* config/rl78/mulsi3.S: Remove a few unneeded moves and branches.
[official-gcc.git] / gcc / combine.c
blob335d3ddd6cd4b31ed27df625651a450cffd22775
1 /* Optimize by combining instructions for GNU compiler.
2 Copyright (C) 1987-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This module is essentially the "combiner" phase of the U. of Arizona
21 Portable Optimizer, but redone to work on our list-structured
22 representation for RTL instead of their string representation.
24 The LOG_LINKS of each insn identify the most recent assignment
25 to each REG used in the insn. It is a list of previous insns,
26 each of which contains a SET for a REG that is used in this insn
27 and not used or set in between. LOG_LINKs never cross basic blocks.
28 They were set up by the preceding pass (lifetime analysis).
30 We try to combine each pair of insns joined by a logical link.
31 We also try to combine triplets of insns A, B and C when C has
32 a link back to B and B has a link back to A. Likewise for a
33 small number of quadruplets of insns A, B, C and D for which
34 there's high likelihood of of success.
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 isn't
53 completely updated (however this is only a local issue since it is
54 regenerated before the next pass that uses it):
56 - reg_live_length is not updated
57 - reg_n_refs is not adjusted in the rare case when a register is
58 no longer required in a computation
59 - there are extremely rare cases (see distribute_notes) when a
60 REG_DEAD note is lost
61 - a LOG_LINKS entry that refers to an insn with multiple SETs may be
62 removed because there is no way to know which register it was
63 linking
65 To simplify substitution, we combine only when the earlier insn(s)
66 consist of only a single assignment. To simplify updating afterward,
67 we never combine when a subroutine call appears in the middle.
69 Since we do not represent assignments to CC0 explicitly except when that
70 is all an insn does, there is no LOG_LINKS entry in an insn that uses
71 the condition code for the insn that set the condition code.
72 Fortunately, these two insns must be consecutive.
73 Therefore, every JUMP_INSN is taken to have an implicit logical link
74 to the preceding insn. This is not quite right, since non-jumps can
75 also use the condition code; but in practice such insns would not
76 combine anyway. */
78 #include "config.h"
79 #include "system.h"
80 #include "coretypes.h"
81 #include "tm.h"
82 #include "rtl.h"
83 #include "tree.h"
84 #include "tm_p.h"
85 #include "flags.h"
86 #include "regs.h"
87 #include "hard-reg-set.h"
88 #include "basic-block.h"
89 #include "insn-config.h"
90 #include "function.h"
91 /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */
92 #include "expr.h"
93 #include "insn-attr.h"
94 #include "recog.h"
95 #include "diagnostic-core.h"
96 #include "target.h"
97 #include "optabs.h"
98 #include "insn-codes.h"
99 #include "rtlhooks-def.h"
100 #include "params.h"
101 #include "tree-pass.h"
102 #include "df.h"
103 #include "valtrack.h"
104 #include "cgraph.h"
105 #include "obstack.h"
107 /* Number of attempts to combine instructions in this function. */
109 static int combine_attempts;
111 /* Number of attempts that got as far as substitution in this function. */
113 static int combine_merges;
115 /* Number of instructions combined with added SETs in this function. */
117 static int combine_extras;
119 /* Number of instructions combined in this function. */
121 static int combine_successes;
123 /* Totals over entire compilation. */
125 static int total_attempts, total_merges, total_extras, total_successes;
127 /* combine_instructions may try to replace the right hand side of the
128 second instruction with the value of an associated REG_EQUAL note
129 before throwing it at try_combine. That is problematic when there
130 is a REG_DEAD note for a register used in the old right hand side
131 and can cause distribute_notes to do wrong things. This is the
132 second instruction if it has been so modified, null otherwise. */
134 static rtx i2mod;
136 /* When I2MOD is nonnull, this is a copy of the old right hand side. */
138 static rtx i2mod_old_rhs;
140 /* When I2MOD is nonnull, this is a copy of the new right hand side. */
142 static rtx i2mod_new_rhs;
144 typedef struct reg_stat_struct {
145 /* Record last point of death of (hard or pseudo) register n. */
146 rtx last_death;
148 /* Record last point of modification of (hard or pseudo) register n. */
149 rtx last_set;
151 /* The next group of fields allows the recording of the last value assigned
152 to (hard or pseudo) register n. We use this information to see if an
153 operation being processed is redundant given a prior operation performed
154 on the register. For example, an `and' with a constant is redundant if
155 all the zero bits are already known to be turned off.
157 We use an approach similar to that used by cse, but change it in the
158 following ways:
160 (1) We do not want to reinitialize at each label.
161 (2) It is useful, but not critical, to know the actual value assigned
162 to a register. Often just its form is helpful.
164 Therefore, we maintain the following fields:
166 last_set_value the last value assigned
167 last_set_label records the value of label_tick when the
168 register was assigned
169 last_set_table_tick records the value of label_tick when a
170 value using the register is assigned
171 last_set_invalid set to nonzero when it is not valid
172 to use the value of this register in some
173 register's value
175 To understand the usage of these tables, it is important to understand
176 the distinction between the value in last_set_value being valid and
177 the register being validly contained in some other expression in the
178 table.
180 (The next two parameters are out of date).
182 reg_stat[i].last_set_value is valid if it is nonzero, and either
183 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
185 Register I may validly appear in any expression returned for the value
186 of another register if reg_n_sets[i] is 1. It may also appear in the
187 value for register J if reg_stat[j].last_set_invalid is zero, or
188 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
190 If an expression is found in the table containing a register which may
191 not validly appear in an expression, the register is replaced by
192 something that won't match, (clobber (const_int 0)). */
194 /* Record last value assigned to (hard or pseudo) register n. */
196 rtx last_set_value;
198 /* Record the value of label_tick when an expression involving register n
199 is placed in last_set_value. */
201 int last_set_table_tick;
203 /* Record the value of label_tick when the value for register n is placed in
204 last_set_value. */
206 int last_set_label;
208 /* These fields are maintained in parallel with last_set_value and are
209 used to store the mode in which the register was last set, the bits
210 that were known to be zero when it was last set, and the number of
211 sign bits copies it was known to have when it was last set. */
213 unsigned HOST_WIDE_INT last_set_nonzero_bits;
214 char last_set_sign_bit_copies;
215 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
217 /* Set nonzero if references to register n in expressions should not be
218 used. last_set_invalid is set nonzero when this register is being
219 assigned to and last_set_table_tick == label_tick. */
221 char last_set_invalid;
223 /* Some registers that are set more than once and used in more than one
224 basic block are nevertheless always set in similar ways. For example,
225 a QImode register may be loaded from memory in two places on a machine
226 where byte loads zero extend.
228 We record in the following fields if a register has some leading bits
229 that are always equal to the sign bit, and what we know about the
230 nonzero bits of a register, specifically which bits are known to be
231 zero.
233 If an entry is zero, it means that we don't know anything special. */
235 unsigned char sign_bit_copies;
237 unsigned HOST_WIDE_INT nonzero_bits;
239 /* Record the value of the label_tick when the last truncation
240 happened. The field truncated_to_mode is only valid if
241 truncation_label == label_tick. */
243 int truncation_label;
245 /* Record the last truncation seen for this register. If truncation
246 is not a nop to this mode we might be able to save an explicit
247 truncation if we know that value already contains a truncated
248 value. */
250 ENUM_BITFIELD(machine_mode) truncated_to_mode : 8;
251 } reg_stat_type;
254 static vec<reg_stat_type> reg_stat;
256 /* Record the luid of the last insn that invalidated memory
257 (anything that writes memory, and subroutine calls, but not pushes). */
259 static int mem_last_set;
261 /* Record the luid of the last CALL_INSN
262 so we can tell whether a potential combination crosses any calls. */
264 static int last_call_luid;
266 /* When `subst' is called, this is the insn that is being modified
267 (by combining in a previous insn). The PATTERN of this insn
268 is still the old pattern partially modified and it should not be
269 looked at, but this may be used to examine the successors of the insn
270 to judge whether a simplification is valid. */
272 static rtx subst_insn;
274 /* This is the lowest LUID that `subst' is currently dealing with.
275 get_last_value will not return a value if the register was set at or
276 after this LUID. If not for this mechanism, we could get confused if
277 I2 or I1 in try_combine were an insn that used the old value of a register
278 to obtain a new value. In that case, we might erroneously get the
279 new value of the register when we wanted the old one. */
281 static int subst_low_luid;
283 /* This contains any hard registers that are used in newpat; reg_dead_at_p
284 must consider all these registers to be always live. */
286 static HARD_REG_SET newpat_used_regs;
288 /* This is an insn to which a LOG_LINKS entry has been added. If this
289 insn is the earlier than I2 or I3, combine should rescan starting at
290 that location. */
292 static rtx added_links_insn;
294 /* Basic block in which we are performing combines. */
295 static basic_block this_basic_block;
296 static bool optimize_this_for_speed_p;
299 /* Length of the currently allocated uid_insn_cost array. */
301 static int max_uid_known;
303 /* The following array records the insn_rtx_cost for every insn
304 in the instruction stream. */
306 static int *uid_insn_cost;
308 /* The following array records the LOG_LINKS for every insn in the
309 instruction stream as struct insn_link pointers. */
311 struct insn_link {
312 rtx insn;
313 struct insn_link *next;
316 static struct insn_link **uid_log_links;
318 #define INSN_COST(INSN) (uid_insn_cost[INSN_UID (INSN)])
319 #define LOG_LINKS(INSN) (uid_log_links[INSN_UID (INSN)])
321 #define FOR_EACH_LOG_LINK(L, INSN) \
322 for ((L) = LOG_LINKS (INSN); (L); (L) = (L)->next)
324 /* Links for LOG_LINKS are allocated from this obstack. */
326 static struct obstack insn_link_obstack;
328 /* Allocate a link. */
330 static inline struct insn_link *
331 alloc_insn_link (rtx insn, struct insn_link *next)
333 struct insn_link *l
334 = (struct insn_link *) obstack_alloc (&insn_link_obstack,
335 sizeof (struct insn_link));
336 l->insn = insn;
337 l->next = next;
338 return l;
341 /* Incremented for each basic block. */
343 static int label_tick;
345 /* Reset to label_tick for each extended basic block in scanning order. */
347 static int label_tick_ebb_start;
349 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
350 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
352 static enum machine_mode nonzero_bits_mode;
354 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
355 be safely used. It is zero while computing them and after combine has
356 completed. This former test prevents propagating values based on
357 previously set values, which can be incorrect if a variable is modified
358 in a loop. */
360 static int nonzero_sign_valid;
363 /* Record one modification to rtl structure
364 to be undone by storing old_contents into *where. */
366 enum undo_kind { UNDO_RTX, UNDO_INT, UNDO_MODE, UNDO_LINKS };
368 struct undo
370 struct undo *next;
371 enum undo_kind kind;
372 union { rtx r; int i; enum machine_mode m; struct insn_link *l; } old_contents;
373 union { rtx *r; int *i; struct insn_link **l; } where;
376 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
377 num_undo says how many are currently recorded.
379 other_insn is nonzero if we have modified some other insn in the process
380 of working on subst_insn. It must be verified too. */
382 struct undobuf
384 struct undo *undos;
385 struct undo *frees;
386 rtx other_insn;
389 static struct undobuf undobuf;
391 /* Number of times the pseudo being substituted for
392 was found and replaced. */
394 static int n_occurrences;
396 static rtx reg_nonzero_bits_for_combine (const_rtx, enum machine_mode, const_rtx,
397 enum machine_mode,
398 unsigned HOST_WIDE_INT,
399 unsigned HOST_WIDE_INT *);
400 static rtx reg_num_sign_bit_copies_for_combine (const_rtx, enum machine_mode, const_rtx,
401 enum machine_mode,
402 unsigned int, unsigned int *);
403 static void do_SUBST (rtx *, rtx);
404 static void do_SUBST_INT (int *, int);
405 static void init_reg_last (void);
406 static void setup_incoming_promotions (rtx);
407 static void set_nonzero_bits_and_sign_copies (rtx, const_rtx, void *);
408 static int cant_combine_insn_p (rtx);
409 static int can_combine_p (rtx, rtx, rtx, rtx, rtx, rtx, rtx *, rtx *);
410 static int combinable_i3pat (rtx, rtx *, rtx, rtx, rtx, int, int, rtx *);
411 static int contains_muldiv (rtx);
412 static rtx try_combine (rtx, rtx, rtx, rtx, int *, rtx);
413 static void undo_all (void);
414 static void undo_commit (void);
415 static rtx *find_split_point (rtx *, rtx, bool);
416 static rtx subst (rtx, rtx, rtx, int, int, int);
417 static rtx combine_simplify_rtx (rtx, enum machine_mode, int, int);
418 static rtx simplify_if_then_else (rtx);
419 static rtx simplify_set (rtx);
420 static rtx simplify_logical (rtx);
421 static rtx expand_compound_operation (rtx);
422 static const_rtx expand_field_assignment (const_rtx);
423 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
424 rtx, unsigned HOST_WIDE_INT, int, int, int);
425 static rtx extract_left_shift (rtx, int);
426 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
427 unsigned HOST_WIDE_INT *);
428 static rtx canon_reg_for_combine (rtx, rtx);
429 static rtx force_to_mode (rtx, enum machine_mode,
430 unsigned HOST_WIDE_INT, int);
431 static rtx if_then_else_cond (rtx, rtx *, rtx *);
432 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
433 static int rtx_equal_for_field_assignment_p (rtx, rtx);
434 static rtx make_field_assignment (rtx);
435 static rtx apply_distributive_law (rtx);
436 static rtx distribute_and_simplify_rtx (rtx, int);
437 static rtx simplify_and_const_int_1 (enum machine_mode, rtx,
438 unsigned HOST_WIDE_INT);
439 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
440 unsigned HOST_WIDE_INT);
441 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
442 HOST_WIDE_INT, enum machine_mode, int *);
443 static rtx simplify_shift_const_1 (enum rtx_code, enum machine_mode, rtx, int);
444 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
445 int);
446 static int recog_for_combine (rtx *, rtx, rtx *);
447 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
448 static enum rtx_code simplify_compare_const (enum rtx_code, rtx, rtx *);
449 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
450 static void update_table_tick (rtx);
451 static void record_value_for_reg (rtx, rtx, rtx);
452 static void check_promoted_subreg (rtx, rtx);
453 static void record_dead_and_set_regs_1 (rtx, const_rtx, void *);
454 static void record_dead_and_set_regs (rtx);
455 static int get_last_value_validate (rtx *, rtx, int, int);
456 static rtx get_last_value (const_rtx);
457 static int use_crosses_set_p (const_rtx, int);
458 static void reg_dead_at_p_1 (rtx, const_rtx, void *);
459 static int reg_dead_at_p (rtx, rtx);
460 static void move_deaths (rtx, rtx, int, rtx, rtx *);
461 static int reg_bitfield_target_p (rtx, rtx);
462 static void distribute_notes (rtx, rtx, rtx, rtx, rtx, rtx, rtx);
463 static void distribute_links (struct insn_link *);
464 static void mark_used_regs_combine (rtx);
465 static void record_promoted_value (rtx, rtx);
466 static int unmentioned_reg_p_1 (rtx *, void *);
467 static bool unmentioned_reg_p (rtx, rtx);
468 static int record_truncated_value (rtx *, void *);
469 static void record_truncated_values (rtx *, void *);
470 static bool reg_truncated_to_mode (enum machine_mode, const_rtx);
471 static rtx gen_lowpart_or_truncate (enum machine_mode, rtx);
474 /* It is not safe to use ordinary gen_lowpart in combine.
475 See comments in gen_lowpart_for_combine. */
476 #undef RTL_HOOKS_GEN_LOWPART
477 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine
479 /* Our implementation of gen_lowpart never emits a new pseudo. */
480 #undef RTL_HOOKS_GEN_LOWPART_NO_EMIT
481 #define RTL_HOOKS_GEN_LOWPART_NO_EMIT gen_lowpart_for_combine
483 #undef RTL_HOOKS_REG_NONZERO_REG_BITS
484 #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine
486 #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES
487 #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine
489 #undef RTL_HOOKS_REG_TRUNCATED_TO_MODE
490 #define RTL_HOOKS_REG_TRUNCATED_TO_MODE reg_truncated_to_mode
492 static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER;
495 /* Convenience wrapper for the canonicalize_comparison target hook.
496 Target hooks cannot use enum rtx_code. */
497 static inline void
498 target_canonicalize_comparison (enum rtx_code *code, rtx *op0, rtx *op1,
499 bool op0_preserve_value)
501 int code_int = (int)*code;
502 targetm.canonicalize_comparison (&code_int, op0, op1, op0_preserve_value);
503 *code = (enum rtx_code)code_int;
506 /* Try to split PATTERN found in INSN. This returns NULL_RTX if
507 PATTERN can not be split. Otherwise, it returns an insn sequence.
508 This is a wrapper around split_insns which ensures that the
509 reg_stat vector is made larger if the splitter creates a new
510 register. */
512 static rtx
513 combine_split_insns (rtx pattern, rtx insn)
515 rtx ret;
516 unsigned int nregs;
518 ret = split_insns (pattern, insn);
519 nregs = max_reg_num ();
520 if (nregs > reg_stat.length ())
521 reg_stat.safe_grow_cleared (nregs);
522 return ret;
525 /* This is used by find_single_use to locate an rtx in LOC that
526 contains exactly one use of DEST, which is typically either a REG
527 or CC0. It returns a pointer to the innermost rtx expression
528 containing DEST. Appearances of DEST that are being used to
529 totally replace it are not counted. */
531 static rtx *
532 find_single_use_1 (rtx dest, rtx *loc)
534 rtx x = *loc;
535 enum rtx_code code = GET_CODE (x);
536 rtx *result = NULL;
537 rtx *this_result;
538 int i;
539 const char *fmt;
541 switch (code)
543 case CONST:
544 case LABEL_REF:
545 case SYMBOL_REF:
546 CASE_CONST_ANY:
547 case CLOBBER:
548 return 0;
550 case SET:
551 /* If the destination is anything other than CC0, PC, a REG or a SUBREG
552 of a REG that occupies all of the REG, the insn uses DEST if
553 it is mentioned in the destination or the source. Otherwise, we
554 need just check the source. */
555 if (GET_CODE (SET_DEST (x)) != CC0
556 && GET_CODE (SET_DEST (x)) != PC
557 && !REG_P (SET_DEST (x))
558 && ! (GET_CODE (SET_DEST (x)) == SUBREG
559 && REG_P (SUBREG_REG (SET_DEST (x)))
560 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
561 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
562 == ((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
563 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))))
564 break;
566 return find_single_use_1 (dest, &SET_SRC (x));
568 case MEM:
569 case SUBREG:
570 return find_single_use_1 (dest, &XEXP (x, 0));
572 default:
573 break;
576 /* If it wasn't one of the common cases above, check each expression and
577 vector of this code. Look for a unique usage of DEST. */
579 fmt = GET_RTX_FORMAT (code);
580 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
582 if (fmt[i] == 'e')
584 if (dest == XEXP (x, i)
585 || (REG_P (dest) && REG_P (XEXP (x, i))
586 && REGNO (dest) == REGNO (XEXP (x, i))))
587 this_result = loc;
588 else
589 this_result = find_single_use_1 (dest, &XEXP (x, i));
591 if (result == NULL)
592 result = this_result;
593 else if (this_result)
594 /* Duplicate usage. */
595 return NULL;
597 else if (fmt[i] == 'E')
599 int j;
601 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
603 if (XVECEXP (x, i, j) == dest
604 || (REG_P (dest)
605 && REG_P (XVECEXP (x, i, j))
606 && REGNO (XVECEXP (x, i, j)) == REGNO (dest)))
607 this_result = loc;
608 else
609 this_result = find_single_use_1 (dest, &XVECEXP (x, i, j));
611 if (result == NULL)
612 result = this_result;
613 else if (this_result)
614 return NULL;
619 return result;
623 /* See if DEST, produced in INSN, is used only a single time in the
624 sequel. If so, return a pointer to the innermost rtx expression in which
625 it is used.
627 If PLOC is nonzero, *PLOC is set to the insn containing the single use.
629 If DEST is cc0_rtx, we look only at the next insn. In that case, we don't
630 care about REG_DEAD notes or LOG_LINKS.
632 Otherwise, we find the single use by finding an insn that has a
633 LOG_LINKS pointing at INSN and has a REG_DEAD note for DEST. If DEST is
634 only referenced once in that insn, we know that it must be the first
635 and last insn referencing DEST. */
637 static rtx *
638 find_single_use (rtx dest, rtx insn, rtx *ploc)
640 basic_block bb;
641 rtx next;
642 rtx *result;
643 struct insn_link *link;
645 #ifdef HAVE_cc0
646 if (dest == cc0_rtx)
648 next = NEXT_INSN (insn);
649 if (next == 0
650 || (!NONJUMP_INSN_P (next) && !JUMP_P (next)))
651 return 0;
653 result = find_single_use_1 (dest, &PATTERN (next));
654 if (result && ploc)
655 *ploc = next;
656 return result;
658 #endif
660 if (!REG_P (dest))
661 return 0;
663 bb = BLOCK_FOR_INSN (insn);
664 for (next = NEXT_INSN (insn);
665 next && BLOCK_FOR_INSN (next) == bb;
666 next = NEXT_INSN (next))
667 if (INSN_P (next) && dead_or_set_p (next, dest))
669 FOR_EACH_LOG_LINK (link, next)
670 if (link->insn == insn)
671 break;
673 if (link)
675 result = find_single_use_1 (dest, &PATTERN (next));
676 if (ploc)
677 *ploc = next;
678 return result;
682 return 0;
685 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
686 insn. The substitution can be undone by undo_all. If INTO is already
687 set to NEWVAL, do not record this change. Because computing NEWVAL might
688 also call SUBST, we have to compute it before we put anything into
689 the undo table. */
691 static void
692 do_SUBST (rtx *into, rtx newval)
694 struct undo *buf;
695 rtx oldval = *into;
697 if (oldval == newval)
698 return;
700 /* We'd like to catch as many invalid transformations here as
701 possible. Unfortunately, there are way too many mode changes
702 that are perfectly valid, so we'd waste too much effort for
703 little gain doing the checks here. Focus on catching invalid
704 transformations involving integer constants. */
705 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
706 && CONST_INT_P (newval))
708 /* Sanity check that we're replacing oldval with a CONST_INT
709 that is a valid sign-extension for the original mode. */
710 gcc_assert (INTVAL (newval)
711 == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval)));
713 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
714 CONST_INT is not valid, because after the replacement, the
715 original mode would be gone. Unfortunately, we can't tell
716 when do_SUBST is called to replace the operand thereof, so we
717 perform this test on oldval instead, checking whether an
718 invalid replacement took place before we got here. */
719 gcc_assert (!(GET_CODE (oldval) == SUBREG
720 && CONST_INT_P (SUBREG_REG (oldval))));
721 gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
722 && CONST_INT_P (XEXP (oldval, 0))));
725 if (undobuf.frees)
726 buf = undobuf.frees, undobuf.frees = buf->next;
727 else
728 buf = XNEW (struct undo);
730 buf->kind = UNDO_RTX;
731 buf->where.r = into;
732 buf->old_contents.r = oldval;
733 *into = newval;
735 buf->next = undobuf.undos, undobuf.undos = buf;
738 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
740 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
741 for the value of a HOST_WIDE_INT value (including CONST_INT) is
742 not safe. */
744 static void
745 do_SUBST_INT (int *into, int newval)
747 struct undo *buf;
748 int oldval = *into;
750 if (oldval == newval)
751 return;
753 if (undobuf.frees)
754 buf = undobuf.frees, undobuf.frees = buf->next;
755 else
756 buf = XNEW (struct undo);
758 buf->kind = UNDO_INT;
759 buf->where.i = into;
760 buf->old_contents.i = oldval;
761 *into = newval;
763 buf->next = undobuf.undos, undobuf.undos = buf;
766 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
768 /* Similar to SUBST, but just substitute the mode. This is used when
769 changing the mode of a pseudo-register, so that any other
770 references to the entry in the regno_reg_rtx array will change as
771 well. */
773 static void
774 do_SUBST_MODE (rtx *into, enum machine_mode newval)
776 struct undo *buf;
777 enum machine_mode oldval = GET_MODE (*into);
779 if (oldval == newval)
780 return;
782 if (undobuf.frees)
783 buf = undobuf.frees, undobuf.frees = buf->next;
784 else
785 buf = XNEW (struct undo);
787 buf->kind = UNDO_MODE;
788 buf->where.r = into;
789 buf->old_contents.m = oldval;
790 adjust_reg_mode (*into, newval);
792 buf->next = undobuf.undos, undobuf.undos = buf;
795 #define SUBST_MODE(INTO, NEWVAL) do_SUBST_MODE(&(INTO), (NEWVAL))
797 #ifndef HAVE_cc0
798 /* Similar to SUBST, but NEWVAL is a LOG_LINKS expression. */
800 static void
801 do_SUBST_LINK (struct insn_link **into, struct insn_link *newval)
803 struct undo *buf;
804 struct insn_link * oldval = *into;
806 if (oldval == newval)
807 return;
809 if (undobuf.frees)
810 buf = undobuf.frees, undobuf.frees = buf->next;
811 else
812 buf = XNEW (struct undo);
814 buf->kind = UNDO_LINKS;
815 buf->where.l = into;
816 buf->old_contents.l = oldval;
817 *into = newval;
819 buf->next = undobuf.undos, undobuf.undos = buf;
822 #define SUBST_LINK(oldval, newval) do_SUBST_LINK (&oldval, newval)
823 #endif
825 /* Subroutine of try_combine. Determine whether the replacement patterns
826 NEWPAT, NEWI2PAT and NEWOTHERPAT are cheaper according to insn_rtx_cost
827 than the original sequence I0, I1, I2, I3 and undobuf.other_insn. Note
828 that I0, I1 and/or NEWI2PAT may be NULL_RTX. Similarly, NEWOTHERPAT and
829 undobuf.other_insn may also both be NULL_RTX. Return false if the cost
830 of all the instructions can be estimated and the replacements are more
831 expensive than the original sequence. */
833 static bool
834 combine_validate_cost (rtx i0, rtx i1, rtx i2, rtx i3, rtx newpat,
835 rtx newi2pat, rtx newotherpat)
837 int i0_cost, i1_cost, i2_cost, i3_cost;
838 int new_i2_cost, new_i3_cost;
839 int old_cost, new_cost;
841 /* Lookup the original insn_rtx_costs. */
842 i2_cost = INSN_COST (i2);
843 i3_cost = INSN_COST (i3);
845 if (i1)
847 i1_cost = INSN_COST (i1);
848 if (i0)
850 i0_cost = INSN_COST (i0);
851 old_cost = (i0_cost > 0 && i1_cost > 0 && i2_cost > 0 && i3_cost > 0
852 ? i0_cost + i1_cost + i2_cost + i3_cost : 0);
854 else
856 old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0
857 ? i1_cost + i2_cost + i3_cost : 0);
858 i0_cost = 0;
861 else
863 old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0;
864 i1_cost = i0_cost = 0;
867 /* Calculate the replacement insn_rtx_costs. */
868 new_i3_cost = insn_rtx_cost (newpat, optimize_this_for_speed_p);
869 if (newi2pat)
871 new_i2_cost = insn_rtx_cost (newi2pat, optimize_this_for_speed_p);
872 new_cost = (new_i2_cost > 0 && new_i3_cost > 0)
873 ? new_i2_cost + new_i3_cost : 0;
875 else
877 new_cost = new_i3_cost;
878 new_i2_cost = 0;
881 if (undobuf.other_insn)
883 int old_other_cost, new_other_cost;
885 old_other_cost = INSN_COST (undobuf.other_insn);
886 new_other_cost = insn_rtx_cost (newotherpat, optimize_this_for_speed_p);
887 if (old_other_cost > 0 && new_other_cost > 0)
889 old_cost += old_other_cost;
890 new_cost += new_other_cost;
892 else
893 old_cost = 0;
896 /* Disallow this combination if both new_cost and old_cost are greater than
897 zero, and new_cost is greater than old cost. */
898 if (old_cost > 0 && new_cost > old_cost)
900 if (dump_file)
902 if (i0)
904 fprintf (dump_file,
905 "rejecting combination of insns %d, %d, %d and %d\n",
906 INSN_UID (i0), INSN_UID (i1), INSN_UID (i2),
907 INSN_UID (i3));
908 fprintf (dump_file, "original costs %d + %d + %d + %d = %d\n",
909 i0_cost, i1_cost, i2_cost, i3_cost, old_cost);
911 else if (i1)
913 fprintf (dump_file,
914 "rejecting combination of insns %d, %d and %d\n",
915 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
916 fprintf (dump_file, "original costs %d + %d + %d = %d\n",
917 i1_cost, i2_cost, i3_cost, old_cost);
919 else
921 fprintf (dump_file,
922 "rejecting combination of insns %d and %d\n",
923 INSN_UID (i2), INSN_UID (i3));
924 fprintf (dump_file, "original costs %d + %d = %d\n",
925 i2_cost, i3_cost, old_cost);
928 if (newi2pat)
930 fprintf (dump_file, "replacement costs %d + %d = %d\n",
931 new_i2_cost, new_i3_cost, new_cost);
933 else
934 fprintf (dump_file, "replacement cost %d\n", new_cost);
937 return false;
940 /* Update the uid_insn_cost array with the replacement costs. */
941 INSN_COST (i2) = new_i2_cost;
942 INSN_COST (i3) = new_i3_cost;
943 if (i1)
945 INSN_COST (i1) = 0;
946 if (i0)
947 INSN_COST (i0) = 0;
950 return true;
954 /* Delete any insns that copy a register to itself. */
956 static void
957 delete_noop_moves (void)
959 rtx insn, next;
960 basic_block bb;
962 FOR_EACH_BB (bb)
964 for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb)); insn = next)
966 next = NEXT_INSN (insn);
967 if (INSN_P (insn) && noop_move_p (insn))
969 if (dump_file)
970 fprintf (dump_file, "deleting noop move %d\n", INSN_UID (insn));
972 delete_insn_and_edges (insn);
979 /* Fill in log links field for all insns. */
981 static void
982 create_log_links (void)
984 basic_block bb;
985 rtx *next_use, insn;
986 df_ref *def_vec, *use_vec;
988 next_use = XCNEWVEC (rtx, max_reg_num ());
990 /* Pass through each block from the end, recording the uses of each
991 register and establishing log links when def is encountered.
992 Note that we do not clear next_use array in order to save time,
993 so we have to test whether the use is in the same basic block as def.
995 There are a few cases below when we do not consider the definition or
996 usage -- these are taken from original flow.c did. Don't ask me why it is
997 done this way; I don't know and if it works, I don't want to know. */
999 FOR_EACH_BB (bb)
1001 FOR_BB_INSNS_REVERSE (bb, insn)
1003 if (!NONDEBUG_INSN_P (insn))
1004 continue;
1006 /* Log links are created only once. */
1007 gcc_assert (!LOG_LINKS (insn));
1009 for (def_vec = DF_INSN_DEFS (insn); *def_vec; def_vec++)
1011 df_ref def = *def_vec;
1012 int regno = DF_REF_REGNO (def);
1013 rtx use_insn;
1015 if (!next_use[regno])
1016 continue;
1018 /* Do not consider if it is pre/post modification in MEM. */
1019 if (DF_REF_FLAGS (def) & DF_REF_PRE_POST_MODIFY)
1020 continue;
1022 /* Do not make the log link for frame pointer. */
1023 if ((regno == FRAME_POINTER_REGNUM
1024 && (! reload_completed || frame_pointer_needed))
1025 #if !HARD_FRAME_POINTER_IS_FRAME_POINTER
1026 || (regno == HARD_FRAME_POINTER_REGNUM
1027 && (! reload_completed || frame_pointer_needed))
1028 #endif
1029 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
1030 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
1031 #endif
1033 continue;
1035 use_insn = next_use[regno];
1036 if (BLOCK_FOR_INSN (use_insn) == bb)
1038 /* flow.c claimed:
1040 We don't build a LOG_LINK for hard registers contained
1041 in ASM_OPERANDs. If these registers get replaced,
1042 we might wind up changing the semantics of the insn,
1043 even if reload can make what appear to be valid
1044 assignments later. */
1045 if (regno >= FIRST_PSEUDO_REGISTER
1046 || asm_noperands (PATTERN (use_insn)) < 0)
1048 /* Don't add duplicate links between instructions. */
1049 struct insn_link *links;
1050 FOR_EACH_LOG_LINK (links, use_insn)
1051 if (insn == links->insn)
1052 break;
1054 if (!links)
1055 LOG_LINKS (use_insn)
1056 = alloc_insn_link (insn, LOG_LINKS (use_insn));
1059 next_use[regno] = NULL_RTX;
1062 for (use_vec = DF_INSN_USES (insn); *use_vec; use_vec++)
1064 df_ref use = *use_vec;
1065 int regno = DF_REF_REGNO (use);
1067 /* Do not consider the usage of the stack pointer
1068 by function call. */
1069 if (DF_REF_FLAGS (use) & DF_REF_CALL_STACK_USAGE)
1070 continue;
1072 next_use[regno] = insn;
1077 free (next_use);
1080 /* Walk the LOG_LINKS of insn B to see if we find a reference to A. Return
1081 true if we found a LOG_LINK that proves that A feeds B. This only works
1082 if there are no instructions between A and B which could have a link
1083 depending on A, since in that case we would not record a link for B.
1084 We also check the implicit dependency created by a cc0 setter/user
1085 pair. */
1087 static bool
1088 insn_a_feeds_b (rtx a, rtx b)
1090 struct insn_link *links;
1091 FOR_EACH_LOG_LINK (links, b)
1092 if (links->insn == a)
1093 return true;
1094 #ifdef HAVE_cc0
1095 if (sets_cc0_p (a))
1096 return true;
1097 #endif
1098 return false;
1101 /* Main entry point for combiner. F is the first insn of the function.
1102 NREGS is the first unused pseudo-reg number.
1104 Return nonzero if the combiner has turned an indirect jump
1105 instruction into a direct jump. */
1106 static int
1107 combine_instructions (rtx f, unsigned int nregs)
1109 rtx insn, next;
1110 #ifdef HAVE_cc0
1111 rtx prev;
1112 #endif
1113 struct insn_link *links, *nextlinks;
1114 rtx first;
1115 basic_block last_bb;
1117 int new_direct_jump_p = 0;
1119 for (first = f; first && !INSN_P (first); )
1120 first = NEXT_INSN (first);
1121 if (!first)
1122 return 0;
1124 combine_attempts = 0;
1125 combine_merges = 0;
1126 combine_extras = 0;
1127 combine_successes = 0;
1129 rtl_hooks = combine_rtl_hooks;
1131 reg_stat.safe_grow_cleared (nregs);
1133 init_recog_no_volatile ();
1135 /* Allocate array for insn info. */
1136 max_uid_known = get_max_uid ();
1137 uid_log_links = XCNEWVEC (struct insn_link *, max_uid_known + 1);
1138 uid_insn_cost = XCNEWVEC (int, max_uid_known + 1);
1139 gcc_obstack_init (&insn_link_obstack);
1141 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
1143 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
1144 problems when, for example, we have j <<= 1 in a loop. */
1146 nonzero_sign_valid = 0;
1147 label_tick = label_tick_ebb_start = 1;
1149 /* Scan all SETs and see if we can deduce anything about what
1150 bits are known to be zero for some registers and how many copies
1151 of the sign bit are known to exist for those registers.
1153 Also set any known values so that we can use it while searching
1154 for what bits are known to be set. */
1156 setup_incoming_promotions (first);
1157 /* Allow the entry block and the first block to fall into the same EBB.
1158 Conceptually the incoming promotions are assigned to the entry block. */
1159 last_bb = ENTRY_BLOCK_PTR;
1161 create_log_links ();
1162 FOR_EACH_BB (this_basic_block)
1164 optimize_this_for_speed_p = optimize_bb_for_speed_p (this_basic_block);
1165 last_call_luid = 0;
1166 mem_last_set = -1;
1168 label_tick++;
1169 if (!single_pred_p (this_basic_block)
1170 || single_pred (this_basic_block) != last_bb)
1171 label_tick_ebb_start = label_tick;
1172 last_bb = this_basic_block;
1174 FOR_BB_INSNS (this_basic_block, insn)
1175 if (INSN_P (insn) && BLOCK_FOR_INSN (insn))
1177 #ifdef AUTO_INC_DEC
1178 rtx links;
1179 #endif
1181 subst_low_luid = DF_INSN_LUID (insn);
1182 subst_insn = insn;
1184 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
1185 insn);
1186 record_dead_and_set_regs (insn);
1188 #ifdef AUTO_INC_DEC
1189 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
1190 if (REG_NOTE_KIND (links) == REG_INC)
1191 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
1192 insn);
1193 #endif
1195 /* Record the current insn_rtx_cost of this instruction. */
1196 if (NONJUMP_INSN_P (insn))
1197 INSN_COST (insn) = insn_rtx_cost (PATTERN (insn),
1198 optimize_this_for_speed_p);
1199 if (dump_file)
1200 fprintf(dump_file, "insn_cost %d: %d\n",
1201 INSN_UID (insn), INSN_COST (insn));
1205 nonzero_sign_valid = 1;
1207 /* Now scan all the insns in forward order. */
1208 label_tick = label_tick_ebb_start = 1;
1209 init_reg_last ();
1210 setup_incoming_promotions (first);
1211 last_bb = ENTRY_BLOCK_PTR;
1213 FOR_EACH_BB (this_basic_block)
1215 rtx last_combined_insn = NULL_RTX;
1216 optimize_this_for_speed_p = optimize_bb_for_speed_p (this_basic_block);
1217 last_call_luid = 0;
1218 mem_last_set = -1;
1220 label_tick++;
1221 if (!single_pred_p (this_basic_block)
1222 || single_pred (this_basic_block) != last_bb)
1223 label_tick_ebb_start = label_tick;
1224 last_bb = this_basic_block;
1226 rtl_profile_for_bb (this_basic_block);
1227 for (insn = BB_HEAD (this_basic_block);
1228 insn != NEXT_INSN (BB_END (this_basic_block));
1229 insn = next ? next : NEXT_INSN (insn))
1231 next = 0;
1232 if (NONDEBUG_INSN_P (insn))
1234 while (last_combined_insn
1235 && INSN_DELETED_P (last_combined_insn))
1236 last_combined_insn = PREV_INSN (last_combined_insn);
1237 if (last_combined_insn == NULL_RTX
1238 || BARRIER_P (last_combined_insn)
1239 || BLOCK_FOR_INSN (last_combined_insn) != this_basic_block
1240 || DF_INSN_LUID (last_combined_insn) <= DF_INSN_LUID (insn))
1241 last_combined_insn = insn;
1243 /* See if we know about function return values before this
1244 insn based upon SUBREG flags. */
1245 check_promoted_subreg (insn, PATTERN (insn));
1247 /* See if we can find hardregs and subreg of pseudos in
1248 narrower modes. This could help turning TRUNCATEs
1249 into SUBREGs. */
1250 note_uses (&PATTERN (insn), record_truncated_values, NULL);
1252 /* Try this insn with each insn it links back to. */
1254 FOR_EACH_LOG_LINK (links, insn)
1255 if ((next = try_combine (insn, links->insn, NULL_RTX,
1256 NULL_RTX, &new_direct_jump_p,
1257 last_combined_insn)) != 0)
1258 goto retry;
1260 /* Try each sequence of three linked insns ending with this one. */
1262 FOR_EACH_LOG_LINK (links, insn)
1264 rtx link = links->insn;
1266 /* If the linked insn has been replaced by a note, then there
1267 is no point in pursuing this chain any further. */
1268 if (NOTE_P (link))
1269 continue;
1271 FOR_EACH_LOG_LINK (nextlinks, link)
1272 if ((next = try_combine (insn, link, nextlinks->insn,
1273 NULL_RTX, &new_direct_jump_p,
1274 last_combined_insn)) != 0)
1275 goto retry;
1278 #ifdef HAVE_cc0
1279 /* Try to combine a jump insn that uses CC0
1280 with a preceding insn that sets CC0, and maybe with its
1281 logical predecessor as well.
1282 This is how we make decrement-and-branch insns.
1283 We need this special code because data flow connections
1284 via CC0 do not get entered in LOG_LINKS. */
1286 if (JUMP_P (insn)
1287 && (prev = prev_nonnote_insn (insn)) != 0
1288 && NONJUMP_INSN_P (prev)
1289 && sets_cc0_p (PATTERN (prev)))
1291 if ((next = try_combine (insn, prev, NULL_RTX, NULL_RTX,
1292 &new_direct_jump_p,
1293 last_combined_insn)) != 0)
1294 goto retry;
1296 FOR_EACH_LOG_LINK (nextlinks, prev)
1297 if ((next = try_combine (insn, prev, nextlinks->insn,
1298 NULL_RTX, &new_direct_jump_p,
1299 last_combined_insn)) != 0)
1300 goto retry;
1303 /* Do the same for an insn that explicitly references CC0. */
1304 if (NONJUMP_INSN_P (insn)
1305 && (prev = prev_nonnote_insn (insn)) != 0
1306 && NONJUMP_INSN_P (prev)
1307 && sets_cc0_p (PATTERN (prev))
1308 && GET_CODE (PATTERN (insn)) == SET
1309 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
1311 if ((next = try_combine (insn, prev, NULL_RTX, NULL_RTX,
1312 &new_direct_jump_p,
1313 last_combined_insn)) != 0)
1314 goto retry;
1316 FOR_EACH_LOG_LINK (nextlinks, prev)
1317 if ((next = try_combine (insn, prev, nextlinks->insn,
1318 NULL_RTX, &new_direct_jump_p,
1319 last_combined_insn)) != 0)
1320 goto retry;
1323 /* Finally, see if any of the insns that this insn links to
1324 explicitly references CC0. If so, try this insn, that insn,
1325 and its predecessor if it sets CC0. */
1326 FOR_EACH_LOG_LINK (links, insn)
1327 if (NONJUMP_INSN_P (links->insn)
1328 && GET_CODE (PATTERN (links->insn)) == SET
1329 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (links->insn)))
1330 && (prev = prev_nonnote_insn (links->insn)) != 0
1331 && NONJUMP_INSN_P (prev)
1332 && sets_cc0_p (PATTERN (prev))
1333 && (next = try_combine (insn, links->insn,
1334 prev, NULL_RTX, &new_direct_jump_p,
1335 last_combined_insn)) != 0)
1336 goto retry;
1337 #endif
1339 /* Try combining an insn with two different insns whose results it
1340 uses. */
1341 FOR_EACH_LOG_LINK (links, insn)
1342 for (nextlinks = links->next; nextlinks;
1343 nextlinks = nextlinks->next)
1344 if ((next = try_combine (insn, links->insn,
1345 nextlinks->insn, NULL_RTX,
1346 &new_direct_jump_p,
1347 last_combined_insn)) != 0)
1348 goto retry;
1350 /* Try four-instruction combinations. */
1351 FOR_EACH_LOG_LINK (links, insn)
1353 struct insn_link *next1;
1354 rtx link = links->insn;
1356 /* If the linked insn has been replaced by a note, then there
1357 is no point in pursuing this chain any further. */
1358 if (NOTE_P (link))
1359 continue;
1361 FOR_EACH_LOG_LINK (next1, link)
1363 rtx link1 = next1->insn;
1364 if (NOTE_P (link1))
1365 continue;
1366 /* I0 -> I1 -> I2 -> I3. */
1367 FOR_EACH_LOG_LINK (nextlinks, link1)
1368 if ((next = try_combine (insn, link, link1,
1369 nextlinks->insn,
1370 &new_direct_jump_p,
1371 last_combined_insn)) != 0)
1372 goto retry;
1373 /* I0, I1 -> I2, I2 -> I3. */
1374 for (nextlinks = next1->next; nextlinks;
1375 nextlinks = nextlinks->next)
1376 if ((next = try_combine (insn, link, link1,
1377 nextlinks->insn,
1378 &new_direct_jump_p,
1379 last_combined_insn)) != 0)
1380 goto retry;
1383 for (next1 = links->next; next1; next1 = next1->next)
1385 rtx link1 = next1->insn;
1386 if (NOTE_P (link1))
1387 continue;
1388 /* I0 -> I2; I1, I2 -> I3. */
1389 FOR_EACH_LOG_LINK (nextlinks, link)
1390 if ((next = try_combine (insn, link, link1,
1391 nextlinks->insn,
1392 &new_direct_jump_p,
1393 last_combined_insn)) != 0)
1394 goto retry;
1395 /* I0 -> I1; I1, I2 -> I3. */
1396 FOR_EACH_LOG_LINK (nextlinks, link1)
1397 if ((next = try_combine (insn, link, link1,
1398 nextlinks->insn,
1399 &new_direct_jump_p,
1400 last_combined_insn)) != 0)
1401 goto retry;
1405 /* Try this insn with each REG_EQUAL note it links back to. */
1406 FOR_EACH_LOG_LINK (links, insn)
1408 rtx set, note;
1409 rtx temp = links->insn;
1410 if ((set = single_set (temp)) != 0
1411 && (note = find_reg_equal_equiv_note (temp)) != 0
1412 && (note = XEXP (note, 0), GET_CODE (note)) != EXPR_LIST
1413 /* Avoid using a register that may already been marked
1414 dead by an earlier instruction. */
1415 && ! unmentioned_reg_p (note, SET_SRC (set))
1416 && (GET_MODE (note) == VOIDmode
1417 ? SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))
1418 : GET_MODE (SET_DEST (set)) == GET_MODE (note)))
1420 /* Temporarily replace the set's source with the
1421 contents of the REG_EQUAL note. The insn will
1422 be deleted or recognized by try_combine. */
1423 rtx orig = SET_SRC (set);
1424 SET_SRC (set) = note;
1425 i2mod = temp;
1426 i2mod_old_rhs = copy_rtx (orig);
1427 i2mod_new_rhs = copy_rtx (note);
1428 next = try_combine (insn, i2mod, NULL_RTX, NULL_RTX,
1429 &new_direct_jump_p,
1430 last_combined_insn);
1431 i2mod = NULL_RTX;
1432 if (next)
1433 goto retry;
1434 SET_SRC (set) = orig;
1438 if (!NOTE_P (insn))
1439 record_dead_and_set_regs (insn);
1441 retry:
1447 default_rtl_profile ();
1448 clear_bb_flags ();
1449 new_direct_jump_p |= purge_all_dead_edges ();
1450 delete_noop_moves ();
1452 /* Clean up. */
1453 obstack_free (&insn_link_obstack, NULL);
1454 free (uid_log_links);
1455 free (uid_insn_cost);
1456 reg_stat.release ();
1459 struct undo *undo, *next;
1460 for (undo = undobuf.frees; undo; undo = next)
1462 next = undo->next;
1463 free (undo);
1465 undobuf.frees = 0;
1468 total_attempts += combine_attempts;
1469 total_merges += combine_merges;
1470 total_extras += combine_extras;
1471 total_successes += combine_successes;
1473 nonzero_sign_valid = 0;
1474 rtl_hooks = general_rtl_hooks;
1476 /* Make recognizer allow volatile MEMs again. */
1477 init_recog ();
1479 return new_direct_jump_p;
1482 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
1484 static void
1485 init_reg_last (void)
1487 unsigned int i;
1488 reg_stat_type *p;
1490 FOR_EACH_VEC_ELT (reg_stat, i, p)
1491 memset (p, 0, offsetof (reg_stat_type, sign_bit_copies));
1494 /* Set up any promoted values for incoming argument registers. */
1496 static void
1497 setup_incoming_promotions (rtx first)
1499 tree arg;
1500 bool strictly_local = false;
1502 for (arg = DECL_ARGUMENTS (current_function_decl); arg;
1503 arg = DECL_CHAIN (arg))
1505 rtx x, reg = DECL_INCOMING_RTL (arg);
1506 int uns1, uns3;
1507 enum machine_mode mode1, mode2, mode3, mode4;
1509 /* Only continue if the incoming argument is in a register. */
1510 if (!REG_P (reg))
1511 continue;
1513 /* Determine, if possible, whether all call sites of the current
1514 function lie within the current compilation unit. (This does
1515 take into account the exporting of a function via taking its
1516 address, and so forth.) */
1517 strictly_local = cgraph_local_info (current_function_decl)->local;
1519 /* The mode and signedness of the argument before any promotions happen
1520 (equal to the mode of the pseudo holding it at that stage). */
1521 mode1 = TYPE_MODE (TREE_TYPE (arg));
1522 uns1 = TYPE_UNSIGNED (TREE_TYPE (arg));
1524 /* The mode and signedness of the argument after any source language and
1525 TARGET_PROMOTE_PROTOTYPES-driven promotions. */
1526 mode2 = TYPE_MODE (DECL_ARG_TYPE (arg));
1527 uns3 = TYPE_UNSIGNED (DECL_ARG_TYPE (arg));
1529 /* The mode and signedness of the argument as it is actually passed,
1530 after any TARGET_PROMOTE_FUNCTION_ARGS-driven ABI promotions. */
1531 mode3 = promote_function_mode (DECL_ARG_TYPE (arg), mode2, &uns3,
1532 TREE_TYPE (cfun->decl), 0);
1534 /* The mode of the register in which the argument is being passed. */
1535 mode4 = GET_MODE (reg);
1537 /* Eliminate sign extensions in the callee when:
1538 (a) A mode promotion has occurred; */
1539 if (mode1 == mode3)
1540 continue;
1541 /* (b) The mode of the register is the same as the mode of
1542 the argument as it is passed; */
1543 if (mode3 != mode4)
1544 continue;
1545 /* (c) There's no language level extension; */
1546 if (mode1 == mode2)
1548 /* (c.1) All callers are from the current compilation unit. If that's
1549 the case we don't have to rely on an ABI, we only have to know
1550 what we're generating right now, and we know that we will do the
1551 mode1 to mode2 promotion with the given sign. */
1552 else if (!strictly_local)
1553 continue;
1554 /* (c.2) The combination of the two promotions is useful. This is
1555 true when the signs match, or if the first promotion is unsigned.
1556 In the later case, (sign_extend (zero_extend x)) is the same as
1557 (zero_extend (zero_extend x)), so make sure to force UNS3 true. */
1558 else if (uns1)
1559 uns3 = true;
1560 else if (uns3)
1561 continue;
1563 /* Record that the value was promoted from mode1 to mode3,
1564 so that any sign extension at the head of the current
1565 function may be eliminated. */
1566 x = gen_rtx_CLOBBER (mode1, const0_rtx);
1567 x = gen_rtx_fmt_e ((uns3 ? ZERO_EXTEND : SIGN_EXTEND), mode3, x);
1568 record_value_for_reg (reg, first, x);
1572 /* Called via note_stores. If X is a pseudo that is narrower than
1573 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
1575 If we are setting only a portion of X and we can't figure out what
1576 portion, assume all bits will be used since we don't know what will
1577 be happening.
1579 Similarly, set how many bits of X are known to be copies of the sign bit
1580 at all locations in the function. This is the smallest number implied
1581 by any set of X. */
1583 static void
1584 set_nonzero_bits_and_sign_copies (rtx x, const_rtx set, void *data)
1586 rtx insn = (rtx) data;
1587 unsigned int num;
1589 if (REG_P (x)
1590 && REGNO (x) >= FIRST_PSEUDO_REGISTER
1591 /* If this register is undefined at the start of the file, we can't
1592 say what its contents were. */
1593 && ! REGNO_REG_SET_P
1594 (DF_LR_IN (ENTRY_BLOCK_PTR->next_bb), REGNO (x))
1595 && HWI_COMPUTABLE_MODE_P (GET_MODE (x)))
1597 reg_stat_type *rsp = &reg_stat[REGNO (x)];
1599 if (set == 0 || GET_CODE (set) == CLOBBER)
1601 rsp->nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1602 rsp->sign_bit_copies = 1;
1603 return;
1606 /* If this register is being initialized using itself, and the
1607 register is uninitialized in this basic block, and there are
1608 no LOG_LINKS which set the register, then part of the
1609 register is uninitialized. In that case we can't assume
1610 anything about the number of nonzero bits.
1612 ??? We could do better if we checked this in
1613 reg_{nonzero_bits,num_sign_bit_copies}_for_combine. Then we
1614 could avoid making assumptions about the insn which initially
1615 sets the register, while still using the information in other
1616 insns. We would have to be careful to check every insn
1617 involved in the combination. */
1619 if (insn
1620 && reg_referenced_p (x, PATTERN (insn))
1621 && !REGNO_REG_SET_P (DF_LR_IN (BLOCK_FOR_INSN (insn)),
1622 REGNO (x)))
1624 struct insn_link *link;
1626 FOR_EACH_LOG_LINK (link, insn)
1627 if (dead_or_set_p (link->insn, x))
1628 break;
1629 if (!link)
1631 rsp->nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1632 rsp->sign_bit_copies = 1;
1633 return;
1637 /* If this is a complex assignment, see if we can convert it into a
1638 simple assignment. */
1639 set = expand_field_assignment (set);
1641 /* If this is a simple assignment, or we have a paradoxical SUBREG,
1642 set what we know about X. */
1644 if (SET_DEST (set) == x
1645 || (paradoxical_subreg_p (SET_DEST (set))
1646 && SUBREG_REG (SET_DEST (set)) == x))
1648 rtx src = SET_SRC (set);
1650 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
1651 /* If X is narrower than a word and SRC is a non-negative
1652 constant that would appear negative in the mode of X,
1653 sign-extend it for use in reg_stat[].nonzero_bits because some
1654 machines (maybe most) will actually do the sign-extension
1655 and this is the conservative approach.
1657 ??? For 2.5, try to tighten up the MD files in this regard
1658 instead of this kludge. */
1660 if (GET_MODE_PRECISION (GET_MODE (x)) < BITS_PER_WORD
1661 && CONST_INT_P (src)
1662 && INTVAL (src) > 0
1663 && val_signbit_known_set_p (GET_MODE (x), INTVAL (src)))
1664 src = GEN_INT (INTVAL (src) | ~GET_MODE_MASK (GET_MODE (x)));
1665 #endif
1667 /* Don't call nonzero_bits if it cannot change anything. */
1668 if (rsp->nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
1669 rsp->nonzero_bits |= nonzero_bits (src, nonzero_bits_mode);
1670 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
1671 if (rsp->sign_bit_copies == 0
1672 || rsp->sign_bit_copies > num)
1673 rsp->sign_bit_copies = num;
1675 else
1677 rsp->nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1678 rsp->sign_bit_copies = 1;
1683 /* See if INSN can be combined into I3. PRED, PRED2, SUCC and SUCC2 are
1684 optionally insns that were previously combined into I3 or that will be
1685 combined into the merger of INSN and I3. The order is PRED, PRED2,
1686 INSN, SUCC, SUCC2, I3.
1688 Return 0 if the combination is not allowed for any reason.
1690 If the combination is allowed, *PDEST will be set to the single
1691 destination of INSN and *PSRC to the single source, and this function
1692 will return 1. */
1694 static int
1695 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED,
1696 rtx pred2 ATTRIBUTE_UNUSED, rtx succ, rtx succ2,
1697 rtx *pdest, rtx *psrc)
1699 int i;
1700 const_rtx set = 0;
1701 rtx src, dest;
1702 rtx p;
1703 #ifdef AUTO_INC_DEC
1704 rtx link;
1705 #endif
1706 bool all_adjacent = true;
1707 int (*is_volatile_p) (const_rtx);
1709 if (succ)
1711 if (succ2)
1713 if (next_active_insn (succ2) != i3)
1714 all_adjacent = false;
1715 if (next_active_insn (succ) != succ2)
1716 all_adjacent = false;
1718 else if (next_active_insn (succ) != i3)
1719 all_adjacent = false;
1720 if (next_active_insn (insn) != succ)
1721 all_adjacent = false;
1723 else if (next_active_insn (insn) != i3)
1724 all_adjacent = false;
1726 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
1727 or a PARALLEL consisting of such a SET and CLOBBERs.
1729 If INSN has CLOBBER parallel parts, ignore them for our processing.
1730 By definition, these happen during the execution of the insn. When it
1731 is merged with another insn, all bets are off. If they are, in fact,
1732 needed and aren't also supplied in I3, they may be added by
1733 recog_for_combine. Otherwise, it won't match.
1735 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
1736 note.
1738 Get the source and destination of INSN. If more than one, can't
1739 combine. */
1741 if (GET_CODE (PATTERN (insn)) == SET)
1742 set = PATTERN (insn);
1743 else if (GET_CODE (PATTERN (insn)) == PARALLEL
1744 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
1746 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1748 rtx elt = XVECEXP (PATTERN (insn), 0, i);
1750 switch (GET_CODE (elt))
1752 /* This is important to combine floating point insns
1753 for the SH4 port. */
1754 case USE:
1755 /* Combining an isolated USE doesn't make sense.
1756 We depend here on combinable_i3pat to reject them. */
1757 /* The code below this loop only verifies that the inputs of
1758 the SET in INSN do not change. We call reg_set_between_p
1759 to verify that the REG in the USE does not change between
1760 I3 and INSN.
1761 If the USE in INSN was for a pseudo register, the matching
1762 insn pattern will likely match any register; combining this
1763 with any other USE would only be safe if we knew that the
1764 used registers have identical values, or if there was
1765 something to tell them apart, e.g. different modes. For
1766 now, we forgo such complicated tests and simply disallow
1767 combining of USES of pseudo registers with any other USE. */
1768 if (REG_P (XEXP (elt, 0))
1769 && GET_CODE (PATTERN (i3)) == PARALLEL)
1771 rtx i3pat = PATTERN (i3);
1772 int i = XVECLEN (i3pat, 0) - 1;
1773 unsigned int regno = REGNO (XEXP (elt, 0));
1777 rtx i3elt = XVECEXP (i3pat, 0, i);
1779 if (GET_CODE (i3elt) == USE
1780 && REG_P (XEXP (i3elt, 0))
1781 && (REGNO (XEXP (i3elt, 0)) == regno
1782 ? reg_set_between_p (XEXP (elt, 0),
1783 PREV_INSN (insn), i3)
1784 : regno >= FIRST_PSEUDO_REGISTER))
1785 return 0;
1787 while (--i >= 0);
1789 break;
1791 /* We can ignore CLOBBERs. */
1792 case CLOBBER:
1793 break;
1795 case SET:
1796 /* Ignore SETs whose result isn't used but not those that
1797 have side-effects. */
1798 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1799 && insn_nothrow_p (insn)
1800 && !side_effects_p (elt))
1801 break;
1803 /* If we have already found a SET, this is a second one and
1804 so we cannot combine with this insn. */
1805 if (set)
1806 return 0;
1808 set = elt;
1809 break;
1811 default:
1812 /* Anything else means we can't combine. */
1813 return 0;
1817 if (set == 0
1818 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1819 so don't do anything with it. */
1820 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1821 return 0;
1823 else
1824 return 0;
1826 if (set == 0)
1827 return 0;
1829 /* The simplification in expand_field_assignment may call back to
1830 get_last_value, so set safe guard here. */
1831 subst_low_luid = DF_INSN_LUID (insn);
1833 set = expand_field_assignment (set);
1834 src = SET_SRC (set), dest = SET_DEST (set);
1836 /* Don't eliminate a store in the stack pointer. */
1837 if (dest == stack_pointer_rtx
1838 /* Don't combine with an insn that sets a register to itself if it has
1839 a REG_EQUAL note. This may be part of a LIBCALL sequence. */
1840 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1841 /* Can't merge an ASM_OPERANDS. */
1842 || GET_CODE (src) == ASM_OPERANDS
1843 /* Can't merge a function call. */
1844 || GET_CODE (src) == CALL
1845 /* Don't eliminate a function call argument. */
1846 || (CALL_P (i3)
1847 && (find_reg_fusage (i3, USE, dest)
1848 || (REG_P (dest)
1849 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1850 && global_regs[REGNO (dest)])))
1851 /* Don't substitute into an incremented register. */
1852 || FIND_REG_INC_NOTE (i3, dest)
1853 || (succ && FIND_REG_INC_NOTE (succ, dest))
1854 || (succ2 && FIND_REG_INC_NOTE (succ2, dest))
1855 /* Don't substitute into a non-local goto, this confuses CFG. */
1856 || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX))
1857 /* Make sure that DEST is not used after SUCC but before I3. */
1858 || (!all_adjacent
1859 && ((succ2
1860 && (reg_used_between_p (dest, succ2, i3)
1861 || reg_used_between_p (dest, succ, succ2)))
1862 || (!succ2 && succ && reg_used_between_p (dest, succ, i3))))
1863 /* Make sure that the value that is to be substituted for the register
1864 does not use any registers whose values alter in between. However,
1865 If the insns are adjacent, a use can't cross a set even though we
1866 think it might (this can happen for a sequence of insns each setting
1867 the same destination; last_set of that register might point to
1868 a NOTE). If INSN has a REG_EQUIV note, the register is always
1869 equivalent to the memory so the substitution is valid even if there
1870 are intervening stores. Also, don't move a volatile asm or
1871 UNSPEC_VOLATILE across any other insns. */
1872 || (! all_adjacent
1873 && (((!MEM_P (src)
1874 || ! find_reg_note (insn, REG_EQUIV, src))
1875 && use_crosses_set_p (src, DF_INSN_LUID (insn)))
1876 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1877 || GET_CODE (src) == UNSPEC_VOLATILE))
1878 /* Don't combine across a CALL_INSN, because that would possibly
1879 change whether the life span of some REGs crosses calls or not,
1880 and it is a pain to update that information.
