usage.adb: Change "pragma inline" to "pragma Inline" in information and error messages
[gcc.git] / gcc / cse.c
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
21
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27
28 #include "rtl.h"
29 #include "tm_p.h"
30 #include "regs.h"
31 #include "hard-reg-set.h"
32 #include "basic-block.h"
33 #include "flags.h"
34 #include "real.h"
35 #include "insn-config.h"
36 #include "recog.h"
37 #include "function.h"
38 #include "expr.h"
39 #include "toplev.h"
40 #include "output.h"
41 #include "ggc.h"
42 #include "timevar.h"
43 #include "except.h"
44 #include "target.h"
45 #include "params.h"
46 #include "rtlhooks-def.h"
47
48 /* The basic idea of common subexpression elimination is to go
49 through the code, keeping a record of expressions that would
50 have the same value at the current scan point, and replacing
51 expressions encountered with the cheapest equivalent expression.
52
53 It is too complicated to keep track of the different possibilities
54 when control paths merge in this code; so, at each label, we forget all
55 that is known and start fresh. This can be described as processing each
56 extended basic block separately. We have a separate pass to perform
57 global CSE.
58
59 Note CSE can turn a conditional or computed jump into a nop or
60 an unconditional jump. When this occurs we arrange to run the jump
61 optimizer after CSE to delete the unreachable code.
62
63 We use two data structures to record the equivalent expressions:
64 a hash table for most expressions, and a vector of "quantity
65 numbers" to record equivalent (pseudo) registers.
66
67 The use of the special data structure for registers is desirable
68 because it is faster. It is possible because registers references
69 contain a fairly small number, the register number, taken from
70 a contiguously allocated series, and two register references are
71 identical if they have the same number. General expressions
72 do not have any such thing, so the only way to retrieve the
73 information recorded on an expression other than a register
74 is to keep it in a hash table.
75
76 Registers and "quantity numbers":
77
78 At the start of each basic block, all of the (hardware and pseudo)
79 registers used in the function are given distinct quantity
80 numbers to indicate their contents. During scan, when the code
81 copies one register into another, we copy the quantity number.
82 When a register is loaded in any other way, we allocate a new
83 quantity number to describe the value generated by this operation.
84 `reg_qty' records what quantity a register is currently thought
85 of as containing.
86
87 All real quantity numbers are greater than or equal to `max_reg'.
88 If register N has not been assigned a quantity, reg_qty[N] will equal N.
89
90 Quantity numbers below `max_reg' do not exist and none of the `qty_table'
91 entries should be referenced with an index below `max_reg'.
92
93 We also maintain a bidirectional chain of registers for each
94 quantity number. The `qty_table` members `first_reg' and `last_reg',
95 and `reg_eqv_table' members `next' and `prev' hold these chains.
96
97 The first register in a chain is the one whose lifespan is least local.
98 Among equals, it is the one that was seen first.
99 We replace any equivalent register with that one.
100
101 If two registers have the same quantity number, it must be true that
102 REG expressions with qty_table `mode' must be in the hash table for both
103 registers and must be in the same class.
104
105 The converse is not true. Since hard registers may be referenced in
106 any mode, two REG expressions might be equivalent in the hash table
107 but not have the same quantity number if the quantity number of one
108 of the registers is not the same mode as those expressions.
109
110 Constants and quantity numbers
111
112 When a quantity has a known constant value, that value is stored
113 in the appropriate qty_table `const_rtx'. This is in addition to
114 putting the constant in the hash table as is usual for non-regs.
115
116 Whether a reg or a constant is preferred is determined by the configuration
117 macro CONST_COSTS and will often depend on the constant value. In any
118 event, expressions containing constants can be simplified, by fold_rtx.
119
120 When a quantity has a known nearly constant value (such as an address
121 of a stack slot), that value is stored in the appropriate qty_table
122 `const_rtx'.
123
124 Integer constants don't have a machine mode. However, cse
125 determines the intended machine mode from the destination
126 of the instruction that moves the constant. The machine mode
127 is recorded in the hash table along with the actual RTL
128 constant expression so that different modes are kept separate.
129
130 Other expressions:
131
132 To record known equivalences among expressions in general
133 we use a hash table called `table'. It has a fixed number of buckets
134 that contain chains of `struct table_elt' elements for expressions.
135 These chains connect the elements whose expressions have the same
136 hash codes.
137
138 Other chains through the same elements connect the elements which
139 currently have equivalent values.
140
141 Register references in an expression are canonicalized before hashing
142 the expression. This is done using `reg_qty' and qty_table `first_reg'.
143 The hash code of a register reference is computed using the quantity
144 number, not the register number.
145
146 When the value of an expression changes, it is necessary to remove from the
147 hash table not just that expression but all expressions whose values
148 could be different as a result.
149
150 1. If the value changing is in memory, except in special cases
151 ANYTHING referring to memory could be changed. That is because
152 nobody knows where a pointer does not point.
153 The function `invalidate_memory' removes what is necessary.
154
155 The special cases are when the address is constant or is
156 a constant plus a fixed register such as the frame pointer
157 or a static chain pointer. When such addresses are stored in,
158 we can tell exactly which other such addresses must be invalidated
159 due to overlap. `invalidate' does this.
160 All expressions that refer to non-constant
161 memory addresses are also invalidated. `invalidate_memory' does this.
162
163 2. If the value changing is a register, all expressions
164 containing references to that register, and only those,
165 must be removed.
166
167 Because searching the entire hash table for expressions that contain
168 a register is very slow, we try to figure out when it isn't necessary.
169 Precisely, this is necessary only when expressions have been
170 entered in the hash table using this register, and then the value has
171 changed, and then another expression wants to be added to refer to
172 the register's new value. This sequence of circumstances is rare
173 within any one basic block.
174
175 The vectors `reg_tick' and `reg_in_table' are used to detect this case.
176 reg_tick[i] is incremented whenever a value is stored in register i.
177 reg_in_table[i] holds -1 if no references to register i have been
178 entered in the table; otherwise, it contains the value reg_tick[i] had
179 when the references were entered. If we want to enter a reference
180 and reg_in_table[i] != reg_tick[i], we must scan and remove old references.
181 Until we want to enter a new entry, the mere fact that the two vectors
182 don't match makes the entries be ignored if anyone tries to match them.
183
184 Registers themselves are entered in the hash table as well as in
185 the equivalent-register chains. However, the vectors `reg_tick'
186 and `reg_in_table' do not apply to expressions which are simple
187 register references. These expressions are removed from the table
188 immediately when they become invalid, and this can be done even if
189 we do not immediately search for all the expressions that refer to
190 the register.
191
192 A CLOBBER rtx in an instruction invalidates its operand for further
193 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
194 invalidates everything that resides in memory.
195
196 Related expressions:
197
198 Constant expressions that differ only by an additive integer
199 are called related. When a constant expression is put in
200 the table, the related expression with no constant term
201 is also entered. These are made to point at each other
202 so that it is possible to find out if there exists any
203 register equivalent to an expression related to a given expression. */
204
205 /* One plus largest register number used in this function. */
206
207 static int max_reg;
208
209 /* One plus largest instruction UID used in this function at time of
210 cse_main call. */
211
212 static int max_insn_uid;
213
214 /* Length of qty_table vector. We know in advance we will not need
215 a quantity number this big. */
216
217 static int max_qty;
218
219 /* Next quantity number to be allocated.
220 This is 1 + the largest number needed so far. */
221
222 static int next_qty;
223
224 /* Per-qty information tracking.
225
226 `first_reg' and `last_reg' track the head and tail of the
227 chain of registers which currently contain this quantity.
228
229 `mode' contains the machine mode of this quantity.
230
231 `const_rtx' holds the rtx of the constant value of this
232 quantity, if known. A summations of the frame/arg pointer
233 and a constant can also be entered here. When this holds
234 a known value, `const_insn' is the insn which stored the
235 constant value.
236
237 `comparison_{code,const,qty}' are used to track when a
238 comparison between a quantity and some constant or register has
239 been passed. In such a case, we know the results of the comparison
240 in case we see it again. These members record a comparison that
241 is known to be true. `comparison_code' holds the rtx code of such
242 a comparison, else it is set to UNKNOWN and the other two
243 comparison members are undefined. `comparison_const' holds
244 the constant being compared against, or zero if the comparison
245 is not against a constant. `comparison_qty' holds the quantity
246 being compared against when the result is known. If the comparison
247 is not with a register, `comparison_qty' is -1. */
248
249 struct qty_table_elem
250 {
251 rtx const_rtx;
252 rtx const_insn;
253 rtx comparison_const;
254 int comparison_qty;
255 unsigned int first_reg, last_reg;
256 /* The sizes of these fields should match the sizes of the
257 code and mode fields of struct rtx_def (see rtl.h). */
258 ENUM_BITFIELD(rtx_code) comparison_code : 16;
259 ENUM_BITFIELD(machine_mode) mode : 8;
260 };
261
262 /* The table of all qtys, indexed by qty number. */
263 static struct qty_table_elem *qty_table;
264
265 #ifdef HAVE_cc0
266 /* For machines that have a CC0, we do not record its value in the hash
267 table since its use is guaranteed to be the insn immediately following
268 its definition and any other insn is presumed to invalidate it.
269
270 Instead, we store below the value last assigned to CC0. If it should
271 happen to be a constant, it is stored in preference to the actual
272 assigned value. In case it is a constant, we store the mode in which
273 the constant should be interpreted. */
274
275 static rtx prev_insn_cc0;
276 static enum machine_mode prev_insn_cc0_mode;
277
278 /* Previous actual insn. 0 if at first insn of basic block. */
279
280 static rtx prev_insn;
281 #endif
282
283 /* Insn being scanned. */
284
285 static rtx this_insn;
286
287 /* Index by register number, gives the number of the next (or
288 previous) register in the chain of registers sharing the same
289 value.
290
291 Or -1 if this register is at the end of the chain.
292
293 If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */
294
295 /* Per-register equivalence chain. */
296 struct reg_eqv_elem
297 {
298 int next, prev;
299 };
300
301 /* The table of all register equivalence chains. */
302 static struct reg_eqv_elem *reg_eqv_table;
303
304 struct cse_reg_info
305 {
306 /* Next in hash chain. */
307 struct cse_reg_info *hash_next;
308
309 /* The next cse_reg_info structure in the free or used list. */
310 struct cse_reg_info *next;
311
312 /* Search key */
313 unsigned int regno;
314
315 /* The quantity number of the register's current contents. */
316 int reg_qty;
317
318 /* The number of times the register has been altered in the current
319 basic block. */
320 int reg_tick;
321
322 /* The REG_TICK value at which rtx's containing this register are
323 valid in the hash table. If this does not equal the current
324 reg_tick value, such expressions existing in the hash table are
325 invalid. */
326 int reg_in_table;
327
328 /* The SUBREG that was set when REG_TICK was last incremented. Set
329 to -1 if the last store was to the whole register, not a subreg. */
330 unsigned int subreg_ticked;
331 };
332
333 /* A free list of cse_reg_info entries. */
334 static struct cse_reg_info *cse_reg_info_free_list;
335
336 /* A used list of cse_reg_info entries. */
337 static struct cse_reg_info *cse_reg_info_used_list;
338 static struct cse_reg_info *cse_reg_info_used_list_end;
339
340 /* A mapping from registers to cse_reg_info data structures. */
341 #define REGHASH_SHIFT 7
342 #define REGHASH_SIZE (1 << REGHASH_SHIFT)
343 #define REGHASH_MASK (REGHASH_SIZE - 1)
344 static struct cse_reg_info *reg_hash[REGHASH_SIZE];
345
346 #define REGHASH_FN(REGNO) \
347 (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK)
348
349 /* The last lookup we did into the cse_reg_info_tree. This allows us
350 to cache repeated lookups. */
351 static unsigned int cached_regno;
352 static struct cse_reg_info *cached_cse_reg_info;
353
354 /* A HARD_REG_SET containing all the hard registers for which there is
355 currently a REG expression in the hash table. Note the difference
356 from the above variables, which indicate if the REG is mentioned in some
357 expression in the table. */
358
359 static HARD_REG_SET hard_regs_in_table;
360
361 /* CUID of insn that starts the basic block currently being cse-processed. */
362
363 static int cse_basic_block_start;
364
365 /* CUID of insn that ends the basic block currently being cse-processed. */
366
367 static int cse_basic_block_end;
368
369 /* Vector mapping INSN_UIDs to cuids.
370 The cuids are like uids but increase monotonically always.
371 We use them to see whether a reg is used outside a given basic block. */
372
373 static int *uid_cuid;
374
375 /* Highest UID in UID_CUID. */
376 static int max_uid;
377
378 /* Get the cuid of an insn. */
379
380 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
381
382 /* Nonzero if this pass has made changes, and therefore it's
383 worthwhile to run the garbage collector. */
384
385 static int cse_altered;
386
387 /* Nonzero if cse has altered conditional jump insns
388 in such a way that jump optimization should be redone. */
389
390 static int cse_jumps_altered;
391
392 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
393 REG_LABEL, we have to rerun jump after CSE to put in the note. */
394 static int recorded_label_ref;
395
396 /* canon_hash stores 1 in do_not_record
397 if it notices a reference to CC0, PC, or some other volatile
398 subexpression. */
399
400 static int do_not_record;
401
402 #ifdef LOAD_EXTEND_OP
403
404 /* Scratch rtl used when looking for load-extended copy of a MEM. */
405 static rtx memory_extend_rtx;
406 #endif
407
408 /* canon_hash stores 1 in hash_arg_in_memory
409 if it notices a reference to memory within the expression being hashed. */
410
411 static int hash_arg_in_memory;
412
413 /* The hash table contains buckets which are chains of `struct table_elt's,
414 each recording one expression's information.
415 That expression is in the `exp' field.
416
417 The canon_exp field contains a canonical (from the point of view of
418 alias analysis) version of the `exp' field.
419
420 Those elements with the same hash code are chained in both directions
421 through the `next_same_hash' and `prev_same_hash' fields.
422
423 Each set of expressions with equivalent values
424 are on a two-way chain through the `next_same_value'
425 and `prev_same_value' fields, and all point with
426 the `first_same_value' field at the first element in
427 that chain. The chain is in order of increasing cost.
428 Each element's cost value is in its `cost' field.
429
430 The `in_memory' field is nonzero for elements that
431 involve any reference to memory. These elements are removed
432 whenever a write is done to an unidentified location in memory.
433 To be safe, we assume that a memory address is unidentified unless
434 the address is either a symbol constant or a constant plus
435 the frame pointer or argument pointer.
436
437 The `related_value' field is used to connect related expressions
438 (that differ by adding an integer).
439 The related expressions are chained in a circular fashion.
440 `related_value' is zero for expressions for which this
441 chain is not useful.
442
443 The `cost' field stores the cost of this element's expression.
444 The `regcost' field stores the value returned by approx_reg_cost for
445 this element's expression.
446
447 The `is_const' flag is set if the element is a constant (including
448 a fixed address).
449
450 The `flag' field is used as a temporary during some search routines.
451
452 The `mode' field is usually the same as GET_MODE (`exp'), but
453 if `exp' is a CONST_INT and has no machine mode then the `mode'
454 field is the mode it was being used as. Each constant is
455 recorded separately for each mode it is used with. */
456
457 struct table_elt
458 {
459 rtx exp;
460 rtx canon_exp;
461 struct table_elt *next_same_hash;
462 struct table_elt *prev_same_hash;
463 struct table_elt *next_same_value;
464 struct table_elt *prev_same_value;
465 struct table_elt *first_same_value;
466 struct table_elt *related_value;
467 int cost;
468 int regcost;
469 /* The size of this field should match the size
470 of the mode field of struct rtx_def (see rtl.h). */
471 ENUM_BITFIELD(machine_mode) mode : 8;
472 char in_memory;
473 char is_const;
474 char flag;
475 };
476
477 /* We don't want a lot of buckets, because we rarely have very many
478 things stored in the hash table, and a lot of buckets slows
479 down a lot of loops that happen frequently. */
480 #define HASH_SHIFT 5
481 #define HASH_SIZE (1 << HASH_SHIFT)
482 #define HASH_MASK (HASH_SIZE - 1)
483
484 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
485 register (hard registers may require `do_not_record' to be set). */
486
487 #define HASH(X, M) \
488 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
489 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
490 : canon_hash (X, M)) & HASH_MASK)
491
492 /* Like HASH, but without side-effects. */
493 #define SAFE_HASH(X, M) \
494 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
495 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
496 : safe_hash (X, M)) & HASH_MASK)
497
498 /* Determine whether register number N is considered a fixed register for the
499 purpose of approximating register costs.
500 It is desirable to replace other regs with fixed regs, to reduce need for
501 non-fixed hard regs.
502 A reg wins if it is either the frame pointer or designated as fixed. */
503 #define FIXED_REGNO_P(N) \
504 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
505 || fixed_regs[N] || global_regs[N])
506
507 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
508 hard registers and pointers into the frame are the cheapest with a cost
509 of 0. Next come pseudos with a cost of one and other hard registers with
510 a cost of 2. Aside from these special cases, call `rtx_cost'. */
511
512 #define CHEAP_REGNO(N) \
513 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
514 || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \
515 || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \
516 || ((N) < FIRST_PSEUDO_REGISTER \
517 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
518
519 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
520 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
521
522 /* Get the info associated with register N. */
523
524 #define GET_CSE_REG_INFO(N) \
525 (((N) == cached_regno && cached_cse_reg_info) \
526 ? cached_cse_reg_info : get_cse_reg_info ((N)))
527
528 /* Get the number of times this register has been updated in this
529 basic block. */
530
531 #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->reg_tick)
532
533 /* Get the point at which REG was recorded in the table. */
534
535 #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table)
536
537 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
538 SUBREG). */
539
540 #define SUBREG_TICKED(N) ((GET_CSE_REG_INFO (N))->subreg_ticked)
541
542 /* Get the quantity number for REG. */
543
544 #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty)
545
546 /* Determine if the quantity number for register X represents a valid index
547 into the qty_table. */
548
549 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) != (int) (N))
550
551 static struct table_elt *table[HASH_SIZE];
552
553 /* Chain of `struct table_elt's made so far for this function
554 but currently removed from the table. */
555
556 static struct table_elt *free_element_chain;
557
558 /* Number of `struct table_elt' structures made so far for this function. */
559
560 static int n_elements_made;
561
562 /* Maximum value `n_elements_made' has had so far in this compilation
563 for functions previously processed. */
564
565 static int max_elements_made;
566
567 /* Set to the cost of a constant pool reference if one was found for a
568 symbolic constant. If this was found, it means we should try to
569 convert constants into constant pool entries if they don't fit in
570 the insn. */
571
572 static int constant_pool_entries_cost;
573 static int constant_pool_entries_regcost;
574
575 /* This data describes a block that will be processed by cse_basic_block. */
576
577 struct cse_basic_block_data
578 {
579 /* Lowest CUID value of insns in block. */
580 int low_cuid;
581 /* Highest CUID value of insns in block. */
582 int high_cuid;
583 /* Total number of SETs in block. */
584 int nsets;
585 /* Last insn in the block. */
586 rtx last;
587 /* Size of current branch path, if any. */
588 int path_size;
589 /* Current branch path, indicating which branches will be taken. */
590 struct branch_path
591 {
592 /* The branch insn. */
593 rtx branch;
594 /* Whether it should be taken or not. AROUND is the same as taken
595 except that it is used when the destination label is not preceded
596 by a BARRIER. */
597 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
598 } *path;
599 };
600
601 static bool fixed_base_plus_p (rtx x);
602 static int notreg_cost (rtx, enum rtx_code);
603 static int approx_reg_cost_1 (rtx *, void *);
604 static int approx_reg_cost (rtx);
605 static int preferable (int, int, int, int);
606 static void new_basic_block (void);
607 static void make_new_qty (unsigned int, enum machine_mode);
608 static void make_regs_eqv (unsigned int, unsigned int);
609 static void delete_reg_equiv (unsigned int);
610 static int mention_regs (rtx);
611 static int insert_regs (rtx, struct table_elt *, int);
612 static void remove_from_table (struct table_elt *, unsigned);
613 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
614 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
615 static rtx lookup_as_function (rtx, enum rtx_code);
616 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
617 enum machine_mode);
618 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
619 static void invalidate (rtx, enum machine_mode);
620 static int cse_rtx_varies_p (rtx, int);
621 static void remove_invalid_refs (unsigned int);
622 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
623 enum machine_mode);
624 static void rehash_using_reg (rtx);
625 static void invalidate_memory (void);
626 static void invalidate_for_call (void);
627 static rtx use_related_value (rtx, struct table_elt *);
628
629 static inline unsigned canon_hash (rtx, enum machine_mode);
630 static inline unsigned safe_hash (rtx, enum machine_mode);
631 static unsigned hash_rtx_string (const char *);
632
633 static rtx canon_reg (rtx, rtx);
634 static void find_best_addr (rtx, rtx *, enum machine_mode);
635 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
636 enum machine_mode *,
637 enum machine_mode *);
638 static rtx fold_rtx (rtx, rtx);
639 static rtx equiv_constant (rtx);
640 static void record_jump_equiv (rtx, int);
641 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
642 int);
643 static void cse_insn (rtx, rtx);
644 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
645 int, int);
646 static int addr_affects_sp_p (rtx);
647 static void invalidate_from_clobbers (rtx);
648 static rtx cse_process_notes (rtx, rtx);
649 static void invalidate_skipped_set (rtx, rtx, void *);
650 static void invalidate_skipped_block (rtx);
651 static rtx cse_basic_block (rtx, rtx, struct branch_path *);
652 static void count_reg_usage (rtx, int *, int);
653 static int check_for_label_ref (rtx *, void *);
654 extern void dump_class (struct table_elt*);
655 static struct cse_reg_info * get_cse_reg_info (unsigned int);
656 static int check_dependence (rtx *, void *);
657
658 static void flush_hash_table (void);
659 static bool insn_live_p (rtx, int *);
660 static bool set_live_p (rtx, rtx, int *);
661 static bool dead_libcall_p (rtx, int *);
662 static int cse_change_cc_mode (rtx *, void *);
663 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
664 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
665 \f
666
667 #undef RTL_HOOKS_GEN_LOWPART
668 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
669
670 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
671 \f
672 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
673 virtual regs here because the simplify_*_operation routines are called
674 by integrate.c, which is called before virtual register instantiation. */
675
676 static bool
677 fixed_base_plus_p (rtx x)
678 {
679 switch (GET_CODE (x))
680 {
681 case REG:
682 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
683 return true;
684 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
685 return true;
686 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
687 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
688 return true;
689 return false;
690
691 case PLUS:
692 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
693 return false;
694 return fixed_base_plus_p (XEXP (x, 0));
695
696 default:
697 return false;
698 }
699 }
700
701 /* Dump the expressions in the equivalence class indicated by CLASSP.
702 This function is used only for debugging. */
703 void
704 dump_class (struct table_elt *classp)
705 {
706 struct table_elt *elt;
707
708 fprintf (stderr, "Equivalence chain for ");
709 print_rtl (stderr, classp->exp);
710 fprintf (stderr, ": \n");
711
712 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
713 {
714 print_rtl (stderr, elt->exp);
715 fprintf (stderr, "\n");
716 }
717 }
718
719 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
720
721 static int
722 approx_reg_cost_1 (rtx *xp, void *data)
723 {
724 rtx x = *xp;
725 int *cost_p = data;
726
727 if (x && REG_P (x))
728 {
729 unsigned int regno = REGNO (x);
730
731 if (! CHEAP_REGNO (regno))
732 {
733 if (regno < FIRST_PSEUDO_REGISTER)
734 {
735 if (SMALL_REGISTER_CLASSES)
736 return 1;
737 *cost_p += 2;
738 }
739 else
740 *cost_p += 1;
741 }
742 }
743
744 return 0;
745 }
746
747 /* Return an estimate of the cost of the registers used in an rtx.
748 This is mostly the number of different REG expressions in the rtx;
749 however for some exceptions like fixed registers we use a cost of
750 0. If any other hard register reference occurs, return MAX_COST. */
751
752 static int
753 approx_reg_cost (rtx x)
754 {
755 int cost = 0;
756
757 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
758 return MAX_COST;
759
760 return cost;
761 }
762
763 /* Return a negative value if an rtx A, whose costs are given by COST_A
764 and REGCOST_A, is more desirable than an rtx B.
765 Return a positive value if A is less desirable, or 0 if the two are
766 equally good. */
767 static int
768 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
769 {
770 /* First, get rid of cases involving expressions that are entirely
771 unwanted. */
772 if (cost_a != cost_b)
773 {
774 if (cost_a == MAX_COST)
775 return 1;
776 if (cost_b == MAX_COST)
777 return -1;
778 }
779
780 /* Avoid extending lifetimes of hardregs. */
781 if (regcost_a != regcost_b)
782 {
783 if (regcost_a == MAX_COST)
784 return 1;
785 if (regcost_b == MAX_COST)
786 return -1;
787 }
788
789 /* Normal operation costs take precedence. */
790 if (cost_a != cost_b)
791 return cost_a - cost_b;
792 /* Only if these are identical consider effects on register pressure. */
793 if (regcost_a != regcost_b)
794 return regcost_a - regcost_b;
795 return 0;
796 }
797
798 /* Internal function, to compute cost when X is not a register; called
799 from COST macro to keep it simple. */
800
801 static int
802 notreg_cost (rtx x, enum rtx_code outer)
803 {
804 return ((GET_CODE (x) == SUBREG
805 && REG_P (SUBREG_REG (x))
806 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
807 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
808 && (GET_MODE_SIZE (GET_MODE (x))
809 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
810 && subreg_lowpart_p (x)
811 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
812 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
813 ? 0
814 : rtx_cost (x, outer) * 2);
815 }
816
817 \f
818 static struct cse_reg_info *
819 get_cse_reg_info (unsigned int regno)
820 {
821 struct cse_reg_info **hash_head = &reg_hash[REGHASH_FN (regno)];
822 struct cse_reg_info *p;
823
824 for (p = *hash_head; p != NULL; p = p->hash_next)
825 if (p->regno == regno)
826 break;
827
828 if (p == NULL)
829 {
830 /* Get a new cse_reg_info structure. */
831 if (cse_reg_info_free_list)
832 {
833 p = cse_reg_info_free_list;
834 cse_reg_info_free_list = p->next;
835 }
836 else
837 p = xmalloc (sizeof (struct cse_reg_info));
838
839 /* Insert into hash table. */
840 p->hash_next = *hash_head;
841 *hash_head = p;
842
843 /* Initialize it. */
844 p->reg_tick = 1;
845 p->reg_in_table = -1;
846 p->subreg_ticked = -1;
847 p->reg_qty = regno;
848 p->regno = regno;
849 p->next = cse_reg_info_used_list;
850 cse_reg_info_used_list = p;
851 if (!cse_reg_info_used_list_end)
852 cse_reg_info_used_list_end = p;
853 }
854
855 /* Cache this lookup; we tend to be looking up information about the
856 same register several times in a row. */
857 cached_regno = regno;
858 cached_cse_reg_info = p;
859
860 return p;
861 }
862
863 /* Clear the hash table and initialize each register with its own quantity,
864 for a new basic block. */
865
866 static void
867 new_basic_block (void)
868 {
869 int i;
870
871 next_qty = max_reg;
872
873 /* Clear out hash table state for this pass. */
874
875 memset (reg_hash, 0, sizeof reg_hash);
876
877 if (cse_reg_info_used_list)
878 {
879 cse_reg_info_used_list_end->next = cse_reg_info_free_list;
880 cse_reg_info_free_list = cse_reg_info_used_list;
881 cse_reg_info_used_list = cse_reg_info_used_list_end = 0;
882 }
883 cached_cse_reg_info = 0;
884
885 CLEAR_HARD_REG_SET (hard_regs_in_table);
886
887 /* The per-quantity values used to be initialized here, but it is
888 much faster to initialize each as it is made in `make_new_qty'. */
889
890 for (i = 0; i < HASH_SIZE; i++)
891 {
892 struct table_elt *first;
893
894 first = table[i];
895 if (first != NULL)
896 {
897 struct table_elt *last = first;
898
899 table[i] = NULL;
900
901 while (last->next_same_hash != NULL)
902 last = last->next_same_hash;
903
904 /* Now relink this hash entire chain into
905 the free element list. */
906
907 last->next_same_hash = free_element_chain;
908 free_element_chain = first;
909 }
910 }
911
912 #ifdef HAVE_cc0
913 prev_insn = 0;
914 prev_insn_cc0 = 0;
915 #endif
916 }
917
918 /* Say that register REG contains a quantity in mode MODE not in any
919 register before and initialize that quantity. */
920
921 static void
922 make_new_qty (unsigned int reg, enum machine_mode mode)
923 {
924 int q;
925 struct qty_table_elem *ent;
926 struct reg_eqv_elem *eqv;
927
928 gcc_assert (next_qty < max_qty);
929
930 q = REG_QTY (reg) = next_qty++;
931 ent = &qty_table[q];
932 ent->first_reg = reg;
933 ent->last_reg = reg;
934 ent->mode = mode;
935 ent->const_rtx = ent->const_insn = NULL_RTX;
936 ent->comparison_code = UNKNOWN;
937
938 eqv = &reg_eqv_table[reg];
939 eqv->next = eqv->prev = -1;
940 }
941
942 /* Make reg NEW equivalent to reg OLD.
