOO, we can replace the PLUS with FOO.
9143 Similarly in the other operand order. This code occurs when
9144 we are computing the size of a variable-size array. */
9146 if ((code == ASHIFTRT || code == LSHIFTRT)
9147 && count < HOST_BITS_PER_WIDE_INT
9148 && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0
9149 && (nonzero_bits (XEXP (varop, 1), result_mode)
9150 & nonzero_bits (XEXP (varop, 0), result_mode)) == 0)
9152 varop = XEXP (varop, 0);
9153 continue;
9155 else if ((code == ASHIFTRT || code == LSHIFTRT)
9156 && count < HOST_BITS_PER_WIDE_INT
9157 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9158 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9159 >> count)
9160 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9161 & nonzero_bits (XEXP (varop, 1),
9162 result_mode)))
9164 varop = XEXP (varop, 1);
9165 continue;
9168 /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */
9169 if (code == ASHIFT
9170 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9171 && (new = simplify_binary_operation (ASHIFT, result_mode,
9172 XEXP (varop, 1),
9173 GEN_INT (count))) != 0
9174 && GET_CODE (new) == CONST_INT
9175 && merge_outer_ops (&outer_op, &outer_const, PLUS,
9176 INTVAL (new), result_mode, &complement_p))
9178 varop = XEXP (varop, 0);
9179 continue;
9182 /* Check for 'PLUS signbit', which is the canonical form of 'XOR
9183 signbit', and attempt to change the PLUS to an XOR and move it to
9184 the outer operation as is done above in the AND/IOR/XOR case
9185 leg for shift(logical). See details in logical handling above
9186 for reasoning in doing so. */
9187 if (code == LSHIFTRT
9188 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9189 && mode_signbit_p (result_mode, XEXP (varop, 1))
9190 && (new = simplify_binary_operation (code, result_mode,
9191 XEXP (varop, 1),
9192 GEN_INT (count))) != 0
9193 && GET_CODE (new) == CONST_INT
9194 && merge_outer_ops (&outer_op, &outer_const, XOR,
9195 INTVAL (new), result_mode, &complement_p))
9197 varop = XEXP (varop, 0);
9198 continue;
9201 break;
9203 case MINUS:
9204 /* If we have (xshiftrt (minus (ashiftrt X C)) X) C)
9205 with C the size of VAROP - 1 and the shift is logical if
9206 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
9207 we have a (gt X 0) operation. If the shift is arithmetic with
9208 STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1,
9209 we have a (neg (gt X 0)) operation. */
9211 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
9212 && GET_CODE (XEXP (varop, 0)) == ASHIFTRT
9213 && count == (unsigned int)
9214 (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
9215 && (code == LSHIFTRT || code == ASHIFTRT)
9216 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9217 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (varop, 0), 1))
9218 == count
9219 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
9221 count = 0;
9222 varop = gen_rtx_GT (GET_MODE (varop), XEXP (varop, 1),
9223 const0_rtx);
9225 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
9226 varop = gen_rtx_NEG (GET_MODE (varop), varop);
9228 continue;
9230 break;
9232 case TRUNCATE:
9233 /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt))
9234 if the truncate does not affect the value. */
9235 if (code == LSHIFTRT
9236 && GET_CODE (XEXP (varop, 0)) == LSHIFTRT
9237 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9238 && (INTVAL (XEXP (XEXP (varop, 0), 1))
9239 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (varop, 0)))
9240 - GET_MODE_BITSIZE (GET_MODE (varop)))))
9242 rtx varop_inner = XEXP (varop, 0);
9244 varop_inner
9245 = gen_rtx_LSHIFTRT (GET_MODE (varop_inner),
9246 XEXP (varop_inner, 0),
9247 GEN_INT
9248 (count + INTVAL (XEXP (varop_inner, 1))));
9249 varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner);
9250 count = 0;
9251 continue;
9253 break;
9255 default:
9256 break;
9259 break;
9262 /* We need to determine what mode to do the shift in. If the shift is
9263 a right shift or ROTATE, we must always do it in the mode it was
9264 originally done in. Otherwise, we can do it in MODE, the widest mode
9265 encountered. The code we care about is that of the shift that will
9266 actually be done, not the shift that was originally requested. */
9267 shift_mode
9268 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
9269 ? result_mode : mode);
9271 /* We have now finished analyzing the shift. The result should be
9272 a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If
9273 OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied
9274 to the result of the shift. OUTER_CONST is the relevant constant,
9275 but we must turn off all bits turned off in the shift.
9277 If we were passed a value for X, see if we can use any pieces of
9278 it. If not, make new rtx. */
9280 if (x && GET_RTX_CLASS (GET_CODE (x)) == RTX_BIN_ARITH
9281 && GET_CODE (XEXP (x, 1)) == CONST_INT
9282 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (x, 1)) == count)
9283 const_rtx = XEXP (x, 1);
9284 else
9285 const_rtx = GEN_INT (count);
9287 if (x && GET_CODE (XEXP (x, 0)) == SUBREG
9288 && GET_MODE (XEXP (x, 0)) == shift_mode
9289 && SUBREG_REG (XEXP (x, 0)) == varop)
9290 varop = XEXP (x, 0);
9291 else if (GET_MODE (varop) != shift_mode)
9292 varop = gen_lowpart (shift_mode, varop);
9294 /* If we can't make the SUBREG, try to return what we were given. */
9295 if (GET_CODE (varop) == CLOBBER)
9296 return x ? x : varop;
9298 new = simplify_binary_operation (code, shift_mode, varop, const_rtx);
9299 if (new != 0)
9300 x = new;
9301 else
9302 x = gen_rtx_fmt_ee (code, shift_mode, varop, const_rtx);
9304 /* If we have an outer operation and we just made a shift, it is
9305 possible that we could have simplified the shift were it not
9306 for the outer operation. So try to do the simplification
9307 recursively. */
9309 if (outer_op != UNKNOWN && GET_CODE (x) == code
9310 && GET_CODE (XEXP (x, 1)) == CONST_INT)
9311 x = simplify_shift_const (x, code, shift_mode, XEXP (x, 0),
9312 INTVAL (XEXP (x, 1)));
9314 /* If we were doing an LSHIFTRT in a wider mode than it was originally,
9315 turn off all the bits that the shift would have turned off. */
9316 if (orig_code == LSHIFTRT && result_mode != shift_mode)
9317 x = simplify_and_const_int (NULL_RTX, shift_mode, x,
9318 GET_MODE_MASK (result_mode) >> orig_count);
9320 /* Do the remainder of the processing in RESULT_MODE. */
9321 x = gen_lowpart (result_mode, x);
9323 /* If COMPLEMENT_P is set, we have to complement X before doing the outer
9324 operation. */
9325 if (complement_p)
9326 x = simplify_gen_unary (NOT, result_mode, x, result_mode);
9328 if (outer_op != UNKNOWN)
9330 if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT)
9331 outer_const = trunc_int_for_mode (outer_const, result_mode);
9333 if (outer_op == AND)
9334 x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const);
9335 else if (outer_op == SET)
9336 /* This means that we have determined that the result is
9337 equivalent to a constant. This should be rare. */
9338 x = GEN_INT (outer_const);
9339 else if (GET_RTX_CLASS (outer_op) == RTX_UNARY)
9340 x = simplify_gen_unary (outer_op, result_mode, x, result_mode);
9341 else
9342 x = simplify_gen_binary (outer_op, result_mode, x,
9343 GEN_INT (outer_const));
9346 return x;
9349 /* Like recog, but we receive the address of a pointer to a new pattern.
9350 We try to match the rtx that the pointer points to.
9351 If that fails, we may try to modify or replace the pattern,
9352 storing the replacement into the same pointer object.
9354 Modifications include deletion or addition of CLOBBERs.
9356 PNOTES is a pointer to a location where any REG_UNUSED notes added for
9357 the CLOBBERs are placed.
9359 The value is the final insn code from the pattern ultimately matched,
9360 or -1. */
9362 static int
9363 recog_for_combine (rtx *pnewpat, rtx insn, rtx *pnotes)
9365 rtx pat = *pnewpat;
9366 int insn_code_number;
9367 int num_clobbers_to_add = 0;
9368 int i;
9369 rtx notes = 0;
9370 rtx old_notes, old_pat;
9372 /* If PAT is a PARALLEL, check to see if it contains the CLOBBER
9373 we use to indicate that something didn't match. If we find such a
9374 thing, force rejection. */
9375 if (GET_CODE (pat) == PARALLEL)
9376 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
9377 if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER
9378 && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx)
9379 return -1;
9381 old_pat = PATTERN (insn);
9382 old_notes = REG_NOTES (insn);
9383 PATTERN (insn) = pat;
9384 REG_NOTES (insn) = 0;
9386 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9388 /* If it isn't, there is the possibility that we previously had an insn
9389 that clobbered some register as a side effect, but the combined
9390 insn doesn't need to do that. So try once more without the clobbers
9391 unless this represents an ASM insn. */
9393 if (insn_code_number < 0 && ! check_asm_operands (pat)
9394 && GET_CODE (pat) == PARALLEL)
9396 int pos;
9398 for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++)
9399 if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER)
9401 if (i != pos)
9402 SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i));
9403 pos++;
9406 SUBST_INT (XVECLEN (pat, 0), pos);
9408 if (pos == 1)
9409 pat = XVECEXP (pat, 0, 0);
9411 PATTERN (insn) = pat;
9412 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9414 PATTERN (insn) = old_pat;
9415 REG_NOTES (insn) = old_notes;
9417 /* Recognize all noop sets, these will be killed by followup pass. */
9418 if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat))
9419 insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0;
9421 /* If we had any clobbers to add, make a new pattern than contains
9422 them. Then check to make sure that all of them are dead. */
9423 if (num_clobbers_to_add)
9425 rtx newpat = gen_rtx_PARALLEL (VOIDmode,
9426 rtvec_alloc (GET_CODE (pat) == PARALLEL
9427 ? (XVECLEN (pat, 0)
9428 + num_clobbers_to_add)
9429 : num_clobbers_to_add + 1));
9431 if (GET_CODE (pat) == PARALLEL)
9432 for (i = 0; i < XVECLEN (pat, 0); i++)
9433 XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i);
9434 else
9435 XVECEXP (newpat, 0, 0) = pat;
9437 add_clobbers (newpat, insn_code_number);
9439 for (i = XVECLEN (newpat, 0) - num_clobbers_to_add;
9440 i < XVECLEN (newpat, 0); i++)
9442 if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))
9443 && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn))
9444 return -1;
9445 notes = gen_rtx_EXPR_LIST (REG_UNUSED,
9446 XEXP (XVECEXP (newpat, 0, i), 0), notes);
9448 pat = newpat;
9451 *pnewpat = pat;
9452 *pnotes = notes;
9454 return insn_code_number;
9457 /* Like gen_lowpart_general but for use by combine. In combine it
9458 is not possible to create any new pseudoregs. However, it is
9459 safe to create invalid memory addresses, because combine will
9460 try to recognize them and all they will do is make the combine
9461 attempt fail.
9463 If for some reason this cannot do its job, an rtx
9464 (clobber (const_int 0)) is returned.
9465 An insn containing that will not be recognized. */
9467 static rtx
9468 gen_lowpart_for_combine (enum machine_mode omode, rtx x)
9470 enum machine_mode imode = GET_MODE (x);
9471 unsigned int osize = GET_MODE_SIZE (omode);
9472 unsigned int isize = GET_MODE_SIZE (imode);
9473 rtx result;
9475 if (omode == imode)
9476 return x;
9478 /* Return identity if this is a CONST or symbolic reference. */
9479 if (omode == Pmode
9480 && (GET_CODE (x) == CONST
9481 || GET_CODE (x) == SYMBOL_REF
9482 || GET_CODE (x) == LABEL_REF))
9483 return x;
9485 /* We can only support MODE being wider than a word if X is a
9486 constant integer or has a mode the same size. */
9487 if (GET_MODE_SIZE (omode) > UNITS_PER_WORD
9488 && ! ((imode == VOIDmode
9489 && (GET_CODE (x) == CONST_INT
9490 || GET_CODE (x) == CONST_DOUBLE))
9491 || isize == osize))
9492 goto fail;
9494 /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart
9495 won't know what to do. So we will strip off the SUBREG here and
9496 process normally. */
9497 if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x)))
9499 x = SUBREG_REG (x);
9501 /* For use in case we fall down into the address adjustments
9502 further below, we need to adjust the known mode and size of
9503 x; imode and isize, since we just adjusted x. */
9504 imode = GET_MODE (x);
9506 if (imode == omode)
9507 return x;
9509 isize = GET_MODE_SIZE (imode);
9512 result = gen_lowpart_common (omode, x);
9514 #ifdef CANNOT_CHANGE_MODE_CLASS
9515 if (result != 0 && GET_CODE (result) == SUBREG)
9516 record_subregs_of_mode (result);
9517 #endif
9519 if (result)
9520 return result;
9522 if (MEM_P (x))
9524 int offset = 0;
9526 /* Refuse to work on a volatile memory ref or one with a mode-dependent
9527 address. */
9528 if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0)))
9529 goto fail;
9531 /* If we want to refer to something bigger than the original memref,
9532 generate a paradoxical subreg instead. That will force a reload
9533 of the original memref X. */
9534 if (isize < osize)
9535 return gen_rtx_SUBREG (omode, x, 0);
9537 if (WORDS_BIG_ENDIAN)
9538 offset = MAX (isize, UNITS_PER_WORD) - MAX (osize, UNITS_PER_WORD);
9540 /* Adjust the address so that the address-after-the-data is
9541 unchanged. */
9542 if (BYTES_BIG_ENDIAN)
9543 offset -= MIN (UNITS_PER_WORD, osize) - MIN (UNITS_PER_WORD, isize);
9545 return adjust_address_nv (x, omode, offset);
9548 /* If X is a comparison operator, rewrite it in a new mode. This
9549 probably won't match, but may allow further simplifications. */
9550 else if (COMPARISON_P (x))
9551 return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1));
9553 /* If we couldn't simplify X any other way, just enclose it in a
9554 SUBREG. Normally, this SUBREG won't match, but some patterns may
9555 include an explicit SUBREG or we may simplify it further in combine. */
9556 else
9558 int offset = 0;
9559 rtx res;
9561 offset = subreg_lowpart_offset (omode, imode);
9562 if (imode == VOIDmode)
9564 imode = int_mode_for_mode (omode);
9565 x = gen_lowpart_common (imode, x);
9566 if (x == NULL)
9567 goto fail;
9569 res = simplify_gen_subreg (omode, x, imode, offset);
9570 if (res)
9571 return res;
9574 fail:
9575 return gen_rtx_CLOBBER (imode, const0_rtx);
9578 /* Simplify a comparison between *POP0 and *POP1 where CODE is the
9579 comparison code that will be tested.