1881 Exception: if source is a constant, moving it later can't hurt.
1882 Accept that as a special case. */
1883 || (DF_INSN_LUID (insn) < last_call_luid && ! CONSTANT_P (src)))
1884 return 0;
1886 /* DEST must either be a REG or CC0. */
1887 if (REG_P (dest))
1889 /* If register alignment is being enforced for multi-word items in all
1890 cases except for parameters, it is possible to have a register copy
1891 insn referencing a hard register that is not allowed to contain the
1892 mode being copied and which would not be valid as an operand of most
1893 insns. Eliminate this problem by not combining with such an insn.
1895 Also, on some machines we don't want to extend the life of a hard
1896 register. */
1898 if (REG_P (src)
1899 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1900 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1901 /* Don't extend the life of a hard register unless it is
1902 user variable (if we have few registers) or it can't
1903 fit into the desired register (meaning something special
1904 is going on).
1905 Also avoid substituting a return register into I3, because
1906 reload can't handle a conflict with constraints of other
1907 inputs. */
1908 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1909 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1910 return 0;
1912 else if (GET_CODE (dest) != CC0)
1913 return 0;
1916 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1917 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1918 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER)
1920 /* Don't substitute for a register intended as a clobberable
1921 operand. */
1922 rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0);
1923 if (rtx_equal_p (reg, dest))
1924 return 0;
1926 /* If the clobber represents an earlyclobber operand, we must not
1927 substitute an expression containing the clobbered register.
1928 As we do not analyze the constraint strings here, we have to
1929 make the conservative assumption. However, if the register is
1930 a fixed hard reg, the clobber cannot represent any operand;
1931 we leave it up to the machine description to either accept or
1932 reject use-and-clobber patterns. */
1933 if (!REG_P (reg)
1934 || REGNO (reg) >= FIRST_PSEUDO_REGISTER
1935 || !fixed_regs[REGNO (reg)])
1936 if (reg_overlap_mentioned_p (reg, src))
1937 return 0;
1940 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1941 or not), reject, unless nothing volatile comes between it and I3 */
1943 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1945 /* Make sure neither succ nor succ2 contains a volatile reference. */
1946 if (succ2 != 0 && volatile_refs_p (PATTERN (succ2)))
1947 return 0;
1948 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1949 return 0;
1950 /* We'll check insns between INSN and I3 below. */
1953 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1954 to be an explicit register variable, and was chosen for a reason. */
1956 if (GET_CODE (src) == ASM_OPERANDS
1957 && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1958 return 0;
1960 /* If INSN contains volatile references (specifically volatile MEMs),
1961 we cannot combine across any other volatile references.
1962 Even if INSN doesn't contain volatile references, any intervening
1963 volatile insn might affect machine state. */
1965 is_volatile_p = volatile_refs_p (PATTERN (insn))
1966 ? volatile_refs_p
1967 : volatile_insn_p;
1969 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1970 if (INSN_P (p) && p != succ && p != succ2 && is_volatile_p (PATTERN (p)))
1971 return 0;
1973 /* If INSN contains an autoincrement or autodecrement, make sure that
1974 register is not used between there and I3, and not already used in
1975 I3 either. Neither must it be used in PRED or SUCC, if they exist.
1976 Also insist that I3 not be a jump; if it were one
1977 and the incremented register were spilled, we would lose. */
1979 #ifdef AUTO_INC_DEC
1980 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1981 if (REG_NOTE_KIND (link) == REG_INC
1982 && (JUMP_P (i3)
1983 || reg_used_between_p (XEXP (link, 0), insn, i3)
1984 || (pred != NULL_RTX
1985 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred)))
1986 || (pred2 != NULL_RTX
1987 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred2)))
1988 || (succ != NULL_RTX
1989 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ)))
1990 || (succ2 != NULL_RTX
1991 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ2)))
1992 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1993 return 0;
1994 #endif
1996 #ifdef HAVE_cc0
1997 /* Don't combine an insn that follows a CC0-setting insn.
1998 An insn that uses CC0 must not be separated from the one that sets it.
1999 We do, however, allow I2 to follow a CC0-setting insn if that insn
2000 is passed as I1; in that case it will be deleted also.
2001 We also allow combining in this case if all the insns are adjacent
2002 because that would leave the two CC0 insns adjacent as well.
2003 It would be more logical to test whether CC0 occurs inside I1 or I2,
2004 but that would be much slower, and this ought to be equivalent. */
2006 p = prev_nonnote_insn (insn);
2007 if (p && p != pred && NONJUMP_INSN_P (p) && sets_cc0_p (PATTERN (p))
2008 && ! all_adjacent)
2009 return 0;
2010 #endif
2012 /* If we get here, we have passed all the tests and the combination is
2013 to be allowed. */
2015 *pdest = dest;
2016 *psrc = src;
2018 return 1;
2021 /* LOC is the location within I3 that contains its pattern or the component
2022 of a PARALLEL of the pattern. We validate that it is valid for combining.
2024 One problem is if I3 modifies its output, as opposed to replacing it
2025 entirely, we can't allow the output to contain I2DEST, I1DEST or I0DEST as
2026 doing so would produce an insn that is not equivalent to the original insns.
2028 Consider:
2030 (set (reg:DI 101) (reg:DI 100))
2031 (set (subreg:SI (reg:DI 101) 0) <foo>)
2033 This is NOT equivalent to:
2035 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
2036 (set (reg:DI 101) (reg:DI 100))])
2038 Not only does this modify 100 (in which case it might still be valid
2039 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
2041 We can also run into a problem if I2 sets a register that I1
2042 uses and I1 gets directly substituted into I3 (not via I2). In that
2043 case, we would be getting the wrong value of I2DEST into I3, so we
2044 must reject the combination. This case occurs when I2 and I1 both
2045 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
2046 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
2047 of a SET must prevent combination from occurring. The same situation
2048 can occur for I0, in which case I0_NOT_IN_SRC is set.
2050 Before doing the above check, we first try to expand a field assignment
2051 into a set of logical operations.
2053 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
2054 we place a register that is both set and used within I3. If more than one
2055 such register is detected, we fail.
2057 Return 1 if the combination is valid, zero otherwise. */
2059 static int
2060 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest, rtx i0dest,
2061 int i1_not_in_src, int i0_not_in_src, rtx *pi3dest_killed)
2063 rtx x = *loc;
2065 if (GET_CODE (x) == SET)
2067 rtx set = x ;
2068 rtx dest = SET_DEST (set);
2069 rtx src = SET_SRC (set);
2070 rtx inner_dest = dest;
2071 rtx subdest;
2073 while (GET_CODE (inner_dest) == STRICT_LOW_PART
2074 || GET_CODE (inner_dest) == SUBREG
2075 || GET_CODE (inner_dest) == ZERO_EXTRACT)
2076 inner_dest = XEXP (inner_dest, 0);
2078 /* Check for the case where I3 modifies its output, as discussed
2079 above. We don't want to prevent pseudos from being combined
2080 into the address of a MEM, so only prevent the combination if
2081 i1 or i2 set the same MEM. */
2082 if ((inner_dest != dest &&
2083 (!MEM_P (inner_dest)
2084 || rtx_equal_p (i2dest, inner_dest)
2085 || (i1dest && rtx_equal_p (i1dest, inner_dest))
2086 || (i0dest && rtx_equal_p (i0dest, inner_dest)))
2087 && (reg_overlap_mentioned_p (i2dest, inner_dest)
2088 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))
2089 || (i0dest && reg_overlap_mentioned_p (i0dest, inner_dest))))
2091 /* This is the same test done in can_combine_p except we can't test
2092 all_adjacent; we don't have to, since this instruction will stay
2093 in place, thus we are not considering increasing the lifetime of
2094 INNER_DEST.
2096 Also, if this insn sets a function argument, combining it with
2097 something that might need a spill could clobber a previous
2098 function argument; the all_adjacent test in can_combine_p also
2099 checks this; here, we do a more specific test for this case. */
2101 || (REG_P (inner_dest)
2102 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
2103 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
2104 GET_MODE (inner_dest))))
2105 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src))
2106 || (i0_not_in_src && reg_overlap_mentioned_p (i0dest, src)))
2107 return 0;
2109 /* If DEST is used in I3, it is being killed in this insn, so
2110 record that for later. We have to consider paradoxical
2111 subregs here, since they kill the whole register, but we
2112 ignore partial subregs, STRICT_LOW_PART, etc.
2113 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
2114 STACK_POINTER_REGNUM, since these are always considered to be
2115 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
2116 subdest = dest;
2117 if (GET_CODE (subdest) == SUBREG
2118 && (GET_MODE_SIZE (GET_MODE (subdest))
2119 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (subdest)))))
2120 subdest = SUBREG_REG (subdest);
2121 if (pi3dest_killed
2122 && REG_P (subdest)
2123 && reg_referenced_p (subdest, PATTERN (i3))
2124 && REGNO (subdest) != FRAME_POINTER_REGNUM
2125 #if !HARD_FRAME_POINTER_IS_FRAME_POINTER
2126 && REGNO (subdest) != HARD_FRAME_POINTER_REGNUM
2127 #endif
2128 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2129 && (REGNO (subdest) != ARG_POINTER_REGNUM
2130 || ! fixed_regs [REGNO (subdest)])
2131 #endif
2132 && REGNO (subdest) != STACK_POINTER_REGNUM)
2134 if (*pi3dest_killed)
2135 return 0;
2137 *pi3dest_killed = subdest;
2141 else if (GET_CODE (x) == PARALLEL)
2143 int i;
2145 for (i = 0; i < XVECLEN (x, 0); i++)
2146 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest, i0dest,
2147 i1_not_in_src, i0_not_in_src, pi3dest_killed))
2148 return 0;
2151 return 1;
2154 /* Return 1 if X is an arithmetic expression that contains a multiplication
2155 and division. We don't count multiplications by powers of two here. */
2157 static int
2158 contains_muldiv (rtx x)
2160 switch (GET_CODE (x))
2162 case MOD: case DIV: case UMOD: case UDIV:
2163 return 1;
2165 case MULT:
2166 return ! (CONST_INT_P (XEXP (x, 1))
2167 && exact_log2 (UINTVAL (XEXP (x, 1))) >= 0);
2168 default:
2169 if (BINARY_P (x))
2170 return contains_muldiv (XEXP (x, 0))
2171 || contains_muldiv (XEXP (x, 1));
2173 if (UNARY_P (x))
2174 return contains_muldiv (XEXP (x, 0));
2176 return 0;
2180 /* Determine whether INSN can be used in a combination. Return nonzero if
2181 not. This is used in try_combine to detect early some cases where we
2182 can't perform combinations. */
2184 static int
2185 cant_combine_insn_p (rtx insn)
2187 rtx set;
2188 rtx src, dest;
2190 /* If this isn't really an insn, we can't do anything.
2191 This can occur when flow deletes an insn that it has merged into an
2192 auto-increment address. */
2193 if (! INSN_P (insn))
2194 return 1;
2196 /* Never combine loads and stores involving hard regs that are likely
2197 to be spilled. The register allocator can usually handle such
2198 reg-reg moves by tying. If we allow the combiner to make
2199 substitutions of likely-spilled regs, reload might die.
2200 As an exception, we allow combinations involving fixed regs; these are
2201 not available to the register allocator so there's no risk involved. */
2203 set = single_set (insn);
2204 if (! set)
2205 return 0;
2206 src = SET_SRC (set);
2207 dest = SET_DEST (set);
2208 if (GET_CODE (src) == SUBREG)
2209 src = SUBREG_REG (src);
2210 if (GET_CODE (dest) == SUBREG)
2211 dest = SUBREG_REG (dest);
2212 if (REG_P (src) && REG_P (dest)
2213 && ((HARD_REGISTER_P (src)
2214 && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (src))
2215 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (src))))
2216 || (HARD_REGISTER_P (dest)
2217 && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (dest))
2218 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (dest))))))
2219 return 1;
2221 return 0;
2224 struct likely_spilled_retval_info
2226 unsigned regno, nregs;
2227 unsigned mask;
2230 /* Called via note_stores by likely_spilled_retval_p. Remove from info->mask
2231 hard registers that are known to be written to / clobbered in full. */
2232 static void
2233 likely_spilled_retval_1 (rtx x, const_rtx set, void *data)
2235 struct likely_spilled_retval_info *const info =
2236 (struct likely_spilled_retval_info *) data;
2237 unsigned regno, nregs;
2238 unsigned new_mask;
2240 if (!REG_P (XEXP (set, 0)))
2241 return;
2242 regno = REGNO (x);
2243 if (regno >= info->regno + info->nregs)
2244 return;
2245 nregs = hard_regno_nregs[regno][GET_MODE (x)];
2246 if (regno + nregs <= info->regno)
2247 return;
2248 new_mask = (2U << (nregs - 1)) - 1;
2249 if (regno < info->regno)
2250 new_mask >>= info->regno - regno;
2251 else
2252 new_mask <<= regno - info->regno;
2253 info->mask &= ~new_mask;
2256 /* Return nonzero iff part of the return value is live during INSN, and
2257 it is likely spilled. This can happen when more than one insn is needed
2258 to copy the return value, e.g. when we consider to combine into the
2259 second copy insn for a complex value. */
2261 static int
2262 likely_spilled_retval_p (rtx insn)
2264 rtx use = BB_END (this_basic_block);
2265 rtx reg, p;
2266 unsigned regno, nregs;
2267 /* We assume here that no machine mode needs more than
2268 32 hard registers when the value overlaps with a register
2269 for which TARGET_FUNCTION_VALUE_REGNO_P is true. */
2270 unsigned mask;
2271 struct likely_spilled_retval_info info;
2273 if (!NONJUMP_INSN_P (use) || GET_CODE (PATTERN (use)) != USE || insn == use)
2274 return 0;
2275 reg = XEXP (PATTERN (use), 0);
2276 if (!REG_P (reg) || !targetm.calls.function_value_regno_p (REGNO (reg)))
2277 return 0;
2278 regno = REGNO (reg);
2279 nregs = hard_regno_nregs[regno][GET_MODE (reg)];
2280 if (nregs == 1)
2281 return 0;
2282 mask = (2U << (nregs - 1)) - 1;
2284 /* Disregard parts of the return value that are set later. */
2285 info.regno = regno;
2286 info.nregs = nregs;
2287 info.mask = mask;
2288 for (p = PREV_INSN (use); info.mask && p != insn; p = PREV_INSN (p))
2289 if (INSN_P (p))
2290 note_stores (PATTERN (p), likely_spilled_retval_1, &info);
2291 mask = info.mask;
2293 /* Check if any of the (probably) live return value registers is
2294 likely spilled. */
2295 nregs --;
2298 if ((mask & 1 << nregs)
2299 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno + nregs)))
2300 return 1;
2301 } while (nregs--);
2302 return 0;
2305 /* Adjust INSN after we made a change to its destination.
2307 Changing the destination can invalidate notes that say something about
2308 the results of the insn and a LOG_LINK pointing to the insn. */
2310 static void
2311 adjust_for_new_dest (rtx insn)
2313 /* For notes, be conservative and simply remove them. */
2314 remove_reg_equal_equiv_notes (insn);
2316 /* The new insn will have a destination that was previously the destination
2317 of an insn just above it. Call distribute_links to make a LOG_LINK from
2318 the next use of that destination. */
2319 distribute_links (alloc_insn_link (insn, NULL));
2321 df_insn_rescan (insn);
2324 /* Return TRUE if combine can reuse reg X in mode MODE.
2325 ADDED_SETS is nonzero if the original set is still required. */
2326 static bool
2327 can_change_dest_mode (rtx x, int added_sets, enum machine_mode mode)
2329 unsigned int regno;
2331 if (!REG_P(x))
2332 return false;
2334 regno = REGNO (x);
2335 /* Allow hard registers if the new mode is legal, and occupies no more
2336 registers than the old mode. */
2337 if (regno < FIRST_PSEUDO_REGISTER)
2338 return (HARD_REGNO_MODE_OK (regno, mode)
2339 && (hard_regno_nregs[regno][GET_MODE (x)]
2340 >= hard_regno_nregs[regno][mode]));
2342 /* Or a pseudo that is only used once. */
2343 return (REG_N_SETS (regno) == 1 && !added_sets
2344 && !REG_USERVAR_P (x));
2348 /* Check whether X, the destination of a set, refers to part of
2349 the register specified by REG. */
2351 static bool
2352 reg_subword_p (rtx x, rtx reg)
2354 /* Check that reg is an integer mode register. */
2355 if (!REG_P (reg) || GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT)
2356 return false;
2358 if (GET_CODE (x) == STRICT_LOW_PART
2359 || GET_CODE (x) == ZERO_EXTRACT)
2360 x = XEXP (x, 0);
2362 return GET_CODE (x) == SUBREG
2363 && SUBREG_REG (x) == reg
2364 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT;
2367 /* Delete the unconditional jump INSN and adjust the CFG correspondingly.
2368 Note that the INSN should be deleted *after* removing dead edges, so
2369 that the kept edge is the fallthrough edge for a (set (pc) (pc))
2370 but not for a (set (pc) (label_ref FOO)). */
2372 static void
2373 update_cfg_for_uncondjump (rtx insn)
2375 basic_block bb = BLOCK_FOR_INSN (insn);
2376 gcc_assert (BB_END (bb) == insn);
2378 purge_dead_edges (bb);
2380 delete_insn (insn);
2381 if (EDGE_COUNT (bb->succs) == 1)
2383 rtx insn;
2385 single_succ_edge (bb)->flags |= EDGE_FALLTHRU;
2387 /* Remove barriers from the footer if there are any. */
2388 for (insn = BB_FOOTER (bb); insn; insn = NEXT_INSN (insn))
2389 if (BARRIER_P (insn))
2391 if (PREV_INSN (insn))
2392 NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (insn);
2393 else
2394 BB_FOOTER (bb) = NEXT_INSN (insn);
2395 if (NEXT_INSN (insn))
2396 PREV_INSN (NEXT_INSN (insn)) = PREV_INSN (insn);
2398 else if (LABEL_P (insn))
2399 break;
2403 /* Try to combine the insns I0, I1 and I2 into I3.
2404 Here I0, I1 and I2 appear earlier than I3.
2405 I0 and I1 can be zero; then we combine just I2 into I3, or I1 and I2 into
2408 If we are combining more than two insns and the resulting insn is not
2409 recognized, try splitting it into two insns. If that happens, I2 and I3
2410 are retained and I1/I0 are pseudo-deleted by turning them into a NOTE.
2411 Otherwise, I0, I1 and I2 are pseudo-deleted.
2413 Return 0 if the combination does not work. Then nothing is changed.
2414 If we did the combination, return the insn at which combine should
2415 resume scanning.
2417 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
2418 new direct jump instruction.
2420 LAST_COMBINED_INSN is either I3, or some insn after I3 that has
2421 been I3 passed to an earlier try_combine within the same basic
2422 block. */
2424 static rtx
2425 try_combine (rtx i3, rtx i2, rtx i1, rtx i0, int *new_direct_jump_p,
2426 rtx last_combined_insn)
2428 /* New patterns for I3 and I2, respectively. */
2429 rtx newpat, newi2pat = 0;
2430 rtvec newpat_vec_with_clobbers = 0;
2431 int substed_i2 = 0, substed_i1 = 0, substed_i0 = 0;
2432 /* Indicates need to preserve SET in I0, I1 or I2 in I3 if it is not
2433 dead. */
2434 int added_sets_0, added_sets_1, added_sets_2;
2435 /* Total number of SETs to put into I3. */
2436 int total_sets;
2437 /* Nonzero if I2's or I1's body now appears in I3. */
2438 int i2_is_used = 0, i1_is_used = 0;
2439 /* INSN_CODEs for new I3, new I2, and user of condition code. */
2440 int insn_code_number, i2_code_number = 0, other_code_number = 0;
2441 /* Contains I3 if the destination of I3 is used in its source, which means
2442 that the old life of I3 is being killed. If that usage is placed into
2443 I2 and not in I3, a REG_DEAD note must be made. */
2444 rtx i3dest_killed = 0;
2445 /* SET_DEST and SET_SRC of I2, I1 and I0. */
2446 rtx i2dest = 0, i2src = 0, i1dest = 0, i1src = 0, i0dest = 0, i0src = 0;
2447 /* Copy of SET_SRC of I1 and I0, if needed. */
2448 rtx i1src_copy = 0, i0src_copy = 0, i0src_copy2 = 0;
2449 /* Set if I2DEST was reused as a scratch register. */
2450 bool i2scratch = false;
2451 /* The PATTERNs of I0, I1, and I2, or a copy of them in certain cases. */
2452 rtx i0pat = 0, i1pat = 0, i2pat = 0;
2453 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
2454 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
2455 int i0dest_in_i0src = 0, i1dest_in_i0src = 0, i2dest_in_i0src = 0;
2456 int i2dest_killed = 0, i1dest_killed = 0, i0dest_killed = 0;
2457 int i1_feeds_i2_n = 0, i0_feeds_i2_n = 0, i0_feeds_i1_n = 0;
2458 /* Notes that must be added to REG_NOTES in I3 and I2. */
2459 rtx new_i3_notes, new_i2_notes;
2460 /* Notes that we substituted I3 into I2 instead of the normal case. */
2461 int i3_subst_into_i2 = 0;
2462 /* Notes that I1, I2 or I3 is a MULT operation. */
2463 int have_mult = 0;
2464 int swap_i2i3 = 0;
2465 int changed_i3_dest = 0;
2467 int maxreg;
2468 rtx temp;
2469 struct insn_link *link;
2470 rtx other_pat = 0;
2471 rtx new_other_notes;
2472 int i;
2474 /* Only try four-insn combinations when there's high likelihood of
2475 success. Look for simple insns, such as loads of constants or
2476 binary operations involving a constant. */
2477 if (i0)
2479 int i;
2480 int ngood = 0;
2481 int nshift = 0;
2483 if (!flag_expensive_optimizations)
2484 return 0;
2486 for (i = 0; i < 4; i++)
2488 rtx insn = i == 0 ? i0 : i == 1 ? i1 : i == 2 ? i2 : i3;
2489 rtx set = single_set (insn);
2490 rtx src;
2491 if (!set)
2492 continue;
2493 src = SET_SRC (set);
2494 if (CONSTANT_P (src))
2496 ngood += 2;
2497 break;
2499 else if (BINARY_P (src) && CONSTANT_P (XEXP (src, 1)))
2500 ngood++;
2501 else if (GET_CODE (src) == ASHIFT || GET_CODE (src) == ASHIFTRT
2502 || GET_CODE (src) == LSHIFTRT)
2503 nshift++;
2505 if (ngood < 2 && nshift < 2)
2506 return 0;
2509 /* Exit early if one of the insns involved can't be used for
2510 combinations. */
2511 if (cant_combine_insn_p (i3)
2512 || cant_combine_insn_p (i2)
2513 || (i1 && cant_combine_insn_p (i1))
2514 || (i0 && cant_combine_insn_p (i0))
2515 || likely_spilled_retval_p (i3))
2516 return 0;
2518 combine_attempts++;
2519 undobuf.other_insn = 0;
2521 /* Reset the hard register usage information. */
2522 CLEAR_HARD_REG_SET (newpat_used_regs);
2524 if (dump_file && (dump_flags & TDF_DETAILS))
2526 if (i0)
2527 fprintf (dump_file, "\nTrying %d, %d, %d -> %d:\n",
2528 INSN_UID (i0), INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
2529 else if (i1)
2530 fprintf (dump_file, "\nTrying %d, %d -> %d:\n",
2531 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
2532 else
2533 fprintf (dump_file, "\nTrying %d -> %d:\n",
2534 INSN_UID (i2), INSN_UID (i3));
2537 /* If multiple insns feed into one of I2 or I3, they can be in any
2538 order. To simplify the code below, reorder them in sequence. */
2539 if (i0 && DF_INSN_LUID (i0) > DF_INSN_LUID (i2))
2540 temp = i2, i2 = i0, i0 = temp;
2541 if (i0 && DF_INSN_LUID (i0) > DF_INSN_LUID (i1))
2542 temp = i1, i1 = i0, i0 = temp;
2543 if (i1 && DF_INSN_LUID (i1) > DF_INSN_LUID (i2))
2544 temp = i1, i1 = i2, i2 = temp;
2546 added_links_insn = 0;
2548 /* First check for one important special case that the code below will
2549 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
2550 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
2551 we may be able to replace that destination with the destination of I3.
2552 This occurs in the common code where we compute both a quotient and
2553 remainder into a structure, in which case we want to do the computation
2554 directly into the structure to avoid register-register copies.
2556 Note that this case handles both multiple sets in I2 and also cases
2557 where I2 has a number of CLOBBERs inside the PARALLEL.
2559 We make very conservative checks below and only try to handle the
2560 most common cases of this. For example, we only handle the case
2561 where I2 and I3 are adjacent to avoid making difficult register
2562 usage tests. */
2564 if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET
2565 && REG_P (SET_SRC (PATTERN (i3)))
2566 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
2567 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
2568 && GET_CODE (PATTERN (i2)) == PARALLEL
2569 && ! side_effects_p (SET_DEST (PATTERN (i3)))
2570 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
2571 below would need to check what is inside (and reg_overlap_mentioned_p
2572 doesn't support those codes anyway). Don't allow those destinations;
2573 the resulting insn isn't likely to be recognized anyway. */
2574 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
2575 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
2576 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
2577 SET_DEST (PATTERN (i3)))
2578 && next_active_insn (i2) == i3)
2580 rtx p2 = PATTERN (i2);
2582 /* Make sure that the destination of I3,
2583 which we are going to substitute into one output of I2,
2584 is not used within another output of I2. We must avoid making this:
2585 (parallel [(set (mem (reg 69)) ...)
2586 (set (reg 69) ...)])
2587 which is not well-defined as to order of actions.
2588 (Besides, reload can't handle output reloads for this.)
2590 The problem can also happen if the dest of I3 is a memory ref,
2591 if another dest in I2 is an indirect memory ref. */
2592 for (i = 0; i < XVECLEN (p2, 0); i++)
2593 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
2594 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
2595 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
2596 SET_DEST (XVECEXP (p2, 0, i))))
2597 break;
2599 if (i == XVECLEN (p2, 0))
2600 for (i = 0; i < XVECLEN (p2, 0); i++)
2601 if (GET_CODE (XVECEXP (p2, 0, i)) == SET
2602 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
2604 combine_merges++;
2606 subst_insn = i3;
2607 subst_low_luid = DF_INSN_LUID (i2);
2609 added_sets_2 = added_sets_1 = added_sets_0 = 0;
2610 i2src = SET_SRC (XVECEXP (p2, 0, i));
2611 i2dest = SET_DEST (XVECEXP (p2, 0, i));
2612 i2dest_killed = dead_or_set_p (i2, i2dest);
2614 /* Replace the dest in I2 with our dest and make the resulting
2615 insn the new pattern for I3. Then skip to where we validate
2616 the pattern. Everything was set up above. */
2617 SUBST (SET_DEST (XVECEXP (p2, 0, i)), SET_DEST (PATTERN (i3)));
2618 newpat = p2;
2619 i3_subst_into_i2 = 1;
2620 goto validate_replacement;
2624 /* If I2 is setting a pseudo to a constant and I3 is setting some
2625 sub-part of it to another constant, merge them by making a new
2626 constant. */
2627 if (i1 == 0
2628 && (temp = single_set (i2)) != 0
2629 && CONST_SCALAR_INT_P (SET_SRC (temp))
2630 && GET_CODE (PATTERN (i3)) == SET
2631 && CONST_SCALAR_INT_P (SET_SRC (PATTERN (i3)))
2632 && reg_subword_p (SET_DEST (PATTERN (i3)), SET_DEST (temp)))
2634 rtx dest = SET_DEST (PATTERN (i3));
2635 int offset = -1;
2636 int width = 0;
2638 if (GET_CODE (dest) == ZERO_EXTRACT)
2640 if (CONST_INT_P (XEXP (dest, 1))
2641 && CONST_INT_P (XEXP (dest, 2)))
2643 width = INTVAL (XEXP (dest, 1));
2644 offset = INTVAL (XEXP (dest, 2));
2645 dest = XEXP (dest, 0);
2646 if (BITS_BIG_ENDIAN)
2647 offset = GET_MODE_PRECISION (GET_MODE (dest)) - width - offset;
2650 else
2652 if (GET_CODE (dest) == STRICT_LOW_PART)
2653 dest = XEXP (dest, 0);
2654 width = GET_MODE_PRECISION (GET_MODE (dest));
2655 offset = 0;
2658 if (offset >= 0)
2660 /* If this is the low part, we're done. */
2661 if (subreg_lowpart_p (dest))
2663 /* Handle the case where inner is twice the size of outer. */
2664 else if (GET_MODE_PRECISION (GET_MODE (SET_DEST (temp)))
2665 == 2 * GET_MODE_PRECISION (GET_MODE (dest)))
2666 offset += GET_MODE_PRECISION (GET_MODE (dest));
2667 /* Otherwise give up for now. */
2668 else
2669 offset = -1;
2672 if (offset >= 0
2673 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (temp)))
2674 <= HOST_BITS_PER_DOUBLE_INT))
2676 double_int m, o, i;
2677 rtx inner = SET_SRC (PATTERN (i3));
2678 rtx outer = SET_SRC (temp);
2680 o = rtx_to_double_int (outer);
2681 i = rtx_to_double_int (inner);
2683 m = double_int::mask (width);
2684 i &= m;
2685 m = m.llshift (offset, HOST_BITS_PER_DOUBLE_INT);
2686 i = i.llshift (offset, HOST_BITS_PER_DOUBLE_INT);
2687 o = o.and_not (m) | i;
2689 combine_merges++;
2690 subst_insn = i3;
2691 subst_low_luid = DF_INSN_LUID (i2);
2692 added_sets_2 = added_sets_1 = added_sets_0 = 0;
2693 i2dest = SET_DEST (temp);
2694 i2dest_killed = dead_or_set_p (i2, i2dest);
2696 /* Replace the source in I2 with the new constant and make the
2697 resulting insn the new pattern for I3. Then skip to where we
2698 validate the pattern. Everything was set up above. */
2699 SUBST (SET_SRC (temp),
2700 immed_double_int_const (o, GET_MODE (SET_DEST (temp))));
2702 newpat = PATTERN (i2);
2704 /* The dest of I3 has been replaced with the dest of I2. */
2705 changed_i3_dest = 1;
2706 goto validate_replacement;
2710 #ifndef HAVE_cc0
2711 /* If we have no I1 and I2 looks like:
2712 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
2713 (set Y OP)])
2714 make up a dummy I1 that is
2715 (set Y OP)
2716 and change I2 to be
2717 (set (reg:CC X) (compare:CC Y (const_int 0)))
2719 (We can ignore any trailing CLOBBERs.)
2721 This undoes a previous combination and allows us to match a branch-and-
2722 decrement insn. */
2724 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
2725 && XVECLEN (PATTERN (i2), 0) >= 2
2726 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
2727 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
2728 == MODE_CC)
2729 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
2730 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
2731 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
2732 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)))
2733 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
2734 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
2736 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
2737 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
2738 break;
2740 if (i == 1)
2742 /* We make I1 with the same INSN_UID as I2. This gives it
2743 the same DF_INSN_LUID for value tracking. Our fake I1 will
2744 never appear in the insn stream so giving it the same INSN_UID
2745 as I2 will not cause a problem. */
2747 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
2748 BLOCK_FOR_INSN (i2), XVECEXP (PATTERN (i2), 0, 1),
2749 INSN_LOCATION (i2), -1, NULL_RTX);
2751 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
2752 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
2753 SET_DEST (PATTERN (i1)));
2754 SUBST_LINK (LOG_LINKS (i2), alloc_insn_link (i1, LOG_LINKS (i2)));
2757 #endif
2759 /* Verify that I2 and I1 are valid for combining. */
2760 if (! can_combine_p (i2, i3, i0, i1, NULL_RTX, NULL_RTX, &i2dest, &i2src)
2761 || (i1 && ! can_combine_p (i1, i3, i0, NULL_RTX, i2, NULL_RTX,
2762 &i1dest, &i1src))
2763 || (i0 && ! can_combine_p (i0, i3, NULL_RTX, NULL_RTX, i1, i2,
2764 &i0dest, &i0src)))
2766 undo_all ();
2767 return 0;
2770 /* Record whether I2DEST is used in I2SRC and similarly for the other
2771 cases. Knowing this will help in register status updating below. */
2772 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
2773 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
2774 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
2775 i0dest_in_i0src = i0 && reg_overlap_mentioned_p (i0dest, i0src);
2776 i1dest_in_i0src = i0 && reg_overlap_mentioned_p (i1dest, i0src);
2777 i2dest_in_i0src = i0 && reg_overlap_mentioned_p (i2dest, i0src);
2778 i2dest_killed = dead_or_set_p (i2, i2dest);
2779 i1dest_killed = i1 && dead_or_set_p (i1, i1dest);
2780 i0dest_killed = i0 && dead_or_set_p (i0, i0dest);
2782 /* For the earlier insns, determine which of the subsequent ones they
2783 feed. */
2784 i1_feeds_i2_n = i1 && insn_a_feeds_b (i1, i2);
2785 i0_feeds_i1_n = i0 && insn_a_feeds_b (i0, i1);
2786 i0_feeds_i2_n = (i0 && (!i0_feeds_i1_n ? insn_a_feeds_b (i0, i2)
2787 : (!reg_overlap_mentioned_p (i1dest, i0dest)
2788 && reg_overlap_mentioned_p (i0dest, i2src))));
2790 /* Ensure that I3's pattern can be the destination of combines. */
2791 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest, i0dest,
2792 i1 && i2dest_in_i1src && !i1_feeds_i2_n,
2793 i0 && ((i2dest_in_i0src && !i0_feeds_i2_n)
2794 || (i1dest_in_i0src && !i0_feeds_i1_n)),
2795 &i3dest_killed))
2797 undo_all ();
2798 return 0;
2801 /* See if any of the insns is a MULT operation. Unless one is, we will
2802 reject a combination that is, since it must be slower. Be conservative
2803 here. */
2804 if (GET_CODE (i2src) == MULT
2805 || (i1 != 0 && GET_CODE (i1src) == MULT)
2806 || (i0 != 0 && GET_CODE (i0src) == MULT)
2807 || (GET_CODE (PATTERN (i3)) == SET
2808 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
2809 have_mult = 1;
2811 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
2812 We used to do this EXCEPT in one case: I3 has a post-inc in an
2813 output operand. However, that exception can give rise to insns like
2814 mov r3,(r3)+
2815 which is a famous insn on the PDP-11 where the value of r3 used as the
2816 source was model-dependent. Avoid this sort of thing. */
2818 #if 0
2819 if (!(GET_CODE (PATTERN (i3)) == SET
2820 && REG_P (SET_SRC (PATTERN (i3)))
2821 && MEM_P (SET_DEST (PATTERN (i3)))
2822 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
2823 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
2824 /* It's not the exception. */
2825 #endif
2826 #ifdef AUTO_INC_DEC
2828 rtx link;
2829 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
2830 if (REG_NOTE_KIND (link) == REG_INC
2831 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
2832 || (i1 != 0
2833 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
2835 undo_all ();
2836 return 0;
2839 #endif
2841 /* See if the SETs in I1 or I2 need to be kept around in the merged
2842 instruction: whenever the value set there is still needed past I3.
2843 For the SET in I2, this is easy: we see if I2DEST dies or is set in I3.
2845 For the SET in I1, we have two cases: if I1 and I2 independently feed
2846 into I3, the set in I1 needs to be kept around unless I1DEST dies
2847 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
2848 in I1 needs to be kept around unless I1DEST dies or is set in either
2849 I2 or I3. The same considerations apply to I0. */
2851 added_sets_2 = !dead_or_set_p (i3, i2dest);
2853 if (i1)
2854 added_sets_1 = !(dead_or_set_p (i3, i1dest)
2855 || (i1_feeds_i2_n && dead_or_set_p (i2, i1dest)));
2856 else
2857 added_sets_1 = 0;
2859 if (i0)
2860 added_sets_0 = !(dead_or_set_p (i3, i0dest)
2861 || (i0_feeds_i1_n && dead_or_set_p (i1, i0dest))
2862 || ((i0_feeds_i2_n || (i0_feeds_i1_n && i1_feeds_i2_n))
2863 && dead_or_set_p (i2, i0dest)));
2864 else
2865 added_sets_0 = 0;
2867 /* We are about to copy insns for the case where they need to be kept
2868 around. Check that they can be copied in the merged instruction. */
2870 if (targetm.cannot_copy_insn_p
2871 && ((added_sets_2 && targetm.cannot_copy_insn_p (i2))
2872 || (i1 && added_sets_1 && targetm.cannot_copy_insn_p (i1))
2873 || (i0 && added_sets_0 && targetm.cannot_copy_insn_p (i0))))
2875 undo_all ();
2876 return 0;
2879 /* If the set in I2 needs to be kept around, we must make a copy of
2880 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
2881 PATTERN (I2), we are only substituting for the original I1DEST, not into
2882 an already-substituted copy. This also prevents making self-referential
2883 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
2884 I2DEST. */
2886 if (added_sets_2)
2888 if (GET_CODE (PATTERN (i2)) == PARALLEL)
2889 i2pat = gen_rtx_SET (VOIDmode, i2dest, copy_rtx (i2src));
2890 else
2891 i2pat = copy_rtx (PATTERN (i2));
2894 if (added_sets_1)
2896 if (GET_CODE (PATTERN (i1)) == PARALLEL)
2897 i1pat = gen_rtx_SET (VOIDmode, i1dest, copy_rtx (i1src));
2898 else
2899 i1pat = copy_rtx (PATTERN (i1));
2902 if (added_sets_0)
2904 if (GET_CODE (PATTERN (i0)) == PARALLEL)
2905 i0pat = gen_rtx_SET (VOIDmode, i0dest, copy_rtx (i0src));
2906 else
2907 i0pat = copy_rtx (PATTERN (i0));
2910 combine_merges++;
2912 /* Substitute in the latest insn for the regs set by the earlier ones. */
2914 maxreg = max_reg_num ();
2916 subst_insn = i3;
2918 #ifndef HAVE_cc0
2919 /* Many machines that don't use CC0 have insns that can both perform an
2920 arithmetic operation and set the condition code. These operations will
2921 be represented as a PARALLEL with the first element of the vector
2922 being a COMPARE of an arithmetic operation with the constant zero.
2923 The second element of the vector will set some pseudo to the result
2924 of the same arithmetic operation. If we simplify the COMPARE, we won't
2925 match such a pattern and so will generate an extra insn. Here we test
2926 for this case, where both the comparison and the operation result are
2927 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
2928 I2SRC. Later we will make the PARALLEL that contains I2. */
2930 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
2931 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
2932 && CONST_INT_P (XEXP (SET_SRC (PATTERN (i3)), 1))
2933 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
2935 rtx newpat_dest;
2936 rtx *cc_use_loc = NULL, cc_use_insn = NULL_RTX;
2937 rtx op0 = i2src, op1 = XEXP (SET_SRC (PATTERN (i3)), 1);
2938 enum machine_mode compare_mode, orig_compare_mode;
2939 enum rtx_code compare_code = UNKNOWN, orig_compare_code = UNKNOWN;
2941 newpat = PATTERN (i3);
2942 newpat_dest = SET_DEST (newpat);
2943 compare_mode = orig_compare_mode = GET_MODE (newpat_dest);
2945 if (undobuf.other_insn == 0
2946 && (cc_use_loc = find_single_use (SET_DEST (newpat), i3,
2947 &cc_use_insn)))
2949 compare_code = orig_compare_code = GET_CODE (*cc_use_loc);
2950 compare_code = simplify_compare_const (compare_code,
2951 op0, &op1);
2952 target_canonicalize_comparison (&compare_code, &op0, &op1, 1);
2955 /* Do the rest only if op1 is const0_rtx, which may be the
2956 result of simplification. */
2957 if (op1 == const0_rtx)
2959 /* If a single use of the CC is found, prepare to modify it
2960 when SELECT_CC_MODE returns a new CC-class mode, or when
2961 the above simplify_compare_const() returned a new comparison
2962 operator. undobuf.other_insn is assigned the CC use insn
2963 when modifying it. */
2964 if (cc_use_loc)
2966 #ifdef SELECT_CC_MODE
2967 enum machine_mode new_mode
2968 = SELECT_CC_MODE (compare_code, op0, op1);
2969 if (new_mode != orig_compare_mode
2970 && can_change_dest_mode (SET_DEST (newpat),
2971 added_sets_2, new_mode))
2973 unsigned int regno = REGNO (newpat_dest);
2974 compare_mode = new_mode;
2975 if (regno < FIRST_PSEUDO_REGISTER)
2976 newpat_dest = gen_rtx_REG (compare_mode, regno);
2977 else
2979 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
2980 newpat_dest = regno_reg_rtx[regno];
2983 #endif
2984 /* Cases for modifying the CC-using comparison. */
2985 if (compare_code != orig_compare_code
2986 /* ??? Do we need to verify the zero rtx? */
2987 && XEXP (*cc_use_loc, 1) == const0_rtx)
2989 /* Replace cc_use_loc with entire new RTX. */
2990 SUBST (*cc_use_loc,
2991 gen_rtx_fmt_ee (compare_code, compare_mode,
2992 newpat_dest, const0_rtx));
2993 undobuf.other_insn = cc_use_insn;
2995 else if (compare_mode != orig_compare_mode)
2997 /* Just replace the CC reg with a new mode. */
2998 SUBST (XEXP (*cc_use_loc, 0), newpat_dest);
2999 undobuf.other_insn = cc_use_insn;
3003 /* Now we modify the current newpat:
3004 First, SET_DEST(newpat) is updated if the CC mode has been
3005 altered. For targets without SELECT_CC_MODE, this should be
3006 optimized away. */
3007 if (compare_mode != orig_compare_mode)
3008 SUBST (SET_DEST (newpat), newpat_dest);
3009 /* This is always done to propagate i2src into newpat. */
3010 SUBST (SET_SRC (newpat),
3011 gen_rtx_COMPARE (compare_mode, op0, op1));
3012 /* Create new version of i2pat if needed; the below PARALLEL
3013 creation needs this to work correctly. */
3014 if (! rtx_equal_p (i2src, op0))
3015 i2pat = gen_rtx_SET (VOIDmode, i2dest, op0);
3016 i2_is_used = 1;
3019 #endif
3021 if (i2_is_used == 0)
3023 /* It is possible that the source of I2 or I1 may be performing
3024 an unneeded operation, such as a ZERO_EXTEND of something
3025 that is known to have the high part zero. Handle that case
3026 by letting subst look at the inner insns.
3028 Another way to do this would be to have a function that tries
3029 to simplify a single insn instead of merging two or more
3030 insns. We don't do this because of the potential of infinite
3031 loops and because of the potential extra memory required.
3032 However, doing it the way we are is a bit of a kludge and
3033 doesn't catch all cases.
3035 But only do this if -fexpensive-optimizations since it slows
3036 things down and doesn't usually win.
3038 This is not done in the COMPARE case above because the
3039 unmodified I2PAT is used in the PARALLEL and so a pattern
3040 with a modified I2SRC would not match. */
3042 if (flag_expensive_optimizations)
3044 /* Pass pc_rtx so no substitutions are done, just
3045 simplifications. */
3046 if (i1)
3048 subst_low_luid = DF_INSN_LUID (i1);
3049 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0, 0);
3052 subst_low_luid = DF_INSN_LUID (i2);
3053 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0, 0);
3056 n_occurrences = 0; /* `subst' counts here */
3057 subst_low_luid = DF_INSN_LUID (i2);
3059 /* If I1 feeds into I2 and I1DEST is in I1SRC, we need to make a unique
3060 copy of I2SRC each time we substitute it, in order to avoid creating
3061 self-referential RTL when we will be substituting I1SRC for I1DEST
3062 later. Likewise if I0 feeds into I2, either directly or indirectly
3063 through I1, and I0DEST is in I0SRC. */
3064 newpat = subst (PATTERN (i3), i2dest, i2src, 0, 0,
3065 (i1_feeds_i2_n && i1dest_in_i1src)
3066 || ((i0_feeds_i2_n || (i0_feeds_i1_n && i1_feeds_i2_n))
3067 && i0dest_in_i0src));
3068 substed_i2 = 1;
3070 /* Record whether I2's body now appears within I3's body. */
3071 i2_is_used = n_occurrences;
3074 /* If we already got a failure, don't try to do more. Otherwise, try to
3075 substitute I1 if we have it. */
3077 if (i1 && GET_CODE (newpat) != CLOBBER)
3079 /* Check that an autoincrement side-effect on I1 has not been lost.
3080 This happens if I1DEST is mentioned in I2 and dies there, and
3081 has disappeared from the new pattern. */
3082 if ((FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
3083 && i1_feeds_i2_n
3084 && dead_or_set_p (i2, i1dest)
3085 && !reg_overlap_mentioned_p (i1dest, newpat))
3086 /* Before we can do this substitution, we must redo the test done
3087 above (see detailed comments there) that ensures I1DEST isn't
3088 mentioned in any SETs in NEWPAT that are field assignments. */
3089 || !combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX, NULL_RTX,
3090 0, 0, 0))
3092 undo_all ();
3093 return 0;
3096 n_occurrences = 0;
3097 subst_low_luid = DF_INSN_LUID (i1);
3099 /* If the following substitution will modify I1SRC, make a copy of it
3100 for the case where it is substituted for I1DEST in I2PAT later. */
3101 if (added_sets_2 && i1_feeds_i2_n)
3102 i1src_copy = copy_rtx (i1src);
3104 /* If I0 feeds into I1 and I0DEST is in I0SRC, we need to make a unique
3105 copy of I1SRC each time we substitute it, in order to avoid creating
3106 self-referential RTL when we will be substituting I0SRC for I0DEST
3107 later. */
3108 newpat = subst (newpat, i1dest, i1src, 0, 0,
3109 i0_feeds_i1_n && i0dest_in_i0src);
3110 substed_i1 = 1;
3112 /* Record whether I1's body now appears within I3's body. */
3113 i1_is_used = n_occurrences;
3116 /* Likewise for I0 if we have it. */
3118 if (i0 && GET_CODE (newpat) != CLOBBER)
3120 if ((FIND_REG_INC_NOTE (i0, NULL_RTX) != 0
3121 && ((i0_feeds_i2_n && dead_or_set_p (i2, i0dest))
3122 || (i0_feeds_i1_n && dead_or_set_p (i1, i0dest)))
3123 && !reg_overlap_mentioned_p (i0dest, newpat))
3124 || !combinable_i3pat (NULL_RTX, &newpat, i0dest, NULL_RTX, NULL_RTX,
3125 0, 0, 0))
3127 undo_all ();
3128 return 0;
3131 /* If the following substitution will modify I0SRC, make a copy of it
3132 for the case where it is substituted for I0DEST in I1PAT later. */
3133 if (added_sets_1 && i0_feeds_i1_n)
3134 i0src_copy = copy_rtx (i0src);
3135 /* And a copy for I0DEST in I2PAT substitution. */
3136 if (added_sets_2 && ((i0_feeds_i1_n && i1_feeds_i2_n)
3137 || (i0_feeds_i2_n)))
3138 i0src_copy2 = copy_rtx (i0src);
3140 n_occurrences = 0;
3141 subst_low_luid = DF_INSN_LUID (i0);
3142 newpat = subst (newpat, i0dest, i0src, 0, 0, 0);
3143 substed_i0 = 1;
3146 /* Fail if an autoincrement side-effect has been duplicated. Be careful
3147 to count all the ways that I2SRC and I1SRC can be used. */
3148 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
3149 && i2_is_used + added_sets_2 > 1)
3150 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
3151 && (i1_is_used + added_sets_1 + (added_sets_2 && i1_feeds_i2_n)
3152 > 1))
3153 || (i0 != 0 && FIND_REG_INC_NOTE (i0, NULL_RTX) != 0
3154 && (n_occurrences + added_sets_0
3155 + (added_sets_1 && i0_feeds_i1_n)
3156 + (added_sets_2 && i0_feeds_i2_n)
3157 > 1))
3158 /* Fail if we tried to make a new register. */
3159 || max_reg_num () != maxreg
3160 /* Fail if we couldn't do something and have a CLOBBER. */
3161 || GET_CODE (newpat) == CLOBBER
3162 /* Fail if this new pattern is a MULT and we didn't have one before
3163 at the outer level. */
3164 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
3165 && ! have_mult))
3167 undo_all ();
3168 return 0;
3171 /* If the actions of the earlier insns must be kept
3172 in addition to substituting them into the latest one,
3173 we must make a new PARALLEL for the latest insn
3174 to hold additional the SETs. */
3176 if (added_sets_0 || added_sets_1 || added_sets_2)
3178 int extra_sets = added_sets_0 + added_sets_1 + added_sets_2;
3179 combine_extras++;
3181 if (GET_CODE (newpat) == PARALLEL)
3183 rtvec old = XVEC (newpat, 0);
3184 total_sets = XVECLEN (newpat, 0) + extra_sets;
3185 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
3186 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
3187 sizeof (old->elem[0]) * old->num_elem);
3189 else
3191 rtx old = newpat;
3192 total_sets = 1 + extra_sets;
3193 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
3194 XVECEXP (newpat, 0, 0) = old;
3197 if (added_sets_0)
3198 XVECEXP (newpat, 0, --total_sets) = i0pat;
3200 if (added_sets_1)
3202 rtx t = i1pat;
3203 if (i0_feeds_i1_n)
3204 t = subst (t, i0dest, i0src_copy ? i0src_copy : i0src, 0, 0, 0);
3206 XVECEXP (newpat, 0, --total_sets) = t;
3208 if (added_sets_2)
3210 rtx t = i2pat;
3211 if (i1_feeds_i2_n)
3212 t = subst (t, i1dest, i1src_copy ? i1src_copy : i1src, 0, 0,
3213 i0_feeds_i1_n && i0dest_in_i0src);
3214 if ((i0_feeds_i1_n && i1_feeds_i2_n) || i0_feeds_i2_n)
3215 t = subst (t, i0dest, i0src_copy2 ? i0src_copy2 : i0src, 0, 0, 0);
3217 XVECEXP (newpat, 0, --total_sets) = t;
3221 validate_replacement:
3223 /* Note which hard regs this insn has as inputs. */
3224 mark_used_regs_combine (newpat);
3226 /* If recog_for_combine fails, it strips existing clobbers. If we'll
3227 consider splitting this pattern, we might need these clobbers. */
3228 if (i1 && GET_CODE (newpat) == PARALLEL
3229 && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER)
3231 int len = XVECLEN (newpat, 0);
3233 newpat_vec_with_clobbers = rtvec_alloc (len);
3234 for (i = 0; i < len; i++)
3235 RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i);
3238 /* Is the result of combination a valid instruction? */
3239 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3241 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
3242 the second SET's destination is a register that is unused and isn't
3243 marked as an instruction that might trap in an EH region. In that case,
3244 we just need the first SET. This can occur when simplifying a divmod
3245 insn. We *must* test for this case here because the code below that
3246 splits two independent SETs doesn't handle this case correctly when it
3247 updates the register status.
3249 It's pointless doing this if we originally had two sets, one from
3250 i3, and one from i2. Combining then splitting the parallel results
3251 in the original i2 again plus an invalid insn (which we delete).
3252 The net effect is only to move instructions around, which makes
3253 debug info less accurate.
3255 Also check the case where the first SET's destination is unused.
3256 That would not cause incorrect code, but does cause an unneeded
3257 insn to remain. */
3259 if (insn_code_number < 0
3260 && !(added_sets_2 && i1 == 0)
3261 && GET_CODE (newpat) == PARALLEL
3262 && XVECLEN (newpat, 0) == 2
3263 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
3264 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
3265 && asm_noperands (newpat) < 0)
3267 rtx set0 = XVECEXP (newpat, 0, 0);
3268 rtx set1 = XVECEXP (newpat, 0, 1);
3270 if (((REG_P (SET_DEST (set1))
3271 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
3272 || (GET_CODE (SET_DEST (set1)) == SUBREG
3273 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
3274 && insn_nothrow_p (i3)
3275 && !side_effects_p (SET_SRC (set1)))
3277 newpat = set0;
3278 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3281 else if (((REG_P (SET_DEST (set0))
3282 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
3283 || (GET_CODE (SET_DEST (set0)) == SUBREG
3284 && find_reg_note (i3, REG_UNUSED,
3285 SUBREG_REG (SET_DEST (set0)))))
3286 && insn_nothrow_p (i3)
3287 && !side_effects_p (SET_SRC (set0)))
3289 newpat = set1;
3290 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3292 if (insn_code_number >= 0)
3293 changed_i3_dest = 1;
3297 /* If we were combining three insns and the result is a simple SET
3298 with no ASM_OPERANDS that wasn't recognized, try to split it into two
3299 insns. There are two ways to do this. It can be split using a
3300 machine-specific method (like when you have an addition of a large
3301 constant) or by combine in the function find_split_point. */
3303 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
3304 && asm_noperands (newpat) < 0)
3306 rtx parallel, m_split, *split;
3308 /* See if the MD file can split NEWPAT. If it can't, see if letting it
3309 use I2DEST as a scratch register will help. In the latter case,
3310 convert I2DEST to the mode of the source of NEWPAT if we can. */
3312 m_split = combine_split_insns (newpat, i3);
3314 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
3315 inputs of NEWPAT. */
3317 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
3318 possible to try that as a scratch reg. This would require adding
3319 more code to make it work though. */
3321 if (m_split == 0 && ! reg_overlap_mentioned_p (i2dest, newpat))
3323 enum machine_mode new_mode = GET_MODE (SET_DEST (newpat));
3325 /* First try to split using the original register as a
3326 scratch register. */
3327 parallel = gen_rtx_PARALLEL (VOIDmode,
3328 gen_rtvec (2, newpat,
3329 gen_rtx_CLOBBER (VOIDmode,
3330 i2dest)));
3331 m_split = combine_split_insns (parallel, i3);
3333 /* If that didn't work, try changing the mode of I2DEST if
3334 we can. */
3335 if (m_split == 0
3336 && new_mode != GET_MODE (i2dest)
3337 && new_mode != VOIDmode
3338 && can_change_dest_mode (i2dest, added_sets_2, new_mode))
3340 enum machine_mode old_mode = GET_MODE (i2dest);
3341 rtx ni2dest;
3343 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
3344 ni2dest = gen_rtx_REG (new_mode, REGNO (i2dest));
3345 else
3347 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], new_mode);
3348 ni2dest = regno_reg_rtx[REGNO (i2dest)];
3351 parallel = (gen_rtx_PARALLEL
3352 (VOIDmode,
3353 gen_rtvec (2, newpat,
3354 gen_rtx_CLOBBER (VOIDmode,
3355 ni2dest))));
3356 m_split = combine_split_insns (parallel, i3);
3358 if (m_split == 0
3359 && REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
3361 struct undo *buf;
3363 adjust_reg_mode (regno_reg_rtx[REGNO (i2dest)], old_mode);
3364 buf = undobuf.undos;
3365 undobuf.undos = buf->next;
3366 buf->next = undobuf.frees;
3367 undobuf.frees = buf;
3371 i2scratch = m_split != 0;
3374 /* If recog_for_combine has discarded clobbers, try to use them
3375 again for the split. */
3376 if (m_split == 0 && newpat_vec_with_clobbers)
3378 parallel = gen_rtx_PARALLEL (VOIDmode, newpat_vec_with_clobbers);
3379 m_split = combine_split_insns (parallel, i3);
3382 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
3384 m_split = PATTERN (m_split);
3385 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
3386 if (insn_code_number >= 0)
3387 newpat = m_split;
3389 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
3390 && (next_nonnote_nondebug_insn (i2) == i3
3391 || ! use_crosses_set_p (PATTERN (m_split), DF_INSN_LUID (i2))))
3393 rtx i2set, i3set;
3394 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
3395 newi2pat = PATTERN (m_split);
3397 i3set = single_set (NEXT_INSN (m_split));
3398 i2set = single_set (m_split);
3400 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
3402 /* If I2 or I3 has multiple SETs, we won't know how to track
3403 register status, so don't use these insns. If I2's destination
3404 is used between I2 and I3, we also can't use these insns. */
3406 if (i2_code_number >= 0 && i2set && i3set
3407 && (next_nonnote_nondebug_insn (i2) == i3
3408 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
3409 insn_code_number = recog_for_combine (&newi3pat, i3,
3410 &new_i3_notes);
3411 if (insn_code_number >= 0)
3412 newpat = newi3pat;
3414 /* It is possible that both insns now set the destination of I3.
3415 If so, we must show an extra use of it. */
3417 if (insn_code_number >= 0)
3419 rtx new_i3_dest = SET_DEST (i3set);
3420 rtx new_i2_dest = SET_DEST (i2set);
3422 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
3423 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
3424 || GET_CODE (new_i3_dest) == SUBREG)
3425 new_i3_dest = XEXP (new_i3_dest, 0);
3427 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
3428 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
3429 || GET_CODE (new_i2_dest) == SUBREG)
3430 new_i2_dest = XEXP (new_i2_dest, 0);
3432 if (REG_P (new_i3_dest)
3433 && REG_P (new_i2_dest)
3434 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
3435 INC_REG_N_SETS (REGNO (new_i2_dest), 1);
3439 /* If we can split it and use I2DEST, go ahead and see if that
3440 helps things be recognized. Verify that none of the registers
3441 are set between I2 and I3. */
3442 if (insn_code_number < 0
3443 && (split = find_split_point (&newpat, i3, false)) != 0
3444 #ifdef HAVE_cc0
3445 && REG_P (i2dest)
3446 #endif
3447 /* We need I2DEST in the proper mode. If it is a hard register
3448 or the only use of a pseudo, we can change its mode.
3449 Make sure we don't change a hard register to have a mode that
3450 isn't valid for it, or change the number of registers. */
3451 && (GET_MODE (*split) == GET_MODE (i2dest)
3452 || GET_MODE (*split) == VOIDmode
3453 || can_change_dest_mode (i2dest, added_sets_2,
3454 GET_MODE (*split)))
3455 && (next_nonnote_nondebug_insn (i2) == i3
3456 || ! use_crosses_set_p (*split, DF_INSN_LUID (i2)))
3457 /* We can't overwrite I2DEST if its value is still used by
3458 NEWPAT. */
3459 && ! reg_referenced_p (i2dest, newpat))
3461 rtx newdest = i2dest;
3462 enum rtx_code split_code = GET_CODE (*split);
3463 enum machine_mode split_mode = GET_MODE (*split);
3464 bool subst_done = false;
3465 newi2pat = NULL_RTX;
3467 i2scratch = true;
3469 /* *SPLIT may be part of I2SRC, so make sure we have the
3470 original expression around for later debug processing.