943 OLD is not changing; NEW is. */
944
945 static void
946 make_regs_eqv (unsigned int new, unsigned int old)
947 {
948 unsigned int lastr, firstr;
949 int q = REG_QTY (old);
950 struct qty_table_elem *ent;
951
952 ent = &qty_table[q];
953
954 /* Nothing should become eqv until it has a "non-invalid" qty number. */
955 gcc_assert (REGNO_QTY_VALID_P (old));
956
957 REG_QTY (new) = q;
958 firstr = ent->first_reg;
959 lastr = ent->last_reg;
960
961 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
962 hard regs. Among pseudos, if NEW will live longer than any other reg
963 of the same qty, and that is beyond the current basic block,
964 make it the new canonical replacement for this qty. */
965 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
966 /* Certain fixed registers might be of the class NO_REGS. This means
967 that not only can they not be allocated by the compiler, but
968 they cannot be used in substitutions or canonicalizations
969 either. */
970 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
971 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
972 || (new >= FIRST_PSEUDO_REGISTER
973 && (firstr < FIRST_PSEUDO_REGISTER
974 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
975 || (uid_cuid[REGNO_FIRST_UID (new)]
976 < cse_basic_block_start))
977 && (uid_cuid[REGNO_LAST_UID (new)]
978 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
979 {
980 reg_eqv_table[firstr].prev = new;
981 reg_eqv_table[new].next = firstr;
982 reg_eqv_table[new].prev = -1;
983 ent->first_reg = new;
984 }
985 else
986 {
987 /* If NEW is a hard reg (known to be non-fixed), insert at end.
988 Otherwise, insert before any non-fixed hard regs that are at the
989 end. Registers of class NO_REGS cannot be used as an
990 equivalent for anything. */
991 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
992 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
993 && new >= FIRST_PSEUDO_REGISTER)
994 lastr = reg_eqv_table[lastr].prev;
995 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
996 if (reg_eqv_table[lastr].next >= 0)
997 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
998 else
999 qty_table[q].last_reg = new;
1000 reg_eqv_table[lastr].next = new;
1001 reg_eqv_table[new].prev = lastr;
1002 }
1003 }
1004
1005 /* Remove REG from its equivalence class. */
1006
1007 static void
1008 delete_reg_equiv (unsigned int reg)
1009 {
1010 struct qty_table_elem *ent;
1011 int q = REG_QTY (reg);
1012 int p, n;
1013
1014 /* If invalid, do nothing. */
1015 if (q == (int) reg)
1016 return;
1017
1018 ent = &qty_table[q];
1019
1020 p = reg_eqv_table[reg].prev;
1021 n = reg_eqv_table[reg].next;
1022
1023 if (n != -1)
1024 reg_eqv_table[n].prev = p;
1025 else
1026 ent->last_reg = p;
1027 if (p != -1)
1028 reg_eqv_table[p].next = n;
1029 else
1030 ent->first_reg = n;
1031
1032 REG_QTY (reg) = reg;
1033 }
1034
1035 /* Remove any invalid expressions from the hash table
1036 that refer to any of the registers contained in expression X.
1037
1038 Make sure that newly inserted references to those registers
1039 as subexpressions will be considered valid.
1040
1041 mention_regs is not called when a register itself
1042 is being stored in the table.
1043
1044 Return 1 if we have done something that may have changed the hash code
1045 of X. */
1046
1047 static int
1048 mention_regs (rtx x)
1049 {
1050 enum rtx_code code;
1051 int i, j;
1052 const char *fmt;
1053 int changed = 0;
1054
1055 if (x == 0)
1056 return 0;
1057
1058 code = GET_CODE (x);
1059 if (code == REG)
1060 {
1061 unsigned int regno = REGNO (x);
1062 unsigned int endregno
1063 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1064 : hard_regno_nregs[regno][GET_MODE (x)]);
1065 unsigned int i;
1066
1067 for (i = regno; i < endregno; i++)
1068 {
1069 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1070 remove_invalid_refs (i);
1071
1072 REG_IN_TABLE (i) = REG_TICK (i);
1073 SUBREG_TICKED (i) = -1;
1074 }
1075
1076 return 0;
1077 }
1078
1079 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1080 pseudo if they don't use overlapping words. We handle only pseudos
1081 here for simplicity. */
1082 if (code == SUBREG && REG_P (SUBREG_REG (x))
1083 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1084 {
1085 unsigned int i = REGNO (SUBREG_REG (x));
1086
1087 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1088 {
1089 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1090 the last store to this register really stored into this
1091 subreg, then remove the memory of this subreg.
1092 Otherwise, remove any memory of the entire register and
1093 all its subregs from the table. */
1094 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1095 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1096 remove_invalid_refs (i);
1097 else
1098 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1099 }
1100
1101 REG_IN_TABLE (i) = REG_TICK (i);
1102 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1103 return 0;
1104 }
1105
1106 /* If X is a comparison or a COMPARE and either operand is a register
1107 that does not have a quantity, give it one. This is so that a later
1108 call to record_jump_equiv won't cause X to be assigned a different
1109 hash code and not found in the table after that call.
1110
1111 It is not necessary to do this here, since rehash_using_reg can
1112 fix up the table later, but doing this here eliminates the need to
1113 call that expensive function in the most common case where the only
1114 use of the register is in the comparison. */
1115
1116 if (code == COMPARE || COMPARISON_P (x))
1117 {
1118 if (REG_P (XEXP (x, 0))
1119 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1120 if (insert_regs (XEXP (x, 0), NULL, 0))
1121 {
1122 rehash_using_reg (XEXP (x, 0));
1123 changed = 1;
1124 }
1125
1126 if (REG_P (XEXP (x, 1))
1127 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1128 if (insert_regs (XEXP (x, 1), NULL, 0))
1129 {
1130 rehash_using_reg (XEXP (x, 1));
1131 changed = 1;
1132 }
1133 }
1134
1135 fmt = GET_RTX_FORMAT (code);
1136 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1137 if (fmt[i] == 'e')
1138 changed |= mention_regs (XEXP (x, i));
1139 else if (fmt[i] == 'E')
1140 for (j = 0; j < XVECLEN (x, i); j++)
1141 changed |= mention_regs (XVECEXP (x, i, j));
1142
1143 return changed;
1144 }
1145
1146 /* Update the register quantities for inserting X into the hash table
1147 with a value equivalent to CLASSP.
1148 (If the class does not contain a REG, it is irrelevant.)
1149 If MODIFIED is nonzero, X is a destination; it is being modified.
1150 Note that delete_reg_equiv should be called on a register
1151 before insert_regs is done on that register with MODIFIED != 0.
1152
1153 Nonzero value means that elements of reg_qty have changed
1154 so X's hash code may be different. */
1155
1156 static int
1157 insert_regs (rtx x, struct table_elt *classp, int modified)
1158 {
1159 if (REG_P (x))
1160 {
1161 unsigned int regno = REGNO (x);
1162 int qty_valid;
1163
1164 /* If REGNO is in the equivalence table already but is of the
1165 wrong mode for that equivalence, don't do anything here. */
1166
1167 qty_valid = REGNO_QTY_VALID_P (regno);
1168 if (qty_valid)
1169 {
1170 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1171
1172 if (ent->mode != GET_MODE (x))
1173 return 0;
1174 }
1175
1176 if (modified || ! qty_valid)
1177 {
1178 if (classp)
1179 for (classp = classp->first_same_value;
1180 classp != 0;
1181 classp = classp->next_same_value)
1182 if (REG_P (classp->exp)
1183 && GET_MODE (classp->exp) == GET_MODE (x))
1184 {
1185 make_regs_eqv (regno, REGNO (classp->exp));
1186 return 1;
1187 }
1188
1189 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1190 than REG_IN_TABLE to find out if there was only a single preceding
1191 invalidation - for the SUBREG - or another one, which would be
1192 for the full register. However, if we find here that REG_TICK
1193 indicates that the register is invalid, it means that it has
1194 been invalidated in a separate operation. The SUBREG might be used
1195 now (then this is a recursive call), or we might use the full REG
1196 now and a SUBREG of it later. So bump up REG_TICK so that
1197 mention_regs will do the right thing. */
1198 if (! modified
1199 && REG_IN_TABLE (regno) >= 0
1200 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1201 REG_TICK (regno)++;
1202 make_new_qty (regno, GET_MODE (x));
1203 return 1;
1204 }
1205
1206 return 0;
1207 }
1208
1209 /* If X is a SUBREG, we will likely be inserting the inner register in the
1210 table. If that register doesn't have an assigned quantity number at
1211 this point but does later, the insertion that we will be doing now will
1212 not be accessible because its hash code will have changed. So assign
1213 a quantity number now. */
1214
1215 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1216 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1217 {
1218 insert_regs (SUBREG_REG (x), NULL, 0);
1219 mention_regs (x);
1220 return 1;
1221 }
1222 else
1223 return mention_regs (x);
1224 }
1225 \f
1226 /* Look in or update the hash table. */
1227
1228 /* Remove table element ELT from use in the table.
1229 HASH is its hash code, made using the HASH macro.
1230 It's an argument because often that is known in advance
1231 and we save much time not recomputing it. */
1232
1233 static void
1234 remove_from_table (struct table_elt *elt, unsigned int hash)
1235 {
1236 if (elt == 0)
1237 return;
1238
1239 /* Mark this element as removed. See cse_insn. */
1240 elt->first_same_value = 0;
1241
1242 /* Remove the table element from its equivalence class. */
1243
1244 {
1245 struct table_elt *prev = elt->prev_same_value;
1246 struct table_elt *next = elt->next_same_value;
1247
1248 if (next)
1249 next->prev_same_value = prev;
1250
1251 if (prev)
1252 prev->next_same_value = next;
1253 else
1254 {
1255 struct table_elt *newfirst = next;
1256 while (next)
1257 {
1258 next->first_same_value = newfirst;
1259 next = next->next_same_value;
1260 }
1261 }
1262 }
1263
1264 /* Remove the table element from its hash bucket. */
1265
1266 {
1267 struct table_elt *prev = elt->prev_same_hash;
1268 struct table_elt *next = elt->next_same_hash;
1269
1270 if (next)
1271 next->prev_same_hash = prev;
1272
1273 if (prev)
1274 prev->next_same_hash = next;
1275 else if (table[hash] == elt)
1276 table[hash] = next;
1277 else
1278 {
1279 /* This entry is not in the proper hash bucket. This can happen
1280 when two classes were merged by `merge_equiv_classes'. Search
1281 for the hash bucket that it heads. This happens only very
1282 rarely, so the cost is acceptable. */
1283 for (hash = 0; hash < HASH_SIZE; hash++)
1284 if (table[hash] == elt)
1285 table[hash] = next;
1286 }
1287 }
1288
1289 /* Remove the table element from its related-value circular chain. */
1290
1291 if (elt->related_value != 0 && elt->related_value != elt)
1292 {
1293 struct table_elt *p = elt->related_value;
1294
1295 while (p->related_value != elt)
1296 p = p->related_value;
1297 p->related_value = elt->related_value;
1298 if (p->related_value == p)
1299 p->related_value = 0;
1300 }
1301
1302 /* Now add it to the free element chain. */
1303 elt->next_same_hash = free_element_chain;
1304 free_element_chain = elt;
1305 }
1306
1307 /* Look up X in the hash table and return its table element,
1308 or 0 if X is not in the table.
1309
1310 MODE is the machine-mode of X, or if X is an integer constant
1311 with VOIDmode then MODE is the mode with which X will be used.
1312
1313 Here we are satisfied to find an expression whose tree structure
1314 looks like X. */
1315
1316 static struct table_elt *
1317 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1318 {
1319 struct table_elt *p;
1320
1321 for (p = table[hash]; p; p = p->next_same_hash)
1322 if (mode == p->mode && ((x == p->exp && REG_P (x))
1323 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1324 return p;
1325
1326 return 0;
1327 }
1328
1329 /* Like `lookup' but don't care whether the table element uses invalid regs.
1330 Also ignore discrepancies in the machine mode of a register. */
1331
1332 static struct table_elt *
1333 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1334 {
1335 struct table_elt *p;
1336
1337 if (REG_P (x))
1338 {
1339 unsigned int regno = REGNO (x);
1340
1341 /* Don't check the machine mode when comparing registers;
1342 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1343 for (p = table[hash]; p; p = p->next_same_hash)
1344 if (REG_P (p->exp)
1345 && REGNO (p->exp) == regno)
1346 return p;
1347 }
1348 else
1349 {
1350 for (p = table[hash]; p; p = p->next_same_hash)
1351 if (mode == p->mode
1352 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1353 return p;
1354 }
1355
1356 return 0;
1357 }
1358
1359 /* Look for an expression equivalent to X and with code CODE.
1360 If one is found, return that expression. */
1361
1362 static rtx
1363 lookup_as_function (rtx x, enum rtx_code code)
1364 {
1365 struct table_elt *p
1366 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1367
1368 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1369 long as we are narrowing. So if we looked in vain for a mode narrower
1370 than word_mode before, look for word_mode now. */
1371 if (p == 0 && code == CONST_INT
1372 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1373 {
1374 x = copy_rtx (x);
1375 PUT_MODE (x, word_mode);
1376 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1377 }
1378
1379 if (p == 0)
1380 return 0;
1381
1382 for (p = p->first_same_value; p; p = p->next_same_value)
1383 if (GET_CODE (p->exp) == code
1384 /* Make sure this is a valid entry in the table. */
1385 && exp_equiv_p (p->exp, p->exp, 1, false))
1386 return p->exp;
1387
1388 return 0;
1389 }
1390
1391 /* Insert X in the hash table, assuming HASH is its hash code
1392 and CLASSP is an element of the class it should go in
1393 (or 0 if a new class should be made).
1394 It is inserted at the proper position to keep the class in
1395 the order cheapest first.
1396
1397 MODE is the machine-mode of X, or if X is an integer constant
1398 with VOIDmode then MODE is the mode with which X will be used.
1399
1400 For elements of equal cheapness, the most recent one
1401 goes in front, except that the first element in the list
1402 remains first unless a cheaper element is added. The order of
1403 pseudo-registers does not matter, as canon_reg will be called to
1404 find the cheapest when a register is retrieved from the table.
1405
1406 The in_memory field in the hash table element is set to 0.
1407 The caller must set it nonzero if appropriate.
1408
1409 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1410 and if insert_regs returns a nonzero value
1411 you must then recompute its hash code before calling here.
1412
1413 If necessary, update table showing constant values of quantities. */
1414
1415 #define CHEAPER(X, Y) \
1416 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1417
1418 static struct table_elt *
1419 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1420 {
1421 struct table_elt *elt;
1422
1423 /* If X is a register and we haven't made a quantity for it,
1424 something is wrong. */
1425 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1426
1427 /* If X is a hard register, show it is being put in the table. */
1428 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1429 {
1430 unsigned int regno = REGNO (x);
1431 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1432 unsigned int i;
1433
1434 for (i = regno; i < endregno; i++)
1435 SET_HARD_REG_BIT (hard_regs_in_table, i);
1436 }
1437
1438 /* Put an element for X into the right hash bucket. */
1439
1440 elt = free_element_chain;
1441 if (elt)
1442 free_element_chain = elt->next_same_hash;
1443 else
1444 {
1445 n_elements_made++;
1446 elt = xmalloc (sizeof (struct table_elt));
1447 }
1448
1449 elt->exp = x;
1450 elt->canon_exp = NULL_RTX;
1451 elt->cost = COST (x);
1452 elt->regcost = approx_reg_cost (x);
1453 elt->next_same_value = 0;
1454 elt->prev_same_value = 0;
1455 elt->next_same_hash = table[hash];
1456 elt->prev_same_hash = 0;
1457 elt->related_value = 0;
1458 elt->in_memory = 0;
1459 elt->mode = mode;
1460 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1461
1462 if (table[hash])
1463 table[hash]->prev_same_hash = elt;
1464 table[hash] = elt;
1465
1466 /* Put it into the proper value-class. */
1467 if (classp)
1468 {
1469 classp = classp->first_same_value;
1470 if (CHEAPER (elt, classp))
1471 /* Insert at the head of the class. */
1472 {
1473 struct table_elt *p;
1474 elt->next_same_value = classp;
1475 classp->prev_same_value = elt;
1476 elt->first_same_value = elt;
1477
1478 for (p = classp; p; p = p->next_same_value)
1479 p->first_same_value = elt;
1480 }
1481 else
1482 {
1483 /* Insert not at head of the class. */
1484 /* Put it after the last element cheaper than X. */
1485 struct table_elt *p, *next;
1486
1487 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1488 p = next);
1489
1490 /* Put it after P and before NEXT. */
1491 elt->next_same_value = next;
1492 if (next)
1493 next->prev_same_value = elt;
1494
1495 elt->prev_same_value = p;
1496 p->next_same_value = elt;
1497 elt->first_same_value = classp;
1498 }
1499 }
1500 else
1501 elt->first_same_value = elt;
1502
1503 /* If this is a constant being set equivalent to a register or a register
1504 being set equivalent to a constant, note the constant equivalence.
1505
1506 If this is a constant, it cannot be equivalent to a different constant,
1507 and a constant is the only thing that can be cheaper than a register. So
1508 we know the register is the head of the class (before the constant was
1509 inserted).
1510
1511 If this is a register that is not already known equivalent to a
1512 constant, we must check the entire class.
1513
1514 If this is a register that is already known equivalent to an insn,
1515 update the qtys `const_insn' to show that `this_insn' is the latest
1516 insn making that quantity equivalent to the constant. */
1517
1518 if (elt->is_const && classp && REG_P (classp->exp)
1519 && !REG_P (x))
1520 {
1521 int exp_q = REG_QTY (REGNO (classp->exp));
1522 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1523
1524 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1525 exp_ent->const_insn = this_insn;
1526 }
1527
1528 else if (REG_P (x)
1529 && classp
1530 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1531 && ! elt->is_const)
1532 {
1533 struct table_elt *p;
1534
1535 for (p = classp; p != 0; p = p->next_same_value)
1536 {
1537 if (p->is_const && !REG_P (p->exp))
1538 {
1539 int x_q = REG_QTY (REGNO (x));
1540 struct qty_table_elem *x_ent = &qty_table[x_q];
1541
1542 x_ent->const_rtx
1543 = gen_lowpart (GET_MODE (x), p->exp);
1544 x_ent->const_insn = this_insn;
1545 break;
1546 }
1547 }
1548 }
1549
1550 else if (REG_P (x)
1551 && qty_table[REG_QTY (REGNO (x))].const_rtx
1552 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1553 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1554
1555 /* If this is a constant with symbolic value,
1556 and it has a term with an explicit integer value,
1557 link it up with related expressions. */
1558 if (GET_CODE (x) == CONST)
1559 {
1560 rtx subexp = get_related_value (x);
1561 unsigned subhash;
1562 struct table_elt *subelt, *subelt_prev;
1563
1564 if (subexp != 0)
1565 {
1566 /* Get the integer-free subexpression in the hash table. */
1567 subhash = SAFE_HASH (subexp, mode);
1568 subelt = lookup (subexp, subhash, mode);
1569 if (subelt == 0)
1570 subelt = insert (subexp, NULL, subhash, mode);
1571 /* Initialize SUBELT's circular chain if it has none. */
1572 if (subelt->related_value == 0)
1573 subelt->related_value = subelt;
1574 /* Find the element in the circular chain that precedes SUBELT. */
1575 subelt_prev = subelt;
1576 while (subelt_prev->related_value != subelt)
1577 subelt_prev = subelt_prev->related_value;
1578 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1579 This way the element that follows SUBELT is the oldest one. */
1580 elt->related_value = subelt_prev->related_value;
1581 subelt_prev->related_value = elt;
1582 }
1583 }
1584
1585 return elt;
1586 }
1587 \f
1588 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1589 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1590 the two classes equivalent.
1591
1592 CLASS1 will be the surviving class; CLASS2 should not be used after this
1593 call.
1594
1595 Any invalid entries in CLASS2 will not be copied. */
1596
1597 static void
1598 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1599 {
1600 struct table_elt *elt, *next, *new;
1601
1602 /* Ensure we start with the head of the classes. */
1603 class1 = class1->first_same_value;
1604 class2 = class2->first_same_value;
1605
1606 /* If they were already equal, forget it. */
1607 if (class1 == class2)
1608 return;
1609
1610 for (elt = class2; elt; elt = next)
1611 {
1612 unsigned int hash;
1613 rtx exp = elt->exp;
1614 enum machine_mode mode = elt->mode;
1615
1616 next = elt->next_same_value;
1617
1618 /* Remove old entry, make a new one in CLASS1's class.
1619 Don't do this for invalid entries as we cannot find their
1620 hash code (it also isn't necessary). */
1621 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1622 {
1623 bool need_rehash = false;
1624
1625 hash_arg_in_memory = 0;
1626 hash = HASH (exp, mode);
1627
1628 if (REG_P (exp))
1629 {
1630 need_rehash = (unsigned) REG_QTY (REGNO (exp)) != REGNO (exp);
1631 delete_reg_equiv (REGNO (exp));
1632 }
1633
1634 remove_from_table (elt, hash);
1635
1636 if (insert_regs (exp, class1, 0) || need_rehash)
1637 {
1638 rehash_using_reg (exp);
1639 hash = HASH (exp, mode);
1640 }
1641 new = insert (exp, class1, hash, mode);
1642 new->in_memory = hash_arg_in_memory;
1643 }
1644 }
1645 }
1646 \f
1647 /* Flush the entire hash table. */
1648
1649 static void
1650 flush_hash_table (void)
1651 {
1652 int i;
1653 struct table_elt *p;
1654
1655 for (i = 0; i < HASH_SIZE; i++)
1656 for (p = table[i]; p; p = table[i])
1657 {
1658 /* Note that invalidate can remove elements
1659 after P in the current hash chain. */
1660 if (REG_P (p->exp))
1661 invalidate (p->exp, p->mode);
1662 else
1663 remove_from_table (p, i);
1664 }
1665 }
1666 \f
1667 /* Function called for each rtx to check whether true dependence exist. */
1668 struct check_dependence_data
1669 {
1670 enum machine_mode mode;
1671 rtx exp;
1672 rtx addr;
1673 };
1674
1675 static int
1676 check_dependence (rtx *x, void *data)
1677 {
1678 struct check_dependence_data *d = (struct check_dependence_data *) data;
1679 if (*x && MEM_P (*x))
1680 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1681 cse_rtx_varies_p);
1682 else
1683 return 0;
1684 }
1685 \f
1686 /* Remove from the hash table, or mark as invalid, all expressions whose
1687 values could be altered by storing in X. X is a register, a subreg, or
1688 a memory reference with nonvarying address (because, when a memory
1689 reference with a varying address is stored in, all memory references are
1690 removed by invalidate_memory so specific invalidation is superfluous).
1691 FULL_MODE, if not VOIDmode, indicates that this much should be
1692 invalidated instead of just the amount indicated by the mode of X. This
1693 is only used for bitfield stores into memory.
1694
1695 A nonvarying address may be just a register or just a symbol reference,
1696 or it may be either of those plus a numeric offset. */
1697
1698 static void
1699 invalidate (rtx x, enum machine_mode full_mode)
1700 {
1701 int i;
1702 struct table_elt *p;
1703 rtx addr;
1704
1705 switch (GET_CODE (x))
1706 {
1707 case REG:
1708 {
1709 /* If X is a register, dependencies on its contents are recorded
1710 through the qty number mechanism. Just change the qty number of
1711 the register, mark it as invalid for expressions that refer to it,
1712 and remove it itself. */
1713 unsigned int regno = REGNO (x);
1714 unsigned int hash = HASH (x, GET_MODE (x));
1715
1716 /* Remove REGNO from any quantity list it might be on and indicate
1717 that its value might have changed. If it is a pseudo, remove its
1718 entry from the hash table.
1719
1720 For a hard register, we do the first two actions above for any
1721 additional hard registers corresponding to X. Then, if any of these
1722 registers are in the table, we must remove any REG entries that
1723 overlap these registers. */
1724
1725 delete_reg_equiv (regno);
1726 REG_TICK (regno)++;
1727 SUBREG_TICKED (regno) = -1;
1728
1729 if (regno >= FIRST_PSEUDO_REGISTER)
1730 {
1731 /* Because a register can be referenced in more than one mode,
1732 we might have to remove more than one table entry. */
1733 struct table_elt *elt;
1734
1735 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1736 remove_from_table (elt, hash);
1737 }
1738 else
1739 {
1740 HOST_WIDE_INT in_table
1741 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1742 unsigned int endregno
1743 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1744 unsigned int tregno, tendregno, rn;
1745 struct table_elt *p, *next;
1746
1747 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1748
1749 for (rn = regno + 1; rn < endregno; rn++)
1750 {
1751 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1752 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1753 delete_reg_equiv (rn);
1754 REG_TICK (rn)++;
1755 SUBREG_TICKED (rn) = -1;
1756 }
1757
1758 if (in_table)
1759 for (hash = 0; hash < HASH_SIZE; hash++)
1760 for (p = table[hash]; p; p = next)
1761 {
1762 next = p->next_same_hash;
1763
1764 if (!REG_P (p->exp)
1765 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1766 continue;
1767
1768 tregno = REGNO (p->exp);
1769 tendregno
1770 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1771 if (tendregno > regno && tregno < endregno)
1772 remove_from_table (p, hash);
1773 }
1774 }
1775 }
1776 return;
1777
1778 case SUBREG:
1779 invalidate (SUBREG_REG (x), VOIDmode);
1780 return;
1781
1782 case PARALLEL:
1783 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1784 invalidate (XVECEXP (x, 0, i), VOIDmode);
1785 return;
1786
1787 case EXPR_LIST:
1788 /* This is part of a disjoint return value; extract the location in
1789 question ignoring the offset. */
1790 invalidate (XEXP (x, 0), VOIDmode);
1791 return;
1792
1793 case MEM:
1794 addr = canon_rtx (get_addr (XEXP (x, 0)));
1795 /* Calculate the canonical version of X here so that
1796 true_dependence doesn't generate new RTL for X on each call. */
1797 x = canon_rtx (x);
1798
1799 /* Remove all hash table elements that refer to overlapping pieces of
1800 memory. */
1801 if (full_mode == VOIDmode)
1802 full_mode = GET_MODE (x);
1803
1804 for (i = 0; i < HASH_SIZE; i++)
1805 {
1806 struct table_elt *next;
1807
1808 for (p = table[i]; p; p = next)
1809 {
1810 next = p->next_same_hash;
1811 if (p->in_memory)
1812 {
1813 struct check_dependence_data d;
1814
1815 /* Just canonicalize the expression once;
1816 otherwise each time we call invalidate
1817 true_dependence will canonicalize the
1818 expression again. */
1819 if (!p->canon_exp)
1820 p->canon_exp = canon_rtx (p->exp);
1821 d.exp = x;
1822 d.addr = addr;
1823 d.mode = full_mode;
1824 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1825 remove_from_table (p, i);
1826 }
1827 }
1828 }
1829 return;
1830
1831 default:
1832 gcc_unreachable ();
1833 }
1834 }
1835 \f
1836 /* Remove all expressions that refer to register REGNO,
1837 since they are already invalid, and we are about to
1838 mark that register valid again and don't want the old
1839 expressions to reappear as valid. */
1840
1841 static void
1842 remove_invalid_refs (unsigned int regno)
1843 {
1844 unsigned int i;
1845 struct table_elt *p, *next;
1846
1847 for (i = 0; i < HASH_SIZE; i++)
1848 for (p = table[i]; p; p = next)
1849 {
1850 next = p->next_same_hash;
1851 if (!REG_P (p->exp)
1852 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1853 remove_from_table (p, i);
1854 }
1855 }
1856
1857 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1858 and mode MODE. */
1859 static void
1860 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1861 enum machine_mode mode)
1862 {
1863 unsigned int i;
1864 struct table_elt *p, *next;
1865 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1866
1867 for (i = 0; i < HASH_SIZE; i++)
1868 for (p = table[i]; p; p = next)
1869 {
1870 rtx exp = p->exp;
1871 next = p->next_same_hash;
1872
1873 if (!REG_P (exp)
1874 && (GET_CODE (exp) != SUBREG
1875 || !REG_P (SUBREG_REG (exp))
1876 || REGNO (SUBREG_REG (exp)) != regno
1877 || (((SUBREG_BYTE (exp)
1878 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1879 && SUBREG_BYTE (exp) <= end))
1880 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1881 remove_from_table (p, i);
1882 }
1883 }
1884 \f
1885 /* Recompute the hash codes of any valid entries in the hash table that
1886 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1887
1888 This is called when we make a jump equivalence. */
1889
1890 static void
1891 rehash_using_reg (rtx x)
1892 {
1893 unsigned int i;
1894 struct table_elt *p, *next;
1895 unsigned hash;
1896
1897 if (GET_CODE (x) == SUBREG)
1898 x = SUBREG_REG (x);
1899
1900 /* If X is not a register or if the register is known not to be in any
1901 valid entries in the table, we have no work to do. */
1902
1903 if (!REG_P (x)
1904 || REG_IN_TABLE (REGNO (x)) < 0
1905 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1906 return;
1907
1908 /* Scan all hash chains looking for valid entries that mention X.