9581 The result is a possibly different comparison code to use. *POP0 and
9582 *POP1 may be updated.
9584 It is possible that we might detect that a comparison is either always
9585 true or always false. However, we do not perform general constant
9586 folding in combine, so this knowledge isn't useful. Such tautologies
9587 should have been detected earlier. Hence we ignore all such cases. */
9589 static enum rtx_code
9590 simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1)
9592 rtx op0 = *pop0;
9593 rtx op1 = *pop1;
9594 rtx tem, tem1;
9595 int i;
9596 enum machine_mode mode, tmode;
9598 /* Try a few ways of applying the same transformation to both operands. */
9599 while (1)
9601 #ifndef WORD_REGISTER_OPERATIONS
9602 /* The test below this one won't handle SIGN_EXTENDs on these machines,
9603 so check specially. */
9604 if (code != GTU && code != GEU && code != LTU && code != LEU
9605 && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT
9606 && GET_CODE (XEXP (op0, 0)) == ASHIFT
9607 && GET_CODE (XEXP (op1, 0)) == ASHIFT
9608 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG
9609 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG
9610 && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))
9611 == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))))
9612 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9613 && XEXP (op0, 1) == XEXP (op1, 1)
9614 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
9615 && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1)
9616 && (INTVAL (XEXP (op0, 1))
9617 == (GET_MODE_BITSIZE (GET_MODE (op0))
9618 - (GET_MODE_BITSIZE
9619 (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))))))))
9621 op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0));
9622 op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0));
9624 #endif
9626 /* If both operands are the same constant shift, see if we can ignore the
9627 shift. We can if the shift is a rotate or if the bits shifted out of
9628 this shift are known to be zero for both inputs and if the type of
9629 comparison is compatible with the shift. */
9630 if (GET_CODE (op0) == GET_CODE (op1)
9631 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
9632 && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ))
9633 || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT)
9634 && (code != GT && code != LT && code != GE && code != LE))
9635 || (GET_CODE (op0) == ASHIFTRT
9636 && (code != GTU && code != LTU
9637 && code != GEU && code != LEU)))
9638 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9639 && INTVAL (XEXP (op0, 1)) >= 0
9640 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
9641 && XEXP (op0, 1) == XEXP (op1, 1))
9643 enum machine_mode mode = GET_MODE (op0);
9644 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9645 int shift_count = INTVAL (XEXP (op0, 1));
9647 if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT)
9648 mask &= (mask >> shift_count) << shift_count;
9649 else if (GET_CODE (op0) == ASHIFT)
9650 mask = (mask & (mask << shift_count)) >> shift_count;
9652 if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0
9653 && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0)
9654 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0);
9655 else
9656 break;
9659 /* If both operands are AND's of a paradoxical SUBREG by constant, the
9660 SUBREGs are of the same mode, and, in both cases, the AND would
9661 be redundant if the comparison was done in the narrower mode,
9662 do the comparison in the narrower mode (e.g., we are AND'ing with 1
9663 and the operand's possibly nonzero bits are 0xffffff01; in that case
9664 if we only care about QImode, we don't need the AND). This case
9665 occurs if the output mode of an scc insn is not SImode and
9666 STORE_FLAG_VALUE == 1 (e.g., the 386).
9668 Similarly, check for a case where the AND's are ZERO_EXTEND
9669 operations from some narrower mode even though a SUBREG is not
9670 present. */
9672 else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND
9673 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9674 && GET_CODE (XEXP (op1, 1)) == CONST_INT)
9676 rtx inner_op0 = XEXP (op0, 0);
9677 rtx inner_op1 = XEXP (op1, 0);
9678 HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1));
9679 HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1));
9680 int changed = 0;
9682 if (GET_CODE (inner_op0) == SUBREG && GET_CODE (inner_op1) == SUBREG
9683 && (GET_MODE_SIZE (GET_MODE (inner_op0))
9684 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner_op0))))
9685 && (GET_MODE (SUBREG_REG (inner_op0))
9686 == GET_MODE (SUBREG_REG (inner_op1)))
9687 && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (inner_op0)))
9688 <= HOST_BITS_PER_WIDE_INT)
9689 && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0),
9690 GET_MODE (SUBREG_REG (inner_op0)))))
9691 && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1),
9692 GET_MODE (SUBREG_REG (inner_op1))))))
9694 op0 = SUBREG_REG (inner_op0);
9695 op1 = SUBREG_REG (inner_op1);
9697 /* The resulting comparison is always unsigned since we masked
9698 off the original sign bit. */
9699 code = unsigned_condition (code);
9701 changed = 1;
9704 else if (c0 == c1)
9705 for (tmode = GET_CLASS_NARROWEST_MODE
9706 (GET_MODE_CLASS (GET_MODE (op0)));
9707 tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode))
9708 if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode))
9710 op0 = gen_lowpart (tmode, inner_op0);
9711 op1 = gen_lowpart (tmode, inner_op1);
9712 code = unsigned_condition (code);
9713 changed = 1;
9714 break;
9717 if (! changed)
9718 break;
9721 /* If both operands are NOT, we can strip off the outer operation
9722 and adjust the comparison code for swapped operands; similarly for
9723 NEG, except that this must be an equality comparison. */
9724 else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT)
9725 || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG
9726 && (code == EQ || code == NE)))
9727 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code);
9729 else
9730 break;
9733 /* If the first operand is a constant, swap the operands and adjust the
9734 comparison code appropriately, but don't do this if the second operand
9735 is already a constant integer. */
9736 if (swap_commutative_operands_p (op0, op1))
9738 tem = op0, op0 = op1, op1 = tem;
9739 code = swap_condition (code);
9742 /* We now enter a loop during which we will try to simplify the comparison.
9743 For the most part, we only are concerned with comparisons with zero,
9744 but some things may really be comparisons with zero but not start
9745 out looking that way. */
9747 while (GET_CODE (op1) == CONST_INT)
9749 enum machine_mode mode = GET_MODE (op0);
9750 unsigned int mode_width = GET_MODE_BITSIZE (mode);
9751 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9752 int equality_comparison_p;
9753 int sign_bit_comparison_p;
9754 int unsigned_comparison_p;
9755 HOST_WIDE_INT const_op;
9757 /* We only want to handle integral modes. This catches VOIDmode,
9758 CCmode, and the floating-point modes. An exception is that we
9759 can handle VOIDmode if OP0 is a COMPARE or a comparison
9760 operation. */
9762 if (GET_MODE_CLASS (mode) != MODE_INT
9763 && ! (mode == VOIDmode
9764 && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0))))
9765 break;
9767 /* Get the constant we are comparing against and turn off all bits
9768 not on in our mode. */
9769 const_op = INTVAL (op1);
9770 if (mode != VOIDmode)
9771 const_op = trunc_int_for_mode (const_op, mode);
9772 op1 = GEN_INT (const_op);
9774 /* If we are comparing against a constant power of two and the value
9775 being compared can only have that single bit nonzero (e.g., it was
9776 `and'ed with that bit), we can replace this with a comparison
9777 with zero. */
9778 if (const_op
9779 && (code == EQ || code == NE || code == GE || code == GEU
9780 || code == LT || code == LTU)
9781 && mode_width <= HOST_BITS_PER_WIDE_INT
9782 && exact_log2 (const_op) >= 0
9783 && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
9785 code = (code == EQ || code == GE || code == GEU ? NE : EQ);
9786 op1 = const0_rtx, const_op = 0;
9789 /* Similarly, if we are comparing a value known to be either -1 or
9790 0 with -1, change it to the opposite comparison against zero. */
9792 if (const_op == -1
9793 && (code == EQ || code == NE || code == GT || code == LE
9794 || code == GEU || code == LTU)
9795 && num_sign_bit_copies (op0, mode) == mode_width)
9797 code = (code == EQ || code == LE || code == GEU ? NE : EQ);
9798 op1 = const0_rtx, const_op = 0;
9801 /* Do some canonicalizations based on the comparison code. We prefer
9802 comparisons against zero and then prefer equality comparisons.
9803 If we can reduce the size of a constant, we will do that too. */
9805 switch (code)
9807 case LT:
9808 /* < C is equivalent to <= (C - 1) */
9809 if (const_op > 0)
9811 const_op -= 1;
9812 op1 = GEN_INT (const_op);
9813 code = LE;
9814 /* ... fall through to LE case below. */
9816 else
9817 break;
9819 case LE:
9820 /* <= C is equivalent to < (C + 1); we do this for C < 0 */
9821 if (const_op < 0)
9823 const_op += 1;
9824 op1 = GEN_INT (const_op);
9825 code = LT;
9828 /* If we are doing a <= 0 comparison on a value known to have
9829 a zero sign bit, we can replace this with == 0. */
9830 else if (const_op == 0
9831 && mode_width <= HOST_BITS_PER_WIDE_INT
9832 && (nonzero_bits (op0, mode)
9833 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9834 code = EQ;
9835 break;
9837 case GE:
9838 /* >= C is equivalent to > (C - 1). */
9839 if (const_op > 0)
9841 const_op -= 1;
9842 op1 = GEN_INT (const_op);
9843 code = GT;
9844 /* ... fall through to GT below. */
9846 else
9847 break;
9849 case GT:
9850 /* > C is equivalent to >= (C + 1); we do this for C < 0. */
9851 if (const_op < 0)
9853 const_op += 1;
9854 op1 = GEN_INT (const_op);
9855 code = GE;
9858 /* If we are doing a > 0 comparison on a value known to have
9859 a zero sign bit, we can replace this with != 0. */
9860 else if (const_op == 0
9861 && mode_width <= HOST_BITS_PER_WIDE_INT
9862 && (nonzero_bits (op0, mode)
9863 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9864 code = NE;
9865 break;
9867 case LTU:
9868 /* < C is equivalent to <= (C - 1). */
9869 if (const_op > 0)
9871 const_op -= 1;
9872 op1 = GEN_INT (const_op);
9873 code = LEU;
9874 /* ... fall through ... */
9877 /* (unsigned) < 0x80000000 is equivalent to >= 0. */
9878 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9879 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9881 const_op = 0, op1 = const0_rtx;
9882 code = GE;
9883 break;
9885 else
9886 break;
9888 case LEU:
9889 /* unsigned <= 0 is equivalent to == 0 */
9890 if (const_op == 0)
9891 code = EQ;
9893 /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */
9894 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9895 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9897 const_op = 0, op1 = const0_rtx;
9898 code = GE;
9900 break;
9902 case GEU:
9903 /* >= C is equivalent to > (C - 1). */
9904 if (const_op > 1)
9906 const_op -= 1;
9907 op1 = GEN_INT (const_op);
9908 code = GTU;
9909 /* ... fall through ... */
9912 /* (unsigned) >= 0x80000000 is equivalent to < 0. */
9913 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9914 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9916 const_op = 0, op1 = const0_rtx;
9917 code = LT;
9918 break;
9920 else
9921 break;
9923 case GTU:
9924 /* unsigned > 0 is equivalent to != 0 */
9925 if (const_op == 0)
9926 code = NE;
9928 /* (unsigned) > 0x7fffffff is equivalent to < 0. */
9929 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9930 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9932 const_op = 0, op1 = const0_rtx;
9933 code = LT;
9935 break;
9937 default:
9938 break;
9941 /* Compute some predicates to simplify code below. */
9943 equality_comparison_p = (code == EQ || code == NE);
9944 sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0);
9945 unsigned_comparison_p = (code == LTU || code == LEU || code == GTU
9946 || code == GEU);
9948 /* If this is a sign bit comparison and we can do arithmetic in
9949 MODE, say that we will only be needing the sign bit of OP0. */
9950 if (sign_bit_comparison_p
9951 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
9952 op0 = force_to_mode (op0, mode,
9953 ((HOST_WIDE_INT) 1
9954 << (GET_MODE_BITSIZE (mode) - 1)),
9955 NULL_RTX, 0);
9957 /* Now try cases based on the opcode of OP0. If none of the cases
9958 does a "continue", we exit this loop immediately after the
9959 switch. */
9961 switch (GET_CODE (op0))
9963 case ZERO_EXTRACT:
9964 /* If we are extracting a single bit from a variable position in
9965 a constant that has only a single bit set and are comparing it
9966 with zero, we can convert this into an equality comparison
9967 between the position and the location of the single bit. */
9968 /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might
9969 have already reduced the shift count modulo the word size. */
9970 if (!SHIFT_COUNT_TRUNCATED
9971 && GET_CODE (XEXP (op0, 0)) == CONST_INT
9972 && XEXP (op0, 1) == const1_rtx
9973 && equality_comparison_p && const_op == 0
9974 && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0)
9976 if (BITS_BIG_ENDIAN)
9978 enum machine_mode new_mode
9979 = mode_for_extraction (EP_extzv, 1);
9980 if (new_mode == MAX_MACHINE_MODE)
9981 i = BITS_PER_WORD - 1 - i;
9982 else
9984 mode = new_mode;
9985 i = (GET_MODE_BITSIZE (mode) - 1 - i);
9989 op0 = XEXP (op0, 2);
9990 op1 = GEN_INT (i);
9991 const_op = i;
9993 /* Result is nonzero iff shift count is equal to I. */
9994 code = reverse_condition (code);
9995 continue;
9998 /* ... fall through ... */
10000 case SIGN_EXTRACT:
10001 tem = expand_compound_operation (op0);
10002 if (tem != op0)
10004 op0 = tem;
10005 continue;