3471 We should not need I2SRC any more in other cases. */
3472 if (MAY_HAVE_DEBUG_INSNS)
3473 i2src = copy_rtx (i2src);
3474 else
3475 i2src = NULL;
3477 /* Get NEWDEST as a register in the proper mode. We have already
3478 validated that we can do this. */
3479 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
3481 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
3482 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
3483 else
3485 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], split_mode);
3486 newdest = regno_reg_rtx[REGNO (i2dest)];
3490 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
3491 an ASHIFT. This can occur if it was inside a PLUS and hence
3492 appeared to be a memory address. This is a kludge. */
3493 if (split_code == MULT
3494 && CONST_INT_P (XEXP (*split, 1))
3495 && INTVAL (XEXP (*split, 1)) > 0
3496 && (i = exact_log2 (UINTVAL (XEXP (*split, 1)))) >= 0)
3498 SUBST (*split, gen_rtx_ASHIFT (split_mode,
3499 XEXP (*split, 0), GEN_INT (i)));
3500 /* Update split_code because we may not have a multiply
3501 anymore. */
3502 split_code = GET_CODE (*split);
3505 #ifdef INSN_SCHEDULING
3506 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
3507 be written as a ZERO_EXTEND. */
3508 if (split_code == SUBREG && MEM_P (SUBREG_REG (*split)))
3510 #ifdef LOAD_EXTEND_OP
3511 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
3512 what it really is. */
3513 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
3514 == SIGN_EXTEND)
3515 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
3516 SUBREG_REG (*split)));
3517 else
3518 #endif
3519 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
3520 SUBREG_REG (*split)));
3522 #endif
3524 /* Attempt to split binary operators using arithmetic identities. */
3525 if (BINARY_P (SET_SRC (newpat))
3526 && split_mode == GET_MODE (SET_SRC (newpat))
3527 && ! side_effects_p (SET_SRC (newpat)))
3529 rtx setsrc = SET_SRC (newpat);
3530 enum machine_mode mode = GET_MODE (setsrc);
3531 enum rtx_code code = GET_CODE (setsrc);
3532 rtx src_op0 = XEXP (setsrc, 0);
3533 rtx src_op1 = XEXP (setsrc, 1);
3535 /* Split "X = Y op Y" as "Z = Y; X = Z op Z". */
3536 if (rtx_equal_p (src_op0, src_op1))
3538 newi2pat = gen_rtx_SET (VOIDmode, newdest, src_op0);
3539 SUBST (XEXP (setsrc, 0), newdest);
3540 SUBST (XEXP (setsrc, 1), newdest);
3541 subst_done = true;
3543 /* Split "((P op Q) op R) op S" where op is PLUS or MULT. */
3544 else if ((code == PLUS || code == MULT)
3545 && GET_CODE (src_op0) == code
3546 && GET_CODE (XEXP (src_op0, 0)) == code
3547 && (INTEGRAL_MODE_P (mode)
3548 || (FLOAT_MODE_P (mode)
3549 && flag_unsafe_math_optimizations)))
3551 rtx p = XEXP (XEXP (src_op0, 0), 0);
3552 rtx q = XEXP (XEXP (src_op0, 0), 1);
3553 rtx r = XEXP (src_op0, 1);
3554 rtx s = src_op1;
3556 /* Split both "((X op Y) op X) op Y" and
3557 "((X op Y) op Y) op X" as "T op T" where T is
3558 "X op Y". */
3559 if ((rtx_equal_p (p,r) && rtx_equal_p (q,s))
3560 || (rtx_equal_p (p,s) && rtx_equal_p (q,r)))
3562 newi2pat = gen_rtx_SET (VOIDmode, newdest,
3563 XEXP (src_op0, 0));
3564 SUBST (XEXP (setsrc, 0), newdest);
3565 SUBST (XEXP (setsrc, 1), newdest);
3566 subst_done = true;
3568 /* Split "((X op X) op Y) op Y)" as "T op T" where
3569 T is "X op Y". */
3570 else if (rtx_equal_p (p,q) && rtx_equal_p (r,s))
3572 rtx tmp = simplify_gen_binary (code, mode, p, r);
3573 newi2pat = gen_rtx_SET (VOIDmode, newdest, tmp);
3574 SUBST (XEXP (setsrc, 0), newdest);
3575 SUBST (XEXP (setsrc, 1), newdest);
3576 subst_done = true;
3581 if (!subst_done)
3583 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
3584 SUBST (*split, newdest);
3587 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
3589 /* recog_for_combine might have added CLOBBERs to newi2pat.
3590 Make sure NEWPAT does not depend on the clobbered regs. */
3591 if (GET_CODE (newi2pat) == PARALLEL)
3592 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
3593 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
3595 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
3596 if (reg_overlap_mentioned_p (reg, newpat))
3598 undo_all ();
3599 return 0;
3603 /* If the split point was a MULT and we didn't have one before,
3604 don't use one now. */
3605 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
3606 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3610 /* Check for a case where we loaded from memory in a narrow mode and
3611 then sign extended it, but we need both registers. In that case,
3612 we have a PARALLEL with both loads from the same memory location.
3613 We can split this into a load from memory followed by a register-register
3614 copy. This saves at least one insn, more if register allocation can
3615 eliminate the copy.
3617 We cannot do this if the destination of the first assignment is a
3618 condition code register or cc0. We eliminate this case by making sure
3619 the SET_DEST and SET_SRC have the same mode.
3621 We cannot do this if the destination of the second assignment is
3622 a register that we have already assumed is zero-extended. Similarly
3623 for a SUBREG of such a register. */
3625 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
3626 && GET_CODE (newpat) == PARALLEL
3627 && XVECLEN (newpat, 0) == 2
3628 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
3629 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
3630 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
3631 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
3632 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
3633 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
3634 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
3635 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
3636 DF_INSN_LUID (i2))
3637 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
3638 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
3639 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
3640 (REG_P (temp)
3641 && reg_stat[REGNO (temp)].nonzero_bits != 0
3642 && GET_MODE_PRECISION (GET_MODE (temp)) < BITS_PER_WORD
3643 && GET_MODE_PRECISION (GET_MODE (temp)) < HOST_BITS_PER_INT
3644 && (reg_stat[REGNO (temp)].nonzero_bits
3645 != GET_MODE_MASK (word_mode))))
3646 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
3647 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
3648 (REG_P (temp)
3649 && reg_stat[REGNO (temp)].nonzero_bits != 0
3650 && GET_MODE_PRECISION (GET_MODE (temp)) < BITS_PER_WORD
3651 && GET_MODE_PRECISION (GET_MODE (temp)) < HOST_BITS_PER_INT
3652 && (reg_stat[REGNO (temp)].nonzero_bits
3653 != GET_MODE_MASK (word_mode)))))
3654 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
3655 SET_SRC (XVECEXP (newpat, 0, 1)))
3656 && ! find_reg_note (i3, REG_UNUSED,
3657 SET_DEST (XVECEXP (newpat, 0, 0))))
3659 rtx ni2dest;
3661 newi2pat = XVECEXP (newpat, 0, 0);
3662 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
3663 newpat = XVECEXP (newpat, 0, 1);
3664 SUBST (SET_SRC (newpat),
3665 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
3666 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
3668 if (i2_code_number >= 0)
3669 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3671 if (insn_code_number >= 0)
3672 swap_i2i3 = 1;
3675 /* Similarly, check for a case where we have a PARALLEL of two independent
3676 SETs but we started with three insns. In this case, we can do the sets
3677 as two separate insns. This case occurs when some SET allows two
3678 other insns to combine, but the destination of that SET is still live. */
3680 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
3681 && GET_CODE (newpat) == PARALLEL
3682 && XVECLEN (newpat, 0) == 2
3683 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
3684 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
3685 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
3686 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
3687 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
3688 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
3689 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
3690 XVECEXP (newpat, 0, 0))
3691 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
3692 XVECEXP (newpat, 0, 1))
3693 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
3694 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1)))))
3696 /* Normally, it doesn't matter which of the two is done first,
3697 but the one that references cc0 can't be the second, and
3698 one which uses any regs/memory set in between i2 and i3 can't
3699 be first. */
3700 if (!use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
3701 DF_INSN_LUID (i2))
3702 #ifdef HAVE_cc0
3703 && !reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0))
3704 #endif
3707 newi2pat = XVECEXP (newpat, 0, 1);
3708 newpat = XVECEXP (newpat, 0, 0);
3710 else if (!use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 0)),
3711 DF_INSN_LUID (i2))
3712 #ifdef HAVE_cc0
3713 && !reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 1))
3714 #endif
3717 newi2pat = XVECEXP (newpat, 0, 0);
3718 newpat = XVECEXP (newpat, 0, 1);
3720 else
3722 undo_all ();
3723 return 0;
3726 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
3728 if (i2_code_number >= 0)
3730 /* recog_for_combine might have added CLOBBERs to newi2pat.
3731 Make sure NEWPAT does not depend on the clobbered regs. */
3732 if (GET_CODE (newi2pat) == PARALLEL)
3734 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
3735 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
3737 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
3738 if (reg_overlap_mentioned_p (reg, newpat))
3740 undo_all ();
3741 return 0;
3746 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
3750 /* If it still isn't recognized, fail and change things back the way they
3751 were. */
3752 if ((insn_code_number < 0
3753 /* Is the result a reasonable ASM_OPERANDS? */
3754 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
3756 undo_all ();
3757 return 0;
3760 /* If we had to change another insn, make sure it is valid also. */
3761 if (undobuf.other_insn)
3763 CLEAR_HARD_REG_SET (newpat_used_regs);
3765 other_pat = PATTERN (undobuf.other_insn);
3766 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
3767 &new_other_notes);
3769 if (other_code_number < 0 && ! check_asm_operands (other_pat))
3771 undo_all ();
3772 return 0;
3776 #ifdef HAVE_cc0
3777 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
3778 they are adjacent to each other or not. */
3780 rtx p = prev_nonnote_insn (i3);
3781 if (p && p != i2 && NONJUMP_INSN_P (p) && newi2pat
3782 && sets_cc0_p (newi2pat))
3784 undo_all ();
3785 return 0;
3788 #endif
3790 /* Only allow this combination if insn_rtx_costs reports that the
3791 replacement instructions are cheaper than the originals. */
3792 if (!combine_validate_cost (i0, i1, i2, i3, newpat, newi2pat, other_pat))
3794 undo_all ();
3795 return 0;
3798 if (MAY_HAVE_DEBUG_INSNS)
3800 struct undo *undo;
3802 for (undo = undobuf.undos; undo; undo = undo->next)
3803 if (undo->kind == UNDO_MODE)
3805 rtx reg = *undo->where.r;
3806 enum machine_mode new_mode = GET_MODE (reg);
3807 enum machine_mode old_mode = undo->old_contents.m;
3809 /* Temporarily revert mode back. */
3810 adjust_reg_mode (reg, old_mode);
3812 if (reg == i2dest && i2scratch)
3814 /* If we used i2dest as a scratch register with a
3815 different mode, substitute it for the original
3816 i2src while its original mode is temporarily
3817 restored, and then clear i2scratch so that we don't
3818 do it again later. */
3819 propagate_for_debug (i2, last_combined_insn, reg, i2src,
3820 this_basic_block);
3821 i2scratch = false;
3822 /* Put back the new mode. */
3823 adjust_reg_mode (reg, new_mode);
3825 else
3827 rtx tempreg = gen_raw_REG (old_mode, REGNO (reg));
3828 rtx first, last;
3830 if (reg == i2dest)
3832 first = i2;
3833 last = last_combined_insn;
3835 else
3837 first = i3;
3838 last = undobuf.other_insn;
3839 gcc_assert (last);
3840 if (DF_INSN_LUID (last)
3841 < DF_INSN_LUID (last_combined_insn))
3842 last = last_combined_insn;
3845 /* We're dealing with a reg that changed mode but not
3846 meaning, so we want to turn it into a subreg for
3847 the new mode. However, because of REG sharing and
3848 because its mode had already changed, we have to do
3849 it in two steps. First, replace any debug uses of
3850 reg, with its original mode temporarily restored,
3851 with this copy we have created; then, replace the
3852 copy with the SUBREG of the original shared reg,
3853 once again changed to the new mode. */
3854 propagate_for_debug (first, last, reg, tempreg,
3855 this_basic_block);
3856 adjust_reg_mode (reg, new_mode);
3857 propagate_for_debug (first, last, tempreg,
3858 lowpart_subreg (old_mode, reg, new_mode),
3859 this_basic_block);
3864 /* If we will be able to accept this, we have made a
3865 change to the destination of I3. This requires us to
3866 do a few adjustments. */
3868 if (changed_i3_dest)
3870 PATTERN (i3) = newpat;
3871 adjust_for_new_dest (i3);
3874 /* We now know that we can do this combination. Merge the insns and
3875 update the status of registers and LOG_LINKS. */
3877 if (undobuf.other_insn)
3879 rtx note, next;
3881 PATTERN (undobuf.other_insn) = other_pat;
3883 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
3884 are still valid. Then add any non-duplicate notes added by
3885 recog_for_combine. */
3886 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
3888 next = XEXP (note, 1);
3890 if (REG_NOTE_KIND (note) == REG_UNUSED
3891 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
3892 remove_note (undobuf.other_insn, note);
3895 distribute_notes (new_other_notes, undobuf.other_insn,
3896 undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX,
3897 NULL_RTX);
3900 if (swap_i2i3)
3902 rtx insn;
3903 struct insn_link *link;
3904 rtx ni2dest;
3906 /* I3 now uses what used to be its destination and which is now
3907 I2's destination. This requires us to do a few adjustments. */
3908 PATTERN (i3) = newpat;
3909 adjust_for_new_dest (i3);
3911 /* We need a LOG_LINK from I3 to I2. But we used to have one,
3912 so we still will.
3914 However, some later insn might be using I2's dest and have
3915 a LOG_LINK pointing at I3. We must remove this link.
3916 The simplest way to remove the link is to point it at I1,
3917 which we know will be a NOTE. */
3919 /* newi2pat is usually a SET here; however, recog_for_combine might
3920 have added some clobbers. */
3921 if (GET_CODE (newi2pat) == PARALLEL)
3922 ni2dest = SET_DEST (XVECEXP (newi2pat, 0, 0));
3923 else
3924 ni2dest = SET_DEST (newi2pat);
3926 for (insn = NEXT_INSN (i3);
3927 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3928 || insn != BB_HEAD (this_basic_block->next_bb));
3929 insn = NEXT_INSN (insn))
3931 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
3933 FOR_EACH_LOG_LINK (link, insn)
3934 if (link->insn == i3)
3935 link->insn = i1;
3937 break;
3943 rtx i3notes, i2notes, i1notes = 0, i0notes = 0;
3944 struct insn_link *i3links, *i2links, *i1links = 0, *i0links = 0;
3945 rtx midnotes = 0;
3946 int from_luid;
3947 /* Compute which registers we expect to eliminate. newi2pat may be setting
3948 either i3dest or i2dest, so we must check it. Also, i1dest may be the
3949 same as i3dest, in which case newi2pat may be setting i1dest. */
3950 rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat))
3951 || i2dest_in_i2src || i2dest_in_i1src || i2dest_in_i0src
3952 || !i2dest_killed
3953 ? 0 : i2dest);
3954 rtx elim_i1 = (i1 == 0 || i1dest_in_i1src || i1dest_in_i0src
3955 || (newi2pat && reg_set_p (i1dest, newi2pat))
3956 || !i1dest_killed
3957 ? 0 : i1dest);
3958 rtx elim_i0 = (i0 == 0 || i0dest_in_i0src
3959 || (newi2pat && reg_set_p (i0dest, newi2pat))
3960 || !i0dest_killed
3961 ? 0 : i0dest);
3963 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
3964 clear them. */
3965 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
3966 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
3967 if (i1)
3968 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
3969 if (i0)
3970 i0notes = REG_NOTES (i0), i0links = LOG_LINKS (i0);
3972 /* Ensure that we do not have something that should not be shared but
3973 occurs multiple times in the new insns. Check this by first
3974 resetting all the `used' flags and then copying anything is shared. */
3976 reset_used_flags (i3notes);
3977 reset_used_flags (i2notes);
3978 reset_used_flags (i1notes);
3979 reset_used_flags (i0notes);
3980 reset_used_flags (newpat);
3981 reset_used_flags (newi2pat);
3982 if (undobuf.other_insn)
3983 reset_used_flags (PATTERN (undobuf.other_insn));
3985 i3notes = copy_rtx_if_shared (i3notes);
3986 i2notes = copy_rtx_if_shared (i2notes);
3987 i1notes = copy_rtx_if_shared (i1notes);
3988 i0notes = copy_rtx_if_shared (i0notes);
3989 newpat = copy_rtx_if_shared (newpat);
3990 newi2pat = copy_rtx_if_shared (newi2pat);
3991 if (undobuf.other_insn)
3992 reset_used_flags (PATTERN (undobuf.other_insn));
3994 INSN_CODE (i3) = insn_code_number;
3995 PATTERN (i3) = newpat;
3997 if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3))
3999 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
4001 reset_used_flags (call_usage);
4002 call_usage = copy_rtx (call_usage);
4004 if (substed_i2)
4006 /* I2SRC must still be meaningful at this point. Some splitting
4007 operations can invalidate I2SRC, but those operations do not
4008 apply to calls. */
4009 gcc_assert (i2src);
4010 replace_rtx (call_usage, i2dest, i2src);
4013 if (substed_i1)
4014 replace_rtx (call_usage, i1dest, i1src);
4015 if (substed_i0)
4016 replace_rtx (call_usage, i0dest, i0src);
4018 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
4021 if (undobuf.other_insn)
4022 INSN_CODE (undobuf.other_insn) = other_code_number;
4024 /* We had one special case above where I2 had more than one set and
4025 we replaced a destination of one of those sets with the destination
4026 of I3. In that case, we have to update LOG_LINKS of insns later
4027 in this basic block. Note that this (expensive) case is rare.
4029 Also, in this case, we must pretend that all REG_NOTEs for I2
4030 actually came from I3, so that REG_UNUSED notes from I2 will be
4031 properly handled. */
4033 if (i3_subst_into_i2)
4035 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
4036 if ((GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == SET
4037 || GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == CLOBBER)
4038 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i)))
4039 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
4040 && ! find_reg_note (i2, REG_UNUSED,
4041 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
4042 for (temp = NEXT_INSN (i2);
4043 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
4044 || BB_HEAD (this_basic_block) != temp);
4045 temp = NEXT_INSN (temp))
4046 if (temp != i3 && INSN_P (temp))
4047 FOR_EACH_LOG_LINK (link, temp)
4048 if (link->insn == i2)
4049 link->insn = i3;
4051 if (i3notes)
4053 rtx link = i3notes;
4054 while (XEXP (link, 1))
4055 link = XEXP (link, 1);
4056 XEXP (link, 1) = i2notes;
4058 else
4059 i3notes = i2notes;
4060 i2notes = 0;
4063 LOG_LINKS (i3) = NULL;
4064 REG_NOTES (i3) = 0;
4065 LOG_LINKS (i2) = NULL;
4066 REG_NOTES (i2) = 0;
4068 if (newi2pat)
4070 if (MAY_HAVE_DEBUG_INSNS && i2scratch)
4071 propagate_for_debug (i2, last_combined_insn, i2dest, i2src,
4072 this_basic_block);
4073 INSN_CODE (i2) = i2_code_number;
4074 PATTERN (i2) = newi2pat;
4076 else
4078 if (MAY_HAVE_DEBUG_INSNS && i2src)
4079 propagate_for_debug (i2, last_combined_insn, i2dest, i2src,
4080 this_basic_block);
4081 SET_INSN_DELETED (i2);
4084 if (i1)
4086 LOG_LINKS (i1) = NULL;
4087 REG_NOTES (i1) = 0;
4088 if (MAY_HAVE_DEBUG_INSNS)
4089 propagate_for_debug (i1, last_combined_insn, i1dest, i1src,
4090 this_basic_block);
4091 SET_INSN_DELETED (i1);
4094 if (i0)
4096 LOG_LINKS (i0) = NULL;
4097 REG_NOTES (i0) = 0;
4098 if (MAY_HAVE_DEBUG_INSNS)
4099 propagate_for_debug (i0, last_combined_insn, i0dest, i0src,
4100 this_basic_block);
4101 SET_INSN_DELETED (i0);
4104 /* Get death notes for everything that is now used in either I3 or
4105 I2 and used to die in a previous insn. If we built two new
4106 patterns, move from I1 to I2 then I2 to I3 so that we get the
4107 proper movement on registers that I2 modifies. */
4109 if (i0)
4110 from_luid = DF_INSN_LUID (i0);
4111 else if (i1)
4112 from_luid = DF_INSN_LUID (i1);
4113 else
4114 from_luid = DF_INSN_LUID (i2);
4115 if (newi2pat)
4116 move_deaths (newi2pat, NULL_RTX, from_luid, i2, &midnotes);
4117 move_deaths (newpat, newi2pat, from_luid, i3, &midnotes);
4119 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
4120 if (i3notes)
4121 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX,
4122 elim_i2, elim_i1, elim_i0);
4123 if (i2notes)
4124 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX,
4125 elim_i2, elim_i1, elim_i0);
4126 if (i1notes)
4127 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX,
4128 elim_i2, elim_i1, elim_i0);
4129 if (i0notes)
4130 distribute_notes (i0notes, i0, i3, newi2pat ? i2 : NULL_RTX,
4131 elim_i2, elim_i1, elim_i0);
4132 if (midnotes)
4133 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
4134 elim_i2, elim_i1, elim_i0);
4136 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
4137 know these are REG_UNUSED and want them to go to the desired insn,
4138 so we always pass it as i3. */
4140 if (newi2pat && new_i2_notes)
4141 distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX,
4142 NULL_RTX);
4144 if (new_i3_notes)
4145 distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX,
4146 NULL_RTX);
4148 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
4149 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
4150 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
4151 in that case, it might delete I2. Similarly for I2 and I1.
4152 Show an additional death due to the REG_DEAD note we make here. If
4153 we discard it in distribute_notes, we will decrement it again. */
4155 if (i3dest_killed)
4157 rtx new_note = alloc_reg_note (REG_DEAD, i3dest_killed, NULL_RTX);
4158 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
4159 distribute_notes (new_note, NULL_RTX, i2, NULL_RTX, elim_i2,
4160 elim_i1, elim_i0);
4161 else
4162 distribute_notes (new_note, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
4163 elim_i2, elim_i1, elim_i0);
4166 if (i2dest_in_i2src)
4168 rtx new_note = alloc_reg_note (REG_DEAD, i2dest, NULL_RTX);
4169 if (newi2pat && reg_set_p (i2dest, newi2pat))
4170 distribute_notes (new_note, NULL_RTX, i2, NULL_RTX, NULL_RTX,
4171 NULL_RTX, NULL_RTX);
4172 else
4173 distribute_notes (new_note, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
4174 NULL_RTX, NULL_RTX, NULL_RTX);
4177 if (i1dest_in_i1src)
4179 rtx new_note = alloc_reg_note (REG_DEAD, i1dest, NULL_RTX);
4180 if (newi2pat && reg_set_p (i1dest, newi2pat))
4181 distribute_notes (new_note, NULL_RTX, i2, NULL_RTX, NULL_RTX,
4182 NULL_RTX, NULL_RTX);
4183 else
4184 distribute_notes (new_note, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
4185 NULL_RTX, NULL_RTX, NULL_RTX);
4188 if (i0dest_in_i0src)
4190 rtx new_note = alloc_reg_note (REG_DEAD, i0dest, NULL_RTX);
4191 if (newi2pat && reg_set_p (i0dest, newi2pat))
4192 distribute_notes (new_note, NULL_RTX, i2, NULL_RTX, NULL_RTX,
4193 NULL_RTX, NULL_RTX);
4194 else
4195 distribute_notes (new_note, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
4196 NULL_RTX, NULL_RTX, NULL_RTX);
4199 distribute_links (i3links);
4200 distribute_links (i2links);
4201 distribute_links (i1links);
4202 distribute_links (i0links);
4204 if (REG_P (i2dest))
4206 struct insn_link *link;
4207 rtx i2_insn = 0, i2_val = 0, set;
4209 /* The insn that used to set this register doesn't exist, and
4210 this life of the register may not exist either. See if one of
4211 I3's links points to an insn that sets I2DEST. If it does,
4212 that is now the last known value for I2DEST. If we don't update
4213 this and I2 set the register to a value that depended on its old
4214 contents, we will get confused. If this insn is used, thing
4215 will be set correctly in combine_instructions. */
4216 FOR_EACH_LOG_LINK (link, i3)
4217 if ((set = single_set (link->insn)) != 0
4218 && rtx_equal_p (i2dest, SET_DEST (set)))
4219 i2_insn = link->insn, i2_val = SET_SRC (set);
4221 record_value_for_reg (i2dest, i2_insn, i2_val);
4223 /* If the reg formerly set in I2 died only once and that was in I3,
4224 zero its use count so it won't make `reload' do any work. */
4225 if (! added_sets_2
4226 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
4227 && ! i2dest_in_i2src)
4228 INC_REG_N_SETS (REGNO (i2dest), -1);
4231 if (i1 && REG_P (i1dest))
4233 struct insn_link *link;
4234 rtx i1_insn = 0, i1_val = 0, set;
4236 FOR_EACH_LOG_LINK (link, i3)
4237 if ((set = single_set (link->insn)) != 0
4238 && rtx_equal_p (i1dest, SET_DEST (set)))
4239 i1_insn = link->insn, i1_val = SET_SRC (set);
4241 record_value_for_reg (i1dest, i1_insn, i1_val);
4243 if (! added_sets_1 && ! i1dest_in_i1src)
4244 INC_REG_N_SETS (REGNO (i1dest), -1);
4247 if (i0 && REG_P (i0dest))
4249 struct insn_link *link;
4250 rtx i0_insn = 0, i0_val = 0, set;
4252 FOR_EACH_LOG_LINK (link, i3)
4253 if ((set = single_set (link->insn)) != 0
4254 && rtx_equal_p (i0dest, SET_DEST (set)))
4255 i0_insn = link->insn, i0_val = SET_SRC (set);
4257 record_value_for_reg (i0dest, i0_insn, i0_val);
4259 if (! added_sets_0 && ! i0dest_in_i0src)
4260 INC_REG_N_SETS (REGNO (i0dest), -1);
4263 /* Update reg_stat[].nonzero_bits et al for any changes that may have
4264 been made to this insn. The order of
4265 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
4266 can affect nonzero_bits of newpat */
4267 if (newi2pat)
4268 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
4269 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
4272 if (undobuf.other_insn != NULL_RTX)
4274 if (dump_file)
4276 fprintf (dump_file, "modifying other_insn ");
4277 dump_insn_slim (dump_file, undobuf.other_insn);
4279 df_insn_rescan (undobuf.other_insn);
4282 if (i0 && !(NOTE_P(i0) && (NOTE_KIND (i0) == NOTE_INSN_DELETED)))
4284 if (dump_file)
4286 fprintf (dump_file, "modifying insn i1 ");
4287 dump_insn_slim (dump_file, i0);
4289 df_insn_rescan (i0);
4292 if (i1 && !(NOTE_P(i1) && (NOTE_KIND (i1) == NOTE_INSN_DELETED)))
4294 if (dump_file)
4296 fprintf (dump_file, "modifying insn i1 ");
4297 dump_insn_slim (dump_file, i1);
4299 df_insn_rescan (i1);
4302 if (i2 && !(NOTE_P(i2) && (NOTE_KIND (i2) == NOTE_INSN_DELETED)))
4304 if (dump_file)
4306 fprintf (dump_file, "modifying insn i2 ");
4307 dump_insn_slim (dump_file, i2);
4309 df_insn_rescan (i2);
4312 if (i3 && !(NOTE_P(i3) && (NOTE_KIND (i3) == NOTE_INSN_DELETED)))
4314 if (dump_file)
4316 fprintf (dump_file, "modifying insn i3 ");
4317 dump_insn_slim (dump_file, i3);
4319 df_insn_rescan (i3);
4322 /* Set new_direct_jump_p if a new return or simple jump instruction
4323 has been created. Adjust the CFG accordingly. */
4325 if (returnjump_p (i3) || any_uncondjump_p (i3))
4327 *new_direct_jump_p = 1;
4328 mark_jump_label (PATTERN (i3), i3, 0);
4329 update_cfg_for_uncondjump (i3);
4332 if (undobuf.other_insn != NULL_RTX
4333 && (returnjump_p (undobuf.other_insn)
4334 || any_uncondjump_p (undobuf.other_insn)))
4336 *new_direct_jump_p = 1;
4337 update_cfg_for_uncondjump (undobuf.other_insn);
4340 /* A noop might also need cleaning up of CFG, if it comes from the
4341 simplification of a jump. */
4342 if (JUMP_P (i3)
4343 && GET_CODE (newpat) == SET
4344 && SET_SRC (newpat) == pc_rtx
4345 && SET_DEST (newpat) == pc_rtx)
4347 *new_direct_jump_p = 1;
4348 update_cfg_for_uncondjump (i3);
4351 if (undobuf.other_insn != NULL_RTX
4352 && JUMP_P (undobuf.other_insn)
4353 && GET_CODE (PATTERN (undobuf.other_insn)) == SET
4354 && SET_SRC (PATTERN (undobuf.other_insn)) == pc_rtx
4355 && SET_DEST (PATTERN (undobuf.other_insn)) == pc_rtx)
4357 *new_direct_jump_p = 1;
4358 update_cfg_for_uncondjump (undobuf.other_insn);
4361 combine_successes++;
4362 undo_commit ();
4364 if (added_links_insn
4365 && (newi2pat == 0 || DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (i2))
4366 && DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (i3))
4367 return added_links_insn;
4368 else
4369 return newi2pat ? i2 : i3;
4372 /* Undo all the modifications recorded in undobuf. */
4374 static void
4375 undo_all (void)
4377 struct undo *undo, *next;
4379 for (undo = undobuf.undos; undo; undo = next)
4381 next = undo->next;
4382 switch (undo->kind)
4384 case UNDO_RTX:
4385 *undo->where.r = undo->old_contents.r;
4386 break;
4387 case UNDO_INT:
4388 *undo->where.i = undo->old_contents.i;
4389 break;
4390 case UNDO_MODE:
4391 adjust_reg_mode (*undo->where.r, undo->old_contents.m);
4392 break;
4393 case UNDO_LINKS:
4394 *undo->where.l = undo->old_contents.l;
4395 break;
4396 default:
4397 gcc_unreachable ();
4400 undo->next = undobuf.frees;
4401 undobuf.frees = undo;
4404 undobuf.undos = 0;
4407 /* We've committed to accepting the changes we made. Move all
4408 of the undos to the free list. */
4410 static void
4411 undo_commit (void)
4413 struct undo *undo, *next;
4415 for (undo = undobuf.undos; undo; undo = next)
4417 next = undo->next;
4418 undo->next = undobuf.frees;
4419 undobuf.frees = undo;
4421 undobuf.undos = 0;
4424 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
4425 where we have an arithmetic expression and return that point. LOC will
4426 be inside INSN.
4428 try_combine will call this function to see if an insn can be split into
4429 two insns. */
4431 static rtx *
4432 find_split_point (rtx *loc, rtx insn, bool set_src)
4434 rtx x = *loc;
4435 enum rtx_code code = GET_CODE (x);
4436 rtx *split;
4437 unsigned HOST_WIDE_INT len = 0;
4438 HOST_WIDE_INT pos = 0;
4439 int unsignedp = 0;
4440 rtx inner = NULL_RTX;
4442 /* First special-case some codes. */
4443 switch (code)
4445 case SUBREG:
4446 #ifdef INSN_SCHEDULING
4447 /* If we are making a paradoxical SUBREG invalid, it becomes a split
4448 point. */
4449 if (MEM_P (SUBREG_REG (x)))
4450 return loc;
4451 #endif
4452 return find_split_point (&SUBREG_REG (x), insn, false);
4454 case MEM:
4455 #ifdef HAVE_lo_sum
4456 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
4457 using LO_SUM and HIGH. */
4458 if (GET_CODE (XEXP (x, 0)) == CONST
4459 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
4461 enum machine_mode address_mode = get_address_mode (x);
4463 SUBST (XEXP (x, 0),
4464 gen_rtx_LO_SUM (address_mode,
4465 gen_rtx_HIGH (address_mode, XEXP (x, 0)),
4466 XEXP (x, 0)));
4467 return &XEXP (XEXP (x, 0), 0);
4469 #endif
4471 /* If we have a PLUS whose second operand is a constant and the
4472 address is not valid, perhaps will can split it up using
4473 the machine-specific way to split large constants. We use
4474 the first pseudo-reg (one of the virtual regs) as a placeholder;
4475 it will not remain in the result. */
4476 if (GET_CODE (XEXP (x, 0)) == PLUS
4477 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
4478 && ! memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0),
4479 MEM_ADDR_SPACE (x)))
4481 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
4482 rtx seq = combine_split_insns (gen_rtx_SET (VOIDmode, reg,
4483 XEXP (x, 0)),
4484 subst_insn);
4486 /* This should have produced two insns, each of which sets our
4487 placeholder. If the source of the second is a valid address,
4488 we can make put both sources together and make a split point
4489 in the middle. */
4491 if (seq
4492 && NEXT_INSN (seq) != NULL_RTX
4493 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
4494 && NONJUMP_INSN_P (seq)
4495 && GET_CODE (PATTERN (seq)) == SET
4496 && SET_DEST (PATTERN (seq)) == reg
4497 && ! reg_mentioned_p (reg,
4498 SET_SRC (PATTERN (seq)))
4499 && NONJUMP_INSN_P (NEXT_INSN (seq))
4500 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
4501 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
4502 && memory_address_addr_space_p
4503 (GET_MODE (x), SET_SRC (PATTERN (NEXT_INSN (seq))),
4504 MEM_ADDR_SPACE (x)))
4506 rtx src1 = SET_SRC (PATTERN (seq));
4507 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
4509 /* Replace the placeholder in SRC2 with SRC1. If we can
4510 find where in SRC2 it was placed, that can become our
4511 split point and we can replace this address with SRC2.
4512 Just try two obvious places. */
4514 src2 = replace_rtx (src2, reg, src1);
4515 split = 0;
4516 if (XEXP (src2, 0) == src1)
4517 split = &XEXP (src2, 0);
4518 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
4519 && XEXP (XEXP (src2, 0), 0) == src1)
4520 split = &XEXP (XEXP (src2, 0), 0);
4522 if (split)
4524 SUBST (XEXP (x, 0), src2);
4525 return split;
4529 /* If that didn't work, perhaps the first operand is complex and
4530 needs to be computed separately, so make a split point there.
4531 This will occur on machines that just support REG + CONST
4532 and have a constant moved through some previous computation. */
4534 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
4535 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
4536 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
4537 return &XEXP (XEXP (x, 0), 0);
4540 /* If we have a PLUS whose first operand is complex, try computing it
4541 separately by making a split there. */
4542 if (GET_CODE (XEXP (x, 0)) == PLUS
4543 && ! memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0),
4544 MEM_ADDR_SPACE (x))
4545 && ! OBJECT_P (XEXP (XEXP (x, 0), 0))
4546 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
4547 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
4548 return &XEXP (XEXP (x, 0), 0);
4549 break;
4551 case SET:
4552 #ifdef HAVE_cc0
4553 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
4554 ZERO_EXTRACT, the most likely reason why this doesn't match is that
4555 we need to put the operand into a register. So split at that
4556 point. */
4558 if (SET_DEST (x) == cc0_rtx
4559 && GET_CODE (SET_SRC (x)) != COMPARE
4560 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
4561 && !OBJECT_P (SET_SRC (x))
4562 && ! (GET_CODE (SET_SRC (x)) == SUBREG
4563 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
4564 return &SET_SRC (x);
4565 #endif
4567 /* See if we can split SET_SRC as it stands. */
4568 split = find_split_point (&SET_SRC (x), insn, true);
4569 if (split && split != &SET_SRC (x))
4570 return split;
4572 /* See if we can split SET_DEST as it stands. */
4573 split = find_split_point (&SET_DEST (x), insn, false);
4574 if (split && split != &SET_DEST (x))
4575 return split;
4577 /* See if this is a bitfield assignment with everything constant. If
4578 so, this is an IOR of an AND, so split it into that. */
4579 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
4580 && HWI_COMPUTABLE_MODE_P (GET_MODE (XEXP (SET_DEST (x), 0)))
4581 && CONST_INT_P (XEXP (SET_DEST (x), 1))
4582 && CONST_INT_P (XEXP (SET_DEST (x), 2))
4583 && CONST_INT_P (SET_SRC (x))
4584 && ((INTVAL (XEXP (SET_DEST (x), 1))
4585 + INTVAL (XEXP (SET_DEST (x), 2)))
4586 <= GET_MODE_PRECISION (GET_MODE (XEXP (SET_DEST (x), 0))))
4587 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
4589 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
4590 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
4591 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
4592 rtx dest = XEXP (SET_DEST (x), 0);
4593 enum machine_mode mode = GET_MODE (dest);
4594 unsigned HOST_WIDE_INT mask
4595 = ((unsigned HOST_WIDE_INT) 1 << len) - 1;
4596 rtx or_mask;
4598 if (BITS_BIG_ENDIAN)
4599 pos = GET_MODE_PRECISION (mode) - len - pos;
4601 or_mask = gen_int_mode (src << pos, mode);
4602 if (src == mask)
4603 SUBST (SET_SRC (x),
4604 simplify_gen_binary (IOR, mode, dest, or_mask));
4605 else
4607 rtx negmask = gen_int_mode (~(mask << pos), mode);
4608 SUBST (SET_SRC (x),
4609 simplify_gen_binary (IOR, mode,
4610 simplify_gen_binary (AND, mode,
4611 dest, negmask),
4612 or_mask));
4615 SUBST (SET_DEST (x), dest);
4617 split = find_split_point (&SET_SRC (x), insn, true);
4618 if (split && split != &SET_SRC (x))
4619 return split;
4622 /* Otherwise, see if this is an operation that we can split into two.
4623 If so, try to split that. */
4624 code = GET_CODE (SET_SRC (x));
4626 switch (code)
4628 case AND:
4629 /* If we are AND'ing with a large constant that is only a single
4630 bit and the result is only being used in a context where we
4631 need to know if it is zero or nonzero, replace it with a bit
4632 extraction. This will avoid the large constant, which might
4633 have taken more than one insn to make. If the constant were
4634 not a valid argument to the AND but took only one insn to make,
4635 this is no worse, but if it took more than one insn, it will
4636 be better. */
4638 if (CONST_INT_P (XEXP (SET_SRC (x), 1))
4639 && REG_P (XEXP (SET_SRC (x), 0))
4640 && (pos = exact_log2 (UINTVAL (XEXP (SET_SRC (x), 1)))) >= 7
4641 && REG_P (SET_DEST (x))
4642 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
4643 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
4644 && XEXP (*split, 0) == SET_DEST (x)
4645 && XEXP (*split, 1) == const0_rtx)
4647 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
4648 XEXP (SET_SRC (x), 0),
4649 pos, NULL_RTX, 1, 1, 0, 0);
4650 if (extraction != 0)
4652 SUBST (SET_SRC (x), extraction);
4653 return find_split_point (loc, insn, false);
4656 break;
4658 case NE:
4659 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
4660 is known to be on, this can be converted into a NEG of a shift. */
4661 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
4662 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
4663 && 1 <= (pos = exact_log2
4664 (nonzero_bits (XEXP (SET_SRC (x), 0),
4665 GET_MODE (XEXP (SET_SRC (x), 0))))))
4667 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
4669 SUBST (SET_SRC (x),
4670 gen_rtx_NEG (mode,
4671 gen_rtx_LSHIFTRT (mode,
4672 XEXP (SET_SRC (x), 0),
4673 GEN_INT (pos))));
4675 split = find_split_point (&SET_SRC (x), insn, true);
4676 if (split && split != &SET_SRC (x))
4677 return split;
4679 break;
4681 case SIGN_EXTEND:
4682 inner = XEXP (SET_SRC (x), 0);
4684 /* We can't optimize if either mode is a partial integer
4685 mode as we don't know how many bits are significant
4686 in those modes. */
4687 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
4688 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
4689 break;
4691 pos = 0;
4692 len = GET_MODE_PRECISION (GET_MODE (inner));
4693 unsignedp = 0;
4694 break;
4696 case SIGN_EXTRACT:
4697 case ZERO_EXTRACT:
4698 if (CONST_INT_P (XEXP (SET_SRC (x), 1))
4699 && CONST_INT_P (XEXP (SET_SRC (x), 2)))
4701 inner = XEXP (SET_SRC (x), 0);
4702 len = INTVAL (XEXP (SET_SRC (x), 1));
4703 pos = INTVAL (XEXP (SET_SRC (x), 2));
4705 if (BITS_BIG_ENDIAN)
4706 pos = GET_MODE_PRECISION (GET_MODE (inner)) - len - pos;
4707 unsignedp = (code == ZERO_EXTRACT);
4709 break;
4711 default:
4712 break;
4715 if (len && pos >= 0
4716 && pos + len <= GET_MODE_PRECISION (GET_MODE (inner)))
4718 enum machine_mode mode = GET_MODE (SET_SRC (x));
4720 /* For unsigned, we have a choice of a shift followed by an
4721 AND or two shifts. Use two shifts for field sizes where the
4722 constant might be too large. We assume here that we can
4723 always at least get 8-bit constants in an AND insn, which is
4724 true for every current RISC. */
4726 if (unsignedp && len <= 8)
4728 unsigned HOST_WIDE_INT mask
4729 = ((unsigned HOST_WIDE_INT) 1 << len) - 1;
4730 SUBST (SET_SRC (x),
4731 gen_rtx_AND (mode,
4732 gen_rtx_LSHIFTRT
4733 (mode, gen_lowpart (mode, inner),
4734 GEN_INT (pos)),
4735 gen_int_mode (mask, mode)));
4737 split = find_split_point (&SET_SRC (x), insn, true);
4738 if (split && split != &SET_SRC (x))
4739 return split;
4741 else
4743 SUBST (SET_SRC (x),
4744 gen_rtx_fmt_ee
4745 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
4746 gen_rtx_ASHIFT (mode,
4747 gen_lowpart (mode, inner),
4748 GEN_INT (GET_MODE_PRECISION (mode)
4749 - len - pos)),
4750 GEN_INT (GET_MODE_PRECISION (mode) - len)));
4752 split = find_split_point (&SET_SRC (x), insn, true);
4753 if (split && split != &SET_SRC (x))
4754 return split;
4758 /* See if this is a simple operation with a constant as the second
4759 operand. It might be that this constant is out of range and hence
4760 could be used as a split point. */
4761 if (BINARY_P (SET_SRC (x))
4762 && CONSTANT_P (XEXP (SET_SRC (x), 1))
4763 && (OBJECT_P (XEXP (SET_SRC (x), 0))
4764 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
4765 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
4766 return &XEXP (SET_SRC (x), 1);
4768 /* Finally, see if this is a simple operation with its first operand
4769 not in a register. The operation might require this operand in a
4770 register, so return it as a split point. We can always do this
4771 because if the first operand were another operation, we would have
4772 already found it as a split point. */
4773 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
4774 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
4775 return &XEXP (SET_SRC (x), 0);
4777 return 0;
4779 case AND:
4780 case IOR:
4781 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
4782 it is better to write this as (not (ior A B)) so we can split it.
4783 Similarly for IOR. */
4784 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
4786 SUBST (*loc,
4787 gen_rtx_NOT (GET_MODE (x),
4788 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
4789 GET_MODE (x),
4790 XEXP (XEXP (x, 0), 0),
4791 XEXP (XEXP (x, 1), 0))));
4792 return find_split_point (loc, insn, set_src);
4795 /* Many RISC machines have a large set of logical insns. If the
4796 second operand is a NOT, put it first so we will try to split the
4797 other operand first. */
4798 if (GET_CODE (XEXP (x, 1)) == NOT)
4800 rtx tem = XEXP (x, 0);
4801 SUBST (XEXP (x, 0), XEXP (x, 1));
4802 SUBST (XEXP (x, 1), tem);
4804 break;
4806 case PLUS:
4807 case MINUS:
4808 /* Canonicalization can produce (minus A (mult B C)), where C is a
4809 constant. It may be better to try splitting (plus (mult B -C) A)
4810 instead if this isn't a multiply by a power of two. */
4811 if (set_src && code == MINUS && GET_CODE (XEXP (x, 1)) == MULT
4812 && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
4813 && exact_log2 (INTVAL (XEXP (XEXP (x, 1), 1))) < 0)
4815 enum machine_mode mode = GET_MODE (x);
4816 unsigned HOST_WIDE_INT this_int = INTVAL (XEXP (XEXP (x, 1), 1));
4817 HOST_WIDE_INT other_int = trunc_int_for_mode (-this_int, mode);
4818 SUBST (*loc, gen_rtx_PLUS (mode,
4819 gen_rtx_MULT (mode,
4820 XEXP (XEXP (x, 1), 0),
4821 gen_int_mode (other_int,
4822 mode)),
4823 XEXP (x, 0)));
4824 return find_split_point (loc, insn, set_src);
4827 /* Split at a multiply-accumulate instruction. However if this is
4828 the SET_SRC, we likely do not have such an instruction and it's
4829 worthless to try this split. */
4830 if (!set_src && GET_CODE (XEXP (x, 0)) == MULT)
4831 return loc;
4833 default:
4834 break;
4837 /* Otherwise, select our actions depending on our rtx class. */
4838 switch (GET_RTX_CLASS (code))
4840 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
4841 case RTX_TERNARY:
4842 split = find_split_point (&XEXP (x, 2), insn, false);
4843 if (split)
4844 return split;
4845 /* ... fall through ... */
4846 case RTX_BIN_ARITH:
4847 case RTX_COMM_ARITH:
4848 case RTX_COMPARE:
4849 case RTX_COMM_COMPARE:
4850 split = find_split_point (&XEXP (x, 1), insn, false);
4851 if (split)
4852 return split;
4853 /* ... fall through ... */
4854 case RTX_UNARY:
4855 /* Some machines have (and (shift ...) ...) insns. If X is not
4856 an AND, but XEXP (X, 0) is, use it as our split point. */
4857 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
4858 return &XEXP (x, 0);
4860 split = find_split_point (&XEXP (x, 0), insn, false);
4861 if (split)
4862 return split;
4863 return loc;
4865 default:
4866 /* Otherwise, we don't have a split point. */
4867 return 0;
4871 /* Throughout X, replace FROM with TO, and return the result.
4872 The result is TO if X is FROM;
4873 otherwise the result is X, but its contents may have been modified.
4874 If they were modified, a record was made in undobuf so that
4875 undo_all will (among other things) return X to its original state.
4877 If the number of changes necessary is too much to record to undo,
4878 the excess changes are not made, so the result is invalid.
4879 The changes already made can still be undone.
4880 undobuf.num_undo is incremented for such changes, so by testing that
4881 the caller can tell whether the result is valid.
4883 `n_occurrences' is incremented each time FROM is replaced.
4885 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
4887 IN_COND is nonzero if we are at the top level of a condition.
4889 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
4890 by copying if `n_occurrences' is nonzero. */
4892 static rtx
4893 subst (rtx x, rtx from, rtx to, int in_dest, int in_cond, int unique_copy)
4895 enum rtx_code code = GET_CODE (x);
4896 enum machine_mode op0_mode = VOIDmode;
4897 const char *fmt;
4898 int len, i;
4899 rtx new_rtx;
4901 /* Two expressions are equal if they are identical copies of a shared
4902 RTX or if they are both registers with the same register number
4903 and mode. */
4905 #define COMBINE_RTX_EQUAL_P(X,Y) \
4906 ((X) == (Y) \
4907 || (REG_P (X) && REG_P (Y) \
4908 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
4910 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
4912 n_occurrences++;
4913 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
4916 /* If X and FROM are the same register but different modes, they
4917 will not have been seen as equal above. However, the log links code
4918 will make a LOG_LINKS entry for that case. If we do nothing, we
4919 will try to rerecognize our original insn and, when it succeeds,
4920 we will delete the feeding insn, which is incorrect.
4922 So force this insn not to match in this (rare) case. */
4923 if (! in_dest && code == REG && REG_P (from)
4924 && reg_overlap_mentioned_p (x, from))
4925 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
4927 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
4928 of which may contain things that can be combined. */
4929 if (code != MEM && code != LO_SUM && OBJECT_P (x))
4930 return x;
4932 /* It is possible to have a subexpression appear twice in the insn.
4933 Suppose that FROM is a register that appears within TO.
4934 Then, after that subexpression has been scanned once by `subst',
4935 the second time it is scanned, TO may be found. If we were
4936 to scan TO here, we would find FROM within it and create a
4937 self-referent rtl structure which is completely wrong. */
4938 if (COMBINE_RTX_EQUAL_P (x, to))
4939 return to;
4941 /* Parallel asm_operands need special attention because all of the
4942 inputs are shared across the arms. Furthermore, unsharing the
4943 rtl results in recognition failures. Failure to handle this case
4944 specially can result in circular rtl.
4946 Solve this by doing a normal pass across the first entry of the
4947 parallel, and only processing the SET_DESTs of the subsequent
4948 entries. Ug. */
4950 if (code == PARALLEL
4951 && GET_CODE (XVECEXP (x, 0, 0)) == SET
4952 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
4954 new_rtx = subst (XVECEXP (x, 0, 0), from, to, 0, 0, unique_copy);
4956 /* If this substitution failed, this whole thing fails. */
4957 if (GET_CODE (new_rtx) == CLOBBER
4958 && XEXP (new_rtx, 0) == const0_rtx)
4959 return new_rtx;
4961 SUBST (XVECEXP (x, 0, 0), new_rtx);
4963 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
4965 rtx dest = SET_DEST (XVECEXP (x, 0, i));
4967 if (!REG_P (dest)
4968 && GET_CODE (dest) != CC0
4969 && GET_CODE (dest) != PC)
4971 new_rtx = subst (dest, from, to, 0, 0, unique_copy);
4973 /* If this substitution failed, this whole thing fails. */
4974 if (GET_CODE (new_rtx) == CLOBBER
4975 && XEXP (new_rtx, 0) == const0_rtx)
4976 return new_rtx;
4978 SUBST (SET_DEST (XVECEXP (x, 0, i)), new_rtx);
4982 else
4984 len = GET_RTX_LENGTH (code);
4985 fmt = GET_RTX_FORMAT (code);
4987 /* We don't need to process a SET_DEST that is a register, CC0,
4988 or PC, so set up to skip this common case. All other cases
4989 where we want to suppress replacing something inside a
4990 SET_SRC are handled via the IN_DEST operand. */
4991 if (code == SET
4992 && (REG_P (SET_DEST (x))
4993 || GET_CODE (SET_DEST (x)) == CC0
4994 || GET_CODE (SET_DEST (x)) == PC))
4995 fmt = "ie";
4997 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
4998 constant. */
4999 if (fmt[0] == 'e')
5000 op0_mode = GET_MODE (XEXP (x, 0));
5002 for (i = 0; i < len; i++)
5004 if (fmt[i] == 'E')
5006 int j;
5007 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5009 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
5011 new_rtx = (unique_copy && n_occurrences
5012 ? copy_rtx (to) : to);
5013 n_occurrences++;
5015 else
5017 new_rtx = subst (XVECEXP (x, i, j), from, to, 0, 0,
5018 unique_copy);
5020 /* If this substitution failed, this whole thing
5021 fails. */
5022 if (GET_CODE (new_rtx) == CLOBBER
5023 && XEXP (new_rtx, 0) == const0_rtx)
5024 return new_rtx;
5027 SUBST (XVECEXP (x, i, j), new_rtx);
5030 else if (fmt[i] == 'e')
5032 /* If this is a register being set, ignore it. */
5033 new_rtx = XEXP (x, i);
5034 if (in_dest
5035 && i == 0
5036 && (((code == SUBREG || code == ZERO_EXTRACT)
5037 && REG_P (new_rtx))
5038 || code == STRICT_LOW_PART))
5041 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
5043 /* In general, don't install a subreg involving two
5044 modes not tieable. It can worsen register
5045 allocation, and can even make invalid reload
5046 insns, since the reg inside may need to be copied
5047 from in the outside mode, and that may be invalid
5048 if it is an fp reg copied in integer mode.
5050 We allow two exceptions to this: It is valid if
5051 it is inside another SUBREG and the mode of that
5052 SUBREG and the mode of the inside of TO is
5053 tieable and it is valid if X is a SET that copies
5054 FROM to CC0. */
5056 if (GET_CODE (to) == SUBREG
5057 && ! MODES_TIEABLE_P (GET_MODE (to),
5058 GET_MODE (SUBREG_REG (to)))
5059 && ! (code == SUBREG
5060 && MODES_TIEABLE_P (GET_MODE (x),
5061 GET_MODE (SUBREG_REG (to))))
5062 #ifdef HAVE_cc0
5063 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
5064 #endif
5066 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
5068 #ifdef CANNOT_CHANGE_MODE_CLASS
5069 if (code == SUBREG
5070 && REG_P (to)
5071 && REGNO (to) < FIRST_PSEUDO_REGISTER
5072 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
5073 GET_MODE (to),
5074 GET_MODE (x)))
5075 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
5076 #endif
5078 new_rtx = (unique_copy && n_occurrences ? copy_rtx (to) : to);
5079 n_occurrences++;
5081 else
5082 /* If we are in a SET_DEST, suppress most cases unless we
5083 have gone inside a MEM, in which case we want to
5084 simplify the address. We assume here that things that
5085 are actually part of the destination have their inner
5086 parts in the first expression. This is true for SUBREG,
5087 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
5088 things aside from REG and MEM that should appear in a
5089 SET_DEST. */
5090 new_rtx = subst (XEXP (x, i), from, to,
5091 (((in_dest
5092 && (code == SUBREG || code == STRICT_LOW_PART
5093 || code == ZERO_EXTRACT))
5094 || code == SET)
5095 && i == 0),
5096 code == IF_THEN_ELSE && i == 0,
5097 unique_copy);
5099 /* If we found that we will have to reject this combination,
5100 indicate that by returning the CLOBBER ourselves, rather than
5101 an expression containing it. This will speed things up as
5102 well as prevent accidents where two CLOBBERs are considered
5103 to be equal, thus producing an incorrect simplification. */
5105 if (GET_CODE (new_rtx) == CLOBBER && XEXP (new_rtx, 0) == const0_rtx)
5106 return new_rtx;
5108 if (GET_CODE (x) == SUBREG && CONST_SCALAR_INT_P (new_rtx))
5110 enum machine_mode mode = GET_MODE (x);
5112 x = simplify_subreg (GET_MODE (x), new_rtx,
5113 GET_MODE (SUBREG_REG (x)),
5114 SUBREG_BYTE (x));
5115 if (! x)
5116 x = gen_rtx_CLOBBER (mode, const0_rtx);
5118 else if (CONST_INT_P (new_rtx)
5119 && GET_CODE (x) == ZERO_EXTEND)
5121 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
5122 new_rtx, GET_MODE (XEXP (x, 0)));
5123 gcc_assert (x);
5125 else
5126 SUBST (XEXP (x, i), new_rtx);
5131 /* Check if we are loading something from the constant pool via float
5132 extension; in this case we would undo compress_float_constant
5133 optimization and degenerate constant load to an immediate value. */
5134 if (GET_CODE (x) == FLOAT_EXTEND
5135 && MEM_P (XEXP (x, 0))
5136 && MEM_READONLY_P (XEXP (x, 0)))
5138 rtx tmp = avoid_constant_pool_reference (x);
5139 if (x != tmp)
5140 return x;
5143 /* Try to simplify X. If the simplification changed the code, it is likely
5144 that further simplification will help, so loop, but limit the number
5145 of repetitions that will be performed. */
5147 for (i = 0; i < 4; i++)
5149 /* If X is sufficiently simple, don't bother trying to do anything
5150 with it. */
5151 if (code != CONST_INT && code != REG && code != CLOBBER)
5152 x = combine_simplify_rtx (x, op0_mode, in_dest, in_cond);
5154 if (GET_CODE (x) == code)
5155 break;
5157 code = GET_CODE (x);
5159 /* We no longer know the original mode of operand 0 since we
5160 have changed the form of X) */
5161 op0_mode = VOIDmode;
5164 return x;
5167 /* Simplify X, a piece of RTL. We just operate on the expression at the
5168 outer level; call `subst' to simplify recursively. Return the new
5169 expression.
5171 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
5172 if we are inside a SET_DEST. IN_COND is nonzero if we are at the top level
5173 of a condition. */
5175 static rtx
5176 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest,
5177 int in_cond)
5179 enum rtx_code code = GET_CODE (x);
5180 enum machine_mode mode = GET_MODE (x);
5181 rtx temp;
5182 int i;
5184 /* If this is a commutative operation, put a constant last and a complex
5185 expression first. We don't need to do this for comparisons here. */
5186 if (COMMUTATIVE_ARITH_P (x)
5187 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
5189 temp = XEXP (x, 0);
5190 SUBST (XEXP (x, 0), XEXP (x, 1));
5191 SUBST (XEXP (x, 1), temp);
5194 /* If this is a simple operation applied to an IF_THEN_ELSE, try
5195 applying it to the arms of the IF_THEN_ELSE. This often simplifies
5196 things. Check for cases where both arms are testing the same
5197 condition.