1909 If we find one and it is in the wrong hash chain, move it. */
1910
1911 for (i = 0; i < HASH_SIZE; i++)
1912 for (p = table[i]; p; p = next)
1913 {
1914 next = p->next_same_hash;
1915 if (reg_mentioned_p (x, p->exp)
1916 && exp_equiv_p (p->exp, p->exp, 1, false)
1917 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1918 {
1919 if (p->next_same_hash)
1920 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1921
1922 if (p->prev_same_hash)
1923 p->prev_same_hash->next_same_hash = p->next_same_hash;
1924 else
1925 table[i] = p->next_same_hash;
1926
1927 p->next_same_hash = table[hash];
1928 p->prev_same_hash = 0;
1929 if (table[hash])
1930 table[hash]->prev_same_hash = p;
1931 table[hash] = p;
1932 }
1933 }
1934 }
1935 \f
1936 /* Remove from the hash table any expression that is a call-clobbered
1937 register. Also update their TICK values. */
1938
1939 static void
1940 invalidate_for_call (void)
1941 {
1942 unsigned int regno, endregno;
1943 unsigned int i;
1944 unsigned hash;
1945 struct table_elt *p, *next;
1946 int in_table = 0;
1947
1948 /* Go through all the hard registers. For each that is clobbered in
1949 a CALL_INSN, remove the register from quantity chains and update
1950 reg_tick if defined. Also see if any of these registers is currently
1951 in the table. */
1952
1953 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1954 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
1955 {
1956 delete_reg_equiv (regno);
1957 if (REG_TICK (regno) >= 0)
1958 {
1959 REG_TICK (regno)++;
1960 SUBREG_TICKED (regno) = -1;
1961 }
1962
1963 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
1964 }
1965
1966 /* In the case where we have no call-clobbered hard registers in the
1967 table, we are done. Otherwise, scan the table and remove any
1968 entry that overlaps a call-clobbered register. */
1969
1970 if (in_table)
1971 for (hash = 0; hash < HASH_SIZE; hash++)
1972 for (p = table[hash]; p; p = next)
1973 {
1974 next = p->next_same_hash;
1975
1976 if (!REG_P (p->exp)
1977 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1978 continue;
1979
1980 regno = REGNO (p->exp);
1981 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
1982
1983 for (i = regno; i < endregno; i++)
1984 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
1985 {
1986 remove_from_table (p, hash);
1987 break;
1988 }
1989 }
1990 }
1991 \f
1992 /* Given an expression X of type CONST,
1993 and ELT which is its table entry (or 0 if it
1994 is not in the hash table),
1995 return an alternate expression for X as a register plus integer.
1996 If none can be found, return 0. */
1997
1998 static rtx
1999 use_related_value (rtx x, struct table_elt *elt)
2000 {
2001 struct table_elt *relt = 0;
2002 struct table_elt *p, *q;
2003 HOST_WIDE_INT offset;
2004
2005 /* First, is there anything related known?
2006 If we have a table element, we can tell from that.
2007 Otherwise, must look it up. */
2008
2009 if (elt != 0 && elt->related_value != 0)
2010 relt = elt;
2011 else if (elt == 0 && GET_CODE (x) == CONST)
2012 {
2013 rtx subexp = get_related_value (x);
2014 if (subexp != 0)
2015 relt = lookup (subexp,
2016 SAFE_HASH (subexp, GET_MODE (subexp)),
2017 GET_MODE (subexp));
2018 }
2019
2020 if (relt == 0)
2021 return 0;
2022
2023 /* Search all related table entries for one that has an
2024 equivalent register. */
2025
2026 p = relt;
2027 while (1)
2028 {
2029 /* This loop is strange in that it is executed in two different cases.
2030 The first is when X is already in the table. Then it is searching
2031 the RELATED_VALUE list of X's class (RELT). The second case is when
2032 X is not in the table. Then RELT points to a class for the related
2033 value.
2034
2035 Ensure that, whatever case we are in, that we ignore classes that have
2036 the same value as X. */
2037
2038 if (rtx_equal_p (x, p->exp))
2039 q = 0;
2040 else
2041 for (q = p->first_same_value; q; q = q->next_same_value)
2042 if (REG_P (q->exp))
2043 break;
2044
2045 if (q)
2046 break;
2047
2048 p = p->related_value;
2049
2050 /* We went all the way around, so there is nothing to be found.
2051 Alternatively, perhaps RELT was in the table for some other reason
2052 and it has no related values recorded. */
2053 if (p == relt || p == 0)
2054 break;
2055 }
2056
2057 if (q == 0)
2058 return 0;
2059
2060 offset = (get_integer_term (x) - get_integer_term (p->exp));
2061 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2062 return plus_constant (q->exp, offset);
2063 }
2064 \f
2065 /* Hash a string. Just add its bytes up. */
2066 static inline unsigned
2067 hash_rtx_string (const char *ps)
2068 {
2069 unsigned hash = 0;
2070 const unsigned char *p = (const unsigned char *) ps;
2071
2072 if (p)
2073 while (*p)
2074 hash += *p++;
2075
2076 return hash;
2077 }
2078
2079 /* Hash an rtx. We are careful to make sure the value is never negative.
2080 Equivalent registers hash identically.
2081 MODE is used in hashing for CONST_INTs only;
2082 otherwise the mode of X is used.
2083
2084 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2085
2086 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2087 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2088
2089 Note that cse_insn knows that the hash code of a MEM expression
2090 is just (int) MEM plus the hash code of the address. */
2091
2092 unsigned
2093 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2094 int *hash_arg_in_memory_p, bool have_reg_qty)
2095 {
2096 int i, j;
2097 unsigned hash = 0;
2098 enum rtx_code code;
2099 const char *fmt;
2100
2101 /* Used to turn recursion into iteration. We can't rely on GCC's
2102 tail-recursion elimination since we need to keep accumulating values
2103 in HASH. */
2104 repeat:
2105 if (x == 0)
2106 return hash;
2107
2108 code = GET_CODE (x);
2109 switch (code)
2110 {
2111 case REG:
2112 {
2113 unsigned int regno = REGNO (x);
2114
2115 if (!reload_completed)
2116 {
2117 /* On some machines, we can't record any non-fixed hard register,
2118 because extending its life will cause reload problems. We
2119 consider ap, fp, sp, gp to be fixed for this purpose.
2120
2121 We also consider CCmode registers to be fixed for this purpose;
2122 failure to do so leads to failure to simplify 0<100 type of
2123 conditionals.
2124
2125 On all machines, we can't record any global registers.
2126 Nor should we record any register that is in a small
2127 class, as defined by CLASS_LIKELY_SPILLED_P. */
2128 bool record;
2129
2130 if (regno >= FIRST_PSEUDO_REGISTER)
2131 record = true;
2132 else if (x == frame_pointer_rtx
2133 || x == hard_frame_pointer_rtx
2134 || x == arg_pointer_rtx
2135 || x == stack_pointer_rtx
2136 || x == pic_offset_table_rtx)
2137 record = true;
2138 else if (global_regs[regno])
2139 record = false;
2140 else if (fixed_regs[regno])
2141 record = true;
2142 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2143 record = true;
2144 else if (SMALL_REGISTER_CLASSES)
2145 record = false;
2146 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2147 record = false;
2148 else
2149 record = true;
2150
2151 if (!record)
2152 {
2153 *do_not_record_p = 1;
2154 return 0;
2155 }
2156 }
2157
2158 hash += ((unsigned int) REG << 7);
2159 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2160 return hash;
2161 }
2162
2163 /* We handle SUBREG of a REG specially because the underlying
2164 reg changes its hash value with every value change; we don't
2165 want to have to forget unrelated subregs when one subreg changes. */
2166 case SUBREG:
2167 {
2168 if (REG_P (SUBREG_REG (x)))
2169 {
2170 hash += (((unsigned int) SUBREG << 7)
2171 + REGNO (SUBREG_REG (x))
2172 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2173 return hash;
2174 }
2175 break;
2176 }
2177
2178 case CONST_INT:
2179 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2180 + (unsigned int) INTVAL (x));
2181 return hash;
2182
2183 case CONST_DOUBLE:
2184 /* This is like the general case, except that it only counts
2185 the integers representing the constant. */
2186 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2187 if (GET_MODE (x) != VOIDmode)
2188 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2189 else
2190 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2191 + (unsigned int) CONST_DOUBLE_HIGH (x));
2192 return hash;
2193
2194 case CONST_VECTOR:
2195 {
2196 int units;
2197 rtx elt;
2198
2199 units = CONST_VECTOR_NUNITS (x);
2200
2201 for (i = 0; i < units; ++i)
2202 {
2203 elt = CONST_VECTOR_ELT (x, i);
2204 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2205 hash_arg_in_memory_p, have_reg_qty);
2206 }
2207
2208 return hash;
2209 }
2210
2211 /* Assume there is only one rtx object for any given label. */
2212 case LABEL_REF:
2213 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2214 differences and differences between each stage's debugging dumps. */
2215 hash += (((unsigned int) LABEL_REF << 7)
2216 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2217 return hash;
2218
2219 case SYMBOL_REF:
2220 {
2221 /* Don't hash on the symbol's address to avoid bootstrap differences.
2222 Different hash values may cause expressions to be recorded in
2223 different orders and thus different registers to be used in the
2224 final assembler. This also avoids differences in the dump files
2225 between various stages. */
2226 unsigned int h = 0;
2227 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2228
2229 while (*p)
2230 h += (h << 7) + *p++; /* ??? revisit */
2231
2232 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2233 return hash;
2234 }
2235
2236 case MEM:
2237 /* We don't record if marked volatile or if BLKmode since we don't
2238 know the size of the move. */
2239 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2240 {
2241 *do_not_record_p = 1;
2242 return 0;
2243 }
2244 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2245 *hash_arg_in_memory_p = 1;
2246
2247 /* Now that we have already found this special case,
2248 might as well speed it up as much as possible. */
2249 hash += (unsigned) MEM;
2250 x = XEXP (x, 0);
2251 goto repeat;
2252
2253 case USE:
2254 /* A USE that mentions non-volatile memory needs special
2255 handling since the MEM may be BLKmode which normally
2256 prevents an entry from being made. Pure calls are
2257 marked by a USE which mentions BLKmode memory.
2258 See calls.c:emit_call_1. */
2259 if (MEM_P (XEXP (x, 0))
2260 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2261 {
2262 hash += (unsigned) USE;
2263 x = XEXP (x, 0);
2264
2265 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2266 *hash_arg_in_memory_p = 1;
2267
2268 /* Now that we have already found this special case,
2269 might as well speed it up as much as possible. */
2270 hash += (unsigned) MEM;
2271 x = XEXP (x, 0);
2272 goto repeat;
2273 }
2274 break;
2275
2276 case PRE_DEC:
2277 case PRE_INC:
2278 case POST_DEC:
2279 case POST_INC:
2280 case PRE_MODIFY:
2281 case POST_MODIFY:
2282 case PC:
2283 case CC0:
2284 case CALL:
2285 case UNSPEC_VOLATILE:
2286 *do_not_record_p = 1;
2287 return 0;
2288
2289 case ASM_OPERANDS:
2290 if (MEM_VOLATILE_P (x))
2291 {
2292 *do_not_record_p = 1;
2293 return 0;
2294 }
2295 else
2296 {
2297 /* We don't want to take the filename and line into account. */
2298 hash += (unsigned) code + (unsigned) GET_MODE (x)
2299 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2300 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2301 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2302
2303 if (ASM_OPERANDS_INPUT_LENGTH (x))
2304 {
2305 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2306 {
2307 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2308 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2309 do_not_record_p, hash_arg_in_memory_p,
2310 have_reg_qty)
2311 + hash_rtx_string
2312 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2313 }
2314
2315 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2316 x = ASM_OPERANDS_INPUT (x, 0);
2317 mode = GET_MODE (x);
2318 goto repeat;
2319 }
2320
2321 return hash;
2322 }
2323 break;
2324
2325 default:
2326 break;
2327 }
2328
2329 i = GET_RTX_LENGTH (code) - 1;
2330 hash += (unsigned) code + (unsigned) GET_MODE (x);
2331 fmt = GET_RTX_FORMAT (code);
2332 for (; i >= 0; i--)
2333 {
2334 switch (fmt[i])
2335 {
2336 case 'e':
2337 /* If we are about to do the last recursive call
2338 needed at this level, change it into iteration.
2339 This function is called enough to be worth it. */
2340 if (i == 0)
2341 {
2342 x = XEXP (x, i);
2343 goto repeat;
2344 }
2345
2346 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2347 hash_arg_in_memory_p, have_reg_qty);
2348 break;
2349
2350 case 'E':
2351 for (j = 0; j < XVECLEN (x, i); j++)
2352 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2353 hash_arg_in_memory_p, have_reg_qty);
2354 break;
2355
2356 case 's':
2357 hash += hash_rtx_string (XSTR (x, i));
2358 break;
2359
2360 case 'i':
2361 hash += (unsigned int) XINT (x, i);
2362 break;
2363
2364 case '0': case 't':
2365 /* Unused. */
2366 break;
2367
2368 default:
2369 gcc_unreachable ();
2370 }
2371 }
2372
2373 return hash;
2374 }
2375
2376 /* Hash an rtx X for cse via hash_rtx.
2377 Stores 1 in do_not_record if any subexpression is volatile.
2378 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2379 does not have the RTX_UNCHANGING_P bit set. */
2380
2381 static inline unsigned
2382 canon_hash (rtx x, enum machine_mode mode)
2383 {
2384 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2385 }
2386
2387 /* Like canon_hash but with no side effects, i.e. do_not_record
2388 and hash_arg_in_memory are not changed. */
2389
2390 static inline unsigned
2391 safe_hash (rtx x, enum machine_mode mode)
2392 {
2393 int dummy_do_not_record;
2394 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2395 }
2396 \f
2397 /* Return 1 iff X and Y would canonicalize into the same thing,
2398 without actually constructing the canonicalization of either one.
2399 If VALIDATE is nonzero,
2400 we assume X is an expression being processed from the rtl
2401 and Y was found in the hash table. We check register refs
2402 in Y for being marked as valid.
2403
2404 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2405
2406 int
2407 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2408 {
2409 int i, j;
2410 enum rtx_code code;
2411 const char *fmt;
2412
2413 /* Note: it is incorrect to assume an expression is equivalent to itself
2414 if VALIDATE is nonzero. */
2415 if (x == y && !validate)
2416 return 1;
2417
2418 if (x == 0 || y == 0)
2419 return x == y;
2420
2421 code = GET_CODE (x);
2422 if (code != GET_CODE (y))
2423 return 0;
2424
2425 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2426 if (GET_MODE (x) != GET_MODE (y))
2427 return 0;
2428
2429 switch (code)
2430 {
2431 case PC:
2432 case CC0:
2433 case CONST_INT:
2434 return x == y;
2435
2436 case LABEL_REF:
2437 return XEXP (x, 0) == XEXP (y, 0);
2438
2439 case SYMBOL_REF:
2440 return XSTR (x, 0) == XSTR (y, 0);
2441
2442 case REG:
2443 if (for_gcse)
2444 return REGNO (x) == REGNO (y);
2445 else
2446 {
2447 unsigned int regno = REGNO (y);
2448 unsigned int i;
2449 unsigned int endregno
2450 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2451 : hard_regno_nregs[regno][GET_MODE (y)]);
2452
2453 /* If the quantities are not the same, the expressions are not
2454 equivalent. If there are and we are not to validate, they
2455 are equivalent. Otherwise, ensure all regs are up-to-date. */
2456
2457 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2458 return 0;
2459
2460 if (! validate)
2461 return 1;
2462
2463 for (i = regno; i < endregno; i++)
2464 if (REG_IN_TABLE (i) != REG_TICK (i))
2465 return 0;
2466
2467 return 1;
2468 }
2469
2470 case MEM:
2471 if (for_gcse)
2472 {
2473 /* Can't merge two expressions in different alias sets, since we
2474 can decide that the expression is transparent in a block when
2475 it isn't, due to it being set with the different alias set. */
2476 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
2477 return 0;
2478
2479 /* A volatile mem should not be considered equivalent to any
2480 other. */
2481 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2482 return 0;
2483 }
2484 break;
2485
2486 /* For commutative operations, check both orders. */
2487 case PLUS:
2488 case MULT:
2489 case AND:
2490 case IOR:
2491 case XOR:
2492 case NE:
2493 case EQ:
2494 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2495 validate, for_gcse)
2496 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2497 validate, for_gcse))
2498 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2499 validate, for_gcse)
2500 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2501 validate, for_gcse)));
2502
2503 case ASM_OPERANDS:
2504 /* We don't use the generic code below because we want to
2505 disregard filename and line numbers. */
2506
2507 /* A volatile asm isn't equivalent to any other. */
2508 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2509 return 0;
2510
2511 if (GET_MODE (x) != GET_MODE (y)
2512 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2513 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2514 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2515 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2516 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2517 return 0;
2518
2519 if (ASM_OPERANDS_INPUT_LENGTH (x))
2520 {
2521 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2522 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2523 ASM_OPERANDS_INPUT (y, i),
2524 validate, for_gcse)
2525 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2526 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2527 return 0;
2528 }
2529
2530 return 1;
2531
2532 default:
2533 break;
2534 }
2535
2536 /* Compare the elements. If any pair of corresponding elements
2537 fail to match, return 0 for the whole thing. */
2538
2539 fmt = GET_RTX_FORMAT (code);
2540 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2541 {
2542 switch (fmt[i])
2543 {
2544 case 'e':
2545 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2546 validate, for_gcse))
2547 return 0;
2548 break;
2549
2550 case 'E':
2551 if (XVECLEN (x, i) != XVECLEN (y, i))
2552 return 0;
2553 for (j = 0; j < XVECLEN (x, i); j++)
2554 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2555 validate, for_gcse))
2556 return 0;
2557 break;
2558
2559 case 's':
2560 if (strcmp (XSTR (x, i), XSTR (y, i)))
2561 return 0;
2562 break;
2563
2564 case 'i':
2565 if (XINT (x, i) != XINT (y, i))
2566 return 0;
2567 break;
2568
2569 case 'w':
2570 if (XWINT (x, i) != XWINT (y, i))
2571 return 0;
2572 break;
2573
2574 case '0':
2575 case 't':
2576 break;
2577
2578 default:
2579 gcc_unreachable ();
2580 }
2581 }
2582
2583 return 1;
2584 }
2585 \f
2586 /* Return 1 if X has a value that can vary even between two
2587 executions of the program. 0 means X can be compared reliably
2588 against certain constants or near-constants. */
2589
2590 static int
2591 cse_rtx_varies_p (rtx x, int from_alias)
2592 {
2593 /* We need not check for X and the equivalence class being of the same
2594 mode because if X is equivalent to a constant in some mode, it
2595 doesn't vary in any mode. */
2596
2597 if (REG_P (x)
2598 && REGNO_QTY_VALID_P (REGNO (x)))
2599 {
2600 int x_q = REG_QTY (REGNO (x));
2601 struct qty_table_elem *x_ent = &qty_table[x_q];
2602
2603 if (GET_MODE (x) == x_ent->mode
2604 && x_ent->const_rtx != NULL_RTX)
2605 return 0;
2606 }
2607
2608 if (GET_CODE (x) == PLUS
2609 && GET_CODE (XEXP (x, 1)) == CONST_INT
2610 && REG_P (XEXP (x, 0))
2611 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2612 {
2613 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2614 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2615
2616 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2617 && x0_ent->const_rtx != NULL_RTX)
2618 return 0;
2619 }
2620
2621 /* This can happen as the result of virtual register instantiation, if
2622 the initial constant is too large to be a valid address. This gives
2623 us a three instruction sequence, load large offset into a register,
2624 load fp minus a constant into a register, then a MEM which is the
2625 sum of the two `constant' registers. */
2626 if (GET_CODE (x) == PLUS
2627 && REG_P (XEXP (x, 0))
2628 && REG_P (XEXP (x, 1))
2629 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2630 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2631 {
2632 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2633 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2634 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2635 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2636
2637 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2638 && x0_ent->const_rtx != NULL_RTX
2639 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2640 && x1_ent->const_rtx != NULL_RTX)
2641 return 0;
2642 }
2643
2644 return rtx_varies_p (x, from_alias);
2645 }
2646 \f
2647 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2648 the result if necessary. INSN is as for canon_reg. */
2649
2650 static void
2651 validate_canon_reg (rtx *xloc, rtx insn)
2652 {
2653 rtx new = canon_reg (*xloc, insn);
2654 int insn_code;
2655
2656 /* If replacing pseudo with hard reg or vice versa, ensure the
2657 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2658 if (insn != 0 && new != 0
2659 && REG_P (new) && REG_P (*xloc)
2660 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2661 != (REGNO (*xloc) < FIRST_PSEUDO_REGISTER))
2662 || GET_MODE (new) != GET_MODE (*xloc)
2663 || (insn_code = recog_memoized (insn)) < 0
2664 || insn_data[insn_code].n_dups > 0))
2665 validate_change (insn, xloc, new, 1);
2666 else
2667 *xloc = new;
2668 }
2669
2670 /* Canonicalize an expression:
2671 replace each register reference inside it
2672 with the "oldest" equivalent register.
2673
2674 If INSN is nonzero and we are replacing a pseudo with a hard register
2675 or vice versa, validate_change is used to ensure that INSN remains valid
2676 after we make our substitution. The calls are made with IN_GROUP nonzero
2677 so apply_change_group must be called upon the outermost return from this
2678 function (unless INSN is zero). The result of apply_change_group can
2679 generally be discarded since the changes we are making are optional. */
2680
2681 static rtx
2682 canon_reg (rtx x, rtx insn)
2683 {
2684 int i;
2685 enum rtx_code code;
2686 const char *fmt;
2687
2688 if (x == 0)
2689 return x;
2690
2691 code = GET_CODE (x);
2692 switch (code)
2693 {
2694 case PC:
2695 case CC0:
2696 case CONST:
2697 case CONST_INT:
2698 case CONST_DOUBLE:
2699 case CONST_VECTOR:
2700 case SYMBOL_REF:
2701 case LABEL_REF:
2702 case ADDR_VEC:
2703 case ADDR_DIFF_VEC:
2704 return x;
2705
2706 case REG:
2707 {
2708 int first;
2709 int q;
2710 struct qty_table_elem *ent;
2711
2712 /* Never replace a hard reg, because hard regs can appear
2713 in more than one machine mode, and we must preserve the mode
2714 of each occurrence. Also, some hard regs appear in
2715 MEMs that are shared and mustn't be altered. Don't try to
2716 replace any reg that maps to a reg of class NO_REGS. */
2717 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2718 || ! REGNO_QTY_VALID_P (REGNO (x)))
2719 return x;
2720
2721 q = REG_QTY (REGNO (x));
2722 ent = &qty_table[q];
2723 first = ent->first_reg;
2724 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2725 : REGNO_REG_CLASS (first) == NO_REGS ? x
2726 : gen_rtx_REG (ent->mode, first));
2727 }
2728
2729 default:
2730 break;
2731 }
2732
2733 fmt = GET_RTX_FORMAT (code);
2734 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2735 {
2736 int j;
2737
2738 if (fmt[i] == 'e')
2739 validate_canon_reg (&XEXP (x, i), insn);
2740 else if (fmt[i] == 'E')
2741 for (j = 0; j < XVECLEN (x, i); j++)
2742 validate_canon_reg (&XVECEXP (x, i, j), insn);
2743 }
2744
2745 return x;
2746 }
2747 \f
2748 /* LOC is a location within INSN that is an operand address (the contents of
2749 a MEM). Find the best equivalent address to use that is valid for this
2750 insn.
2751
2752 On most CISC machines, complicated address modes are costly, and rtx_cost
2753 is a good approximation for that cost. However, most RISC machines have
2754 only a few (usually only one) memory reference formats. If an address is
2755 valid at all, it is often just as cheap as any other address. Hence, for
2756 RISC machines, we use `address_cost' to compare the costs of various
2757 addresses. For two addresses of equal cost, choose the one with the
2758 highest `rtx_cost' value as that has the potential of eliminating the
2759 most insns. For equal costs, we choose the first in the equivalence
2760 class. Note that we ignore the fact that pseudo registers are cheaper than
2761 hard registers here because we would also prefer the pseudo registers. */
2762
2763 static void
2764 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2765 {
2766 struct table_elt *elt;
2767 rtx addr = *loc;
2768 struct table_elt *p;
2769 int found_better = 1;
2770 int save_do_not_record = do_not_record;
2771 int save_hash_arg_in_memory = hash_arg_in_memory;
2772 int addr_volatile;
2773 int regno;
2774 unsigned hash;
2775
2776 /* Do not try to replace constant addresses or addresses of local and
2777 argument slots. These MEM expressions are made only once and inserted
2778 in many instructions, as well as being used to control symbol table
2779 output. It is not safe to clobber them.
2780
2781 There are some uncommon cases where the address is already in a register
2782 for some reason, but we cannot take advantage of that because we have
2783 no easy way to unshare the MEM. In addition, looking up all stack
2784 addresses is costly. */
2785 if ((GET_CODE (addr) == PLUS
2786 && REG_P (XEXP (addr, 0))
2787 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2788 && (regno = REGNO (XEXP (addr, 0)),
2789 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2790 || regno == ARG_POINTER_REGNUM))
2791 || (REG_P (addr)
2792 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2793 || regno == HARD_FRAME_POINTER_REGNUM
2794 || regno == ARG_POINTER_REGNUM))
2795 || CONSTANT_ADDRESS_P (addr))
2796 return;
2797
2798 /* If this address is not simply a register, try to fold it. This will
2799 sometimes simplify the expression. Many simplifications
2800 will not be valid, but some, usually applying the associative rule, will
2801 be valid and produce better code. */
2802 if (!REG_P (addr))
2803 {
2804 rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX);
2805 int addr_folded_cost = address_cost (folded, mode);
2806 int addr_cost = address_cost (addr, mode);
2807
2808 if ((addr_folded_cost < addr_cost
2809 || (addr_folded_cost == addr_cost
2810 /* ??? The rtx_cost comparison is left over from an older
2811 version of this code. It is probably no longer helpful. */
2812 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2813 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2814 && validate_change (insn, loc, folded, 0))
2815 addr = folded;
2816 }
2817
2818 /* If this address is not in the hash table, we can't look for equivalences
2819 of the whole address. Also, ignore if volatile. */
2820
2821 do_not_record = 0;
2822 hash = HASH (addr, Pmode);
2823 addr_volatile = do_not_record;
2824 do_not_record = save_do_not_record;
2825 hash_arg_in_memory = save_hash_arg_in_memory;
2826
2827 if (addr_volatile)
2828 return;
2829
2830 elt = lookup (addr, hash, Pmode);
2831
2832 if (elt)
2833 {
2834 /* We need to find the best (under the criteria documented above) entry
2835 in the class that is valid. We use the `flag' field to indicate
2836 choices that were invalid and iterate until we can't find a better
2837 one that hasn't already been tried. */
2838
2839 for (p = elt->first_same_value; p; p = p->next_same_value)
2840 p->flag = 0;
2841
2842 while (found_better)
2843 {
2844 int best_addr_cost = address_cost (*loc, mode);
2845 int best_rtx_cost = (elt->cost + 1) >> 1;
2846 int exp_cost;
2847 struct table_elt *best_elt = elt;
2848
2849 found_better = 0;
2850 for (p = elt->first_same_value; p; p = p->next_same_value)
2851 if (! p->flag)
2852 {
2853 if ((REG_P (p->exp)
2854 || exp_equiv_p (p->exp, p->exp, 1, false))
2855 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2856 || (exp_cost == best_addr_cost
2857 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2858 {
2859 found_better = 1;
2860 best_addr_cost = exp_cost;
2861 best_rtx_cost = (p->cost + 1) >> 1;
2862 best_elt = p;
2863 }
2864 }
2865
2866 if (found_better)
2867 {
2868 if (validate_change (insn, loc,
2869 canon_reg (copy_rtx (best_elt->exp),
2870 NULL_RTX), 0))
2871 return;
2872 else
2873 best_elt->flag = 1;
2874 }
2875 }
2876 }
2877
2878 /* If the address is a binary operation with the first operand a register
2879 and the second a constant, do the same as above, but looking for
2880 equivalences of the register. Then try to simplify before checking for
2881 the best address to use. This catches a few cases: First is when we
2882 have REG+const and the register is another REG+const. We can often merge
2883 the constants and eliminate one insn and one register. It may also be
2884 that a machine has a cheap REG+REG+const. Finally, this improves the
2885 code on the Alpha for unaligned byte stores. */
2886
2887 if (flag_expensive_optimizations
2888 && ARITHMETIC_P (*loc)
2889 && REG_P (XEXP (*loc, 0)))
2890 {
2891 rtx op1 = XEXP (*loc, 1);
2892
2893 do_not_record = 0;
2894 hash = HASH (XEXP (*loc, 0), Pmode);
2895 do_not_record = save_do_not_record;
2896 hash_arg_in_memory = save_hash_arg_in_memory;
2897
2898 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2899 if (elt == 0)
2900 return;
2901
2902 /* We need to find the best (under the criteria documented above) entry
2903 in the class that is valid. We use the `flag' field to indicate
2904 choices that were invalid and iterate until we can't find a better
2905 one that hasn't already been tried. */
2906
2907 for (p = elt->first_same_value; p; p = p->next_same_value)
2908 p->flag = 0;
2909
2910 while (found_better)
2911 {
2912 int best_addr_cost = address_cost (*loc, mode);
2913 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2914 struct table_elt *best_elt = elt;
2915 rtx best_rtx = *loc;
2916 int count;
2917
2918 /* This is at worst case an O(n^2) algorithm, so limit our search
2919 to the first 32 elements on the list. This avoids trouble
2920 compiling code with very long basic blocks that can easily
2921 call simplify_gen_binary so many times that we run out of
2922 memory. */
2923
2924 found_better = 0;
2925 for (p = elt->first_same_value, count = 0;
2926 p && count < 32;
2927 p = p->next_same_value, count++)
2928 if (! p->flag
2929 && (REG_P (p->exp)
2930 || exp_equiv_p (p->exp, p->exp, 1, false)))
2931 {
2932 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
2933 p->exp, op1);
2934 int new_cost;
2935 new_cost = address_cost (new, mode);
2936
2937 if (new_cost < best_addr_cost
2938 || (new_cost == best_addr_cost
2939 && (COST (new) + 1) >> 1 > best_rtx_cost))
2940 {
2941 found_better = 1;
2942 best_addr_cost = new_cost;
2943 best_rtx_cost = (COST (new) + 1) >> 1;
2944 best_elt = p;
2945 best_rtx = new;
2946 }
2947 }
2948
2949 if (found_better)
2950 {
2951 if (validate_change (insn, loc,
2952 canon_reg (copy_rtx (best_rtx),
2953 NULL_RTX), 0))
2954 return;
2955 else
2956 best_elt->flag = 1;
2957 }
2958 }
2959 }
2960 }
2961 \f
2962 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2963 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2964 what values are being compared.