10007 break;
10009 case NOT:
10010 /* If testing for equality, we can take the NOT of the constant. */
10011 if (equality_comparison_p
10012 && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0)
10014 op0 = XEXP (op0, 0);
10015 op1 = tem;
10016 continue;
10019 /* If just looking at the sign bit, reverse the sense of the
10020 comparison. */
10021 if (sign_bit_comparison_p)
10023 op0 = XEXP (op0, 0);
10024 code = (code == GE ? LT : GE);
10025 continue;
10027 break;
10029 case NEG:
10030 /* If testing for equality, we can take the NEG of the constant. */
10031 if (equality_comparison_p
10032 && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0)
10034 op0 = XEXP (op0, 0);
10035 op1 = tem;
10036 continue;
10039 /* The remaining cases only apply to comparisons with zero. */
10040 if (const_op != 0)
10041 break;
10043 /* When X is ABS or is known positive,
10044 (neg X) is < 0 if and only if X != 0. */
10046 if (sign_bit_comparison_p
10047 && (GET_CODE (XEXP (op0, 0)) == ABS
10048 || (mode_width <= HOST_BITS_PER_WIDE_INT
10049 && (nonzero_bits (XEXP (op0, 0), mode)
10050 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)))
10052 op0 = XEXP (op0, 0);
10053 code = (code == LT ? NE : EQ);
10054 continue;
10057 /* If we have NEG of something whose two high-order bits are the
10058 same, we know that "(-a) < 0" is equivalent to "a > 0". */
10059 if (num_sign_bit_copies (op0, mode) >= 2)
10061 op0 = XEXP (op0, 0);
10062 code = swap_condition (code);
10063 continue;
10065 break;
10067 case ROTATE:
10068 /* If we are testing equality and our count is a constant, we
10069 can perform the inverse operation on our RHS. */
10070 if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10071 && (tem = simplify_binary_operation (ROTATERT, mode,
10072 op1, XEXP (op0, 1))) != 0)
10074 op0 = XEXP (op0, 0);
10075 op1 = tem;
10076 continue;
10079 /* If we are doing a < 0 or >= 0 comparison, it means we are testing
10080 a particular bit. Convert it to an AND of a constant of that
10081 bit. This will be converted into a ZERO_EXTRACT. */
10082 if (const_op == 0 && sign_bit_comparison_p
10083 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10084 && mode_width <= HOST_BITS_PER_WIDE_INT)
10086 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10087 ((HOST_WIDE_INT) 1
10088 << (mode_width - 1
10089 - INTVAL (XEXP (op0, 1)))));
10090 code = (code == LT ? NE : EQ);
10091 continue;
10094 /* Fall through. */
10096 case ABS:
10097 /* ABS is ignorable inside an equality comparison with zero. */
10098 if (const_op == 0 && equality_comparison_p)
10100 op0 = XEXP (op0, 0);
10101 continue;
10103 break;
10105 case SIGN_EXTEND:
10106 /* Can simplify (compare (zero/sign_extend FOO) CONST) to
10107 (compare FOO CONST) if CONST fits in FOO's mode and we
10108 are either testing inequality or have an unsigned
10109 comparison with ZERO_EXTEND or a signed comparison with
10110 SIGN_EXTEND. But don't do it if we don't have a compare
10111 insn of the given mode, since we'd have to revert it
10112 later on, and then we wouldn't know whether to sign- or
10113 zero-extend. */
10114 mode = GET_MODE (XEXP (op0, 0));
10115 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10116 && ! unsigned_comparison_p
10117 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10118 && ((unsigned HOST_WIDE_INT) const_op
10119 < (((unsigned HOST_WIDE_INT) 1
10120 << (GET_MODE_BITSIZE (mode) - 1))))
10121 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10123 op0 = XEXP (op0, 0);
10124 continue;
10126 break;
10128 case SUBREG:
10129 /* Check for the case where we are comparing A - C1 with C2, that is
10131 (subreg:MODE (plus (A) (-C1))) op (C2)
10133 with C1 a constant, and try to lift the SUBREG, i.e. to do the
10134 comparison in the wider mode. One of the following two conditions
10135 must be true in order for this to be valid:
10137 1. The mode extension results in the same bit pattern being added
10138 on both sides and the comparison is equality or unsigned. As
10139 C2 has been truncated to fit in MODE, the pattern can only be
10140 all 0s or all 1s.
10142 2. The mode extension results in the sign bit being copied on
10143 each side.
10145 The difficulty here is that we have predicates for A but not for
10146 (A - C1) so we need to check that C1 is within proper bounds so
10147 as to perturbate A as little as possible. */
10149 if (mode_width <= HOST_BITS_PER_WIDE_INT
10150 && subreg_lowpart_p (op0)
10151 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) > mode_width
10152 && GET_CODE (SUBREG_REG (op0)) == PLUS
10153 && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT)
10155 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
10156 rtx a = XEXP (SUBREG_REG (op0), 0);
10157 HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1));
10159 if ((c1 > 0
10160 && (unsigned HOST_WIDE_INT) c1
10161 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)
10162 && (equality_comparison_p || unsigned_comparison_p)
10163 /* (A - C1) zero-extends if it is positive and sign-extends
10164 if it is negative, C2 both zero- and sign-extends. */
10165 && ((0 == (nonzero_bits (a, inner_mode)
10166 & ~GET_MODE_MASK (mode))
10167 && const_op >= 0)
10168 /* (A - C1) sign-extends if it is positive and 1-extends
10169 if it is negative, C2 both sign- and 1-extends. */
10170 || (num_sign_bit_copies (a, inner_mode)
10171 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10172 - mode_width)
10173 && const_op < 0)))
10174 || ((unsigned HOST_WIDE_INT) c1
10175 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 2)
10176 /* (A - C1) always sign-extends, like C2. */
10177 && num_sign_bit_copies (a, inner_mode)
10178 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10179 - mode_width - 1)))
10181 op0 = SUBREG_REG (op0);
10182 continue;
10186 /* If the inner mode is narrower and we are extracting the low part,
10187 we can treat the SUBREG as if it were a ZERO_EXTEND. */
10188 if (subreg_lowpart_p (op0)
10189 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width)
10190 /* Fall through */ ;
10191 else
10192 break;
10194 /* ... fall through ... */
10196 case ZERO_EXTEND:
10197 mode = GET_MODE (XEXP (op0, 0));
10198 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10199 && (unsigned_comparison_p || equality_comparison_p)
10200 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10201 && ((unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode))
10202 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10204 op0 = XEXP (op0, 0);
10205 continue;
10207 break;
10209 case PLUS:
10210 /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do
10211 this for equality comparisons due to pathological cases involving
10212 overflows. */
10213 if (equality_comparison_p
10214 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10215 op1, XEXP (op0, 1))))
10217 op0 = XEXP (op0, 0);
10218 op1 = tem;
10219 continue;
10222 /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */
10223 if (const_op == 0 && XEXP (op0, 1) == constm1_rtx
10224 && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p)
10226 op0 = XEXP (XEXP (op0, 0), 0);
10227 code = (code == LT ? EQ : NE);
10228 continue;
10230 break;
10232 case MINUS:
10233 /* We used to optimize signed comparisons against zero, but that
10234 was incorrect. Unsigned comparisons against zero (GTU, LEU)
10235 arrive here as equality comparisons, or (GEU, LTU) are
10236 optimized away. No need to special-case them. */
10238 /* (eq (minus A B) C) -> (eq A (plus B C)) or
10239 (eq B (minus A C)), whichever simplifies. We can only do
10240 this for equality comparisons due to pathological cases involving
10241 overflows. */
10242 if (equality_comparison_p
10243 && 0 != (tem = simplify_binary_operation (PLUS, mode,
10244 XEXP (op0, 1), op1)))
10246 op0 = XEXP (op0, 0);
10247 op1 = tem;
10248 continue;
10251 if (equality_comparison_p
10252 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10253 XEXP (op0, 0), op1)))
10255 op0 = XEXP (op0, 1);
10256 op1 = tem;
10257 continue;
10260 /* The sign bit of (minus (ashiftrt X C) X), where C is the number
10261 of bits in X minus 1, is one iff X > 0. */
10262 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT
10263 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10264 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (op0, 0), 1))
10265 == mode_width - 1
10266 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10268 op0 = XEXP (op0, 1);
10269 code = (code == GE ? LE : GT);
10270 continue;
10272 break;
10274 case XOR:
10275 /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification
10276 if C is zero or B is a constant. */
10277 if (equality_comparison_p
10278 && 0 != (tem = simplify_binary_operation (XOR, mode,
10279 XEXP (op0, 1), op1)))
10281 op0 = XEXP (op0, 0);
10282 op1 = tem;
10283 continue;
10285 break;
10287 case EQ: case NE:
10288 case UNEQ: case LTGT:
10289 case LT: case LTU: case UNLT: case LE: case LEU: case UNLE:
10290 case GT: case GTU: case UNGT: case GE: case GEU: case UNGE:
10291 case UNORDERED: case ORDERED:
10292 /* We can't do anything if OP0 is a condition code value, rather
10293 than an actual data value. */
10294 if (const_op != 0
10295 || CC0_P (XEXP (op0, 0))
10296 || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC)
10297 break;
10299 /* Get the two operands being compared. */
10300 if (GET_CODE (XEXP (op0, 0)) == COMPARE)
10301 tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1);
10302 else
10303 tem = XEXP (op0, 0), tem1 = XEXP (op0, 1);
10305 /* Check for the cases where we simply want the result of the
10306 earlier test or the opposite of that result. */
10307 if (code == NE || code == EQ
10308 || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
10309 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10310 && (STORE_FLAG_VALUE
10311 & (((HOST_WIDE_INT) 1
10312 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
10313 && (code == LT || code == GE)))
10315 enum rtx_code new_code;
10316 if (code == LT || code == NE)
10317 new_code = GET_CODE (op0);
10318 else
10319 new_code = reversed_comparison_code (op0, NULL);
10321 if (new_code != UNKNOWN)
10323 code = new_code;
10324 op0 = tem;
10325 op1 = tem1;
10326 continue;
10329 break;
10331 case IOR:
10332 /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero
10333 iff X <= 0. */
10334 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS
10335 && XEXP (XEXP (op0, 0), 1) == constm1_rtx
10336 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10338 op0 = XEXP (op0, 1);
10339 code = (code == GE ? GT : LE);
10340 continue;
10342 break;
10344 case AND:
10345 /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This
10346 will be converted to a ZERO_EXTRACT later. */
10347 if (const_op == 0 && equality_comparison_p
10348 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10349 && XEXP (XEXP (op0, 0), 0) == const1_rtx)
10351 op0 = simplify_and_const_int
10352 (op0, mode, gen_rtx_LSHIFTRT (mode,
10353 XEXP (op0, 1),
10354 XEXP (XEXP (op0, 0), 1)),
10355 (HOST_WIDE_INT) 1);
10356 continue;
10359 /* If we are comparing (and (lshiftrt X C1) C2) for equality with
10360 zero and X is a comparison and C1 and C2 describe only bits set
10361 in STORE_FLAG_VALUE, we can compare with X. */
10362 if (const_op == 0 && equality_comparison_p
10363 && mode_width <= HOST_BITS_PER_WIDE_INT
10364 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10365 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT
10366 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10367 && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0
10368 && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT)
10370 mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10371 << INTVAL (XEXP (XEXP (op0, 0), 1)));
10372 if ((~STORE_FLAG_VALUE & mask) == 0
10373 && (COMPARISON_P (XEXP (XEXP (op0, 0), 0))
10374 || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0
10375 && COMPARISON_P (tem))))
10377 op0 = XEXP (XEXP (op0, 0), 0);
10378 continue;
10382 /* If we are doing an equality comparison of an AND of a bit equal
10383 to the sign bit, replace this with a LT or GE comparison of
10384 the underlying value. */
10385 if (equality_comparison_p
10386 && const_op == 0
10387 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10388 && mode_width <= HOST_BITS_PER_WIDE_INT
10389 && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10390 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10392 op0 = XEXP (op0, 0);
10393 code = (code == EQ ? GE : LT);
10394 continue;
10397 /* If this AND operation is really a ZERO_EXTEND from a narrower
10398 mode, the constant fits within that mode, and this is either an
10399 equality or unsigned comparison, try to do this comparison in
10400 the narrower mode. */
10401 if ((equality_comparison_p || unsigned_comparison_p)
10402 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10403 && (i = exact_log2 ((INTVAL (XEXP (op0, 1))
10404 & GET_MODE_MASK (mode))
10405 + 1)) >= 0
10406 && const_op >> i == 0
10407 && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode)
10409 op0 = gen_lowpart (tmode, XEXP (op0, 0));
10410 continue;
10413 /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1
10414 fits in both M1 and M2 and the SUBREG is either paradoxical
10415 or represents the low part, permute the SUBREG and the AND
10416 and try again. */
10417 if (GET_CODE (XEXP (op0, 0)) == SUBREG)
10419 unsigned HOST_WIDE_INT c1;
10420 tmode = GET_MODE (SUBREG_REG (XEXP (op0, 0)));
10421 /* Require an integral mode, to avoid creating something like
10422 (AND:SF ...). */
10423 if (SCALAR_INT_MODE_P (tmode)
10424 /* It is unsafe to commute the AND into the SUBREG if the
10425 SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is
10426 not defined. As originally written the upper bits
10427 have a defined value due to the AND operation.