5199 Don't do anything if all operands are very simple. */
5201 if ((BINARY_P (x)
5202 && ((!OBJECT_P (XEXP (x, 0))
5203 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
5204 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
5205 || (!OBJECT_P (XEXP (x, 1))
5206 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
5207 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
5208 || (UNARY_P (x)
5209 && (!OBJECT_P (XEXP (x, 0))
5210 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
5211 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
5213 rtx cond, true_rtx, false_rtx;
5215 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
5216 if (cond != 0
5217 /* If everything is a comparison, what we have is highly unlikely
5218 to be simpler, so don't use it. */
5219 && ! (COMPARISON_P (x)
5220 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
5222 rtx cop1 = const0_rtx;
5223 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
5225 if (cond_code == NE && COMPARISON_P (cond))
5226 return x;
5228 /* Simplify the alternative arms; this may collapse the true and
5229 false arms to store-flag values. Be careful to use copy_rtx
5230 here since true_rtx or false_rtx might share RTL with x as a
5231 result of the if_then_else_cond call above. */
5232 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0, 0);
5233 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0, 0);
5235 /* If true_rtx and false_rtx are not general_operands, an if_then_else
5236 is unlikely to be simpler. */
5237 if (general_operand (true_rtx, VOIDmode)
5238 && general_operand (false_rtx, VOIDmode))
5240 enum rtx_code reversed;
5242 /* Restarting if we generate a store-flag expression will cause
5243 us to loop. Just drop through in this case. */
5245 /* If the result values are STORE_FLAG_VALUE and zero, we can
5246 just make the comparison operation. */
5247 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
5248 x = simplify_gen_relational (cond_code, mode, VOIDmode,
5249 cond, cop1);
5250 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
5251 && ((reversed = reversed_comparison_code_parts
5252 (cond_code, cond, cop1, NULL))
5253 != UNKNOWN))
5254 x = simplify_gen_relational (reversed, mode, VOIDmode,
5255 cond, cop1);
5257 /* Likewise, we can make the negate of a comparison operation
5258 if the result values are - STORE_FLAG_VALUE and zero. */
5259 else if (CONST_INT_P (true_rtx)
5260 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
5261 && false_rtx == const0_rtx)
5262 x = simplify_gen_unary (NEG, mode,
5263 simplify_gen_relational (cond_code,
5264 mode, VOIDmode,
5265 cond, cop1),
5266 mode);
5267 else if (CONST_INT_P (false_rtx)
5268 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
5269 && true_rtx == const0_rtx
5270 && ((reversed = reversed_comparison_code_parts
5271 (cond_code, cond, cop1, NULL))
5272 != UNKNOWN))
5273 x = simplify_gen_unary (NEG, mode,
5274 simplify_gen_relational (reversed,
5275 mode, VOIDmode,
5276 cond, cop1),
5277 mode);
5278 else
5279 return gen_rtx_IF_THEN_ELSE (mode,
5280 simplify_gen_relational (cond_code,
5281 mode,
5282 VOIDmode,
5283 cond,
5284 cop1),
5285 true_rtx, false_rtx);
5287 code = GET_CODE (x);
5288 op0_mode = VOIDmode;
5293 /* Try to fold this expression in case we have constants that weren't
5294 present before. */
5295 temp = 0;
5296 switch (GET_RTX_CLASS (code))
5298 case RTX_UNARY:
5299 if (op0_mode == VOIDmode)
5300 op0_mode = GET_MODE (XEXP (x, 0));
5301 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
5302 break;
5303 case RTX_COMPARE:
5304 case RTX_COMM_COMPARE:
5306 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
5307 if (cmp_mode == VOIDmode)
5309 cmp_mode = GET_MODE (XEXP (x, 1));
5310 if (cmp_mode == VOIDmode)
5311 cmp_mode = op0_mode;
5313 temp = simplify_relational_operation (code, mode, cmp_mode,
5314 XEXP (x, 0), XEXP (x, 1));
5316 break;
5317 case RTX_COMM_ARITH:
5318 case RTX_BIN_ARITH:
5319 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
5320 break;
5321 case RTX_BITFIELD_OPS:
5322 case RTX_TERNARY:
5323 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
5324 XEXP (x, 1), XEXP (x, 2));
5325 break;
5326 default:
5327 break;
5330 if (temp)
5332 x = temp;
5333 code = GET_CODE (temp);
5334 op0_mode = VOIDmode;
5335 mode = GET_MODE (temp);
5338 /* First see if we can apply the inverse distributive law. */
5339 if (code == PLUS || code == MINUS
5340 || code == AND || code == IOR || code == XOR)
5342 x = apply_distributive_law (x);
5343 code = GET_CODE (x);
5344 op0_mode = VOIDmode;
5347 /* If CODE is an associative operation not otherwise handled, see if we
5348 can associate some operands. This can win if they are constants or
5349 if they are logically related (i.e. (a & b) & a). */
5350 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
5351 || code == AND || code == IOR || code == XOR
5352 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
5353 && ((INTEGRAL_MODE_P (mode) && code != DIV)
5354 || (flag_associative_math && FLOAT_MODE_P (mode))))
5356 if (GET_CODE (XEXP (x, 0)) == code)
5358 rtx other = XEXP (XEXP (x, 0), 0);
5359 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
5360 rtx inner_op1 = XEXP (x, 1);
5361 rtx inner;
5363 /* Make sure we pass the constant operand if any as the second
5364 one if this is a commutative operation. */
5365 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
5367 rtx tem = inner_op0;
5368 inner_op0 = inner_op1;
5369 inner_op1 = tem;
5371 inner = simplify_binary_operation (code == MINUS ? PLUS
5372 : code == DIV ? MULT
5373 : code,
5374 mode, inner_op0, inner_op1);
5376 /* For commutative operations, try the other pair if that one
5377 didn't simplify. */
5378 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
5380 other = XEXP (XEXP (x, 0), 1);
5381 inner = simplify_binary_operation (code, mode,
5382 XEXP (XEXP (x, 0), 0),
5383 XEXP (x, 1));
5386 if (inner)
5387 return simplify_gen_binary (code, mode, other, inner);
5391 /* A little bit of algebraic simplification here. */
5392 switch (code)
5394 case MEM:
5395 /* Ensure that our address has any ASHIFTs converted to MULT in case
5396 address-recognizing predicates are called later. */
5397 temp = make_compound_operation (XEXP (x, 0), MEM);
5398 SUBST (XEXP (x, 0), temp);
5399 break;
5401 case SUBREG:
5402 if (op0_mode == VOIDmode)
5403 op0_mode = GET_MODE (SUBREG_REG (x));
5405 /* See if this can be moved to simplify_subreg. */
5406 if (CONSTANT_P (SUBREG_REG (x))
5407 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
5408 /* Don't call gen_lowpart if the inner mode
5409 is VOIDmode and we cannot simplify it, as SUBREG without
5410 inner mode is invalid. */
5411 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
5412 || gen_lowpart_common (mode, SUBREG_REG (x))))
5413 return gen_lowpart (mode, SUBREG_REG (x));
5415 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
5416 break;
5418 rtx temp;
5419 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,
5420 SUBREG_BYTE (x));
5421 if (temp)
5422 return temp;
5424 /* If op is known to have all lower bits zero, the result is zero. */
5425 if (!in_dest
5426 && SCALAR_INT_MODE_P (mode)
5427 && SCALAR_INT_MODE_P (op0_mode)
5428 && GET_MODE_PRECISION (mode) < GET_MODE_PRECISION (op0_mode)
5429 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
5430 && HWI_COMPUTABLE_MODE_P (op0_mode)
5431 && (nonzero_bits (SUBREG_REG (x), op0_mode)
5432 & GET_MODE_MASK (mode)) == 0)
5433 return CONST0_RTX (mode);
5436 /* Don't change the mode of the MEM if that would change the meaning
5437 of the address. */
5438 if (MEM_P (SUBREG_REG (x))
5439 && (MEM_VOLATILE_P (SUBREG_REG (x))
5440 || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0),
5441 MEM_ADDR_SPACE (SUBREG_REG (x)))))
5442 return gen_rtx_CLOBBER (mode, const0_rtx);
5444 /* Note that we cannot do any narrowing for non-constants since
5445 we might have been counting on using the fact that some bits were
5446 zero. We now do this in the SET. */
5448 break;
5450 case NEG:
5451 temp = expand_compound_operation (XEXP (x, 0));
5453 /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be
5454 replaced by (lshiftrt X C). This will convert
5455 (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */
5457 if (GET_CODE (temp) == ASHIFTRT
5458 && CONST_INT_P (XEXP (temp, 1))
5459 && INTVAL (XEXP (temp, 1)) == GET_MODE_PRECISION (mode) - 1)
5460 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (temp, 0),
5461 INTVAL (XEXP (temp, 1)));
5463 /* If X has only a single bit that might be nonzero, say, bit I, convert
5464 (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of
5465 MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to
5466 (sign_extract X 1 Y). But only do this if TEMP isn't a register
5467 or a SUBREG of one since we'd be making the expression more
5468 complex if it was just a register. */
5470 if (!REG_P (temp)
5471 && ! (GET_CODE (temp) == SUBREG
5472 && REG_P (SUBREG_REG (temp)))
5473 && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0)
5475 rtx temp1 = simplify_shift_const
5476 (NULL_RTX, ASHIFTRT, mode,
5477 simplify_shift_const (NULL_RTX, ASHIFT, mode, temp,
5478 GET_MODE_PRECISION (mode) - 1 - i),
5479 GET_MODE_PRECISION (mode) - 1 - i);
5481 /* If all we did was surround TEMP with the two shifts, we
5482 haven't improved anything, so don't use it. Otherwise,
5483 we are better off with TEMP1. */
5484 if (GET_CODE (temp1) != ASHIFTRT
5485 || GET_CODE (XEXP (temp1, 0)) != ASHIFT
5486 || XEXP (XEXP (temp1, 0), 0) != temp)
5487 return temp1;
5489 break;
5491 case TRUNCATE:
5492 /* We can't handle truncation to a partial integer mode here
5493 because we don't know the real bitsize of the partial
5494 integer mode. */
5495 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
5496 break;
5498 if (HWI_COMPUTABLE_MODE_P (mode))
5499 SUBST (XEXP (x, 0),
5500 force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5501 GET_MODE_MASK (mode), 0));
5503 /* We can truncate a constant value and return it. */
5504 if (CONST_INT_P (XEXP (x, 0)))
5505 return gen_int_mode (INTVAL (XEXP (x, 0)), mode);
5507 /* Similarly to what we do in simplify-rtx.c, a truncate of a register
5508 whose value is a comparison can be replaced with a subreg if
5509 STORE_FLAG_VALUE permits. */
5510 if (HWI_COMPUTABLE_MODE_P (mode)
5511 && (STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0
5512 && (temp = get_last_value (XEXP (x, 0)))
5513 && COMPARISON_P (temp))
5514 return gen_lowpart (mode, XEXP (x, 0));
5515 break;
5517 case CONST:
5518 /* (const (const X)) can become (const X). Do it this way rather than
5519 returning the inner CONST since CONST can be shared with a
5520 REG_EQUAL note. */
5521 if (GET_CODE (XEXP (x, 0)) == CONST)
5522 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
5523 break;
5525 #ifdef HAVE_lo_sum
5526 case LO_SUM:
5527 /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we
5528 can add in an offset. find_split_point will split this address up
5529 again if it doesn't match. */
5530 if (GET_CODE (XEXP (x, 0)) == HIGH
5531 && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
5532 return XEXP (x, 1);
5533 break;
5534 #endif
5536 case PLUS:
5537 /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>)
5538 when c is (const_int (pow2 + 1) / 2) is a sign extension of a
5539 bit-field and can be replaced by either a sign_extend or a
5540 sign_extract. The `and' may be a zero_extend and the two
5541 <c>, -<c> constants may be reversed. */
5542 if (GET_CODE (XEXP (x, 0)) == XOR
5543 && CONST_INT_P (XEXP (x, 1))
5544 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
5545 && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1))
5546 && ((i = exact_log2 (UINTVAL (XEXP (XEXP (x, 0), 1)))) >= 0
5547 || (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0)
5548 && HWI_COMPUTABLE_MODE_P (mode)
5549 && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND
5550 && CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 0), 1))
5551 && (UINTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))
5552 == ((unsigned HOST_WIDE_INT) 1 << (i + 1)) - 1))
5553 || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND
5554 && (GET_MODE_PRECISION (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))
5555 == (unsigned int) i + 1))))
5556 return simplify_shift_const
5557 (NULL_RTX, ASHIFTRT, mode,
5558 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5559 XEXP (XEXP (XEXP (x, 0), 0), 0),
5560 GET_MODE_PRECISION (mode) - (i + 1)),
5561 GET_MODE_PRECISION (mode) - (i + 1));
5563 /* If only the low-order bit of X is possibly nonzero, (plus x -1)
5564 can become (ashiftrt (ashift (xor x 1) C) C) where C is
5565 the bitsize of the mode - 1. This allows simplification of
5566 "a = (b & 8) == 0;" */
5567 if (XEXP (x, 1) == constm1_rtx
5568 && !REG_P (XEXP (x, 0))
5569 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
5570 && REG_P (SUBREG_REG (XEXP (x, 0))))
5571 && nonzero_bits (XEXP (x, 0), mode) == 1)
5572 return simplify_shift_const (NULL_RTX, ASHIFTRT, mode,
5573 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5574 gen_rtx_XOR (mode, XEXP (x, 0), const1_rtx),
5575 GET_MODE_PRECISION (mode) - 1),
5576 GET_MODE_PRECISION (mode) - 1);
5578 /* If we are adding two things that have no bits in common, convert
5579 the addition into an IOR. This will often be further simplified,
5580 for example in cases like ((a & 1) + (a & 2)), which can
5581 become a & 3. */
5583 if (HWI_COMPUTABLE_MODE_P (mode)
5584 && (nonzero_bits (XEXP (x, 0), mode)
5585 & nonzero_bits (XEXP (x, 1), mode)) == 0)
5587 /* Try to simplify the expression further. */
5588 rtx tor = simplify_gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1));
5589 temp = combine_simplify_rtx (tor, VOIDmode, in_dest, 0);
5591 /* If we could, great. If not, do not go ahead with the IOR
5592 replacement, since PLUS appears in many special purpose
5593 address arithmetic instructions. */
5594 if (GET_CODE (temp) != CLOBBER
5595 && (GET_CODE (temp) != IOR
5596 || ((XEXP (temp, 0) != XEXP (x, 0)
5597 || XEXP (temp, 1) != XEXP (x, 1))
5598 && (XEXP (temp, 0) != XEXP (x, 1)
5599 || XEXP (temp, 1) != XEXP (x, 0)))))
5600 return temp;
5602 break;
5604 case MINUS:
5605 /* (minus <foo> (and <foo> (const_int -pow2))) becomes
5606 (and <foo> (const_int pow2-1)) */
5607 if (GET_CODE (XEXP (x, 1)) == AND
5608 && CONST_INT_P (XEXP (XEXP (x, 1), 1))
5609 && exact_log2 (-UINTVAL (XEXP (XEXP (x, 1), 1))) >= 0
5610 && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
5611 return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0),
5612 -INTVAL (XEXP (XEXP (x, 1), 1)) - 1);
5613 break;
5615 case MULT:
5616 /* If we have (mult (plus A B) C), apply the distributive law and then
5617 the inverse distributive law to see if things simplify. This
5618 occurs mostly in addresses, often when unrolling loops. */
5620 if (GET_CODE (XEXP (x, 0)) == PLUS)
5622 rtx result = distribute_and_simplify_rtx (x, 0);
5623 if (result)
5624 return result;
5627 /* Try simplify a*(b/c) as (a*b)/c. */
5628 if (FLOAT_MODE_P (mode) && flag_associative_math
5629 && GET_CODE (XEXP (x, 0)) == DIV)
5631 rtx tem = simplify_binary_operation (MULT, mode,
5632 XEXP (XEXP (x, 0), 0),
5633 XEXP (x, 1));
5634 if (tem)
5635 return simplify_gen_binary (DIV, mode, tem, XEXP (XEXP (x, 0), 1));
5637 break;
5639 case UDIV:
5640 /* If this is a divide by a power of two, treat it as a shift if
5641 its first operand is a shift. */
5642 if (CONST_INT_P (XEXP (x, 1))
5643 && (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0
5644 && (GET_CODE (XEXP (x, 0)) == ASHIFT
5645 || GET_CODE (XEXP (x, 0)) == LSHIFTRT
5646 || GET_CODE (XEXP (x, 0)) == ASHIFTRT
5647 || GET_CODE (XEXP (x, 0)) == ROTATE
5648 || GET_CODE (XEXP (x, 0)) == ROTATERT))
5649 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i);
5650 break;
5652 case EQ: case NE:
5653 case GT: case GTU: case GE: case GEU:
5654 case LT: case LTU: case LE: case LEU:
5655 case UNEQ: case LTGT:
5656 case UNGT: case UNGE:
5657 case UNLT: case UNLE:
5658 case UNORDERED: case ORDERED:
5659 /* If the first operand is a condition code, we can't do anything
5660 with it. */
5661 if (GET_CODE (XEXP (x, 0)) == COMPARE
5662 || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC
5663 && ! CC0_P (XEXP (x, 0))))
5665 rtx op0 = XEXP (x, 0);
5666 rtx op1 = XEXP (x, 1);
5667 enum rtx_code new_code;
5669 if (GET_CODE (op0) == COMPARE)
5670 op1 = XEXP (op0, 1), op0 = XEXP (op0, 0);
5672 /* Simplify our comparison, if possible. */
5673 new_code = simplify_comparison (code, &op0, &op1);
5675 /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X
5676 if only the low-order bit is possibly nonzero in X (such as when
5677 X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to
5678 (xor X 1) or (minus 1 X); we use the former. Finally, if X is
5679 known to be either 0 or -1, NE becomes a NEG and EQ becomes
5680 (plus X 1).
5682 Remove any ZERO_EXTRACT we made when thinking this was a
5683 comparison. It may now be simpler to use, e.g., an AND. If a
5684 ZERO_EXTRACT is indeed appropriate, it will be placed back by
5685 the call to make_compound_operation in the SET case.
5687 Don't apply these optimizations if the caller would
5688 prefer a comparison rather than a value.
5689 E.g., for the condition in an IF_THEN_ELSE most targets need
5690 an explicit comparison. */
5692 if (in_cond)
5695 else if (STORE_FLAG_VALUE == 1
5696 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
5697 && op1 == const0_rtx
5698 && mode == GET_MODE (op0)
5699 && nonzero_bits (op0, mode) == 1)
5700 return gen_lowpart (mode,
5701 expand_compound_operation (op0));
5703 else if (STORE_FLAG_VALUE == 1
5704 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
5705 && op1 == const0_rtx
5706 && mode == GET_MODE (op0)
5707 && (num_sign_bit_copies (op0, mode)
5708 == GET_MODE_PRECISION (mode)))
5710 op0 = expand_compound_operation (op0);
5711 return simplify_gen_unary (NEG, mode,
5712 gen_lowpart (mode, op0),
5713 mode);
5716 else if (STORE_FLAG_VALUE == 1
5717 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
5718 && op1 == const0_rtx
5719 && mode == GET_MODE (op0)
5720 && nonzero_bits (op0, mode) == 1)
5722 op0 = expand_compound_operation (op0);
5723 return simplify_gen_binary (XOR, mode,
5724 gen_lowpart (mode, op0),
5725 const1_rtx);
5728 else if (STORE_FLAG_VALUE == 1
5729 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
5730 && op1 == const0_rtx
5731 && mode == GET_MODE (op0)
5732 && (num_sign_bit_copies (op0, mode)
5733 == GET_MODE_PRECISION (mode)))
5735 op0 = expand_compound_operation (op0);
5736 return plus_constant (mode, gen_lowpart (mode, op0), 1);
5739 /* If STORE_FLAG_VALUE is -1, we have cases similar to
5740 those above. */
5741 if (in_cond)
5744 else if (STORE_FLAG_VALUE == -1
5745 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
5746 && op1 == const0_rtx
5747 && (num_sign_bit_copies (op0, mode)
5748 == GET_MODE_PRECISION (mode)))
5749 return gen_lowpart (mode,
5750 expand_compound_operation (op0));
5752 else if (STORE_FLAG_VALUE == -1
5753 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
5754 && op1 == const0_rtx
5755 && mode == GET_MODE (op0)
5756 && nonzero_bits (op0, mode) == 1)
5758 op0 = expand_compound_operation (op0);
5759 return simplify_gen_unary (NEG, mode,
5760 gen_lowpart (mode, op0),
5761 mode);
5764 else if (STORE_FLAG_VALUE == -1
5765 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
5766 && op1 == const0_rtx
5767 && mode == GET_MODE (op0)
5768 && (num_sign_bit_copies (op0, mode)
5769 == GET_MODE_PRECISION (mode)))
5771 op0 = expand_compound_operation (op0);
5772 return simplify_gen_unary (NOT, mode,
5773 gen_lowpart (mode, op0),
5774 mode);
5777 /* If X is 0/1, (eq X 0) is X-1. */
5778 else if (STORE_FLAG_VALUE == -1
5779 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
5780 && op1 == const0_rtx
5781 && mode == GET_MODE (op0)
5782 && nonzero_bits (op0, mode) == 1)
5784 op0 = expand_compound_operation (op0);
5785 return plus_constant (mode, gen_lowpart (mode, op0), -1);
5788 /* If STORE_FLAG_VALUE says to just test the sign bit and X has just
5789 one bit that might be nonzero, we can convert (ne x 0) to
5790 (ashift x c) where C puts the bit in the sign bit. Remove any
5791 AND with STORE_FLAG_VALUE when we are done, since we are only
5792 going to test the sign bit. */
5793 if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
5794 && HWI_COMPUTABLE_MODE_P (mode)
5795 && val_signbit_p (mode, STORE_FLAG_VALUE)
5796 && op1 == const0_rtx
5797 && mode == GET_MODE (op0)
5798 && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0)
5800 x = simplify_shift_const (NULL_RTX, ASHIFT, mode,
5801 expand_compound_operation (op0),
5802 GET_MODE_PRECISION (mode) - 1 - i);
5803 if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx)
5804 return XEXP (x, 0);
5805 else
5806 return x;
5809 /* If the code changed, return a whole new comparison.
5810 We also need to avoid using SUBST in cases where
5811 simplify_comparison has widened a comparison with a CONST_INT,
5812 since in that case the wider CONST_INT may fail the sanity
5813 checks in do_SUBST. */
5814 if (new_code != code
5815 || (CONST_INT_P (op1)
5816 && GET_MODE (op0) != GET_MODE (XEXP (x, 0))
5817 && GET_MODE (op0) != GET_MODE (XEXP (x, 1))))
5818 return gen_rtx_fmt_ee (new_code, mode, op0, op1);
5820 /* Otherwise, keep this operation, but maybe change its operands.
5821 This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */
5822 SUBST (XEXP (x, 0), op0);
5823 SUBST (XEXP (x, 1), op1);
5825 break;
5827 case IF_THEN_ELSE:
5828 return simplify_if_then_else (x);
5830 case ZERO_EXTRACT:
5831 case SIGN_EXTRACT:
5832 case ZERO_EXTEND:
5833 case SIGN_EXTEND:
5834 /* If we are processing SET_DEST, we are done. */
5835 if (in_dest)
5836 return x;
5838 return expand_compound_operation (x);
5840 case SET:
5841 return simplify_set (x);
5843 case AND:
5844 case IOR:
5845 return simplify_logical (x);
5847 case ASHIFT:
5848 case LSHIFTRT:
5849 case ASHIFTRT:
5850 case ROTATE:
5851 case ROTATERT:
5852 /* If this is a shift by a constant amount, simplify it. */
5853 if (CONST_INT_P (XEXP (x, 1)))
5854 return simplify_shift_const (x, code, mode, XEXP (x, 0),
5855 INTVAL (XEXP (x, 1)));
5857 else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1)))
5858 SUBST (XEXP (x, 1),
5859 force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)),
5860 ((unsigned HOST_WIDE_INT) 1
5861 << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))))
5862 - 1,
5863 0));
5864 break;
5866 default:
5867 break;
5870 return x;
5873 /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */
5875 static rtx
5876 simplify_if_then_else (rtx x)
5878 enum machine_mode mode = GET_MODE (x);
5879 rtx cond = XEXP (x, 0);
5880 rtx true_rtx = XEXP (x, 1);
5881 rtx false_rtx = XEXP (x, 2);
5882 enum rtx_code true_code = GET_CODE (cond);
5883 int comparison_p = COMPARISON_P (cond);
5884 rtx temp;
5885 int i;
5886 enum rtx_code false_code;
5887 rtx reversed;
5889 /* Simplify storing of the truth value. */
5890 if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx)
5891 return simplify_gen_relational (true_code, mode, VOIDmode,
5892 XEXP (cond, 0), XEXP (cond, 1));
5894 /* Also when the truth value has to be reversed. */
5895 if (comparison_p
5896 && true_rtx == const0_rtx && false_rtx == const_true_rtx
5897 && (reversed = reversed_comparison (cond, mode)))
5898 return reversed;
5900 /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used
5901 in it is being compared against certain values. Get the true and false
5902 comparisons and see if that says anything about the value of each arm. */
5904 if (comparison_p
5905 && ((false_code = reversed_comparison_code (cond, NULL))
5906 != UNKNOWN)
5907 && REG_P (XEXP (cond, 0)))
5909 HOST_WIDE_INT nzb;
5910 rtx from = XEXP (cond, 0);
5911 rtx true_val = XEXP (cond, 1);
5912 rtx false_val = true_val;
5913 int swapped = 0;
5915 /* If FALSE_CODE is EQ, swap the codes and arms. */
5917 if (false_code == EQ)
5919 swapped = 1, true_code = EQ, false_code = NE;
5920 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
5923 /* If we are comparing against zero and the expression being tested has
5924 only a single bit that might be nonzero, that is its value when it is
5925 not equal to zero. Similarly if it is known to be -1 or 0. */
5927 if (true_code == EQ && true_val == const0_rtx
5928 && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0)
5930 false_code = EQ;
5931 false_val = gen_int_mode (nzb, GET_MODE (from));
5933 else if (true_code == EQ && true_val == const0_rtx
5934 && (num_sign_bit_copies (from, GET_MODE (from))
5935 == GET_MODE_PRECISION (GET_MODE (from))))
5937 false_code = EQ;
5938 false_val = constm1_rtx;
5941 /* Now simplify an arm if we know the value of the register in the
5942 branch and it is used in the arm. Be careful due to the potential
5943 of locally-shared RTL. */
5945 if (reg_mentioned_p (from, true_rtx))
5946 true_rtx = subst (known_cond (copy_rtx (true_rtx), true_code,
5947 from, true_val),
5948 pc_rtx, pc_rtx, 0, 0, 0);
5949 if (reg_mentioned_p (from, false_rtx))
5950 false_rtx = subst (known_cond (copy_rtx (false_rtx), false_code,
5951 from, false_val),
5952 pc_rtx, pc_rtx, 0, 0, 0);
5954 SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx);
5955 SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx);
5957 true_rtx = XEXP (x, 1);
5958 false_rtx = XEXP (x, 2);
5959 true_code = GET_CODE (cond);
5962 /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be
5963 reversed, do so to avoid needing two sets of patterns for
5964 subtract-and-branch insns. Similarly if we have a constant in the true
5965 arm, the false arm is the same as the first operand of the comparison, or
5966 the false arm is more complicated than the true arm. */
5968 if (comparison_p
5969 && reversed_comparison_code (cond, NULL) != UNKNOWN
5970 && (true_rtx == pc_rtx
5971 || (CONSTANT_P (true_rtx)
5972 && !CONST_INT_P (false_rtx) && false_rtx != pc_rtx)
5973 || true_rtx == const0_rtx
5974 || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx))
5975 || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx))
5976 && !OBJECT_P (false_rtx))
5977 || reg_mentioned_p (true_rtx, false_rtx)
5978 || rtx_equal_p (false_rtx, XEXP (cond, 0))))
5980 true_code = reversed_comparison_code (cond, NULL);
5981 SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond)));
5982 SUBST (XEXP (x, 1), false_rtx);
5983 SUBST (XEXP (x, 2), true_rtx);
5985 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
5986 cond = XEXP (x, 0);
5988 /* It is possible that the conditional has been simplified out. */
5989 true_code = GET_CODE (cond);
5990 comparison_p = COMPARISON_P (cond);
5993 /* If the two arms are identical, we don't need the comparison. */
5995 if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond))
5996 return true_rtx;
5998 /* Convert a == b ? b : a to "a". */
5999 if (true_code == EQ && ! side_effects_p (cond)
6000 && !HONOR_NANS (mode)
6001 && rtx_equal_p (XEXP (cond, 0), false_rtx)
6002 && rtx_equal_p (XEXP (cond, 1), true_rtx))
6003 return false_rtx;
6004 else if (true_code == NE && ! side_effects_p (cond)
6005 && !HONOR_NANS (mode)
6006 && rtx_equal_p (XEXP (cond, 0), true_rtx)
6007 && rtx_equal_p (XEXP (cond, 1), false_rtx))
6008 return true_rtx;
6010 /* Look for cases where we have (abs x) or (neg (abs X)). */
6012 if (GET_MODE_CLASS (mode) == MODE_INT
6013 && comparison_p
6014 && XEXP (cond, 1) == const0_rtx
6015 && GET_CODE (false_rtx) == NEG
6016 && rtx_equal_p (true_rtx, XEXP (false_rtx, 0))
6017 && rtx_equal_p (true_rtx, XEXP (cond, 0))
6018 && ! side_effects_p (true_rtx))
6019 switch (true_code)
6021 case GT:
6022 case GE:
6023 return simplify_gen_unary (ABS, mode, true_rtx, mode);
6024 case LT:
6025 case LE:
6026 return
6027 simplify_gen_unary (NEG, mode,
6028 simplify_gen_unary (ABS, mode, true_rtx, mode),
6029 mode);
6030 default:
6031 break;
6034 /* Look for MIN or MAX. */
6036 if ((! FLOAT_MODE_P (mode) || flag_unsafe_math_optimizations)
6037 && comparison_p
6038 && rtx_equal_p (XEXP (cond, 0), true_rtx)
6039 && rtx_equal_p (XEXP (cond, 1), false_rtx)
6040 && ! side_effects_p (cond))
6041 switch (true_code)
6043 case GE:
6044 case GT:
6045 return simplify_gen_binary (SMAX, mode, true_rtx, false_rtx);
6046 case LE:
6047 case LT:
6048 return simplify_gen_binary (SMIN, mode, true_rtx, false_rtx);
6049 case GEU:
6050 case GTU:
6051 return simplify_gen_binary (UMAX, mode, true_rtx, false_rtx);
6052 case LEU:
6053 case LTU:
6054 return simplify_gen_binary (UMIN, mode, true_rtx, false_rtx);
6055 default:
6056 break;
6059 /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its
6060 second operand is zero, this can be done as (OP Z (mult COND C2)) where
6061 C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or
6062 SIGN_EXTEND as long as Z is already extended (so we don't destroy it).
6063 We can do this kind of thing in some cases when STORE_FLAG_VALUE is
6064 neither 1 or -1, but it isn't worth checking for. */
6066 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
6067 && comparison_p
6068 && GET_MODE_CLASS (mode) == MODE_INT
6069 && ! side_effects_p (x))
6071 rtx t = make_compound_operation (true_rtx, SET);
6072 rtx f = make_compound_operation (false_rtx, SET);
6073 rtx cond_op0 = XEXP (cond, 0);
6074 rtx cond_op1 = XEXP (cond, 1);
6075 enum rtx_code op = UNKNOWN, extend_op = UNKNOWN;
6076 enum machine_mode m = mode;
6077 rtx z = 0, c1 = NULL_RTX;
6079 if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS
6080 || GET_CODE (t) == IOR || GET_CODE (t) == XOR
6081 || GET_CODE (t) == ASHIFT
6082 || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT)
6083 && rtx_equal_p (XEXP (t, 0), f))
6084 c1 = XEXP (t, 1), op = GET_CODE (t), z = f;
6086 /* If an identity-zero op is commutative, check whether there
6087 would be a match if we swapped the operands. */
6088 else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR
6089 || GET_CODE (t) == XOR)
6090 && rtx_equal_p (XEXP (t, 1), f))
6091 c1 = XEXP (t, 0), op = GET_CODE (t), z = f;
6092 else if (GET_CODE (t) == SIGN_EXTEND
6093 && (GET_CODE (XEXP (t, 0)) == PLUS
6094 || GET_CODE (XEXP (t, 0)) == MINUS
6095 || GET_CODE (XEXP (t, 0)) == IOR
6096 || GET_CODE (XEXP (t, 0)) == XOR
6097 || GET_CODE (XEXP (t, 0)) == ASHIFT
6098 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
6099 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
6100 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
6101 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
6102 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
6103 && (num_sign_bit_copies (f, GET_MODE (f))
6104 > (unsigned int)
6105 (GET_MODE_PRECISION (mode)
6106 - GET_MODE_PRECISION (GET_MODE (XEXP (XEXP (t, 0), 0))))))
6108 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
6109 extend_op = SIGN_EXTEND;
6110 m = GET_MODE (XEXP (t, 0));
6112 else if (GET_CODE (t) == SIGN_EXTEND
6113 && (GET_CODE (XEXP (t, 0)) == PLUS
6114 || GET_CODE (XEXP (t, 0)) == IOR
6115 || GET_CODE (XEXP (t, 0)) == XOR)
6116 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
6117 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
6118 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
6119 && (num_sign_bit_copies (f, GET_MODE (f))
6120 > (unsigned int)
6121 (GET_MODE_PRECISION (mode)
6122 - GET_MODE_PRECISION (GET_MODE (XEXP (XEXP (t, 0), 1))))))
6124 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
6125 extend_op = SIGN_EXTEND;
6126 m = GET_MODE (XEXP (t, 0));
6128 else if (GET_CODE (t) == ZERO_EXTEND
6129 && (GET_CODE (XEXP (t, 0)) == PLUS
6130 || GET_CODE (XEXP (t, 0)) == MINUS
6131 || GET_CODE (XEXP (t, 0)) == IOR
6132 || GET_CODE (XEXP (t, 0)) == XOR
6133 || GET_CODE (XEXP (t, 0)) == ASHIFT
6134 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
6135 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
6136 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
6137 && HWI_COMPUTABLE_MODE_P (mode)
6138 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
6139 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
6140 && ((nonzero_bits (f, GET_MODE (f))
6141 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0))))
6142 == 0))
6144 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
6145 extend_op = ZERO_EXTEND;
6146 m = GET_MODE (XEXP (t, 0));
6148 else if (GET_CODE (t) == ZERO_EXTEND
6149 && (GET_CODE (XEXP (t, 0)) == PLUS
6150 || GET_CODE (XEXP (t, 0)) == IOR
6151 || GET_CODE (XEXP (t, 0)) == XOR)
6152 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
6153 && HWI_COMPUTABLE_MODE_P (mode)
6154 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
6155 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
6156 && ((nonzero_bits (f, GET_MODE (f))
6157 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1))))
6158 == 0))
6160 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
6161 extend_op = ZERO_EXTEND;
6162 m = GET_MODE (XEXP (t, 0));
6165 if (z)
6167 temp = subst (simplify_gen_relational (true_code, m, VOIDmode,
6168 cond_op0, cond_op1),
6169 pc_rtx, pc_rtx, 0, 0, 0);
6170 temp = simplify_gen_binary (MULT, m, temp,
6171 simplify_gen_binary (MULT, m, c1,
6172 const_true_rtx));
6173 temp = subst (temp, pc_rtx, pc_rtx, 0, 0, 0);
6174 temp = simplify_gen_binary (op, m, gen_lowpart (m, z), temp);
6176 if (extend_op != UNKNOWN)
6177 temp = simplify_gen_unary (extend_op, mode, temp, m);
6179 return temp;
6183 /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or
6184 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the
6185 negation of a single bit, we can convert this operation to a shift. We
6186 can actually do this more generally, but it doesn't seem worth it. */
6188 if (true_code == NE && XEXP (cond, 1) == const0_rtx
6189 && false_rtx == const0_rtx && CONST_INT_P (true_rtx)
6190 && ((1 == nonzero_bits (XEXP (cond, 0), mode)
6191 && (i = exact_log2 (UINTVAL (true_rtx))) >= 0)
6192 || ((num_sign_bit_copies (XEXP (cond, 0), mode)
6193 == GET_MODE_PRECISION (mode))
6194 && (i = exact_log2 (-UINTVAL (true_rtx))) >= 0)))
6195 return
6196 simplify_shift_const (NULL_RTX, ASHIFT, mode,
6197 gen_lowpart (mode, XEXP (cond, 0)), i);
6199 /* (IF_THEN_ELSE (NE REG 0) (0) (8)) is REG for nonzero_bits (REG) == 8. */
6200 if (true_code == NE && XEXP (cond, 1) == const0_rtx
6201 && false_rtx == const0_rtx && CONST_INT_P (true_rtx)
6202 && GET_MODE (XEXP (cond, 0)) == mode
6203 && (UINTVAL (true_rtx) & GET_MODE_MASK (mode))
6204 == nonzero_bits (XEXP (cond, 0), mode)
6205 && (i = exact_log2 (UINTVAL (true_rtx) & GET_MODE_MASK (mode))) >= 0)
6206 return XEXP (cond, 0);
6208 return x;
6211 /* Simplify X, a SET expression. Return the new expression. */
6213 static rtx
6214 simplify_set (rtx x)
6216 rtx src = SET_SRC (x);
6217 rtx dest = SET_DEST (x);
6218 enum machine_mode mode
6219 = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest);
6220 rtx other_insn;
6221 rtx *cc_use;
6223 /* (set (pc) (return)) gets written as (return). */
6224 if (GET_CODE (dest) == PC && ANY_RETURN_P (src))
6225 return src;
6227 /* Now that we know for sure which bits of SRC we are using, see if we can
6228 simplify the expression for the object knowing that we only need the
6229 low-order bits. */
6231 if (GET_MODE_CLASS (mode) == MODE_INT && HWI_COMPUTABLE_MODE_P (mode))
6233 src = force_to_mode (src, mode, ~(unsigned HOST_WIDE_INT) 0, 0);
6234 SUBST (SET_SRC (x), src);
6237 /* If we are setting CC0 or if the source is a COMPARE, look for the use of
6238 the comparison result and try to simplify it unless we already have used
6239 undobuf.other_insn. */
6240 if ((GET_MODE_CLASS (mode) == MODE_CC
6241 || GET_CODE (src) == COMPARE
6242 || CC0_P (dest))
6243 && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0
6244 && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn)
6245 && COMPARISON_P (*cc_use)
6246 && rtx_equal_p (XEXP (*cc_use, 0), dest))
6248 enum rtx_code old_code = GET_CODE (*cc_use);
6249 enum rtx_code new_code;
6250 rtx op0, op1, tmp;
6251 int other_changed = 0;
6252 rtx inner_compare = NULL_RTX;
6253 enum machine_mode compare_mode = GET_MODE (dest);
6255 if (GET_CODE (src) == COMPARE)
6257 op0 = XEXP (src, 0), op1 = XEXP (src, 1);
6258 if (GET_CODE (op0) == COMPARE && op1 == const0_rtx)
6260 inner_compare = op0;
6261 op0 = XEXP (inner_compare, 0), op1 = XEXP (inner_compare, 1);
6264 else
6265 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
6267 tmp = simplify_relational_operation (old_code, compare_mode, VOIDmode,
6268 op0, op1);
6269 if (!tmp)
6270 new_code = old_code;
6271 else if (!CONSTANT_P (tmp))
6273 new_code = GET_CODE (tmp);
6274 op0 = XEXP (tmp, 0);
6275 op1 = XEXP (tmp, 1);
6277 else
6279 rtx pat = PATTERN (other_insn);
6280 undobuf.other_insn = other_insn;
6281 SUBST (*cc_use, tmp);
6283 /* Attempt to simplify CC user. */
6284 if (GET_CODE (pat) == SET)
6286 rtx new_rtx = simplify_rtx (SET_SRC (pat));
6287 if (new_rtx != NULL_RTX)
6288 SUBST (SET_SRC (pat), new_rtx);
6291 /* Convert X into a no-op move. */
6292 SUBST (SET_DEST (x), pc_rtx);
6293 SUBST (SET_SRC (x), pc_rtx);
6294 return x;
6297 /* Simplify our comparison, if possible. */
6298 new_code = simplify_comparison (new_code, &op0, &op1);
6300 #ifdef SELECT_CC_MODE
6301 /* If this machine has CC modes other than CCmode, check to see if we
6302 need to use a different CC mode here. */
6303 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6304 compare_mode = GET_MODE (op0);
6305 else if (inner_compare
6306 && GET_MODE_CLASS (GET_MODE (inner_compare)) == MODE_CC
6307 && new_code == old_code
6308 && op0 == XEXP (inner_compare, 0)
6309 && op1 == XEXP (inner_compare, 1))
6310 compare_mode = GET_MODE (inner_compare);
6311 else
6312 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
6314 #ifndef HAVE_cc0
6315 /* If the mode changed, we have to change SET_DEST, the mode in the
6316 compare, and the mode in the place SET_DEST is used. If SET_DEST is
6317 a hard register, just build new versions with the proper mode. If it
6318 is a pseudo, we lose unless it is only time we set the pseudo, in
6319 which case we can safely change its mode. */
6320 if (compare_mode != GET_MODE (dest))
6322 if (can_change_dest_mode (dest, 0, compare_mode))
6324 unsigned int regno = REGNO (dest);
6325 rtx new_dest;
6327 if (regno < FIRST_PSEUDO_REGISTER)
6328 new_dest = gen_rtx_REG (compare_mode, regno);
6329 else
6331 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
6332 new_dest = regno_reg_rtx[regno];
6335 SUBST (SET_DEST (x), new_dest);
6336 SUBST (XEXP (*cc_use, 0), new_dest);
6337 other_changed = 1;
6339 dest = new_dest;
6342 #endif /* cc0 */
6343 #endif /* SELECT_CC_MODE */
6345 /* If the code changed, we have to build a new comparison in
6346 undobuf.other_insn. */
6347 if (new_code != old_code)
6349 int other_changed_previously = other_changed;
6350 unsigned HOST_WIDE_INT mask;
6351 rtx old_cc_use = *cc_use;
6353 SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use),
6354 dest, const0_rtx));
6355 other_changed = 1;
6357 /* If the only change we made was to change an EQ into an NE or
6358 vice versa, OP0 has only one bit that might be nonzero, and OP1
6359 is zero, check if changing the user of the condition code will
6360 produce a valid insn. If it won't, we can keep the original code
6361 in that insn by surrounding our operation with an XOR. */
6363 if (((old_code == NE && new_code == EQ)
6364 || (old_code == EQ && new_code == NE))
6365 && ! other_changed_previously && op1 == const0_rtx
6366 && HWI_COMPUTABLE_MODE_P (GET_MODE (op0))
6367 && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0)
6369 rtx pat = PATTERN (other_insn), note = 0;
6371 if ((recog_for_combine (&pat, other_insn, &note) < 0
6372 && ! check_asm_operands (pat)))
6374 *cc_use = old_cc_use;
6375 other_changed = 0;
6377 op0 = simplify_gen_binary (XOR, GET_MODE (op0), op0,
6378 gen_int_mode (mask,
6379 GET_MODE (op0)));
6384 if (other_changed)
6385 undobuf.other_insn = other_insn;
6387 /* Otherwise, if we didn't previously have a COMPARE in the
6388 correct mode, we need one. */
6389 if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode)
6391 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
6392 src = SET_SRC (x);
6394 else if (GET_MODE (op0) == compare_mode && op1 == const0_rtx)
6396 SUBST (SET_SRC (x), op0);
6397 src = SET_SRC (x);
6399 /* Otherwise, update the COMPARE if needed. */
6400 else if (XEXP (src, 0) != op0 || XEXP (src, 1) != op1)
6402 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
6403 src = SET_SRC (x);
6406 else
6408 /* Get SET_SRC in a form where we have placed back any
6409 compound expressions. Then do the checks below. */
6410 src = make_compound_operation (src, SET);
6411 SUBST (SET_SRC (x), src);
6414 /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation,
6415 and X being a REG or (subreg (reg)), we may be able to convert this to
6416 (set (subreg:m2 x) (op)).
6418 We can always do this if M1 is narrower than M2 because that means that
6419 we only care about the low bits of the result.
6421 However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot
6422 perform a narrower operation than requested since the high-order bits will
6423 be undefined. On machine where it is defined, this transformation is safe
6424 as long as M1 and M2 have the same number of words. */
6426 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
6427 && !OBJECT_P (SUBREG_REG (src))
6428 && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1))
6429 / UNITS_PER_WORD)
6430 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
6431 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))
6432 #ifndef WORD_REGISTER_OPERATIONS
6433 && (GET_MODE_SIZE (GET_MODE (src))
6434 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
6435 #endif
6436 #ifdef CANNOT_CHANGE_MODE_CLASS
6437 && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER
6438 && REG_CANNOT_CHANGE_MODE_P (REGNO (dest),
6439 GET_MODE (SUBREG_REG (src)),
6440 GET_MODE (src)))
6441 #endif
6442 && (REG_P (dest)
6443 || (GET_CODE (dest) == SUBREG
6444 && REG_P (SUBREG_REG (dest)))))
6446 SUBST (SET_DEST (x),
6447 gen_lowpart (GET_MODE (SUBREG_REG (src)),
6448 dest));
6449 SUBST (SET_SRC (x), SUBREG_REG (src));
6451 src = SET_SRC (x), dest = SET_DEST (x);
6454 #ifdef HAVE_cc0
6455 /* If we have (set (cc0) (subreg ...)), we try to remove the subreg
6456 in SRC. */
6457 if (dest == cc0_rtx
6458 && GET_CODE (src) == SUBREG
6459 && subreg_lowpart_p (src)
6460 && (GET_MODE_PRECISION (GET_MODE (src))
6461 < GET_MODE_PRECISION (GET_MODE (SUBREG_REG (src)))))
6463 rtx inner = SUBREG_REG (src);
6464 enum machine_mode inner_mode = GET_MODE (inner);
6466 /* Here we make sure that we don't have a sign bit on. */
6467 if (val_signbit_known_clear_p (GET_MODE (src),
6468 nonzero_bits (inner, inner_mode)))
6470 SUBST (SET_SRC (x), inner);
6471 src = SET_SRC (x);
6474 #endif
6476 #ifdef LOAD_EXTEND_OP
6477 /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this
6478 would require a paradoxical subreg. Replace the subreg with a
6479 zero_extend to avoid the reload that would otherwise be required. */
6481 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
6482 && INTEGRAL_MODE_P (GET_MODE (SUBREG_REG (src)))
6483 && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != UNKNOWN
6484 && SUBREG_BYTE (src) == 0
6485 && paradoxical_subreg_p (src)
6486 && MEM_P (SUBREG_REG (src)))
6488 SUBST (SET_SRC (x),
6489 gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))),
6490 GET_MODE (src), SUBREG_REG (src)));
6492 src = SET_SRC (x);
6494 #endif
6496 /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we
6497 are comparing an item known to be 0 or -1 against 0, use a logical
6498 operation instead. Check for one of the arms being an IOR of the other
6499 arm with some value. We compute three terms to be IOR'ed together. In
6500 practice, at most two will be nonzero. Then we do the IOR's. */
6502 if (GET_CODE (dest) != PC
6503 && GET_CODE (src) == IF_THEN_ELSE
6504 && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT
6505 && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE)
6506 && XEXP (XEXP (src, 0), 1) == const0_rtx
6507 && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0))
6508 #ifdef HAVE_conditional_move
6509 && ! can_conditionally_move_p (GET_MODE (src))
6510 #endif
6511 && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0),
6512 GET_MODE (XEXP (XEXP (src, 0), 0)))
6513 == GET_MODE_PRECISION (GET_MODE (XEXP (XEXP (src, 0), 0))))
6514 && ! side_effects_p (src))
6516 rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE
6517 ? XEXP (src, 1) : XEXP (src, 2));
6518 rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE
6519 ? XEXP (src, 2) : XEXP (src, 1));
6520 rtx term1 = const0_rtx, term2, term3;
6522 if (GET_CODE (true_rtx) == IOR
6523 && rtx_equal_p (XEXP (true_rtx, 0), false_rtx))
6524 term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx;
6525 else if (GET_CODE (true_rtx) == IOR
6526 && rtx_equal_p (XEXP (true_rtx, 1), false_rtx))
6527 term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx;
6528 else if (GET_CODE (false_rtx) == IOR
6529 && rtx_equal_p (XEXP (false_rtx, 0), true_rtx))
6530 term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx;
6531 else if (GET_CODE (false_rtx) == IOR
6532 && rtx_equal_p (XEXP (false_rtx, 1), true_rtx))
6533 term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx;
6535 term2 = simplify_gen_binary (AND, GET_MODE (src),
6536 XEXP (XEXP (src, 0), 0), true_rtx);
6537 term3 = simplify_gen_binary (AND, GET_MODE (src),
6538 simplify_gen_unary (NOT, GET_MODE (src),
6539 XEXP (XEXP (src, 0), 0),
6540 GET_MODE (src)),
6541 false_rtx);
6543 SUBST (SET_SRC (x),
6544 simplify_gen_binary (IOR, GET_MODE (src),
6545 simplify_gen_binary (IOR, GET_MODE (src),
6546 term1, term2),
6547 term3));
6549 src = SET_SRC (x);
6552 /* If either SRC or DEST is a CLOBBER of (const_int 0), make this
6553 whole thing fail. */
6554 if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx)
6555 return src;
6556 else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx)
6557 return dest;
6558 else
6559 /* Convert this into a field assignment operation, if possible. */
6560 return make_field_assignment (x);
6563 /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified
6564 result. */
6566 static rtx
6567 simplify_logical (rtx x)
6569 enum machine_mode mode = GET_MODE (x);
6570 rtx op0 = XEXP (x, 0);
6571 rtx op1 = XEXP (x, 1);
6573 switch (GET_CODE (x))
6575 case AND:
6576 /* We can call simplify_and_const_int only if we don't lose
6577 any (sign) bits when converting INTVAL (op1) to
6578 "unsigned HOST_WIDE_INT". */
6579 if (CONST_INT_P (op1)
6580 && (HWI_COMPUTABLE_MODE_P (mode)
6581 || INTVAL (op1) > 0))
6583 x = simplify_and_const_int (x, mode, op0, INTVAL (op1));
6584 if (GET_CODE (x) != AND)
6585 return x;
6587 op0 = XEXP (x, 0);
6588 op1 = XEXP (x, 1);
6591 /* If we have any of (and (ior A B) C) or (and (xor A B) C),
6592 apply the distributive law and then the inverse distributive
6593 law to see if things simplify. */
6594 if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR)
6596 rtx result = distribute_and_simplify_rtx (x, 0);
6597 if (result)
6598 return result;
6600 if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR)
6602 rtx result = distribute_and_simplify_rtx (x, 1);
6603 if (result)
6604 return result;
6606 break;
6608 case IOR:
6609 /* If we have (ior (and A B) C), apply the distributive law and then
6610 the inverse distributive law to see if things simplify. */
6612 if (GET_CODE (op0) == AND)
6614 rtx result = distribute_and_simplify_rtx (x, 0);
6615 if (result)
6616 return result;
6619 if (GET_CODE (op1) == AND)
6621 rtx result = distribute_and_simplify_rtx (x, 1);
6622 if (result)
6623 return result;
6625 break;
6627 default:
6628 gcc_unreachable ();
6631 return x;
6634 /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound
6635 operations" because they can be replaced with two more basic operations.
6636 ZERO_EXTEND is also considered "compound" because it can be replaced with
6637 an AND operation, which is simpler, though only one operation.
6639 The function expand_compound_operation is called with an rtx expression
6640 and will convert it to the appropriate shifts and AND operations,
6641 simplifying at each stage.
6643 The function make_compound_operation is called to convert an expression
6644 consisting of shifts and ANDs into the equivalent compound expression.
6645 It is the inverse of this function, loosely speaking. */
6647 static rtx
6648 expand_compound_operation (rtx x)
6650 unsigned HOST_WIDE_INT pos = 0, len;
6651 int unsignedp = 0;
6652 unsigned int modewidth;
6653 rtx tem;
6655 switch (GET_CODE (x))
6657 case ZERO_EXTEND:
6658 unsignedp = 1;
6659 case SIGN_EXTEND:
6660 /* We can't necessarily use a const_int for a multiword mode;
6661 it depends on implicitly extending the value.
6662 Since we don't know the right way to extend it,
6663 we can't tell whether the implicit way is right.
6665 Even for a mode that is no wider than a const_int,
6666 we can't win, because we need to sign extend one of its bits through
6667 the rest of it, and we don't know which bit. */
6668 if (CONST_INT_P (XEXP (x, 0)))
6669 return x;
6671 /* Return if (subreg:MODE FROM 0) is not a safe replacement for
6672 (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM
6673 because (SUBREG (MEM...)) is guaranteed to cause the MEM to be
6674 reloaded. If not for that, MEM's would very rarely be safe.
6676 Reject MODEs bigger than a word, because we might not be able
6677 to reference a two-register group starting with an arbitrary register
6678 (and currently gen_lowpart might crash for a SUBREG). */
6680 if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD)
6681 return x;
6683 /* Reject MODEs that aren't scalar integers because turning vector
6684 or complex modes into shifts causes problems. */
6686 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
6687 return x;
6689 len = GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)));
6690 /* If the inner object has VOIDmode (the only way this can happen
6691 is if it is an ASM_OPERANDS), we can't do anything since we don't
6692 know how much masking to do. */
6693 if (len == 0)
6694 return x;
6696 break;
6698 case ZERO_EXTRACT:
6699 unsignedp = 1;
6701 /* ... fall through ... */
6703 case SIGN_EXTRACT:
6704 /* If the operand is a CLOBBER, just return it. */
6705 if (GET_CODE (XEXP (x, 0)) == CLOBBER)
6706 return XEXP (x, 0);
6708 if (!CONST_INT_P (XEXP (x, 1))
6709 || !CONST_INT_P (XEXP (x, 2))
6710 || GET_MODE (XEXP (x, 0)) == VOIDmode)
6711 return x;
6713 /* Reject MODEs that aren't scalar integers because turning vector
6714 or complex modes into shifts causes problems. */
6716 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
6717 return x;
6719 len = INTVAL (XEXP (x, 1));
6720 pos = INTVAL (XEXP (x, 2));
6722 /* This should stay within the object being extracted, fail otherwise. */
6723 if (len + pos > GET_MODE_PRECISION (GET_MODE (XEXP (x, 0))))
6724 return x;
6726 if (BITS_BIG_ENDIAN)
6727 pos = GET_MODE_PRECISION (GET_MODE (XEXP (x, 0))) - len - pos;
6729 break;
6731 default:
6732 return x;
6734 /* Convert sign extension to zero extension, if we know that the high
6735 bit is not set, as this is easier to optimize. It will be converted
6736 back to cheaper alternative in make_extraction. */
6737 if (GET_CODE (x) == SIGN_EXTEND
6738 && (HWI_COMPUTABLE_MODE_P (GET_MODE (x))
6739 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
6740 & ~(((unsigned HOST_WIDE_INT)
6741 GET_MODE_MASK (GET_MODE (XEXP (x, 0))))
6742 >> 1))
6743 == 0)))
6745 rtx temp = gen_rtx_ZERO_EXTEND (GET_MODE (x), XEXP (x, 0));
6746 rtx temp2 = expand_compound_operation (temp);
6748 /* Make sure this is a profitable operation. */
6749 if (set_src_cost (x, optimize_this_for_speed_p)
6750 > set_src_cost (temp2, optimize_this_for_speed_p))
6751 return temp2;
6752 else if (set_src_cost (x, optimize_this_for_speed_p)
6753 > set_src_cost (temp, optimize_this_for_speed_p))
6754 return temp;
6755 else
6756 return x;
6759 /* We can optimize some special cases of ZERO_EXTEND. */
6760 if (GET_CODE (x) == ZERO_EXTEND)
6762 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we
6763 know that the last value didn't have any inappropriate bits
6764 set. */
6765 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
6766 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
6767 && HWI_COMPUTABLE_MODE_P (GET_MODE (x))
6768 && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x))
6769 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
6770 return XEXP (XEXP (x, 0), 0);
6772 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
6773 if (GET_CODE (XEXP (x, 0)) == SUBREG
6774 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
6775 && subreg_lowpart_p (XEXP (x, 0))
6776 && HWI_COMPUTABLE_MODE_P (GET_MODE (x))
6777 && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x))
6778 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
6779 return SUBREG_REG (XEXP (x, 0));
6781 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo
6782 is a comparison and STORE_FLAG_VALUE permits. This is like
6783 the first case, but it works even when GET_MODE (x) is larger
6784 than HOST_WIDE_INT. */
6785 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
6786 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
6787 && COMPARISON_P (XEXP (XEXP (x, 0), 0))
6788 && (GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
6789 <= HOST_BITS_PER_WIDE_INT)
6790 && (STORE_FLAG_VALUE & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
6791 return XEXP (XEXP (x, 0), 0);
6793 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
6794 if (GET_CODE (XEXP (x, 0)) == SUBREG
6795 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
6796 && subreg_lowpart_p (XEXP (x, 0))
6797 && COMPARISON_P (SUBREG_REG (XEXP (x, 0)))
6798 && (GET_MODE_PRECISION (GET_MODE (XEXP (x, 0)))
6799 <= HOST_BITS_PER_WIDE_INT)
6800 && (STORE_FLAG_VALUE & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
6801 return SUBREG_REG (XEXP (x, 0));
6805 /* If we reach here, we want to return a pair of shifts. The inner
6806 shift is a left shift of BITSIZE - POS - LEN bits. The outer
6807 shift is a right shift of BITSIZE - LEN bits. It is arithmetic or
6808 logical depending on the value of UNSIGNEDP.
6810 If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be
6811 converted into an AND of a shift.
6813 We must check for the case where the left shift would have a negative
6814 count. This can happen in a case like (x >> 31) & 255 on machines
6815 that can't shift by a constant. On those machines, we would first
6816 combine the shift with the AND to produce a variable-position
6817 extraction. Then the constant of 31 would be substituted in
6818 to produce such a position. */
6820 modewidth = GET_MODE_PRECISION (GET_MODE (x));
6821 if (modewidth >= pos + len)
6823 enum machine_mode mode = GET_MODE (x);
6824 tem = gen_lowpart (mode, XEXP (x, 0));
6825 if (!tem || GET_CODE (tem) == CLOBBER)
6826 return x;
6827 tem = simplify_shift_const (NULL_RTX, ASHIFT, mode,
6828 tem, modewidth - pos - len);
6829 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT,
6830 mode, tem, modewidth - len);
6832 else if (unsignedp && len < HOST_BITS_PER_WIDE_INT)
6833 tem = simplify_and_const_int (NULL_RTX, GET_MODE (x),
6834 simplify_shift_const (NULL_RTX, LSHIFTRT,
6835 GET_MODE (x),
6836 XEXP (x, 0), pos),
6837 ((unsigned HOST_WIDE_INT) 1 << len) - 1);
6838 else
6839 /* Any other cases we can't handle. */
6840 return x;
6842 /* If we couldn't do this for some reason, return the original
6843 expression. */
6844 if (GET_CODE (tem) == CLOBBER)
6845 return x;
6847 return tem;
6850 /* X is a SET which contains an assignment of one object into
6851 a part of another (such as a bit-field assignment, STRICT_LOW_PART,
6852 or certain SUBREGS). If possible, convert it into a series of
6853 logical operations.
6855 We half-heartedly support variable positions, but do not at all
6856 support variable lengths. */
6858 static const_rtx
6859 expand_field_assignment (const_rtx x)
6861 rtx inner;
6862 rtx pos; /* Always counts from low bit. */
6863 int len;
6864 rtx mask, cleared, masked;
6865 enum machine_mode compute_mode;
6867 /* Loop until we find something we can't simplify. */
6868 while (1)
6870 if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
6871 && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG)
6873 inner = SUBREG_REG (XEXP (SET_DEST (x), 0));
6874 len = GET_MODE_PRECISION (GET_MODE (XEXP (SET_DEST (x), 0)));
6875 pos = GEN_INT (subreg_lsb (XEXP (SET_DEST (x), 0)));
6877 else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
6878 && CONST_INT_P (XEXP (SET_DEST (x), 1)))
6880 inner = XEXP (SET_DEST (x), 0);
6881 len = INTVAL (XEXP (SET_DEST (x), 1));
6882 pos = XEXP (SET_DEST (x), 2);
6884 /* A constant position should stay within the width of INNER. */
6885 if (CONST_INT_P (pos)
6886 && INTVAL (pos) + len > GET_MODE_PRECISION (GET_MODE (inner)))
6887 break;
6889 if (BITS_BIG_ENDIAN)
6891 if (CONST_INT_P (pos))
6892 pos = GEN_INT (GET_MODE_PRECISION (GET_MODE (inner)) - len
6893 - INTVAL (pos));
6894 else if (GET_CODE (pos) == MINUS
6895 && CONST_INT_P (XEXP (pos, 1))
6896 && (INTVAL (XEXP (pos, 1))
6897 == GET_MODE_PRECISION (GET_MODE (inner)) - len))
6898 /* If position is ADJUST - X, new position is X. */
6899 pos = XEXP (pos, 0);
6900 else
6902 HOST_WIDE_INT prec = GET_MODE_PRECISION (GET_MODE (inner));
6903 pos = simplify_gen_binary (MINUS, GET_MODE (pos),
6904 gen_int_mode (prec - len,
6905 GET_MODE (pos)),
6906 pos);
6911 /* A SUBREG between two modes that occupy the same numbers of words
6912 can be done by moving the SUBREG to the source. */
6913 else if (GET_CODE (SET_DEST (x)) == SUBREG
6914 /* We need SUBREGs to compute nonzero_bits properly. */
6915 && nonzero_sign_valid
6916 && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
6917 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
6918 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
6919 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))
6921 x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)),
6922 gen_lowpart
6923 (GET_MODE (SUBREG_REG (SET_DEST (x))),
6924 SET_SRC (x)));
6925 continue;
6927 else
6928 break;
6930 while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
6931 inner = SUBREG_REG (inner);
6933 compute_mode = GET_MODE (inner);
6935 /* Don't attempt bitwise arithmetic on non scalar integer modes. */
6936 if (! SCALAR_INT_MODE_P (compute_mode))
6938 enum machine_mode imode;
6940 /* Don't do anything for vector or complex integral types. */
6941 if (! FLOAT_MODE_P (compute_mode))
6942 break;
6944 /* Try to find an integral mode to pun with. */
6945 imode = mode_for_size (GET_MODE_BITSIZE (compute_mode), MODE_INT, 0);
6946 if (imode == BLKmode)
6947 break;
6949 compute_mode = imode;
6950 inner = gen_lowpart (imode, inner);
6953 /* Compute a mask of LEN bits, if we can do this on the host machine. */
6954 if (len >= HOST_BITS_PER_WIDE_INT)
6955 break;
6957 /* Now compute the equivalent expression. Make a copy of INNER
6958 for the SET_DEST in case it is a MEM into which we will substitute;
6959 we don't want shared RTL in that case. */
6960 mask = gen_int_mode (((unsigned HOST_WIDE_INT) 1 << len) - 1,
6961 compute_mode);
6962 cleared = simplify_gen_binary (AND, compute_mode,
6963 simplify_gen_unary (NOT, compute_mode,
6964 simplify_gen_binary (ASHIFT,
6965 compute_mode,
6966 mask, pos),
6967 compute_mode),
6968 inner);
6969 masked = simplify_gen_binary (ASHIFT, compute_mode,
6970 simplify_gen_binary (
6971 AND, compute_mode,
6972 gen_lowpart (compute_mode, SET_SRC (x)),
6973 mask),
6974 pos);
6976 x = gen_rtx_SET (VOIDmode, copy_rtx (inner),
6977 simplify_gen_binary (IOR, compute_mode,
6978 cleared, masked));
6981 return x;
6984 /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero,
6985 it is an RTX that represents the (variable) starting position; otherwise,
6986 POS is the (constant) starting bit position. Both are counted from the LSB.
6988 UNSIGNEDP is nonzero for an unsigned reference and zero for a signed one.
6990 IN_DEST is nonzero if this is a reference in the destination of a SET.
6991 This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero,
6992 a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will
6993 be used.