2965
2966 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2967 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2968 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2969 compared to produce cc0.
2970
2971 The return value is the comparison operator and is either the code of
2972 A or the code corresponding to the inverse of the comparison. */
2973
2974 static enum rtx_code
2975 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2976 enum machine_mode *pmode1, enum machine_mode *pmode2)
2977 {
2978 rtx arg1, arg2;
2979
2980 arg1 = *parg1, arg2 = *parg2;
2981
2982 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2983
2984 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2985 {
2986 /* Set nonzero when we find something of interest. */
2987 rtx x = 0;
2988 int reverse_code = 0;
2989 struct table_elt *p = 0;
2990
2991 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2992 On machines with CC0, this is the only case that can occur, since
2993 fold_rtx will return the COMPARE or item being compared with zero
2994 when given CC0. */
2995
2996 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2997 x = arg1;
2998
2999 /* If ARG1 is a comparison operator and CODE is testing for
3000 STORE_FLAG_VALUE, get the inner arguments. */
3001
3002 else if (COMPARISON_P (arg1))
3003 {
3004 #ifdef FLOAT_STORE_FLAG_VALUE
3005 REAL_VALUE_TYPE fsfv;
3006 #endif
3007
3008 if (code == NE
3009 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3010 && code == LT && STORE_FLAG_VALUE == -1)
3011 #ifdef FLOAT_STORE_FLAG_VALUE
3012 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3013 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3014 REAL_VALUE_NEGATIVE (fsfv)))
3015 #endif
3016 )
3017 x = arg1;
3018 else if (code == EQ
3019 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3020 && code == GE && STORE_FLAG_VALUE == -1)
3021 #ifdef FLOAT_STORE_FLAG_VALUE
3022 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3023 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3024 REAL_VALUE_NEGATIVE (fsfv)))
3025 #endif
3026 )
3027 x = arg1, reverse_code = 1;
3028 }
3029
3030 /* ??? We could also check for
3031
3032 (ne (and (eq (...) (const_int 1))) (const_int 0))
3033
3034 and related forms, but let's wait until we see them occurring. */
3035
3036 if (x == 0)
3037 /* Look up ARG1 in the hash table and see if it has an equivalence
3038 that lets us see what is being compared. */
3039 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3040 if (p)
3041 {
3042 p = p->first_same_value;
3043
3044 /* If what we compare is already known to be constant, that is as
3045 good as it gets.
3046 We need to break the loop in this case, because otherwise we
3047 can have an infinite loop when looking at a reg that is known
3048 to be a constant which is the same as a comparison of a reg
3049 against zero which appears later in the insn stream, which in
3050 turn is constant and the same as the comparison of the first reg
3051 against zero... */
3052 if (p->is_const)
3053 break;
3054 }
3055
3056 for (; p; p = p->next_same_value)
3057 {
3058 enum machine_mode inner_mode = GET_MODE (p->exp);
3059 #ifdef FLOAT_STORE_FLAG_VALUE
3060 REAL_VALUE_TYPE fsfv;
3061 #endif
3062
3063 /* If the entry isn't valid, skip it. */
3064 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3065 continue;
3066
3067 if (GET_CODE (p->exp) == COMPARE
3068 /* Another possibility is that this machine has a compare insn
3069 that includes the comparison code. In that case, ARG1 would
3070 be equivalent to a comparison operation that would set ARG1 to
3071 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3072 ORIG_CODE is the actual comparison being done; if it is an EQ,
3073 we must reverse ORIG_CODE. On machine with a negative value
3074 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3075 || ((code == NE
3076 || (code == LT
3077 && GET_MODE_CLASS (inner_mode) == MODE_INT
3078 && (GET_MODE_BITSIZE (inner_mode)
3079 <= HOST_BITS_PER_WIDE_INT)
3080 && (STORE_FLAG_VALUE
3081 & ((HOST_WIDE_INT) 1
3082 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3083 #ifdef FLOAT_STORE_FLAG_VALUE
3084 || (code == LT
3085 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3086 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3087 REAL_VALUE_NEGATIVE (fsfv)))
3088 #endif
3089 )
3090 && COMPARISON_P (p->exp)))
3091 {
3092 x = p->exp;
3093 break;
3094 }
3095 else if ((code == EQ
3096 || (code == GE
3097 && GET_MODE_CLASS (inner_mode) == MODE_INT
3098 && (GET_MODE_BITSIZE (inner_mode)
3099 <= HOST_BITS_PER_WIDE_INT)
3100 && (STORE_FLAG_VALUE
3101 & ((HOST_WIDE_INT) 1
3102 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3103 #ifdef FLOAT_STORE_FLAG_VALUE
3104 || (code == GE
3105 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3106 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3107 REAL_VALUE_NEGATIVE (fsfv)))
3108 #endif
3109 )
3110 && COMPARISON_P (p->exp))
3111 {
3112 reverse_code = 1;
3113 x = p->exp;
3114 break;
3115 }
3116
3117 /* If this non-trapping address, e.g. fp + constant, the
3118 equivalent is a better operand since it may let us predict
3119 the value of the comparison. */
3120 else if (!rtx_addr_can_trap_p (p->exp))
3121 {
3122 arg1 = p->exp;
3123 continue;
3124 }
3125 }
3126
3127 /* If we didn't find a useful equivalence for ARG1, we are done.
3128 Otherwise, set up for the next iteration. */
3129 if (x == 0)
3130 break;
3131
3132 /* If we need to reverse the comparison, make sure that that is
3133 possible -- we can't necessarily infer the value of GE from LT
3134 with floating-point operands. */
3135 if (reverse_code)
3136 {
3137 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3138 if (reversed == UNKNOWN)
3139 break;
3140 else
3141 code = reversed;
3142 }
3143 else if (COMPARISON_P (x))
3144 code = GET_CODE (x);
3145 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3146 }
3147
3148 /* Return our results. Return the modes from before fold_rtx
3149 because fold_rtx might produce const_int, and then it's too late. */
3150 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3151 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3152
3153 return code;
3154 }
3155 \f
3156 /* If X is a nontrivial arithmetic operation on an argument
3157 for which a constant value can be determined, return
3158 the result of operating on that value, as a constant.
3159 Otherwise, return X, possibly with one or more operands
3160 modified by recursive calls to this function.
3161
3162 If X is a register whose contents are known, we do NOT
3163 return those contents here. equiv_constant is called to
3164 perform that task.
3165
3166 INSN is the insn that we may be modifying. If it is 0, make a copy
3167 of X before modifying it. */
3168
3169 static rtx
3170 fold_rtx (rtx x, rtx insn)
3171 {
3172 enum rtx_code code;
3173 enum machine_mode mode;
3174 const char *fmt;
3175 int i;
3176 rtx new = 0;
3177 int copied = 0;
3178 int must_swap = 0;
3179
3180 /* Folded equivalents of first two operands of X. */
3181 rtx folded_arg0;
3182 rtx folded_arg1;
3183
3184 /* Constant equivalents of first three operands of X;
3185 0 when no such equivalent is known. */
3186 rtx const_arg0;
3187 rtx const_arg1;
3188 rtx const_arg2;
3189
3190 /* The mode of the first operand of X. We need this for sign and zero
3191 extends. */
3192 enum machine_mode mode_arg0;
3193
3194 if (x == 0)
3195 return x;
3196
3197 mode = GET_MODE (x);
3198 code = GET_CODE (x);
3199 switch (code)
3200 {
3201 case CONST:
3202 case CONST_INT:
3203 case CONST_DOUBLE:
3204 case CONST_VECTOR:
3205 case SYMBOL_REF:
3206 case LABEL_REF:
3207 case REG:
3208 /* No use simplifying an EXPR_LIST
3209 since they are used only for lists of args
3210 in a function call's REG_EQUAL note. */
3211 case EXPR_LIST:
3212 return x;
3213
3214 #ifdef HAVE_cc0
3215 case CC0:
3216 return prev_insn_cc0;
3217 #endif
3218
3219 case PC:
3220 /* If the next insn is a CODE_LABEL followed by a jump table,
3221 PC's value is a LABEL_REF pointing to that label. That
3222 lets us fold switch statements on the VAX. */
3223 {
3224 rtx next;
3225 if (insn && tablejump_p (insn, &next, NULL))
3226 return gen_rtx_LABEL_REF (Pmode, next);
3227 }
3228 break;
3229
3230 case SUBREG:
3231 /* See if we previously assigned a constant value to this SUBREG. */
3232 if ((new = lookup_as_function (x, CONST_INT)) != 0
3233 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3234 return new;
3235
3236 /* If this is a paradoxical SUBREG, we have no idea what value the
3237 extra bits would have. However, if the operand is equivalent
3238 to a SUBREG whose operand is the same as our mode, and all the
3239 modes are within a word, we can just use the inner operand
3240 because these SUBREGs just say how to treat the register.
3241
3242 Similarly if we find an integer constant. */
3243
3244 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3245 {
3246 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3247 struct table_elt *elt;
3248
3249 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3250 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3251 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3252 imode)) != 0)
3253 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3254 {
3255 if (CONSTANT_P (elt->exp)
3256 && GET_MODE (elt->exp) == VOIDmode)
3257 return elt->exp;
3258
3259 if (GET_CODE (elt->exp) == SUBREG
3260 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3261 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3262 return copy_rtx (SUBREG_REG (elt->exp));
3263 }
3264
3265 return x;
3266 }
3267
3268 /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG.
3269 We might be able to if the SUBREG is extracting a single word in an
3270 integral mode or extracting the low part. */
3271
3272 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3273 const_arg0 = equiv_constant (folded_arg0);
3274 if (const_arg0)
3275 folded_arg0 = const_arg0;
3276
3277 if (folded_arg0 != SUBREG_REG (x))
3278 {
3279 new = simplify_subreg (mode, folded_arg0,
3280 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3281 if (new)
3282 return new;
3283 }
3284
3285 if (REG_P (folded_arg0)
3286 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3287 {
3288 struct table_elt *elt;
3289
3290 elt = lookup (folded_arg0,
3291 HASH (folded_arg0, GET_MODE (folded_arg0)),
3292 GET_MODE (folded_arg0));
3293
3294 if (elt)
3295 elt = elt->first_same_value;
3296
3297 if (subreg_lowpart_p (x))
3298 /* If this is a narrowing SUBREG and our operand is a REG, see
3299 if we can find an equivalence for REG that is an arithmetic
3300 operation in a wider mode where both operands are paradoxical
3301 SUBREGs from objects of our result mode. In that case, we
3302 couldn-t report an equivalent value for that operation, since we
3303 don't know what the extra bits will be. But we can find an
3304 equivalence for this SUBREG by folding that operation in the
3305 narrow mode. This allows us to fold arithmetic in narrow modes
3306 when the machine only supports word-sized arithmetic.
3307
3308 Also look for a case where we have a SUBREG whose operand
3309 is the same as our result. If both modes are smaller
3310 than a word, we are simply interpreting a register in
3311 different modes and we can use the inner value. */
3312
3313 for (; elt; elt = elt->next_same_value)
3314 {
3315 enum rtx_code eltcode = GET_CODE (elt->exp);
3316
3317 /* Just check for unary and binary operations. */
3318 if (UNARY_P (elt->exp)
3319 && eltcode != SIGN_EXTEND
3320 && eltcode != ZERO_EXTEND
3321 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3322 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3323 && (GET_MODE_CLASS (mode)
3324 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3325 {
3326 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3327
3328 if (!REG_P (op0) && ! CONSTANT_P (op0))
3329 op0 = fold_rtx (op0, NULL_RTX);
3330
3331 op0 = equiv_constant (op0);
3332 if (op0)
3333 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3334 op0, mode);
3335 }
3336 else if (ARITHMETIC_P (elt->exp)
3337 && eltcode != DIV && eltcode != MOD
3338 && eltcode != UDIV && eltcode != UMOD
3339 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3340 && eltcode != ROTATE && eltcode != ROTATERT
3341 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3342 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3343 == mode))
3344 || CONSTANT_P (XEXP (elt->exp, 0)))
3345 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3346 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3347 == mode))
3348 || CONSTANT_P (XEXP (elt->exp, 1))))
3349 {
3350 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3351 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3352
3353 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3354 op0 = fold_rtx (op0, NULL_RTX);
3355
3356 if (op0)
3357 op0 = equiv_constant (op0);
3358
3359 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3360 op1 = fold_rtx (op1, NULL_RTX);
3361
3362 if (op1)
3363 op1 = equiv_constant (op1);
3364
3365 /* If we are looking for the low SImode part of
3366 (ashift:DI c (const_int 32)), it doesn't work
3367 to compute that in SImode, because a 32-bit shift
3368 in SImode is unpredictable. We know the value is 0. */
3369 if (op0 && op1
3370 && GET_CODE (elt->exp) == ASHIFT
3371 && GET_CODE (op1) == CONST_INT
3372 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3373 {
3374 if (INTVAL (op1)
3375 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3376 /* If the count fits in the inner mode's width,
3377 but exceeds the outer mode's width,
3378 the value will get truncated to 0
3379 by the subreg. */
3380 new = CONST0_RTX (mode);
3381 else
3382 /* If the count exceeds even the inner mode's width,
3383 don't fold this expression. */
3384 new = 0;
3385 }
3386 else if (op0 && op1)
3387 new = simplify_binary_operation (GET_CODE (elt->exp), mode, op0, op1);
3388 }
3389
3390 else if (GET_CODE (elt->exp) == SUBREG
3391 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3392 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3393 <= UNITS_PER_WORD)
3394 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3395 new = copy_rtx (SUBREG_REG (elt->exp));
3396
3397 if (new)
3398 return new;
3399 }
3400 else
3401 /* A SUBREG resulting from a zero extension may fold to zero if
3402 it extracts higher bits than the ZERO_EXTEND's source bits.
3403 FIXME: if combine tried to, er, combine these instructions,
3404 this transformation may be moved to simplify_subreg. */
3405 for (; elt; elt = elt->next_same_value)
3406 {
3407 if (GET_CODE (elt->exp) == ZERO_EXTEND
3408 && subreg_lsb (x)
3409 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3410 return CONST0_RTX (mode);
3411 }
3412 }
3413
3414 return x;
3415
3416 case NOT:
3417 case NEG:
3418 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3419 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3420 new = lookup_as_function (XEXP (x, 0), code);
3421 if (new)
3422 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3423 break;
3424
3425 case MEM:
3426 /* If we are not actually processing an insn, don't try to find the
3427 best address. Not only don't we care, but we could modify the
3428 MEM in an invalid way since we have no insn to validate against. */
3429 if (insn != 0)
3430 find_best_addr (insn, &XEXP (x, 0), GET_MODE (x));
3431
3432 {
3433 /* Even if we don't fold in the insn itself,
3434 we can safely do so here, in hopes of getting a constant. */
3435 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3436 rtx base = 0;
3437 HOST_WIDE_INT offset = 0;
3438
3439 if (REG_P (addr)
3440 && REGNO_QTY_VALID_P (REGNO (addr)))
3441 {
3442 int addr_q = REG_QTY (REGNO (addr));
3443 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3444
3445 if (GET_MODE (addr) == addr_ent->mode
3446 && addr_ent->const_rtx != NULL_RTX)
3447 addr = addr_ent->const_rtx;
3448 }
3449
3450 /* If address is constant, split it into a base and integer offset. */
3451 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3452 base = addr;
3453 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3454 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3455 {
3456 base = XEXP (XEXP (addr, 0), 0);
3457 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3458 }
3459 else if (GET_CODE (addr) == LO_SUM
3460 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3461 base = XEXP (addr, 1);
3462
3463 /* If this is a constant pool reference, we can fold it into its
3464 constant to allow better value tracking. */
3465 if (base && GET_CODE (base) == SYMBOL_REF
3466 && CONSTANT_POOL_ADDRESS_P (base))
3467 {
3468 rtx constant = get_pool_constant (base);
3469 enum machine_mode const_mode = get_pool_mode (base);
3470 rtx new;
3471
3472 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3473 {
3474 constant_pool_entries_cost = COST (constant);
3475 constant_pool_entries_regcost = approx_reg_cost (constant);
3476 }
3477
3478 /* If we are loading the full constant, we have an equivalence. */
3479 if (offset == 0 && mode == const_mode)
3480 return constant;
3481
3482 /* If this actually isn't a constant (weird!), we can't do
3483 anything. Otherwise, handle the two most common cases:
3484 extracting a word from a multi-word constant, and extracting
3485 the low-order bits. Other cases don't seem common enough to
3486 worry about. */
3487 if (! CONSTANT_P (constant))
3488 return x;
3489
3490 if (GET_MODE_CLASS (mode) == MODE_INT
3491 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3492 && offset % UNITS_PER_WORD == 0
3493 && (new = operand_subword (constant,
3494 offset / UNITS_PER_WORD,
3495 0, const_mode)) != 0)
3496 return new;
3497
3498 if (((BYTES_BIG_ENDIAN
3499 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3500 || (! BYTES_BIG_ENDIAN && offset == 0))
3501 && (new = gen_lowpart (mode, constant)) != 0)
3502 return new;
3503 }
3504
3505 /* If this is a reference to a label at a known position in a jump
3506 table, we also know its value. */
3507 if (base && GET_CODE (base) == LABEL_REF)
3508 {
3509 rtx label = XEXP (base, 0);
3510 rtx table_insn = NEXT_INSN (label);
3511
3512 if (table_insn && JUMP_P (table_insn)
3513 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3514 {
3515 rtx table = PATTERN (table_insn);
3516
3517 if (offset >= 0
3518 && (offset / GET_MODE_SIZE (GET_MODE (table))
3519 < XVECLEN (table, 0)))
3520 return XVECEXP (table, 0,
3521 offset / GET_MODE_SIZE (GET_MODE (table)));
3522 }
3523 if (table_insn && JUMP_P (table_insn)
3524 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3525 {
3526 rtx table = PATTERN (table_insn);
3527
3528 if (offset >= 0
3529 && (offset / GET_MODE_SIZE (GET_MODE (table))
3530 < XVECLEN (table, 1)))
3531 {
3532 offset /= GET_MODE_SIZE (GET_MODE (table));
3533 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3534 XEXP (table, 0));
3535
3536 if (GET_MODE (table) != Pmode)
3537 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3538
3539 /* Indicate this is a constant. This isn't a
3540 valid form of CONST, but it will only be used
3541 to fold the next insns and then discarded, so
3542 it should be safe.
3543
3544 Note this expression must be explicitly discarded,
3545 by cse_insn, else it may end up in a REG_EQUAL note
3546 and "escape" to cause problems elsewhere. */
3547 return gen_rtx_CONST (GET_MODE (new), new);
3548 }
3549 }
3550 }
3551
3552 return x;
3553 }
3554
3555 #ifdef NO_FUNCTION_CSE
3556 case CALL:
3557 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3558 return x;
3559 break;
3560 #endif
3561
3562 case ASM_OPERANDS:
3563 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3564 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3565 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3566 break;
3567
3568 default:
3569 break;
3570 }
3571
3572 const_arg0 = 0;
3573 const_arg1 = 0;
3574 const_arg2 = 0;
3575 mode_arg0 = VOIDmode;
3576
3577 /* Try folding our operands.
3578 Then see which ones have constant values known. */
3579
3580 fmt = GET_RTX_FORMAT (code);
3581 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3582 if (fmt[i] == 'e')
3583 {
3584 rtx arg = XEXP (x, i);
3585 rtx folded_arg = arg, const_arg = 0;
3586 enum machine_mode mode_arg = GET_MODE (arg);
3587 rtx cheap_arg, expensive_arg;
3588 rtx replacements[2];
3589 int j;
3590 int old_cost = COST_IN (XEXP (x, i), code);
3591
3592 /* Most arguments are cheap, so handle them specially. */
3593 switch (GET_CODE (arg))
3594 {
3595 case REG:
3596 /* This is the same as calling equiv_constant; it is duplicated
3597 here for speed. */
3598 if (REGNO_QTY_VALID_P (REGNO (arg)))
3599 {
3600 int arg_q = REG_QTY (REGNO (arg));
3601 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3602
3603 if (arg_ent->const_rtx != NULL_RTX
3604 && !REG_P (arg_ent->const_rtx)
3605 && GET_CODE (arg_ent->const_rtx) != PLUS)
3606 const_arg
3607 = gen_lowpart (GET_MODE (arg),
3608 arg_ent->const_rtx);
3609 }
3610 break;
3611
3612 case CONST:
3613 case CONST_INT:
3614 case SYMBOL_REF:
3615 case LABEL_REF:
3616 case CONST_DOUBLE:
3617 case CONST_VECTOR:
3618 const_arg = arg;
3619 break;
3620
3621 #ifdef HAVE_cc0
3622 case CC0:
3623 folded_arg = prev_insn_cc0;
3624 mode_arg = prev_insn_cc0_mode;
3625 const_arg = equiv_constant (folded_arg);
3626 break;
3627 #endif
3628
3629 default:
3630 folded_arg = fold_rtx (arg, insn);
3631 const_arg = equiv_constant (folded_arg);
3632 }
3633
3634 /* For the first three operands, see if the operand
3635 is constant or equivalent to a constant. */
3636 switch (i)
3637 {
3638 case 0:
3639 folded_arg0 = folded_arg;
3640 const_arg0 = const_arg;
3641 mode_arg0 = mode_arg;
3642 break;
3643 case 1:
3644 folded_arg1 = folded_arg;
3645 const_arg1 = const_arg;
3646 break;
3647 case 2:
3648 const_arg2 = const_arg;
3649 break;
3650 }
3651
3652 /* Pick the least expensive of the folded argument and an
3653 equivalent constant argument. */
3654 if (const_arg == 0 || const_arg == folded_arg
3655 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3656 cheap_arg = folded_arg, expensive_arg = const_arg;
3657 else
3658 cheap_arg = const_arg, expensive_arg = folded_arg;
3659
3660 /* Try to replace the operand with the cheapest of the two
3661 possibilities. If it doesn't work and this is either of the first
3662 two operands of a commutative operation, try swapping them.
3663 If THAT fails, try the more expensive, provided it is cheaper
3664 than what is already there. */
3665
3666 if (cheap_arg == XEXP (x, i))
3667 continue;
3668
3669 if (insn == 0 && ! copied)
3670 {
3671 x = copy_rtx (x);
3672 copied = 1;
3673 }
3674
3675 /* Order the replacements from cheapest to most expensive. */
3676 replacements[0] = cheap_arg;
3677 replacements[1] = expensive_arg;
3678
3679 for (j = 0; j < 2 && replacements[j]; j++)
3680 {
3681 int new_cost = COST_IN (replacements[j], code);
3682
3683 /* Stop if what existed before was cheaper. Prefer constants
3684 in the case of a tie. */
3685 if (new_cost > old_cost
3686 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3687 break;
3688
3689 /* It's not safe to substitute the operand of a conversion
3690 operator with a constant, as the conversion's identity
3691 depends upon the mode of it's operand. This optimization
3692 is handled by the call to simplify_unary_operation. */
3693 if (GET_RTX_CLASS (code) == RTX_UNARY
3694 && GET_MODE (replacements[j]) != mode_arg0
3695 && (code == ZERO_EXTEND
3696 || code == SIGN_EXTEND
3697 || code == TRUNCATE
3698 || code == FLOAT_TRUNCATE
3699 || code == FLOAT_EXTEND
3700 || code == FLOAT
3701 || code == FIX
3702 || code == UNSIGNED_FLOAT
3703 || code == UNSIGNED_FIX))
3704 continue;
3705
3706 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3707 break;
3708
3709 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3710 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3711 {
3712 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3713 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3714
3715 if (apply_change_group ())
3716 {
3717 /* Swap them back to be invalid so that this loop can
3718 continue and flag them to be swapped back later. */
3719 rtx tem;
3720
3721 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3722 XEXP (x, 1) = tem;
3723 must_swap = 1;
3724 break;
3725 }
3726 }
3727 }
3728 }
3729
3730 else
3731 {
3732 if (fmt[i] == 'E')
3733 /* Don't try to fold inside of a vector of expressions.