10428 However, if we commute the AND inside the SUBREG then
10429 they no longer have defined values and the meaning of
10430 the code has been changed. */
10431 && (0
10432 #ifdef WORD_REGISTER_OPERATIONS
10433 || (mode_width > GET_MODE_BITSIZE (tmode)
10434 && mode_width <= BITS_PER_WORD)
10435 #endif
10436 || (mode_width <= GET_MODE_BITSIZE (tmode)
10437 && subreg_lowpart_p (XEXP (op0, 0))))
10438 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10439 && mode_width <= HOST_BITS_PER_WIDE_INT
10440 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
10441 && ((c1 = INTVAL (XEXP (op0, 1))) & ~mask) == 0
10442 && (c1 & ~GET_MODE_MASK (tmode)) == 0
10443 && c1 != mask
10444 && c1 != GET_MODE_MASK (tmode))
10446 op0 = simplify_gen_binary (AND, tmode,
10447 SUBREG_REG (XEXP (op0, 0)),
10448 gen_int_mode (c1, tmode));
10449 op0 = gen_lowpart (mode, op0);
10450 continue;
10454 /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */
10455 if (const_op == 0 && equality_comparison_p
10456 && XEXP (op0, 1) == const1_rtx
10457 && GET_CODE (XEXP (op0, 0)) == NOT)
10459 op0 = simplify_and_const_int
10460 (NULL_RTX, mode, XEXP (XEXP (op0, 0), 0), (HOST_WIDE_INT) 1);
10461 code = (code == NE ? EQ : NE);
10462 continue;
10465 /* Convert (ne (and (lshiftrt (not X)) 1) 0) to
10466 (eq (and (lshiftrt X) 1) 0).
10467 Also handle the case where (not X) is expressed using xor. */
10468 if (const_op == 0 && equality_comparison_p
10469 && XEXP (op0, 1) == const1_rtx
10470 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT)
10472 rtx shift_op = XEXP (XEXP (op0, 0), 0);
10473 rtx shift_count = XEXP (XEXP (op0, 0), 1);
10475 if (GET_CODE (shift_op) == NOT
10476 || (GET_CODE (shift_op) == XOR
10477 && GET_CODE (XEXP (shift_op, 1)) == CONST_INT
10478 && GET_CODE (shift_count) == CONST_INT
10479 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
10480 && (INTVAL (XEXP (shift_op, 1))
10481 == (HOST_WIDE_INT) 1 << INTVAL (shift_count))))
10483 op0 = simplify_and_const_int
10484 (NULL_RTX, mode,
10485 gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count),
10486 (HOST_WIDE_INT) 1);
10487 code = (code == NE ? EQ : NE);
10488 continue;
10491 break;
10493 case ASHIFT:
10494 /* If we have (compare (ashift FOO N) (const_int C)) and
10495 the high order N bits of FOO (N+1 if an inequality comparison)
10496 are known to be zero, we can do this by comparing FOO with C
10497 shifted right N bits so long as the low-order N bits of C are
10498 zero. */
10499 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10500 && INTVAL (XEXP (op0, 1)) >= 0
10501 && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p)
10502 < HOST_BITS_PER_WIDE_INT)
10503 && ((const_op
10504 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0)
10505 && mode_width <= HOST_BITS_PER_WIDE_INT
10506 && (nonzero_bits (XEXP (op0, 0), mode)
10507 & ~(mask >> (INTVAL (XEXP (op0, 1))
10508 + ! equality_comparison_p))) == 0)
10510 /* We must perform a logical shift, not an arithmetic one,
10511 as we want the top N bits of C to be zero. */
10512 unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode);
10514 temp >>= INTVAL (XEXP (op0, 1));
10515 op1 = gen_int_mode (temp, mode);
10516 op0 = XEXP (op0, 0);
10517 continue;
10520 /* If we are doing a sign bit comparison, it means we are testing
10521 a particular bit. Convert it to the appropriate AND. */
10522 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10523 && mode_width <= HOST_BITS_PER_WIDE_INT)
10525 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10526 ((HOST_WIDE_INT) 1
10527 << (mode_width - 1
10528 - INTVAL (XEXP (op0, 1)))));
10529 code = (code == LT ? NE : EQ);
10530 continue;
10533 /* If this an equality comparison with zero and we are shifting
10534 the low bit to the sign bit, we can convert this to an AND of the
10535 low-order bit. */
10536 if (const_op == 0 && equality_comparison_p
10537 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10538 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10539 == mode_width - 1)
10541 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10542 (HOST_WIDE_INT) 1);
10543 continue;
10545 break;
10547 case ASHIFTRT:
10548 /* If this is an equality comparison with zero, we can do this
10549 as a logical shift, which might be much simpler. */
10550 if (equality_comparison_p && const_op == 0
10551 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
10553 op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode,
10554 XEXP (op0, 0),
10555 INTVAL (XEXP (op0, 1)));
10556 continue;
10559 /* If OP0 is a sign extension and CODE is not an unsigned comparison,
10560 do the comparison in a narrower mode. */
10561 if (! unsigned_comparison_p
10562 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10563 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10564 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
10565 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10566 MODE_INT, 1)) != BLKmode
10567 && (((unsigned HOST_WIDE_INT) const_op
10568 + (GET_MODE_MASK (tmode) >> 1) + 1)
10569 <= GET_MODE_MASK (tmode)))
10571 op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0));
10572 continue;
10575 /* Likewise if OP0 is a PLUS of a sign extension with a
10576 constant, which is usually represented with the PLUS
10577 between the shifts. */
10578 if (! unsigned_comparison_p
10579 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10580 && GET_CODE (XEXP (op0, 0)) == PLUS
10581 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10582 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT
10583 && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1)
10584 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10585 MODE_INT, 1)) != BLKmode
10586 && (((unsigned HOST_WIDE_INT) const_op
10587 + (GET_MODE_MASK (tmode) >> 1) + 1)
10588 <= GET_MODE_MASK (tmode)))
10590 rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0);
10591 rtx add_const = XEXP (XEXP (op0, 0), 1);
10592 rtx new_const = simplify_gen_binary (ASHIFTRT, GET_MODE (op0),
10593 add_const, XEXP (op0, 1));
10595 op0 = simplify_gen_binary (PLUS, tmode,
10596 gen_lowpart (tmode, inner),
10597 new_const);
10598 continue;
10601 /* ... fall through ... */
10602 case LSHIFTRT:
10603 /* If we have (compare (xshiftrt FOO N) (const_int C)) and
10604 the low order N bits of FOO are known to be zero, we can do this
10605 by comparing FOO with C shifted left N bits so long as no
10606 overflow occurs. */
10607 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10608 && INTVAL (XEXP (op0, 1)) >= 0
10609 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
10610 && mode_width <= HOST_BITS_PER_WIDE_INT
10611 && (nonzero_bits (XEXP (op0, 0), mode)
10612 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0
10613 && (((unsigned HOST_WIDE_INT) const_op
10614 + (GET_CODE (op0) != LSHIFTRT
10615 ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1)
10616 + 1)
10617 : 0))
10618 <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1))))
10620 /* If the shift was logical, then we must make the condition
10621 unsigned. */
10622 if (GET_CODE (op0) == LSHIFTRT)
10623 code = unsigned_condition (code);
10625 const_op <<= INTVAL (XEXP (op0, 1));
10626 op1 = GEN_INT (const_op);
10627 op0 = XEXP (op0, 0);
10628 continue;
10631 /* If we are using this shift to extract just the sign bit, we
10632 can replace this with an LT or GE comparison. */
10633 if (const_op == 0
10634 && (equality_comparison_p || sign_bit_comparison_p)
10635 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10636 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10637 == mode_width - 1)
10639 op0 = XEXP (op0, 0);
10640 code = (code == NE || code == GT ? LT : GE);
10641 continue;
10643 break;
10645 default:
10646 break;
10649 break;
10652 /* Now make any compound operations involved in this comparison. Then,
10653 check for an outmost SUBREG on OP0 that is not doing anything or is
10654 paradoxical. The latter transformation must only be performed when
10655 it is known that the "extra" bits will be the same in op0 and op1 or
10656 that they don't matter. There are three cases to consider:
10658 1. SUBREG_REG (op0) is a register. In this case the bits are don't
10659 care bits and we can assume they have any convenient value. So
10660 making the transformation is safe.
10662 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not defined.
10663 In this case the upper bits of op0 are undefined. We should not make
10664 the simplification in that case as we do not know the contents of
10665 those bits.
10667 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined and not
10668 UNKNOWN. In that case we know those bits are zeros or ones. We must
10669 also be sure that they are the same as the upper bits of op1.
10671 We can never remove a SUBREG for a non-equality comparison because
10672 the sign bit is in a different place in the underlying object. */
10674 op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET);
10675 op1 = make_compound_operation (op1, SET);
10677 if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0)
10678 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10679 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op0))) == MODE_INT
10680 && (code == NE || code == EQ))
10682 if (GET_MODE_SIZE (GET_MODE (op0))
10683 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))
10685 /* For paradoxical subregs, allow case 1 as above. Case 3 isn't
10686 implemented. */
10687 if (REG_P (SUBREG_REG (op0)))
10689 op0 = SUBREG_REG (op0);
10690 op1 = gen_lowpart (GET_MODE (op0), op1);
10693 else if ((GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0)))
10694 <= HOST_BITS_PER_WIDE_INT)
10695 && (nonzero_bits (SUBREG_REG (op0),
10696 GET_MODE (SUBREG_REG (op0)))
10697 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10699 tem = gen_lowpart (GET_MODE (SUBREG_REG (op0)), op1);
10701 if ((nonzero_bits (tem, GET_MODE (SUBREG_REG (op0)))
10702 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10703 op0 = SUBREG_REG (op0), op1 = tem;
10707 /* We now do the opposite procedure: Some machines don't have compare
10708 insns in all modes. If OP0's mode is an integer mode smaller than a
10709 word and we can't do a compare in that mode, see if there is a larger
10710 mode for which we can do the compare. There are a number of cases in
10711 which we can use the wider mode. */
10713 mode = GET_MODE (op0);
10714 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10715 && GET_MODE_SIZE (mode) < UNITS_PER_WORD
10716 && ! have_insn_for (COMPARE, mode))
10717 for (tmode = GET_MODE_WIDER_MODE (mode);
10718 (tmode != VOIDmode
10719 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT);
10720 tmode = GET_MODE_WIDER_MODE (tmode))
10721 if (have_insn_for (COMPARE, tmode))
10723 int zero_extended;
10725 /* If the only nonzero bits in OP0 and OP1 are those in the
10726 narrower mode and this is an equality or unsigned comparison,
10727 we can use the wider mode. Similarly for sign-extended
10728 values, in which case it is true for all comparisons. */
10729 zero_extended = ((code == EQ || code == NE
10730 || code == GEU || code == GTU
10731 || code == LEU || code == LTU)
10732 && (nonzero_bits (op0, tmode)
10733 & ~GET_MODE_MASK (mode)) == 0
10734 && ((GET_CODE (op1) == CONST_INT
10735 || (nonzero_bits (op1, tmode)
10736 & ~GET_MODE_MASK (mode)) == 0)));
10738 if (zero_extended
10739 || ((num_sign_bit_copies (op0, tmode)
10740 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10741 - GET_MODE_BITSIZE (mode)))
10742 && (num_sign_bit_copies (op1, tmode)
10743 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10744 - GET_MODE_BITSIZE (mode)))))
10746 /* If OP0 is an AND and we don't have an AND in MODE either,
10747 make a new AND in the proper mode. */
10748 if (GET_CODE (op0) == AND
10749 && !have_insn_for (AND, mode))
10750 op0 = simplify_gen_binary (AND, tmode,
10751 gen_lowpart (tmode,
10752 XEXP (op0, 0)),
10753 gen_lowpart (tmode,
10754 XEXP (op0, 1)));
10756 op0 = gen_lowpart (tmode, op0);
10757 if (zero_extended && GET_CODE (op1) == CONST_INT)
10758 op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (mode));
10759 op1 = gen_lowpart (tmode, op1);
10760 break;
10763 /* If this is a test for negative, we can make an explicit
10764 test of the sign bit. */
10766 if (op1 == const0_rtx && (code == LT || code == GE)
10767 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10769 op0 = simplify_gen_binary (AND, tmode,
10770 gen_lowpart (tmode, op0),
10771 GEN_INT ((HOST_WIDE_INT) 1
10772 << (GET_MODE_BITSIZE (mode)
10773 - 1)));
10774 code = (code == LT) ? NE : EQ;
10775 break;
10779 #ifdef CANONICALIZE_COMPARISON