6995 IN_COMPARE is nonzero if we are in a COMPARE. This means that a
6996 ZERO_EXTRACT should be built even for bits starting at bit 0.
6998 MODE is the desired mode of the result (if IN_DEST == 0).
7000 The result is an RTX for the extraction or NULL_RTX if the target
7001 can't handle it. */
7003 static rtx
7004 make_extraction (enum machine_mode mode, rtx inner, HOST_WIDE_INT pos,
7005 rtx pos_rtx, unsigned HOST_WIDE_INT len, int unsignedp,
7006 int in_dest, int in_compare)
7008 /* This mode describes the size of the storage area
7009 to fetch the overall value from. Within that, we
7010 ignore the POS lowest bits, etc. */
7011 enum machine_mode is_mode = GET_MODE (inner);
7012 enum machine_mode inner_mode;
7013 enum machine_mode wanted_inner_mode;
7014 enum machine_mode wanted_inner_reg_mode = word_mode;
7015 enum machine_mode pos_mode = word_mode;
7016 enum machine_mode extraction_mode = word_mode;
7017 enum machine_mode tmode = mode_for_size (len, MODE_INT, 1);
7018 rtx new_rtx = 0;
7019 rtx orig_pos_rtx = pos_rtx;
7020 HOST_WIDE_INT orig_pos;
7022 if (pos_rtx && CONST_INT_P (pos_rtx))
7023 pos = INTVAL (pos_rtx), pos_rtx = 0;
7025 if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
7027 /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...),
7028 consider just the QI as the memory to extract from.
7029 The subreg adds or removes high bits; its mode is
7030 irrelevant to the meaning of this extraction,
7031 since POS and LEN count from the lsb. */
7032 if (MEM_P (SUBREG_REG (inner)))
7033 is_mode = GET_MODE (SUBREG_REG (inner));
7034 inner = SUBREG_REG (inner);
7036 else if (GET_CODE (inner) == ASHIFT
7037 && CONST_INT_P (XEXP (inner, 1))
7038 && pos_rtx == 0 && pos == 0
7039 && len > UINTVAL (XEXP (inner, 1)))
7041 /* We're extracting the least significant bits of an rtx
7042 (ashift X (const_int C)), where LEN > C. Extract the
7043 least significant (LEN - C) bits of X, giving an rtx
7044 whose mode is MODE, then shift it left C times. */
7045 new_rtx = make_extraction (mode, XEXP (inner, 0),
7046 0, 0, len - INTVAL (XEXP (inner, 1)),
7047 unsignedp, in_dest, in_compare);
7048 if (new_rtx != 0)
7049 return gen_rtx_ASHIFT (mode, new_rtx, XEXP (inner, 1));
7051 else if (GET_CODE (inner) == TRUNCATE)
7052 inner = XEXP (inner, 0);
7054 inner_mode = GET_MODE (inner);
7056 /* See if this can be done without an extraction. We never can if the
7057 width of the field is not the same as that of some integer mode. For
7058 registers, we can only avoid the extraction if the position is at the
7059 low-order bit and this is either not in the destination or we have the
7060 appropriate STRICT_LOW_PART operation available.
7062 For MEM, we can avoid an extract if the field starts on an appropriate
7063 boundary and we can change the mode of the memory reference. */
7065 if (tmode != BLKmode
7066 && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
7067 && !MEM_P (inner)
7068 && (inner_mode == tmode
7069 || !REG_P (inner)
7070 || TRULY_NOOP_TRUNCATION_MODES_P (tmode, inner_mode)
7071 || reg_truncated_to_mode (tmode, inner))
7072 && (! in_dest
7073 || (REG_P (inner)
7074 && have_insn_for (STRICT_LOW_PART, tmode))))
7075 || (MEM_P (inner) && pos_rtx == 0
7076 && (pos
7077 % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode)
7078 : BITS_PER_UNIT)) == 0
7079 /* We can't do this if we are widening INNER_MODE (it
7080 may not be aligned, for one thing). */
7081 && GET_MODE_PRECISION (inner_mode) >= GET_MODE_PRECISION (tmode)
7082 && (inner_mode == tmode
7083 || (! mode_dependent_address_p (XEXP (inner, 0),
7084 MEM_ADDR_SPACE (inner))
7085 && ! MEM_VOLATILE_P (inner))))))
7087 /* If INNER is a MEM, make a new MEM that encompasses just the desired
7088 field. If the original and current mode are the same, we need not
7089 adjust the offset. Otherwise, we do if bytes big endian.
7091 If INNER is not a MEM, get a piece consisting of just the field
7092 of interest (in this case POS % BITS_PER_WORD must be 0). */
7094 if (MEM_P (inner))
7096 HOST_WIDE_INT offset;
7098 /* POS counts from lsb, but make OFFSET count in memory order. */
7099 if (BYTES_BIG_ENDIAN)
7100 offset = (GET_MODE_PRECISION (is_mode) - len - pos) / BITS_PER_UNIT;
7101 else
7102 offset = pos / BITS_PER_UNIT;
7104 new_rtx = adjust_address_nv (inner, tmode, offset);
7106 else if (REG_P (inner))
7108 if (tmode != inner_mode)
7110 /* We can't call gen_lowpart in a DEST since we
7111 always want a SUBREG (see below) and it would sometimes
7112 return a new hard register. */
7113 if (pos || in_dest)
7115 HOST_WIDE_INT final_word = pos / BITS_PER_WORD;
7117 if (WORDS_BIG_ENDIAN
7118 && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD)
7119 final_word = ((GET_MODE_SIZE (inner_mode)
7120 - GET_MODE_SIZE (tmode))
7121 / UNITS_PER_WORD) - final_word;
7123 final_word *= UNITS_PER_WORD;
7124 if (BYTES_BIG_ENDIAN &&
7125 GET_MODE_SIZE (inner_mode) > GET_MODE_SIZE (tmode))
7126 final_word += (GET_MODE_SIZE (inner_mode)
7127 - GET_MODE_SIZE (tmode)) % UNITS_PER_WORD;
7129 /* Avoid creating invalid subregs, for example when
7130 simplifying (x>>32)&255. */
7131 if (!validate_subreg (tmode, inner_mode, inner, final_word))
7132 return NULL_RTX;
7134 new_rtx = gen_rtx_SUBREG (tmode, inner, final_word);
7136 else
7137 new_rtx = gen_lowpart (tmode, inner);
7139 else
7140 new_rtx = inner;
7142 else
7143 new_rtx = force_to_mode (inner, tmode,
7144 len >= HOST_BITS_PER_WIDE_INT
7145 ? ~(unsigned HOST_WIDE_INT) 0
7146 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
7149 /* If this extraction is going into the destination of a SET,
7150 make a STRICT_LOW_PART unless we made a MEM. */
7152 if (in_dest)
7153 return (MEM_P (new_rtx) ? new_rtx
7154 : (GET_CODE (new_rtx) != SUBREG
7155 ? gen_rtx_CLOBBER (tmode, const0_rtx)
7156 : gen_rtx_STRICT_LOW_PART (VOIDmode, new_rtx)));
7158 if (mode == tmode)
7159 return new_rtx;
7161 if (CONST_SCALAR_INT_P (new_rtx))
7162 return simplify_unary_operation (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
7163 mode, new_rtx, tmode);
7165 /* If we know that no extraneous bits are set, and that the high
7166 bit is not set, convert the extraction to the cheaper of
7167 sign and zero extension, that are equivalent in these cases. */
7168 if (flag_expensive_optimizations
7169 && (HWI_COMPUTABLE_MODE_P (tmode)
7170 && ((nonzero_bits (new_rtx, tmode)
7171 & ~(((unsigned HOST_WIDE_INT)GET_MODE_MASK (tmode)) >> 1))
7172 == 0)))
7174 rtx temp = gen_rtx_ZERO_EXTEND (mode, new_rtx);
7175 rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new_rtx);
7177 /* Prefer ZERO_EXTENSION, since it gives more information to
7178 backends. */
7179 if (set_src_cost (temp, optimize_this_for_speed_p)
7180 <= set_src_cost (temp1, optimize_this_for_speed_p))
7181 return temp;
7182 return temp1;
7185 /* Otherwise, sign- or zero-extend unless we already are in the
7186 proper mode. */
7188 return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
7189 mode, new_rtx));
7192 /* Unless this is a COMPARE or we have a funny memory reference,
7193 don't do anything with zero-extending field extracts starting at
7194 the low-order bit since they are simple AND operations. */
7195 if (pos_rtx == 0 && pos == 0 && ! in_dest
7196 && ! in_compare && unsignedp)
7197 return 0;
7199 /* Unless INNER is not MEM, reject this if we would be spanning bytes or
7200 if the position is not a constant and the length is not 1. In all
7201 other cases, we would only be going outside our object in cases when
7202 an original shift would have been undefined. */
7203 if (MEM_P (inner)
7204 && ((pos_rtx == 0 && pos + len > GET_MODE_PRECISION (is_mode))
7205 || (pos_rtx != 0 && len != 1)))
7206 return 0;
7208 enum extraction_pattern pattern = (in_dest ? EP_insv
7209 : unsignedp ? EP_extzv : EP_extv);
7211 /* If INNER is not from memory, we want it to have the mode of a register
7212 extraction pattern's structure operand, or word_mode if there is no
7213 such pattern. The same applies to extraction_mode and pos_mode
7214 and their respective operands.
7216 For memory, assume that the desired extraction_mode and pos_mode
7217 are the same as for a register operation, since at present we don't
7218 have named patterns for aligned memory structures. */
7219 struct extraction_insn insn;
7220 if (get_best_reg_extraction_insn (&insn, pattern,
7221 GET_MODE_BITSIZE (inner_mode), mode))
7223 wanted_inner_reg_mode = insn.struct_mode;
7224 pos_mode = insn.pos_mode;
7225 extraction_mode = insn.field_mode;
7228 /* Never narrow an object, since that might not be safe. */
7230 if (mode != VOIDmode
7231 && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode))
7232 extraction_mode = mode;
7234 if (!MEM_P (inner))
7235 wanted_inner_mode = wanted_inner_reg_mode;
7236 else
7238 /* Be careful not to go beyond the extracted object and maintain the
7239 natural alignment of the memory. */
7240 wanted_inner_mode = smallest_mode_for_size (len, MODE_INT);
7241 while (pos % GET_MODE_BITSIZE (wanted_inner_mode) + len
7242 > GET_MODE_BITSIZE (wanted_inner_mode))
7244 wanted_inner_mode = GET_MODE_WIDER_MODE (wanted_inner_mode);
7245 gcc_assert (wanted_inner_mode != VOIDmode);
7249 orig_pos = pos;
7251 if (BITS_BIG_ENDIAN)
7253 /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to
7254 BITS_BIG_ENDIAN style. If position is constant, compute new
7255 position. Otherwise, build subtraction.
7256 Note that POS is relative to the mode of the original argument.
7257 If it's a MEM we need to recompute POS relative to that.
7258 However, if we're extracting from (or inserting into) a register,
7259 we want to recompute POS relative to wanted_inner_mode. */
7260 int width = (MEM_P (inner)
7261 ? GET_MODE_BITSIZE (is_mode)
7262 : GET_MODE_BITSIZE (wanted_inner_mode));
7264 if (pos_rtx == 0)
7265 pos = width - len - pos;
7266 else
7267 pos_rtx
7268 = gen_rtx_MINUS (GET_MODE (pos_rtx),
7269 gen_int_mode (width - len, GET_MODE (pos_rtx)),
7270 pos_rtx);
7271 /* POS may be less than 0 now, but we check for that below.
7272 Note that it can only be less than 0 if !MEM_P (inner). */
7275 /* If INNER has a wider mode, and this is a constant extraction, try to
7276 make it smaller and adjust the byte to point to the byte containing
7277 the value. */
7278 if (wanted_inner_mode != VOIDmode
7279 && inner_mode != wanted_inner_mode
7280 && ! pos_rtx
7281 && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode)
7282 && MEM_P (inner)
7283 && ! mode_dependent_address_p (XEXP (inner, 0), MEM_ADDR_SPACE (inner))
7284 && ! MEM_VOLATILE_P (inner))
7286 int offset = 0;
7288 /* The computations below will be correct if the machine is big
7289 endian in both bits and bytes or little endian in bits and bytes.
7290 If it is mixed, we must adjust. */
7292 /* If bytes are big endian and we had a paradoxical SUBREG, we must
7293 adjust OFFSET to compensate. */
7294 if (BYTES_BIG_ENDIAN
7295 && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode))
7296 offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode);
7298 /* We can now move to the desired byte. */
7299 offset += (pos / GET_MODE_BITSIZE (wanted_inner_mode))
7300 * GET_MODE_SIZE (wanted_inner_mode);
7301 pos %= GET_MODE_BITSIZE (wanted_inner_mode);
7303 if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN
7304 && is_mode != wanted_inner_mode)
7305 offset = (GET_MODE_SIZE (is_mode)
7306 - GET_MODE_SIZE (wanted_inner_mode) - offset);
7308 inner = adjust_address_nv (inner, wanted_inner_mode, offset);
7311 /* If INNER is not memory, get it into the proper mode. If we are changing
7312 its mode, POS must be a constant and smaller than the size of the new
7313 mode. */
7314 else if (!MEM_P (inner))
7316 /* On the LHS, don't create paradoxical subregs implicitely truncating
7317 the register unless TRULY_NOOP_TRUNCATION. */
7318 if (in_dest
7319 && !TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (inner),
7320 wanted_inner_mode))
7321 return NULL_RTX;
7323 if (GET_MODE (inner) != wanted_inner_mode
7324 && (pos_rtx != 0
7325 || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode)))
7326 return NULL_RTX;
7328 if (orig_pos < 0)
7329 return NULL_RTX;
7331 inner = force_to_mode (inner, wanted_inner_mode,
7332 pos_rtx
7333 || len + orig_pos >= HOST_BITS_PER_WIDE_INT
7334 ? ~(unsigned HOST_WIDE_INT) 0
7335 : ((((unsigned HOST_WIDE_INT) 1 << len) - 1)
7336 << orig_pos),
7340 /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we
7341 have to zero extend. Otherwise, we can just use a SUBREG. */
7342 if (pos_rtx != 0
7343 && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx)))
7345 rtx temp = simplify_gen_unary (ZERO_EXTEND, pos_mode, pos_rtx,
7346 GET_MODE (pos_rtx));
7348 /* If we know that no extraneous bits are set, and that the high
7349 bit is not set, convert extraction to cheaper one - either
7350 SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these
7351 cases. */
7352 if (flag_expensive_optimizations
7353 && (HWI_COMPUTABLE_MODE_P (GET_MODE (pos_rtx))
7354 && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx))
7355 & ~(((unsigned HOST_WIDE_INT)
7356 GET_MODE_MASK (GET_MODE (pos_rtx)))
7357 >> 1))
7358 == 0)))
7360 rtx temp1 = simplify_gen_unary (SIGN_EXTEND, pos_mode, pos_rtx,
7361 GET_MODE (pos_rtx));
7363 /* Prefer ZERO_EXTENSION, since it gives more information to
7364 backends. */
7365 if (set_src_cost (temp1, optimize_this_for_speed_p)
7366 < set_src_cost (temp, optimize_this_for_speed_p))
7367 temp = temp1;
7369 pos_rtx = temp;
7372 /* Make POS_RTX unless we already have it and it is correct. If we don't
7373 have a POS_RTX but we do have an ORIG_POS_RTX, the latter must
7374 be a CONST_INT. */
7375 if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos)
7376 pos_rtx = orig_pos_rtx;
7378 else if (pos_rtx == 0)
7379 pos_rtx = GEN_INT (pos);
7381 /* Make the required operation. See if we can use existing rtx. */
7382 new_rtx = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT,
7383 extraction_mode, inner, GEN_INT (len), pos_rtx);
7384 if (! in_dest)
7385 new_rtx = gen_lowpart (mode, new_rtx);
7387 return new_rtx;
7390 /* See if X contains an ASHIFT of COUNT or more bits that can be commuted
7391 with any other operations in X. Return X without that shift if so. */
7393 static rtx
7394 extract_left_shift (rtx x, int count)
7396 enum rtx_code code = GET_CODE (x);
7397 enum machine_mode mode = GET_MODE (x);
7398 rtx tem;
7400 switch (code)
7402 case ASHIFT:
7403 /* This is the shift itself. If it is wide enough, we will return
7404 either the value being shifted if the shift count is equal to
7405 COUNT or a shift for the difference. */
7406 if (CONST_INT_P (XEXP (x, 1))
7407 && INTVAL (XEXP (x, 1)) >= count)
7408 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0),
7409 INTVAL (XEXP (x, 1)) - count);
7410 break;
7412 case NEG: case NOT:
7413 if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0)
7414 return simplify_gen_unary (code, mode, tem, mode);
7416 break;
7418 case PLUS: case IOR: case XOR: case AND:
7419 /* If we can safely shift this constant and we find the inner shift,
7420 make a new operation. */
7421 if (CONST_INT_P (XEXP (x, 1))
7422 && (UINTVAL (XEXP (x, 1))
7423 & ((((unsigned HOST_WIDE_INT) 1 << count)) - 1)) == 0
7424 && (tem = extract_left_shift (XEXP (x, 0), count)) != 0)
7426 HOST_WIDE_INT val = INTVAL (XEXP (x, 1)) >> count;
7427 return simplify_gen_binary (code, mode, tem,
7428 gen_int_mode (val, mode));
7430 break;
7432 default:
7433 break;
7436 return 0;
7439 /* Look at the expression rooted at X. Look for expressions
7440 equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND.
7441 Form these expressions.
7443 Return the new rtx, usually just X.
7445 Also, for machines like the VAX that don't have logical shift insns,
7446 try to convert logical to arithmetic shift operations in cases where
7447 they are equivalent. This undoes the canonicalizations to logical
7448 shifts done elsewhere.
7450 We try, as much as possible, to re-use rtl expressions to save memory.
7452 IN_CODE says what kind of expression we are processing. Normally, it is
7453 SET. In a memory address (inside a MEM, PLUS or minus, the latter two
7454 being kludges), it is MEM. When processing the arguments of a comparison
7455 or a COMPARE against zero, it is COMPARE. */
7458 make_compound_operation (rtx x, enum rtx_code in_code)
7460 enum rtx_code code = GET_CODE (x);
7461 enum machine_mode mode = GET_MODE (x);
7462 int mode_width = GET_MODE_PRECISION (mode);
7463 rtx rhs, lhs;
7464 enum rtx_code next_code;
7465 int i, j;
7466 rtx new_rtx = 0;
7467 rtx tem;
7468 const char *fmt;
7470 /* Select the code to be used in recursive calls. Once we are inside an
7471 address, we stay there. If we have a comparison, set to COMPARE,
7472 but once inside, go back to our default of SET. */
7474 next_code = (code == MEM ? MEM
7475 : ((code == PLUS || code == MINUS)
7476 && SCALAR_INT_MODE_P (mode)) ? MEM
7477 : ((code == COMPARE || COMPARISON_P (x))
7478 && XEXP (x, 1) == const0_rtx) ? COMPARE
7479 : in_code == COMPARE ? SET : in_code);
7481 /* Process depending on the code of this operation. If NEW is set
7482 nonzero, it will be returned. */
7484 switch (code)
7486 case ASHIFT:
7487 /* Convert shifts by constants into multiplications if inside
7488 an address. */
7489 if (in_code == MEM && CONST_INT_P (XEXP (x, 1))
7490 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
7491 && INTVAL (XEXP (x, 1)) >= 0
7492 && SCALAR_INT_MODE_P (mode))
7494 HOST_WIDE_INT count = INTVAL (XEXP (x, 1));
7495 HOST_WIDE_INT multval = (HOST_WIDE_INT) 1 << count;
7497 new_rtx = make_compound_operation (XEXP (x, 0), next_code);
7498 if (GET_CODE (new_rtx) == NEG)
7500 new_rtx = XEXP (new_rtx, 0);
7501 multval = -multval;
7503 multval = trunc_int_for_mode (multval, mode);
7504 new_rtx = gen_rtx_MULT (mode, new_rtx, gen_int_mode (multval, mode));
7506 break;
7508 case PLUS:
7509 lhs = XEXP (x, 0);
7510 rhs = XEXP (x, 1);
7511 lhs = make_compound_operation (lhs, next_code);
7512 rhs = make_compound_operation (rhs, next_code);
7513 if (GET_CODE (lhs) == MULT && GET_CODE (XEXP (lhs, 0)) == NEG
7514 && SCALAR_INT_MODE_P (mode))
7516 tem = simplify_gen_binary (MULT, mode, XEXP (XEXP (lhs, 0), 0),
7517 XEXP (lhs, 1));
7518 new_rtx = simplify_gen_binary (MINUS, mode, rhs, tem);
7520 else if (GET_CODE (lhs) == MULT
7521 && (CONST_INT_P (XEXP (lhs, 1)) && INTVAL (XEXP (lhs, 1)) < 0))
7523 tem = simplify_gen_binary (MULT, mode, XEXP (lhs, 0),
7524 simplify_gen_unary (NEG, mode,
7525 XEXP (lhs, 1),
7526 mode));
7527 new_rtx = simplify_gen_binary (MINUS, mode, rhs, tem);
7529 else
7531 SUBST (XEXP (x, 0), lhs);
7532 SUBST (XEXP (x, 1), rhs);
7533 goto maybe_swap;
7535 x = gen_lowpart (mode, new_rtx);
7536 goto maybe_swap;
7538 case MINUS:
7539 lhs = XEXP (x, 0);
7540 rhs = XEXP (x, 1);
7541 lhs = make_compound_operation (lhs, next_code);
7542 rhs = make_compound_operation (rhs, next_code);
7543 if (GET_CODE (rhs) == MULT && GET_CODE (XEXP (rhs, 0)) == NEG
7544 && SCALAR_INT_MODE_P (mode))
7546 tem = simplify_gen_binary (MULT, mode, XEXP (XEXP (rhs, 0), 0),
7547 XEXP (rhs, 1));
7548 new_rtx = simplify_gen_binary (PLUS, mode, tem, lhs);
7550 else if (GET_CODE (rhs) == MULT
7551 && (CONST_INT_P (XEXP (rhs, 1)) && INTVAL (XEXP (rhs, 1)) < 0))
7553 tem = simplify_gen_binary (MULT, mode, XEXP (rhs, 0),
7554 simplify_gen_unary (NEG, mode,
7555 XEXP (rhs, 1),
7556 mode));
7557 new_rtx = simplify_gen_binary (PLUS, mode, tem, lhs);
7559 else
7561 SUBST (XEXP (x, 0), lhs);
7562 SUBST (XEXP (x, 1), rhs);
7563 return x;
7565 return gen_lowpart (mode, new_rtx);
7567 case AND:
7568 /* If the second operand is not a constant, we can't do anything
7569 with it. */
7570 if (!CONST_INT_P (XEXP (x, 1)))
7571 break;
7573 /* If the constant is a power of two minus one and the first operand
7574 is a logical right shift, make an extraction. */
7575 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7576 && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0)
7578 new_rtx = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
7579 new_rtx = make_extraction (mode, new_rtx, 0, XEXP (XEXP (x, 0), 1), i, 1,
7580 0, in_code == COMPARE);
7583 /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */
7584 else if (GET_CODE (XEXP (x, 0)) == SUBREG
7585 && subreg_lowpart_p (XEXP (x, 0))
7586 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT
7587 && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0)
7589 new_rtx = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0),
7590 next_code);
7591 new_rtx = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new_rtx, 0,
7592 XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1,
7593 0, in_code == COMPARE);
7595 /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */
7596 else if ((GET_CODE (XEXP (x, 0)) == XOR
7597 || GET_CODE (XEXP (x, 0)) == IOR)
7598 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT
7599 && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT
7600 && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0)
7602 /* Apply the distributive law, and then try to make extractions. */
7603 new_rtx = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode,
7604 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0),
7605 XEXP (x, 1)),
7606 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1),
7607 XEXP (x, 1)));
7608 new_rtx = make_compound_operation (new_rtx, in_code);
7611 /* If we are have (and (rotate X C) M) and C is larger than the number
7612 of bits in M, this is an extraction. */
7614 else if (GET_CODE (XEXP (x, 0)) == ROTATE
7615 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
7616 && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0
7617 && i <= INTVAL (XEXP (XEXP (x, 0), 1)))
7619 new_rtx = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
7620 new_rtx = make_extraction (mode, new_rtx,
7621 (GET_MODE_PRECISION (mode)
7622 - INTVAL (XEXP (XEXP (x, 0), 1))),
7623 NULL_RTX, i, 1, 0, in_code == COMPARE);
7626 /* On machines without logical shifts, if the operand of the AND is
7627 a logical shift and our mask turns off all the propagated sign
7628 bits, we can replace the logical shift with an arithmetic shift. */
7629 else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7630 && !have_insn_for (LSHIFTRT, mode)
7631 && have_insn_for (ASHIFTRT, mode)
7632 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
7633 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7634 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
7635 && mode_width <= HOST_BITS_PER_WIDE_INT)
7637 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
7639 mask >>= INTVAL (XEXP (XEXP (x, 0), 1));
7640 if ((INTVAL (XEXP (x, 1)) & ~mask) == 0)
7641 SUBST (XEXP (x, 0),
7642 gen_rtx_ASHIFTRT (mode,
7643 make_compound_operation
7644 (XEXP (XEXP (x, 0), 0), next_code),
7645 XEXP (XEXP (x, 0), 1)));
7648 /* If the constant is one less than a power of two, this might be
7649 representable by an extraction even if no shift is present.
7650 If it doesn't end up being a ZERO_EXTEND, we will ignore it unless
7651 we are in a COMPARE. */
7652 else if ((i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0)
7653 new_rtx = make_extraction (mode,
7654 make_compound_operation (XEXP (x, 0),
7655 next_code),
7656 0, NULL_RTX, i, 1, 0, in_code == COMPARE);
7658 /* If we are in a comparison and this is an AND with a power of two,
7659 convert this into the appropriate bit extract. */
7660 else if (in_code == COMPARE
7661 && (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0)
7662 new_rtx = make_extraction (mode,
7663 make_compound_operation (XEXP (x, 0),
7664 next_code),
7665 i, NULL_RTX, 1, 1, 0, 1);
7667 break;
7669 case LSHIFTRT:
7670 /* If the sign bit is known to be zero, replace this with an
7671 arithmetic shift. */
7672 if (have_insn_for (ASHIFTRT, mode)
7673 && ! have_insn_for (LSHIFTRT, mode)
7674 && mode_width <= HOST_BITS_PER_WIDE_INT
7675 && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0)
7677 new_rtx = gen_rtx_ASHIFTRT (mode,
7678 make_compound_operation (XEXP (x, 0),
7679 next_code),
7680 XEXP (x, 1));
7681 break;
7684 /* ... fall through ... */
7686 case ASHIFTRT:
7687 lhs = XEXP (x, 0);
7688 rhs = XEXP (x, 1);
7690 /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1,
7691 this is a SIGN_EXTRACT. */
7692 if (CONST_INT_P (rhs)
7693 && GET_CODE (lhs) == ASHIFT
7694 && CONST_INT_P (XEXP (lhs, 1))
7695 && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1))
7696 && INTVAL (XEXP (lhs, 1)) >= 0
7697 && INTVAL (rhs) < mode_width)
7699 new_rtx = make_compound_operation (XEXP (lhs, 0), next_code);
7700 new_rtx = make_extraction (mode, new_rtx,
7701 INTVAL (rhs) - INTVAL (XEXP (lhs, 1)),
7702 NULL_RTX, mode_width - INTVAL (rhs),
7703 code == LSHIFTRT, 0, in_code == COMPARE);
7704 break;
7707 /* See if we have operations between an ASHIFTRT and an ASHIFT.
7708 If so, try to merge the shifts into a SIGN_EXTEND. We could
7709 also do this for some cases of SIGN_EXTRACT, but it doesn't
7710 seem worth the effort; the case checked for occurs on Alpha. */
7712 if (!OBJECT_P (lhs)
7713 && ! (GET_CODE (lhs) == SUBREG
7714 && (OBJECT_P (SUBREG_REG (lhs))))
7715 && CONST_INT_P (rhs)
7716 && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT
7717 && INTVAL (rhs) < mode_width
7718 && (new_rtx = extract_left_shift (lhs, INTVAL (rhs))) != 0)
7719 new_rtx = make_extraction (mode, make_compound_operation (new_rtx, next_code),
7720 0, NULL_RTX, mode_width - INTVAL (rhs),
7721 code == LSHIFTRT, 0, in_code == COMPARE);
7723 break;
7725 case SUBREG:
7726 /* Call ourselves recursively on the inner expression. If we are
7727 narrowing the object and it has a different RTL code from
7728 what it originally did, do this SUBREG as a force_to_mode. */
7730 rtx inner = SUBREG_REG (x), simplified;
7731 enum rtx_code subreg_code = in_code;
7733 /* If in_code is COMPARE, it isn't always safe to pass it through
7734 to the recursive make_compound_operation call. */
7735 if (subreg_code == COMPARE
7736 && (!subreg_lowpart_p (x)
7737 || GET_CODE (inner) == SUBREG
7738 /* (subreg:SI (and:DI (reg:DI) (const_int 0x800000000)) 0)
7739 is (const_int 0), rather than
7740 (subreg:SI (lshiftrt:DI (reg:DI) (const_int 35)) 0). */
7741 || (GET_CODE (inner) == AND
7742 && CONST_INT_P (XEXP (inner, 1))
7743 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (inner))
7744 && exact_log2 (UINTVAL (XEXP (inner, 1)))
7745 >= GET_MODE_BITSIZE (mode))))
7746 subreg_code = SET;
7748 tem = make_compound_operation (inner, subreg_code);
7750 simplified
7751 = simplify_subreg (mode, tem, GET_MODE (inner), SUBREG_BYTE (x));
7752 if (simplified)
7753 tem = simplified;
7755 if (GET_CODE (tem) != GET_CODE (inner)
7756 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (inner))
7757 && subreg_lowpart_p (x))
7759 rtx newer
7760 = force_to_mode (tem, mode, ~(unsigned HOST_WIDE_INT) 0, 0);
7762 /* If we have something other than a SUBREG, we might have
7763 done an expansion, so rerun ourselves. */
7764 if (GET_CODE (newer) != SUBREG)
7765 newer = make_compound_operation (newer, in_code);
7767 /* force_to_mode can expand compounds. If it just re-expanded the
7768 compound, use gen_lowpart to convert to the desired mode. */
7769 if (rtx_equal_p (newer, x)
7770 /* Likewise if it re-expanded the compound only partially.
7771 This happens for SUBREG of ZERO_EXTRACT if they extract
7772 the same number of bits. */
7773 || (GET_CODE (newer) == SUBREG
7774 && (GET_CODE (SUBREG_REG (newer)) == LSHIFTRT
7775 || GET_CODE (SUBREG_REG (newer)) == ASHIFTRT)
7776 && GET_CODE (inner) == AND
7777 && rtx_equal_p (SUBREG_REG (newer), XEXP (inner, 0))))
7778 return gen_lowpart (GET_MODE (x), tem);
7780 return newer;
7783 if (simplified)
7784 return tem;
7786 break;
7788 default:
7789 break;
7792 if (new_rtx)
7794 x = gen_lowpart (mode, new_rtx);
7795 code = GET_CODE (x);
7798 /* Now recursively process each operand of this operation. We need to
7799 handle ZERO_EXTEND specially so that we don't lose track of the
7800 inner mode. */
7801 if (GET_CODE (x) == ZERO_EXTEND)
7803 new_rtx = make_compound_operation (XEXP (x, 0), next_code);
7804 tem = simplify_const_unary_operation (ZERO_EXTEND, GET_MODE (x),
7805 new_rtx, GET_MODE (XEXP (x, 0)));
7806 if (tem)
7807 return tem;
7808 SUBST (XEXP (x, 0), new_rtx);
7809 return x;
7812 fmt = GET_RTX_FORMAT (code);
7813 for (i = 0; i < GET_RTX_LENGTH (code); i++)
7814 if (fmt[i] == 'e')
7816 new_rtx = make_compound_operation (XEXP (x, i), next_code);
7817 SUBST (XEXP (x, i), new_rtx);
7819 else if (fmt[i] == 'E')
7820 for (j = 0; j < XVECLEN (x, i); j++)
7822 new_rtx = make_compound_operation (XVECEXP (x, i, j), next_code);
7823 SUBST (XVECEXP (x, i, j), new_rtx);
7826 maybe_swap:
7827 /* If this is a commutative operation, the changes to the operands
7828 may have made it noncanonical. */
7829 if (COMMUTATIVE_ARITH_P (x)
7830 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
7832 tem = XEXP (x, 0);
7833 SUBST (XEXP (x, 0), XEXP (x, 1));
7834 SUBST (XEXP (x, 1), tem);
7837 return x;
7840 /* Given M see if it is a value that would select a field of bits
7841 within an item, but not the entire word. Return -1 if not.
7842 Otherwise, return the starting position of the field, where 0 is the
7843 low-order bit.
7845 *PLEN is set to the length of the field. */
7847 static int
7848 get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen)
7850 /* Get the bit number of the first 1 bit from the right, -1 if none. */
7851 int pos = m ? ctz_hwi (m) : -1;
7852 int len = 0;
7854 if (pos >= 0)
7855 /* Now shift off the low-order zero bits and see if we have a
7856 power of two minus 1. */
7857 len = exact_log2 ((m >> pos) + 1);
7859 if (len <= 0)
7860 pos = -1;
7862 *plen = len;
7863 return pos;
7866 /* If X refers to a register that equals REG in value, replace these
7867 references with REG. */
7868 static rtx
7869 canon_reg_for_combine (rtx x, rtx reg)
7871 rtx op0, op1, op2;
7872 const char *fmt;
7873 int i;
7874 bool copied;
7876 enum rtx_code code = GET_CODE (x);
7877 switch (GET_RTX_CLASS (code))
7879 case RTX_UNARY:
7880 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
7881 if (op0 != XEXP (x, 0))
7882 return simplify_gen_unary (GET_CODE (x), GET_MODE (x), op0,
7883 GET_MODE (reg));
7884 break;
7886 case RTX_BIN_ARITH:
7887 case RTX_COMM_ARITH:
7888 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
7889 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
7890 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7891 return simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1);
7892 break;
7894 case RTX_COMPARE:
7895 case RTX_COMM_COMPARE:
7896 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
7897 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
7898 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7899 return simplify_gen_relational (GET_CODE (x), GET_MODE (x),
7900 GET_MODE (op0), op0, op1);
7901 break;
7903 case RTX_TERNARY:
7904 case RTX_BITFIELD_OPS:
7905 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
7906 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
7907 op2 = canon_reg_for_combine (XEXP (x, 2), reg);
7908 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1) || op2 != XEXP (x, 2))
7909 return simplify_gen_ternary (GET_CODE (x), GET_MODE (x),
7910 GET_MODE (op0), op0, op1, op2);
7912 case RTX_OBJ:
7913 if (REG_P (x))
7915 if (rtx_equal_p (get_last_value (reg), x)
7916 || rtx_equal_p (reg, get_last_value (x)))
7917 return reg;
7918 else
7919 break;
7922 /* fall through */
7924 default:
7925 fmt = GET_RTX_FORMAT (code);
7926 copied = false;
7927 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7928 if (fmt[i] == 'e')
7930 rtx op = canon_reg_for_combine (XEXP (x, i), reg);
7931 if (op != XEXP (x, i))
7933 if (!copied)
7935 copied = true;
7936 x = copy_rtx (x);
7938 XEXP (x, i) = op;
7941 else if (fmt[i] == 'E')
7943 int j;
7944 for (j = 0; j < XVECLEN (x, i); j++)
7946 rtx op = canon_reg_for_combine (XVECEXP (x, i, j), reg);
7947 if (op != XVECEXP (x, i, j))
7949 if (!copied)
7951 copied = true;
7952 x = copy_rtx (x);
7954 XVECEXP (x, i, j) = op;
7959 break;
7962 return x;
7965 /* Return X converted to MODE. If the value is already truncated to
7966 MODE we can just return a subreg even though in the general case we
7967 would need an explicit truncation. */
7969 static rtx
7970 gen_lowpart_or_truncate (enum machine_mode mode, rtx x)
7972 if (!CONST_INT_P (x)
7973 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (x))
7974 && !TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (x))
7975 && !(REG_P (x) && reg_truncated_to_mode (mode, x)))
7977 /* Bit-cast X into an integer mode. */
7978 if (!SCALAR_INT_MODE_P (GET_MODE (x)))
7979 x = gen_lowpart (int_mode_for_mode (GET_MODE (x)), x);
7980 x = simplify_gen_unary (TRUNCATE, int_mode_for_mode (mode),
7981 x, GET_MODE (x));
7984 return gen_lowpart (mode, x);
7987 /* See if X can be simplified knowing that we will only refer to it in
7988 MODE and will only refer to those bits that are nonzero in MASK.
7989 If other bits are being computed or if masking operations are done
7990 that select a superset of the bits in MASK, they can sometimes be
7991 ignored.
7993 Return a possibly simplified expression, but always convert X to
7994 MODE. If X is a CONST_INT, AND the CONST_INT with MASK.
7996 If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK
7997 are all off in X. This is used when X will be complemented, by either
7998 NOT, NEG, or XOR. */
8000 static rtx
8001 force_to_mode (rtx x, enum machine_mode mode, unsigned HOST_WIDE_INT mask,
8002 int just_select)
8004 enum rtx_code code = GET_CODE (x);
8005 int next_select = just_select || code == XOR || code == NOT || code == NEG;
8006 enum machine_mode op_mode;
8007 unsigned HOST_WIDE_INT fuller_mask, nonzero;
8008 rtx op0, op1, temp;
8010 /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the
8011 code below will do the wrong thing since the mode of such an
8012 expression is VOIDmode.
8014 Also do nothing if X is a CLOBBER; this can happen if X was
8015 the return value from a call to gen_lowpart. */
8016 if (code == CALL || code == ASM_OPERANDS || code == CLOBBER)
8017 return x;
8019 /* We want to perform the operation is its present mode unless we know
8020 that the operation is valid in MODE, in which case we do the operation
8021 in MODE. */
8022 op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x))
8023 && have_insn_for (code, mode))
8024 ? mode : GET_MODE (x));
8026 /* It is not valid to do a right-shift in a narrower mode
8027 than the one it came in with. */
8028 if ((code == LSHIFTRT || code == ASHIFTRT)
8029 && GET_MODE_PRECISION (mode) < GET_MODE_PRECISION (GET_MODE (x)))
8030 op_mode = GET_MODE (x);
8032 /* Truncate MASK to fit OP_MODE. */
8033 if (op_mode)
8034 mask &= GET_MODE_MASK (op_mode);
8036 /* When we have an arithmetic operation, or a shift whose count we
8037 do not know, we need to assume that all bits up to the highest-order
8038 bit in MASK will be needed. This is how we form such a mask. */
8039 if (mask & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)))
8040 fuller_mask = ~(unsigned HOST_WIDE_INT) 0;
8041 else
8042 fuller_mask = (((unsigned HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1))
8043 - 1);
8045 /* Determine what bits of X are guaranteed to be (non)zero. */
8046 nonzero = nonzero_bits (x, mode);
8048 /* If none of the bits in X are needed, return a zero. */
8049 if (!just_select && (nonzero & mask) == 0 && !side_effects_p (x))
8050 x = const0_rtx;
8052 /* If X is a CONST_INT, return a new one. Do this here since the
8053 test below will fail. */
8054 if (CONST_INT_P (x))
8056 if (SCALAR_INT_MODE_P (mode))
8057 return gen_int_mode (INTVAL (x) & mask, mode);
8058 else
8060 x = GEN_INT (INTVAL (x) & mask);
8061 return gen_lowpart_common (mode, x);
8065 /* If X is narrower than MODE and we want all the bits in X's mode, just
8066 get X in the proper mode. */
8067 if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)
8068 && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0)
8069 return gen_lowpart (mode, x);
8071 /* We can ignore the effect of a SUBREG if it narrows the mode or
8072 if the constant masks to zero all the bits the mode doesn't have. */
8073 if (GET_CODE (x) == SUBREG
8074 && subreg_lowpart_p (x)
8075 && ((GET_MODE_SIZE (GET_MODE (x))
8076 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
8077 || (0 == (mask
8078 & GET_MODE_MASK (GET_MODE (x))
8079 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))))))
8080 return force_to_mode (SUBREG_REG (x), mode, mask, next_select);
8082 /* The arithmetic simplifications here only work for scalar integer modes. */
8083 if (!SCALAR_INT_MODE_P (mode) || !SCALAR_INT_MODE_P (GET_MODE (x)))
8084 return gen_lowpart_or_truncate (mode, x);
8086 switch (code)
8088 case CLOBBER:
8089 /* If X is a (clobber (const_int)), return it since we know we are
8090 generating something that won't match. */
8091 return x;
8093 case SIGN_EXTEND:
8094 case ZERO_EXTEND:
8095 case ZERO_EXTRACT:
8096 case SIGN_EXTRACT:
8097 x = expand_compound_operation (x);
8098 if (GET_CODE (x) != code)
8099 return force_to_mode (x, mode, mask, next_select);
8100 break;
8102 case TRUNCATE:
8103 /* Similarly for a truncate. */
8104 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
8106 case AND:
8107 /* If this is an AND with a constant, convert it into an AND
8108 whose constant is the AND of that constant with MASK. If it
8109 remains an AND of MASK, delete it since it is redundant. */
8111 if (CONST_INT_P (XEXP (x, 1)))
8113 x = simplify_and_const_int (x, op_mode, XEXP (x, 0),
8114 mask & INTVAL (XEXP (x, 1)));
8116 /* If X is still an AND, see if it is an AND with a mask that
8117 is just some low-order bits. If so, and it is MASK, we don't
8118 need it. */
8120 if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))
8121 && ((INTVAL (XEXP (x, 1)) & GET_MODE_MASK (GET_MODE (x)))
8122 == mask))
8123 x = XEXP (x, 0);
8125 /* If it remains an AND, try making another AND with the bits
8126 in the mode mask that aren't in MASK turned on. If the
8127 constant in the AND is wide enough, this might make a
8128 cheaper constant. */
8130 if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))
8131 && GET_MODE_MASK (GET_MODE (x)) != mask
8132 && HWI_COMPUTABLE_MODE_P (GET_MODE (x)))
8134 unsigned HOST_WIDE_INT cval
8135 = UINTVAL (XEXP (x, 1))
8136 | (GET_MODE_MASK (GET_MODE (x)) & ~mask);
8137 rtx y;
8139 y = simplify_gen_binary (AND, GET_MODE (x), XEXP (x, 0),
8140 gen_int_mode (cval, GET_MODE (x)));
8141 if (set_src_cost (y, optimize_this_for_speed_p)
8142 < set_src_cost (x, optimize_this_for_speed_p))
8143 x = y;
8146 break;
8149 goto binop;
8151 case PLUS:
8152 /* In (and (plus FOO C1) M), if M is a mask that just turns off
8153 low-order bits (as in an alignment operation) and FOO is already
8154 aligned to that boundary, mask C1 to that boundary as well.
8155 This may eliminate that PLUS and, later, the AND. */
8158 unsigned int width = GET_MODE_PRECISION (mode);
8159 unsigned HOST_WIDE_INT smask = mask;
8161 /* If MODE is narrower than HOST_WIDE_INT and mask is a negative
8162 number, sign extend it. */
8164 if (width < HOST_BITS_PER_WIDE_INT
8165 && (smask & (HOST_WIDE_INT_1U << (width - 1))) != 0)
8166 smask |= HOST_WIDE_INT_M1U << width;
8168 if (CONST_INT_P (XEXP (x, 1))
8169 && exact_log2 (- smask) >= 0
8170 && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0
8171 && (INTVAL (XEXP (x, 1)) & ~smask) != 0)
8172 return force_to_mode (plus_constant (GET_MODE (x), XEXP (x, 0),
8173 (INTVAL (XEXP (x, 1)) & smask)),
8174 mode, smask, next_select);
8177 /* ... fall through ... */
8179 case MULT:
8180 /* For PLUS, MINUS and MULT, we need any bits less significant than the
8181 most significant bit in MASK since carries from those bits will
8182 affect the bits we are interested in. */
8183 mask = fuller_mask;
8184 goto binop;
8186 case MINUS:
8187 /* If X is (minus C Y) where C's least set bit is larger than any bit
8188 in the mask, then we may replace with (neg Y). */
8189 if (CONST_INT_P (XEXP (x, 0))
8190 && (((unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 0))
8191 & -INTVAL (XEXP (x, 0))))
8192 > mask))
8194 x = simplify_gen_unary (NEG, GET_MODE (x), XEXP (x, 1),
8195 GET_MODE (x));
8196 return force_to_mode (x, mode, mask, next_select);
8199 /* Similarly, if C contains every bit in the fuller_mask, then we may
8200 replace with (not Y). */
8201 if (CONST_INT_P (XEXP (x, 0))
8202 && ((UINTVAL (XEXP (x, 0)) | fuller_mask) == UINTVAL (XEXP (x, 0))))
8204 x = simplify_gen_unary (NOT, GET_MODE (x),
8205 XEXP (x, 1), GET_MODE (x));
8206 return force_to_mode (x, mode, mask, next_select);
8209 mask = fuller_mask;
8210 goto binop;
8212 case IOR:
8213 case XOR:
8214 /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and
8215 LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...)
8216 operation which may be a bitfield extraction. Ensure that the
8217 constant we form is not wider than the mode of X. */
8219 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
8220 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
8221 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
8222 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
8223 && CONST_INT_P (XEXP (x, 1))
8224 && ((INTVAL (XEXP (XEXP (x, 0), 1))
8225 + floor_log2 (INTVAL (XEXP (x, 1))))
8226 < GET_MODE_PRECISION (GET_MODE (x)))
8227 && (UINTVAL (XEXP (x, 1))
8228 & ~nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0)
8230 temp = gen_int_mode ((INTVAL (XEXP (x, 1)) & mask)
8231 << INTVAL (XEXP (XEXP (x, 0), 1)),
8232 GET_MODE (x));
8233 temp = simplify_gen_binary (GET_CODE (x), GET_MODE (x),
8234 XEXP (XEXP (x, 0), 0), temp);
8235 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), temp,
8236 XEXP (XEXP (x, 0), 1));
8237 return force_to_mode (x, mode, mask, next_select);
8240 binop:
8241 /* For most binary operations, just propagate into the operation and
8242 change the mode if we have an operation of that mode. */
8244 op0 = force_to_mode (XEXP (x, 0), mode, mask, next_select);
8245 op1 = force_to_mode (XEXP (x, 1), mode, mask, next_select);
8247 /* If we ended up truncating both operands, truncate the result of the
8248 operation instead. */
8249 if (GET_CODE (op0) == TRUNCATE
8250 && GET_CODE (op1) == TRUNCATE)
8252 op0 = XEXP (op0, 0);
8253 op1 = XEXP (op1, 0);
8256 op0 = gen_lowpart_or_truncate (op_mode, op0);
8257 op1 = gen_lowpart_or_truncate (op_mode, op1);
8259 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
8260 x = simplify_gen_binary (code, op_mode, op0, op1);
8261 break;
8263 case ASHIFT:
8264 /* For left shifts, do the same, but just for the first operand.
8265 However, we cannot do anything with shifts where we cannot
8266 guarantee that the counts are smaller than the size of the mode
8267 because such a count will have a different meaning in a
8268 wider mode. */
8270 if (! (CONST_INT_P (XEXP (x, 1))
8271 && INTVAL (XEXP (x, 1)) >= 0
8272 && INTVAL (XEXP (x, 1)) < GET_MODE_PRECISION (mode))
8273 && ! (GET_MODE (XEXP (x, 1)) != VOIDmode
8274 && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1)))
8275 < (unsigned HOST_WIDE_INT) GET_MODE_PRECISION (mode))))
8276 break;
8278 /* If the shift count is a constant and we can do arithmetic in
8279 the mode of the shift, refine which bits we need. Otherwise, use the
8280 conservative form of the mask. */
8281 if (CONST_INT_P (XEXP (x, 1))
8282 && INTVAL (XEXP (x, 1)) >= 0
8283 && INTVAL (XEXP (x, 1)) < GET_MODE_PRECISION (op_mode)
8284 && HWI_COMPUTABLE_MODE_P (op_mode))
8285 mask >>= INTVAL (XEXP (x, 1));
8286 else
8287 mask = fuller_mask;
8289 op0 = gen_lowpart_or_truncate (op_mode,
8290 force_to_mode (XEXP (x, 0), op_mode,
8291 mask, next_select));
8293 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
8294 x = simplify_gen_binary (code, op_mode, op0, XEXP (x, 1));
8295 break;
8297 case LSHIFTRT:
8298 /* Here we can only do something if the shift count is a constant,
8299 this shift constant is valid for the host, and we can do arithmetic
8300 in OP_MODE. */
8302 if (CONST_INT_P (XEXP (x, 1))
8303 && INTVAL (XEXP (x, 1)) >= 0
8304 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
8305 && HWI_COMPUTABLE_MODE_P (op_mode))
8307 rtx inner = XEXP (x, 0);
8308 unsigned HOST_WIDE_INT inner_mask;
8310 /* Select the mask of the bits we need for the shift operand. */
8311 inner_mask = mask << INTVAL (XEXP (x, 1));
8313 /* We can only change the mode of the shift if we can do arithmetic
8314 in the mode of the shift and INNER_MASK is no wider than the
8315 width of X's mode. */
8316 if ((inner_mask & ~GET_MODE_MASK (GET_MODE (x))) != 0)
8317 op_mode = GET_MODE (x);
8319 inner = force_to_mode (inner, op_mode, inner_mask, next_select);
8321 if (GET_MODE (x) != op_mode || inner != XEXP (x, 0))
8322 x = simplify_gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1));
8325 /* If we have (and (lshiftrt FOO C1) C2) where the combination of the
8326 shift and AND produces only copies of the sign bit (C2 is one less
8327 than a power of two), we can do this with just a shift. */
8329 if (GET_CODE (x) == LSHIFTRT
8330 && CONST_INT_P (XEXP (x, 1))
8331 /* The shift puts one of the sign bit copies in the least significant
8332 bit. */
8333 && ((INTVAL (XEXP (x, 1))
8334 + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))))
8335 >= GET_MODE_PRECISION (GET_MODE (x)))
8336 && exact_log2 (mask + 1) >= 0
8337 /* Number of bits left after the shift must be more than the mask
8338 needs. */
8339 && ((INTVAL (XEXP (x, 1)) + exact_log2 (mask + 1))
8340 <= GET_MODE_PRECISION (GET_MODE (x)))
8341 /* Must be more sign bit copies than the mask needs. */
8342 && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
8343 >= exact_log2 (mask + 1)))
8344 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0),
8345 GEN_INT (GET_MODE_PRECISION (GET_MODE (x))
8346 - exact_log2 (mask + 1)));
8348 goto shiftrt;
8350 case ASHIFTRT:
8351 /* If we are just looking for the sign bit, we don't need this shift at
8352 all, even if it has a variable count. */
8353 if (val_signbit_p (GET_MODE (x), mask))
8354 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
8356 /* If this is a shift by a constant, get a mask that contains those bits
8357 that are not copies of the sign bit. We then have two cases: If
8358 MASK only includes those bits, this can be a logical shift, which may
8359 allow simplifications. If MASK is a single-bit field not within
8360 those bits, we are requesting a copy of the sign bit and hence can
8361 shift the sign bit to the appropriate location. */
8363 if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) >= 0
8364 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT)
8366 int i;
8368 /* If the considered data is wider than HOST_WIDE_INT, we can't
8369 represent a mask for all its bits in a single scalar.
8370 But we only care about the lower bits, so calculate these. */
8372 if (GET_MODE_PRECISION (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT)
8374 nonzero = ~(unsigned HOST_WIDE_INT) 0;
8376 /* GET_MODE_PRECISION (GET_MODE (x)) - INTVAL (XEXP (x, 1))
8377 is the number of bits a full-width mask would have set.
8378 We need only shift if these are fewer than nonzero can
8379 hold. If not, we must keep all bits set in nonzero. */
8381 if (GET_MODE_PRECISION (GET_MODE (x)) - INTVAL (XEXP (x, 1))
8382 < HOST_BITS_PER_WIDE_INT)
8383 nonzero >>= INTVAL (XEXP (x, 1))
8384 + HOST_BITS_PER_WIDE_INT
8385 - GET_MODE_PRECISION (GET_MODE (x)) ;
8387 else
8389 nonzero = GET_MODE_MASK (GET_MODE (x));
8390 nonzero >>= INTVAL (XEXP (x, 1));
8393 if ((mask & ~nonzero) == 0)
8395 x = simplify_shift_const (NULL_RTX, LSHIFTRT, GET_MODE (x),
8396 XEXP (x, 0), INTVAL (XEXP (x, 1)));
8397 if (GET_CODE (x) != ASHIFTRT)
8398 return force_to_mode (x, mode, mask, next_select);
8401 else if ((i = exact_log2 (mask)) >= 0)
8403 x = simplify_shift_const
8404 (NULL_RTX, LSHIFTRT, GET_MODE (x), XEXP (x, 0),
8405 GET_MODE_PRECISION (GET_MODE (x)) - 1 - i);
8407 if (GET_CODE (x) != ASHIFTRT)
8408 return force_to_mode (x, mode, mask, next_select);
8412 /* If MASK is 1, convert this to an LSHIFTRT. This can be done
8413 even if the shift count isn't a constant. */
8414 if (mask == 1)
8415 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
8416 XEXP (x, 0), XEXP (x, 1));
8418 shiftrt:
8420 /* If this is a zero- or sign-extension operation that just affects bits
8421 we don't care about, remove it. Be sure the call above returned
8422 something that is still a shift. */
8424 if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT)
8425 && CONST_INT_P (XEXP (x, 1))
8426 && INTVAL (XEXP (x, 1)) >= 0
8427 && (INTVAL (XEXP (x, 1))
8428 <= GET_MODE_PRECISION (GET_MODE (x)) - (floor_log2 (mask) + 1))
8429 && GET_CODE (XEXP (x, 0)) == ASHIFT
8430 && XEXP (XEXP (x, 0), 1) == XEXP (x, 1))
8431 return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask,
8432 next_select);
8434 break;
8436 case ROTATE:
8437 case ROTATERT:
8438 /* If the shift count is constant and we can do computations
8439 in the mode of X, compute where the bits we care about are.
8440 Otherwise, we can't do anything. Don't change the mode of
8441 the shift or propagate MODE into the shift, though. */
8442 if (CONST_INT_P (XEXP (x, 1))
8443 && INTVAL (XEXP (x, 1)) >= 0)
8445 temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE,
8446 GET_MODE (x),
8447 gen_int_mode (mask, GET_MODE (x)),
8448 XEXP (x, 1));
8449 if (temp && CONST_INT_P (temp))
8450 SUBST (XEXP (x, 0),
8451 force_to_mode (XEXP (x, 0), GET_MODE (x),
8452 INTVAL (temp), next_select));
8454 break;
8456 case NEG:
8457 /* If we just want the low-order bit, the NEG isn't needed since it
8458 won't change the low-order bit. */
8459 if (mask == 1)
8460 return force_to_mode (XEXP (x, 0), mode, mask, just_select);
8462 /* We need any bits less significant than the most significant bit in
8463 MASK since carries from those bits will affect the bits we are
8464 interested in. */
8465 mask = fuller_mask;
8466 goto unop;
8468 case NOT:
8469 /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the
8470 same as the XOR case above. Ensure that the constant we form is not
8471 wider than the mode of X. */
8473 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
8474 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
8475 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
8476 && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask)
8477 < GET_MODE_PRECISION (GET_MODE (x)))
8478 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT)
8480 temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)),
8481 GET_MODE (x));
8482 temp = simplify_gen_binary (XOR, GET_MODE (x),
8483 XEXP (XEXP (x, 0), 0), temp);
8484 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
8485 temp, XEXP (XEXP (x, 0), 1));
8487 return force_to_mode (x, mode, mask, next_select);
8490 /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must
8491 use the full mask inside the NOT. */
8492 mask = fuller_mask;
8494 unop:
8495 op0 = gen_lowpart_or_truncate (op_mode,
8496 force_to_mode (XEXP (x, 0), mode, mask,
8497 next_select));
8498 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
8499 x = simplify_gen_unary (code, op_mode, op0, op_mode);
8500 break;
8502 case NE:
8503 /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included
8504 in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero,
8505 which is equal to STORE_FLAG_VALUE. */
8506 if ((mask & ~STORE_FLAG_VALUE) == 0
8507 && XEXP (x, 1) == const0_rtx
8508 && GET_MODE (XEXP (x, 0)) == mode
8509 && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0
8510 && (nonzero_bits (XEXP (x, 0), mode)
8511 == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE))
8512 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
8514 break;
8516 case IF_THEN_ELSE:
8517 /* We have no way of knowing if the IF_THEN_ELSE can itself be
8518 written in a narrower mode. We play it safe and do not do so. */
8520 SUBST (XEXP (x, 1),
8521 gen_lowpart_or_truncate (GET_MODE (x),
8522 force_to_mode (XEXP (x, 1), mode,
8523 mask, next_select)));
8524 SUBST (XEXP (x, 2),
8525 gen_lowpart_or_truncate (GET_MODE (x),
8526 force_to_mode (XEXP (x, 2), mode,
8527 mask, next_select)));
8528 break;
8530 default:
8531 break;
8534 /* Ensure we return a value of the proper mode. */
8535 return gen_lowpart_or_truncate (mode, x);
8538 /* Return nonzero if X is an expression that has one of two values depending on
8539 whether some other value is zero or nonzero. In that case, we return the
8540 value that is being tested, *PTRUE is set to the value if the rtx being
8541 returned has a nonzero value, and *PFALSE is set to the other alternative.