3734 Doing nothing is harmless. */
3735 {;}
3736 }
3737
3738 /* If a commutative operation, place a constant integer as the second
3739 operand unless the first operand is also a constant integer. Otherwise,
3740 place any constant second unless the first operand is also a constant. */
3741
3742 if (COMMUTATIVE_P (x))
3743 {
3744 if (must_swap
3745 || swap_commutative_operands_p (const_arg0 ? const_arg0
3746 : XEXP (x, 0),
3747 const_arg1 ? const_arg1
3748 : XEXP (x, 1)))
3749 {
3750 rtx tem = XEXP (x, 0);
3751
3752 if (insn == 0 && ! copied)
3753 {
3754 x = copy_rtx (x);
3755 copied = 1;
3756 }
3757
3758 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3759 validate_change (insn, &XEXP (x, 1), tem, 1);
3760 if (apply_change_group ())
3761 {
3762 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3763 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3764 }
3765 }
3766 }
3767
3768 /* If X is an arithmetic operation, see if we can simplify it. */
3769
3770 switch (GET_RTX_CLASS (code))
3771 {
3772 case RTX_UNARY:
3773 {
3774 int is_const = 0;
3775
3776 /* We can't simplify extension ops unless we know the
3777 original mode. */
3778 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3779 && mode_arg0 == VOIDmode)
3780 break;
3781
3782 /* If we had a CONST, strip it off and put it back later if we
3783 fold. */
3784 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3785 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3786
3787 new = simplify_unary_operation (code, mode,
3788 const_arg0 ? const_arg0 : folded_arg0,
3789 mode_arg0);
3790 /* NEG of PLUS could be converted into MINUS, but that causes
3791 expressions of the form
3792 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3793 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3794 FIXME: those ports should be fixed. */
3795 if (new != 0 && is_const
3796 && GET_CODE (new) == PLUS
3797 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3798 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3799 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3800 new = gen_rtx_CONST (mode, new);
3801 }
3802 break;
3803
3804 case RTX_COMPARE:
3805 case RTX_COMM_COMPARE:
3806 /* See what items are actually being compared and set FOLDED_ARG[01]
3807 to those values and CODE to the actual comparison code. If any are
3808 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3809 do anything if both operands are already known to be constant. */
3810
3811 if (const_arg0 == 0 || const_arg1 == 0)
3812 {
3813 struct table_elt *p0, *p1;
3814 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3815 enum machine_mode mode_arg1;
3816
3817 #ifdef FLOAT_STORE_FLAG_VALUE
3818 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3819 {
3820 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3821 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3822 false_rtx = CONST0_RTX (mode);
3823 }
3824 #endif
3825
3826 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3827 &mode_arg0, &mode_arg1);
3828 const_arg0 = equiv_constant (folded_arg0);
3829 const_arg1 = equiv_constant (folded_arg1);
3830
3831 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3832 what kinds of things are being compared, so we can't do
3833 anything with this comparison. */
3834
3835 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3836 break;
3837
3838 /* If we do not now have two constants being compared, see
3839 if we can nevertheless deduce some things about the
3840 comparison. */
3841 if (const_arg0 == 0 || const_arg1 == 0)
3842 {
3843 /* Some addresses are known to be nonzero. We don't know
3844 their sign, but equality comparisons are known. */
3845 if (const_arg1 == const0_rtx
3846 && nonzero_address_p (folded_arg0))
3847 {
3848 if (code == EQ)
3849 return false_rtx;
3850 else if (code == NE)
3851 return true_rtx;
3852 }
3853
3854 /* See if the two operands are the same. */
3855
3856 if (folded_arg0 == folded_arg1
3857 || (REG_P (folded_arg0)
3858 && REG_P (folded_arg1)
3859 && (REG_QTY (REGNO (folded_arg0))
3860 == REG_QTY (REGNO (folded_arg1))))
3861 || ((p0 = lookup (folded_arg0,
3862 SAFE_HASH (folded_arg0, mode_arg0),
3863 mode_arg0))
3864 && (p1 = lookup (folded_arg1,
3865 SAFE_HASH (folded_arg1, mode_arg0),
3866 mode_arg0))
3867 && p0->first_same_value == p1->first_same_value))
3868 {
3869 /* Sadly two equal NaNs are not equivalent. */
3870 if (!HONOR_NANS (mode_arg0))
3871 return ((code == EQ || code == LE || code == GE
3872 || code == LEU || code == GEU || code == UNEQ
3873 || code == UNLE || code == UNGE
3874 || code == ORDERED)
3875 ? true_rtx : false_rtx);
3876 /* Take care for the FP compares we can resolve. */
3877 if (code == UNEQ || code == UNLE || code == UNGE)
3878 return true_rtx;
3879 if (code == LTGT || code == LT || code == GT)
3880 return false_rtx;
3881 }
3882
3883 /* If FOLDED_ARG0 is a register, see if the comparison we are
3884 doing now is either the same as we did before or the reverse
3885 (we only check the reverse if not floating-point). */
3886 else if (REG_P (folded_arg0))
3887 {
3888 int qty = REG_QTY (REGNO (folded_arg0));
3889
3890 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3891 {
3892 struct qty_table_elem *ent = &qty_table[qty];
3893
3894 if ((comparison_dominates_p (ent->comparison_code, code)
3895 || (! FLOAT_MODE_P (mode_arg0)
3896 && comparison_dominates_p (ent->comparison_code,
3897 reverse_condition (code))))
3898 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3899 || (const_arg1
3900 && rtx_equal_p (ent->comparison_const,
3901 const_arg1))
3902 || (REG_P (folded_arg1)
3903 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3904 return (comparison_dominates_p (ent->comparison_code, code)
3905 ? true_rtx : false_rtx);
3906 }
3907 }
3908 }
3909 }
3910
3911 /* If we are comparing against zero, see if the first operand is
3912 equivalent to an IOR with a constant. If so, we may be able to
3913 determine the result of this comparison. */
3914
3915 if (const_arg1 == const0_rtx)
3916 {
3917 rtx y = lookup_as_function (folded_arg0, IOR);
3918 rtx inner_const;
3919
3920 if (y != 0
3921 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3922 && GET_CODE (inner_const) == CONST_INT
3923 && INTVAL (inner_const) != 0)
3924 {
3925 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
3926 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
3927 && (INTVAL (inner_const)
3928 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
3929 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3930
3931 #ifdef FLOAT_STORE_FLAG_VALUE
3932 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3933 {
3934 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3935 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3936 false_rtx = CONST0_RTX (mode);
3937 }
3938 #endif
3939
3940 switch (code)
3941 {
3942 case EQ:
3943 return false_rtx;
3944 case NE:
3945 return true_rtx;
3946 case LT: case LE:
3947 if (has_sign)
3948 return true_rtx;
3949 break;
3950 case GT: case GE:
3951 if (has_sign)
3952 return false_rtx;
3953 break;
3954 default:
3955 break;
3956 }
3957 }
3958 }
3959
3960 {
3961 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3962 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3963 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3964 }
3965 break;
3966
3967 case RTX_BIN_ARITH:
3968 case RTX_COMM_ARITH:
3969 switch (code)
3970 {
3971 case PLUS:
3972 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3973 with that LABEL_REF as its second operand. If so, the result is
3974 the first operand of that MINUS. This handles switches with an
3975 ADDR_DIFF_VEC table. */
3976 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3977 {
3978 rtx y
3979 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3980 : lookup_as_function (folded_arg0, MINUS);
3981
3982 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3983 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3984 return XEXP (y, 0);
3985
3986 /* Now try for a CONST of a MINUS like the above. */
3987 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3988 : lookup_as_function (folded_arg0, CONST))) != 0
3989 && GET_CODE (XEXP (y, 0)) == MINUS
3990 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3991 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3992 return XEXP (XEXP (y, 0), 0);
3993 }
3994
3995 /* Likewise if the operands are in the other order. */
3996 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3997 {
3998 rtx y
3999 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4000 : lookup_as_function (folded_arg1, MINUS);
4001
4002 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4003 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4004 return XEXP (y, 0);
4005
4006 /* Now try for a CONST of a MINUS like the above. */
4007 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4008 : lookup_as_function (folded_arg1, CONST))) != 0
4009 && GET_CODE (XEXP (y, 0)) == MINUS
4010 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4011 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4012 return XEXP (XEXP (y, 0), 0);
4013 }
4014
4015 /* If second operand is a register equivalent to a negative
4016 CONST_INT, see if we can find a register equivalent to the
4017 positive constant. Make a MINUS if so. Don't do this for
4018 a non-negative constant since we might then alternate between
4019 choosing positive and negative constants. Having the positive
4020 constant previously-used is the more common case. Be sure
4021 the resulting constant is non-negative; if const_arg1 were
4022 the smallest negative number this would overflow: depending
4023 on the mode, this would either just be the same value (and
4024 hence not save anything) or be incorrect. */
4025 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4026 && INTVAL (const_arg1) < 0
4027 /* This used to test
4028
4029 -INTVAL (const_arg1) >= 0
4030
4031 But The Sun V5.0 compilers mis-compiled that test. So
4032 instead we test for the problematic value in a more direct
4033 manner and hope the Sun compilers get it correct. */
4034 && INTVAL (const_arg1) !=
4035 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4036 && REG_P (folded_arg1))
4037 {
4038 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4039 struct table_elt *p
4040 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4041
4042 if (p)
4043 for (p = p->first_same_value; p; p = p->next_same_value)
4044 if (REG_P (p->exp))
4045 return simplify_gen_binary (MINUS, mode, folded_arg0,
4046 canon_reg (p->exp, NULL_RTX));
4047 }
4048 goto from_plus;
4049
4050 case MINUS:
4051 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4052 If so, produce (PLUS Z C2-C). */
4053 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4054 {
4055 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4056 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4057 return fold_rtx (plus_constant (copy_rtx (y),
4058 -INTVAL (const_arg1)),
4059 NULL_RTX);
4060 }
4061
4062 /* Fall through. */
4063
4064 from_plus:
4065 case SMIN: case SMAX: case UMIN: case UMAX:
4066 case IOR: case AND: case XOR:
4067 case MULT:
4068 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4069 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4070 is known to be of similar form, we may be able to replace the
4071 operation with a combined operation. This may eliminate the
4072 intermediate operation if every use is simplified in this way.
4073 Note that the similar optimization done by combine.c only works
4074 if the intermediate operation's result has only one reference. */
4075
4076 if (REG_P (folded_arg0)
4077 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4078 {
4079 int is_shift
4080 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4081 rtx y = lookup_as_function (folded_arg0, code);
4082 rtx inner_const;
4083 enum rtx_code associate_code;
4084 rtx new_const;
4085
4086 if (y == 0
4087 || 0 == (inner_const
4088 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4089 || GET_CODE (inner_const) != CONST_INT
4090 /* If we have compiled a statement like
4091 "if (x == (x & mask1))", and now are looking at
4092 "x & mask2", we will have a case where the first operand
4093 of Y is the same as our first operand. Unless we detect
4094 this case, an infinite loop will result. */
4095 || XEXP (y, 0) == folded_arg0)
4096 break;
4097
4098 /* Don't associate these operations if they are a PLUS with the
4099 same constant and it is a power of two. These might be doable
4100 with a pre- or post-increment. Similarly for two subtracts of
4101 identical powers of two with post decrement. */
4102
4103 if (code == PLUS && const_arg1 == inner_const
4104 && ((HAVE_PRE_INCREMENT
4105 && exact_log2 (INTVAL (const_arg1)) >= 0)
4106 || (HAVE_POST_INCREMENT
4107 && exact_log2 (INTVAL (const_arg1)) >= 0)
4108 || (HAVE_PRE_DECREMENT
4109 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4110 || (HAVE_POST_DECREMENT
4111 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4112 break;
4113
4114 /* Compute the code used to compose the constants. For example,
4115 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4116
4117 associate_code = (is_shift || code == MINUS ? PLUS : code);
4118
4119 new_const = simplify_binary_operation (associate_code, mode,
4120 const_arg1, inner_const);
4121
4122 if (new_const == 0)
4123 break;
4124
4125 /* If we are associating shift operations, don't let this
4126 produce a shift of the size of the object or larger.
4127 This could occur when we follow a sign-extend by a right
4128 shift on a machine that does a sign-extend as a pair
4129 of shifts. */
4130
4131 if (is_shift && GET_CODE (new_const) == CONST_INT
4132 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4133 {
4134 /* As an exception, we can turn an ASHIFTRT of this
4135 form into a shift of the number of bits - 1. */
4136 if (code == ASHIFTRT)
4137 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4138 else
4139 break;
4140 }
4141
4142 y = copy_rtx (XEXP (y, 0));
4143
4144 /* If Y contains our first operand (the most common way this
4145 can happen is if Y is a MEM), we would do into an infinite
4146 loop if we tried to fold it. So don't in that case. */
4147
4148 if (! reg_mentioned_p (folded_arg0, y))
4149 y = fold_rtx (y, insn);
4150
4151 return simplify_gen_binary (code, mode, y, new_const);
4152 }
4153 break;
4154
4155 case DIV: case UDIV:
4156 /* ??? The associative optimization performed immediately above is
4157 also possible for DIV and UDIV using associate_code of MULT.
4158 However, we would need extra code to verify that the
4159 multiplication does not overflow, that is, there is no overflow
4160 in the calculation of new_const. */
4161 break;
4162
4163 default:
4164 break;
4165 }
4166
4167 new = simplify_binary_operation (code, mode,
4168 const_arg0 ? const_arg0 : folded_arg0,
4169 const_arg1 ? const_arg1 : folded_arg1);
4170 break;
4171
4172 case RTX_OBJ:
4173 /* (lo_sum (high X) X) is simply X. */
4174 if (code == LO_SUM && const_arg0 != 0
4175 && GET_CODE (const_arg0) == HIGH
4176 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4177 return const_arg1;
4178 break;
4179
4180 case RTX_TERNARY:
4181 case RTX_BITFIELD_OPS:
4182 new = simplify_ternary_operation (code, mode, mode_arg0,
4183 const_arg0 ? const_arg0 : folded_arg0,
4184 const_arg1 ? const_arg1 : folded_arg1,
4185 const_arg2 ? const_arg2 : XEXP (x, 2));
4186 break;
4187
4188 default:
4189 break;
4190 }
4191
4192 return new ? new : x;
4193 }
4194 \f
4195 /* Return a constant value currently equivalent to X.
4196 Return 0 if we don't know one. */
4197
4198 static rtx
4199 equiv_constant (rtx x)
4200 {
4201 if (REG_P (x)
4202 && REGNO_QTY_VALID_P (REGNO (x)))
4203 {
4204 int x_q = REG_QTY (REGNO (x));
4205 struct qty_table_elem *x_ent = &qty_table[x_q];
4206
4207 if (x_ent->const_rtx)
4208 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4209 }
4210
4211 if (x == 0 || CONSTANT_P (x))
4212 return x;
4213
4214 /* If X is a MEM, try to fold it outside the context of any insn to see if
4215 it might be equivalent to a constant. That handles the case where it
4216 is a constant-pool reference. Then try to look it up in the hash table
4217 in case it is something whose value we have seen before. */
4218
4219 if (MEM_P (x))
4220 {
4221 struct table_elt *elt;
4222
4223 x = fold_rtx (x, NULL_RTX);
4224 if (CONSTANT_P (x))
4225 return x;
4226
4227 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4228 if (elt == 0)
4229 return 0;
4230
4231 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4232 if (elt->is_const && CONSTANT_P (elt->exp))
4233 return elt->exp;
4234 }
4235
4236 return 0;
4237 }
4238 \f
4239 /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point
4240 number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
4241 least-significant part of X.
4242 MODE specifies how big a part of X to return.
4243
4244 If the requested operation cannot be done, 0 is returned.
4245
4246 This is similar to gen_lowpart_general in emit-rtl.c. */
4247
4248 rtx
4249 gen_lowpart_if_possible (enum machine_mode mode, rtx x)
4250 {
4251 rtx result = gen_lowpart_common (mode, x);
4252
4253 if (result)
4254 return result;
4255 else if (MEM_P (x))
4256 {
4257 /* This is the only other case we handle. */
4258 int offset = 0;
4259 rtx new;
4260
4261 if (WORDS_BIG_ENDIAN)
4262 offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD)
4263 - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD));
4264 if (BYTES_BIG_ENDIAN)
4265 /* Adjust the address so that the address-after-the-data is
4266 unchanged. */
4267 offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode))
4268 - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x))));
4269
4270 new = adjust_address_nv (x, mode, offset);
4271 if (! memory_address_p (mode, XEXP (new, 0)))
4272 return 0;
4273
4274 return new;
4275 }
4276 else
4277 return 0;
4278 }
4279 \f
4280 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4281 branch. It will be zero if not.
4282
4283 In certain cases, this can cause us to add an equivalence. For example,
4284 if we are following the taken case of
4285 if (i == 2)
4286 we can add the fact that `i' and '2' are now equivalent.
4287
4288 In any case, we can record that this comparison was passed. If the same
4289 comparison is seen later, we will know its value. */
4290
4291 static void
4292 record_jump_equiv (rtx insn, int taken)
4293 {
4294 int cond_known_true;
4295 rtx op0, op1;
4296 rtx set;
4297 enum machine_mode mode, mode0, mode1;
4298 int reversed_nonequality = 0;
4299 enum rtx_code code;
4300
4301 /* Ensure this is the right kind of insn. */
4302 if (! any_condjump_p (insn))
4303 return;
4304 set = pc_set (insn);
4305
4306 /* See if this jump condition is known true or false. */
4307 if (taken)
4308 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4309 else
4310 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4311
4312 /* Get the type of comparison being done and the operands being compared.
4313 If we had to reverse a non-equality condition, record that fact so we
4314 know that it isn't valid for floating-point. */
4315 code = GET_CODE (XEXP (SET_SRC (set), 0));
4316 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4317 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4318
4319 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4320 if (! cond_known_true)
4321 {
4322 code = reversed_comparison_code_parts (code, op0, op1, insn);
4323
4324 /* Don't remember if we can't find the inverse. */
4325 if (code == UNKNOWN)
4326 return;
4327 }
4328
4329 /* The mode is the mode of the non-constant. */
4330 mode = mode0;
4331 if (mode1 != VOIDmode)
4332 mode = mode1;
4333
4334 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4335 }
4336
4337 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4338 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4339 Make any useful entries we can with that information. Called from
4340 above function and called recursively. */
4341
4342 static void
4343 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4344 rtx op1, int reversed_nonequality)
4345 {
4346 unsigned op0_hash, op1_hash;
4347 int op0_in_memory, op1_in_memory;
4348 struct table_elt *op0_elt, *op1_elt;
4349
4350 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4351 we know that they are also equal in the smaller mode (this is also
4352 true for all smaller modes whether or not there is a SUBREG, but
4353 is not worth testing for with no SUBREG). */
4354
4355 /* Note that GET_MODE (op0) may not equal MODE. */
4356 if (code == EQ && GET_CODE (op0) == SUBREG
4357 && (GET_MODE_SIZE (GET_MODE (op0))
4358 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4359 {
4360 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4361 rtx tem = gen_lowpart (inner_mode, op1);
4362
4363 record_jump_cond (code, mode, SUBREG_REG (op0),
4364 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4365 reversed_nonequality);
4366 }
4367
4368 if (code == EQ && GET_CODE (op1) == SUBREG
4369 && (GET_MODE_SIZE (GET_MODE (op1))
4370 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4371 {
4372 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4373 rtx tem = gen_lowpart (inner_mode, op0);
4374
4375 record_jump_cond (code, mode, SUBREG_REG (op1),
4376 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4377 reversed_nonequality);
4378 }
4379
4380 /* Similarly, if this is an NE comparison, and either is a SUBREG
4381 making a smaller mode, we know the whole thing is also NE. */
4382
4383 /* Note that GET_MODE (op0) may not equal MODE;
4384 if we test MODE instead, we can get an infinite recursion
4385 alternating between two modes each wider than MODE. */
4386
4387 if (code == NE && GET_CODE (op0) == SUBREG
4388 && subreg_lowpart_p (op0)
4389 && (GET_MODE_SIZE (GET_MODE (op0))
4390 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4391 {
4392 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4393 rtx tem = gen_lowpart (inner_mode, op1);
4394
4395 record_jump_cond (code, mode, SUBREG_REG (op0),
4396 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4397 reversed_nonequality);
4398 }
4399
4400 if (code == NE && GET_CODE (op1) == SUBREG
4401 && subreg_lowpart_p (op1)
4402 && (GET_MODE_SIZE (GET_MODE (op1))
4403 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4404 {
4405 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4406 rtx tem = gen_lowpart (inner_mode, op0);
4407
4408 record_jump_cond (code, mode, SUBREG_REG (op1),
4409 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4410 reversed_nonequality);
4411 }
4412
4413 /* Hash both operands. */
4414
4415 do_not_record = 0;
4416 hash_arg_in_memory = 0;
4417 op0_hash = HASH (op0, mode);
4418 op0_in_memory = hash_arg_in_memory;
4419
4420 if (do_not_record)
4421 return;
4422
4423 do_not_record = 0;
4424 hash_arg_in_memory = 0;
4425 op1_hash = HASH (op1, mode);
4426 op1_in_memory = hash_arg_in_memory;
4427
4428 if (do_not_record)
4429 return;
4430
4431 /* Look up both operands. */
4432 op0_elt = lookup (op0, op0_hash, mode);
4433 op1_elt = lookup (op1, op1_hash, mode);
4434
4435 /* If both operands are already equivalent or if they are not in the
4436 table but are identical, do nothing. */
4437 if ((op0_elt != 0 && op1_elt != 0
4438 && op0_elt->first_same_value == op1_elt->first_same_value)
4439 || op0 == op1 || rtx_equal_p (op0, op1))
4440 return;
4441
4442 /* If we aren't setting two things equal all we can do is save this
4443 comparison. Similarly if this is floating-point. In the latter
4444 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4445 If we record the equality, we might inadvertently delete code
4446 whose intent was to change -0 to +0. */
4447
4448 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4449 {
4450 struct qty_table_elem *ent;
4451 int qty;
4452
4453 /* If we reversed a floating-point comparison, if OP0 is not a
4454 register, or if OP1 is neither a register or constant, we can't
4455 do anything. */
4456
4457 if (!REG_P (op1))
4458 op1 = equiv_constant (op1);
4459
4460 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4461 || !REG_P (op0) || op1 == 0)
4462 return;
4463
4464 /* Put OP0 in the hash table if it isn't already. This gives it a
4465 new quantity number. */
4466 if (op0_elt == 0)
4467 {
4468 if (insert_regs (op0, NULL, 0))
4469 {
4470 rehash_using_reg (op0);
4471 op0_hash = HASH (op0, mode);
4472
4473 /* If OP0 is contained in OP1, this changes its hash code
4474 as well. Faster to rehash than to check, except
4475 for the simple case of a constant. */
4476 if (! CONSTANT_P (op1))
4477 op1_hash = HASH (op1,mode);
4478 }
4479
4480 op0_elt = insert (op0, NULL, op0_hash, mode);
4481 op0_elt->in_memory = op0_in_memory;
4482 }
4483
4484 qty = REG_QTY (REGNO (op0));
4485 ent = &qty_table[qty];
4486
4487 ent->comparison_code = code;
4488 if (REG_P (op1))
4489 {
4490 /* Look it up again--in case op0 and op1 are the same. */
4491 op1_elt = lookup (op1, op1_hash, mode);
4492
4493 /* Put OP1 in the hash table so it gets a new quantity number. */
4494 if (op1_elt == 0)
4495 {
4496 if (insert_regs (op1, NULL, 0))
4497 {
4498 rehash_using_reg (op1);
4499 op1_hash = HASH (op1, mode);
4500 }
4501
4502 op1_elt = insert (op1, NULL, op1_hash, mode);
4503 op1_elt->in_memory = op1_in_memory;
4504 }
4505
4506 ent->comparison_const = NULL_RTX;
4507 ent->comparison_qty = REG_QTY (REGNO (op1));
4508 }
4509 else
4510 {
4511 ent->comparison_const = op1;
4512 ent->comparison_qty = -1;
4513 }
4514
4515 return;
4516 }
4517
4518 /* If either side is still missing an equivalence, make it now,
4519 then merge the equivalences. */
4520
4521 if (op0_elt == 0)
4522 {
4523 if (insert_regs (op0, NULL, 0))
4524 {
4525 rehash_using_reg (op0);
4526 op0_hash = HASH (op0, mode);
4527 }
4528
4529 op0_elt = insert (op0, NULL, op0_hash, mode);
4530 op0_elt->in_memory = op0_in_memory;
4531 }
4532
4533 if (op1_elt == 0)
4534 {
4535 if (insert_regs (op1, NULL, 0))
4536 {
4537 rehash_using_reg (op1);
4538 op1_hash = HASH (op1, mode);
4539 }
4540
4541 op1_elt = insert (op1, NULL, op1_hash, mode);
4542 op1_elt->in_memory = op1_in_memory;
4543 }
4544
4545 merge_equiv_classes (op0_elt, op1_elt);
4546 }
4547 \f
4548 /* CSE processing for one instruction.
4549 First simplify sources and addresses of all assignments
4550 in the instruction, using previously-computed equivalents values.
4551 Then install the new sources and destinations in the table
4552 of available values.
4553
4554 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4555 the insn. It means that INSN is inside libcall block. In this
4556 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4557
4558 /* Data on one SET contained in the instruction. */
4559
4560 struct set
4561 {
4562 /* The SET rtx itself. */
4563 rtx rtl;
4564 /* The SET_SRC of the rtx (the original value, if it is changing). */
4565 rtx src;
4566 /* The hash-table element for the SET_SRC of the SET. */
4567 struct table_elt *src_elt;
4568 /* Hash value for the SET_SRC. */
4569 unsigned src_hash;
4570 /* Hash value for the SET_DEST. */
4571 unsigned dest_hash;
4572 /* The SET_DEST, with SUBREG, etc., stripped. */
4573 rtx inner_dest;
4574 /* Nonzero if the SET_SRC is in memory. */
4575 char src_in_memory;
4576 /* Nonzero if the SET_SRC contains something
4577 whose value cannot be predicted and understood. */
4578 char src_volatile;
4579 /* Original machine mode, in case it becomes a CONST_INT.
4580 The size of this field should match the size of the mode
4581 field of struct rtx_def (see rtl.h). */
4582 ENUM_BITFIELD(machine_mode) mode : 8;
4583 /* A constant equivalent for SET_SRC, if any. */
4584 rtx src_const;
4585 /* Original SET_SRC value used for libcall notes. */
4586 rtx orig_src;
4587 /* Hash value of constant equivalent for SET_SRC. */
4588 unsigned src_const_hash;
4589 /* Table entry for constant equivalent for SET_SRC, if any. */
4590 struct table_elt *src_const_elt;
4591 };
4592
4593 static void
4594 cse_insn (rtx insn, rtx libcall_insn)
4595 {
4596 rtx x = PATTERN (insn);
4597 int i;
4598 rtx tem;
4599 int n_sets = 0;
4600
4601 #ifdef HAVE_cc0
4602 /* Records what this insn does to set CC0. */
4603 rtx this_insn_cc0 = 0;
4604 enum machine_mode this_insn_cc0_mode = VOIDmode;
4605 #endif
4606
4607 rtx src_eqv = 0;
4608 struct table_elt *src_eqv_elt = 0;
4609 int src_eqv_volatile = 0;
4610 int src_eqv_in_memory = 0;
4611 unsigned src_eqv_hash = 0;
4612
4613 struct set *sets = (struct set *) 0;
4614
4615 this_insn = insn;
4616
4617 /* Find all the SETs and CLOBBERs in this instruction.
4618 Record all the SETs in the array `set' and count them.
4619 Also determine whether there is a CLOBBER that invalidates
4620 all memory references, or all references at varying addresses. */
4621
4622 if (CALL_P (insn))
4623 {
4624 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4625 {
4626 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4627 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4628 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4629 }
4630 }
4631
4632 if (GET_CODE (x) == SET)
4633 {
4634 sets = alloca (sizeof (struct set));
4635 sets[0].rtl = x;
4636
4637 /* Ignore SETs that are unconditional jumps.
4638 They never need cse processing, so this does not hurt.
4639 The reason is not efficiency but rather
4640 so that we can test at the end for instructions
4641 that have been simplified to unconditional jumps
4642 and not be misled by unchanged instructions
4643 that were unconditional jumps to begin with. */
4644 if (SET_DEST (x) == pc_rtx
4645 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4646 ;
4647
4648 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4649 The hard function value register is used only once, to copy to
4650 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4651 Ensure we invalidate the destination register. On the 80386 no
4652 other code would invalidate it since it is a fixed_reg.
4653 We need not check the return of apply_change_group; see canon_reg. */
4654
4655 else if (GET_CODE (SET_SRC (x)) == CALL)
4656 {
4657 canon_reg (SET_SRC (x), insn);
4658 apply_change_group ();
4659 fold_rtx (SET_SRC (x), insn);
4660 invalidate (SET_DEST (x), VOIDmode);
4661 }
4662 else
4663 n_sets = 1;
4664 }
4665 else if (GET_CODE (x) == PARALLEL)
4666 {
4667 int lim = XVECLEN (x, 0);
4668
4669 sets = alloca (lim * sizeof (struct set));
4670
4671 /* Find all regs explicitly clobbered in this insn,
4672 and ensure they are not replaced with any other regs
4673 elsewhere in this insn.
4674 When a reg that is clobbered is also used for input,
4675 we should presume that that is for a reason,
4676 and we should not substitute some other register
4677 which is not supposed to be clobbered.
4678 Therefore, this loop cannot be merged into the one below
4679 because a CALL may precede a CLOBBER and refer to the
4680 value clobbered. We must not let a canonicalization do
4681 anything in that case. */
4682 for (i = 0; i < lim; i++)
4683 {
4684 rtx y = XVECEXP (x, 0, i);
4685 if (GET_CODE (y) == CLOBBER)
4686 {
4687 rtx clobbered = XEXP (y, 0);
4688
4689 if (REG_P (clobbered)
4690 || GET_CODE (clobbered) == SUBREG)
4691 invalidate (clobbered, VOIDmode);
4692 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4693 || GET_CODE (clobbered) == ZERO_EXTRACT)
4694 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4695 }
4696 }
4697
4698 for (i = 0; i < lim; i++)
4699 {
4700 rtx y = XVECEXP (x, 0, i);
4701 if (GET_CODE (y) == SET)
4702 {
4703 /* As above, we ignore unconditional jumps and call-insns and
4704 ignore the result of apply_change_group. */
4705 if (GET_CODE (SET_SRC (y)) == CALL)
4706 {
4707 canon_reg (SET_SRC (y), insn);
4708 apply_change_group ();
4709 fold_rtx (SET_SRC (y), insn);
4710 invalidate (SET_DEST (y), VOIDmode);
4711 }
4712 else if (SET_DEST (y) == pc_rtx
4713 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4714 ;
4715 else
4716 sets[n_sets++].rtl = y;
4717 }
4718 else if (GET_CODE (y) == CLOBBER)
4719 {
4720 /* If we clobber memory, canon the address.
4721 This does nothing when a register is clobbered
4722 because we have already invalidated the reg. */
4723 if (MEM_P (XEXP (y, 0)))
4724 canon_reg (XEXP (y, 0), NULL_RTX);
4725 }
4726 else if (GET_CODE (y) == USE
4727 && ! (REG_P (XEXP (y, 0))
4728 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4729 canon_reg (y, NULL_RTX);
4730 else if (GET_CODE (y) == CALL)
4731 {
4732 /* The result of apply_change_group can be ignored; see
4733 canon_reg. */
4734 canon_reg (y, insn);
4735 apply_change_group ();
4736 fold_rtx (y, insn);
4737 }
4738 }
4739 }
4740 else if (GET_CODE (x) == CLOBBER)
4741 {
4742 if (MEM_P (XEXP (x, 0)))
4743 canon_reg (XEXP (x, 0), NULL_RTX);
4744 }
4745
4746 /* Canonicalize a USE of a pseudo register or memory location. */
4747 else if (GET_CODE (x) == USE
4748 && ! (REG_P (XEXP (x, 0))
4749 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4750 canon_reg (XEXP (x, 0), NULL_RTX);
4751 else if (GET_CODE (x) == CALL)
4752 {
4753 /* The result of apply_change_group can be ignored; see canon_reg. */
4754 canon_reg (x, insn);
4755 apply_change_group ();
4756 fold_rtx (x, insn);
4757 }
4758
4759 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4760 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4761 is handled specially for this case, and if it isn't set, then there will
4762 be no equivalence for the destination. */
4763 if (n_sets == 1 && REG_NOTES (insn) != 0
4764 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4765 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4766 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4767 {
4768 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4769 XEXP (tem, 0) = src_eqv;
4770 }
4771
4772 /* Canonicalize sources and addresses of destinations.
4773 We do this in a separate pass to avoid problems when a MATCH_DUP is
4774 present in the insn pattern. In that case, we want to ensure that
4775 we don't break the duplicate nature of the pattern. So we will replace
4776 both operands at the same time. Otherwise, we would fail to find an
4777 equivalent substitution in the loop calling validate_change below.
4778
4779 We used to suppress canonicalization of DEST if it appears in SRC,
4780 but we don't do this any more. */
4781
4782 for (i = 0; i < n_sets; i++)
4783 {
4784 rtx dest = SET_DEST (sets[i].rtl);
4785 rtx src = SET_SRC (sets[i].rtl);
4786 rtx new = canon_reg (src, insn);
4787 int insn_code;
4788
4789 sets[i].orig_src = src;
4790 if ((REG_P (new) && REG_P (src)
4791 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4792 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4793 || (insn_code = recog_memoized (insn)) < 0
4794 || insn_data[insn_code].n_dups > 0)
4795 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4796 else
4797 SET_SRC (sets[i].rtl) = new;
4798
4799 if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT)
4800 {
4801 validate_change (insn, &XEXP (dest, 1),
4802 canon_reg (XEXP (dest, 1), insn), 1);
4803 validate_change (insn, &XEXP (dest, 2),
4804 canon_reg (XEXP (dest, 2), insn), 1);
4805 }
4806
4807 while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART
4808 || GET_CODE (dest) == ZERO_EXTRACT
4809 || GET_CODE (dest) == SIGN_EXTRACT)
4810 dest = XEXP (dest, 0);
4811
4812 if (MEM_P (dest))
4813 canon_reg (dest, insn);
4814 }
4815
4816 /* Now that we have done all the replacements, we can apply the change
4817 group and see if they all work. Note that this will cause some
4818 canonicalizations that would have worked individually not to be applied
4819 because some other canonicalization didn't work, but this should not
4820 occur often.