10780 /* If this machine only supports a subset of valid comparisons, see if we
10781 can convert an unsupported one into a supported one. */
10782 CANONICALIZE_COMPARISON (code, op0, op1);
10783 #endif
10785 *pop0 = op0;
10786 *pop1 = op1;
10788 return code;
10791 /* Utility function for record_value_for_reg. Count number of
10792 rtxs in X. */
10793 static int
10794 count_rtxs (rtx x)
10796 enum rtx_code code = GET_CODE (x);
10797 const char *fmt;
10798 int i, ret = 1;
10800 if (GET_RTX_CLASS (code) == '2'
10801 || GET_RTX_CLASS (code) == 'c')
10803 rtx x0 = XEXP (x, 0);
10804 rtx x1 = XEXP (x, 1);
10806 if (x0 == x1)
10807 return 1 + 2 * count_rtxs (x0);
10809 if ((GET_RTX_CLASS (GET_CODE (x1)) == '2'
10810 || GET_RTX_CLASS (GET_CODE (x1)) == 'c')
10811 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10812 return 2 + 2 * count_rtxs (x0)
10813 + count_rtxs (x == XEXP (x1, 0)
10814 ? XEXP (x1, 1) : XEXP (x1, 0));
10816 if ((GET_RTX_CLASS (GET_CODE (x0)) == '2'
10817 || GET_RTX_CLASS (GET_CODE (x0)) == 'c')
10818 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10819 return 2 + 2 * count_rtxs (x1)
10820 + count_rtxs (x == XEXP (x0, 0)
10821 ? XEXP (x0, 1) : XEXP (x0, 0));
10824 fmt = GET_RTX_FORMAT (code);
10825 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10826 if (fmt[i] == 'e')
10827 ret += count_rtxs (XEXP (x, i));
10829 return ret;
10832 /* Utility function for following routine. Called when X is part of a value
10833 being stored into last_set_value. Sets last_set_table_tick
10834 for each register mentioned. Similar to mention_regs in cse.c */
10836 static void
10837 update_table_tick (rtx x)
10839 enum rtx_code code = GET_CODE (x);
10840 const char *fmt = GET_RTX_FORMAT (code);
10841 int i;
10843 if (code == REG)
10845 unsigned int regno = REGNO (x);
10846 unsigned int endregno
10847 = regno + (regno < FIRST_PSEUDO_REGISTER
10848 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
10849 unsigned int r;
10851 for (r = regno; r < endregno; r++)
10852 reg_stat[r].last_set_table_tick = label_tick;
10854 return;
10857 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10858 /* Note that we can't have an "E" in values stored; see
10859 get_last_value_validate. */
10860 if (fmt[i] == 'e')
10862 /* Check for identical subexpressions. If x contains
10863 identical subexpression we only have to traverse one of
10864 them. */
10865 if (i == 0 && ARITHMETIC_P (x))
10867 /* Note that at this point x1 has already been
10868 processed. */
10869 rtx x0 = XEXP (x, 0);
10870 rtx x1 = XEXP (x, 1);
10872 /* If x0 and x1 are identical then there is no need to
10873 process x0. */
10874 if (x0 == x1)
10875 break;
10877 /* If x0 is identical to a subexpression of x1 then while
10878 processing x1, x0 has already been processed. Thus we
10879 are done with x. */
10880 if (ARITHMETIC_P (x1)
10881 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10882 break;
10884 /* If x1 is identical to a subexpression of x0 then we
10885 still have to process the rest of x0. */
10886 if (ARITHMETIC_P (x0)
10887 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10889 update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0));
10890 break;
10894 update_table_tick (XEXP (x, i));
10898 /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we
10899 are saying that the register is clobbered and we no longer know its
10900 value. If INSN is zero, don't update reg_stat[].last_set; this is
10901 only permitted with VALUE also zero and is used to invalidate the
10902 register. */
10904 static void
10905 record_value_for_reg (rtx reg, rtx insn, rtx value)
10907 unsigned int regno = REGNO (reg);
10908 unsigned int endregno
10909 = regno + (regno < FIRST_PSEUDO_REGISTER
10910 ? hard_regno_nregs[regno][GET_MODE (reg)] : 1);
10911 unsigned int i;
10913 /* If VALUE contains REG and we have a previous value for REG, substitute
10914 the previous value. */
10915 if (value && insn && reg_overlap_mentioned_p (reg, value))
10917 rtx tem;
10919 /* Set things up so get_last_value is allowed to see anything set up to
10920 our insn. */
10921 subst_low_cuid = INSN_CUID (insn);
10922 tem = get_last_value (reg);
10924 /* If TEM is simply a binary operation with two CLOBBERs as operands,
10925 it isn't going to be useful and will take a lot of time to process,
10926 so just use the CLOBBER. */
10928 if (tem)
10930 if (ARITHMETIC_P (tem)
10931 && GET_CODE (XEXP (tem, 0)) == CLOBBER
10932 && GET_CODE (XEXP (tem, 1)) == CLOBBER)
10933 tem = XEXP (tem, 0);
10934 else if (count_occurrences (value, reg, 1) >= 2)
10936 /* If there are two or more occurrences of REG in VALUE,
10937 prevent the value from growing too much. */
10938 if (count_rtxs (tem) > MAX_LAST_VALUE_RTL)
10939 tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx);
10942 value = replace_rtx (copy_rtx (value), reg, tem);
10946 /* For each register modified, show we don't know its value, that
10947 we don't know about its bitwise content, that its value has been
10948 updated, and that we don't know the location of the death of the
10949 register. */
10950 for (i = regno; i < endregno; i++)
10952 if (insn)
10953 reg_stat[i].last_set = insn;
10955 reg_stat[i].last_set_value = 0;
10956 reg_stat[i].last_set_mode = 0;
10957 reg_stat[i].last_set_nonzero_bits = 0;
10958 reg_stat[i].last_set_sign_bit_copies = 0;
10959 reg_stat[i].last_death = 0;
10962 /* Mark registers that are being referenced in this value. */
10963 if (value)
10964 update_table_tick (value);
10966 /* Now update the status of each register being set.
10967 If someone is using this register in this block, set this register
10968 to invalid since we will get confused between the two lives in this
10969 basic block. This makes using this register always invalid. In cse, we
10970 scan the table to invalidate all entries using this register, but this
10971 is too much work for us. */
10973 for (i = regno; i < endregno; i++)
10975 reg_stat[i].last_set_label = label_tick;
10976 if (value && reg_stat[i].last_set_table_tick == label_tick)
10977 reg_stat[i].last_set_invalid = 1;
10978 else
10979 reg_stat[i].last_set_invalid = 0;
10982 /* The value being assigned might refer to X (like in "x++;"). In that
10983 case, we must replace it with (clobber (const_int 0)) to prevent
10984 infinite loops. */
10985 if (value && ! get_last_value_validate (&value, insn,
10986 reg_stat[regno].last_set_label, 0))
10988 value = copy_rtx (value);
10989 if (! get_last_value_validate (&value, insn,
10990 reg_stat[regno].last_set_label, 1))
10991 value = 0;
10994 /* For the main register being modified, update the value, the mode, the
10995 nonzero bits, and the number of sign bit copies. */
10997 reg_stat[regno].last_set_value = value;
10999 if (value)
11001 enum machine_mode mode = GET_MODE (reg);
11002 subst_low_cuid = INSN_CUID (insn);
11003 reg_stat[regno].last_set_mode = mode;
11004 if (GET_MODE_CLASS (mode) == MODE_INT
11005 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
11006 mode = nonzero_bits_mode;
11007 reg_stat[regno].last_set_nonzero_bits = nonzero_bits (value, mode);
11008 reg_stat[regno].last_set_sign_bit_copies
11009 = num_sign_bit_copies (value, GET_MODE (reg));
11013 /* Called via note_stores from record_dead_and_set_regs to handle one
11014 SET or CLOBBER in an insn. DATA is the instruction in which the
11015 set is occurring. */
11017 static void
11018 record_dead_and_set_regs_1 (rtx dest, rtx setter, void *data)
11020 rtx record_dead_insn = (rtx) data;
11022 if (GET_CODE (dest) == SUBREG)
11023 dest = SUBREG_REG (dest);
11025 if (REG_P (dest))
11027 /* If we are setting the whole register, we know its value. Otherwise
11028 show that we don't know the value. We can handle SUBREG in
11029 some cases. */
11030 if (GET_CODE (setter) == SET && dest == SET_DEST (setter))
11031 record_value_for_reg (dest, record_dead_insn, SET_SRC (setter));
11032 else if (GET_CODE (setter) == SET
11033 && GET_CODE (SET_DEST (setter)) == SUBREG
11034 && SUBREG_REG (SET_DEST (setter)) == dest
11035 && GET_MODE_BITSIZE (GET_MODE (dest)) <= BITS_PER_WORD
11036 && subreg_lowpart_p (SET_DEST (setter)))
11037 record_value_for_reg (dest, record_dead_insn,
11038 gen_lowpart (GET_MODE (dest),
11039 SET_SRC (setter)));
11040 else
11041 record_value_for_reg (dest, record_dead_insn, NULL_RTX);
11043 else if (MEM_P (dest)
11044 /* Ignore pushes, they clobber nothing. */
11045 && ! push_operand (dest, GET_MODE (dest)))
11046 mem_last_set = INSN_CUID (record_dead_insn);
11049 /* Update the records of when each REG was most recently set or killed
11050 for the things done by INSN. This is the last thing done in processing
11051 INSN in the combiner loop.
11053 We update reg_stat[], in particular fields last_set, last_set_value,
11054 last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies,
11055 last_death, and also the similar information mem_last_set (which insn
11056 most recently modified memory) and last_call_cuid (which insn was the
11057 most recent subroutine call). */
11059 static void
11060 record_dead_and_set_regs (rtx insn)
11062 rtx link;
11063 unsigned int i;
11065 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
11067 if (REG_NOTE_KIND (link) == REG_DEAD
11068 && REG_P (XEXP (link, 0)))
11070 unsigned int regno = REGNO (XEXP (link, 0));
11071 unsigned int endregno
11072 = regno + (regno < FIRST_PSEUDO_REGISTER
11073 ? hard_regno_nregs[regno][GET_MODE (XEXP (link, 0))]
11074 : 1);
11076 for (i = regno; i < endregno; i++)
11077 reg_stat[i].last_death = insn;
11079 else if (REG_NOTE_KIND (link) == REG_INC)
11080 record_value_for_reg (XEXP (link, 0), insn, NULL_RTX);
11083 if (CALL_P (insn))
11085 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
11086 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
11088 reg_stat[i].last_set_value = 0;
11089 reg_stat[i].last_set_mode = 0;
11090 reg_stat[i].last_set_nonzero_bits = 0;
11091 reg_stat[i].last_set_sign_bit_copies = 0;
11092 reg_stat[i].last_death = 0;
11095 last_call_cuid = mem_last_set = INSN_CUID (insn);
11097 /* Don't bother recording what this insn does. It might set the
11098 return value register, but we can't combine into a call
11099 pattern anyway, so there's no point trying (and it may cause
11100 a crash, if e.g. we wind up asking for last_set_value of a
11101 SUBREG of the return value register). */
11102 return;
11105 note_stores (PATTERN (insn), record_dead_and_set_regs_1, insn);
11108 /* If a SUBREG has the promoted bit set, it is in fact a property of the
11109 register present in the SUBREG, so for each such SUBREG go back and
11110 adjust nonzero and sign bit information of the registers that are
11111 known to have some zero/sign bits set.
11113 This is needed because when combine blows the SUBREGs away, the
11114 information on zero/sign bits is lost and further combines can be
11115 missed because of that. */
11117 static void
11118 record_promoted_value (rtx insn, rtx subreg)
11120 rtx links, set;
11121 unsigned int regno = REGNO (SUBREG_REG (subreg));
11122 enum machine_mode mode = GET_MODE (subreg);
11124 if (GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
11125 return;
11127 for (links = LOG_LINKS (insn); links;)
11129 insn = XEXP (links, 0);
11130 set = single_set (insn);
11132 if (! set || !REG_P (SET_DEST (set))
11133 || REGNO (SET_DEST (set)) != regno
11134 || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg)))
11136 links = XEXP (links, 1);
11137 continue;
11140 if (reg_stat[regno].last_set == insn)
11142 if (SUBREG_PROMOTED_UNSIGNED_P (subreg) > 0)
11143 reg_stat[regno].last_set_nonzero_bits &= GET_MODE_MASK (mode);
11146 if (REG_P (SET_SRC (set)))
11148 regno = REGNO (SET_SRC (set));
11149 links = LOG_LINKS (insn);
11151 else
11152 break;
11156 /* Scan X for promoted SUBREGs. For each one found,
11157 note what it implies to the registers used in it. */
11159 static void
11160 check_promoted_subreg (rtx insn, rtx x)
11162 if (GET_CODE (x) == SUBREG && SUBREG_PROMOTED_VAR_P (x)
11163 && REG_P (SUBREG_REG (x)))
11164 record_promoted_value (insn, x);
11165 else
11167 const char *format = GET_RTX_FORMAT (GET_CODE (x));
11168 int i, j;
11170 for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++)
11171 switch (format[i])
11173 case 'e':
11174 check_promoted_subreg (insn, XEXP (x, i));
11175 break;
11176 case 'V':
11177 case 'E':
11178 if (XVEC (x, i) != 0)
11179 for (j = 0; j < XVECLEN (x, i); j++)
11180 check_promoted_subreg (insn, XVECEXP (x, i, j));
11181 break;
11186 /* Utility routine for the following function. Verify that all the registers
11187 mentioned in *LOC are valid when *LOC was part of a value set when
11188 label_tick == TICK. Return 0 if some are not.