8543 If we return zero, we set *PTRUE and *PFALSE to X. */
8545 static rtx
8546 if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse)
8548 enum machine_mode mode = GET_MODE (x);
8549 enum rtx_code code = GET_CODE (x);
8550 rtx cond0, cond1, true0, true1, false0, false1;
8551 unsigned HOST_WIDE_INT nz;
8553 /* If we are comparing a value against zero, we are done. */
8554 if ((code == NE || code == EQ)
8555 && XEXP (x, 1) == const0_rtx)
8557 *ptrue = (code == NE) ? const_true_rtx : const0_rtx;
8558 *pfalse = (code == NE) ? const0_rtx : const_true_rtx;
8559 return XEXP (x, 0);
8562 /* If this is a unary operation whose operand has one of two values, apply
8563 our opcode to compute those values. */
8564 else if (UNARY_P (x)
8565 && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0)
8567 *ptrue = simplify_gen_unary (code, mode, true0, GET_MODE (XEXP (x, 0)));
8568 *pfalse = simplify_gen_unary (code, mode, false0,
8569 GET_MODE (XEXP (x, 0)));
8570 return cond0;
8573 /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would
8574 make can't possibly match and would suppress other optimizations. */
8575 else if (code == COMPARE)
8578 /* If this is a binary operation, see if either side has only one of two
8579 values. If either one does or if both do and they are conditional on
8580 the same value, compute the new true and false values. */
8581 else if (BINARY_P (x))
8583 cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0);
8584 cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1);
8586 if ((cond0 != 0 || cond1 != 0)
8587 && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1)))
8589 /* If if_then_else_cond returned zero, then true/false are the
8590 same rtl. We must copy one of them to prevent invalid rtl
8591 sharing. */
8592 if (cond0 == 0)
8593 true0 = copy_rtx (true0);
8594 else if (cond1 == 0)
8595 true1 = copy_rtx (true1);
8597 if (COMPARISON_P (x))
8599 *ptrue = simplify_gen_relational (code, mode, VOIDmode,
8600 true0, true1);
8601 *pfalse = simplify_gen_relational (code, mode, VOIDmode,
8602 false0, false1);
8604 else
8606 *ptrue = simplify_gen_binary (code, mode, true0, true1);
8607 *pfalse = simplify_gen_binary (code, mode, false0, false1);
8610 return cond0 ? cond0 : cond1;
8613 /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the
8614 operands is zero when the other is nonzero, and vice-versa,
8615 and STORE_FLAG_VALUE is 1 or -1. */
8617 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
8618 && (code == PLUS || code == IOR || code == XOR || code == MINUS
8619 || code == UMAX)
8620 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
8622 rtx op0 = XEXP (XEXP (x, 0), 1);
8623 rtx op1 = XEXP (XEXP (x, 1), 1);
8625 cond0 = XEXP (XEXP (x, 0), 0);
8626 cond1 = XEXP (XEXP (x, 1), 0);
8628 if (COMPARISON_P (cond0)
8629 && COMPARISON_P (cond1)
8630 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
8631 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
8632 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
8633 || ((swap_condition (GET_CODE (cond0))
8634 == reversed_comparison_code (cond1, NULL))
8635 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
8636 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
8637 && ! side_effects_p (x))
8639 *ptrue = simplify_gen_binary (MULT, mode, op0, const_true_rtx);
8640 *pfalse = simplify_gen_binary (MULT, mode,
8641 (code == MINUS
8642 ? simplify_gen_unary (NEG, mode,
8643 op1, mode)
8644 : op1),
8645 const_true_rtx);
8646 return cond0;
8650 /* Similarly for MULT, AND and UMIN, except that for these the result
8651 is always zero. */
8652 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
8653 && (code == MULT || code == AND || code == UMIN)
8654 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
8656 cond0 = XEXP (XEXP (x, 0), 0);
8657 cond1 = XEXP (XEXP (x, 1), 0);
8659 if (COMPARISON_P (cond0)
8660 && COMPARISON_P (cond1)
8661 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
8662 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
8663 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
8664 || ((swap_condition (GET_CODE (cond0))
8665 == reversed_comparison_code (cond1, NULL))
8666 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
8667 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
8668 && ! side_effects_p (x))
8670 *ptrue = *pfalse = const0_rtx;
8671 return cond0;
8676 else if (code == IF_THEN_ELSE)
8678 /* If we have IF_THEN_ELSE already, extract the condition and
8679 canonicalize it if it is NE or EQ. */
8680 cond0 = XEXP (x, 0);
8681 *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2);
8682 if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx)
8683 return XEXP (cond0, 0);
8684 else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx)
8686 *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1);
8687 return XEXP (cond0, 0);
8689 else
8690 return cond0;
8693 /* If X is a SUBREG, we can narrow both the true and false values
8694 if the inner expression, if there is a condition. */
8695 else if (code == SUBREG
8696 && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x),
8697 &true0, &false0)))
8699 true0 = simplify_gen_subreg (mode, true0,
8700 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
8701 false0 = simplify_gen_subreg (mode, false0,
8702 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
8703 if (true0 && false0)
8705 *ptrue = true0;
8706 *pfalse = false0;
8707 return cond0;
8711 /* If X is a constant, this isn't special and will cause confusions
8712 if we treat it as such. Likewise if it is equivalent to a constant. */
8713 else if (CONSTANT_P (x)
8714 || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0)))
8717 /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that
8718 will be least confusing to the rest of the compiler. */
8719 else if (mode == BImode)
8721 *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx;
8722 return x;
8725 /* If X is known to be either 0 or -1, those are the true and
8726 false values when testing X. */
8727 else if (x == constm1_rtx || x == const0_rtx
8728 || (mode != VOIDmode
8729 && num_sign_bit_copies (x, mode) == GET_MODE_PRECISION (mode)))
8731 *ptrue = constm1_rtx, *pfalse = const0_rtx;
8732 return x;
8735 /* Likewise for 0 or a single bit. */
8736 else if (HWI_COMPUTABLE_MODE_P (mode)
8737 && exact_log2 (nz = nonzero_bits (x, mode)) >= 0)
8739 *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx;
8740 return x;
8743 /* Otherwise fail; show no condition with true and false values the same. */
8744 *ptrue = *pfalse = x;
8745 return 0;
8748 /* Return the value of expression X given the fact that condition COND
8749 is known to be true when applied to REG as its first operand and VAL
8750 as its second. X is known to not be shared and so can be modified in
8751 place.
8753 We only handle the simplest cases, and specifically those cases that
8754 arise with IF_THEN_ELSE expressions. */
8756 static rtx
8757 known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val)
8759 enum rtx_code code = GET_CODE (x);
8760 rtx temp;
8761 const char *fmt;
8762 int i, j;
8764 if (side_effects_p (x))
8765 return x;
8767 /* If either operand of the condition is a floating point value,
8768 then we have to avoid collapsing an EQ comparison. */
8769 if (cond == EQ
8770 && rtx_equal_p (x, reg)
8771 && ! FLOAT_MODE_P (GET_MODE (x))
8772 && ! FLOAT_MODE_P (GET_MODE (val)))
8773 return val;
8775 if (cond == UNEQ && rtx_equal_p (x, reg))
8776 return val;
8778 /* If X is (abs REG) and we know something about REG's relationship
8779 with zero, we may be able to simplify this. */
8781 if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx)
8782 switch (cond)
8784 case GE: case GT: case EQ:
8785 return XEXP (x, 0);
8786 case LT: case LE:
8787 return simplify_gen_unary (NEG, GET_MODE (XEXP (x, 0)),
8788 XEXP (x, 0),
8789 GET_MODE (XEXP (x, 0)));
8790 default:
8791 break;
8794 /* The only other cases we handle are MIN, MAX, and comparisons if the
8795 operands are the same as REG and VAL. */
8797 else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x))
8799 if (rtx_equal_p (XEXP (x, 0), val))
8800 cond = swap_condition (cond), temp = val, val = reg, reg = temp;
8802 if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val))
8804 if (COMPARISON_P (x))
8806 if (comparison_dominates_p (cond, code))
8807 return const_true_rtx;
8809 code = reversed_comparison_code (x, NULL);
8810 if (code != UNKNOWN
8811 && comparison_dominates_p (cond, code))
8812 return const0_rtx;
8813 else
8814 return x;
8816 else if (code == SMAX || code == SMIN
8817 || code == UMIN || code == UMAX)
8819 int unsignedp = (code == UMIN || code == UMAX);
8821 /* Do not reverse the condition when it is NE or EQ.
8822 This is because we cannot conclude anything about
8823 the value of 'SMAX (x, y)' when x is not equal to y,
8824 but we can when x equals y. */
8825 if ((code == SMAX || code == UMAX)
8826 && ! (cond == EQ || cond == NE))
8827 cond = reverse_condition (cond);
8829 switch (cond)
8831 case GE: case GT:
8832 return unsignedp ? x : XEXP (x, 1);
8833 case LE: case LT:
8834 return unsignedp ? x : XEXP (x, 0);
8835 case GEU: case GTU:
8836 return unsignedp ? XEXP (x, 1) : x;
8837 case LEU: case LTU:
8838 return unsignedp ? XEXP (x, 0) : x;
8839 default:
8840 break;
8845 else if (code == SUBREG)
8847 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (x));
8848 rtx new_rtx, r = known_cond (SUBREG_REG (x), cond, reg, val);
8850 if (SUBREG_REG (x) != r)
8852 /* We must simplify subreg here, before we lose track of the
8853 original inner_mode. */
8854 new_rtx = simplify_subreg (GET_MODE (x), r,
8855 inner_mode, SUBREG_BYTE (x));
8856 if (new_rtx)
8857 return new_rtx;
8858 else
8859 SUBST (SUBREG_REG (x), r);
8862 return x;
8864 /* We don't have to handle SIGN_EXTEND here, because even in the
8865 case of replacing something with a modeless CONST_INT, a
8866 CONST_INT is already (supposed to be) a valid sign extension for
8867 its narrower mode, which implies it's already properly
8868 sign-extended for the wider mode. Now, for ZERO_EXTEND, the
8869 story is different. */
8870 else if (code == ZERO_EXTEND)
8872 enum machine_mode inner_mode = GET_MODE (XEXP (x, 0));
8873 rtx new_rtx, r = known_cond (XEXP (x, 0), cond, reg, val);
8875 if (XEXP (x, 0) != r)
8877 /* We must simplify the zero_extend here, before we lose
8878 track of the original inner_mode. */
8879 new_rtx = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
8880 r, inner_mode);
8881 if (new_rtx)
8882 return new_rtx;
8883 else
8884 SUBST (XEXP (x, 0), r);
8887 return x;
8890 fmt = GET_RTX_FORMAT (code);
8891 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8893 if (fmt[i] == 'e')
8894 SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val));
8895 else if (fmt[i] == 'E')
8896 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8897 SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j),
8898 cond, reg, val));
8901 return x;
8904 /* See if X and Y are equal for the purposes of seeing if we can rewrite an
8905 assignment as a field assignment. */
8907 static int
8908 rtx_equal_for_field_assignment_p (rtx x, rtx y)
8910 if (x == y || rtx_equal_p (x, y))
8911 return 1;
8913 if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y))
8914 return 0;
8916 /* Check for a paradoxical SUBREG of a MEM compared with the MEM.
8917 Note that all SUBREGs of MEM are paradoxical; otherwise they
8918 would have been rewritten. */
8919 if (MEM_P (x) && GET_CODE (y) == SUBREG
8920 && MEM_P (SUBREG_REG (y))
8921 && rtx_equal_p (SUBREG_REG (y),
8922 gen_lowpart (GET_MODE (SUBREG_REG (y)), x)))
8923 return 1;
8925 if (MEM_P (y) && GET_CODE (x) == SUBREG
8926 && MEM_P (SUBREG_REG (x))
8927 && rtx_equal_p (SUBREG_REG (x),
8928 gen_lowpart (GET_MODE (SUBREG_REG (x)), y)))
8929 return 1;
8931 /* We used to see if get_last_value of X and Y were the same but that's
8932 not correct. In one direction, we'll cause the assignment to have
8933 the wrong destination and in the case, we'll import a register into this
8934 insn that might have already have been dead. So fail if none of the
8935 above cases are true. */
8936 return 0;
8939 /* See if X, a SET operation, can be rewritten as a bit-field assignment.
8940 Return that assignment if so.
8942 We only handle the most common cases. */
8944 static rtx
8945 make_field_assignment (rtx x)
8947 rtx dest = SET_DEST (x);
8948 rtx src = SET_SRC (x);
8949 rtx assign;
8950 rtx rhs, lhs;
8951 HOST_WIDE_INT c1;
8952 HOST_WIDE_INT pos;
8953 unsigned HOST_WIDE_INT len;
8954 rtx other;
8955 enum machine_mode mode;
8957 /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is
8958 a clear of a one-bit field. We will have changed it to
8959 (and (rotate (const_int -2) POS) DEST), so check for that. Also check
8960 for a SUBREG. */
8962 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE
8963 && CONST_INT_P (XEXP (XEXP (src, 0), 0))
8964 && INTVAL (XEXP (XEXP (src, 0), 0)) == -2
8965 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
8967 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
8968 1, 1, 1, 0);
8969 if (assign != 0)
8970 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
8971 return x;
8974 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG
8975 && subreg_lowpart_p (XEXP (src, 0))
8976 && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0)))
8977 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0)))))
8978 && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE
8979 && CONST_INT_P (XEXP (SUBREG_REG (XEXP (src, 0)), 0))
8980 && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2
8981 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
8983 assign = make_extraction (VOIDmode, dest, 0,
8984 XEXP (SUBREG_REG (XEXP (src, 0)), 1),
8985 1, 1, 1, 0);
8986 if (assign != 0)
8987 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
8988 return x;
8991 /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a
8992 one-bit field. */
8993 if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT
8994 && XEXP (XEXP (src, 0), 0) == const1_rtx
8995 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
8997 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
8998 1, 1, 1, 0);
8999 if (assign != 0)
9000 return gen_rtx_SET (VOIDmode, assign, const1_rtx);
9001 return x;
9004 /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the
9005 SRC is an AND with all bits of that field set, then we can discard
9006 the AND. */
9007 if (GET_CODE (dest) == ZERO_EXTRACT
9008 && CONST_INT_P (XEXP (dest, 1))
9009 && GET_CODE (src) == AND
9010 && CONST_INT_P (XEXP (src, 1)))
9012 HOST_WIDE_INT width = INTVAL (XEXP (dest, 1));
9013 unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1));
9014 unsigned HOST_WIDE_INT ze_mask;
9016 if (width >= HOST_BITS_PER_WIDE_INT)
9017 ze_mask = -1;
9018 else
9019 ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1;
9021 /* Complete overlap. We can remove the source AND. */
9022 if ((and_mask & ze_mask) == ze_mask)
9023 return gen_rtx_SET (VOIDmode, dest, XEXP (src, 0));
9025 /* Partial overlap. We can reduce the source AND. */
9026 if ((and_mask & ze_mask) != and_mask)
9028 mode = GET_MODE (src);
9029 src = gen_rtx_AND (mode, XEXP (src, 0),
9030 gen_int_mode (and_mask & ze_mask, mode));
9031 return gen_rtx_SET (VOIDmode, dest, src);
9035 /* The other case we handle is assignments into a constant-position
9036 field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents
9037 a mask that has all one bits except for a group of zero bits and
9038 OTHER is known to have zeros where C1 has ones, this is such an
9039 assignment. Compute the position and length from C1. Shift OTHER
9040 to the appropriate position, force it to the required mode, and
9041 make the extraction. Check for the AND in both operands. */
9043 if (GET_CODE (src) != IOR && GET_CODE (src) != XOR)
9044 return x;
9046 rhs = expand_compound_operation (XEXP (src, 0));
9047 lhs = expand_compound_operation (XEXP (src, 1));
9049 if (GET_CODE (rhs) == AND
9050 && CONST_INT_P (XEXP (rhs, 1))
9051 && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest))
9052 c1 = INTVAL (XEXP (rhs, 1)), other = lhs;
9053 else if (GET_CODE (lhs) == AND
9054 && CONST_INT_P (XEXP (lhs, 1))
9055 && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest))
9056 c1 = INTVAL (XEXP (lhs, 1)), other = rhs;
9057 else
9058 return x;
9060 pos = get_pos_from_mask ((~c1) & GET_MODE_MASK (GET_MODE (dest)), &len);
9061 if (pos < 0 || pos + len > GET_MODE_PRECISION (GET_MODE (dest))
9062 || GET_MODE_PRECISION (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT
9063 || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0)
9064 return x;
9066 assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0);
9067 if (assign == 0)
9068 return x;
9070 /* The mode to use for the source is the mode of the assignment, or of
9071 what is inside a possible STRICT_LOW_PART. */
9072 mode = (GET_CODE (assign) == STRICT_LOW_PART
9073 ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign));
9075 /* Shift OTHER right POS places and make it the source, restricting it
9076 to the proper length and mode. */
9078 src = canon_reg_for_combine (simplify_shift_const (NULL_RTX, LSHIFTRT,
9079 GET_MODE (src),
9080 other, pos),
9081 dest);
9082 src = force_to_mode (src, mode,
9083 GET_MODE_PRECISION (mode) >= HOST_BITS_PER_WIDE_INT
9084 ? ~(unsigned HOST_WIDE_INT) 0
9085 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
9088 /* If SRC is masked by an AND that does not make a difference in
9089 the value being stored, strip it. */
9090 if (GET_CODE (assign) == ZERO_EXTRACT
9091 && CONST_INT_P (XEXP (assign, 1))
9092 && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT
9093 && GET_CODE (src) == AND
9094 && CONST_INT_P (XEXP (src, 1))
9095 && UINTVAL (XEXP (src, 1))
9096 == ((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (assign, 1))) - 1)
9097 src = XEXP (src, 0);
9099 return gen_rtx_SET (VOIDmode, assign, src);
9102 /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c)
9103 if so. */
9105 static rtx
9106 apply_distributive_law (rtx x)
9108 enum rtx_code code = GET_CODE (x);
9109 enum rtx_code inner_code;
9110 rtx lhs, rhs, other;
9111 rtx tem;
9113 /* Distributivity is not true for floating point as it can change the
9114 value. So we don't do it unless -funsafe-math-optimizations. */
9115 if (FLOAT_MODE_P (GET_MODE (x))
9116 && ! flag_unsafe_math_optimizations)
9117 return x;
9119 /* The outer operation can only be one of the following: */
9120 if (code != IOR && code != AND && code != XOR
9121 && code != PLUS && code != MINUS)
9122 return x;
9124 lhs = XEXP (x, 0);
9125 rhs = XEXP (x, 1);
9127 /* If either operand is a primitive we can't do anything, so get out
9128 fast. */
9129 if (OBJECT_P (lhs) || OBJECT_P (rhs))
9130 return x;
9132 lhs = expand_compound_operation (lhs);
9133 rhs = expand_compound_operation (rhs);
9134 inner_code = GET_CODE (lhs);
9135 if (inner_code != GET_CODE (rhs))
9136 return x;
9138 /* See if the inner and outer operations distribute. */
9139 switch (inner_code)
9141 case LSHIFTRT:
9142 case ASHIFTRT:
9143 case AND:
9144 case IOR:
9145 /* These all distribute except over PLUS. */
9146 if (code == PLUS || code == MINUS)
9147 return x;
9148 break;
9150 case MULT:
9151 if (code != PLUS && code != MINUS)
9152 return x;
9153 break;
9155 case ASHIFT:
9156 /* This is also a multiply, so it distributes over everything. */
9157 break;
9159 /* This used to handle SUBREG, but this turned out to be counter-
9160 productive, since (subreg (op ...)) usually is not handled by
9161 insn patterns, and this "optimization" therefore transformed
9162 recognizable patterns into unrecognizable ones. Therefore the
9163 SUBREG case was removed from here.
9165 It is possible that distributing SUBREG over arithmetic operations
9166 leads to an intermediate result than can then be optimized further,
9167 e.g. by moving the outer SUBREG to the other side of a SET as done
9168 in simplify_set. This seems to have been the original intent of
9169 handling SUBREGs here.
9171 However, with current GCC this does not appear to actually happen,
9172 at least on major platforms. If some case is found where removing
9173 the SUBREG case here prevents follow-on optimizations, distributing
9174 SUBREGs ought to be re-added at that place, e.g. in simplify_set. */
9176 default:
9177 return x;
9180 /* Set LHS and RHS to the inner operands (A and B in the example
9181 above) and set OTHER to the common operand (C in the example).
9182 There is only one way to do this unless the inner operation is
9183 commutative. */
9184 if (COMMUTATIVE_ARITH_P (lhs)
9185 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0)))
9186 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1);
9187 else if (COMMUTATIVE_ARITH_P (lhs)
9188 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1)))
9189 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0);
9190 else if (COMMUTATIVE_ARITH_P (lhs)
9191 && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0)))
9192 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1);
9193 else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1)))
9194 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0);
9195 else
9196 return x;
9198 /* Form the new inner operation, seeing if it simplifies first. */
9199 tem = simplify_gen_binary (code, GET_MODE (x), lhs, rhs);
9201 /* There is one exception to the general way of distributing:
9202 (a | c) ^ (b | c) -> (a ^ b) & ~c */
9203 if (code == XOR && inner_code == IOR)
9205 inner_code = AND;
9206 other = simplify_gen_unary (NOT, GET_MODE (x), other, GET_MODE (x));
9209 /* We may be able to continuing distributing the result, so call
9210 ourselves recursively on the inner operation before forming the
9211 outer operation, which we return. */
9212 return simplify_gen_binary (inner_code, GET_MODE (x),
9213 apply_distributive_law (tem), other);
9216 /* See if X is of the form (* (+ A B) C), and if so convert to
9217 (+ (* A C) (* B C)) and try to simplify.
9219 Most of the time, this results in no change. However, if some of
9220 the operands are the same or inverses of each other, simplifications
9221 will result.
9223 For example, (and (ior A B) (not B)) can occur as the result of
9224 expanding a bit field assignment. When we apply the distributive
9225 law to this, we get (ior (and (A (not B))) (and (B (not B)))),
9226 which then simplifies to (and (A (not B))).
9228 Note that no checks happen on the validity of applying the inverse
9229 distributive law. This is pointless since we can do it in the
9230 few places where this routine is called.
9232 N is the index of the term that is decomposed (the arithmetic operation,
9233 i.e. (+ A B) in the first example above). !N is the index of the term that
9234 is distributed, i.e. of C in the first example above. */
9235 static rtx
9236 distribute_and_simplify_rtx (rtx x, int n)
9238 enum machine_mode mode;
9239 enum rtx_code outer_code, inner_code;
9240 rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp;
9242 /* Distributivity is not true for floating point as it can change the
9243 value. So we don't do it unless -funsafe-math-optimizations. */
9244 if (FLOAT_MODE_P (GET_MODE (x))
9245 && ! flag_unsafe_math_optimizations)
9246 return NULL_RTX;
9248 decomposed = XEXP (x, n);
9249 if (!ARITHMETIC_P (decomposed))
9250 return NULL_RTX;
9252 mode = GET_MODE (x);
9253 outer_code = GET_CODE (x);
9254 distributed = XEXP (x, !n);
9256 inner_code = GET_CODE (decomposed);
9257 inner_op0 = XEXP (decomposed, 0);
9258 inner_op1 = XEXP (decomposed, 1);
9260 /* Special case (and (xor B C) (not A)), which is equivalent to
9261 (xor (ior A B) (ior A C)) */
9262 if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT)
9264 distributed = XEXP (distributed, 0);
9265 outer_code = IOR;
9268 if (n == 0)
9270 /* Distribute the second term. */
9271 new_op0 = simplify_gen_binary (outer_code, mode, inner_op0, distributed);
9272 new_op1 = simplify_gen_binary (outer_code, mode, inner_op1, distributed);
9274 else
9276 /* Distribute the first term. */
9277 new_op0 = simplify_gen_binary (outer_code, mode, distributed, inner_op0);
9278 new_op1 = simplify_gen_binary (outer_code, mode, distributed, inner_op1);
9281 tmp = apply_distributive_law (simplify_gen_binary (inner_code, mode,
9282 new_op0, new_op1));
9283 if (GET_CODE (tmp) != outer_code
9284 && (set_src_cost (tmp, optimize_this_for_speed_p)
9285 < set_src_cost (x, optimize_this_for_speed_p)))
9286 return tmp;
9288 return NULL_RTX;
9291 /* Simplify a logical `and' of VAROP with the constant CONSTOP, to be done
9292 in MODE. Return an equivalent form, if different from (and VAROP
9293 (const_int CONSTOP)). Otherwise, return NULL_RTX. */
9295 static rtx
9296 simplify_and_const_int_1 (enum machine_mode mode, rtx varop,
9297 unsigned HOST_WIDE_INT constop)
9299 unsigned HOST_WIDE_INT nonzero;
9300 unsigned HOST_WIDE_INT orig_constop;
9301 rtx orig_varop;
9302 int i;
9304 orig_varop = varop;
9305 orig_constop = constop;
9306 if (GET_CODE (varop) == CLOBBER)
9307 return NULL_RTX;
9309 /* Simplify VAROP knowing that we will be only looking at some of the
9310 bits in it.
9312 Note by passing in CONSTOP, we guarantee that the bits not set in
9313 CONSTOP are not significant and will never be examined. We must
9314 ensure that is the case by explicitly masking out those bits
9315 before returning. */
9316 varop = force_to_mode (varop, mode, constop, 0);
9318 /* If VAROP is a CLOBBER, we will fail so return it. */
9319 if (GET_CODE (varop) == CLOBBER)
9320 return varop;
9322 /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP
9323 to VAROP and return the new constant. */
9324 if (CONST_INT_P (varop))
9325 return gen_int_mode (INTVAL (varop) & constop, mode);
9327 /* See what bits may be nonzero in VAROP. Unlike the general case of
9328 a call to nonzero_bits, here we don't care about bits outside
9329 MODE. */
9331 nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
9333 /* Turn off all bits in the constant that are known to already be zero.
9334 Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
9335 which is tested below. */
9337 constop &= nonzero;
9339 /* If we don't have any bits left, return zero. */
9340 if (constop == 0)
9341 return const0_rtx;
9343 /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is
9344 a power of two, we can replace this with an ASHIFT. */
9345 if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1
9346 && (i = exact_log2 (constop)) >= 0)
9347 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i);
9349 /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR
9350 or XOR, then try to apply the distributive law. This may eliminate
9351 operations if either branch can be simplified because of the AND.
9352 It may also make some cases more complex, but those cases probably
9353 won't match a pattern either with or without this. */
9355 if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR)
9356 return
9357 gen_lowpart
9358 (mode,
9359 apply_distributive_law
9360 (simplify_gen_binary (GET_CODE (varop), GET_MODE (varop),
9361 simplify_and_const_int (NULL_RTX,
9362 GET_MODE (varop),
9363 XEXP (varop, 0),
9364 constop),
9365 simplify_and_const_int (NULL_RTX,
9366 GET_MODE (varop),
9367 XEXP (varop, 1),
9368 constop))));
9370 /* If VAROP is PLUS, and the constant is a mask of low bits, distribute
9371 the AND and see if one of the operands simplifies to zero. If so, we
9372 may eliminate it. */
9374 if (GET_CODE (varop) == PLUS
9375 && exact_log2 (constop + 1) >= 0)
9377 rtx o0, o1;
9379 o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop);
9380 o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop);
9381 if (o0 == const0_rtx)
9382 return o1;
9383 if (o1 == const0_rtx)
9384 return o0;
9387 /* Make a SUBREG if necessary. If we can't make it, fail. */
9388 varop = gen_lowpart (mode, varop);
9389 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
9390 return NULL_RTX;
9392 /* If we are only masking insignificant bits, return VAROP. */
9393 if (constop == nonzero)
9394 return varop;
9396 if (varop == orig_varop && constop == orig_constop)
9397 return NULL_RTX;
9399 /* Otherwise, return an AND. */
9400 return simplify_gen_binary (AND, mode, varop, gen_int_mode (constop, mode));
9404 /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done
9405 in MODE.
9407 Return an equivalent form, if different from X. Otherwise, return X. If
9408 X is zero, we are to always construct the equivalent form. */
9410 static rtx
9411 simplify_and_const_int (rtx x, enum machine_mode mode, rtx varop,
9412 unsigned HOST_WIDE_INT constop)
9414 rtx tem = simplify_and_const_int_1 (mode, varop, constop);
9415 if (tem)
9416 return tem;
9418 if (!x)
9419 x = simplify_gen_binary (AND, GET_MODE (varop), varop,
9420 gen_int_mode (constop, mode));
9421 if (GET_MODE (x) != mode)
9422 x = gen_lowpart (mode, x);
9423 return x;
9426 /* Given a REG, X, compute which bits in X can be nonzero.
9427 We don't care about bits outside of those defined in MODE.
9429 For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is
9430 a shift, AND, or zero_extract, we can do better. */
9432 static rtx
9433 reg_nonzero_bits_for_combine (const_rtx x, enum machine_mode mode,
9434 const_rtx known_x ATTRIBUTE_UNUSED,
9435 enum machine_mode known_mode ATTRIBUTE_UNUSED,
9436 unsigned HOST_WIDE_INT known_ret ATTRIBUTE_UNUSED,
9437 unsigned HOST_WIDE_INT *nonzero)
9439 rtx tem;
9440 reg_stat_type *rsp;
9442 /* If X is a register whose nonzero bits value is current, use it.
9443 Otherwise, if X is a register whose value we can find, use that
9444 value. Otherwise, use the previously-computed global nonzero bits
9445 for this register. */
9447 rsp = &reg_stat[REGNO (x)];
9448 if (rsp->last_set_value != 0
9449 && (rsp->last_set_mode == mode
9450 || (GET_MODE_CLASS (rsp->last_set_mode) == MODE_INT
9451 && GET_MODE_CLASS (mode) == MODE_INT))
9452 && ((rsp->last_set_label >= label_tick_ebb_start
9453 && rsp->last_set_label < label_tick)
9454 || (rsp->last_set_label == label_tick
9455 && DF_INSN_LUID (rsp->last_set) < subst_low_luid)
9456 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
9457 && REG_N_SETS (REGNO (x)) == 1
9458 && !REGNO_REG_SET_P
9459 (DF_LR_IN (ENTRY_BLOCK_PTR->next_bb), REGNO (x)))))
9461 *nonzero &= rsp->last_set_nonzero_bits;
9462 return NULL;
9465 tem = get_last_value (x);
9467 if (tem)
9469 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
9470 /* If X is narrower than MODE and TEM is a non-negative
9471 constant that would appear negative in the mode of X,
9472 sign-extend it for use in reg_nonzero_bits because some
9473 machines (maybe most) will actually do the sign-extension
9474 and this is the conservative approach.
9476 ??? For 2.5, try to tighten up the MD files in this regard
9477 instead of this kludge. */
9479 if (GET_MODE_PRECISION (GET_MODE (x)) < GET_MODE_PRECISION (mode)
9480 && CONST_INT_P (tem)
9481 && INTVAL (tem) > 0
9482 && val_signbit_known_set_p (GET_MODE (x), INTVAL (tem)))
9483 tem = GEN_INT (INTVAL (tem) | ~GET_MODE_MASK (GET_MODE (x)));
9484 #endif
9485 return tem;
9487 else if (nonzero_sign_valid && rsp->nonzero_bits)
9489 unsigned HOST_WIDE_INT mask = rsp->nonzero_bits;
9491 if (GET_MODE_PRECISION (GET_MODE (x)) < GET_MODE_PRECISION (mode))
9492 /* We don't know anything about the upper bits. */
9493 mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (GET_MODE (x));
9494 *nonzero &= mask;
9497 return NULL;
9500 /* Return the number of bits at the high-order end of X that are known to
9501 be equal to the sign bit. X will be used in mode MODE; if MODE is
9502 VOIDmode, X will be used in its own mode. The returned value will always
9503 be between 1 and the number of bits in MODE. */
9505 static rtx
9506 reg_num_sign_bit_copies_for_combine (const_rtx x, enum machine_mode mode,
9507 const_rtx known_x ATTRIBUTE_UNUSED,
9508 enum machine_mode known_mode
9509 ATTRIBUTE_UNUSED,
9510 unsigned int known_ret ATTRIBUTE_UNUSED,
9511 unsigned int *result)
9513 rtx tem;
9514 reg_stat_type *rsp;
9516 rsp = &reg_stat[REGNO (x)];
9517 if (rsp->last_set_value != 0
9518 && rsp->last_set_mode == mode
9519 && ((rsp->last_set_label >= label_tick_ebb_start
9520 && rsp->last_set_label < label_tick)
9521 || (rsp->last_set_label == label_tick
9522 && DF_INSN_LUID (rsp->last_set) < subst_low_luid)
9523 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
9524 && REG_N_SETS (REGNO (x)) == 1
9525 && !REGNO_REG_SET_P
9526 (DF_LR_IN (ENTRY_BLOCK_PTR->next_bb), REGNO (x)))))
9528 *result = rsp->last_set_sign_bit_copies;
9529 return NULL;
9532 tem = get_last_value (x);
9533 if (tem != 0)
9534 return tem;
9536 if (nonzero_sign_valid && rsp->sign_bit_copies != 0
9537 && GET_MODE_PRECISION (GET_MODE (x)) == GET_MODE_PRECISION (mode))
9538 *result = rsp->sign_bit_copies;
9540 return NULL;
9543 /* Return the number of "extended" bits there are in X, when interpreted
9544 as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For
9545 unsigned quantities, this is the number of high-order zero bits.
9546 For signed quantities, this is the number of copies of the sign bit
9547 minus 1. In both case, this function returns the number of "spare"
9548 bits. For example, if two quantities for which this function returns
9549 at least 1 are added, the addition is known not to overflow.
9551 This function will always return 0 unless called during combine, which
9552 implies that it must be called from a define_split. */
9554 unsigned int
9555 extended_count (const_rtx x, enum machine_mode mode, int unsignedp)
9557 if (nonzero_sign_valid == 0)
9558 return 0;
9560 return (unsignedp
9561 ? (HWI_COMPUTABLE_MODE_P (mode)
9562 ? (unsigned int) (GET_MODE_PRECISION (mode) - 1
9563 - floor_log2 (nonzero_bits (x, mode)))
9564 : 0)
9565 : num_sign_bit_copies (x, mode) - 1);
9568 /* This function is called from `simplify_shift_const' to merge two
9569 outer operations. Specifically, we have already found that we need
9570 to perform operation *POP0 with constant *PCONST0 at the outermost
9571 position. We would now like to also perform OP1 with constant CONST1
9572 (with *POP0 being done last).
9574 Return 1 if we can do the operation and update *POP0 and *PCONST0 with
9575 the resulting operation. *PCOMP_P is set to 1 if we would need to
9576 complement the innermost operand, otherwise it is unchanged.
9578 MODE is the mode in which the operation will be done. No bits outside
9579 the width of this mode matter. It is assumed that the width of this mode
9580 is smaller than or equal to HOST_BITS_PER_WIDE_INT.
9582 If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS,
9583 IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper
9584 result is simply *PCONST0.
9586 If the resulting operation cannot be expressed as one operation, we
9587 return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */
9589 static int
9590 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)
9592 enum rtx_code op0 = *pop0;
9593 HOST_WIDE_INT const0 = *pconst0;
9595 const0 &= GET_MODE_MASK (mode);
9596 const1 &= GET_MODE_MASK (mode);
9598 /* If OP0 is an AND, clear unimportant bits in CONST1. */
9599 if (op0 == AND)
9600 const1 &= const0;
9602 /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or
9603 if OP0 is SET. */
9605 if (op1 == UNKNOWN || op0 == SET)
9606 return 1;
9608 else if (op0 == UNKNOWN)
9609 op0 = op1, const0 = const1;
9611 else if (op0 == op1)
9613 switch (op0)
9615 case AND:
9616 const0 &= const1;
9617 break;
9618 case IOR:
9619 const0 |= const1;
9620 break;
9621 case XOR:
9622 const0 ^= const1;
9623 break;
9624 case PLUS:
9625 const0 += const1;
9626 break;
9627 case NEG:
9628 op0 = UNKNOWN;
9629 break;
9630 default:
9631 break;
9635 /* Otherwise, if either is a PLUS or NEG, we can't do anything. */
9636 else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG)
9637 return 0;
9639 /* If the two constants aren't the same, we can't do anything. The
9640 remaining six cases can all be done. */
9641 else if (const0 != const1)
9642 return 0;
9644 else
9645 switch (op0)
9647 case IOR:
9648 if (op1 == AND)
9649 /* (a & b) | b == b */
9650 op0 = SET;
9651 else /* op1 == XOR */
9652 /* (a ^ b) | b == a | b */
9654 break;
9656 case XOR:
9657 if (op1 == AND)
9658 /* (a & b) ^ b == (~a) & b */
9659 op0 = AND, *pcomp_p = 1;
9660 else /* op1 == IOR */
9661 /* (a | b) ^ b == a & ~b */
9662 op0 = AND, const0 = ~const0;
9663 break;
9665 case AND:
9666 if (op1 == IOR)
9667 /* (a | b) & b == b */
9668 op0 = SET;
9669 else /* op1 == XOR */
9670 /* (a ^ b) & b) == (~a) & b */
9671 *pcomp_p = 1;
9672 break;
9673 default:
9674 break;
9677 /* Check for NO-OP cases. */
9678 const0 &= GET_MODE_MASK (mode);
9679 if (const0 == 0
9680 && (op0 == IOR || op0 == XOR || op0 == PLUS))
9681 op0 = UNKNOWN;
9682 else if (const0 == 0 && op0 == AND)
9683 op0 = SET;
9684 else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode)
9685 && op0 == AND)
9686 op0 = UNKNOWN;
9688 *pop0 = op0;
9690 /* ??? Slightly redundant with the above mask, but not entirely.
9691 Moving this above means we'd have to sign-extend the mode mask
9692 for the final test. */
9693 if (op0 != UNKNOWN && op0 != NEG)
9694 *pconst0 = trunc_int_for_mode (const0, mode);
9696 return 1;
9699 /* A helper to simplify_shift_const_1 to determine the mode we can perform
9700 the shift in. The original shift operation CODE is performed on OP in
9701 ORIG_MODE. Return the wider mode MODE if we can perform the operation
9702 in that mode. Return ORIG_MODE otherwise. We can also assume that the
9703 result of the shift is subject to operation OUTER_CODE with operand
9704 OUTER_CONST. */
9706 static enum machine_mode
9707 try_widen_shift_mode (enum rtx_code code, rtx op, int count,
9708 enum machine_mode orig_mode, enum machine_mode mode,
9709 enum rtx_code outer_code, HOST_WIDE_INT outer_const)
9711 if (orig_mode == mode)
9712 return mode;
9713 gcc_assert (GET_MODE_PRECISION (mode) > GET_MODE_PRECISION (orig_mode));
9715 /* In general we can't perform in wider mode for right shift and rotate. */
9716 switch (code)
9718 case ASHIFTRT:
9719 /* We can still widen if the bits brought in from the left are identical
9720 to the sign bit of ORIG_MODE. */
9721 if (num_sign_bit_copies (op, mode)
9722 > (unsigned) (GET_MODE_PRECISION (mode)
9723 - GET_MODE_PRECISION (orig_mode)))
9724 return mode;
9725 return orig_mode;
9727 case LSHIFTRT:
9728 /* Similarly here but with zero bits. */
9729 if (HWI_COMPUTABLE_MODE_P (mode)
9730 && (nonzero_bits (op, mode) & ~GET_MODE_MASK (orig_mode)) == 0)
9731 return mode;
9733 /* We can also widen if the bits brought in will be masked off. This
9734 operation is performed in ORIG_MODE. */
9735 if (outer_code == AND)
9737 int care_bits = low_bitmask_len (orig_mode, outer_const);
9739 if (care_bits >= 0
9740 && GET_MODE_PRECISION (orig_mode) - care_bits >= count)
9741 return mode;
9743 /* fall through */
9745 case ROTATE:
9746 return orig_mode;
9748 case ROTATERT:
9749 gcc_unreachable ();
9751 default:
9752 return mode;
9756 /* Simplify a shift of VAROP by ORIG_COUNT bits. CODE says what kind
9757 of shift. The result of the shift is RESULT_MODE. Return NULL_RTX
9758 if we cannot simplify it. Otherwise, return a simplified value.
9760 The shift is normally computed in the widest mode we find in VAROP, as
9761 long as it isn't a different number of words than RESULT_MODE. Exceptions
9762 are ASHIFTRT and ROTATE, which are always done in their original mode. */
9764 static rtx
9765 simplify_shift_const_1 (enum rtx_code code, enum machine_mode result_mode,
9766 rtx varop, int orig_count)
9768 enum rtx_code orig_code = code;
9769 rtx orig_varop = varop;
9770 int count;
9771 enum machine_mode mode = result_mode;
9772 enum machine_mode shift_mode, tmode;
9773 unsigned int mode_words
9774 = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
9775 /* We form (outer_op (code varop count) (outer_const)). */
9776 enum rtx_code outer_op = UNKNOWN;
9777 HOST_WIDE_INT outer_const = 0;
9778 int complement_p = 0;
9779 rtx new_rtx, x;
9781 /* Make sure and truncate the "natural" shift on the way in. We don't
9782 want to do this inside the loop as it makes it more difficult to
9783 combine shifts. */
9784 if (SHIFT_COUNT_TRUNCATED)
9785 orig_count &= GET_MODE_BITSIZE (mode) - 1;
9787 /* If we were given an invalid count, don't do anything except exactly
9788 what was requested. */
9790 if (orig_count < 0 || orig_count >= (int) GET_MODE_PRECISION (mode))
9791 return NULL_RTX;
9793 count = orig_count;
9795 /* Unless one of the branches of the `if' in this loop does a `continue',
9796 we will `break' the loop after the `if'. */
9798 while (count != 0)
9800 /* If we have an operand of (clobber (const_int 0)), fail. */
9801 if (GET_CODE (varop) == CLOBBER)
9802 return NULL_RTX;
9804 /* Convert ROTATERT to ROTATE. */
9805 if (code == ROTATERT)
9807 unsigned int bitsize = GET_MODE_PRECISION (result_mode);
9808 code = ROTATE;
9809 if (VECTOR_MODE_P (result_mode))
9810 count = bitsize / GET_MODE_NUNITS (result_mode) - count;
9811 else
9812 count = bitsize - count;
9815 shift_mode = try_widen_shift_mode (code, varop, count, result_mode,
9816 mode, outer_op, outer_const);
9818 /* Handle cases where the count is greater than the size of the mode
9819 minus 1. For ASHIFT, use the size minus one as the count (this can
9820 occur when simplifying (lshiftrt (ashiftrt ..))). For rotates,
9821 take the count modulo the size. For other shifts, the result is
9822 zero.
9824 Since these shifts are being produced by the compiler by combining
9825 multiple operations, each of which are defined, we know what the
9826 result is supposed to be. */
9828 if (count > (GET_MODE_PRECISION (shift_mode) - 1))
9830 if (code == ASHIFTRT)
9831 count = GET_MODE_PRECISION (shift_mode) - 1;
9832 else if (code == ROTATE || code == ROTATERT)
9833 count %= GET_MODE_PRECISION (shift_mode);
9834 else
9836 /* We can't simply return zero because there may be an
9837 outer op. */
9838 varop = const0_rtx;
9839 count = 0;
9840 break;
9844 /* If we discovered we had to complement VAROP, leave. Making a NOT
9845 here would cause an infinite loop. */
9846 if (complement_p)
9847 break;
9849 /* An arithmetic right shift of a quantity known to be -1 or 0
9850 is a no-op. */
9851 if (code == ASHIFTRT
9852 && (num_sign_bit_copies (varop, shift_mode)
9853 == GET_MODE_PRECISION (shift_mode)))
9855 count = 0;
9856 break;
9859 /* If we are doing an arithmetic right shift and discarding all but
9860 the sign bit copies, this is equivalent to doing a shift by the
9861 bitsize minus one. Convert it into that shift because it will often
9862 allow other simplifications. */
9864 if (code == ASHIFTRT
9865 && (count + num_sign_bit_copies (varop, shift_mode)
9866 >= GET_MODE_PRECISION (shift_mode)))
9867 count = GET_MODE_PRECISION (shift_mode) - 1;
9869 /* We simplify the tests below and elsewhere by converting
9870 ASHIFTRT to LSHIFTRT if we know the sign bit is clear.
9871 `make_compound_operation' will convert it to an ASHIFTRT for
9872 those machines (such as VAX) that don't have an LSHIFTRT. */
9873 if (code == ASHIFTRT
9874 && val_signbit_known_clear_p (shift_mode,
9875 nonzero_bits (varop, shift_mode)))
9876 code = LSHIFTRT;
9878 if (((code == LSHIFTRT
9879 && HWI_COMPUTABLE_MODE_P (shift_mode)
9880 && !(nonzero_bits (varop, shift_mode) >> count))
9881 || (code == ASHIFT
9882 && HWI_COMPUTABLE_MODE_P (shift_mode)
9883 && !((nonzero_bits (varop, shift_mode) << count)
9884 & GET_MODE_MASK (shift_mode))))
9885 && !side_effects_p (varop))
9886 varop = const0_rtx;
9888 switch (GET_CODE (varop))
9890 case SIGN_EXTEND:
9891 case ZERO_EXTEND:
9892 case SIGN_EXTRACT:
9893 case ZERO_EXTRACT:
9894 new_rtx = expand_compound_operation (varop);
9895 if (new_rtx != varop)
9897 varop = new_rtx;
9898 continue;
9900 break;
9902 case MEM:
9903 /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH
9904 minus the width of a smaller mode, we can do this with a
9905 SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */
9906 if ((code == ASHIFTRT || code == LSHIFTRT)
9907 && ! mode_dependent_address_p (XEXP (varop, 0),
9908 MEM_ADDR_SPACE (varop))
9909 && ! MEM_VOLATILE_P (varop)
9910 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
9911 MODE_INT, 1)) != BLKmode)
9913 new_rtx = adjust_address_nv (varop, tmode,
9914 BYTES_BIG_ENDIAN ? 0
9915 : count / BITS_PER_UNIT);
9917 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
9918 : ZERO_EXTEND, mode, new_rtx);
9919 count = 0;
9920 continue;
9922 break;
9924 case SUBREG:
9925 /* If VAROP is a SUBREG, strip it as long as the inner operand has
9926 the same number of words as what we've seen so far. Then store
9927 the widest mode in MODE. */
9928 if (subreg_lowpart_p (varop)
9929 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
9930 > GET_MODE_SIZE (GET_MODE (varop)))
9931 && (unsigned int) ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
9932 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
9933 == mode_words
9934 && GET_MODE_CLASS (GET_MODE (varop)) == MODE_INT
9935 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (varop))) == MODE_INT)
9937 varop = SUBREG_REG (varop);
9938 if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode))
9939 mode = GET_MODE (varop);
9940 continue;
9942 break;
9944 case MULT:
9945 /* Some machines use MULT instead of ASHIFT because MULT
9946 is cheaper. But it is still better on those machines to
9947 merge two shifts into one. */
9948 if (CONST_INT_P (XEXP (varop, 1))
9949 && exact_log2 (UINTVAL (XEXP (varop, 1))) >= 0)
9951 varop
9952 = simplify_gen_binary (ASHIFT, GET_MODE (varop),
9953 XEXP (varop, 0),
9954 GEN_INT (exact_log2 (
9955 UINTVAL (XEXP (varop, 1)))));
9956 continue;
9958 break;
9960 case UDIV:
9961 /* Similar, for when divides are cheaper. */
9962 if (CONST_INT_P (XEXP (varop, 1))
9963 && exact_log2 (UINTVAL (XEXP (varop, 1))) >= 0)
9965 varop
9966 = simplify_gen_binary (LSHIFTRT, GET_MODE (varop),
9967 XEXP (varop, 0),
9968 GEN_INT (exact_log2 (
9969 UINTVAL (XEXP (varop, 1)))));
9970 continue;
9972 break;
9974 case ASHIFTRT:
9975 /* If we are extracting just the sign bit of an arithmetic
9976 right shift, that shift is not needed. However, the sign
9977 bit of a wider mode may be different from what would be
9978 interpreted as the sign bit in a narrower mode, so, if
9979 the result is narrower, don't discard the shift. */
9980 if (code == LSHIFTRT
9981 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9982 && (GET_MODE_BITSIZE (result_mode)
9983 >= GET_MODE_BITSIZE (GET_MODE (varop))))
9985 varop = XEXP (varop, 0);
9986 continue;
9989 /* ... fall through ... */
9991 case LSHIFTRT:
9992 case ASHIFT:
9993 case ROTATE:
9994 /* Here we have two nested shifts. The result is usually the
9995 AND of a new shift with a mask. We compute the result below. */
9996 if (CONST_INT_P (XEXP (varop, 1))
9997 && INTVAL (XEXP (varop, 1)) >= 0
9998 && INTVAL (XEXP (varop, 1)) < GET_MODE_PRECISION (GET_MODE (varop))
9999 && HWI_COMPUTABLE_MODE_P (result_mode)
10000 && HWI_COMPUTABLE_MODE_P (mode)
10001 && !VECTOR_MODE_P (result_mode))
10003 enum rtx_code first_code = GET_CODE (varop);
10004 unsigned int first_count = INTVAL (XEXP (varop, 1));
10005 unsigned HOST_WIDE_INT mask;
10006 rtx mask_rtx;
10008 /* We have one common special case. We can't do any merging if
10009 the inner code is an ASHIFTRT of a smaller mode. However, if
10010 we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2)
10011 with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2),
10012 we can convert it to
10013 (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0) C3) C2) C1).
10014 This simplifies certain SIGN_EXTEND operations. */
10015 if (code == ASHIFT && first_code == ASHIFTRT
10016 && count == (GET_MODE_PRECISION (result_mode)
10017 - GET_MODE_PRECISION (GET_MODE (varop))))
10019 /* C3 has the low-order C1 bits zero. */
10021 mask = GET_MODE_MASK (mode)
10022 & ~(((unsigned HOST_WIDE_INT) 1 << first_count) - 1);
10024 varop = simplify_and_const_int (NULL_RTX, result_mode,
10025 XEXP (varop, 0), mask);
10026 varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode,
10027 varop, count);
10028 count = first_count;
10029 code = ASHIFTRT;
10030 continue;
10033 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more
10034 than C1 high-order bits equal to the sign bit, we can convert
10035 this to either an ASHIFT or an ASHIFTRT depending on the
10036 two counts.
10038 We cannot do this if VAROP's mode is not SHIFT_MODE. */
10040 if (code == ASHIFTRT && first_code == ASHIFT
10041 && GET_MODE (varop) == shift_mode
10042 && (num_sign_bit_copies (XEXP (varop, 0), shift_mode)
10043 > first_count))
10045 varop = XEXP (varop, 0);
10046 count -= first_count;
10047 if (count < 0)
10049 count = -count;
10050 code = ASHIFT;
10053 continue;
10056 /* There are some cases we can't do. If CODE is ASHIFTRT,
10057 we can only do this if FIRST_CODE is also ASHIFTRT.
10059 We can't do the case when CODE is ROTATE and FIRST_CODE is
10060 ASHIFTRT.
10062 If the mode of this shift is not the mode of the outer shift,
10063 we can't do this if either shift is a right shift or ROTATE.
10065 Finally, we can't do any of these if the mode is too wide
10066 unless the codes are the same.
10068 Handle the case where the shift codes are the same
10069 first. */
10071 if (code == first_code)
10073 if (GET_MODE (varop) != result_mode
10074 && (code == ASHIFTRT || code == LSHIFTRT
10075 || code == ROTATE))
10076 break;
10078 count += first_count;
10079 varop = XEXP (varop, 0);
10080 continue;
10083 if (code == ASHIFTRT
10084 || (code == ROTATE && first_code == ASHIFTRT)
10085 || GET_MODE_PRECISION (mode) > HOST_BITS_PER_WIDE_INT
10086 || (GET_MODE (varop) != result_mode
10087 && (first_code == ASHIFTRT || first_code == LSHIFTRT
10088 || first_code == ROTATE
10089 || code == ROTATE)))
10090 break;
10092 /* To compute the mask to apply after the shift, shift the
10093 nonzero bits of the inner shift the same way the
10094 outer shift will. */
10096 mask_rtx = gen_int_mode (nonzero_bits (varop, GET_MODE (varop)),
10097 result_mode);
10099 mask_rtx
10100 = simplify_const_binary_operation (code, result_mode, mask_rtx,
10101 GEN_INT (count));
10103 /* Give up if we can't compute an outer operation to use. */
10104 if (mask_rtx == 0
10105 || !CONST_INT_P (mask_rtx)
10106 || ! merge_outer_ops (&outer_op, &outer_const, AND,
10107 INTVAL (mask_rtx),
10108 result_mode, &complement_p))
10109 break;
10111 /* If the shifts are in the same direction, we add the
10112 counts. Otherwise, we subtract them. */
10113 if ((code == ASHIFTRT || code == LSHIFTRT)
10114 == (first_code == ASHIFTRT || first_code == LSHIFTRT))
10115 count += first_count;
10116 else
10117 count -= first_count;
10119 /* If COUNT is positive, the new shift is usually CODE,
10120 except for the two exceptions below, in which case it is
10121 FIRST_CODE. If the count is negative, FIRST_CODE should
10122 always be used */
10123 if (count > 0
10124 && ((first_code == ROTATE && code == ASHIFT)
10125 || (first_code == ASHIFTRT && code == LSHIFTRT)))
10126 code = first_code;
10127 else if (count < 0)
10128 code = first_code, count = -count;
10130 varop = XEXP (varop, 0);
10131 continue;
10134 /* If we have (A << B << C) for any shift, we can convert this to
10135 (A << C << B). This wins if A is a constant. Only try this if
10136 B is not a constant. */
10138 else if (GET_CODE (varop) == code
10139 && CONST_INT_P (XEXP (varop, 0))
10140 && !CONST_INT_P (XEXP (varop, 1)))
10142 rtx new_rtx = simplify_const_binary_operation (code, mode,
10143 XEXP (varop, 0),
10144 GEN_INT (count));
10145 varop = gen_rtx_fmt_ee (code, mode, new_rtx, XEXP (varop, 1));
10146 count = 0;
10147 continue;
10149 break;
10151 case NOT:
10152 if (VECTOR_MODE_P (mode))
10153 break;
10155 /* Make this fit the case below. */
10156 varop = gen_rtx_XOR (mode, XEXP (varop, 0), constm1_rtx);
10157 continue;
10159 case IOR:
10160 case AND:
10161 case XOR:
10162 /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C)
10163 with C the size of VAROP - 1 and the shift is logical if
10164 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
10165 we have an (le X 0) operation. If we have an arithmetic shift
10166 and STORE_FLAG_VALUE is 1 or we have a logical shift with
10167 STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */
10169 if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS
10170 && XEXP (XEXP (varop, 0), 1) == constm1_rtx
10171 && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
10172 && (code == LSHIFTRT || code == ASHIFTRT)
10173 && count == (GET_MODE_PRECISION (GET_MODE (varop)) - 1)
10174 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
10176 count = 0;
10177 varop = gen_rtx_LE (GET_MODE (varop), XEXP (varop, 1),
10178 const0_rtx);
10180 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
10181 varop = gen_rtx_NEG (GET_MODE (varop), varop);
10183 continue;
10186 /* If we have (shift (logical)), move the logical to the outside
10187 to allow it to possibly combine with another logical and the
10188 shift to combine with another shift. This also canonicalizes to
10189 what a ZERO_EXTRACT looks like. Also, some machines have
10190 (and (shift)) insns. */
10192 if (CONST_INT_P (XEXP (varop, 1))
10193 /* We can't do this if we have (ashiftrt (xor)) and the
10194 constant has its sign bit set in shift_mode. */
10195 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
10196 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
10197 shift_mode))
10198 && (new_rtx = simplify_const_binary_operation
10199 (code, result_mode,
10200 gen_int_mode (INTVAL (XEXP (varop, 1)), result_mode),
10201 GEN_INT (count))) != 0
10202 && CONST_INT_P (new_rtx)
10203 && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop),
10204 INTVAL (new_rtx), result_mode, &complement_p))
10206 varop = XEXP (varop, 0);
10207 continue;
10210 /* If we can't do that, try to simplify the shift in each arm of the
10211 logical expression, make a new logical expression, and apply
10212 the inverse distributive law. This also can't be done
10213 for some (ashiftrt (xor)). */
10214 if (CONST_INT_P (XEXP (varop, 1))
10215 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
10216 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
10217 shift_mode)))
10219 rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode,
10220 XEXP (varop, 0), count);
10221 rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode,
10222 XEXP (varop, 1), count);
10224 varop = simplify_gen_binary (GET_CODE (varop), shift_mode,
10225 lhs, rhs);
10226 varop = apply_distributive_law (varop);
10228 count = 0;
10229 continue;
10231 break;
10233 case EQ:
10234 /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE
10235 says that the sign bit can be tested, FOO has mode MODE, C is
10236 GET_MODE_PRECISION (MODE) - 1, and FOO has only its low-order bit
10237 that may be nonzero. */
10238 if (code == LSHIFTRT
10239 && XEXP (varop, 1) == const0_rtx
10240 && GET_MODE (XEXP (varop, 0)) == result_mode
10241 && count == (GET_MODE_PRECISION (result_mode) - 1)
10242 && HWI_COMPUTABLE_MODE_P (result_mode)
10243 && STORE_FLAG_VALUE == -1
10244 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
10245 && merge_outer_ops (&outer_op, &outer_const, XOR, 1, result_mode,
10246 &complement_p))
10248 varop = XEXP (varop, 0);
10249 count = 0;
10250 continue;
10252 break;
10254 case NEG:
10255 /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less
10256 than the number of bits in the mode is equivalent to A. */
10257 if (code == LSHIFTRT
10258 && count == (GET_MODE_PRECISION (result_mode) - 1)
10259 && nonzero_bits (XEXP (varop, 0), result_mode) == 1)
10261 varop = XEXP (varop, 0);
10262 count = 0;
10263 continue;
10266 /* NEG commutes with ASHIFT since it is multiplication. Move the
10267 NEG outside to allow shifts to combine. */
10268 if (code == ASHIFT
10269 && merge_outer_ops (&outer_op, &outer_const, NEG, 0, result_mode,
10270 &complement_p))
10272 varop = XEXP (varop, 0);
10273 continue;
10275 break;
10277 case PLUS:
10278 /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C
10279 is one less than the number of bits in the mode is
10280 equivalent to (xor A 1). */
10281 if (code == LSHIFTRT
10282 && count == (GET_MODE_PRECISION (result_mode) - 1)
10283 && XEXP (varop, 1) == constm1_rtx
10284 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
10285 && merge_outer_ops (&outer_op, &outer_const, XOR, 1, result_mode,
10286 &complement_p))
10288 count = 0;
10289 varop = XEXP (varop, 0);
10290 continue;
10293 /* If we have (xshiftrt (plus FOO BAR) C), and the only bits
10294 that might be nonzero in BAR are those being shifted out and those
10295 bits are known zero in FOO, we can replace the PLUS with FOO.