4821
4822 The result of apply_change_group can be ignored; see canon_reg. */
4823
4824 apply_change_group ();
4825
4826 /* Set sets[i].src_elt to the class each source belongs to.
4827 Detect assignments from or to volatile things
4828 and set set[i] to zero so they will be ignored
4829 in the rest of this function.
4830
4831 Nothing in this loop changes the hash table or the register chains. */
4832
4833 for (i = 0; i < n_sets; i++)
4834 {
4835 rtx src, dest;
4836 rtx src_folded;
4837 struct table_elt *elt = 0, *p;
4838 enum machine_mode mode;
4839 rtx src_eqv_here;
4840 rtx src_const = 0;
4841 rtx src_related = 0;
4842 struct table_elt *src_const_elt = 0;
4843 int src_cost = MAX_COST;
4844 int src_eqv_cost = MAX_COST;
4845 int src_folded_cost = MAX_COST;
4846 int src_related_cost = MAX_COST;
4847 int src_elt_cost = MAX_COST;
4848 int src_regcost = MAX_COST;
4849 int src_eqv_regcost = MAX_COST;
4850 int src_folded_regcost = MAX_COST;
4851 int src_related_regcost = MAX_COST;
4852 int src_elt_regcost = MAX_COST;
4853 /* Set nonzero if we need to call force_const_mem on with the
4854 contents of src_folded before using it. */
4855 int src_folded_force_flag = 0;
4856
4857 dest = SET_DEST (sets[i].rtl);
4858 src = SET_SRC (sets[i].rtl);
4859
4860 /* If SRC is a constant that has no machine mode,
4861 hash it with the destination's machine mode.
4862 This way we can keep different modes separate. */
4863
4864 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4865 sets[i].mode = mode;
4866
4867 if (src_eqv)
4868 {
4869 enum machine_mode eqvmode = mode;
4870 if (GET_CODE (dest) == STRICT_LOW_PART)
4871 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4872 do_not_record = 0;
4873 hash_arg_in_memory = 0;
4874 src_eqv_hash = HASH (src_eqv, eqvmode);
4875
4876 /* Find the equivalence class for the equivalent expression. */
4877
4878 if (!do_not_record)
4879 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4880
4881 src_eqv_volatile = do_not_record;
4882 src_eqv_in_memory = hash_arg_in_memory;
4883 }
4884
4885 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4886 value of the INNER register, not the destination. So it is not
4887 a valid substitution for the source. But save it for later. */
4888 if (GET_CODE (dest) == STRICT_LOW_PART)
4889 src_eqv_here = 0;
4890 else
4891 src_eqv_here = src_eqv;
4892
4893 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4894 simplified result, which may not necessarily be valid. */
4895 src_folded = fold_rtx (src, insn);
4896
4897 #if 0
4898 /* ??? This caused bad code to be generated for the m68k port with -O2.
4899 Suppose src is (CONST_INT -1), and that after truncation src_folded
4900 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4901 At the end we will add src and src_const to the same equivalence
4902 class. We now have 3 and -1 on the same equivalence class. This
4903 causes later instructions to be mis-optimized. */
4904 /* If storing a constant in a bitfield, pre-truncate the constant
4905 so we will be able to record it later. */
4906 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
4907 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
4908 {
4909 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4910
4911 if (GET_CODE (src) == CONST_INT
4912 && GET_CODE (width) == CONST_INT
4913 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4914 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4915 src_folded
4916 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4917 << INTVAL (width)) - 1));
4918 }
4919 #endif
4920
4921 /* Compute SRC's hash code, and also notice if it
4922 should not be recorded at all. In that case,
4923 prevent any further processing of this assignment. */
4924 do_not_record = 0;
4925 hash_arg_in_memory = 0;
4926
4927 sets[i].src = src;
4928 sets[i].src_hash = HASH (src, mode);
4929 sets[i].src_volatile = do_not_record;
4930 sets[i].src_in_memory = hash_arg_in_memory;
4931
4932 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4933 a pseudo, do not record SRC. Using SRC as a replacement for
4934 anything else will be incorrect in that situation. Note that
4935 this usually occurs only for stack slots, in which case all the
4936 RTL would be referring to SRC, so we don't lose any optimization
4937 opportunities by not having SRC in the hash table. */
4938
4939 if (MEM_P (src)
4940 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4941 && REG_P (dest)
4942 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4943 sets[i].src_volatile = 1;
4944
4945 #if 0
4946 /* It is no longer clear why we used to do this, but it doesn't
4947 appear to still be needed. So let's try without it since this
4948 code hurts cse'ing widened ops. */
4949 /* If source is a paradoxical subreg (such as QI treated as an SI),
4950 treat it as volatile. It may do the work of an SI in one context
4951 where the extra bits are not being used, but cannot replace an SI
4952 in general. */
4953 if (GET_CODE (src) == SUBREG
4954 && (GET_MODE_SIZE (GET_MODE (src))
4955 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
4956 sets[i].src_volatile = 1;
4957 #endif
4958
4959 /* Locate all possible equivalent forms for SRC. Try to replace
4960 SRC in the insn with each cheaper equivalent.
4961
4962 We have the following types of equivalents: SRC itself, a folded
4963 version, a value given in a REG_EQUAL note, or a value related
4964 to a constant.
4965
4966 Each of these equivalents may be part of an additional class
4967 of equivalents (if more than one is in the table, they must be in
4968 the same class; we check for this).
4969
4970 If the source is volatile, we don't do any table lookups.
4971
4972 We note any constant equivalent for possible later use in a
4973 REG_NOTE. */
4974
4975 if (!sets[i].src_volatile)
4976 elt = lookup (src, sets[i].src_hash, mode);
4977
4978 sets[i].src_elt = elt;
4979
4980 if (elt && src_eqv_here && src_eqv_elt)
4981 {
4982 if (elt->first_same_value != src_eqv_elt->first_same_value)
4983 {
4984 /* The REG_EQUAL is indicating that two formerly distinct
4985 classes are now equivalent. So merge them. */
4986 merge_equiv_classes (elt, src_eqv_elt);
4987 src_eqv_hash = HASH (src_eqv, elt->mode);
4988 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4989 }
4990
4991 src_eqv_here = 0;
4992 }
4993
4994 else if (src_eqv_elt)
4995 elt = src_eqv_elt;
4996
4997 /* Try to find a constant somewhere and record it in `src_const'.
4998 Record its table element, if any, in `src_const_elt'. Look in
4999 any known equivalences first. (If the constant is not in the
5000 table, also set `sets[i].src_const_hash'). */
5001 if (elt)
5002 for (p = elt->first_same_value; p; p = p->next_same_value)
5003 if (p->is_const)
5004 {
5005 src_const = p->exp;
5006 src_const_elt = elt;
5007 break;
5008 }
5009
5010 if (src_const == 0
5011 && (CONSTANT_P (src_folded)
5012 /* Consider (minus (label_ref L1) (label_ref L2)) as
5013 "constant" here so we will record it. This allows us
5014 to fold switch statements when an ADDR_DIFF_VEC is used. */
5015 || (GET_CODE (src_folded) == MINUS
5016 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5017 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5018 src_const = src_folded, src_const_elt = elt;
5019 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5020 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5021
5022 /* If we don't know if the constant is in the table, get its
5023 hash code and look it up. */
5024 if (src_const && src_const_elt == 0)
5025 {
5026 sets[i].src_const_hash = HASH (src_const, mode);
5027 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5028 }
5029
5030 sets[i].src_const = src_const;
5031 sets[i].src_const_elt = src_const_elt;
5032
5033 /* If the constant and our source are both in the table, mark them as
5034 equivalent. Otherwise, if a constant is in the table but the source
5035 isn't, set ELT to it. */
5036 if (src_const_elt && elt
5037 && src_const_elt->first_same_value != elt->first_same_value)
5038 merge_equiv_classes (elt, src_const_elt);
5039 else if (src_const_elt && elt == 0)
5040 elt = src_const_elt;
5041
5042 /* See if there is a register linearly related to a constant
5043 equivalent of SRC. */
5044 if (src_const
5045 && (GET_CODE (src_const) == CONST
5046 || (src_const_elt && src_const_elt->related_value != 0)))
5047 {
5048 src_related = use_related_value (src_const, src_const_elt);
5049 if (src_related)
5050 {
5051 struct table_elt *src_related_elt
5052 = lookup (src_related, HASH (src_related, mode), mode);
5053 if (src_related_elt && elt)
5054 {
5055 if (elt->first_same_value
5056 != src_related_elt->first_same_value)
5057 /* This can occur when we previously saw a CONST
5058 involving a SYMBOL_REF and then see the SYMBOL_REF
5059 twice. Merge the involved classes. */
5060 merge_equiv_classes (elt, src_related_elt);
5061
5062 src_related = 0;
5063 src_related_elt = 0;
5064 }
5065 else if (src_related_elt && elt == 0)
5066 elt = src_related_elt;
5067 }
5068 }
5069
5070 /* See if we have a CONST_INT that is already in a register in a
5071 wider mode. */
5072
5073 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5074 && GET_MODE_CLASS (mode) == MODE_INT
5075 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5076 {
5077 enum machine_mode wider_mode;
5078
5079 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5080 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5081 && src_related == 0;
5082 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5083 {
5084 struct table_elt *const_elt
5085 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5086
5087 if (const_elt == 0)
5088 continue;
5089
5090 for (const_elt = const_elt->first_same_value;
5091 const_elt; const_elt = const_elt->next_same_value)
5092 if (REG_P (const_elt->exp))
5093 {
5094 src_related = gen_lowpart (mode,
5095 const_elt->exp);
5096 break;
5097 }
5098 }
5099 }
5100
5101 /* Another possibility is that we have an AND with a constant in
5102 a mode narrower than a word. If so, it might have been generated
5103 as part of an "if" which would narrow the AND. If we already
5104 have done the AND in a wider mode, we can use a SUBREG of that
5105 value. */
5106
5107 if (flag_expensive_optimizations && ! src_related
5108 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5109 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5110 {
5111 enum machine_mode tmode;
5112 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5113
5114 for (tmode = GET_MODE_WIDER_MODE (mode);
5115 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5116 tmode = GET_MODE_WIDER_MODE (tmode))
5117 {
5118 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5119 struct table_elt *larger_elt;
5120
5121 if (inner)
5122 {
5123 PUT_MODE (new_and, tmode);
5124 XEXP (new_and, 0) = inner;
5125 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5126 if (larger_elt == 0)
5127 continue;
5128
5129 for (larger_elt = larger_elt->first_same_value;
5130 larger_elt; larger_elt = larger_elt->next_same_value)
5131 if (REG_P (larger_elt->exp))
5132 {
5133 src_related
5134 = gen_lowpart (mode, larger_elt->exp);
5135 break;
5136 }
5137
5138 if (src_related)
5139 break;
5140 }
5141 }
5142 }
5143
5144 #ifdef LOAD_EXTEND_OP
5145 /* See if a MEM has already been loaded with a widening operation;
5146 if it has, we can use a subreg of that. Many CISC machines
5147 also have such operations, but this is only likely to be
5148 beneficial on these machines. */
5149
5150 if (flag_expensive_optimizations && src_related == 0
5151 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5152 && GET_MODE_CLASS (mode) == MODE_INT
5153 && MEM_P (src) && ! do_not_record
5154 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5155 {
5156 enum machine_mode tmode;
5157
5158 /* Set what we are trying to extend and the operation it might
5159 have been extended with. */
5160 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5161 XEXP (memory_extend_rtx, 0) = src;
5162
5163 for (tmode = GET_MODE_WIDER_MODE (mode);
5164 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5165 tmode = GET_MODE_WIDER_MODE (tmode))
5166 {
5167 struct table_elt *larger_elt;
5168
5169 PUT_MODE (memory_extend_rtx, tmode);
5170 larger_elt = lookup (memory_extend_rtx,
5171 HASH (memory_extend_rtx, tmode), tmode);
5172 if (larger_elt == 0)
5173 continue;
5174
5175 for (larger_elt = larger_elt->first_same_value;
5176 larger_elt; larger_elt = larger_elt->next_same_value)
5177 if (REG_P (larger_elt->exp))
5178 {
5179 src_related = gen_lowpart (mode,
5180 larger_elt->exp);
5181 break;
5182 }
5183
5184 if (src_related)
5185 break;
5186 }
5187 }
5188 #endif /* LOAD_EXTEND_OP */
5189
5190 if (src == src_folded)
5191 src_folded = 0;
5192
5193 /* At this point, ELT, if nonzero, points to a class of expressions
5194 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5195 and SRC_RELATED, if nonzero, each contain additional equivalent
5196 expressions. Prune these latter expressions by deleting expressions
5197 already in the equivalence class.
5198
5199 Check for an equivalent identical to the destination. If found,
5200 this is the preferred equivalent since it will likely lead to
5201 elimination of the insn. Indicate this by placing it in
5202 `src_related'. */
5203
5204 if (elt)
5205 elt = elt->first_same_value;
5206 for (p = elt; p; p = p->next_same_value)
5207 {
5208 enum rtx_code code = GET_CODE (p->exp);
5209
5210 /* If the expression is not valid, ignore it. Then we do not
5211 have to check for validity below. In most cases, we can use
5212 `rtx_equal_p', since canonicalization has already been done. */
5213 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5214 continue;
5215
5216 /* Also skip paradoxical subregs, unless that's what we're
5217 looking for. */
5218 if (code == SUBREG
5219 && (GET_MODE_SIZE (GET_MODE (p->exp))
5220 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5221 && ! (src != 0
5222 && GET_CODE (src) == SUBREG
5223 && GET_MODE (src) == GET_MODE (p->exp)
5224 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5225 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5226 continue;
5227
5228 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5229 src = 0;
5230 else if (src_folded && GET_CODE (src_folded) == code
5231 && rtx_equal_p (src_folded, p->exp))
5232 src_folded = 0;
5233 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5234 && rtx_equal_p (src_eqv_here, p->exp))
5235 src_eqv_here = 0;
5236 else if (src_related && GET_CODE (src_related) == code
5237 && rtx_equal_p (src_related, p->exp))
5238 src_related = 0;
5239
5240 /* This is the same as the destination of the insns, we want
5241 to prefer it. Copy it to src_related. The code below will
5242 then give it a negative cost. */
5243 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5244 src_related = dest;
5245 }
5246
5247 /* Find the cheapest valid equivalent, trying all the available
5248 possibilities. Prefer items not in the hash table to ones
5249 that are when they are equal cost. Note that we can never
5250 worsen an insn as the current contents will also succeed.
5251 If we find an equivalent identical to the destination, use it as best,
5252 since this insn will probably be eliminated in that case. */
5253 if (src)
5254 {
5255 if (rtx_equal_p (src, dest))
5256 src_cost = src_regcost = -1;
5257 else
5258 {
5259 src_cost = COST (src);
5260 src_regcost = approx_reg_cost (src);
5261 }
5262 }
5263
5264 if (src_eqv_here)
5265 {
5266 if (rtx_equal_p (src_eqv_here, dest))
5267 src_eqv_cost = src_eqv_regcost = -1;
5268 else
5269 {
5270 src_eqv_cost = COST (src_eqv_here);
5271 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5272 }
5273 }
5274
5275 if (src_folded)
5276 {
5277 if (rtx_equal_p (src_folded, dest))
5278 src_folded_cost = src_folded_regcost = -1;
5279 else
5280 {
5281 src_folded_cost = COST (src_folded);
5282 src_folded_regcost = approx_reg_cost (src_folded);
5283 }
5284 }
5285
5286 if (src_related)
5287 {
5288 if (rtx_equal_p (src_related, dest))
5289 src_related_cost = src_related_regcost = -1;
5290 else
5291 {
5292 src_related_cost = COST (src_related);
5293 src_related_regcost = approx_reg_cost (src_related);
5294 }
5295 }
5296
5297 /* If this was an indirect jump insn, a known label will really be
5298 cheaper even though it looks more expensive. */
5299 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5300 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5301
5302 /* Terminate loop when replacement made. This must terminate since
5303 the current contents will be tested and will always be valid. */
5304 while (1)
5305 {
5306 rtx trial;
5307
5308 /* Skip invalid entries. */
5309 while (elt && !REG_P (elt->exp)
5310 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5311 elt = elt->next_same_value;
5312
5313 /* A paradoxical subreg would be bad here: it'll be the right
5314 size, but later may be adjusted so that the upper bits aren't
5315 what we want. So reject it. */
5316 if (elt != 0
5317 && GET_CODE (elt->exp) == SUBREG
5318 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5319 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5320 /* It is okay, though, if the rtx we're trying to match
5321 will ignore any of the bits we can't predict. */
5322 && ! (src != 0
5323 && GET_CODE (src) == SUBREG
5324 && GET_MODE (src) == GET_MODE (elt->exp)
5325 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5326 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5327 {
5328 elt = elt->next_same_value;
5329 continue;
5330 }
5331
5332 if (elt)
5333 {
5334 src_elt_cost = elt->cost;
5335 src_elt_regcost = elt->regcost;
5336 }
5337
5338 /* Find cheapest and skip it for the next time. For items
5339 of equal cost, use this order:
5340 src_folded, src, src_eqv, src_related and hash table entry. */
5341 if (src_folded
5342 && preferable (src_folded_cost, src_folded_regcost,
5343 src_cost, src_regcost) <= 0
5344 && preferable (src_folded_cost, src_folded_regcost,
5345 src_eqv_cost, src_eqv_regcost) <= 0
5346 && preferable (src_folded_cost, src_folded_regcost,
5347 src_related_cost, src_related_regcost) <= 0
5348 && preferable (src_folded_cost, src_folded_regcost,
5349 src_elt_cost, src_elt_regcost) <= 0)
5350 {
5351 trial = src_folded, src_folded_cost = MAX_COST;
5352 if (src_folded_force_flag)
5353 {
5354 rtx forced = force_const_mem (mode, trial);
5355 if (forced)
5356 trial = forced;
5357 }
5358 }
5359 else if (src
5360 && preferable (src_cost, src_regcost,
5361 src_eqv_cost, src_eqv_regcost) <= 0
5362 && preferable (src_cost, src_regcost,
5363 src_related_cost, src_related_regcost) <= 0
5364 && preferable (src_cost, src_regcost,
5365 src_elt_cost, src_elt_regcost) <= 0)
5366 trial = src, src_cost = MAX_COST;
5367 else if (src_eqv_here
5368 && preferable (src_eqv_cost, src_eqv_regcost,
5369 src_related_cost, src_related_regcost) <= 0
5370 && preferable (src_eqv_cost, src_eqv_regcost,
5371 src_elt_cost, src_elt_regcost) <= 0)
5372 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5373 else if (src_related
5374 && preferable (src_related_cost, src_related_regcost,
5375 src_elt_cost, src_elt_regcost) <= 0)
5376 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5377 else
5378 {
5379 trial = copy_rtx (elt->exp);
5380 elt = elt->next_same_value;
5381 src_elt_cost = MAX_COST;
5382 }
5383
5384 /* We don't normally have an insn matching (set (pc) (pc)), so
5385 check for this separately here. We will delete such an
5386 insn below.
5387
5388 For other cases such as a table jump or conditional jump
5389 where we know the ultimate target, go ahead and replace the
5390 operand. While that may not make a valid insn, we will
5391 reemit the jump below (and also insert any necessary
5392 barriers). */
5393 if (n_sets == 1 && dest == pc_rtx
5394 && (trial == pc_rtx
5395 || (GET_CODE (trial) == LABEL_REF
5396 && ! condjump_p (insn))))
5397 {
5398 /* Don't substitute non-local labels, this confuses CFG. */
5399 if (GET_CODE (trial) == LABEL_REF
5400 && LABEL_REF_NONLOCAL_P (trial))
5401 continue;
5402
5403 SET_SRC (sets[i].rtl) = trial;
5404 cse_jumps_altered = 1;
5405 break;
5406 }
5407
5408 /* Look for a substitution that makes a valid insn. */
5409 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5410 {
5411 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5412
5413 /* If we just made a substitution inside a libcall, then we
5414 need to make the same substitution in any notes attached
5415 to the RETVAL insn. */
5416 if (libcall_insn
5417 && (REG_P (sets[i].orig_src)
5418 || GET_CODE (sets[i].orig_src) == SUBREG
5419 || MEM_P (sets[i].orig_src)))
5420 {
5421 rtx note = find_reg_equal_equiv_note (libcall_insn);
5422 if (note != 0)
5423 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5424 sets[i].orig_src,
5425 copy_rtx (new));
5426 }
5427
5428 /* The result of apply_change_group can be ignored; see
5429 canon_reg. */
5430
5431 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5432 apply_change_group ();
5433 break;
5434 }
5435
5436 /* If we previously found constant pool entries for
5437 constants and this is a constant, try making a
5438 pool entry. Put it in src_folded unless we already have done
5439 this since that is where it likely came from. */
5440
5441 else if (constant_pool_entries_cost
5442 && CONSTANT_P (trial)
5443 /* Reject cases that will abort in decode_rtx_const.
5444 On the alpha when simplifying a switch, we get
5445 (const (truncate (minus (label_ref) (label_ref)))). */
5446 && ! (GET_CODE (trial) == CONST
5447 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5448 /* Likewise on IA-64, except without the truncate. */
5449 && ! (GET_CODE (trial) == CONST
5450 && GET_CODE (XEXP (trial, 0)) == MINUS
5451 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5452 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5453 && (src_folded == 0
5454 || (!MEM_P (src_folded)
5455 && ! src_folded_force_flag))
5456 && GET_MODE_CLASS (mode) != MODE_CC
5457 && mode != VOIDmode)
5458 {
5459 src_folded_force_flag = 1;
5460 src_folded = trial;
5461 src_folded_cost = constant_pool_entries_cost;
5462 src_folded_regcost = constant_pool_entries_regcost;
5463 }
5464 }
5465
5466 src = SET_SRC (sets[i].rtl);
5467
5468 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5469 However, there is an important exception: If both are registers
5470 that are not the head of their equivalence class, replace SET_SRC
5471 with the head of the class. If we do not do this, we will have
5472 both registers live over a portion of the basic block. This way,
5473 their lifetimes will likely abut instead of overlapping. */
5474 if (REG_P (dest)
5475 && REGNO_QTY_VALID_P (REGNO (dest)))
5476 {
5477 int dest_q = REG_QTY (REGNO (dest));
5478 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5479
5480 if (dest_ent->mode == GET_MODE (dest)
5481 && dest_ent->first_reg != REGNO (dest)
5482 && REG_P (src) && REGNO (src) == REGNO (dest)
5483 /* Don't do this if the original insn had a hard reg as
5484 SET_SRC or SET_DEST. */
5485 && (!REG_P (sets[i].src)
5486 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5487 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5488 /* We can't call canon_reg here because it won't do anything if
5489 SRC is a hard register. */
5490 {
5491 int src_q = REG_QTY (REGNO (src));
5492 struct qty_table_elem *src_ent = &qty_table[src_q];
5493 int first = src_ent->first_reg;
5494 rtx new_src
5495 = (first >= FIRST_PSEUDO_REGISTER
5496 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5497
5498 /* We must use validate-change even for this, because this
5499 might be a special no-op instruction, suitable only to
5500 tag notes onto. */
5501 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5502 {
5503 src = new_src;
5504 /* If we had a constant that is cheaper than what we are now
5505 setting SRC to, use that constant. We ignored it when we
5506 thought we could make this into a no-op. */
5507 if (src_const && COST (src_const) < COST (src)
5508 && validate_change (insn, &SET_SRC (sets[i].rtl),
5509 src_const, 0))
5510 src = src_const;
5511 }
5512 }
5513 }
5514
5515 /* If we made a change, recompute SRC values. */
5516 if (src != sets[i].src)
5517 {
5518 cse_altered = 1;
5519 do_not_record = 0;
5520 hash_arg_in_memory = 0;
5521 sets[i].src = src;
5522 sets[i].src_hash = HASH (src, mode);
5523 sets[i].src_volatile = do_not_record;
5524 sets[i].src_in_memory = hash_arg_in_memory;
5525 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5526 }
5527
5528 /* If this is a single SET, we are setting a register, and we have an
5529 equivalent constant, we want to add a REG_NOTE. We don't want
5530 to write a REG_EQUAL note for a constant pseudo since verifying that
5531 that pseudo hasn't been eliminated is a pain. Such a note also
5532 won't help anything.
5533
5534 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5535 which can be created for a reference to a compile time computable
5536 entry in a jump table. */
5537
5538 if (n_sets == 1 && src_const && REG_P (dest)
5539 && !REG_P (src_const)
5540 && ! (GET_CODE (src_const) == CONST
5541 && GET_CODE (XEXP (src_const, 0)) == MINUS
5542 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5543 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5544 {
5545 /* We only want a REG_EQUAL note if src_const != src. */
5546 if (! rtx_equal_p (src, src_const))
5547 {
5548 /* Make sure that the rtx is not shared. */
5549 src_const = copy_rtx (src_const);
5550
5551 /* Record the actual constant value in a REG_EQUAL note,
5552 making a new one if one does not already exist. */
5553 set_unique_reg_note (insn, REG_EQUAL, src_const);
5554 }
5555 }
5556
5557 /* Now deal with the destination. */
5558 do_not_record = 0;
5559
5560 /* Look within any SIGN_EXTRACT or ZERO_EXTRACT
5561 to the MEM or REG within it. */
5562 while (GET_CODE (dest) == SIGN_EXTRACT
5563 || GET_CODE (dest) == ZERO_EXTRACT
5564 || GET_CODE (dest) == SUBREG
5565 || GET_CODE (dest) == STRICT_LOW_PART)
5566 dest = XEXP (dest, 0);
5567
5568 sets[i].inner_dest = dest;
5569
5570 if (MEM_P (dest))
5571 {
5572 #ifdef PUSH_ROUNDING
5573 /* Stack pushes invalidate the stack pointer. */
5574 rtx addr = XEXP (dest, 0);
5575 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5576 && XEXP (addr, 0) == stack_pointer_rtx)
5577 invalidate (stack_pointer_rtx, Pmode);
5578 #endif
5579 dest = fold_rtx (dest, insn);
5580 }
5581
5582 /* Compute the hash code of the destination now,
5583 before the effects of this instruction are recorded,
5584 since the register values used in the address computation
5585 are those before this instruction. */
5586 sets[i].dest_hash = HASH (dest, mode);
5587
5588 /* Don't enter a bit-field in the hash table
5589 because the value in it after the store
5590 may not equal what was stored, due to truncation. */
5591
5592 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5593 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
5594 {
5595 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5596
5597 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5598 && GET_CODE (width) == CONST_INT
5599 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5600 && ! (INTVAL (src_const)
5601 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5602 /* Exception: if the value is constant,
5603 and it won't be truncated, record it. */
5604 ;
5605 else
5606 {
5607 /* This is chosen so that the destination will be invalidated
5608 but no new value will be recorded.
5609 We must invalidate because sometimes constant
5610 values can be recorded for bitfields. */
5611 sets[i].src_elt = 0;
5612 sets[i].src_volatile = 1;
5613 src_eqv = 0;
5614 src_eqv_elt = 0;
5615 }
5616 }
5617
5618 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5619 the insn. */
5620 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5621 {
5622 /* One less use of the label this insn used to jump to. */
5623 delete_insn (insn);
5624 cse_jumps_altered = 1;
5625 /* No more processing for this set. */
5626 sets[i].rtl = 0;
5627 }
5628
5629 /* If this SET is now setting PC to a label, we know it used to
5630 be a conditional or computed branch. */
5631 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5632 && !LABEL_REF_NONLOCAL_P (src))
5633 {
5634 /* Now emit a BARRIER after the unconditional jump. */
5635 if (NEXT_INSN (insn) == 0
5636 || !BARRIER_P (NEXT_INSN (insn)))
5637 emit_barrier_after (insn);
5638
5639 /* We reemit the jump in as many cases as possible just in
5640 case the form of an unconditional jump is significantly
5641 different than a computed jump or conditional jump.
5642
5643 If this insn has multiple sets, then reemitting the
5644 jump is nontrivial. So instead we just force rerecognition
5645 and hope for the best. */
5646 if (n_sets == 1)
5647 {
5648 rtx new, note;
5649
5650 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5651 JUMP_LABEL (new) = XEXP (src, 0);
5652 LABEL_NUSES (XEXP (src, 0))++;
5653
5654 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5655 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5656 if (note)
5657 {
5658 XEXP (note, 1) = NULL_RTX;
5659 REG_NOTES (new) = note;
5660 }
5661
5662 delete_insn (insn);
5663 insn = new;
5664
5665 /* Now emit a BARRIER after the unconditional jump. */
5666 if (NEXT_INSN (insn) == 0
5667 || !BARRIER_P (NEXT_INSN (insn)))
5668 emit_barrier_after (insn);
5669 }
5670 else
5671 INSN_CODE (insn) = -1;
5672
5673 /* Do not bother deleting any unreachable code,
5674 let jump/flow do that. */
5675
5676 cse_jumps_altered = 1;
5677 sets[i].rtl = 0;
5678 }
5679
5680 /* If destination is volatile, invalidate it and then do no further
5681 processing for this assignment. */
5682
5683 else if (do_not_record)
5684 {
5685 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5686 invalidate (dest, VOIDmode);
5687 else if (MEM_P (dest))
5688 {
5689 /* Outgoing arguments for a libcall don't
5690 affect any recorded expressions. */
5691 if (! libcall_insn || insn == libcall_insn)
5692 invalidate (dest, VOIDmode);
5693 }
5694 else if (GET_CODE (dest) == STRICT_LOW_PART
5695 || GET_CODE (dest) == ZERO_EXTRACT)
5696 invalidate (XEXP (dest, 0), GET_MODE (dest));
5697 sets[i].rtl = 0;
5698 }
5699
5700 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5701 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5702
5703 #ifdef HAVE_cc0
5704 /* If setting CC0, record what it was set to, or a constant, if it
5705 is equivalent to a constant. If it is being set to a floating-point
5706 value, make a COMPARE with the appropriate constant of 0. If we
5707 don't do this, later code can interpret this as a test against
5708 const0_rtx, which can cause problems if we try to put it into an
5709 insn as a floating-point operand. */
5710 if (dest == cc0_rtx)
5711 {
5712 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5713 this_insn_cc0_mode = mode;
5714 if (FLOAT_MODE_P (mode))
5715 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5716 CONST0_RTX (mode));
5717 }
5718 #endif
5719 }
5720
5721 /* Now enter all non-volatile source expressions in the hash table
5722 if they are not already present.