11190 If REPLACE is nonzero, replace the invalid reference with
11191 (clobber (const_int 0)) and return 1. This replacement is useful because
11192 we often can get useful information about the form of a value (e.g., if
11193 it was produced by a shift that always produces -1 or 0) even though
11194 we don't know exactly what registers it was produced from. */
11196 static int
11197 get_last_value_validate (rtx *loc, rtx insn, int tick, int replace)
11199 rtx x = *loc;
11200 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
11201 int len = GET_RTX_LENGTH (GET_CODE (x));
11202 int i;
11204 if (REG_P (x))
11206 unsigned int regno = REGNO (x);
11207 unsigned int endregno
11208 = regno + (regno < FIRST_PSEUDO_REGISTER
11209 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11210 unsigned int j;
11212 for (j = regno; j < endregno; j++)
11213 if (reg_stat[j].last_set_invalid
11214 /* If this is a pseudo-register that was only set once and not
11215 live at the beginning of the function, it is always valid. */
11216 || (! (regno >= FIRST_PSEUDO_REGISTER
11217 && REG_N_SETS (regno) == 1
11218 && (! REGNO_REG_SET_P
11219 (ENTRY_BLOCK_PTR->next_bb->global_live_at_start, regno)))
11220 && reg_stat[j].last_set_label > tick))
11222 if (replace)
11223 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11224 return replace;
11227 return 1;
11229 /* If this is a memory reference, make sure that there were
11230 no stores after it that might have clobbered the value. We don't
11231 have alias info, so we assume any store invalidates it. */
11232 else if (MEM_P (x) && !MEM_READONLY_P (x)
11233 && INSN_CUID (insn) <= mem_last_set)
11235 if (replace)
11236 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11237 return replace;
11240 for (i = 0; i < len; i++)
11242 if (fmt[i] == 'e')
11244 /* Check for identical subexpressions. If x contains
11245 identical subexpression we only have to traverse one of
11246 them. */
11247 if (i == 1 && ARITHMETIC_P (x))
11249 /* Note that at this point x0 has already been checked
11250 and found valid. */
11251 rtx x0 = XEXP (x, 0);
11252 rtx x1 = XEXP (x, 1);
11254 /* If x0 and x1 are identical then x is also valid. */
11255 if (x0 == x1)
11256 return 1;
11258 /* If x1 is identical to a subexpression of x0 then
11259 while checking x0, x1 has already been checked. Thus
11260 it is valid and so as x. */
11261 if (ARITHMETIC_P (x0)
11262 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
11263 return 1;
11265 /* If x0 is identical to a subexpression of x1 then x is
11266 valid iff the rest of x1 is valid. */
11267 if (ARITHMETIC_P (x1)
11268 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
11269 return
11270 get_last_value_validate (&XEXP (x1,
11271 x0 == XEXP (x1, 0) ? 1 : 0),
11272 insn, tick, replace);
11275 if (get_last_value_validate (&XEXP (x, i), insn, tick,
11276 replace) == 0)
11277 return 0;
11279 /* Don't bother with these. They shouldn't occur anyway. */
11280 else if (fmt[i] == 'E')
11281 return 0;
11284 /* If we haven't found a reason for it to be invalid, it is valid. */
11285 return 1;
11288 /* Get the last value assigned to X, if known. Some registers
11289 in the value may be replaced with (clobber (const_int 0)) if their value
11290 is known longer known reliably. */
11292 static rtx
11293 get_last_value (rtx x)
11295 unsigned int regno;
11296 rtx value;
11298 /* If this is a non-paradoxical SUBREG, get the value of its operand and
11299 then convert it to the desired mode. If this is a paradoxical SUBREG,
11300 we cannot predict what values the "extra" bits might have. */
11301 if (GET_CODE (x) == SUBREG
11302 && subreg_lowpart_p (x)
11303 && (GET_MODE_SIZE (GET_MODE (x))
11304 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
11305 && (value = get_last_value (SUBREG_REG (x))) != 0)
11306 return gen_lowpart (GET_MODE (x), value);
11308 if (!REG_P (x))
11309 return 0;
11311 regno = REGNO (x);
11312 value = reg_stat[regno].last_set_value;
11314 /* If we don't have a value, or if it isn't for this basic block and
11315 it's either a hard register, set more than once, or it's a live
11316 at the beginning of the function, return 0.
11318 Because if it's not live at the beginning of the function then the reg
11319 is always set before being used (is never used without being set).
11320 And, if it's set only once, and it's always set before use, then all
11321 uses must have the same last value, even if it's not from this basic
11322 block. */
11324 if (value == 0
11325 || (reg_stat[regno].last_set_label != label_tick
11326 && (regno < FIRST_PSEUDO_REGISTER
11327 || REG_N_SETS (regno) != 1
11328 || (REGNO_REG_SET_P
11329 (ENTRY_BLOCK_PTR->next_bb->global_live_at_start, regno)))))
11330 return 0;
11332 /* If the value was set in a later insn than the ones we are processing,
11333 we can't use it even if the register was only set once. */
11334 if (INSN_CUID (reg_stat[regno].last_set) >= subst_low_cuid)
11335 return 0;
11337 /* If the value has all its registers valid, return it. */
11338 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11339 reg_stat[regno].last_set_label, 0))
11340 return value;
11342 /* Otherwise, make a copy and replace any invalid register with
11343 (clobber (const_int 0)). If that fails for some reason, return 0. */
11345 value = copy_rtx (value);
11346 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11347 reg_stat[regno].last_set_label, 1))
11348 return value;
11350 return 0;
11353 /* Return nonzero if expression X refers to a REG or to memory
11354 that is set in an instruction more recent than FROM_CUID. */
11356 static int
11357 use_crosses_set_p (rtx x, int from_cuid)
11359 const char *fmt;
11360 int i;
11361 enum rtx_code code = GET_CODE (x);
11363 if (code == REG)
11365 unsigned int regno = REGNO (x);
11366 unsigned endreg = regno + (regno < FIRST_PSEUDO_REGISTER
11367 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11369 #ifdef PUSH_ROUNDING
11370 /* Don't allow uses of the stack pointer to be moved,
11371 because we don't know whether the move crosses a push insn. */
11372 if (regno == STACK_POINTER_REGNUM && PUSH_ARGS)
11373 return 1;
11374 #endif
11375 for (; regno < endreg; regno++)
11376 if (reg_stat[regno].last_set
11377 && INSN_CUID (reg_stat[regno].last_set) > from_cuid)
11378 return 1;
11379 return 0;
11382 if (code == MEM && mem_last_set > from_cuid)
11383 return 1;
11385 fmt = GET_RTX_FORMAT (code);
11387 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11389 if (fmt[i] == 'E')
11391 int j;
11392 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11393 if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid))
11394 return 1;
11396 else if (fmt[i] == 'e'
11397 && use_crosses_set_p (XEXP (x, i), from_cuid))
11398 return 1;
11400 return 0;
11403 /* Define three variables used for communication between the following
11404 routines. */
11406 static unsigned int reg_dead_regno, reg_dead_endregno;
11407 static int reg_dead_flag;
11409 /* Function called via note_stores from reg_dead_at_p.
11411 If DEST is within [reg_dead_regno, reg_dead_endregno), set
11412 reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */
11414 static void
11415 reg_dead_at_p_1 (rtx dest, rtx x, void *data ATTRIBUTE_UNUSED)
11417 unsigned int regno, endregno;
11419 if (!REG_P (dest))
11420 return;
11422 regno = REGNO (dest);
11423 endregno = regno + (regno < FIRST_PSEUDO_REGISTER
11424 ? hard_regno_nregs[regno][GET_MODE (dest)] : 1);
11426 if (reg_dead_endregno > regno && reg_dead_regno < endregno)
11427 reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1;
11430 /* Return nonzero if REG is known to be dead at INSN.
11432 We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER
11433 referencing REG, it is dead. If we hit a SET referencing REG, it is
11434 live. Otherwise, see if it is live or dead at the start of the basic
11435 block we are in. Hard regs marked as being live in NEWPAT_USED_REGS
11436 must be assumed to be always live. */
11438 static int
11439 reg_dead_at_p (rtx reg, rtx insn)
11441 basic_block block;
11442 unsigned int i;
11444 /* Set variables for reg_dead_at_p_1. */
11445 reg_dead_regno = REGNO (reg);
11446 reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER
11447 ? hard_regno_nregs[reg_dead_regno]
11448 [GET_MODE (reg)]
11449 : 1);
11451 reg_dead_flag = 0;
11453 /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers
11454 we allow the machine description to decide whether use-and-clobber
11455 patterns are OK. */
11456 if (reg_dead_regno < FIRST_PSEUDO_REGISTER)
11458 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11459 if (!fixed_regs[i] && TEST_HARD_REG_BIT (newpat_used_regs, i))
11460 return 0;
11463 /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or
11464 beginning of function. */
11465 for (; insn && !LABEL_P (insn) && !BARRIER_P (insn);
11466 insn = prev_nonnote_insn (insn))
11468 note_stores (PATTERN (insn), reg_dead_at_p_1, NULL);
11469 if (reg_dead_flag)
11470 return reg_dead_flag == 1 ? 1 : 0;
11472 if (find_regno_note (insn, REG_DEAD, reg_dead_regno))
11473 return 1;
11476 /* Get the basic block that we were in. */
11477 if (insn == 0)
11478 block = ENTRY_BLOCK_PTR->next_bb;
11479 else
11481 FOR_EACH_BB (block)
11482 if (insn == BB_HEAD (block))
11483 break;
11485 if (block == EXIT_BLOCK_PTR)
11486 return 0;
11489 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11490 if (REGNO_REG_SET_P (block->global_live_at_start, i))
11491 return 0;
11493 return 1;
11496 /* Note hard registers in X that are used. This code is similar to
11497 that in flow.c, but much simpler since we don't care about pseudos. */
11499 static void
11500 mark_used_regs_combine (rtx x)
11502 RTX_CODE code = GET_CODE (x);
11503 unsigned int regno;
11504 int i;
11506 switch (code)
11508 case LABEL_REF:
11509 case SYMBOL_REF:
11510 case CONST_INT:
11511 case CONST:
11512 case CONST_DOUBLE:
11513 case CONST_VECTOR:
11514 case PC:
11515 case ADDR_VEC:
11516 case ADDR_DIFF_VEC:
11517 case ASM_INPUT:
11518 #ifdef HAVE_cc0
11519 /* CC0 must die in the insn after it is set, so we don't need to take
11520 special note of it here. */
11521 case CC0:
11522 #endif
11523 return;
11525 case CLOBBER:
11526 /* If we are clobbering a MEM, mark any hard registers inside the
11527 address as used. */
11528 if (MEM_P (XEXP (x, 0)))
11529 mark_used_regs_combine (XEXP (XEXP (x, 0), 0));
11530 return;
11532 case REG:
11533 regno = REGNO (x);
11534 /* A hard reg in a wide mode may really be multiple registers.
11535 If so, mark all of them just like the first. */
11536 if (regno < FIRST_PSEUDO_REGISTER)
11538 unsigned int endregno, r;
11540 /* None of this applies to the stack, frame or arg pointers. */
11541 if (regno == STACK_POINTER_REGNUM
11542 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
11543 || regno == HARD_FRAME_POINTER_REGNUM
11544 #endif
11545 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
11546 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
11547 #endif
11548 || regno == FRAME_POINTER_REGNUM)
11549 return;
11551 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11552 for (r = regno; r < endregno; r++)
11553 SET_HARD_REG_BIT (newpat_used_regs, r);
11555 return;
11557 case SET:
11559 /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in
11560 the address. */
11561 rtx testreg = SET_DEST (x);
11563 while (GET_CODE (testreg) == SUBREG
11564 || GET_CODE (testreg) == ZERO_EXTRACT
11565 || GET_CODE (testreg) == STRICT_LOW_PART)
11566 testreg = XEXP (testreg, 0);
11568 if (MEM_P (testreg))
11569 mark_used_regs_combine (XEXP (testreg, 0));
11571 mark_used_regs_combine (SET_SRC (x));
11573 return;
11575 default:
11576 break;
11579 /* Recursively scan the operands of this expression. */
11582 const char *fmt = GET_RTX_FORMAT (code);
11584 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11586 if (fmt[i] == 'e')
11587 mark_used_regs_combine (XEXP (x, i));
11588 else if (fmt[i] == 'E')
11590 int j;
11592 for (j = 0; j < XVECLEN (x, i); j++)
11593 mark_used_regs_combine (XVECEXP (x, i, j));
11599 /* Remove register number REGNO from the dead registers list of INSN.
11601 Return the note used to record the death, if there was one. */
11604 remove_death (unsigned int regno, rtx insn)
11606 rtx note = find_regno_note (insn, REG_DEAD, regno);
11608 if (note)
11610 REG_N_DEATHS (regno)--;
11611 remove_note (insn, note);
11614 return note;
11617 /* For each register (hardware or pseudo) used within expression X, if its
11618 death is in an instruction with cuid between FROM_CUID (inclusive) and
11619 TO_INSN (exclusive), put a REG_DEAD note for that register in the
11620 list headed by PNOTES.
11622 That said, don't move registers killed by maybe_kill_insn.
11624 This is done when X is being merged by combination into TO_INSN. These
11625 notes will then be distributed as needed. */
11627 static void
11628 move_deaths (rtx x, rtx maybe_kill_insn, int from_cuid, rtx to_insn,
11629 rtx *pnotes)
11631 const char *fmt;
11632 int len, i;
11633 enum rtx_code code = GET_CODE (x);
11635 if (code == REG)
11637 unsigned int regno = REGNO (x);
11638 rtx where_dead = reg_stat[regno].last_death;
11639 rtx before_dead, after_dead;
11641 /* Don't move the register if it gets killed in between from and to. */
11642 if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn)
11643 && ! reg_referenced_p (x, maybe_kill_insn))
11644 return;
11646 /* WHERE_DEAD could be a USE insn made by combine, so first we
11647 make sure that we have insns with valid INSN_CUID values. */
11648 before_dead = where_dead;
11649 while (before_dead && INSN_UID (before_dead) > max_uid_cuid)
11650 before_dead = PREV_INSN (before_dead);
11652 after_dead = where_dead;
11653 while (after_dead && INSN_UID (after_dead) > max_uid_cuid)
11654 after_dead = NEXT_INSN (after_dead);
11656 if (before_dead && after_dead
11657 && INSN_CUID (before_dead) >= from_cuid
11658 && (INSN_CUID (after_dead) < INSN_CUID (to_insn)
11659 || (where_dead != after_dead
11660 && INSN_CUID (after_dead) == INSN_CUID (to_insn))))
11662 rtx note = remove_death (regno, where_dead);
11664 /* It is possible for the call above to return 0. This can occur
11665 when last_death points to I2 or I1 that we combined with.