10296 Similarly in the other operand order. This code occurs when
10297 we are computing the size of a variable-size array. */
10299 if ((code == ASHIFTRT || code == LSHIFTRT)
10300 && count < HOST_BITS_PER_WIDE_INT
10301 && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0
10302 && (nonzero_bits (XEXP (varop, 1), result_mode)
10303 & nonzero_bits (XEXP (varop, 0), result_mode)) == 0)
10305 varop = XEXP (varop, 0);
10306 continue;
10308 else if ((code == ASHIFTRT || code == LSHIFTRT)
10309 && count < HOST_BITS_PER_WIDE_INT
10310 && HWI_COMPUTABLE_MODE_P (result_mode)
10311 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
10312 >> count)
10313 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
10314 & nonzero_bits (XEXP (varop, 1),
10315 result_mode)))
10317 varop = XEXP (varop, 1);
10318 continue;
10321 /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */
10322 if (code == ASHIFT
10323 && CONST_INT_P (XEXP (varop, 1))
10324 && (new_rtx = simplify_const_binary_operation (ASHIFT, result_mode,
10325 XEXP (varop, 1),
10326 GEN_INT (count))) != 0
10327 && CONST_INT_P (new_rtx)
10328 && merge_outer_ops (&outer_op, &outer_const, PLUS,
10329 INTVAL (new_rtx), result_mode, &complement_p))
10331 varop = XEXP (varop, 0);
10332 continue;
10335 /* Check for 'PLUS signbit', which is the canonical form of 'XOR
10336 signbit', and attempt to change the PLUS to an XOR and move it to
10337 the outer operation as is done above in the AND/IOR/XOR case
10338 leg for shift(logical). See details in logical handling above
10339 for reasoning in doing so. */
10340 if (code == LSHIFTRT
10341 && CONST_INT_P (XEXP (varop, 1))
10342 && mode_signbit_p (result_mode, XEXP (varop, 1))
10343 && (new_rtx = simplify_const_binary_operation (code, result_mode,
10344 XEXP (varop, 1),
10345 GEN_INT (count))) != 0
10346 && CONST_INT_P (new_rtx)
10347 && merge_outer_ops (&outer_op, &outer_const, XOR,
10348 INTVAL (new_rtx), result_mode, &complement_p))
10350 varop = XEXP (varop, 0);
10351 continue;
10354 break;
10356 case MINUS:
10357 /* If we have (xshiftrt (minus (ashiftrt X C)) X) C)
10358 with C the size of VAROP - 1 and the shift is logical if
10359 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
10360 we have a (gt X 0) operation. If the shift is arithmetic with
10361 STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1,
10362 we have a (neg (gt X 0)) operation. */
10364 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
10365 && GET_CODE (XEXP (varop, 0)) == ASHIFTRT
10366 && count == (GET_MODE_PRECISION (GET_MODE (varop)) - 1)
10367 && (code == LSHIFTRT || code == ASHIFTRT)
10368 && CONST_INT_P (XEXP (XEXP (varop, 0), 1))
10369 && INTVAL (XEXP (XEXP (varop, 0), 1)) == count
10370 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
10372 count = 0;
10373 varop = gen_rtx_GT (GET_MODE (varop), XEXP (varop, 1),
10374 const0_rtx);
10376 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
10377 varop = gen_rtx_NEG (GET_MODE (varop), varop);
10379 continue;
10381 break;
10383 case TRUNCATE:
10384 /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt))
10385 if the truncate does not affect the value. */
10386 if (code == LSHIFTRT
10387 && GET_CODE (XEXP (varop, 0)) == LSHIFTRT
10388 && CONST_INT_P (XEXP (XEXP (varop, 0), 1))
10389 && (INTVAL (XEXP (XEXP (varop, 0), 1))
10390 >= (GET_MODE_PRECISION (GET_MODE (XEXP (varop, 0)))
10391 - GET_MODE_PRECISION (GET_MODE (varop)))))
10393 rtx varop_inner = XEXP (varop, 0);
10395 varop_inner
10396 = gen_rtx_LSHIFTRT (GET_MODE (varop_inner),
10397 XEXP (varop_inner, 0),
10398 GEN_INT
10399 (count + INTVAL (XEXP (varop_inner, 1))));
10400 varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner);
10401 count = 0;
10402 continue;
10404 break;
10406 default:
10407 break;
10410 break;
10413 shift_mode = try_widen_shift_mode (code, varop, count, result_mode, mode,
10414 outer_op, outer_const);
10416 /* We have now finished analyzing the shift. The result should be
10417 a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If
10418 OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied
10419 to the result of the shift. OUTER_CONST is the relevant constant,
10420 but we must turn off all bits turned off in the shift. */
10422 if (outer_op == UNKNOWN
10423 && orig_code == code && orig_count == count
10424 && varop == orig_varop
10425 && shift_mode == GET_MODE (varop))
10426 return NULL_RTX;
10428 /* Make a SUBREG if necessary. If we can't make it, fail. */
10429 varop = gen_lowpart (shift_mode, varop);
10430 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
10431 return NULL_RTX;
10433 /* If we have an outer operation and we just made a shift, it is
10434 possible that we could have simplified the shift were it not
10435 for the outer operation. So try to do the simplification
10436 recursively. */
10438 if (outer_op != UNKNOWN)
10439 x = simplify_shift_const_1 (code, shift_mode, varop, count);
10440 else
10441 x = NULL_RTX;
10443 if (x == NULL_RTX)
10444 x = simplify_gen_binary (code, shift_mode, varop, GEN_INT (count));
10446 /* If we were doing an LSHIFTRT in a wider mode than it was originally,
10447 turn off all the bits that the shift would have turned off. */
10448 if (orig_code == LSHIFTRT && result_mode != shift_mode)
10449 x = simplify_and_const_int (NULL_RTX, shift_mode, x,
10450 GET_MODE_MASK (result_mode) >> orig_count);
10452 /* Do the remainder of the processing in RESULT_MODE. */
10453 x = gen_lowpart_or_truncate (result_mode, x);
10455 /* If COMPLEMENT_P is set, we have to complement X before doing the outer
10456 operation. */
10457 if (complement_p)
10458 x = simplify_gen_unary (NOT, result_mode, x, result_mode);
10460 if (outer_op != UNKNOWN)
10462 if (GET_RTX_CLASS (outer_op) != RTX_UNARY
10463 && GET_MODE_PRECISION (result_mode) < HOST_BITS_PER_WIDE_INT)
10464 outer_const = trunc_int_for_mode (outer_const, result_mode);
10466 if (outer_op == AND)
10467 x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const);
10468 else if (outer_op == SET)
10470 /* This means that we have determined that the result is
10471 equivalent to a constant. This should be rare. */
10472 if (!side_effects_p (x))
10473 x = GEN_INT (outer_const);
10475 else if (GET_RTX_CLASS (outer_op) == RTX_UNARY)
10476 x = simplify_gen_unary (outer_op, result_mode, x, result_mode);
10477 else
10478 x = simplify_gen_binary (outer_op, result_mode, x,
10479 GEN_INT (outer_const));
10482 return x;
10485 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
10486 The result of the shift is RESULT_MODE. If we cannot simplify it,
10487 return X or, if it is NULL, synthesize the expression with
10488 simplify_gen_binary. Otherwise, return a simplified value.
10490 The shift is normally computed in the widest mode we find in VAROP, as
10491 long as it isn't a different number of words than RESULT_MODE. Exceptions
10492 are ASHIFTRT and ROTATE, which are always done in their original mode. */
10494 static rtx
10495 simplify_shift_const (rtx x, enum rtx_code code, enum machine_mode result_mode,
10496 rtx varop, int count)
10498 rtx tem = simplify_shift_const_1 (code, result_mode, varop, count);
10499 if (tem)
10500 return tem;
10502 if (!x)
10503 x = simplify_gen_binary (code, GET_MODE (varop), varop, GEN_INT (count));
10504 if (GET_MODE (x) != result_mode)
10505 x = gen_lowpart (result_mode, x);
10506 return x;
10510 /* Like recog, but we receive the address of a pointer to a new pattern.
10511 We try to match the rtx that the pointer points to.
10512 If that fails, we may try to modify or replace the pattern,
10513 storing the replacement into the same pointer object.
10515 Modifications include deletion or addition of CLOBBERs.
10517 PNOTES is a pointer to a location where any REG_UNUSED notes added for
10518 the CLOBBERs are placed.
10520 The value is the final insn code from the pattern ultimately matched,
10521 or -1. */
10523 static int
10524 recog_for_combine (rtx *pnewpat, rtx insn, rtx *pnotes)
10526 rtx pat = *pnewpat;
10527 rtx pat_without_clobbers;
10528 int insn_code_number;
10529 int num_clobbers_to_add = 0;
10530 int i;
10531 rtx notes = NULL_RTX;
10532 rtx old_notes, old_pat;
10533 int old_icode;
10535 /* If PAT is a PARALLEL, check to see if it contains the CLOBBER
10536 we use to indicate that something didn't match. If we find such a
10537 thing, force rejection. */
10538 if (GET_CODE (pat) == PARALLEL)
10539 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
10540 if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER
10541 && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx)
10542 return -1;
10544 old_pat = PATTERN (insn);
10545 old_notes = REG_NOTES (insn);
10546 PATTERN (insn) = pat;
10547 REG_NOTES (insn) = NULL_RTX;
10549 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
10550 if (dump_file && (dump_flags & TDF_DETAILS))
10552 if (insn_code_number < 0)
10553 fputs ("Failed to match this instruction:\n", dump_file);
10554 else
10555 fputs ("Successfully matched this instruction:\n", dump_file);
10556 print_rtl_single (dump_file, pat);
10559 /* If it isn't, there is the possibility that we previously had an insn
10560 that clobbered some register as a side effect, but the combined
10561 insn doesn't need to do that. So try once more without the clobbers
10562 unless this represents an ASM insn. */
10564 if (insn_code_number < 0 && ! check_asm_operands (pat)
10565 && GET_CODE (pat) == PARALLEL)
10567 int pos;
10569 for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++)
10570 if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER)
10572 if (i != pos)
10573 SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i));
10574 pos++;
10577 SUBST_INT (XVECLEN (pat, 0), pos);
10579 if (pos == 1)
10580 pat = XVECEXP (pat, 0, 0);
10582 PATTERN (insn) = pat;
10583 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
10584 if (dump_file && (dump_flags & TDF_DETAILS))
10586 if (insn_code_number < 0)
10587 fputs ("Failed to match this instruction:\n", dump_file);
10588 else
10589 fputs ("Successfully matched this instruction:\n", dump_file);
10590 print_rtl_single (dump_file, pat);
10594 pat_without_clobbers = pat;
10596 PATTERN (insn) = old_pat;
10597 REG_NOTES (insn) = old_notes;
10599 /* Recognize all noop sets, these will be killed by followup pass. */
10600 if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat))
10601 insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0;
10603 /* If we had any clobbers to add, make a new pattern than contains
10604 them. Then check to make sure that all of them are dead. */
10605 if (num_clobbers_to_add)
10607 rtx newpat = gen_rtx_PARALLEL (VOIDmode,
10608 rtvec_alloc (GET_CODE (pat) == PARALLEL
10609 ? (XVECLEN (pat, 0)
10610 + num_clobbers_to_add)
10611 : num_clobbers_to_add + 1));
10613 if (GET_CODE (pat) == PARALLEL)
10614 for (i = 0; i < XVECLEN (pat, 0); i++)
10615 XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i);
10616 else
10617 XVECEXP (newpat, 0, 0) = pat;
10619 add_clobbers (newpat, insn_code_number);
10621 for (i = XVECLEN (newpat, 0) - num_clobbers_to_add;
10622 i < XVECLEN (newpat, 0); i++)
10624 if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))
10625 && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn))
10626 return -1;
10627 if (GET_CODE (XEXP (XVECEXP (newpat, 0, i), 0)) != SCRATCH)
10629 gcc_assert (REG_P (XEXP (XVECEXP (newpat, 0, i), 0)));
10630 notes = alloc_reg_note (REG_UNUSED,
10631 XEXP (XVECEXP (newpat, 0, i), 0), notes);
10634 pat = newpat;
10637 if (insn_code_number >= 0
10638 && insn_code_number != NOOP_MOVE_INSN_CODE)
10640 old_pat = PATTERN (insn);
10641 old_notes = REG_NOTES (insn);
10642 old_icode = INSN_CODE (insn);
10643 PATTERN (insn) = pat;
10644 REG_NOTES (insn) = notes;
10646 /* Allow targets to reject combined insn. */
10647 if (!targetm.legitimate_combined_insn (insn))
10649 if (dump_file && (dump_flags & TDF_DETAILS))
10650 fputs ("Instruction not appropriate for target.",
10651 dump_file);
10653 /* Callers expect recog_for_combine to strip
10654 clobbers from the pattern on failure. */
10655 pat = pat_without_clobbers;
10656 notes = NULL_RTX;
10658 insn_code_number = -1;
10661 PATTERN (insn) = old_pat;
10662 REG_NOTES (insn) = old_notes;
10663 INSN_CODE (insn) = old_icode;
10666 *pnewpat = pat;
10667 *pnotes = notes;
10669 return insn_code_number;
10672 /* Like gen_lowpart_general but for use by combine. In combine it
10673 is not possible to create any new pseudoregs. However, it is
10674 safe to create invalid memory addresses, because combine will
10675 try to recognize them and all they will do is make the combine
10676 attempt fail.
10678 If for some reason this cannot do its job, an rtx
10679 (clobber (const_int 0)) is returned.
10680 An insn containing that will not be recognized. */
10682 static rtx
10683 gen_lowpart_for_combine (enum machine_mode omode, rtx x)
10685 enum machine_mode imode = GET_MODE (x);
10686 unsigned int osize = GET_MODE_SIZE (omode);
10687 unsigned int isize = GET_MODE_SIZE (imode);
10688 rtx result;
10690 if (omode == imode)
10691 return x;
10693 /* We can only support MODE being wider than a word if X is a
10694 constant integer or has a mode the same size. */
10695 if (GET_MODE_SIZE (omode) > UNITS_PER_WORD
10696 && ! (CONST_SCALAR_INT_P (x) || isize == osize))
10697 goto fail;
10699 /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart
10700 won't know what to do. So we will strip off the SUBREG here and
10701 process normally. */
10702 if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x)))
10704 x = SUBREG_REG (x);
10706 /* For use in case we fall down into the address adjustments
10707 further below, we need to adjust the known mode and size of
10708 x; imode and isize, since we just adjusted x. */
10709 imode = GET_MODE (x);
10711 if (imode == omode)
10712 return x;
10714 isize = GET_MODE_SIZE (imode);
10717 result = gen_lowpart_common (omode, x);
10719 if (result)
10720 return result;
10722 if (MEM_P (x))
10724 int offset = 0;
10726 /* Refuse to work on a volatile memory ref or one with a mode-dependent
10727 address. */
10728 if (MEM_VOLATILE_P (x)
10729 || mode_dependent_address_p (XEXP (x, 0), MEM_ADDR_SPACE (x)))
10730 goto fail;
10732 /* If we want to refer to something bigger than the original memref,
10733 generate a paradoxical subreg instead. That will force a reload
10734 of the original memref X. */
10735 if (isize < osize)
10736 return gen_rtx_SUBREG (omode, x, 0);
10738 if (WORDS_BIG_ENDIAN)
10739 offset = MAX (isize, UNITS_PER_WORD) - MAX (osize, UNITS_PER_WORD);
10741 /* Adjust the address so that the address-after-the-data is
10742 unchanged. */
10743 if (BYTES_BIG_ENDIAN)
10744 offset -= MIN (UNITS_PER_WORD, osize) - MIN (UNITS_PER_WORD, isize);
10746 return adjust_address_nv (x, omode, offset);
10749 /* If X is a comparison operator, rewrite it in a new mode. This
10750 probably won't match, but may allow further simplifications. */
10751 else if (COMPARISON_P (x))
10752 return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1));
10754 /* If we couldn't simplify X any other way, just enclose it in a
10755 SUBREG. Normally, this SUBREG won't match, but some patterns may
10756 include an explicit SUBREG or we may simplify it further in combine. */
10757 else
10759 int offset = 0;
10760 rtx res;
10762 offset = subreg_lowpart_offset (omode, imode);
10763 if (imode == VOIDmode)
10765 imode = int_mode_for_mode (omode);
10766 x = gen_lowpart_common (imode, x);
10767 if (x == NULL)
10768 goto fail;
10770 res = simplify_gen_subreg (omode, x, imode, offset);
10771 if (res)
10772 return res;
10775 fail:
10776 return gen_rtx_CLOBBER (omode, const0_rtx);
10779 /* Try to simplify a comparison between OP0 and a constant OP1,
10780 where CODE is the comparison code that will be tested, into a
10781 (CODE OP0 const0_rtx) form.
10783 The result is a possibly different comparison code to use.
10784 *POP1 may be updated. */
10786 static enum rtx_code
10787 simplify_compare_const (enum rtx_code code, rtx op0, rtx *pop1)
10789 enum machine_mode mode = GET_MODE (op0);
10790 unsigned int mode_width = GET_MODE_PRECISION (mode);
10791 HOST_WIDE_INT const_op = INTVAL (*pop1);
10793 /* Get the constant we are comparing against and turn off all bits
10794 not on in our mode. */
10795 if (mode != VOIDmode)
10796 const_op = trunc_int_for_mode (const_op, mode);
10798 /* If we are comparing against a constant power of two and the value
10799 being compared can only have that single bit nonzero (e.g., it was
10800 `and'ed with that bit), we can replace this with a comparison
10801 with zero. */
10802 if (const_op
10803 && (code == EQ || code == NE || code == GE || code == GEU
10804 || code == LT || code == LTU)
10805 && mode_width <= HOST_BITS_PER_WIDE_INT
10806 && exact_log2 (const_op & GET_MODE_MASK (mode)) >= 0
10807 && (nonzero_bits (op0, mode)
10808 == (unsigned HOST_WIDE_INT) (const_op & GET_MODE_MASK (mode))))
10810 code = (code == EQ || code == GE || code == GEU ? NE : EQ);
10811 const_op = 0;
10814 /* Similarly, if we are comparing a value known to be either -1 or
10815 0 with -1, change it to the opposite comparison against zero. */
10816 if (const_op == -1
10817 && (code == EQ || code == NE || code == GT || code == LE
10818 || code == GEU || code == LTU)
10819 && num_sign_bit_copies (op0, mode) == mode_width)
10821 code = (code == EQ || code == LE || code == GEU ? NE : EQ);
10822 const_op = 0;
10825 /* Do some canonicalizations based on the comparison code. We prefer
10826 comparisons against zero and then prefer equality comparisons.
10827 If we can reduce the size of a constant, we will do that too. */
10828 switch (code)
10830 case LT:
10831 /* < C is equivalent to <= (C - 1) */
10832 if (const_op > 0)
10834 const_op -= 1;
10835 code = LE;
10836 /* ... fall through to LE case below. */
10838 else
10839 break;
10841 case LE:
10842 /* <= C is equivalent to < (C + 1); we do this for C < 0 */
10843 if (const_op < 0)
10845 const_op += 1;
10846 code = LT;
10849 /* If we are doing a <= 0 comparison on a value known to have
10850 a zero sign bit, we can replace this with == 0. */
10851 else if (const_op == 0
10852 && mode_width <= HOST_BITS_PER_WIDE_INT
10853 && (nonzero_bits (op0, mode)
10854 & ((unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10855 == 0)
10856 code = EQ;
10857 break;
10859 case GE:
10860 /* >= C is equivalent to > (C - 1). */
10861 if (const_op > 0)
10863 const_op -= 1;
10864 code = GT;
10865 /* ... fall through to GT below. */
10867 else
10868 break;
10870 case GT:
10871 /* > C is equivalent to >= (C + 1); we do this for C < 0. */
10872 if (const_op < 0)
10874 const_op += 1;
10875 code = GE;
10878 /* If we are doing a > 0 comparison on a value known to have
10879 a zero sign bit, we can replace this with != 0. */
10880 else if (const_op == 0
10881 && mode_width <= HOST_BITS_PER_WIDE_INT
10882 && (nonzero_bits (op0, mode)
10883 & ((unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10884 == 0)
10885 code = NE;
10886 break;
10888 case LTU:
10889 /* < C is equivalent to <= (C - 1). */
10890 if (const_op > 0)
10892 const_op -= 1;
10893 code = LEU;
10894 /* ... fall through ... */
10896 /* (unsigned) < 0x80000000 is equivalent to >= 0. */
10897 else if (mode_width <= HOST_BITS_PER_WIDE_INT
10898 && (unsigned HOST_WIDE_INT) const_op
10899 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1))
10901 const_op = 0;
10902 code = GE;
10903 break;
10905 else
10906 break;
10908 case LEU:
10909 /* unsigned <= 0 is equivalent to == 0 */
10910 if (const_op == 0)
10911 code = EQ;
10912 /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */
10913 else if (mode_width <= HOST_BITS_PER_WIDE_INT
10914 && (unsigned HOST_WIDE_INT) const_op
10915 == ((unsigned HOST_WIDE_INT) 1 << (mode_width - 1)) - 1)
10917 const_op = 0;
10918 code = GE;
10920 break;
10922 case GEU:
10923 /* >= C is equivalent to > (C - 1). */
10924 if (const_op > 1)
10926 const_op -= 1;
10927 code = GTU;
10928 /* ... fall through ... */
10931 /* (unsigned) >= 0x80000000 is equivalent to < 0. */
10932 else if (mode_width <= HOST_BITS_PER_WIDE_INT
10933 && (unsigned HOST_WIDE_INT) const_op
10934 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1))
10936 const_op = 0;
10937 code = LT;
10938 break;
10940 else
10941 break;
10943 case GTU:
10944 /* unsigned > 0 is equivalent to != 0 */
10945 if (const_op == 0)
10946 code = NE;
10947 /* (unsigned) > 0x7fffffff is equivalent to < 0. */
10948 else if (mode_width <= HOST_BITS_PER_WIDE_INT
10949 && (unsigned HOST_WIDE_INT) const_op
10950 == ((unsigned HOST_WIDE_INT) 1 << (mode_width - 1)) - 1)
10952 const_op = 0;
10953 code = LT;
10955 break;
10957 default:
10958 break;
10961 *pop1 = GEN_INT (const_op);
10962 return code;
10965 /* Simplify a comparison between *POP0 and *POP1 where CODE is the
10966 comparison code that will be tested.
10968 The result is a possibly different comparison code to use. *POP0 and
10969 *POP1 may be updated.
10971 It is possible that we might detect that a comparison is either always
10972 true or always false. However, we do not perform general constant
10973 folding in combine, so this knowledge isn't useful. Such tautologies
10974 should have been detected earlier. Hence we ignore all such cases. */
10976 static enum rtx_code
10977 simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1)
10979 rtx op0 = *pop0;
10980 rtx op1 = *pop1;
10981 rtx tem, tem1;
10982 int i;
10983 enum machine_mode mode, tmode;
10985 /* Try a few ways of applying the same transformation to both operands. */
10986 while (1)
10988 #ifndef WORD_REGISTER_OPERATIONS
10989 /* The test below this one won't handle SIGN_EXTENDs on these machines,
10990 so check specially. */
10991 if (code != GTU && code != GEU && code != LTU && code != LEU
10992 && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT
10993 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10994 && GET_CODE (XEXP (op1, 0)) == ASHIFT
10995 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG
10996 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG
10997 && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))
10998 == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))))
10999 && CONST_INT_P (XEXP (op0, 1))
11000 && XEXP (op0, 1) == XEXP (op1, 1)
11001 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
11002 && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1)
11003 && (INTVAL (XEXP (op0, 1))
11004 == (GET_MODE_PRECISION (GET_MODE (op0))
11005 - (GET_MODE_PRECISION
11006 (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))))))))
11008 op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0));
11009 op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0));
11011 #endif
11013 /* If both operands are the same constant shift, see if we can ignore the
11014 shift. We can if the shift is a rotate or if the bits shifted out of
11015 this shift are known to be zero for both inputs and if the type of
11016 comparison is compatible with the shift. */
11017 if (GET_CODE (op0) == GET_CODE (op1)
11018 && HWI_COMPUTABLE_MODE_P (GET_MODE(op0))
11019 && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ))
11020 || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT)
11021 && (code != GT && code != LT && code != GE && code != LE))
11022 || (GET_CODE (op0) == ASHIFTRT
11023 && (code != GTU && code != LTU
11024 && code != GEU && code != LEU)))
11025 && CONST_INT_P (XEXP (op0, 1))
11026 && INTVAL (XEXP (op0, 1)) >= 0
11027 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
11028 && XEXP (op0, 1) == XEXP (op1, 1))
11030 enum machine_mode mode = GET_MODE (op0);
11031 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
11032 int shift_count = INTVAL (XEXP (op0, 1));
11034 if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT)
11035 mask &= (mask >> shift_count) << shift_count;
11036 else if (GET_CODE (op0) == ASHIFT)
11037 mask = (mask & (mask << shift_count)) >> shift_count;
11039 if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0
11040 && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0)
11041 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0);
11042 else
11043 break;
11046 /* If both operands are AND's of a paradoxical SUBREG by constant, the
11047 SUBREGs are of the same mode, and, in both cases, the AND would
11048 be redundant if the comparison was done in the narrower mode,
11049 do the comparison in the narrower mode (e.g., we are AND'ing with 1
11050 and the operand's possibly nonzero bits are 0xffffff01; in that case
11051 if we only care about QImode, we don't need the AND). This case
11052 occurs if the output mode of an scc insn is not SImode and
11053 STORE_FLAG_VALUE == 1 (e.g., the 386).
11055 Similarly, check for a case where the AND's are ZERO_EXTEND
11056 operations from some narrower mode even though a SUBREG is not
11057 present. */
11059 else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND
11060 && CONST_INT_P (XEXP (op0, 1))
11061 && CONST_INT_P (XEXP (op1, 1)))
11063 rtx inner_op0 = XEXP (op0, 0);
11064 rtx inner_op1 = XEXP (op1, 0);
11065 HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1));
11066 HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1));
11067 int changed = 0;
11069 if (paradoxical_subreg_p (inner_op0)
11070 && GET_CODE (inner_op1) == SUBREG
11071 && (GET_MODE (SUBREG_REG (inner_op0))
11072 == GET_MODE (SUBREG_REG (inner_op1)))
11073 && (GET_MODE_PRECISION (GET_MODE (SUBREG_REG (inner_op0)))
11074 <= HOST_BITS_PER_WIDE_INT)
11075 && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0),
11076 GET_MODE (SUBREG_REG (inner_op0)))))
11077 && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1),
11078 GET_MODE (SUBREG_REG (inner_op1))))))
11080 op0 = SUBREG_REG (inner_op0);
11081 op1 = SUBREG_REG (inner_op1);
11083 /* The resulting comparison is always unsigned since we masked
11084 off the original sign bit. */
11085 code = unsigned_condition (code);
11087 changed = 1;
11090 else if (c0 == c1)
11091 for (tmode = GET_CLASS_NARROWEST_MODE
11092 (GET_MODE_CLASS (GET_MODE (op0)));
11093 tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode))
11094 if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode))
11096 op0 = gen_lowpart (tmode, inner_op0);
11097 op1 = gen_lowpart (tmode, inner_op1);
11098 code = unsigned_condition (code);
11099 changed = 1;
11100 break;
11103 if (! changed)
11104 break;
11107 /* If both operands are NOT, we can strip off the outer operation
11108 and adjust the comparison code for swapped operands; similarly for
11109 NEG, except that this must be an equality comparison. */
11110 else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT)
11111 || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG
11112 && (code == EQ || code == NE)))
11113 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code);
11115 else
11116 break;
11119 /* If the first operand is a constant, swap the operands and adjust the
11120 comparison code appropriately, but don't do this if the second operand
11121 is already a constant integer. */
11122 if (swap_commutative_operands_p (op0, op1))
11124 tem = op0, op0 = op1, op1 = tem;
11125 code = swap_condition (code);
11128 /* We now enter a loop during which we will try to simplify the comparison.
11129 For the most part, we only are concerned with comparisons with zero,
11130 but some things may really be comparisons with zero but not start
11131 out looking that way. */
11133 while (CONST_INT_P (op1))
11135 enum machine_mode mode = GET_MODE (op0);
11136 unsigned int mode_width = GET_MODE_PRECISION (mode);
11137 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
11138 int equality_comparison_p;
11139 int sign_bit_comparison_p;
11140 int unsigned_comparison_p;
11141 HOST_WIDE_INT const_op;
11143 /* We only want to handle integral modes. This catches VOIDmode,
11144 CCmode, and the floating-point modes. An exception is that we
11145 can handle VOIDmode if OP0 is a COMPARE or a comparison
11146 operation. */
11148 if (GET_MODE_CLASS (mode) != MODE_INT
11149 && ! (mode == VOIDmode
11150 && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0))))
11151 break;
11153 /* Try to simplify the compare to constant, possibly changing the
11154 comparison op, and/or changing op1 to zero. */
11155 code = simplify_compare_const (code, op0, &op1);
11156 const_op = INTVAL (op1);
11158 /* Compute some predicates to simplify code below. */
11160 equality_comparison_p = (code == EQ || code == NE);
11161 sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0);
11162 unsigned_comparison_p = (code == LTU || code == LEU || code == GTU
11163 || code == GEU);
11165 /* If this is a sign bit comparison and we can do arithmetic in
11166 MODE, say that we will only be needing the sign bit of OP0. */
11167 if (sign_bit_comparison_p && HWI_COMPUTABLE_MODE_P (mode))
11168 op0 = force_to_mode (op0, mode,
11169 (unsigned HOST_WIDE_INT) 1
11170 << (GET_MODE_PRECISION (mode) - 1),
11173 /* Now try cases based on the opcode of OP0. If none of the cases
11174 does a "continue", we exit this loop immediately after the
11175 switch. */
11177 switch (GET_CODE (op0))
11179 case ZERO_EXTRACT:
11180 /* If we are extracting a single bit from a variable position in
11181 a constant that has only a single bit set and are comparing it
11182 with zero, we can convert this into an equality comparison
11183 between the position and the location of the single bit. */
11184 /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might
11185 have already reduced the shift count modulo the word size. */
11186 if (!SHIFT_COUNT_TRUNCATED
11187 && CONST_INT_P (XEXP (op0, 0))
11188 && XEXP (op0, 1) == const1_rtx
11189 && equality_comparison_p && const_op == 0
11190 && (i = exact_log2 (UINTVAL (XEXP (op0, 0)))) >= 0)
11192 if (BITS_BIG_ENDIAN)
11193 i = BITS_PER_WORD - 1 - i;
11195 op0 = XEXP (op0, 2);
11196 op1 = GEN_INT (i);
11197 const_op = i;
11199 /* Result is nonzero iff shift count is equal to I. */
11200 code = reverse_condition (code);
11201 continue;
11204 /* ... fall through ... */
11206 case SIGN_EXTRACT:
11207 tem = expand_compound_operation (op0);
11208 if (tem != op0)
11210 op0 = tem;
11211 continue;
11213 break;
11215 case NOT:
11216 /* If testing for equality, we can take the NOT of the constant. */
11217 if (equality_comparison_p
11218 && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0)
11220 op0 = XEXP (op0, 0);
11221 op1 = tem;
11222 continue;
11225 /* If just looking at the sign bit, reverse the sense of the
11226 comparison. */
11227 if (sign_bit_comparison_p)
11229 op0 = XEXP (op0, 0);
11230 code = (code == GE ? LT : GE);
11231 continue;
11233 break;
11235 case NEG:
11236 /* If testing for equality, we can take the NEG of the constant. */
11237 if (equality_comparison_p
11238 && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0)
11240 op0 = XEXP (op0, 0);
11241 op1 = tem;
11242 continue;
11245 /* The remaining cases only apply to comparisons with zero. */
11246 if (const_op != 0)
11247 break;
11249 /* When X is ABS or is known positive,
11250 (neg X) is < 0 if and only if X != 0. */
11252 if (sign_bit_comparison_p
11253 && (GET_CODE (XEXP (op0, 0)) == ABS
11254 || (mode_width <= HOST_BITS_PER_WIDE_INT
11255 && (nonzero_bits (XEXP (op0, 0), mode)
11256 & ((unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
11257 == 0)))
11259 op0 = XEXP (op0, 0);
11260 code = (code == LT ? NE : EQ);
11261 continue;
11264 /* If we have NEG of something whose two high-order bits are the
11265 same, we know that "(-a) < 0" is equivalent to "a > 0". */
11266 if (num_sign_bit_copies (op0, mode) >= 2)
11268 op0 = XEXP (op0, 0);
11269 code = swap_condition (code);
11270 continue;
11272 break;
11274 case ROTATE:
11275 /* If we are testing equality and our count is a constant, we
11276 can perform the inverse operation on our RHS. */
11277 if (equality_comparison_p && CONST_INT_P (XEXP (op0, 1))
11278 && (tem = simplify_binary_operation (ROTATERT, mode,
11279 op1, XEXP (op0, 1))) != 0)
11281 op0 = XEXP (op0, 0);
11282 op1 = tem;
11283 continue;
11286 /* If we are doing a < 0 or >= 0 comparison, it means we are testing
11287 a particular bit. Convert it to an AND of a constant of that
11288 bit. This will be converted into a ZERO_EXTRACT. */
11289 if (const_op == 0 && sign_bit_comparison_p
11290 && CONST_INT_P (XEXP (op0, 1))
11291 && mode_width <= HOST_BITS_PER_WIDE_INT)
11293 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
11294 ((unsigned HOST_WIDE_INT) 1
11295 << (mode_width - 1
11296 - INTVAL (XEXP (op0, 1)))));
11297 code = (code == LT ? NE : EQ);
11298 continue;
11301 /* Fall through. */
11303 case ABS:
11304 /* ABS is ignorable inside an equality comparison with zero. */
11305 if (const_op == 0 && equality_comparison_p)
11307 op0 = XEXP (op0, 0);
11308 continue;
11310 break;
11312 case SIGN_EXTEND:
11313 /* Can simplify (compare (zero/sign_extend FOO) CONST) to
11314 (compare FOO CONST) if CONST fits in FOO's mode and we
11315 are either testing inequality or have an unsigned
11316 comparison with ZERO_EXTEND or a signed comparison with
11317 SIGN_EXTEND. But don't do it if we don't have a compare
11318 insn of the given mode, since we'd have to revert it
11319 later on, and then we wouldn't know whether to sign- or
11320 zero-extend. */
11321 mode = GET_MODE (XEXP (op0, 0));
11322 if (GET_MODE_CLASS (mode) == MODE_INT
11323 && ! unsigned_comparison_p
11324 && HWI_COMPUTABLE_MODE_P (mode)
11325 && trunc_int_for_mode (const_op, mode) == const_op
11326 && have_insn_for (COMPARE, mode))
11328 op0 = XEXP (op0, 0);
11329 continue;
11331 break;
11333 case SUBREG:
11334 /* Check for the case where we are comparing A - C1 with C2, that is
11336 (subreg:MODE (plus (A) (-C1))) op (C2)
11338 with C1 a constant, and try to lift the SUBREG, i.e. to do the
11339 comparison in the wider mode. One of the following two conditions
11340 must be true in order for this to be valid:
11342 1. The mode extension results in the same bit pattern being added
11343 on both sides and the comparison is equality or unsigned. As
11344 C2 has been truncated to fit in MODE, the pattern can only be
11345 all 0s or all 1s.
11347 2. The mode extension results in the sign bit being copied on
11348 each side.
11350 The difficulty here is that we have predicates for A but not for
11351 (A - C1) so we need to check that C1 is within proper bounds so
11352 as to perturbate A as little as possible. */
11354 if (mode_width <= HOST_BITS_PER_WIDE_INT
11355 && subreg_lowpart_p (op0)
11356 && GET_MODE_PRECISION (GET_MODE (SUBREG_REG (op0))) > mode_width
11357 && GET_CODE (SUBREG_REG (op0)) == PLUS
11358 && CONST_INT_P (XEXP (SUBREG_REG (op0), 1)))
11360 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
11361 rtx a = XEXP (SUBREG_REG (op0), 0);
11362 HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1));
11364 if ((c1 > 0
11365 && (unsigned HOST_WIDE_INT) c1
11366 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)
11367 && (equality_comparison_p || unsigned_comparison_p)
11368 /* (A - C1) zero-extends if it is positive and sign-extends
11369 if it is negative, C2 both zero- and sign-extends. */
11370 && ((0 == (nonzero_bits (a, inner_mode)
11371 & ~GET_MODE_MASK (mode))
11372 && const_op >= 0)
11373 /* (A - C1) sign-extends if it is positive and 1-extends
11374 if it is negative, C2 both sign- and 1-extends. */
11375 || (num_sign_bit_copies (a, inner_mode)
11376 > (unsigned int) (GET_MODE_PRECISION (inner_mode)
11377 - mode_width)
11378 && const_op < 0)))
11379 || ((unsigned HOST_WIDE_INT) c1
11380 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 2)
11381 /* (A - C1) always sign-extends, like C2. */
11382 && num_sign_bit_copies (a, inner_mode)
11383 > (unsigned int) (GET_MODE_PRECISION (inner_mode)
11384 - (mode_width - 1))))
11386 op0 = SUBREG_REG (op0);
11387 continue;
11391 /* If the inner mode is narrower and we are extracting the low part,
11392 we can treat the SUBREG as if it were a ZERO_EXTEND. */
11393 if (subreg_lowpart_p (op0)
11394 && GET_MODE_PRECISION (GET_MODE (SUBREG_REG (op0))) < mode_width)
11395 /* Fall through */ ;
11396 else
11397 break;
11399 /* ... fall through ... */
11401 case ZERO_EXTEND:
11402 mode = GET_MODE (XEXP (op0, 0));
11403 if (GET_MODE_CLASS (mode) == MODE_INT
11404 && (unsigned_comparison_p || equality_comparison_p)
11405 && HWI_COMPUTABLE_MODE_P (mode)
11406 && (unsigned HOST_WIDE_INT) const_op <= GET_MODE_MASK (mode)
11407 && const_op >= 0
11408 && have_insn_for (COMPARE, mode))
11410 op0 = XEXP (op0, 0);
11411 continue;
11413 break;
11415 case PLUS:
11416 /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do
11417 this for equality comparisons due to pathological cases involving
11418 overflows. */
11419 if (equality_comparison_p
11420 && 0 != (tem = simplify_binary_operation (MINUS, mode,
11421 op1, XEXP (op0, 1))))
11423 op0 = XEXP (op0, 0);
11424 op1 = tem;
11425 continue;
11428 /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */
11429 if (const_op == 0 && XEXP (op0, 1) == constm1_rtx
11430 && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p)
11432 op0 = XEXP (XEXP (op0, 0), 0);
11433 code = (code == LT ? EQ : NE);
11434 continue;
11436 break;
11438 case MINUS:
11439 /* We used to optimize signed comparisons against zero, but that
11440 was incorrect. Unsigned comparisons against zero (GTU, LEU)
11441 arrive here as equality comparisons, or (GEU, LTU) are
11442 optimized away. No need to special-case them. */
11444 /* (eq (minus A B) C) -> (eq A (plus B C)) or
11445 (eq B (minus A C)), whichever simplifies. We can only do
11446 this for equality comparisons due to pathological cases involving
11447 overflows. */
11448 if (equality_comparison_p
11449 && 0 != (tem = simplify_binary_operation (PLUS, mode,
11450 XEXP (op0, 1), op1)))
11452 op0 = XEXP (op0, 0);
11453 op1 = tem;
11454 continue;
11457 if (equality_comparison_p
11458 && 0 != (tem = simplify_binary_operation (MINUS, mode,
11459 XEXP (op0, 0), op1)))
11461 op0 = XEXP (op0, 1);
11462 op1 = tem;
11463 continue;
11466 /* The sign bit of (minus (ashiftrt X C) X), where C is the number
11467 of bits in X minus 1, is one iff X > 0. */
11468 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT
11469 && CONST_INT_P (XEXP (XEXP (op0, 0), 1))
11470 && UINTVAL (XEXP (XEXP (op0, 0), 1)) == mode_width - 1
11471 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
11473 op0 = XEXP (op0, 1);
11474 code = (code == GE ? LE : GT);
11475 continue;
11477 break;
11479 case XOR:
11480 /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification
11481 if C is zero or B is a constant. */
11482 if (equality_comparison_p
11483 && 0 != (tem = simplify_binary_operation (XOR, mode,
11484 XEXP (op0, 1), op1)))
11486 op0 = XEXP (op0, 0);
11487 op1 = tem;
11488 continue;
11490 break;
11492 case EQ: case NE:
11493 case UNEQ: case LTGT:
11494 case LT: case LTU: case UNLT: case LE: case LEU: case UNLE:
11495 case GT: case GTU: case UNGT: case GE: case GEU: case UNGE:
11496 case UNORDERED: case ORDERED:
11497 /* We can't do anything if OP0 is a condition code value, rather
11498 than an actual data value. */
11499 if (const_op != 0
11500 || CC0_P (XEXP (op0, 0))
11501 || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC)
11502 break;
11504 /* Get the two operands being compared. */
11505 if (GET_CODE (XEXP (op0, 0)) == COMPARE)
11506 tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1);
11507 else
11508 tem = XEXP (op0, 0), tem1 = XEXP (op0, 1);
11510 /* Check for the cases where we simply want the result of the
11511 earlier test or the opposite of that result. */
11512 if (code == NE || code == EQ
11513 || (val_signbit_known_set_p (GET_MODE (op0), STORE_FLAG_VALUE)
11514 && (code == LT || code == GE)))
11516 enum rtx_code new_code;
11517 if (code == LT || code == NE)
11518 new_code = GET_CODE (op0);
11519 else
11520 new_code = reversed_comparison_code (op0, NULL);
11522 if (new_code != UNKNOWN)
11524 code = new_code;
11525 op0 = tem;
11526 op1 = tem1;
11527 continue;
11530 break;
11532 case IOR:
11533 /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero
11534 iff X <= 0. */
11535 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS
11536 && XEXP (XEXP (op0, 0), 1) == constm1_rtx
11537 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
11539 op0 = XEXP (op0, 1);
11540 code = (code == GE ? GT : LE);
11541 continue;
11543 break;
11545 case AND:
11546 /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This
11547 will be converted to a ZERO_EXTRACT later. */
11548 if (const_op == 0 && equality_comparison_p
11549 && GET_CODE (XEXP (op0, 0)) == ASHIFT
11550 && XEXP (XEXP (op0, 0), 0) == const1_rtx)
11552 op0 = gen_rtx_LSHIFTRT (mode, XEXP (op0, 1),
11553 XEXP (XEXP (op0, 0), 1));
11554 op0 = simplify_and_const_int (NULL_RTX, mode, op0, 1);
11555 continue;
11558 /* If we are comparing (and (lshiftrt X C1) C2) for equality with
11559 zero and X is a comparison and C1 and C2 describe only bits set
11560 in STORE_FLAG_VALUE, we can compare with X. */
11561 if (const_op == 0 && equality_comparison_p
11562 && mode_width <= HOST_BITS_PER_WIDE_INT
11563 && CONST_INT_P (XEXP (op0, 1))
11564 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT
11565 && CONST_INT_P (XEXP (XEXP (op0, 0), 1))
11566 && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0
11567 && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT)
11569 mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
11570 << INTVAL (XEXP (XEXP (op0, 0), 1)));
11571 if ((~STORE_FLAG_VALUE & mask) == 0
11572 && (COMPARISON_P (XEXP (XEXP (op0, 0), 0))
11573 || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0
11574 && COMPARISON_P (tem))))
11576 op0 = XEXP (XEXP (op0, 0), 0);
11577 continue;
11581 /* If we are doing an equality comparison of an AND of a bit equal
11582 to the sign bit, replace this with a LT or GE comparison of
11583 the underlying value. */
11584 if (equality_comparison_p
11585 && const_op == 0
11586 && CONST_INT_P (XEXP (op0, 1))
11587 && mode_width <= HOST_BITS_PER_WIDE_INT
11588 && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
11589 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
11591 op0 = XEXP (op0, 0);
11592 code = (code == EQ ? GE : LT);
11593 continue;
11596 /* If this AND operation is really a ZERO_EXTEND from a narrower
11597 mode, the constant fits within that mode, and this is either an
11598 equality or unsigned comparison, try to do this comparison in
11599 the narrower mode.
11601 Note that in:
11603 (ne:DI (and:DI (reg:DI 4) (const_int 0xffffffff)) (const_int 0))
11604 -> (ne:DI (reg:SI 4) (const_int 0))
11606 unless TRULY_NOOP_TRUNCATION allows it or the register is
11607 known to hold a value of the required mode the
11608 transformation is invalid. */
11609 if ((equality_comparison_p || unsigned_comparison_p)
11610 && CONST_INT_P (XEXP (op0, 1))
11611 && (i = exact_log2 ((UINTVAL (XEXP (op0, 1))
11612 & GET_MODE_MASK (mode))
11613 + 1)) >= 0
11614 && const_op >> i == 0
11615 && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode
11616 && (TRULY_NOOP_TRUNCATION_MODES_P (tmode, GET_MODE (op0))
11617 || (REG_P (XEXP (op0, 0))
11618 && reg_truncated_to_mode (tmode, XEXP (op0, 0)))))
11620 op0 = gen_lowpart (tmode, XEXP (op0, 0));
11621 continue;
11624 /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1
11625 fits in both M1 and M2 and the SUBREG is either paradoxical
11626 or represents the low part, permute the SUBREG and the AND
11627 and try again. */
11628 if (GET_CODE (XEXP (op0, 0)) == SUBREG)
11630 unsigned HOST_WIDE_INT c1;
11631 tmode = GET_MODE (SUBREG_REG (XEXP (op0, 0)));
11632 /* Require an integral mode, to avoid creating something like
11633 (AND:SF ...). */
11634 if (SCALAR_INT_MODE_P (tmode)
11635 /* It is unsafe to commute the AND into the SUBREG if the
11636 SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is
11637 not defined. As originally written the upper bits
11638 have a defined value due to the AND operation.
11639 However, if we commute the AND inside the SUBREG then
11640 they no longer have defined values and the meaning of
11641 the code has been changed. */
11642 && (0
11643 #ifdef WORD_REGISTER_OPERATIONS
11644 || (mode_width > GET_MODE_PRECISION (tmode)
11645 && mode_width <= BITS_PER_WORD)
11646 #endif
11647 || (mode_width <= GET_MODE_PRECISION (tmode)
11648 && subreg_lowpart_p (XEXP (op0, 0))))
11649 && CONST_INT_P (XEXP (op0, 1))
11650 && mode_width <= HOST_BITS_PER_WIDE_INT
11651 && HWI_COMPUTABLE_MODE_P (tmode)
11652 && ((c1 = INTVAL (XEXP (op0, 1))) & ~mask) == 0
11653 && (c1 & ~GET_MODE_MASK (tmode)) == 0
11654 && c1 != mask
11655 && c1 != GET_MODE_MASK (tmode))
11657 op0 = simplify_gen_binary (AND, tmode,
11658 SUBREG_REG (XEXP (op0, 0)),
11659 gen_int_mode (c1, tmode));
11660 op0 = gen_lowpart (mode, op0);
11661 continue;
11665 /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */
11666 if (const_op == 0 && equality_comparison_p
11667 && XEXP (op0, 1) == const1_rtx
11668 && GET_CODE (XEXP (op0, 0)) == NOT)
11670 op0 = simplify_and_const_int (NULL_RTX, mode,
11671 XEXP (XEXP (op0, 0), 0), 1);
11672 code = (code == NE ? EQ : NE);
11673 continue;
11676 /* Convert (ne (and (lshiftrt (not X)) 1) 0) to
11677 (eq (and (lshiftrt X) 1) 0).
11678 Also handle the case where (not X) is expressed using xor. */
11679 if (const_op == 0 && equality_comparison_p
11680 && XEXP (op0, 1) == const1_rtx
11681 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT)
11683 rtx shift_op = XEXP (XEXP (op0, 0), 0);
11684 rtx shift_count = XEXP (XEXP (op0, 0), 1);
11686 if (GET_CODE (shift_op) == NOT
11687 || (GET_CODE (shift_op) == XOR
11688 && CONST_INT_P (XEXP (shift_op, 1))
11689 && CONST_INT_P (shift_count)
11690 && HWI_COMPUTABLE_MODE_P (mode)
11691 && (UINTVAL (XEXP (shift_op, 1))
11692 == (unsigned HOST_WIDE_INT) 1
11693 << INTVAL (shift_count))))
11696 = gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count);
11697 op0 = simplify_and_const_int (NULL_RTX, mode, op0, 1);
11698 code = (code == NE ? EQ : NE);
11699 continue;
11702 break;
11704 case ASHIFT:
11705 /* If we have (compare (ashift FOO N) (const_int C)) and
11706 the high order N bits of FOO (N+1 if an inequality comparison)
11707 are known to be zero, we can do this by comparing FOO with C
11708 shifted right N bits so long as the low-order N bits of C are
11709 zero. */
11710 if (CONST_INT_P (XEXP (op0, 1))
11711 && INTVAL (XEXP (op0, 1)) >= 0
11712 && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p)
11713 < HOST_BITS_PER_WIDE_INT)
11714 && (((unsigned HOST_WIDE_INT) const_op
11715 & (((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1)))
11716 - 1)) == 0)
11717 && mode_width <= HOST_BITS_PER_WIDE_INT
11718 && (nonzero_bits (XEXP (op0, 0), mode)
11719 & ~(mask >> (INTVAL (XEXP (op0, 1))
11720 + ! equality_comparison_p))) == 0)
11722 /* We must perform a logical shift, not an arithmetic one,
11723 as we want the top N bits of C to be zero. */
11724 unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode);
11726 temp >>= INTVAL (XEXP (op0, 1));
11727 op1 = gen_int_mode (temp, mode);
11728 op0 = XEXP (op0, 0);
11729 continue;
11732 /* If we are doing a sign bit comparison, it means we are testing
11733 a particular bit. Convert it to the appropriate AND. */
11734 if (sign_bit_comparison_p && CONST_INT_P (XEXP (op0, 1))
11735 && mode_width <= HOST_BITS_PER_WIDE_INT)
11737 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
11738 ((unsigned HOST_WIDE_INT) 1
11739 << (mode_width - 1
11740 - INTVAL (XEXP (op0, 1)))));
11741 code = (code == LT ? NE : EQ);
11742 continue;
11745 /* If this an equality comparison with zero and we are shifting
11746 the low bit to the sign bit, we can convert this to an AND of the
11747 low-order bit. */
11748 if (const_op == 0 && equality_comparison_p
11749 && CONST_INT_P (XEXP (op0, 1))
11750 && UINTVAL (XEXP (op0, 1)) == mode_width - 1)
11752 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), 1);
11753 continue;
11755 break;
11757 case ASHIFTRT:
11758 /* If this is an equality comparison with zero, we can do this
11759 as a logical shift, which might be much simpler. */
11760 if (equality_comparison_p && const_op == 0
11761 && CONST_INT_P (XEXP (op0, 1)))
11763 op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode,
11764 XEXP (op0, 0),
11765 INTVAL (XEXP (op0, 1)));
11766 continue;
11769 /* If OP0 is a sign extension and CODE is not an unsigned comparison,
11770 do the comparison in a narrower mode. */
11771 if (! unsigned_comparison_p
11772 && CONST_INT_P (XEXP (op0, 1))
11773 && GET_CODE (XEXP (op0, 0)) == ASHIFT
11774 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
11775 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
11776 MODE_INT, 1)) != BLKmode
11777 && (((unsigned HOST_WIDE_INT) const_op
11778 + (GET_MODE_MASK (tmode) >> 1) + 1)
11779 <= GET_MODE_MASK (tmode)))
11781 op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0));
11782 continue;
11785 /* Likewise if OP0 is a PLUS of a sign extension with a
11786 constant, which is usually represented with the PLUS
11787 between the shifts. */
11788 if (! unsigned_comparison_p
11789 && CONST_INT_P (XEXP (op0, 1))
11790 && GET_CODE (XEXP (op0, 0)) == PLUS
11791 && CONST_INT_P (XEXP (XEXP (op0, 0), 1))
11792 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT
11793 && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1)
11794 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
11795 MODE_INT, 1)) != BLKmode
11796 && (((unsigned HOST_WIDE_INT) const_op
11797 + (GET_MODE_MASK (tmode) >> 1) + 1)
11798 <= GET_MODE_MASK (tmode)))
11800 rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0);
11801 rtx add_const = XEXP (XEXP (op0, 0), 1);
11802 rtx new_const = simplify_gen_binary (ASHIFTRT, GET_MODE (op0),
11803 add_const, XEXP (op0, 1));
11805 op0 = simplify_gen_binary (PLUS, tmode,
11806 gen_lowpart (tmode, inner),
11807 new_const);
11808 continue;
11811 /* ... fall through ... */
11812 case LSHIFTRT:
11813 /* If we have (compare (xshiftrt FOO N) (const_int C)) and
11814 the low order N bits of FOO are known to be zero, we can do this
11815 by comparing FOO with C shifted left N bits so long as no
11816 overflow occurs. Even if the low order N bits of FOO aren't known
11817 to be zero, if the comparison is >= or < we can use the same
11818 optimization and for > or <= by setting all the low
11819 order N bits in the comparison constant. */
11820 if (CONST_INT_P (XEXP (op0, 1))
11821 && INTVAL (XEXP (op0, 1)) > 0
11822 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
11823 && mode_width <= HOST_BITS_PER_WIDE_INT
11824 && (((unsigned HOST_WIDE_INT) const_op
11825 + (GET_CODE (op0) != LSHIFTRT
11826 ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1)
11827 + 1)
11828 : 0))
11829 <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1))))
11831 unsigned HOST_WIDE_INT low_bits
11832 = (nonzero_bits (XEXP (op0, 0), mode)
11833 & (((unsigned HOST_WIDE_INT) 1
11834 << INTVAL (XEXP (op0, 1))) - 1));
11835 if (low_bits == 0 || !equality_comparison_p)
11837 /* If the shift was logical, then we must make the condition
11838 unsigned. */
11839 if (GET_CODE (op0) == LSHIFTRT)
11840 code = unsigned_condition (code);
11842 const_op <<= INTVAL (XEXP (op0, 1));
11843 if (low_bits != 0
11844 && (code == GT || code == GTU
11845 || code == LE || code == LEU))
11846 const_op
11847 |= (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1);
11848 op1 = GEN_INT (const_op);
11849 op0 = XEXP (op0, 0);
11850 continue;
11854 /* If we are using this shift to extract just the sign bit, we
11855 can replace this with an LT or GE comparison. */
11856 if (const_op == 0
11857 && (equality_comparison_p || sign_bit_comparison_p)
11858 && CONST_INT_P (XEXP (op0, 1))
11859 && UINTVAL (XEXP (op0, 1)) == mode_width - 1)
11861 op0 = XEXP (op0, 0);
11862 code = (code == NE || code == GT ? LT : GE);
11863 continue;
11865 break;
11867 default:
11868 break;
11871 break;
11874 /* Now make any compound operations involved in this comparison. Then,
11875 check for an outmost SUBREG on OP0 that is not doing anything or is
11876 paradoxical. The latter transformation must only be performed when
11877 it is known that the "extra" bits will be the same in op0 and op1 or
11878 that they don't matter. There are three cases to consider:
11880 1. SUBREG_REG (op0) is a register. In this case the bits are don't
11881 care bits and we can assume they have any convenient value. So
11882 making the transformation is safe.
11884 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not defined.
11885 In this case the upper bits of op0 are undefined. We should not make
11886 the simplification in that case as we do not know the contents of
11887 those bits.
11889 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined and not
11890 UNKNOWN. In that case we know those bits are zeros or ones. We must
11891 also be sure that they are the same as the upper bits of op1.