5723 Record their equivalence classes in src_elt.
5724 This way we can insert the corresponding destinations into
5725 the same classes even if the actual sources are no longer in them
5726 (having been invalidated). */
5727
5728 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5729 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5730 {
5731 struct table_elt *elt;
5732 struct table_elt *classp = sets[0].src_elt;
5733 rtx dest = SET_DEST (sets[0].rtl);
5734 enum machine_mode eqvmode = GET_MODE (dest);
5735
5736 if (GET_CODE (dest) == STRICT_LOW_PART)
5737 {
5738 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5739 classp = 0;
5740 }
5741 if (insert_regs (src_eqv, classp, 0))
5742 {
5743 rehash_using_reg (src_eqv);
5744 src_eqv_hash = HASH (src_eqv, eqvmode);
5745 }
5746 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5747 elt->in_memory = src_eqv_in_memory;
5748 src_eqv_elt = elt;
5749
5750 /* Check to see if src_eqv_elt is the same as a set source which
5751 does not yet have an elt, and if so set the elt of the set source
5752 to src_eqv_elt. */
5753 for (i = 0; i < n_sets; i++)
5754 if (sets[i].rtl && sets[i].src_elt == 0
5755 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5756 sets[i].src_elt = src_eqv_elt;
5757 }
5758
5759 for (i = 0; i < n_sets; i++)
5760 if (sets[i].rtl && ! sets[i].src_volatile
5761 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5762 {
5763 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5764 {
5765 /* REG_EQUAL in setting a STRICT_LOW_PART
5766 gives an equivalent for the entire destination register,
5767 not just for the subreg being stored in now.
5768 This is a more interesting equivalence, so we arrange later
5769 to treat the entire reg as the destination. */
5770 sets[i].src_elt = src_eqv_elt;
5771 sets[i].src_hash = src_eqv_hash;
5772 }
5773 else
5774 {
5775 /* Insert source and constant equivalent into hash table, if not
5776 already present. */
5777 struct table_elt *classp = src_eqv_elt;
5778 rtx src = sets[i].src;
5779 rtx dest = SET_DEST (sets[i].rtl);
5780 enum machine_mode mode
5781 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5782
5783 /* It's possible that we have a source value known to be
5784 constant but don't have a REG_EQUAL note on the insn.
5785 Lack of a note will mean src_eqv_elt will be NULL. This
5786 can happen where we've generated a SUBREG to access a
5787 CONST_INT that is already in a register in a wider mode.
5788 Ensure that the source expression is put in the proper
5789 constant class. */
5790 if (!classp)
5791 classp = sets[i].src_const_elt;
5792
5793 if (sets[i].src_elt == 0)
5794 {
5795 /* Don't put a hard register source into the table if this is
5796 the last insn of a libcall. In this case, we only need
5797 to put src_eqv_elt in src_elt. */
5798 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5799 {
5800 struct table_elt *elt;
5801
5802 /* Note that these insert_regs calls cannot remove
5803 any of the src_elt's, because they would have failed to
5804 match if not still valid. */
5805 if (insert_regs (src, classp, 0))
5806 {
5807 rehash_using_reg (src);
5808 sets[i].src_hash = HASH (src, mode);
5809 }
5810 elt = insert (src, classp, sets[i].src_hash, mode);
5811 elt->in_memory = sets[i].src_in_memory;
5812 sets[i].src_elt = classp = elt;
5813 }
5814 else
5815 sets[i].src_elt = classp;
5816 }
5817 if (sets[i].src_const && sets[i].src_const_elt == 0
5818 && src != sets[i].src_const
5819 && ! rtx_equal_p (sets[i].src_const, src))
5820 sets[i].src_elt = insert (sets[i].src_const, classp,
5821 sets[i].src_const_hash, mode);
5822 }
5823 }
5824 else if (sets[i].src_elt == 0)
5825 /* If we did not insert the source into the hash table (e.g., it was
5826 volatile), note the equivalence class for the REG_EQUAL value, if any,
5827 so that the destination goes into that class. */
5828 sets[i].src_elt = src_eqv_elt;
5829
5830 invalidate_from_clobbers (x);
5831
5832 /* Some registers are invalidated by subroutine calls. Memory is
5833 invalidated by non-constant calls. */
5834
5835 if (CALL_P (insn))
5836 {
5837 if (! CONST_OR_PURE_CALL_P (insn))
5838 invalidate_memory ();
5839 invalidate_for_call ();
5840 }
5841
5842 /* Now invalidate everything set by this instruction.
5843 If a SUBREG or other funny destination is being set,
5844 sets[i].rtl is still nonzero, so here we invalidate the reg
5845 a part of which is being set. */
5846
5847 for (i = 0; i < n_sets; i++)
5848 if (sets[i].rtl)
5849 {
5850 /* We can't use the inner dest, because the mode associated with
5851 a ZERO_EXTRACT is significant. */
5852 rtx dest = SET_DEST (sets[i].rtl);
5853
5854 /* Needed for registers to remove the register from its
5855 previous quantity's chain.
5856 Needed for memory if this is a nonvarying address, unless
5857 we have just done an invalidate_memory that covers even those. */
5858 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5859 invalidate (dest, VOIDmode);
5860 else if (MEM_P (dest))
5861 {
5862 /* Outgoing arguments for a libcall don't
5863 affect any recorded expressions. */
5864 if (! libcall_insn || insn == libcall_insn)
5865 invalidate (dest, VOIDmode);
5866 }
5867 else if (GET_CODE (dest) == STRICT_LOW_PART
5868 || GET_CODE (dest) == ZERO_EXTRACT)
5869 invalidate (XEXP (dest, 0), GET_MODE (dest));
5870 }
5871
5872 /* A volatile ASM invalidates everything. */
5873 if (NONJUMP_INSN_P (insn)
5874 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5875 && MEM_VOLATILE_P (PATTERN (insn)))
5876 flush_hash_table ();
5877
5878 /* Make sure registers mentioned in destinations
5879 are safe for use in an expression to be inserted.
5880 This removes from the hash table
5881 any invalid entry that refers to one of these registers.
5882
5883 We don't care about the return value from mention_regs because
5884 we are going to hash the SET_DEST values unconditionally. */
5885
5886 for (i = 0; i < n_sets; i++)
5887 {
5888 if (sets[i].rtl)
5889 {
5890 rtx x = SET_DEST (sets[i].rtl);
5891
5892 if (!REG_P (x))
5893 mention_regs (x);
5894 else
5895 {
5896 /* We used to rely on all references to a register becoming
5897 inaccessible when a register changes to a new quantity,
5898 since that changes the hash code. However, that is not
5899 safe, since after HASH_SIZE new quantities we get a
5900 hash 'collision' of a register with its own invalid
5901 entries. And since SUBREGs have been changed not to
5902 change their hash code with the hash code of the register,
5903 it wouldn't work any longer at all. So we have to check
5904 for any invalid references lying around now.
5905 This code is similar to the REG case in mention_regs,
5906 but it knows that reg_tick has been incremented, and
5907 it leaves reg_in_table as -1 . */
5908 unsigned int regno = REGNO (x);
5909 unsigned int endregno
5910 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
5911 : hard_regno_nregs[regno][GET_MODE (x)]);
5912 unsigned int i;
5913
5914 for (i = regno; i < endregno; i++)
5915 {
5916 if (REG_IN_TABLE (i) >= 0)
5917 {
5918 remove_invalid_refs (i);
5919 REG_IN_TABLE (i) = -1;
5920 }
5921 }
5922 }
5923 }
5924 }
5925
5926 /* We may have just removed some of the src_elt's from the hash table.
5927 So replace each one with the current head of the same class. */
5928
5929 for (i = 0; i < n_sets; i++)
5930 if (sets[i].rtl)
5931 {
5932 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5933 /* If elt was removed, find current head of same class,
5934 or 0 if nothing remains of that class. */
5935 {
5936 struct table_elt *elt = sets[i].src_elt;
5937
5938 while (elt && elt->prev_same_value)
5939 elt = elt->prev_same_value;
5940
5941 while (elt && elt->first_same_value == 0)
5942 elt = elt->next_same_value;
5943 sets[i].src_elt = elt ? elt->first_same_value : 0;
5944 }
5945 }
5946
5947 /* Now insert the destinations into their equivalence classes. */
5948
5949 for (i = 0; i < n_sets; i++)
5950 if (sets[i].rtl)
5951 {
5952 rtx dest = SET_DEST (sets[i].rtl);
5953 struct table_elt *elt;
5954
5955 /* Don't record value if we are not supposed to risk allocating
5956 floating-point values in registers that might be wider than
5957 memory. */
5958 if ((flag_float_store
5959 && MEM_P (dest)
5960 && FLOAT_MODE_P (GET_MODE (dest)))
5961 /* Don't record BLKmode values, because we don't know the
5962 size of it, and can't be sure that other BLKmode values
5963 have the same or smaller size. */
5964 || GET_MODE (dest) == BLKmode
5965 /* Don't record values of destinations set inside a libcall block
5966 since we might delete the libcall. Things should have been set
5967 up so we won't want to reuse such a value, but we play it safe
5968 here. */
5969 || libcall_insn
5970 /* If we didn't put a REG_EQUAL value or a source into the hash
5971 table, there is no point is recording DEST. */
5972 || sets[i].src_elt == 0
5973 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5974 or SIGN_EXTEND, don't record DEST since it can cause
5975 some tracking to be wrong.
5976
5977 ??? Think about this more later. */
5978 || (GET_CODE (dest) == SUBREG
5979 && (GET_MODE_SIZE (GET_MODE (dest))
5980 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5981 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5982 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5983 continue;
5984
5985 /* STRICT_LOW_PART isn't part of the value BEING set,
5986 and neither is the SUBREG inside it.
5987 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5988 if (GET_CODE (dest) == STRICT_LOW_PART)
5989 dest = SUBREG_REG (XEXP (dest, 0));
5990
5991 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5992 /* Registers must also be inserted into chains for quantities. */
5993 if (insert_regs (dest, sets[i].src_elt, 1))
5994 {
5995 /* If `insert_regs' changes something, the hash code must be
5996 recalculated. */
5997 rehash_using_reg (dest);
5998 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5999 }
6000
6001 elt = insert (dest, sets[i].src_elt,
6002 sets[i].dest_hash, GET_MODE (dest));
6003
6004 elt->in_memory = (MEM_P (sets[i].inner_dest)
6005 && !MEM_READONLY_P (sets[i].inner_dest));
6006
6007 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6008 narrower than M2, and both M1 and M2 are the same number of words,
6009 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6010 make that equivalence as well.
6011
6012 However, BAR may have equivalences for which gen_lowpart
6013 will produce a simpler value than gen_lowpart applied to
6014 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6015 BAR's equivalences. If we don't get a simplified form, make
6016 the SUBREG. It will not be used in an equivalence, but will
6017 cause two similar assignments to be detected.
6018
6019 Note the loop below will find SUBREG_REG (DEST) since we have
6020 already entered SRC and DEST of the SET in the table. */
6021
6022 if (GET_CODE (dest) == SUBREG
6023 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6024 / UNITS_PER_WORD)
6025 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6026 && (GET_MODE_SIZE (GET_MODE (dest))
6027 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6028 && sets[i].src_elt != 0)
6029 {
6030 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6031 struct table_elt *elt, *classp = 0;
6032
6033 for (elt = sets[i].src_elt->first_same_value; elt;
6034 elt = elt->next_same_value)
6035 {
6036 rtx new_src = 0;
6037 unsigned src_hash;
6038 struct table_elt *src_elt;
6039 int byte = 0;
6040
6041 /* Ignore invalid entries. */
6042 if (!REG_P (elt->exp)
6043 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6044 continue;
6045
6046 /* We may have already been playing subreg games. If the
6047 mode is already correct for the destination, use it. */
6048 if (GET_MODE (elt->exp) == new_mode)
6049 new_src = elt->exp;
6050 else
6051 {
6052 /* Calculate big endian correction for the SUBREG_BYTE.
6053 We have already checked that M1 (GET_MODE (dest))
6054 is not narrower than M2 (new_mode). */
6055 if (BYTES_BIG_ENDIAN)
6056 byte = (GET_MODE_SIZE (GET_MODE (dest))
6057 - GET_MODE_SIZE (new_mode));
6058
6059 new_src = simplify_gen_subreg (new_mode, elt->exp,
6060 GET_MODE (dest), byte);
6061 }
6062
6063 /* The call to simplify_gen_subreg fails if the value
6064 is VOIDmode, yet we can't do any simplification, e.g.
6065 for EXPR_LISTs denoting function call results.
6066 It is invalid to construct a SUBREG with a VOIDmode
6067 SUBREG_REG, hence a zero new_src means we can't do
6068 this substitution. */
6069 if (! new_src)
6070 continue;
6071
6072 src_hash = HASH (new_src, new_mode);
6073 src_elt = lookup (new_src, src_hash, new_mode);
6074
6075 /* Put the new source in the hash table is if isn't
6076 already. */
6077 if (src_elt == 0)
6078 {
6079 if (insert_regs (new_src, classp, 0))
6080 {
6081 rehash_using_reg (new_src);
6082 src_hash = HASH (new_src, new_mode);
6083 }
6084 src_elt = insert (new_src, classp, src_hash, new_mode);
6085 src_elt->in_memory = elt->in_memory;
6086 }
6087 else if (classp && classp != src_elt->first_same_value)
6088 /* Show that two things that we've seen before are
6089 actually the same. */
6090 merge_equiv_classes (src_elt, classp);
6091
6092 classp = src_elt->first_same_value;
6093 /* Ignore invalid entries. */
6094 while (classp
6095 && !REG_P (classp->exp)
6096 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6097 classp = classp->next_same_value;
6098 }
6099 }
6100 }
6101
6102 /* Special handling for (set REG0 REG1) where REG0 is the
6103 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6104 be used in the sequel, so (if easily done) change this insn to
6105 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6106 that computed their value. Then REG1 will become a dead store
6107 and won't cloud the situation for later optimizations.
6108
6109 Do not make this change if REG1 is a hard register, because it will
6110 then be used in the sequel and we may be changing a two-operand insn
6111 into a three-operand insn.
6112
6113 Also do not do this if we are operating on a copy of INSN.
6114
6115 Also don't do this if INSN ends a libcall; this would cause an unrelated
6116 register to be set in the middle of a libcall, and we then get bad code
6117 if the libcall is deleted. */
6118
6119 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6120 && NEXT_INSN (PREV_INSN (insn)) == insn
6121 && REG_P (SET_SRC (sets[0].rtl))
6122 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6123 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6124 {
6125 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6126 struct qty_table_elem *src_ent = &qty_table[src_q];
6127
6128 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6129 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6130 {
6131 rtx prev = insn;
6132 /* Scan for the previous nonnote insn, but stop at a basic
6133 block boundary. */
6134 do
6135 {
6136 prev = PREV_INSN (prev);
6137 }
6138 while (prev && NOTE_P (prev)
6139 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6140
6141 /* Do not swap the registers around if the previous instruction
6142 attaches a REG_EQUIV note to REG1.
6143
6144 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6145 from the pseudo that originally shadowed an incoming argument
6146 to another register. Some uses of REG_EQUIV might rely on it
6147 being attached to REG1 rather than REG2.
6148
6149 This section previously turned the REG_EQUIV into a REG_EQUAL
6150 note. We cannot do that because REG_EQUIV may provide an
6151 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6152
6153 if (prev != 0 && NONJUMP_INSN_P (prev)
6154 && GET_CODE (PATTERN (prev)) == SET
6155 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6156 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6157 {
6158 rtx dest = SET_DEST (sets[0].rtl);
6159 rtx src = SET_SRC (sets[0].rtl);
6160 rtx note;
6161
6162 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6163 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6164 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6165 apply_change_group ();
6166
6167 /* If INSN has a REG_EQUAL note, and this note mentions
6168 REG0, then we must delete it, because the value in
6169 REG0 has changed. If the note's value is REG1, we must
6170 also delete it because that is now this insn's dest. */
6171 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6172 if (note != 0
6173 && (reg_mentioned_p (dest, XEXP (note, 0))
6174 || rtx_equal_p (src, XEXP (note, 0))))
6175 remove_note (insn, note);
6176 }
6177 }
6178 }
6179
6180 /* If this is a conditional jump insn, record any known equivalences due to
6181 the condition being tested. */
6182
6183 if (JUMP_P (insn)
6184 && n_sets == 1 && GET_CODE (x) == SET
6185 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6186 record_jump_equiv (insn, 0);
6187
6188 #ifdef HAVE_cc0
6189 /* If the previous insn set CC0 and this insn no longer references CC0,
6190 delete the previous insn. Here we use the fact that nothing expects CC0
6191 to be valid over an insn, which is true until the final pass. */
6192 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6193 && (tem = single_set (prev_insn)) != 0
6194 && SET_DEST (tem) == cc0_rtx
6195 && ! reg_mentioned_p (cc0_rtx, x))
6196 delete_insn (prev_insn);
6197
6198 prev_insn_cc0 = this_insn_cc0;
6199 prev_insn_cc0_mode = this_insn_cc0_mode;
6200 prev_insn = insn;
6201 #endif
6202 }
6203 \f
6204 /* Remove from the hash table all expressions that reference memory. */
6205
6206 static void
6207 invalidate_memory (void)
6208 {
6209 int i;
6210 struct table_elt *p, *next;
6211
6212 for (i = 0; i < HASH_SIZE; i++)
6213 for (p = table[i]; p; p = next)
6214 {
6215 next = p->next_same_hash;
6216 if (p->in_memory)
6217 remove_from_table (p, i);
6218 }
6219 }
6220
6221 /* If ADDR is an address that implicitly affects the stack pointer, return
6222 1 and update the register tables to show the effect. Else, return 0. */
6223
6224 static int
6225 addr_affects_sp_p (rtx addr)
6226 {
6227 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6228 && REG_P (XEXP (addr, 0))
6229 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6230 {
6231 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6232 {
6233 REG_TICK (STACK_POINTER_REGNUM)++;
6234 /* Is it possible to use a subreg of SP? */
6235 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6236 }
6237
6238 /* This should be *very* rare. */
6239 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6240 invalidate (stack_pointer_rtx, VOIDmode);
6241
6242 return 1;
6243 }
6244
6245 return 0;
6246 }
6247
6248 /* Perform invalidation on the basis of everything about an insn
6249 except for invalidating the actual places that are SET in it.
6250 This includes the places CLOBBERed, and anything that might
6251 alias with something that is SET or CLOBBERed.
6252
6253 X is the pattern of the insn. */
6254
6255 static void
6256 invalidate_from_clobbers (rtx x)
6257 {
6258 if (GET_CODE (x) == CLOBBER)
6259 {
6260 rtx ref = XEXP (x, 0);
6261 if (ref)
6262 {
6263 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6264 || MEM_P (ref))
6265 invalidate (ref, VOIDmode);
6266 else if (GET_CODE (ref) == STRICT_LOW_PART
6267 || GET_CODE (ref) == ZERO_EXTRACT)
6268 invalidate (XEXP (ref, 0), GET_MODE (ref));
6269 }
6270 }
6271 else if (GET_CODE (x) == PARALLEL)
6272 {
6273 int i;
6274 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6275 {
6276 rtx y = XVECEXP (x, 0, i);
6277 if (GET_CODE (y) == CLOBBER)
6278 {
6279 rtx ref = XEXP (y, 0);
6280 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6281 || MEM_P (ref))
6282 invalidate (ref, VOIDmode);
6283 else if (GET_CODE (ref) == STRICT_LOW_PART
6284 || GET_CODE (ref) == ZERO_EXTRACT)
6285 invalidate (XEXP (ref, 0), GET_MODE (ref));
6286 }
6287 }
6288 }
6289 }
6290 \f
6291 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6292 and replace any registers in them with either an equivalent constant
6293 or the canonical form of the register. If we are inside an address,
6294 only do this if the address remains valid.
6295
6296 OBJECT is 0 except when within a MEM in which case it is the MEM.
6297
6298 Return the replacement for X. */
6299
6300 static rtx
6301 cse_process_notes (rtx x, rtx object)
6302 {
6303 enum rtx_code code = GET_CODE (x);
6304 const char *fmt = GET_RTX_FORMAT (code);
6305 int i;
6306
6307 switch (code)
6308 {
6309 case CONST_INT:
6310 case CONST:
6311 case SYMBOL_REF:
6312 case LABEL_REF:
6313 case CONST_DOUBLE:
6314 case CONST_VECTOR:
6315 case PC:
6316 case CC0:
6317 case LO_SUM:
6318 return x;
6319
6320 case MEM:
6321 validate_change (x, &XEXP (x, 0),
6322 cse_process_notes (XEXP (x, 0), x), 0);
6323 return x;
6324
6325 case EXPR_LIST:
6326 case INSN_LIST:
6327 if (REG_NOTE_KIND (x) == REG_EQUAL)
6328 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6329 if (XEXP (x, 1))
6330 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6331 return x;
6332
6333 case SIGN_EXTEND:
6334 case ZERO_EXTEND:
6335 case SUBREG:
6336 {
6337 rtx new = cse_process_notes (XEXP (x, 0), object);
6338 /* We don't substitute VOIDmode constants into these rtx,
6339 since they would impede folding. */
6340 if (GET_MODE (new) != VOIDmode)
6341 validate_change (object, &XEXP (x, 0), new, 0);
6342 return x;
6343 }
6344
6345 case REG:
6346 i = REG_QTY (REGNO (x));
6347
6348 /* Return a constant or a constant register. */
6349 if (REGNO_QTY_VALID_P (REGNO (x)))
6350 {
6351 struct qty_table_elem *ent = &qty_table[i];
6352
6353 if (ent->const_rtx != NULL_RTX
6354 && (CONSTANT_P (ent->const_rtx)
6355 || REG_P (ent->const_rtx)))
6356 {
6357 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6358 if (new)
6359 return new;
6360 }
6361 }
6362
6363 /* Otherwise, canonicalize this register. */
6364 return canon_reg (x, NULL_RTX);
6365
6366 default:
6367 break;
6368 }
6369
6370 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6371 if (fmt[i] == 'e')
6372 validate_change (object, &XEXP (x, i),
6373 cse_process_notes (XEXP (x, i), object), 0);
6374
6375 return x;
6376 }
6377 \f
6378 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6379 since they are done elsewhere. This function is called via note_stores. */
6380
6381 static void
6382 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6383 {
6384 enum rtx_code code = GET_CODE (dest);
6385
6386 if (code == MEM
6387 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6388 /* There are times when an address can appear varying and be a PLUS
6389 during this scan when it would be a fixed address were we to know
6390 the proper equivalences. So invalidate all memory if there is
6391 a BLKmode or nonscalar memory reference or a reference to a
6392 variable address. */
6393 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6394 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6395 {
6396 invalidate_memory ();
6397 return;
6398 }
6399
6400 if (GET_CODE (set) == CLOBBER
6401 || CC0_P (dest)
6402 || dest == pc_rtx)
6403 return;
6404
6405 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6406 invalidate (XEXP (dest, 0), GET_MODE (dest));
6407 else if (code == REG || code == SUBREG || code == MEM)
6408 invalidate (dest, VOIDmode);
6409 }
6410
6411 /* Invalidate all insns from START up to the end of the function or the
6412 next label. This called when we wish to CSE around a block that is
6413 conditionally executed. */
6414
6415 static void
6416 invalidate_skipped_block (rtx start)
6417 {
6418 rtx insn;
6419
6420 for (insn = start; insn && !LABEL_P (insn);
6421 insn = NEXT_INSN (insn))
6422 {
6423 if (! INSN_P (insn))
6424 continue;
6425
6426 if (CALL_P (insn))
6427 {
6428 if (! CONST_OR_PURE_CALL_P (insn))
6429 invalidate_memory ();
6430 invalidate_for_call ();
6431 }
6432
6433 invalidate_from_clobbers (PATTERN (insn));
6434 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6435 }
6436 }
6437 \f
6438 /* Find the end of INSN's basic block and return its range,
6439 the total number of SETs in all the insns of the block, the last insn of the
6440 block, and the branch path.
6441
6442 The branch path indicates which branches should be followed. If a nonzero
6443 path size is specified, the block should be rescanned and a different set
6444 of branches will be taken. The branch path is only used if
6445 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6446
6447 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6448 used to describe the block. It is filled in with the information about
6449 the current block. The incoming structure's branch path, if any, is used
6450 to construct the output branch path. */
6451
6452 static void
6453 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6454 int follow_jumps, int skip_blocks)
6455 {
6456 rtx p = insn, q;
6457 int nsets = 0;
6458 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6459 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6460 int path_size = data->path_size;
6461 int path_entry = 0;
6462 int i;
6463
6464 /* Update the previous branch path, if any. If the last branch was
6465 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6466 If it was previously PATH_NOT_TAKEN,
6467 shorten the path by one and look at the previous branch. We know that
6468 at least one branch must have been taken if PATH_SIZE is nonzero. */
6469 while (path_size > 0)
6470 {
6471 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6472 {
6473 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6474 break;
6475 }
6476 else
6477 path_size--;
6478 }
6479
6480 /* If the first instruction is marked with QImode, that means we've
6481 already processed this block. Our caller will look at DATA->LAST
6482 to figure out where to go next. We want to return the next block
6483 in the instruction stream, not some branched-to block somewhere
6484 else. We accomplish this by pretending our called forbid us to
6485 follow jumps, or skip blocks. */
6486 if (GET_MODE (insn) == QImode)
6487 follow_jumps = skip_blocks = 0;
6488
6489 /* Scan to end of this basic block. */
6490 while (p && !LABEL_P (p))
6491 {
6492 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6493 the regs restored by the longjmp come from
6494 a later time than the setjmp. */
6495 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6496 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6497 break;
6498
6499 /* A PARALLEL can have lots of SETs in it,
6500 especially if it is really an ASM_OPERANDS. */
6501 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6502 nsets += XVECLEN (PATTERN (p), 0);
6503 else if (!NOTE_P (p))
6504 nsets += 1;
6505
6506 /* Ignore insns made by CSE; they cannot affect the boundaries of
6507 the basic block. */
6508
6509 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6510 high_cuid = INSN_CUID (p);
6511 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6512 low_cuid = INSN_CUID (p);
6513
6514 /* See if this insn is in our branch path. If it is and we are to
6515 take it, do so. */
6516 if (path_entry < path_size && data->path[path_entry].branch == p)
6517 {
6518 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6519 p = JUMP_LABEL (p);
6520
6521 /* Point to next entry in path, if any. */
6522 path_entry++;
6523 }
6524
6525 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6526 was specified, we haven't reached our maximum path length, there are
6527 insns following the target of the jump, this is the only use of the
6528 jump label, and the target label is preceded by a BARRIER.
6529
6530 Alternatively, we can follow the jump if it branches around a
6531 block of code and there are no other branches into the block.
6532 In this case invalidate_skipped_block will be called to invalidate any
6533 registers set in the block when following the jump. */
6534
6535 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6536 && JUMP_P (p)
6537 && GET_CODE (PATTERN (p)) == SET
6538 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6539 && JUMP_LABEL (p) != 0
6540 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6541 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6542 {
6543 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6544 if ((!NOTE_P (q)
6545 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6546 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6547 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6548 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6549 break;
6550
6551 /* If we ran into a BARRIER, this code is an extension of the
6552 basic block when the branch is taken. */
6553 if (follow_jumps && q != 0 && BARRIER_P (q))
6554 {
6555 /* Don't allow ourself to keep walking around an
6556 always-executed loop. */
6557 if (next_real_insn (q) == next)
6558 {
6559 p = NEXT_INSN (p);
6560 continue;
6561 }
6562
6563 /* Similarly, don't put a branch in our path more than once. */
6564 for (i = 0; i < path_entry; i++)
6565 if (data->path[i].branch == p)
6566 break;
6567
6568 if (i != path_entry)
6569 break;
6570
6571 data->path[path_entry].branch = p;
6572 data->path[path_entry++].status = PATH_TAKEN;
6573
6574 /* This branch now ends our path. It was possible that we
6575 didn't see this branch the last time around (when the
6576 insn in front of the target was a JUMP_INSN that was
6577 turned into a no-op). */
6578 path_size = path_entry;
6579
6580 p = JUMP_LABEL (p);
6581 /* Mark block so we won't scan it again later. */
6582 PUT_MODE (NEXT_INSN (p), QImode);
6583 }
6584 /* Detect a branch around a block of code. */
6585 else if (skip_blocks && q != 0 && !LABEL_P (q))
6586 {
6587 rtx tmp;
6588
6589 if (next_real_insn (q) == next)
6590 {
6591 p = NEXT_INSN (p);
6592 continue;
6593 }
6594
6595 for (i = 0; i < path_entry; i++)
6596 if (data->path[i].branch == p)
6597 break;
6598
6599 if (i != path_entry)
6600 break;
6601
6602 /* This is no_labels_between_p (p, q) with an added check for
6603 reaching the end of a function (in case Q precedes P). */
6604 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6605 if (LABEL_P (tmp))
6606 break;
6607
6608 if (tmp == q)
6609 {
6610 data->path[path_entry].branch = p;
6611 data->path[path_entry++].status = PATH_AROUND;
6612
6613 path_size = path_entry;
6614
6615 p = JUMP_LABEL (p);
6616 /* Mark block so we won't scan it again later. */
6617 PUT_MODE (NEXT_INSN (p), QImode);
6618 }
6619 }
6620 }
6621 p = NEXT_INSN (p);
6622 }
6623
6624 data->low_cuid = low_cuid;
6625 data->high_cuid = high_cuid;
6626 data->nsets = nsets;
6627 data->last = p;
6628
6629 /* If all jumps in the path are not taken, set our path length to zero
6630 so a rescan won't be done. */
6631 for (i = path_size - 1; i >= 0; i--)
6632 if (data->path[i].status != PATH_NOT_TAKEN)
6633 break;
6634
6635 if (i == -1)
6636 data->path_size = 0;
6637 else
6638 data->path_size = path_size;
6639
6640 /* End the current branch path. */
6641 data->path[path_size].branch = 0;
6642 }
6643 \f
6644 /* Perform cse on the instructions of a function.