11666 In that case make a new note.
11668 We must also check for the case where X is a hard register
11669 and NOTE is a death note for a range of hard registers
11670 including X. In that case, we must put REG_DEAD notes for
11671 the remaining registers in place of NOTE. */
11673 if (note != 0 && regno < FIRST_PSEUDO_REGISTER
11674 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11675 > GET_MODE_SIZE (GET_MODE (x))))
11677 unsigned int deadregno = REGNO (XEXP (note, 0));
11678 unsigned int deadend
11679 = (deadregno + hard_regno_nregs[deadregno]
11680 [GET_MODE (XEXP (note, 0))]);
11681 unsigned int ourend
11682 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11683 unsigned int i;
11685 for (i = deadregno; i < deadend; i++)
11686 if (i < regno || i >= ourend)
11687 REG_NOTES (where_dead)
11688 = gen_rtx_EXPR_LIST (REG_DEAD,
11689 regno_reg_rtx[i],
11690 REG_NOTES (where_dead));
11693 /* If we didn't find any note, or if we found a REG_DEAD note that
11694 covers only part of the given reg, and we have a multi-reg hard
11695 register, then to be safe we must check for REG_DEAD notes
11696 for each register other than the first. They could have
11697 their own REG_DEAD notes lying around. */
11698 else if ((note == 0
11699 || (note != 0
11700 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11701 < GET_MODE_SIZE (GET_MODE (x)))))
11702 && regno < FIRST_PSEUDO_REGISTER
11703 && hard_regno_nregs[regno][GET_MODE (x)] > 1)
11705 unsigned int ourend
11706 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11707 unsigned int i, offset;
11708 rtx oldnotes = 0;
11710 if (note)
11711 offset = hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))];
11712 else
11713 offset = 1;
11715 for (i = regno + offset; i < ourend; i++)
11716 move_deaths (regno_reg_rtx[i],
11717 maybe_kill_insn, from_cuid, to_insn, &oldnotes);
11720 if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x))
11722 XEXP (note, 1) = *pnotes;
11723 *pnotes = note;
11725 else
11726 *pnotes = gen_rtx_EXPR_LIST (REG_DEAD, x, *pnotes);
11728 REG_N_DEATHS (regno)++;
11731 return;
11734 else if (GET_CODE (x) == SET)
11736 rtx dest = SET_DEST (x);
11738 move_deaths (SET_SRC (x), maybe_kill_insn, from_cuid, to_insn, pnotes);
11740 /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG
11741 that accesses one word of a multi-word item, some
11742 piece of everything register in the expression is used by
11743 this insn, so remove any old death. */
11744 /* ??? So why do we test for equality of the sizes? */
11746 if (GET_CODE (dest) == ZERO_EXTRACT
11747 || GET_CODE (dest) == STRICT_LOW_PART
11748 || (GET_CODE (dest) == SUBREG
11749 && (((GET_MODE_SIZE (GET_MODE (dest))
11750 + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
11751 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))
11752 + UNITS_PER_WORD - 1) / UNITS_PER_WORD))))
11754 move_deaths (dest, maybe_kill_insn, from_cuid, to_insn, pnotes);
11755 return;
11758 /* If this is some other SUBREG, we know it replaces the entire
11759 value, so use that as the destination. */
11760 if (GET_CODE (dest) == SUBREG)
11761 dest = SUBREG_REG (dest);
11763 /* If this is a MEM, adjust deaths of anything used in the address.
11764 For a REG (the only other possibility), the entire value is
11765 being replaced so the old value is not used in this insn. */
11767 if (MEM_P (dest))
11768 move_deaths (XEXP (dest, 0), maybe_kill_insn, from_cuid,
11769 to_insn, pnotes);
11770 return;
11773 else if (GET_CODE (x) == CLOBBER)
11774 return;
11776 len = GET_RTX_LENGTH (code);
11777 fmt = GET_RTX_FORMAT (code);
11779 for (i = 0; i < len; i++)
11781 if (fmt[i] == 'E')
11783 int j;
11784 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11785 move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_cuid,
11786 to_insn, pnotes);
11788 else if (fmt[i] == 'e')
11789 move_deaths (XEXP (x, i), maybe_kill_insn, from_cuid, to_insn, pnotes);
11793 /* Return 1 if X is the target of a bit-field assignment in BODY, the
11794 pattern of an insn. X must be a REG. */
11796 static int
11797 reg_bitfield_target_p (rtx x, rtx body)
11799 int i;
11801 if (GET_CODE (body) == SET)
11803 rtx dest = SET_DEST (body);
11804 rtx target;
11805 unsigned int regno, tregno, endregno, endtregno;
11807 if (GET_CODE (dest) == ZERO_EXTRACT)
11808 target = XEXP (dest, 0);
11809 else if (GET_CODE (dest) == STRICT_LOW_PART)
11810 target = SUBREG_REG (XEXP (dest, 0));
11811 else
11812 return 0;
11814 if (GET_CODE (target) == SUBREG)
11815 target = SUBREG_REG (target);
11817 if (!REG_P (target))
11818 return 0;
11820 tregno = REGNO (target), regno = REGNO (x);
11821 if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER)
11822 return target == x;
11824 endtregno = tregno + hard_regno_nregs[tregno][GET_MODE (target)];
11825 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11827 return endregno > tregno && regno < endtregno;
11830 else if (GET_CODE (body) == PARALLEL)
11831 for (i = XVECLEN (body, 0) - 1; i >= 0; i--)
11832 if (reg_bitfield_target_p (x, XVECEXP (body, 0, i)))
11833 return 1;
11835 return 0;
11838 /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them
11839 as appropriate. I3 and I2 are the insns resulting from the combination
11840 insns including FROM (I2 may be zero).
11842 Each note in the list is either ignored or placed on some insns, depending
11843 on the type of note. */
11845 static void
11846 distribute_notes (rtx notes, rtx from_insn, rtx i3, rtx i2)
11848 rtx note, next_note;
11849 rtx tem;
11851 for (note = notes; note; note = next_note)
11853 rtx place = 0, place2 = 0;
11855 /* If this NOTE references a pseudo register, ensure it references
11856 the latest copy of that register. */
11857 if (XEXP (note, 0) && REG_P (XEXP (note, 0))
11858 && REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER)
11859 XEXP (note, 0) = regno_reg_rtx[REGNO (XEXP (note, 0))];
11861 next_note = XEXP (note, 1);
11862 switch (REG_NOTE_KIND (note))
11864 case REG_BR_PROB:
11865 case REG_BR_PRED:
11866 /* Doesn't matter much where we put this, as long as it's somewhere.
11867 It is preferable to keep these notes on branches, which is most
11868 likely to be i3. */
11869 place = i3;
11870 break;
11872 case REG_VALUE_PROFILE:
11873 /* Just get rid of this note, as it is unused later anyway. */
11874 break;
11876 case REG_NON_LOCAL_GOTO:
11877 if (JUMP_P (i3))
11878 place = i3;
11879 else
11881 gcc_assert (i2 && JUMP_P (i2));
11882 place = i2;
11884 break;
11886 case REG_EH_REGION:
11887 /* These notes must remain with the call or trapping instruction. */
11888 if (CALL_P (i3))
11889 place = i3;
11890 else if (i2 && CALL_P (i2))
11891 place = i2;
11892 else
11894 gcc_assert (flag_non_call_exceptions);
11895 if (may_trap_p (i3))
11896 place = i3;
11897 else if (i2 && may_trap_p (i2))
11898 place = i2;
11899 /* ??? Otherwise assume we've combined things such that we
11900 can now prove that the instructions can't trap. Drop the
11901 note in this case. */
11903 break;
11905 case REG_NORETURN:
11906 case REG_SETJMP:
11907 /* These notes must remain with the call. It should not be
11908 possible for both I2 and I3 to be a call. */
11909 if (CALL_P (i3))
11910 place = i3;
11911 else
11913 gcc_assert (i2 && CALL_P (i2));
11914 place = i2;
11916 break;
11918 case REG_UNUSED:
11919 /* Any clobbers for i3 may still exist, and so we must process
11920 REG_UNUSED notes from that insn.
11922 Any clobbers from i2 or i1 can only exist if they were added by
11923 recog_for_combine. In that case, recog_for_combine created the
11924 necessary REG_UNUSED notes. Trying to keep any original
11925 REG_UNUSED notes from these insns can cause incorrect output
11926 if it is for the same register as the original i3 dest.
11927 In that case, we will notice that the register is set in i3,
11928 and then add a REG_UNUSED note for the destination of i3, which
11929 is wrong. However, it is possible to have REG_UNUSED notes from
11930 i2 or i1 for register which were both used and clobbered, so
11931 we keep notes from i2 or i1 if they will turn into REG_DEAD
11932 notes. */
11934 /* If this register is set or clobbered in I3, put the note there
11935 unless there is one already. */
11936 if (reg_set_p (XEXP (note, 0), PATTERN (i3)))
11938 if (from_insn != i3)
11939 break;
11941 if (! (REG_P (XEXP (note, 0))
11942 ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0)))
11943 : find_reg_note (i3, REG_UNUSED, XEXP (note, 0))))
11944 place = i3;
11946 /* Otherwise, if this register is used by I3, then this register
11947 now dies here, so we must put a REG_DEAD note here unless there
11948 is one already. */
11949 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))
11950 && ! (REG_P (XEXP (note, 0))
11951 ? find_regno_note (i3, REG_DEAD,
11952 REGNO (XEXP (note, 0)))
11953 : find_reg_note (i3, REG_DEAD, XEXP (note, 0))))
11955 PUT_REG_NOTE_KIND (note, REG_DEAD);
11956 place = i3;
11958 break;
11960 case REG_EQUAL:
11961 case REG_EQUIV:
11962 case REG_NOALIAS:
11963 /* These notes say something about results of an insn. We can
11964 only support them if they used to be on I3 in which case they
11965 remain on I3. Otherwise they are ignored.
11967 If the note refers to an expression that is not a constant, we
11968 must also ignore the note since we cannot tell whether the
11969 equivalence is still true. It might be possible to do
11970 slightly better than this (we only have a problem if I2DEST
11971 or I1DEST is present in the expression), but it doesn't
11972 seem worth the trouble. */
11974 if (from_insn == i3
11975 && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0))))
11976 place = i3;
11977 break;
11979 case REG_INC:
11980 case REG_NO_CONFLICT:
11981 /* These notes say something about how a register is used. They must
11982 be present on any use of the register in I2 or I3. */
11983 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3)))
11984 place = i3;
11986 if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2)))
11988 if (place)
11989 place2 = i2;
11990 else
11991 place = i2;
11993 break;
11995 case REG_LABEL:
11996 /* This can show up in several ways -- either directly in the
11997 pattern, or hidden off in the constant pool with (or without?)
11998 a REG_EQUAL note. */
11999 /* ??? Ignore the without-reg_equal-note problem for now. */
12000 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))
12001 || ((tem = find_reg_note (i3, REG_EQUAL, NULL_RTX))
12002 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
12003 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0)))
12004 place = i3;
12006 if (i2
12007 && (reg_mentioned_p (XEXP (note, 0), PATTERN (i2))
12008 || ((tem = find_reg_note (i2, REG_EQUAL, NULL_RTX))
12009 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
12010 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0))))
12012 if (place)
12013 place2 = i2;
12014 else
12015 place = i2;
12018 /* Don't attach REG_LABEL note to a JUMP_INSN. Add
12019 a JUMP_LABEL instead or decrement LABEL_NUSES. */
12020 if (place && JUMP_P (place))
12022 rtx label = JUMP_LABEL (place);
12024 if (!label)
12025 JUMP_LABEL (place) = XEXP (note, 0);
12026 else
12028 gcc_assert (label == XEXP (note, 0));
12029 if (LABEL_P (label))
12030 LABEL_NUSES (label)--;
12032 place = 0;
12034 if (place2 && JUMP_P (place2))
12036 rtx label = JUMP_LABEL (place2);
12038 if (!label)
12039 JUMP_LABEL (place2) = XEXP (note, 0);
12040 else
12042 gcc_assert (label == XEXP (note, 0));
12043 if (LABEL_P (label))
12044 LABEL_NUSES (label)--;
12046 place2 = 0;
12048 break;
12050 case REG_NONNEG:
12051 /* This note says something about the value of a register prior
12052 to the execution of an insn. It is too much trouble to see
12053 if the note is still correct in all situations. It is better
12054 to simply delete it. */
12055 break;
12057 case REG_RETVAL:
12058 /* If the insn previously containing this note still exists,
12059 put it back where it was. Otherwise move it to the previous
12060 insn. Adjust the corresponding REG_LIBCALL note. */
12061 if (!NOTE_P (from_insn))
12062 place = from_insn;
12063 else
12065 tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX);
12066 place = prev_real_insn (from_insn);
12067 if (tem && place)
12068 XEXP (tem, 0) = place;
12069 /* If we're deleting the last remaining instruction of a
12070 libcall sequence, don't add the notes. */
12071 else if (XEXP (note, 0) == from_insn)
12072 tem = place = 0;
12073 /* Don't add the dangling REG_RETVAL note. */
12074 else if (! tem)
12075 place = 0;
12077 break;
12079 case REG_LIBCALL:
12080 /* This is handled similarly to REG_RETVAL. */
12081 if (!NOTE_P (from_insn))
12082 place = from_insn;
12083 else
12085 tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX);
12086 place = next_real_insn (from_insn);
12087 if (tem && place)
12088 XEXP (tem, 0) = place;
12089 /* If we're deleting the last remaining instruction of a
12090 libcall sequence, don't add the notes. */
12091 else if (XEXP (note, 0) == from_insn)
12092 tem = place = 0;
12093 /* Don't add the dangling REG_LIBCALL note. */
12094 else if (! tem)
12095 place = 0;
12097 break;
12099 case REG_DEAD:
12100 /* If the register is used as an input in I3, it dies there.
12101 Similarly for I2, if it is nonzero and adjacent to I3.
12103 If the register is not used as an input in either I3 or I2
12104 and it is not one of the registers we were supposed to eliminate,
12105 there are two possibilities. We might have a non-adjacent I2
12106 or we might have somehow eliminated an additional register
12107 from a computation. For example, we might have had A & B where
12108 we discover that B will always be zero. In this case we will
12109 eliminate the reference to A.