11893 We can never remove a SUBREG for a non-equality comparison because
11894 the sign bit is in a different place in the underlying object. */
11896 op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET);
11897 op1 = make_compound_operation (op1, SET);
11899 if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0)
11900 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
11901 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op0))) == MODE_INT
11902 && (code == NE || code == EQ))
11904 if (paradoxical_subreg_p (op0))
11906 /* For paradoxical subregs, allow case 1 as above. Case 3 isn't
11907 implemented. */
11908 if (REG_P (SUBREG_REG (op0)))
11910 op0 = SUBREG_REG (op0);
11911 op1 = gen_lowpart (GET_MODE (op0), op1);
11914 else if ((GET_MODE_PRECISION (GET_MODE (SUBREG_REG (op0)))
11915 <= HOST_BITS_PER_WIDE_INT)
11916 && (nonzero_bits (SUBREG_REG (op0),
11917 GET_MODE (SUBREG_REG (op0)))
11918 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
11920 tem = gen_lowpart (GET_MODE (SUBREG_REG (op0)), op1);
11922 if ((nonzero_bits (tem, GET_MODE (SUBREG_REG (op0)))
11923 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
11924 op0 = SUBREG_REG (op0), op1 = tem;
11928 /* We now do the opposite procedure: Some machines don't have compare
11929 insns in all modes. If OP0's mode is an integer mode smaller than a
11930 word and we can't do a compare in that mode, see if there is a larger
11931 mode for which we can do the compare. There are a number of cases in
11932 which we can use the wider mode. */
11934 mode = GET_MODE (op0);
11935 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
11936 && GET_MODE_SIZE (mode) < UNITS_PER_WORD
11937 && ! have_insn_for (COMPARE, mode))
11938 for (tmode = GET_MODE_WIDER_MODE (mode);
11939 (tmode != VOIDmode && HWI_COMPUTABLE_MODE_P (tmode));
11940 tmode = GET_MODE_WIDER_MODE (tmode))
11941 if (have_insn_for (COMPARE, tmode))
11943 int zero_extended;
11945 /* If this is a test for negative, we can make an explicit
11946 test of the sign bit. Test this first so we can use
11947 a paradoxical subreg to extend OP0. */
11949 if (op1 == const0_rtx && (code == LT || code == GE)
11950 && HWI_COMPUTABLE_MODE_P (mode))
11952 unsigned HOST_WIDE_INT sign
11953 = (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1);
11954 op0 = simplify_gen_binary (AND, tmode,
11955 gen_lowpart (tmode, op0),
11956 gen_int_mode (sign, mode));
11957 code = (code == LT) ? NE : EQ;
11958 break;
11961 /* If the only nonzero bits in OP0 and OP1 are those in the
11962 narrower mode and this is an equality or unsigned comparison,
11963 we can use the wider mode. Similarly for sign-extended
11964 values, in which case it is true for all comparisons. */
11965 zero_extended = ((code == EQ || code == NE
11966 || code == GEU || code == GTU
11967 || code == LEU || code == LTU)
11968 && (nonzero_bits (op0, tmode)
11969 & ~GET_MODE_MASK (mode)) == 0
11970 && ((CONST_INT_P (op1)
11971 || (nonzero_bits (op1, tmode)
11972 & ~GET_MODE_MASK (mode)) == 0)));
11974 if (zero_extended
11975 || ((num_sign_bit_copies (op0, tmode)
11976 > (unsigned int) (GET_MODE_PRECISION (tmode)
11977 - GET_MODE_PRECISION (mode)))
11978 && (num_sign_bit_copies (op1, tmode)
11979 > (unsigned int) (GET_MODE_PRECISION (tmode)
11980 - GET_MODE_PRECISION (mode)))))
11982 /* If OP0 is an AND and we don't have an AND in MODE either,
11983 make a new AND in the proper mode. */
11984 if (GET_CODE (op0) == AND
11985 && !have_insn_for (AND, mode))
11986 op0 = simplify_gen_binary (AND, tmode,
11987 gen_lowpart (tmode,
11988 XEXP (op0, 0)),
11989 gen_lowpart (tmode,
11990 XEXP (op0, 1)));
11991 else
11993 if (zero_extended)
11995 op0 = simplify_gen_unary (ZERO_EXTEND, tmode, op0, mode);
11996 op1 = simplify_gen_unary (ZERO_EXTEND, tmode, op1, mode);
11998 else
12000 op0 = simplify_gen_unary (SIGN_EXTEND, tmode, op0, mode);
12001 op1 = simplify_gen_unary (SIGN_EXTEND, tmode, op1, mode);
12003 break;
12008 /* We may have changed the comparison operands. Re-canonicalize. */
12009 if (swap_commutative_operands_p (op0, op1))
12011 tem = op0, op0 = op1, op1 = tem;
12012 code = swap_condition (code);
12015 /* If this machine only supports a subset of valid comparisons, see if we
12016 can convert an unsupported one into a supported one. */
12017 target_canonicalize_comparison (&code, &op0, &op1, 0);
12019 *pop0 = op0;
12020 *pop1 = op1;
12022 return code;
12025 /* Utility function for record_value_for_reg. Count number of
12026 rtxs in X. */
12027 static int
12028 count_rtxs (rtx x)
12030 enum rtx_code code = GET_CODE (x);
12031 const char *fmt;
12032 int i, j, ret = 1;
12034 if (GET_RTX_CLASS (code) == RTX_BIN_ARITH
12035 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
12037 rtx x0 = XEXP (x, 0);
12038 rtx x1 = XEXP (x, 1);
12040 if (x0 == x1)
12041 return 1 + 2 * count_rtxs (x0);
12043 if ((GET_RTX_CLASS (GET_CODE (x1)) == RTX_BIN_ARITH
12044 || GET_RTX_CLASS (GET_CODE (x1)) == RTX_COMM_ARITH)
12045 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
12046 return 2 + 2 * count_rtxs (x0)
12047 + count_rtxs (x == XEXP (x1, 0)
12048 ? XEXP (x1, 1) : XEXP (x1, 0));
12050 if ((GET_RTX_CLASS (GET_CODE (x0)) == RTX_BIN_ARITH
12051 || GET_RTX_CLASS (GET_CODE (x0)) == RTX_COMM_ARITH)
12052 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
12053 return 2 + 2 * count_rtxs (x1)
12054 + count_rtxs (x == XEXP (x0, 0)
12055 ? XEXP (x0, 1) : XEXP (x0, 0));
12058 fmt = GET_RTX_FORMAT (code);
12059 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
12060 if (fmt[i] == 'e')
12061 ret += count_rtxs (XEXP (x, i));
12062 else if (fmt[i] == 'E')
12063 for (j = 0; j < XVECLEN (x, i); j++)
12064 ret += count_rtxs (XVECEXP (x, i, j));
12066 return ret;
12069 /* Utility function for following routine. Called when X is part of a value
12070 being stored into last_set_value. Sets last_set_table_tick
12071 for each register mentioned. Similar to mention_regs in cse.c */
12073 static void
12074 update_table_tick (rtx x)
12076 enum rtx_code code = GET_CODE (x);
12077 const char *fmt = GET_RTX_FORMAT (code);
12078 int i, j;
12080 if (code == REG)
12082 unsigned int regno = REGNO (x);
12083 unsigned int endregno = END_REGNO (x);
12084 unsigned int r;
12086 for (r = regno; r < endregno; r++)
12088 reg_stat_type *rsp = &reg_stat[r];
12089 rsp->last_set_table_tick = label_tick;
12092 return;
12095 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
12096 if (fmt[i] == 'e')
12098 /* Check for identical subexpressions. If x contains
12099 identical subexpression we only have to traverse one of
12100 them. */
12101 if (i == 0 && ARITHMETIC_P (x))
12103 /* Note that at this point x1 has already been
12104 processed. */
12105 rtx x0 = XEXP (x, 0);
12106 rtx x1 = XEXP (x, 1);
12108 /* If x0 and x1 are identical then there is no need to
12109 process x0. */
12110 if (x0 == x1)
12111 break;
12113 /* If x0 is identical to a subexpression of x1 then while
12114 processing x1, x0 has already been processed. Thus we
12115 are done with x. */
12116 if (ARITHMETIC_P (x1)
12117 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
12118 break;
12120 /* If x1 is identical to a subexpression of x0 then we
12121 still have to process the rest of x0. */
12122 if (ARITHMETIC_P (x0)
12123 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
12125 update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0));
12126 break;
12130 update_table_tick (XEXP (x, i));
12132 else if (fmt[i] == 'E')
12133 for (j = 0; j < XVECLEN (x, i); j++)
12134 update_table_tick (XVECEXP (x, i, j));
12137 /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we
12138 are saying that the register is clobbered and we no longer know its
12139 value. If INSN is zero, don't update reg_stat[].last_set; this is
12140 only permitted with VALUE also zero and is used to invalidate the
12141 register. */
12143 static void
12144 record_value_for_reg (rtx reg, rtx insn, rtx value)
12146 unsigned int regno = REGNO (reg);
12147 unsigned int endregno = END_REGNO (reg);
12148 unsigned int i;
12149 reg_stat_type *rsp;
12151 /* If VALUE contains REG and we have a previous value for REG, substitute
12152 the previous value. */
12153 if (value && insn && reg_overlap_mentioned_p (reg, value))
12155 rtx tem;
12157 /* Set things up so get_last_value is allowed to see anything set up to
12158 our insn. */
12159 subst_low_luid = DF_INSN_LUID (insn);
12160 tem = get_last_value (reg);
12162 /* If TEM is simply a binary operation with two CLOBBERs as operands,
12163 it isn't going to be useful and will take a lot of time to process,
12164 so just use the CLOBBER. */
12166 if (tem)
12168 if (ARITHMETIC_P (tem)
12169 && GET_CODE (XEXP (tem, 0)) == CLOBBER
12170 && GET_CODE (XEXP (tem, 1)) == CLOBBER)
12171 tem = XEXP (tem, 0);
12172 else if (count_occurrences (value, reg, 1) >= 2)
12174 /* If there are two or more occurrences of REG in VALUE,
12175 prevent the value from growing too much. */
12176 if (count_rtxs (tem) > MAX_LAST_VALUE_RTL)
12177 tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx);
12180 value = replace_rtx (copy_rtx (value), reg, tem);
12184 /* For each register modified, show we don't know its value, that
12185 we don't know about its bitwise content, that its value has been
12186 updated, and that we don't know the location of the death of the
12187 register. */
12188 for (i = regno; i < endregno; i++)
12190 rsp = &reg_stat[i];
12192 if (insn)
12193 rsp->last_set = insn;
12195 rsp->last_set_value = 0;
12196 rsp->last_set_mode = VOIDmode;
12197 rsp->last_set_nonzero_bits = 0;
12198 rsp->last_set_sign_bit_copies = 0;
12199 rsp->last_death = 0;
12200 rsp->truncated_to_mode = VOIDmode;
12203 /* Mark registers that are being referenced in this value. */
12204 if (value)
12205 update_table_tick (value);
12207 /* Now update the status of each register being set.
12208 If someone is using this register in this block, set this register
12209 to invalid since we will get confused between the two lives in this
12210 basic block. This makes using this register always invalid. In cse, we
12211 scan the table to invalidate all entries using this register, but this
12212 is too much work for us. */
12214 for (i = regno; i < endregno; i++)
12216 rsp = &reg_stat[i];
12217 rsp->last_set_label = label_tick;
12218 if (!insn
12219 || (value && rsp->last_set_table_tick >= label_tick_ebb_start))
12220 rsp->last_set_invalid = 1;
12221 else
12222 rsp->last_set_invalid = 0;
12225 /* The value being assigned might refer to X (like in "x++;"). In that
12226 case, we must replace it with (clobber (const_int 0)) to prevent
12227 infinite loops. */
12228 rsp = &reg_stat[regno];
12229 if (value && !get_last_value_validate (&value, insn, label_tick, 0))
12231 value = copy_rtx (value);
12232 if (!get_last_value_validate (&value, insn, label_tick, 1))
12233 value = 0;
12236 /* For the main register being modified, update the value, the mode, the
12237 nonzero bits, and the number of sign bit copies. */
12239 rsp->last_set_value = value;
12241 if (value)
12243 enum machine_mode mode = GET_MODE (reg);
12244 subst_low_luid = DF_INSN_LUID (insn);
12245 rsp->last_set_mode = mode;
12246 if (GET_MODE_CLASS (mode) == MODE_INT
12247 && HWI_COMPUTABLE_MODE_P (mode))
12248 mode = nonzero_bits_mode;
12249 rsp->last_set_nonzero_bits = nonzero_bits (value, mode);
12250 rsp->last_set_sign_bit_copies
12251 = num_sign_bit_copies (value, GET_MODE (reg));
12255 /* Called via note_stores from record_dead_and_set_regs to handle one
12256 SET or CLOBBER in an insn. DATA is the instruction in which the
12257 set is occurring. */
12259 static void
12260 record_dead_and_set_regs_1 (rtx dest, const_rtx setter, void *data)
12262 rtx record_dead_insn = (rtx) data;
12264 if (GET_CODE (dest) == SUBREG)
12265 dest = SUBREG_REG (dest);
12267 if (!record_dead_insn)
12269 if (REG_P (dest))
12270 record_value_for_reg (dest, NULL_RTX, NULL_RTX);
12271 return;
12274 if (REG_P (dest))
12276 /* If we are setting the whole register, we know its value. Otherwise
12277 show that we don't know the value. We can handle SUBREG in
12278 some cases. */
12279 if (GET_CODE (setter) == SET && dest == SET_DEST (setter))
12280 record_value_for_reg (dest, record_dead_insn, SET_SRC (setter));
12281 else if (GET_CODE (setter) == SET
12282 && GET_CODE (SET_DEST (setter)) == SUBREG
12283 && SUBREG_REG (SET_DEST (setter)) == dest
12284 && GET_MODE_PRECISION (GET_MODE (dest)) <= BITS_PER_WORD
12285 && subreg_lowpart_p (SET_DEST (setter)))
12286 record_value_for_reg (dest, record_dead_insn,
12287 gen_lowpart (GET_MODE (dest),
12288 SET_SRC (setter)));
12289 else
12290 record_value_for_reg (dest, record_dead_insn, NULL_RTX);
12292 else if (MEM_P (dest)
12293 /* Ignore pushes, they clobber nothing. */
12294 && ! push_operand (dest, GET_MODE (dest)))
12295 mem_last_set = DF_INSN_LUID (record_dead_insn);
12298 /* Update the records of when each REG was most recently set or killed
12299 for the things done by INSN. This is the last thing done in processing
12300 INSN in the combiner loop.
12302 We update reg_stat[], in particular fields last_set, last_set_value,
12303 last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies,
12304 last_death, and also the similar information mem_last_set (which insn
12305 most recently modified memory) and last_call_luid (which insn was the
12306 most recent subroutine call). */
12308 static void
12309 record_dead_and_set_regs (rtx insn)
12311 rtx link;
12312 unsigned int i;
12314 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
12316 if (REG_NOTE_KIND (link) == REG_DEAD
12317 && REG_P (XEXP (link, 0)))
12319 unsigned int regno = REGNO (XEXP (link, 0));
12320 unsigned int endregno = END_REGNO (XEXP (link, 0));
12322 for (i = regno; i < endregno; i++)
12324 reg_stat_type *rsp;
12326 rsp = &reg_stat[i];
12327 rsp->last_death = insn;
12330 else if (REG_NOTE_KIND (link) == REG_INC)
12331 record_value_for_reg (XEXP (link, 0), insn, NULL_RTX);
12334 if (CALL_P (insn))
12336 hard_reg_set_iterator hrsi;
12337 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, i, hrsi)
12339 reg_stat_type *rsp;
12341 rsp = &reg_stat[i];
12342 rsp->last_set_invalid = 1;
12343 rsp->last_set = insn;
12344 rsp->last_set_value = 0;
12345 rsp->last_set_mode = VOIDmode;
12346 rsp->last_set_nonzero_bits = 0;
12347 rsp->last_set_sign_bit_copies = 0;
12348 rsp->last_death = 0;
12349 rsp->truncated_to_mode = VOIDmode;
12352 last_call_luid = mem_last_set = DF_INSN_LUID (insn);
12354 /* We can't combine into a call pattern. Remember, though, that
12355 the return value register is set at this LUID. We could
12356 still replace a register with the return value from the
12357 wrong subroutine call! */
12358 note_stores (PATTERN (insn), record_dead_and_set_regs_1, NULL_RTX);
12360 else
12361 note_stores (PATTERN (insn), record_dead_and_set_regs_1, insn);
12364 /* If a SUBREG has the promoted bit set, it is in fact a property of the
12365 register present in the SUBREG, so for each such SUBREG go back and
12366 adjust nonzero and sign bit information of the registers that are
12367 known to have some zero/sign bits set.
12369 This is needed because when combine blows the SUBREGs away, the
12370 information on zero/sign bits is lost and further combines can be
12371 missed because of that. */
12373 static void
12374 record_promoted_value (rtx insn, rtx subreg)
12376 struct insn_link *links;
12377 rtx set;
12378 unsigned int regno = REGNO (SUBREG_REG (subreg));
12379 enum machine_mode mode = GET_MODE (subreg);
12381 if (GET_MODE_PRECISION (mode) > HOST_BITS_PER_WIDE_INT)
12382 return;
12384 for (links = LOG_LINKS (insn); links;)
12386 reg_stat_type *rsp;
12388 insn = links->insn;
12389 set = single_set (insn);
12391 if (! set || !REG_P (SET_DEST (set))
12392 || REGNO (SET_DEST (set)) != regno
12393 || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg)))
12395 links = links->next;
12396 continue;
12399 rsp = &reg_stat[regno];
12400 if (rsp->last_set == insn)
12402 if (SUBREG_PROMOTED_UNSIGNED_P (subreg) > 0)
12403 rsp->last_set_nonzero_bits &= GET_MODE_MASK (mode);
12406 if (REG_P (SET_SRC (set)))
12408 regno = REGNO (SET_SRC (set));
12409 links = LOG_LINKS (insn);
12411 else
12412 break;
12416 /* Check if X, a register, is known to contain a value already
12417 truncated to MODE. In this case we can use a subreg to refer to
12418 the truncated value even though in the generic case we would need
12419 an explicit truncation. */
12421 static bool
12422 reg_truncated_to_mode (enum machine_mode mode, const_rtx x)
12424 reg_stat_type *rsp = &reg_stat[REGNO (x)];
12425 enum machine_mode truncated = rsp->truncated_to_mode;
12427 if (truncated == 0
12428 || rsp->truncation_label < label_tick_ebb_start)
12429 return false;
12430 if (GET_MODE_SIZE (truncated) <= GET_MODE_SIZE (mode))
12431 return true;
12432 if (TRULY_NOOP_TRUNCATION_MODES_P (mode, truncated))
12433 return true;
12434 return false;
12437 /* Callback for for_each_rtx. If *P is a hard reg or a subreg record the mode
12438 that the register is accessed in. For non-TRULY_NOOP_TRUNCATION targets we
12439 might be able to turn a truncate into a subreg using this information.
12440 Return -1 if traversing *P is complete or 0 otherwise. */
12442 static int
12443 record_truncated_value (rtx *p, void *data ATTRIBUTE_UNUSED)
12445 rtx x = *p;
12446 enum machine_mode truncated_mode;
12447 reg_stat_type *rsp;
12449 if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)))
12451 enum machine_mode original_mode = GET_MODE (SUBREG_REG (x));
12452 truncated_mode = GET_MODE (x);
12454 if (GET_MODE_SIZE (original_mode) <= GET_MODE_SIZE (truncated_mode))
12455 return -1;
12457 if (TRULY_NOOP_TRUNCATION_MODES_P (truncated_mode, original_mode))
12458 return -1;
12460 x = SUBREG_REG (x);
12462 /* ??? For hard-regs we now record everything. We might be able to
12463 optimize this using last_set_mode. */
12464 else if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
12465 truncated_mode = GET_MODE (x);
12466 else
12467 return 0;
12469 rsp = &reg_stat[REGNO (x)];
12470 if (rsp->truncated_to_mode == 0
12471 || rsp->truncation_label < label_tick_ebb_start
12472 || (GET_MODE_SIZE (truncated_mode)
12473 < GET_MODE_SIZE (rsp->truncated_to_mode)))
12475 rsp->truncated_to_mode = truncated_mode;
12476 rsp->truncation_label = label_tick;
12479 return -1;
12482 /* Callback for note_uses. Find hardregs and subregs of pseudos and
12483 the modes they are used in. This can help truning TRUNCATEs into
12484 SUBREGs. */
12486 static void
12487 record_truncated_values (rtx *x, void *data ATTRIBUTE_UNUSED)
12489 for_each_rtx (x, record_truncated_value, NULL);
12492 /* Scan X for promoted SUBREGs. For each one found,
12493 note what it implies to the registers used in it. */
12495 static void
12496 check_promoted_subreg (rtx insn, rtx x)
12498 if (GET_CODE (x) == SUBREG
12499 && SUBREG_PROMOTED_VAR_P (x)
12500 && REG_P (SUBREG_REG (x)))
12501 record_promoted_value (insn, x);
12502 else
12504 const char *format = GET_RTX_FORMAT (GET_CODE (x));
12505 int i, j;
12507 for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++)
12508 switch (format[i])
12510 case 'e':
12511 check_promoted_subreg (insn, XEXP (x, i));
12512 break;
12513 case 'V':
12514 case 'E':
12515 if (XVEC (x, i) != 0)
12516 for (j = 0; j < XVECLEN (x, i); j++)
12517 check_promoted_subreg (insn, XVECEXP (x, i, j));
12518 break;
12523 /* Verify that all the registers and memory references mentioned in *LOC are
12524 still valid. *LOC was part of a value set in INSN when label_tick was
12525 equal to TICK. Return 0 if some are not. If REPLACE is nonzero, replace
12526 the invalid references with (clobber (const_int 0)) and return 1. This
12527 replacement is useful because we often can get useful information about
12528 the form of a value (e.g., if it was produced by a shift that always
12529 produces -1 or 0) even though we don't know exactly what registers it
12530 was produced from. */
12532 static int
12533 get_last_value_validate (rtx *loc, rtx insn, int tick, int replace)
12535 rtx x = *loc;
12536 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
12537 int len = GET_RTX_LENGTH (GET_CODE (x));
12538 int i, j;
12540 if (REG_P (x))
12542 unsigned int regno = REGNO (x);
12543 unsigned int endregno = END_REGNO (x);
12544 unsigned int j;
12546 for (j = regno; j < endregno; j++)
12548 reg_stat_type *rsp = &reg_stat[j];
12549 if (rsp->last_set_invalid
12550 /* If this is a pseudo-register that was only set once and not
12551 live at the beginning of the function, it is always valid. */
12552 || (! (regno >= FIRST_PSEUDO_REGISTER
12553 && REG_N_SETS (regno) == 1
12554 && (!REGNO_REG_SET_P
12555 (DF_LR_IN (ENTRY_BLOCK_PTR->next_bb), regno)))
12556 && rsp->last_set_label > tick))
12558 if (replace)
12559 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
12560 return replace;
12564 return 1;
12566 /* If this is a memory reference, make sure that there were no stores after
12567 it that might have clobbered the value. We don't have alias info, so we
12568 assume any store invalidates it. Moreover, we only have local UIDs, so
12569 we also assume that there were stores in the intervening basic blocks. */
12570 else if (MEM_P (x) && !MEM_READONLY_P (x)
12571 && (tick != label_tick || DF_INSN_LUID (insn) <= mem_last_set))
12573 if (replace)
12574 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
12575 return replace;
12578 for (i = 0; i < len; i++)
12580 if (fmt[i] == 'e')
12582 /* Check for identical subexpressions. If x contains
12583 identical subexpression we only have to traverse one of
12584 them. */
12585 if (i == 1 && ARITHMETIC_P (x))
12587 /* Note that at this point x0 has already been checked
12588 and found valid. */
12589 rtx x0 = XEXP (x, 0);
12590 rtx x1 = XEXP (x, 1);
12592 /* If x0 and x1 are identical then x is also valid. */
12593 if (x0 == x1)
12594 return 1;
12596 /* If x1 is identical to a subexpression of x0 then
12597 while checking x0, x1 has already been checked. Thus
12598 it is valid and so as x. */
12599 if (ARITHMETIC_P (x0)
12600 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
12601 return 1;
12603 /* If x0 is identical to a subexpression of x1 then x is
12604 valid iff the rest of x1 is valid. */
12605 if (ARITHMETIC_P (x1)
12606 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
12607 return
12608 get_last_value_validate (&XEXP (x1,
12609 x0 == XEXP (x1, 0) ? 1 : 0),
12610 insn, tick, replace);
12613 if (get_last_value_validate (&XEXP (x, i), insn, tick,
12614 replace) == 0)
12615 return 0;
12617 else if (fmt[i] == 'E')
12618 for (j = 0; j < XVECLEN (x, i); j++)
12619 if (get_last_value_validate (&XVECEXP (x, i, j),
12620 insn, tick, replace) == 0)
12621 return 0;
12624 /* If we haven't found a reason for it to be invalid, it is valid. */
12625 return 1;
12628 /* Get the last value assigned to X, if known. Some registers
12629 in the value may be replaced with (clobber (const_int 0)) if their value
12630 is known longer known reliably. */
12632 static rtx
12633 get_last_value (const_rtx x)
12635 unsigned int regno;
12636 rtx value;
12637 reg_stat_type *rsp;
12639 /* If this is a non-paradoxical SUBREG, get the value of its operand and
12640 then convert it to the desired mode. If this is a paradoxical SUBREG,
12641 we cannot predict what values the "extra" bits might have. */
12642 if (GET_CODE (x) == SUBREG
12643 && subreg_lowpart_p (x)
12644 && !paradoxical_subreg_p (x)
12645 && (value = get_last_value (SUBREG_REG (x))) != 0)
12646 return gen_lowpart (GET_MODE (x), value);
12648 if (!REG_P (x))
12649 return 0;
12651 regno = REGNO (x);
12652 rsp = &reg_stat[regno];
12653 value = rsp->last_set_value;
12655 /* If we don't have a value, or if it isn't for this basic block and
12656 it's either a hard register, set more than once, or it's a live
12657 at the beginning of the function, return 0.
12659 Because if it's not live at the beginning of the function then the reg
12660 is always set before being used (is never used without being set).
12661 And, if it's set only once, and it's always set before use, then all
12662 uses must have the same last value, even if it's not from this basic
12663 block. */
12665 if (value == 0
12666 || (rsp->last_set_label < label_tick_ebb_start
12667 && (regno < FIRST_PSEUDO_REGISTER
12668 || REG_N_SETS (regno) != 1
12669 || REGNO_REG_SET_P
12670 (DF_LR_IN (ENTRY_BLOCK_PTR->next_bb), regno))))
12671 return 0;
12673 /* If the value was set in a later insn than the ones we are processing,
12674 we can't use it even if the register was only set once. */
12675 if (rsp->last_set_label == label_tick
12676 && DF_INSN_LUID (rsp->last_set) >= subst_low_luid)
12677 return 0;
12679 /* If the value has all its registers valid, return it. */
12680 if (get_last_value_validate (&value, rsp->last_set, rsp->last_set_label, 0))
12681 return value;
12683 /* Otherwise, make a copy and replace any invalid register with
12684 (clobber (const_int 0)). If that fails for some reason, return 0. */
12686 value = copy_rtx (value);
12687 if (get_last_value_validate (&value, rsp->last_set, rsp->last_set_label, 1))
12688 return value;
12690 return 0;
12693 /* Return nonzero if expression X refers to a REG or to memory
12694 that is set in an instruction more recent than FROM_LUID. */
12696 static int
12697 use_crosses_set_p (const_rtx x, int from_luid)
12699 const char *fmt;
12700 int i;
12701 enum rtx_code code = GET_CODE (x);
12703 if (code == REG)
12705 unsigned int regno = REGNO (x);
12706 unsigned endreg = END_REGNO (x);
12708 #ifdef PUSH_ROUNDING
12709 /* Don't allow uses of the stack pointer to be moved,
12710 because we don't know whether the move crosses a push insn. */
12711 if (regno == STACK_POINTER_REGNUM && PUSH_ARGS)
12712 return 1;
12713 #endif
12714 for (; regno < endreg; regno++)
12716 reg_stat_type *rsp = &reg_stat[regno];
12717 if (rsp->last_set
12718 && rsp->last_set_label == label_tick
12719 && DF_INSN_LUID (rsp->last_set) > from_luid)
12720 return 1;
12722 return 0;
12725 if (code == MEM && mem_last_set > from_luid)
12726 return 1;
12728 fmt = GET_RTX_FORMAT (code);
12730 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
12732 if (fmt[i] == 'E')
12734 int j;
12735 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
12736 if (use_crosses_set_p (XVECEXP (x, i, j), from_luid))
12737 return 1;
12739 else if (fmt[i] == 'e'
12740 && use_crosses_set_p (XEXP (x, i), from_luid))
12741 return 1;
12743 return 0;
12746 /* Define three variables used for communication between the following
12747 routines. */
12749 static unsigned int reg_dead_regno, reg_dead_endregno;
12750 static int reg_dead_flag;
12752 /* Function called via note_stores from reg_dead_at_p.
12754 If DEST is within [reg_dead_regno, reg_dead_endregno), set
12755 reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */
12757 static void
12758 reg_dead_at_p_1 (rtx dest, const_rtx x, void *data ATTRIBUTE_UNUSED)
12760 unsigned int regno, endregno;
12762 if (!REG_P (dest))
12763 return;
12765 regno = REGNO (dest);
12766 endregno = END_REGNO (dest);
12767 if (reg_dead_endregno > regno && reg_dead_regno < endregno)
12768 reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1;
12771 /* Return nonzero if REG is known to be dead at INSN.
12773 We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER
12774 referencing REG, it is dead. If we hit a SET referencing REG, it is
12775 live. Otherwise, see if it is live or dead at the start of the basic
12776 block we are in. Hard regs marked as being live in NEWPAT_USED_REGS
12777 must be assumed to be always live. */
12779 static int
12780 reg_dead_at_p (rtx reg, rtx insn)
12782 basic_block block;
12783 unsigned int i;
12785 /* Set variables for reg_dead_at_p_1. */
12786 reg_dead_regno = REGNO (reg);
12787 reg_dead_endregno = END_REGNO (reg);
12789 reg_dead_flag = 0;
12791 /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers
12792 we allow the machine description to decide whether use-and-clobber
12793 patterns are OK. */
12794 if (reg_dead_regno < FIRST_PSEUDO_REGISTER)
12796 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
12797 if (!fixed_regs[i] && TEST_HARD_REG_BIT (newpat_used_regs, i))
12798 return 0;
12801 /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, or
12802 beginning of basic block. */
12803 block = BLOCK_FOR_INSN (insn);
12804 for (;;)
12806 if (INSN_P (insn))
12808 note_stores (PATTERN (insn), reg_dead_at_p_1, NULL);
12809 if (reg_dead_flag)
12810 return reg_dead_flag == 1 ? 1 : 0;
12812 if (find_regno_note (insn, REG_DEAD, reg_dead_regno))
12813 return 1;
12816 if (insn == BB_HEAD (block))
12817 break;
12819 insn = PREV_INSN (insn);
12822 /* Look at live-in sets for the basic block that we were in. */
12823 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
12824 if (REGNO_REG_SET_P (df_get_live_in (block), i))
12825 return 0;
12827 return 1;
12830 /* Note hard registers in X that are used. */
12832 static void
12833 mark_used_regs_combine (rtx x)
12835 RTX_CODE code = GET_CODE (x);
12836 unsigned int regno;
12837 int i;
12839 switch (code)
12841 case LABEL_REF:
12842 case SYMBOL_REF:
12843 case CONST:
12844 CASE_CONST_ANY:
12845 case PC:
12846 case ADDR_VEC:
12847 case ADDR_DIFF_VEC:
12848 case ASM_INPUT:
12849 #ifdef HAVE_cc0
12850 /* CC0 must die in the insn after it is set, so we don't need to take
12851 special note of it here. */
12852 case CC0:
12853 #endif
12854 return;
12856 case CLOBBER:
12857 /* If we are clobbering a MEM, mark any hard registers inside the
12858 address as used. */
12859 if (MEM_P (XEXP (x, 0)))
12860 mark_used_regs_combine (XEXP (XEXP (x, 0), 0));
12861 return;
12863 case REG:
12864 regno = REGNO (x);
12865 /* A hard reg in a wide mode may really be multiple registers.
12866 If so, mark all of them just like the first. */
12867 if (regno < FIRST_PSEUDO_REGISTER)
12869 /* None of this applies to the stack, frame or arg pointers. */
12870 if (regno == STACK_POINTER_REGNUM
12871 #if !HARD_FRAME_POINTER_IS_FRAME_POINTER
12872 || regno == HARD_FRAME_POINTER_REGNUM
12873 #endif
12874 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
12875 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
12876 #endif
12877 || regno == FRAME_POINTER_REGNUM)
12878 return;
12880 add_to_hard_reg_set (&newpat_used_regs, GET_MODE (x), regno);
12882 return;
12884 case SET:
12886 /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in
12887 the address. */
12888 rtx testreg = SET_DEST (x);
12890 while (GET_CODE (testreg) == SUBREG
12891 || GET_CODE (testreg) == ZERO_EXTRACT
12892 || GET_CODE (testreg) == STRICT_LOW_PART)
12893 testreg = XEXP (testreg, 0);
12895 if (MEM_P (testreg))
12896 mark_used_regs_combine (XEXP (testreg, 0));
12898 mark_used_regs_combine (SET_SRC (x));
12900 return;
12902 default:
12903 break;
12906 /* Recursively scan the operands of this expression. */
12909 const char *fmt = GET_RTX_FORMAT (code);
12911 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
12913 if (fmt[i] == 'e')
12914 mark_used_regs_combine (XEXP (x, i));
12915 else if (fmt[i] == 'E')
12917 int j;
12919 for (j = 0; j < XVECLEN (x, i); j++)
12920 mark_used_regs_combine (XVECEXP (x, i, j));
12926 /* Remove register number REGNO from the dead registers list of INSN.
12928 Return the note used to record the death, if there was one. */
12931 remove_death (unsigned int regno, rtx insn)
12933 rtx note = find_regno_note (insn, REG_DEAD, regno);
12935 if (note)
12936 remove_note (insn, note);
12938 return note;
12941 /* For each register (hardware or pseudo) used within expression X, if its
12942 death is in an instruction with luid between FROM_LUID (inclusive) and
12943 TO_INSN (exclusive), put a REG_DEAD note for that register in the
12944 list headed by PNOTES.
12946 That said, don't move registers killed by maybe_kill_insn.
12948 This is done when X is being merged by combination into TO_INSN. These
12949 notes will then be distributed as needed. */
12951 static void
12952 move_deaths (rtx x, rtx maybe_kill_insn, int from_luid, rtx to_insn,
12953 rtx *pnotes)
12955 const char *fmt;
12956 int len, i;
12957 enum rtx_code code = GET_CODE (x);
12959 if (code == REG)
12961 unsigned int regno = REGNO (x);
12962 rtx where_dead = reg_stat[regno].last_death;
12964 /* Don't move the register if it gets killed in between from and to. */
12965 if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn)
12966 && ! reg_referenced_p (x, maybe_kill_insn))
12967 return;
12969 if (where_dead
12970 && BLOCK_FOR_INSN (where_dead) == BLOCK_FOR_INSN (to_insn)
12971 && DF_INSN_LUID (where_dead) >= from_luid
12972 && DF_INSN_LUID (where_dead) < DF_INSN_LUID (to_insn))
12974 rtx note = remove_death (regno, where_dead);
12976 /* It is possible for the call above to return 0. This can occur
12977 when last_death points to I2 or I1 that we combined with.
12978 In that case make a new note.
12980 We must also check for the case where X is a hard register
12981 and NOTE is a death note for a range of hard registers
12982 including X. In that case, we must put REG_DEAD notes for
12983 the remaining registers in place of NOTE. */
12985 if (note != 0 && regno < FIRST_PSEUDO_REGISTER
12986 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
12987 > GET_MODE_SIZE (GET_MODE (x))))
12989 unsigned int deadregno = REGNO (XEXP (note, 0));
12990 unsigned int deadend = END_HARD_REGNO (XEXP (note, 0));
12991 unsigned int ourend = END_HARD_REGNO (x);
12992 unsigned int i;
12994 for (i = deadregno; i < deadend; i++)
12995 if (i < regno || i >= ourend)
12996 add_reg_note (where_dead, REG_DEAD, regno_reg_rtx[i]);
12999 /* If we didn't find any note, or if we found a REG_DEAD note that
13000 covers only part of the given reg, and we have a multi-reg hard
13001 register, then to be safe we must check for REG_DEAD notes
13002 for each register other than the first. They could have
13003 their own REG_DEAD notes lying around. */
13004 else if ((note == 0
13005 || (note != 0
13006 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
13007 < GET_MODE_SIZE (GET_MODE (x)))))
13008 && regno < FIRST_PSEUDO_REGISTER
13009 && hard_regno_nregs[regno][GET_MODE (x)] > 1)
13011 unsigned int ourend = END_HARD_REGNO (x);
13012 unsigned int i, offset;
13013 rtx oldnotes = 0;
13015 if (note)
13016 offset = hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))];
13017 else
13018 offset = 1;
13020 for (i = regno + offset; i < ourend; i++)
13021 move_deaths (regno_reg_rtx[i],
13022 maybe_kill_insn, from_luid, to_insn, &oldnotes);
13025 if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x))
13027 XEXP (note, 1) = *pnotes;
13028 *pnotes = note;
13030 else
13031 *pnotes = alloc_reg_note (REG_DEAD, x, *pnotes);
13034 return;
13037 else if (GET_CODE (x) == SET)
13039 rtx dest = SET_DEST (x);
13041 move_deaths (SET_SRC (x), maybe_kill_insn, from_luid, to_insn, pnotes);
13043 /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG
13044 that accesses one word of a multi-word item, some
13045 piece of everything register in the expression is used by
13046 this insn, so remove any old death. */
13047 /* ??? So why do we test for equality of the sizes? */
13049 if (GET_CODE (dest) == ZERO_EXTRACT
13050 || GET_CODE (dest) == STRICT_LOW_PART
13051 || (GET_CODE (dest) == SUBREG
13052 && (((GET_MODE_SIZE (GET_MODE (dest))
13053 + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
13054 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))
13055 + UNITS_PER_WORD - 1) / UNITS_PER_WORD))))
13057 move_deaths (dest, maybe_kill_insn, from_luid, to_insn, pnotes);
13058 return;
13061 /* If this is some other SUBREG, we know it replaces the entire
13062 value, so use that as the destination. */
13063 if (GET_CODE (dest) == SUBREG)
13064 dest = SUBREG_REG (dest);
13066 /* If this is a MEM, adjust deaths of anything used in the address.
13067 For a REG (the only other possibility), the entire value is
13068 being replaced so the old value is not used in this insn. */
13070 if (MEM_P (dest))
13071 move_deaths (XEXP (dest, 0), maybe_kill_insn, from_luid,
13072 to_insn, pnotes);
13073 return;
13076 else if (GET_CODE (x) == CLOBBER)
13077 return;
13079 len = GET_RTX_LENGTH (code);
13080 fmt = GET_RTX_FORMAT (code);
13082 for (i = 0; i < len; i++)
13084 if (fmt[i] == 'E')
13086 int j;
13087 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
13088 move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_luid,
13089 to_insn, pnotes);
13091 else if (fmt[i] == 'e')
13092 move_deaths (XEXP (x, i), maybe_kill_insn, from_luid, to_insn, pnotes);
13096 /* Return 1 if X is the target of a bit-field assignment in BODY, the
13097 pattern of an insn. X must be a REG. */
13099 static int
13100 reg_bitfield_target_p (rtx x, rtx body)
13102 int i;
13104 if (GET_CODE (body) == SET)
13106 rtx dest = SET_DEST (body);
13107 rtx target;
13108 unsigned int regno, tregno, endregno, endtregno;
13110 if (GET_CODE (dest) == ZERO_EXTRACT)
13111 target = XEXP (dest, 0);
13112 else if (GET_CODE (dest) == STRICT_LOW_PART)
13113 target = SUBREG_REG (XEXP (dest, 0));
13114 else
13115 return 0;
13117 if (GET_CODE (target) == SUBREG)
13118 target = SUBREG_REG (target);
13120 if (!REG_P (target))
13121 return 0;
13123 tregno = REGNO (target), regno = REGNO (x);
13124 if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER)
13125 return target == x;
13127 endtregno = end_hard_regno (GET_MODE (target), tregno);
13128 endregno = end_hard_regno (GET_MODE (x), regno);
13130 return endregno > tregno && regno < endtregno;
13133 else if (GET_CODE (body) == PARALLEL)
13134 for (i = XVECLEN (body, 0) - 1; i >= 0; i--)
13135 if (reg_bitfield_target_p (x, XVECEXP (body, 0, i)))
13136 return 1;
13138 return 0;
13141 /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them
13142 as appropriate. I3 and I2 are the insns resulting from the combination
13143 insns including FROM (I2 may be zero).
13145 ELIM_I2 and ELIM_I1 are either zero or registers that we know will
13146 not need REG_DEAD notes because they are being substituted for. This
13147 saves searching in the most common cases.
13149 Each note in the list is either ignored or placed on some insns, depending
13150 on the type of note. */
13152 static void
13153 distribute_notes (rtx notes, rtx from_insn, rtx i3, rtx i2, rtx elim_i2,
13154 rtx elim_i1, rtx elim_i0)
13156 rtx note, next_note;
13157 rtx tem;
13159 for (note = notes; note; note = next_note)
13161 rtx place = 0, place2 = 0;
13163 next_note = XEXP (note, 1);
13164 switch (REG_NOTE_KIND (note))
13166 case REG_BR_PROB:
13167 case REG_BR_PRED:
13168 /* Doesn't matter much where we put this, as long as it's somewhere.
13169 It is preferable to keep these notes on branches, which is most
13170 likely to be i3. */
13171 place = i3;
13172 break;
13174 case REG_NON_LOCAL_GOTO:
13175 if (JUMP_P (i3))
13176 place = i3;
13177 else
13179 gcc_assert (i2 && JUMP_P (i2));
13180 place = i2;
13182 break;
13184 case REG_EH_REGION:
13185 /* These notes must remain with the call or trapping instruction. */
13186 if (CALL_P (i3))
13187 place = i3;
13188 else if (i2 && CALL_P (i2))
13189 place = i2;
13190 else
13192 gcc_assert (cfun->can_throw_non_call_exceptions);
13193 if (may_trap_p (i3))
13194 place = i3;
13195 else if (i2 && may_trap_p (i2))
13196 place = i2;
13197 /* ??? Otherwise assume we've combined things such that we
13198 can now prove that the instructions can't trap. Drop the
13199 note in this case. */
13201 break;
13203 case REG_ARGS_SIZE:
13204 /* ??? How to distribute between i3-i1. Assume i3 contains the
13205 entire adjustment. Assert i3 contains at least some adjust. */
13206 if (!noop_move_p (i3))
13208 int old_size, args_size = INTVAL (XEXP (note, 0));
13209 /* fixup_args_size_notes looks at REG_NORETURN note,
13210 so ensure the note is placed there first. */
13211 if (CALL_P (i3))
13213 rtx *np;
13214 for (np = &next_note; *np; np = &XEXP (*np, 1))
13215 if (REG_NOTE_KIND (*np) == REG_NORETURN)
13217 rtx n = *np;
13218 *np = XEXP (n, 1);
13219 XEXP (n, 1) = REG_NOTES (i3);
13220 REG_NOTES (i3) = n;
13221 break;
13224 old_size = fixup_args_size_notes (PREV_INSN (i3), i3, args_size);
13225 /* emit_call_1 adds for !ACCUMULATE_OUTGOING_ARGS
13226 REG_ARGS_SIZE note to all noreturn calls, allow that here. */
13227 gcc_assert (old_size != args_size
13228 || (CALL_P (i3)
13229 && !ACCUMULATE_OUTGOING_ARGS
13230 && find_reg_note (i3, REG_NORETURN, NULL_RTX)));
13232 break;
13234 case REG_NORETURN:
13235 case REG_SETJMP:
13236 case REG_TM:
13237 /* These notes must remain with the call. It should not be
13238 possible for both I2 and I3 to be a call. */
13239 if (CALL_P (i3))
13240 place = i3;
13241 else
13243 gcc_assert (i2 && CALL_P (i2));
13244 place = i2;
13246 break;
13248 case REG_UNUSED:
13249 /* Any clobbers for i3 may still exist, and so we must process
13250 REG_UNUSED notes from that insn.
13252 Any clobbers from i2 or i1 can only exist if they were added by
13253 recog_for_combine. In that case, recog_for_combine created the
13254 necessary REG_UNUSED notes. Trying to keep any original
13255 REG_UNUSED notes from these insns can cause incorrect output
13256 if it is for the same register as the original i3 dest.
13257 In that case, we will notice that the register is set in i3,
13258 and then add a REG_UNUSED note for the destination of i3, which
13259 is wrong. However, it is possible to have REG_UNUSED notes from
13260 i2 or i1 for register which were both used and clobbered, so
13261 we keep notes from i2 or i1 if they will turn into REG_DEAD
13262 notes. */
13264 /* If this register is set or clobbered in I3, put the note there
13265 unless there is one already. */
13266 if (reg_set_p (XEXP (note, 0), PATTERN (i3)))
13268 if (from_insn != i3)
13269 break;
13271 if (! (REG_P (XEXP (note, 0))
13272 ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0)))
13273 : find_reg_note (i3, REG_UNUSED, XEXP (note, 0))))
13274 place = i3;
13276 /* Otherwise, if this register is used by I3, then this register
13277 now dies here, so we must put a REG_DEAD note here unless there
13278 is one already. */
13279 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))
13280 && ! (REG_P (XEXP (note, 0))
13281 ? find_regno_note (i3, REG_DEAD,
13282 REGNO (XEXP (note, 0)))
13283 : find_reg_note (i3, REG_DEAD, XEXP (note, 0))))
13285 PUT_REG_NOTE_KIND (note, REG_DEAD);
13286 place = i3;
13288 break;
13290 case REG_EQUAL:
13291 case REG_EQUIV:
13292 case REG_NOALIAS:
13293 /* These notes say something about results of an insn. We can
13294 only support them if they used to be on I3 in which case they
13295 remain on I3. Otherwise they are ignored.
13297 If the note refers to an expression that is not a constant, we
13298 must also ignore the note since we cannot tell whether the
13299 equivalence is still true. It might be possible to do
13300 slightly better than this (we only have a problem if I2DEST
13301 or I1DEST is present in the expression), but it doesn't
13302 seem worth the trouble. */
13304 if (from_insn == i3
13305 && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0))))
13306 place = i3;
13307 break;
13309 case REG_INC:
13310 /* These notes say something about how a register is used. They must
13311 be present on any use of the register in I2 or I3. */
13312 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3)))
13313 place = i3;
13315 if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2)))
13317 if (place)
13318 place2 = i2;
13319 else
13320 place = i2;
13322 break;
13324 case REG_LABEL_TARGET:
13325 case REG_LABEL_OPERAND:
13326 /* This can show up in several ways -- either directly in the
13327 pattern, or hidden off in the constant pool with (or without?)
13328 a REG_EQUAL note. */
13329 /* ??? Ignore the without-reg_equal-note problem for now. */
13330 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))
13331 || ((tem = find_reg_note (i3, REG_EQUAL, NULL_RTX))
13332 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
13333 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0)))
13334 place = i3;
13336 if (i2
13337 && (reg_mentioned_p (XEXP (note, 0), PATTERN (i2))
13338 || ((tem = find_reg_note (i2, REG_EQUAL, NULL_RTX))
13339 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
13340 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0))))
13342 if (place)
13343 place2 = i2;
13344 else
13345 place = i2;
13348 /* For REG_LABEL_TARGET on a JUMP_P, we prefer to put the note
13349 as a JUMP_LABEL or decrement LABEL_NUSES if it's already
13350 there. */
13351 if (place && JUMP_P (place)
13352 && REG_NOTE_KIND (note) == REG_LABEL_TARGET
13353 && (JUMP_LABEL (place) == NULL
13354 || JUMP_LABEL (place) == XEXP (note, 0)))
13356 rtx label = JUMP_LABEL (place);
13358 if (!label)
13359 JUMP_LABEL (place) = XEXP (note, 0);
13360 else if (LABEL_P (label))
13361 LABEL_NUSES (label)--;
13364 if (place2 && JUMP_P (place2)
13365 && REG_NOTE_KIND (note) == REG_LABEL_TARGET
13366 && (JUMP_LABEL (place2) == NULL
13367 || JUMP_LABEL (place2) == XEXP (note, 0)))
13369 rtx label = JUMP_LABEL (place2);
13371 if (!label)
13372 JUMP_LABEL (place2) = XEXP (note, 0);
13373 else if (LABEL_P (label))
13374 LABEL_NUSES (label)--;
13375 place2 = 0;
13377 break;
13379 case REG_NONNEG:
13380 /* This note says something about the value of a register prior
13381 to the execution of an insn. It is too much trouble to see
13382 if the note is still correct in all situations. It is better
13383 to simply delete it. */
13384 break;
13386 case REG_DEAD:
13387 /* If we replaced the right hand side of FROM_INSN with a
13388 REG_EQUAL note, the original use of the dying register
13389 will not have been combined into I3 and I2. In such cases,
13390 FROM_INSN is guaranteed to be the first of the combined
13391 instructions, so we simply need to search back before
13392 FROM_INSN for the previous use or set of this register,
13393 then alter the notes there appropriately.
13395 If the register is used as an input in I3, it dies there.
13396 Similarly for I2, if it is nonzero and adjacent to I3.
13398 If the register is not used as an input in either I3 or I2
13399 and it is not one of the registers we were supposed to eliminate,
13400 there are two possibilities. We might have a non-adjacent I2
13401 or we might have somehow eliminated an additional register
13402 from a computation. For example, we might have had A & B where
13403 we discover that B will always be zero. In this case we will
13404 eliminate the reference to A.
13406 In both cases, we must search to see if we can find a previous
13407 use of A and put the death note there. */
13409 if (from_insn
13410 && from_insn == i2mod
13411 && !reg_overlap_mentioned_p (XEXP (note, 0), i2mod_new_rhs))
13412 tem = from_insn;
13413 else
13415 if (from_insn
13416 && CALL_P (from_insn)
13417 && find_reg_fusage (from_insn, USE, XEXP (note, 0)))
13418 place = from_insn;
13419 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)))
13420 place = i3;
13421 else if (i2 != 0 && next_nonnote_nondebug_insn (i2) == i3
13422 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
13423 place = i2;
13424 else if ((rtx_equal_p (XEXP (note, 0), elim_i2)
13425 && !(i2mod
13426 && reg_overlap_mentioned_p (XEXP (note, 0),
13427 i2mod_old_rhs)))
13428 || rtx_equal_p (XEXP (note, 0), elim_i1)
13429 || rtx_equal_p (XEXP (note, 0), elim_i0))
13430 break;
13431 tem = i3;
13434 if (place == 0)
13436 basic_block bb = this_basic_block;
13438 for (tem = PREV_INSN (tem); place == 0; tem = PREV_INSN (tem))
13440 if (!NONDEBUG_INSN_P (tem))
13442 if (tem == BB_HEAD (bb))
13443 break;
13444 continue;
13447 /* If the register is being set at TEM, see if that is all
13448 TEM is doing. If so, delete TEM. Otherwise, make this
13449 into a REG_UNUSED note instead. Don't delete sets to
13450 global register vars. */
13451 if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER
13452 || !global_regs[REGNO (XEXP (note, 0))])
13453 && reg_set_p (XEXP (note, 0), PATTERN (tem)))
13455 rtx set = single_set (tem);
13456 rtx inner_dest = 0;
13457 #ifdef HAVE_cc0
13458 rtx cc0_setter = NULL_RTX;
13459 #endif
13461 if (set != 0)
13462 for (inner_dest = SET_DEST (set);
13463 (GET_CODE (inner_dest) == STRICT_LOW_PART
13464 || GET_CODE (inner_dest) == SUBREG
13465 || GET_CODE (inner_dest) == ZERO_EXTRACT);
13466 inner_dest = XEXP (inner_dest, 0))
13469 /* Verify that it was the set, and not a clobber that
13470 modified the register.
13472 CC0 targets must be careful to maintain setter/user
13473 pairs. If we cannot delete the setter due to side
13474 effects, mark the user with an UNUSED note instead
13475 of deleting it. */
13477 if (set != 0 && ! side_effects_p (SET_SRC (set))
13478 && rtx_equal_p (XEXP (note, 0), inner_dest)
13479 #ifdef HAVE_cc0
13480 && (! reg_mentioned_p (cc0_rtx, SET_SRC (set))
13481 || ((cc0_setter = prev_cc0_setter (tem)) != NULL
13482 && sets_cc0_p (PATTERN (cc0_setter)) > 0))
13483 #endif
13486 /* Move the notes and links of TEM elsewhere.
13487 This might delete other dead insns recursively.
13488 First set the pattern to something that won't use
13489 any register. */
13490 rtx old_notes = REG_NOTES (tem);
13492 PATTERN (tem) = pc_rtx;
13493 REG_NOTES (tem) = NULL;
13495 distribute_notes (old_notes, tem, tem, NULL_RTX,
13496 NULL_RTX, NULL_RTX, NULL_RTX);
13497 distribute_links (LOG_LINKS (tem));
13499 SET_INSN_DELETED (tem);
13500 if (tem == i2)
13501 i2 = NULL_RTX;
13503 #ifdef HAVE_cc0
13504 /* Delete the setter too. */
13505 if (cc0_setter)
13507 PATTERN (cc0_setter) = pc_rtx;
13508 old_notes = REG_NOTES (cc0_setter);
13509 REG_NOTES (cc0_setter) = NULL;
13511 distribute_notes (old_notes, cc0_setter,
13512 cc0_setter, NULL_RTX,
13513 NULL_RTX, NULL_RTX, NULL_RTX);
13514 distribute_links (LOG_LINKS (cc0_setter));
13516 SET_INSN_DELETED (cc0_setter);
13517 if (cc0_setter == i2)
13518 i2 = NULL_RTX;
13520 #endif
13522 else
13524 PUT_REG_NOTE_KIND (note, REG_UNUSED);
13526 /* If there isn't already a REG_UNUSED note, put one
13527 here. Do not place a REG_DEAD note, even if
13528 the register is also used here; that would not
13529 match the algorithm used in lifetime analysis
13530 and can cause the consistency check in the
13531 scheduler to fail. */
13532 if (! find_regno_note (tem, REG_UNUSED,
13533 REGNO (XEXP (note, 0))))
13534 place = tem;
13535 break;
13538 else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))
13539 || (CALL_P (tem)
13540 && find_reg_fusage (tem, USE, XEXP (note, 0))))
13542 place = tem;
13544 /* If we are doing a 3->2 combination, and we have a
13545 register which formerly died in i3 and was not used
13546 by i2, which now no longer dies in i3 and is used in
13547 i2 but does not die in i2, and place is between i2
13548 and i3, then we may need to move a link from place to
13549 i2. */
13550 if (i2 && DF_INSN_LUID (place) > DF_INSN_LUID (i2)
13551 && from_insn
13552 && DF_INSN_LUID (from_insn) > DF_INSN_LUID (i2)
13553 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
13555 struct insn_link *links = LOG_LINKS (place);
13556 LOG_LINKS (place) = NULL;
13557 distribute_links (links);
13559 break;
13562 if (tem == BB_HEAD (bb))
13563 break;
13568 /* If the register is set or already dead at PLACE, we needn't do
13569 anything with this note if it is still a REG_DEAD note.
13570 We check here if it is set at all, not if is it totally replaced,
13571 which is what `dead_or_set_p' checks, so also check for it being
13572 set partially. */
13574 if (place && REG_NOTE_KIND (note) == REG_DEAD)
13576 unsigned int regno = REGNO (XEXP (note, 0));
13577 reg_stat_type *rsp = &reg_stat[regno];
13579 if (dead_or_set_p (place, XEXP (note, 0))
13580 || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place)))
13582 /* Unless the register previously died in PLACE, clear
13583 last_death. [I no longer understand why this is
13584 being done.] */
13585 if (rsp->last_death != place)
13586 rsp->last_death = 0;
13587 place = 0;
13589 else
13590 rsp->last_death = place;
13592 /* If this is a death note for a hard reg that is occupying
13593 multiple registers, ensure that we are still using all
13594 parts of the object. If we find a piece of the object
13595 that is unused, we must arrange for an appropriate REG_DEAD
13596 note to be added for it. However, we can't just emit a USE
13597 and tag the note to it, since the register might actually
13598 be dead; so we recourse, and the recursive call then finds
13599 the previous insn that used this register. */
13601 if (place && regno < FIRST_PSEUDO_REGISTER
13602 && hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))] > 1)
13604 unsigned int endregno = END_HARD_REGNO (XEXP (note, 0));
13605 bool all_used = true;
13606 unsigned int i;
13608 for (i = regno; i < endregno; i++)
13609 if ((! refers_to_regno_p (i, i + 1, PATTERN (place), 0)
13610 && ! find_regno_fusage (place, USE, i))
13611 || dead_or_set_regno_p (place, i))
13613 all_used = false;
13614 break;
13617 if (! all_used)
13619 /* Put only REG_DEAD notes for pieces that are
13620 not already dead or set. */
13622 for (i = regno; i < endregno;
13623 i += hard_regno_nregs[i][reg_raw_mode[i]])
13625 rtx piece = regno_reg_rtx[i];
13626 basic_block bb = this_basic_block;
13628 if (! dead_or_set_p (place, piece)
13629 && ! reg_bitfield_target_p (piece,
13630 PATTERN (place)))
13632 rtx new_note = alloc_reg_note (REG_DEAD, piece,
13633 NULL_RTX);
13635 distribute_notes (new_note, place, place,
13636 NULL_RTX, NULL_RTX, NULL_RTX,
13637 NULL_RTX);
13639 else if (! refers_to_regno_p (i, i + 1,
13640 PATTERN (place), 0)
13641 && ! find_regno_fusage (place, USE, i))
13642 for (tem = PREV_INSN (place); ;
13643 tem = PREV_INSN (tem))
13645 if (!NONDEBUG_INSN_P (tem))
13647 if (tem == BB_HEAD (bb))
13648 break;
13649 continue;
13651 if (dead_or_set_p (tem, piece)
13652 || reg_bitfield_target_p (piece,
13653 PATTERN (tem)))
13655 add_reg_note (tem, REG_UNUSED, piece);
13656 break;
13661 place = 0;
13665 break;
13667 default:
13668 /* Any other notes should not be present at this point in the
13669 compilation. */
13670 gcc_unreachable ();
13673 if (place)
13675 XEXP (note, 1) = REG_NOTES (place);
13676 REG_NOTES (place) = note;
13679 if (place2)
13680 add_reg_note (place2, REG_NOTE_KIND (note), XEXP (note, 0));
13684 /* Similarly to above, distribute the LOG_LINKS that used to be present on
13685 I3, I2, and I1 to new locations. This is also called to add a link
13686 pointing at I3 when I3's destination is changed. */
13688 static void
13689 distribute_links (struct insn_link *links)
13691 struct insn_link *link, *next_link;
13693 for (link = links; link; link = next_link)
13695 rtx place = 0;
13696 rtx insn;
13697 rtx set, reg;
13699 next_link = link->next;
13701 /* If the insn that this link points to is a NOTE or isn't a single
13702 set, ignore it. In the latter case, it isn't clear what we
13703 can do other than ignore the link, since we can't tell which
13704 register it was for. Such links wouldn't be used by combine
13705 anyway.
13707 It is not possible for the destination of the target of the link to
13708 have been changed by combine. The only potential of this is if we
13709 replace I3, I2, and I1 by I3 and I2. But in that case the
13710 destination of I2 also remains unchanged. */
13712 if (NOTE_P (link->insn)
13713 || (set = single_set (link->insn)) == 0)
13714 continue;
13716 reg = SET_DEST (set);
13717 while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
13718 || GET_CODE (reg) == STRICT_LOW_PART)
13719 reg = XEXP (reg, 0);
13721 /* A LOG_LINK is defined as being placed on the first insn that uses
13722 a register and points to the insn that sets the register. Start
13723 searching at the next insn after the target of the link and stop
13724 when we reach a set of the register or the end of the basic block.
13726 Note that this correctly handles the link that used to point from
13727 I3 to I2. Also note that not much searching is typically done here
13728 since most links don't point very far away. */
13730 for (insn = NEXT_INSN (link->insn);
13731 (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
13732 || BB_HEAD (this_basic_block->next_bb) != insn));
13733 insn = NEXT_INSN (insn))
13734 if (DEBUG_INSN_P (insn))
13735 continue;
13736 else if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
13738 if (reg_referenced_p (reg, PATTERN (insn)))
13739 place = insn;
13740 break;
13742 else if (CALL_P (insn)
13743 && find_reg_fusage (insn, USE, reg))
13745 place = insn;
13746 break;
13748 else if (INSN_P (insn) && reg_set_p (reg, insn))
13749 break;
13751 /* If we found a place to put the link, place it there unless there
13752 is already a link to the same insn as LINK at that point. */
13754 if (place)
13756 struct insn_link *link2;
13758 FOR_EACH_LOG_LINK (link2, place)
13759 if (link2->insn == link->insn)
13760 break;
13762 if (link2 == NULL)
13764 link->next = LOG_LINKS (place);
13765 LOG_LINKS (place) = link;
13767 /* Set added_links_insn to the earliest insn we added a
13768 link to. */
13769 if (added_links_insn == 0
13770 || DF_INSN_LUID (added_links_insn) > DF_INSN_LUID (place))
13771 added_links_insn = place;
13777 /* Subroutine of unmentioned_reg_p and callback from for_each_rtx.
13778 Check whether the expression pointer to by LOC is a register or
13779 memory, and if so return 1 if it isn't mentioned in the rtx EXPR.
13780 Otherwise return zero. */
13782 static int
13783 unmentioned_reg_p_1 (rtx *loc, void *expr)
13785 rtx x = *loc;
13787 if (x != NULL_RTX
13788 && (REG_P (x) || MEM_P (x))
13789 && ! reg_mentioned_p (x, (rtx) expr))
13790 return 1;
13791 return 0;
13794 /* Check for any register or memory mentioned in EQUIV that is not
13795 mentioned in EXPR. This is used to restrict EQUIV to "specializations"
13796 of EXPR where some registers may have been replaced by constants. */
13798 static bool
13799 unmentioned_reg_p (rtx equiv, rtx expr)
13801 return for_each_rtx (&equiv, unmentioned_reg_p_1, expr);
13804 DEBUG_FUNCTION void
13805 dump_combine_stats (FILE *file)
13807 fprintf
13808 (file,
13809 ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n",
13810 combine_attempts, combine_merges, combine_extras, combine_successes);
13813 void
13814 dump_combine_total_stats (FILE *file)
13816 fprintf
13817 (file,
13818 "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n",
13819 total_attempts, total_merges, total_extras, total_successes);
13822 static bool
13823 gate_handle_combine (void)
13825 return (optimize > 0);
13828 /* Try combining insns through substitution. */
13829 static unsigned int
13830 rest_of_handle_combine (void)
13832 int rebuild_jump_labels_after_combine;
13834 df_set_flags (DF_LR_RUN_DCE + DF_DEFER_INSN_RESCAN);
13835 df_note_add_problem ();
13836 df_analyze ();
13838 regstat_init_n_sets_and_refs ();
13840 rebuild_jump_labels_after_combine
13841 = combine_instructions (get_insns (), max_reg_num ());
13843 /* Combining insns may have turned an indirect jump into a
13844 direct jump. Rebuild the JUMP_LABEL fields of jumping
13845 instructions. */
13846 if (rebuild_jump_labels_after_combine)
13848 timevar_push (TV_JUMP);
13849 rebuild_jump_labels (get_insns ());
13850 cleanup_cfg (0);
13851 timevar_pop (TV_JUMP);
13854 regstat_free_n_sets_and_refs ();
13855 return 0;
13858 namespace {
13860 const pass_data pass_data_combine =
13862 RTL_PASS, /* type */
13863 "combine", /* name */
13864 OPTGROUP_NONE, /* optinfo_flags */
13865 true, /* has_gate */
13866 true, /* has_execute */
13867 TV_COMBINE, /* tv_id */
13868 PROP_cfglayout, /* properties_required */
13869 0, /* properties_provided */
13870 0, /* properties_destroyed */
13871 0, /* todo_flags_start */
13872 ( TODO_df_finish | TODO_verify_rtl_sharing ), /* todo_flags_finish */
13875 class pass_combine : public rtl_opt_pass
13877 public:
13878 pass_combine(gcc::context *ctxt)
13879 : rtl_opt_pass(pass_data_combine, ctxt)
13882 /* opt_pass methods: */
13883 bool gate () { return gate_handle_combine (); }
13884 unsigned int execute () { return rest_of_handle_combine (); }
13886 }; // class pass_combine
13888 } // anon namespace
13890 rtl_opt_pass *
13891 make_pass_combine (gcc::context *ctxt)
13893 return new pass_combine (ctxt);