12111 In both cases, we must search to see if we can find a previous
12112 use of A and put the death note there. */
12114 if (from_insn
12115 && CALL_P (from_insn)
12116 && find_reg_fusage (from_insn, USE, XEXP (note, 0)))
12117 place = from_insn;
12118 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)))
12119 place = i3;
12120 else if (i2 != 0 && next_nonnote_insn (i2) == i3
12121 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12122 place = i2;
12124 if (place == 0)
12126 basic_block bb = this_basic_block;
12128 for (tem = PREV_INSN (i3); place == 0; tem = PREV_INSN (tem))
12130 if (! INSN_P (tem))
12132 if (tem == BB_HEAD (bb))
12133 break;
12134 continue;
12137 /* If the register is being set at TEM, see if that is all
12138 TEM is doing. If so, delete TEM. Otherwise, make this
12139 into a REG_UNUSED note instead. Don't delete sets to
12140 global register vars. */
12141 if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER
12142 || !global_regs[REGNO (XEXP (note, 0))])
12143 && reg_set_p (XEXP (note, 0), PATTERN (tem)))
12145 rtx set = single_set (tem);
12146 rtx inner_dest = 0;
12147 #ifdef HAVE_cc0
12148 rtx cc0_setter = NULL_RTX;
12149 #endif
12151 if (set != 0)
12152 for (inner_dest = SET_DEST (set);
12153 (GET_CODE (inner_dest) == STRICT_LOW_PART
12154 || GET_CODE (inner_dest) == SUBREG
12155 || GET_CODE (inner_dest) == ZERO_EXTRACT);
12156 inner_dest = XEXP (inner_dest, 0))
12159 /* Verify that it was the set, and not a clobber that
12160 modified the register.
12162 CC0 targets must be careful to maintain setter/user
12163 pairs. If we cannot delete the setter due to side
12164 effects, mark the user with an UNUSED note instead
12165 of deleting it. */
12167 if (set != 0 && ! side_effects_p (SET_SRC (set))
12168 && rtx_equal_p (XEXP (note, 0), inner_dest)
12169 #ifdef HAVE_cc0
12170 && (! reg_mentioned_p (cc0_rtx, SET_SRC (set))
12171 || ((cc0_setter = prev_cc0_setter (tem)) != NULL
12172 && sets_cc0_p (PATTERN (cc0_setter)) > 0))
12173 #endif
12176 /* Move the notes and links of TEM elsewhere.
12177 This might delete other dead insns recursively.
12178 First set the pattern to something that won't use
12179 any register. */
12180 rtx old_notes = REG_NOTES (tem);
12182 PATTERN (tem) = pc_rtx;
12183 REG_NOTES (tem) = NULL;
12185 distribute_notes (old_notes, tem, tem, NULL_RTX);
12186 distribute_links (LOG_LINKS (tem));
12188 SET_INSN_DELETED (tem);
12190 #ifdef HAVE_cc0
12191 /* Delete the setter too. */
12192 if (cc0_setter)
12194 PATTERN (cc0_setter) = pc_rtx;
12195 old_notes = REG_NOTES (cc0_setter);
12196 REG_NOTES (cc0_setter) = NULL;
12198 distribute_notes (old_notes, cc0_setter,
12199 cc0_setter, NULL_RTX);
12200 distribute_links (LOG_LINKS (cc0_setter));
12202 SET_INSN_DELETED (cc0_setter);
12204 #endif
12206 else
12208 PUT_REG_NOTE_KIND (note, REG_UNUSED);
12210 /* If there isn't already a REG_UNUSED note, put one
12211 here. Do not place a REG_DEAD note, even if
12212 the register is also used here; that would not
12213 match the algorithm used in lifetime analysis
12214 and can cause the consistency check in the
12215 scheduler to fail. */
12216 if (! find_regno_note (tem, REG_UNUSED,
12217 REGNO (XEXP (note, 0))))
12218 place = tem;
12219 break;
12222 else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))
12223 || (CALL_P (tem)
12224 && find_reg_fusage (tem, USE, XEXP (note, 0))))
12226 place = tem;
12228 /* If we are doing a 3->2 combination, and we have a
12229 register which formerly died in i3 and was not used
12230 by i2, which now no longer dies in i3 and is used in
12231 i2 but does not die in i2, and place is between i2
12232 and i3, then we may need to move a link from place to
12233 i2. */
12234 if (i2 && INSN_UID (place) <= max_uid_cuid
12235 && INSN_CUID (place) > INSN_CUID (i2)
12236 && from_insn
12237 && INSN_CUID (from_insn) > INSN_CUID (i2)
12238 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12240 rtx links = LOG_LINKS (place);
12241 LOG_LINKS (place) = 0;
12242 distribute_links (links);
12244 break;
12247 if (tem == BB_HEAD (bb))
12248 break;
12251 /* We haven't found an insn for the death note and it
12252 is still a REG_DEAD note, but we have hit the beginning
12253 of the block. If the existing life info says the reg
12254 was dead, there's nothing left to do. Otherwise, we'll
12255 need to do a global life update after combine. */
12256 if (REG_NOTE_KIND (note) == REG_DEAD && place == 0
12257 && REGNO_REG_SET_P (bb->global_live_at_start,
12258 REGNO (XEXP (note, 0))))
12259 SET_BIT (refresh_blocks, this_basic_block->index);
12262 /* If the register is set or already dead at PLACE, we needn't do
12263 anything with this note if it is still a REG_DEAD note.
12264 We check here if it is set at all, not if is it totally replaced,
12265 which is what `dead_or_set_p' checks, so also check for it being
12266 set partially. */
12268 if (place && REG_NOTE_KIND (note) == REG_DEAD)
12270 unsigned int regno = REGNO (XEXP (note, 0));
12272 /* Similarly, if the instruction on which we want to place
12273 the note is a noop, we'll need do a global live update
12274 after we remove them in delete_noop_moves. */
12275 if (noop_move_p (place))
12276 SET_BIT (refresh_blocks, this_basic_block->index);
12278 if (dead_or_set_p (place, XEXP (note, 0))
12279 || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place)))
12281 /* Unless the register previously died in PLACE, clear
12282 last_death. [I no longer understand why this is
12283 being done.] */
12284 if (reg_stat[regno].last_death != place)
12285 reg_stat[regno].last_death = 0;
12286 place = 0;
12288 else
12289 reg_stat[regno].last_death = place;
12291 /* If this is a death note for a hard reg that is occupying
12292 multiple registers, ensure that we are still using all
12293 parts of the object. If we find a piece of the object
12294 that is unused, we must arrange for an appropriate REG_DEAD
12295 note to be added for it. However, we can't just emit a USE
12296 and tag the note to it, since the register might actually
12297 be dead; so we recourse, and the recursive call then finds
12298 the previous insn that used this register. */
12300 if (place && regno < FIRST_PSEUDO_REGISTER
12301 && hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))] > 1)
12303 unsigned int endregno
12304 = regno + hard_regno_nregs[regno]
12305 [GET_MODE (XEXP (note, 0))];
12306 int all_used = 1;
12307 unsigned int i;
12309 for (i = regno; i < endregno; i++)
12310 if ((! refers_to_regno_p (i, i + 1, PATTERN (place), 0)
12311 && ! find_regno_fusage (place, USE, i))
12312 || dead_or_set_regno_p (place, i))
12313 all_used = 0;
12315 if (! all_used)
12317 /* Put only REG_DEAD notes for pieces that are
12318 not already dead or set. */
12320 for (i = regno; i < endregno;
12321 i += hard_regno_nregs[i][reg_raw_mode[i]])
12323 rtx piece = regno_reg_rtx[i];
12324 basic_block bb = this_basic_block;
12326 if (! dead_or_set_p (place, piece)
12327 && ! reg_bitfield_target_p (piece,
12328 PATTERN (place)))
12330 rtx new_note
12331 = gen_rtx_EXPR_LIST (REG_DEAD, piece, NULL_RTX);
12333 distribute_notes (new_note, place, place,
12334 NULL_RTX);
12336 else if (! refers_to_regno_p (i, i + 1,
12337 PATTERN (place), 0)
12338 && ! find_regno_fusage (place, USE, i))
12339 for (tem = PREV_INSN (place); ;
12340 tem = PREV_INSN (tem))
12342 if (! INSN_P (tem))
12344 if (tem == BB_HEAD (bb))
12346 SET_BIT (refresh_blocks,
12347 this_basic_block->index);
12348 break;
12350 continue;
12352 if (dead_or_set_p (tem, piece)
12353 || reg_bitfield_target_p (piece,
12354 PATTERN (tem)))
12356 REG_NOTES (tem)
12357 = gen_rtx_EXPR_LIST (REG_UNUSED, piece,
12358 REG_NOTES (tem));
12359 break;
12365 place = 0;
12369 break;
12371 default:
12372 /* Any other notes should not be present at this point in the
12373 compilation. */
12374 gcc_unreachable ();
12377 if (place)
12379 XEXP (note, 1) = REG_NOTES (place);
12380 REG_NOTES (place) = note;
12382 else if ((REG_NOTE_KIND (note) == REG_DEAD
12383 || REG_NOTE_KIND (note) == REG_UNUSED)
12384 && REG_P (XEXP (note, 0)))
12385 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
12387 if (place2)
12389 if ((REG_NOTE_KIND (note) == REG_DEAD
12390 || REG_NOTE_KIND (note) == REG_UNUSED)
12391 && REG_P (XEXP (note, 0)))
12392 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
12394 REG_NOTES (place2) = gen_rtx_fmt_ee (GET_CODE (note),
12395 REG_NOTE_KIND (note),
12396 XEXP (note, 0),
12397 REG_NOTES (place2));
12402 /* Similarly to above, distribute the LOG_LINKS that used to be present on
12403 I3, I2, and I1 to new locations. This is also called to add a link
12404 pointing at I3 when I3's destination is changed. */
12406 static void
12407 distribute_links (rtx links)
12409 rtx link, next_link;
12411 for (link = links; link; link = next_link)
12413 rtx place = 0;
12414 rtx insn;
12415 rtx set, reg;
12417 next_link = XEXP (link, 1);
12419 /* If the insn that this link points to is a NOTE or isn't a single
12420 set, ignore it. In the latter case, it isn't clear what we
12421 can do other than ignore the link, since we can't tell which
12422 register it was for. Such links wouldn't be used by combine
12423 anyway.
12425 It is not possible for the destination of the target of the link to
12426 have been changed by combine. The only potential of this is if we
12427 replace I3, I2, and I1 by I3 and I2. But in that case the
12428 destination of I2 also remains unchanged. */
12430 if (NOTE_P (XEXP (link, 0))
12431 || (set = single_set (XEXP (link, 0))) == 0)
12432 continue;
12434 reg = SET_DEST (set);
12435 while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
12436 || GET_CODE (reg) == STRICT_LOW_PART)
12437 reg = XEXP (reg, 0);
12439 /* A LOG_LINK is defined as being placed on the first insn that uses
12440 a register and points to the insn that sets the register. Start
12441 searching at the next insn after the target of the link and stop
12442 when we reach a set of the register or the end of the basic block.
12444 Note that this correctly handles the link that used to point from
12445 I3 to I2. Also note that not much searching is typically done here
12446 since most links don't point very far away. */
12448 for (insn = NEXT_INSN (XEXP (link, 0));
12449 (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
12450 || BB_HEAD (this_basic_block->next_bb) != insn));
12451 insn = NEXT_INSN (insn))
12452 if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
12454 if (reg_referenced_p (reg, PATTERN (insn)))
12455 place = insn;
12456 break;
12458 else if (CALL_P (insn)
12459 && find_reg_fusage (insn, USE, reg))
12461 place = insn;
12462 break;
12464 else if (INSN_P (insn) && reg_set_p (reg, insn))
12465 break;
12467 /* If we found a place to put the link, place it there unless there
12468 is already a link to the same insn as LINK at that point. */
12470 if (place)
12472 rtx link2;
12474 for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1))
12475 if (XEXP (link2, 0) == XEXP (link, 0))
12476 break;
12478 if (link2 == 0)
12480 XEXP (link, 1) = LOG_LINKS (place);
12481 LOG_LINKS (place) = link;
12483 /* Set added_links_insn to the earliest insn we added a
12484 link to. */
12485 if (added_links_insn == 0
12486 || INSN_CUID (added_links_insn) > INSN_CUID (place))
12487 added_links_insn = place;
12493 /* Subroutine of unmentioned_reg_p and callback from for_each_rtx.
12494 Check whether the expression pointer to by LOC is a register or
12495 memory, and if so return 1 if it isn't mentioned in the rtx EXPR.
12496 Otherwise return zero. */
12498 static int
12499 unmentioned_reg_p_1 (rtx *loc, void *expr)
12501 rtx x = *loc;
12503 if (x != NULL_RTX
12504 && (REG_P (x) || MEM_P (x))
12505 && ! reg_mentioned_p (x, (rtx) expr))
12506 return 1;
12507 return 0;
12510 /* Check for any register or memory mentioned in EQUIV that is not
12511 mentioned in EXPR. This is used to restrict EQUIV to "specializations"
12512 of EXPR where some registers may have been replaced by constants. */
12514 static bool
12515 unmentioned_reg_p (rtx equiv, rtx expr)
12517 return for_each_rtx (&equiv, unmentioned_reg_p_1, expr);
12520 /* Compute INSN_CUID for INSN, which is an insn made by combine. */
12522 static int
12523 insn_cuid (rtx insn)
12525 while (insn != 0 && INSN_UID (insn) > max_uid_cuid
12526 && NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE)
12527 insn = NEXT_INSN (insn);
12529 gcc_assert (INSN_UID (insn) <= max_uid_cuid);
12531 return INSN_CUID (insn);
12534 void
12535 dump_combine_stats (FILE *file)
12537 fnotice
12538 (file,
12539 ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n",
12540 combine_attempts, combine_merges, combine_extras, combine_successes);
12543 void
12544 dump_combine_total_stats (FILE *file)
12546 fnotice
12547 (file,
12548 "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n",
12549 total_attempts, total_merges, total_extras, total_successes);