6645 F is the first instruction.
6646 NREGS is one plus the highest pseudo-reg number used in the instruction.
6647
6648 Returns 1 if jump_optimize should be redone due to simplifications
6649 in conditional jump instructions. */
6650
6651 int
6652 cse_main (rtx f, int nregs, FILE *file)
6653 {
6654 struct cse_basic_block_data val;
6655 rtx insn = f;
6656 int i;
6657
6658 val.path = xmalloc (sizeof (struct branch_path)
6659 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6660
6661 cse_jumps_altered = 0;
6662 recorded_label_ref = 0;
6663 constant_pool_entries_cost = 0;
6664 constant_pool_entries_regcost = 0;
6665 val.path_size = 0;
6666 rtl_hooks = cse_rtl_hooks;
6667
6668 init_recog ();
6669 init_alias_analysis ();
6670
6671 max_reg = nregs;
6672
6673 max_insn_uid = get_max_uid ();
6674
6675 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6676
6677 #ifdef LOAD_EXTEND_OP
6678
6679 /* Allocate scratch rtl here. cse_insn will fill in the memory reference
6680 and change the code and mode as appropriate. */
6681 memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX);
6682 #endif
6683
6684 /* Reset the counter indicating how many elements have been made
6685 thus far. */
6686 n_elements_made = 0;
6687
6688 /* Find the largest uid. */
6689
6690 max_uid = get_max_uid ();
6691 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6692
6693 /* Compute the mapping from uids to cuids.
6694 CUIDs are numbers assigned to insns, like uids,
6695 except that cuids increase monotonically through the code.
6696 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6697 between two insns is not affected by -g. */
6698
6699 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6700 {
6701 if (!NOTE_P (insn)
6702 || NOTE_LINE_NUMBER (insn) < 0)
6703 INSN_CUID (insn) = ++i;
6704 else
6705 /* Give a line number note the same cuid as preceding insn. */
6706 INSN_CUID (insn) = i;
6707 }
6708
6709 ggc_push_context ();
6710
6711 /* Loop over basic blocks.
6712 Compute the maximum number of qty's needed for each basic block
6713 (which is 2 for each SET). */
6714 insn = f;
6715 while (insn)
6716 {
6717 cse_altered = 0;
6718 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6719 flag_cse_skip_blocks);
6720
6721 /* If this basic block was already processed or has no sets, skip it. */
6722 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6723 {
6724 PUT_MODE (insn, VOIDmode);
6725 insn = (val.last ? NEXT_INSN (val.last) : 0);
6726 val.path_size = 0;
6727 continue;
6728 }
6729
6730 cse_basic_block_start = val.low_cuid;
6731 cse_basic_block_end = val.high_cuid;
6732 max_qty = val.nsets * 2;
6733
6734 if (file)
6735 fnotice (file, ";; Processing block from %d to %d, %d sets.\n",
6736 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6737 val.nsets);
6738
6739 /* Make MAX_QTY bigger to give us room to optimize
6740 past the end of this basic block, if that should prove useful. */
6741 if (max_qty < 500)
6742 max_qty = 500;
6743
6744 max_qty += max_reg;
6745
6746 /* If this basic block is being extended by following certain jumps,
6747 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6748 Otherwise, we start after this basic block. */
6749 if (val.path_size > 0)
6750 cse_basic_block (insn, val.last, val.path);
6751 else
6752 {
6753 int old_cse_jumps_altered = cse_jumps_altered;
6754 rtx temp;
6755
6756 /* When cse changes a conditional jump to an unconditional
6757 jump, we want to reprocess the block, since it will give
6758 us a new branch path to investigate. */
6759 cse_jumps_altered = 0;
6760 temp = cse_basic_block (insn, val.last, val.path);
6761 if (cse_jumps_altered == 0
6762 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6763 insn = temp;
6764
6765 cse_jumps_altered |= old_cse_jumps_altered;
6766 }
6767
6768 if (cse_altered)
6769 ggc_collect ();
6770
6771 #ifdef USE_C_ALLOCA
6772 alloca (0);
6773 #endif
6774 }
6775
6776 ggc_pop_context ();
6777
6778 if (max_elements_made < n_elements_made)
6779 max_elements_made = n_elements_made;
6780
6781 /* Clean up. */
6782 end_alias_analysis ();
6783 free (uid_cuid);
6784 free (reg_eqv_table);
6785 free (val.path);
6786 rtl_hooks = general_rtl_hooks;
6787
6788 return cse_jumps_altered || recorded_label_ref;
6789 }
6790
6791 /* Process a single basic block. FROM and TO and the limits of the basic
6792 block. NEXT_BRANCH points to the branch path when following jumps or
6793 a null path when not following jumps.
6794
6795 AROUND_LOOP is nonzero if we are to try to cse around to the start of a
6796 loop. This is true when we are being called for the last time on a
6797 block and this CSE pass is before loop.c. */
6798
6799 static rtx
6800 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
6801 {
6802 rtx insn;
6803 int to_usage = 0;
6804 rtx libcall_insn = NULL_RTX;
6805 int num_insns = 0;
6806 int no_conflict = 0;
6807
6808 /* This array is undefined before max_reg, so only allocate
6809 the space actually needed and adjust the start. */
6810
6811 qty_table = xmalloc ((max_qty - max_reg) * sizeof (struct qty_table_elem));
6812 qty_table -= max_reg;
6813
6814 new_basic_block ();
6815
6816 /* TO might be a label. If so, protect it from being deleted. */
6817 if (to != 0 && LABEL_P (to))
6818 ++LABEL_NUSES (to);
6819
6820 for (insn = from; insn != to; insn = NEXT_INSN (insn))
6821 {
6822 enum rtx_code code = GET_CODE (insn);
6823
6824 /* If we have processed 1,000 insns, flush the hash table to
6825 avoid extreme quadratic behavior. We must not include NOTEs
6826 in the count since there may be more of them when generating
6827 debugging information. If we clear the table at different
6828 times, code generated with -g -O might be different than code
6829 generated with -O but not -g.
6830
6831 ??? This is a real kludge and needs to be done some other way.
6832 Perhaps for 2.9. */
6833 if (code != NOTE && num_insns++ > 1000)
6834 {
6835 flush_hash_table ();
6836 num_insns = 0;
6837 }
6838
6839 /* See if this is a branch that is part of the path. If so, and it is
6840 to be taken, do so. */
6841 if (next_branch->branch == insn)
6842 {
6843 enum taken status = next_branch++->status;
6844 if (status != PATH_NOT_TAKEN)
6845 {
6846 if (status == PATH_TAKEN)
6847 record_jump_equiv (insn, 1);
6848 else
6849 invalidate_skipped_block (NEXT_INSN (insn));
6850
6851 /* Set the last insn as the jump insn; it doesn't affect cc0.
6852 Then follow this branch. */
6853 #ifdef HAVE_cc0
6854 prev_insn_cc0 = 0;
6855 prev_insn = insn;
6856 #endif
6857 insn = JUMP_LABEL (insn);
6858 continue;
6859 }
6860 }
6861
6862 if (GET_MODE (insn) == QImode)
6863 PUT_MODE (insn, VOIDmode);
6864
6865 if (GET_RTX_CLASS (code) == RTX_INSN)
6866 {
6867 rtx p;
6868
6869 /* Process notes first so we have all notes in canonical forms when
6870 looking for duplicate operations. */
6871
6872 if (REG_NOTES (insn))
6873 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
6874
6875 /* Track when we are inside in LIBCALL block. Inside such a block,
6876 we do not want to record destinations. The last insn of a
6877 LIBCALL block is not considered to be part of the block, since
6878 its destination is the result of the block and hence should be
6879 recorded. */
6880
6881 if (REG_NOTES (insn) != 0)
6882 {
6883 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
6884 libcall_insn = XEXP (p, 0);
6885 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
6886 {
6887 /* Keep libcall_insn for the last SET insn of a no-conflict
6888 block to prevent changing the destination. */
6889 if (! no_conflict)
6890 libcall_insn = 0;
6891 else
6892 no_conflict = -1;
6893 }
6894 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
6895 no_conflict = 1;
6896 }
6897
6898 cse_insn (insn, libcall_insn);
6899
6900 if (no_conflict == -1)
6901 {
6902 libcall_insn = 0;
6903 no_conflict = 0;
6904 }
6905
6906 /* If we haven't already found an insn where we added a LABEL_REF,
6907 check this one. */
6908 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
6909 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6910 (void *) insn))
6911 recorded_label_ref = 1;
6912 }
6913
6914 /* If INSN is now an unconditional jump, skip to the end of our
6915 basic block by pretending that we just did the last insn in the
6916 basic block. If we are jumping to the end of our block, show
6917 that we can have one usage of TO. */
6918
6919 if (any_uncondjump_p (insn))
6920 {
6921 if (to == 0)
6922 {
6923 free (qty_table + max_reg);
6924 return 0;
6925 }
6926
6927 if (JUMP_LABEL (insn) == to)
6928 to_usage = 1;
6929
6930 /* Maybe TO was deleted because the jump is unconditional.
6931 If so, there is nothing left in this basic block. */
6932 /* ??? Perhaps it would be smarter to set TO
6933 to whatever follows this insn,
6934 and pretend the basic block had always ended here. */
6935 if (INSN_DELETED_P (to))
6936 break;
6937
6938 insn = PREV_INSN (to);
6939 }
6940
6941 /* See if it is ok to keep on going past the label
6942 which used to end our basic block. Remember that we incremented
6943 the count of that label, so we decrement it here. If we made
6944 a jump unconditional, TO_USAGE will be one; in that case, we don't
6945 want to count the use in that jump. */
6946
6947 if (to != 0 && NEXT_INSN (insn) == to
6948 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
6949 {
6950 struct cse_basic_block_data val;
6951 rtx prev;
6952
6953 insn = NEXT_INSN (to);
6954
6955 /* If TO was the last insn in the function, we are done. */
6956 if (insn == 0)
6957 {
6958 free (qty_table + max_reg);
6959 return 0;
6960 }
6961
6962 /* If TO was preceded by a BARRIER we are done with this block
6963 because it has no continuation. */
6964 prev = prev_nonnote_insn (to);
6965 if (prev && BARRIER_P (prev))
6966 {
6967 free (qty_table + max_reg);
6968 return insn;
6969 }
6970
6971 /* Find the end of the following block. Note that we won't be
6972 following branches in this case. */
6973 to_usage = 0;
6974 val.path_size = 0;
6975 val.path = xmalloc (sizeof (struct branch_path)
6976 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6977 cse_end_of_basic_block (insn, &val, 0, 0);
6978 free (val.path);
6979
6980 /* If the tables we allocated have enough space left
6981 to handle all the SETs in the next basic block,
6982 continue through it. Otherwise, return,
6983 and that block will be scanned individually. */
6984 if (val.nsets * 2 + next_qty > max_qty)
6985 break;
6986
6987 cse_basic_block_start = val.low_cuid;
6988 cse_basic_block_end = val.high_cuid;
6989 to = val.last;
6990
6991 /* Prevent TO from being deleted if it is a label. */
6992 if (to != 0 && LABEL_P (to))
6993 ++LABEL_NUSES (to);
6994
6995 /* Back up so we process the first insn in the extension. */
6996 insn = PREV_INSN (insn);
6997 }
6998 }
6999
7000 gcc_assert (next_qty <= max_qty);
7001
7002 free (qty_table + max_reg);
7003
7004 return to ? NEXT_INSN (to) : 0;
7005 }
7006 \f
7007 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7008 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7009
7010 static int
7011 check_for_label_ref (rtx *rtl, void *data)
7012 {
7013 rtx insn = (rtx) data;
7014
7015 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7016 we must rerun jump since it needs to place the note. If this is a
7017 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7018 since no REG_LABEL will be added. */
7019 return (GET_CODE (*rtl) == LABEL_REF
7020 && ! LABEL_REF_NONLOCAL_P (*rtl)
7021 && LABEL_P (XEXP (*rtl, 0))
7022 && INSN_UID (XEXP (*rtl, 0)) != 0
7023 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7024 }
7025 \f
7026 /* Count the number of times registers are used (not set) in X.
7027 COUNTS is an array in which we accumulate the count, INCR is how much
7028 we count each register usage. */
7029
7030 static void
7031 count_reg_usage (rtx x, int *counts, int incr)
7032 {
7033 enum rtx_code code;
7034 rtx note;
7035 const char *fmt;
7036 int i, j;
7037
7038 if (x == 0)
7039 return;
7040
7041 switch (code = GET_CODE (x))
7042 {
7043 case REG:
7044 counts[REGNO (x)] += incr;
7045 return;
7046
7047 case PC:
7048 case CC0:
7049 case CONST:
7050 case CONST_INT:
7051 case CONST_DOUBLE:
7052 case CONST_VECTOR:
7053 case SYMBOL_REF:
7054 case LABEL_REF:
7055 return;
7056
7057 case CLOBBER:
7058 /* If we are clobbering a MEM, mark any registers inside the address
7059 as being used. */
7060 if (MEM_P (XEXP (x, 0)))
7061 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, incr);
7062 return;
7063
7064 case SET:
7065 /* Unless we are setting a REG, count everything in SET_DEST. */
7066 if (!REG_P (SET_DEST (x)))
7067 count_reg_usage (SET_DEST (x), counts, incr);
7068 count_reg_usage (SET_SRC (x), counts, incr);
7069 return;
7070
7071 case CALL_INSN:
7072 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, incr);
7073 /* Fall through. */
7074
7075 case INSN:
7076 case JUMP_INSN:
7077 count_reg_usage (PATTERN (x), counts, incr);
7078
7079 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7080 use them. */
7081
7082 note = find_reg_equal_equiv_note (x);
7083 if (note)
7084 {
7085 rtx eqv = XEXP (note, 0);
7086
7087 if (GET_CODE (eqv) == EXPR_LIST)
7088 /* This REG_EQUAL note describes the result of a function call.
7089 Process all the arguments. */
7090 do
7091 {
7092 count_reg_usage (XEXP (eqv, 0), counts, incr);
7093 eqv = XEXP (eqv, 1);
7094 }
7095 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7096 else
7097 count_reg_usage (eqv, counts, incr);
7098 }
7099 return;
7100
7101 case EXPR_LIST:
7102 if (REG_NOTE_KIND (x) == REG_EQUAL
7103 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7104 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7105 involving registers in the address. */
7106 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7107 count_reg_usage (XEXP (x, 0), counts, incr);
7108
7109 count_reg_usage (XEXP (x, 1), counts, incr);
7110 return;
7111
7112 case ASM_OPERANDS:
7113 /* Iterate over just the inputs, not the constraints as well. */
7114 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7115 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, incr);
7116 return;
7117
7118 case INSN_LIST:
7119 gcc_unreachable ();
7120
7121 default:
7122 break;
7123 }
7124
7125 fmt = GET_RTX_FORMAT (code);
7126 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7127 {
7128 if (fmt[i] == 'e')
7129 count_reg_usage (XEXP (x, i), counts, incr);
7130 else if (fmt[i] == 'E')
7131 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7132 count_reg_usage (XVECEXP (x, i, j), counts, incr);
7133 }
7134 }
7135 \f
7136 /* Return true if set is live. */
7137 static bool
7138 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7139 int *counts)
7140 {
7141 #ifdef HAVE_cc0
7142 rtx tem;
7143 #endif
7144
7145 if (set_noop_p (set))
7146 ;
7147
7148 #ifdef HAVE_cc0
7149 else if (GET_CODE (SET_DEST (set)) == CC0
7150 && !side_effects_p (SET_SRC (set))
7151 && ((tem = next_nonnote_insn (insn)) == 0
7152 || !INSN_P (tem)
7153 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7154 return false;
7155 #endif
7156 else if (!REG_P (SET_DEST (set))
7157 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7158 || counts[REGNO (SET_DEST (set))] != 0
7159 || side_effects_p (SET_SRC (set)))
7160 return true;
7161 return false;
7162 }
7163
7164 /* Return true if insn is live. */
7165
7166 static bool
7167 insn_live_p (rtx insn, int *counts)
7168 {
7169 int i;
7170 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7171 return true;
7172 else if (GET_CODE (PATTERN (insn)) == SET)
7173 return set_live_p (PATTERN (insn), insn, counts);
7174 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7175 {
7176 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7177 {
7178 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7179
7180 if (GET_CODE (elt) == SET)
7181 {
7182 if (set_live_p (elt, insn, counts))
7183 return true;
7184 }
7185 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7186 return true;
7187 }
7188 return false;
7189 }
7190 else
7191 return true;
7192 }
7193
7194 /* Return true if libcall is dead as a whole. */
7195
7196 static bool
7197 dead_libcall_p (rtx insn, int *counts)
7198 {
7199 rtx note, set, new;
7200
7201 /* See if there's a REG_EQUAL note on this insn and try to
7202 replace the source with the REG_EQUAL expression.
7203
7204 We assume that insns with REG_RETVALs can only be reg->reg
7205 copies at this point. */
7206 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7207 if (!note)
7208 return false;
7209
7210 set = single_set (insn);
7211 if (!set)
7212 return false;
7213
7214 new = simplify_rtx (XEXP (note, 0));
7215 if (!new)
7216 new = XEXP (note, 0);
7217
7218 /* While changing insn, we must update the counts accordingly. */
7219 count_reg_usage (insn, counts, -1);
7220
7221 if (validate_change (insn, &SET_SRC (set), new, 0))
7222 {
7223 count_reg_usage (insn, counts, 1);
7224 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7225 remove_note (insn, note);
7226 return true;
7227 }
7228
7229 if (CONSTANT_P (new))
7230 {
7231 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7232 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7233 {
7234 count_reg_usage (insn, counts, 1);
7235 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7236 remove_note (insn, note);
7237 return true;
7238 }
7239 }
7240
7241 count_reg_usage (insn, counts, 1);
7242 return false;
7243 }
7244
7245 /* Scan all the insns and delete any that are dead; i.e., they store a register
7246 that is never used or they copy a register to itself.
7247
7248 This is used to remove insns made obviously dead by cse, loop or other
7249 optimizations. It improves the heuristics in loop since it won't try to
7250 move dead invariants out of loops or make givs for dead quantities. The
7251 remaining passes of the compilation are also sped up. */
7252
7253 int
7254 delete_trivially_dead_insns (rtx insns, int nreg)
7255 {
7256 int *counts;
7257 rtx insn, prev;
7258 int in_libcall = 0, dead_libcall = 0;
7259 int ndead = 0, nlastdead, niterations = 0;
7260
7261 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7262 /* First count the number of times each register is used. */
7263 counts = xcalloc (nreg, sizeof (int));
7264 for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn))
7265 count_reg_usage (insn, counts, 1);
7266
7267 do
7268 {
7269 nlastdead = ndead;
7270 niterations++;
7271 /* Go from the last insn to the first and delete insns that only set unused
7272 registers or copy a register to itself. As we delete an insn, remove
7273 usage counts for registers it uses.
7274
7275 The first jump optimization pass may leave a real insn as the last
7276 insn in the function. We must not skip that insn or we may end
7277 up deleting code that is not really dead. */
7278 insn = get_last_insn ();
7279 if (! INSN_P (insn))
7280 insn = prev_real_insn (insn);
7281
7282 for (; insn; insn = prev)
7283 {
7284 int live_insn = 0;
7285
7286 prev = prev_real_insn (insn);
7287
7288 /* Don't delete any insns that are part of a libcall block unless
7289 we can delete the whole libcall block.
7290
7291 Flow or loop might get confused if we did that. Remember
7292 that we are scanning backwards. */
7293 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7294 {
7295 in_libcall = 1;
7296 live_insn = 1;
7297 dead_libcall = dead_libcall_p (insn, counts);
7298 }
7299 else if (in_libcall)
7300 live_insn = ! dead_libcall;
7301 else
7302 live_insn = insn_live_p (insn, counts);
7303
7304 /* If this is a dead insn, delete it and show registers in it aren't
7305 being used. */
7306
7307 if (! live_insn)
7308 {
7309 count_reg_usage (insn, counts, -1);
7310 delete_insn_and_edges (insn);
7311 ndead++;
7312 }
7313
7314 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7315 {
7316 in_libcall = 0;
7317 dead_libcall = 0;
7318 }
7319 }
7320 }
7321 while (ndead != nlastdead);
7322
7323 if (dump_file && ndead)
7324 fprintf (dump_file, "Deleted %i trivially dead insns; %i iterations\n",
7325 ndead, niterations);
7326 /* Clean up. */
7327 free (counts);
7328 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7329 return ndead;
7330 }
7331
7332 /* This function is called via for_each_rtx. The argument, NEWREG, is
7333 a condition code register with the desired mode. If we are looking
7334 at the same register in a different mode, replace it with
7335 NEWREG. */
7336
7337 static int
7338 cse_change_cc_mode (rtx *loc, void *data)
7339 {
7340 rtx newreg = (rtx) data;
7341
7342 if (*loc
7343 && REG_P (*loc)
7344 && REGNO (*loc) == REGNO (newreg)
7345 && GET_MODE (*loc) != GET_MODE (newreg))
7346 {
7347 *loc = newreg;
7348 return -1;
7349 }
7350 return 0;
7351 }
7352
7353 /* Change the mode of any reference to the register REGNO (NEWREG) to
7354 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7355 any instruction which modifies NEWREG. */
7356
7357 static void
7358 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7359 {
7360 rtx insn;
7361
7362 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7363 {
7364 if (! INSN_P (insn))
7365 continue;
7366
7367 if (reg_set_p (newreg, insn))
7368 return;
7369
7370 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, newreg);
7371 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, newreg);
7372 }
7373 }
7374
7375 /* BB is a basic block which finishes with CC_REG as a condition code
7376 register which is set to CC_SRC. Look through the successors of BB
7377 to find blocks which have a single predecessor (i.e., this one),
7378 and look through those blocks for an assignment to CC_REG which is
7379 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7380 permitted to change the mode of CC_SRC to a compatible mode. This
7381 returns VOIDmode if no equivalent assignments were found.
7382 Otherwise it returns the mode which CC_SRC should wind up with.
7383
7384 The main complexity in this function is handling the mode issues.
7385 We may have more than one duplicate which we can eliminate, and we
7386 try to find a mode which will work for multiple duplicates. */
7387
7388 static enum machine_mode
7389 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7390 {
7391 bool found_equiv;
7392 enum machine_mode mode;
7393 unsigned int insn_count;
7394 edge e;
7395 rtx insns[2];
7396 enum machine_mode modes[2];
7397 rtx last_insns[2];
7398 unsigned int i;
7399 rtx newreg;
7400 edge_iterator ei;
7401
7402 /* We expect to have two successors. Look at both before picking
7403 the final mode for the comparison. If we have more successors
7404 (i.e., some sort of table jump, although that seems unlikely),
7405 then we require all beyond the first two to use the same
7406 mode. */
7407
7408 found_equiv = false;
7409 mode = GET_MODE (cc_src);
7410 insn_count = 0;
7411 FOR_EACH_EDGE (e, ei, bb->succs)
7412 {
7413 rtx insn;
7414 rtx end;
7415
7416 if (e->flags & EDGE_COMPLEX)
7417 continue;
7418
7419 if (EDGE_COUNT (e->dest->preds) != 1
7420 || e->dest == EXIT_BLOCK_PTR)
7421 continue;
7422
7423 end = NEXT_INSN (BB_END (e->dest));
7424 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7425 {
7426 rtx set;
7427
7428 if (! INSN_P (insn))
7429 continue;
7430
7431 /* If CC_SRC is modified, we have to stop looking for
7432 something which uses it. */
7433 if (modified_in_p (cc_src, insn))
7434 break;
7435
7436 /* Check whether INSN sets CC_REG to CC_SRC. */
7437 set = single_set (insn);
7438 if (set
7439 && REG_P (SET_DEST (set))
7440 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7441 {
7442 bool found;
7443 enum machine_mode set_mode;
7444 enum machine_mode comp_mode;
7445
7446 found = false;
7447 set_mode = GET_MODE (SET_SRC (set));
7448 comp_mode = set_mode;
7449 if (rtx_equal_p (cc_src, SET_SRC (set)))
7450 found = true;
7451 else if (GET_CODE (cc_src) == COMPARE
7452 && GET_CODE (SET_SRC (set)) == COMPARE
7453 && mode != set_mode
7454 && rtx_equal_p (XEXP (cc_src, 0),
7455 XEXP (SET_SRC (set), 0))
7456 && rtx_equal_p (XEXP (cc_src, 1),
7457 XEXP (SET_SRC (set), 1)))
7458
7459 {
7460 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7461 if (comp_mode != VOIDmode
7462 && (can_change_mode || comp_mode == mode))
7463 found = true;
7464 }
7465
7466 if (found)
7467 {
7468 found_equiv = true;
7469 if (insn_count < ARRAY_SIZE (insns))
7470 {
7471 insns[insn_count] = insn;
7472 modes[insn_count] = set_mode;
7473 last_insns[insn_count] = end;
7474 ++insn_count;
7475
7476 if (mode != comp_mode)
7477 {
7478 gcc_assert (can_change_mode);
7479 mode = comp_mode;
7480 PUT_MODE (cc_src, mode);
7481 }
7482 }
7483 else
7484 {
7485 if (set_mode != mode)
7486 {
7487 /* We found a matching expression in the
7488 wrong mode, but we don't have room to
7489 store it in the array. Punt. This case
7490 should be rare. */
7491 break;
7492 }
7493 /* INSN sets CC_REG to a value equal to CC_SRC
7494 with the right mode. We can simply delete
7495 it. */
7496 delete_insn (insn);
7497 }
7498
7499 /* We found an instruction to delete. Keep looking,
7500 in the hopes of finding a three-way jump. */
7501 continue;
7502 }
7503
7504 /* We found an instruction which sets the condition
7505 code, so don't look any farther. */
7506 break;
7507 }
7508
7509 /* If INSN sets CC_REG in some other way, don't look any
7510 farther. */
7511 if (reg_set_p (cc_reg, insn))
7512 break;
7513 }
7514
7515 /* If we fell off the bottom of the block, we can keep looking
7516 through successors. We pass CAN_CHANGE_MODE as false because
7517 we aren't prepared to handle compatibility between the
7518 further blocks and this block. */
7519 if (insn == end)
7520 {
7521 enum machine_mode submode;
7522
7523 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7524 if (submode != VOIDmode)
7525 {
7526 gcc_assert (submode == mode);
7527 found_equiv = true;
7528 can_change_mode = false;
7529 }
7530 }
7531 }
7532
7533 if (! found_equiv)
7534 return VOIDmode;
7535
7536 /* Now INSN_COUNT is the number of instructions we found which set
7537 CC_REG to a value equivalent to CC_SRC. The instructions are in
7538 INSNS. The modes used by those instructions are in MODES. */
7539
7540 newreg = NULL_RTX;
7541 for (i = 0; i < insn_count; ++i)
7542 {
7543 if (modes[i] != mode)
7544 {
7545 /* We need to change the mode of CC_REG in INSNS[i] and
7546 subsequent instructions. */
7547 if (! newreg)
7548 {
7549 if (GET_MODE (cc_reg) == mode)
7550 newreg = cc_reg;
7551 else
7552 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7553 }
7554 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7555 newreg);
7556 }
7557
7558 delete_insn (insns[i]);
7559 }
7560
7561 return mode;
7562 }
7563
7564 /* If we have a fixed condition code register (or two), walk through
7565 the instructions and try to eliminate duplicate assignments. */
7566
7567 void
7568 cse_condition_code_reg (void)
7569 {
7570 unsigned int cc_regno_1;
7571 unsigned int cc_regno_2;
7572 rtx cc_reg_1;
7573 rtx cc_reg_2;
7574 basic_block bb;
7575
7576 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7577 return;
7578
7579 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7580 if (cc_regno_2 != INVALID_REGNUM)
7581 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7582 else
7583 cc_reg_2 = NULL_RTX;
7584
7585 FOR_EACH_BB (bb)
7586 {
7587 rtx last_insn;
7588 rtx cc_reg;
7589 rtx insn;
7590 rtx cc_src_insn;
7591 rtx cc_src;
7592 enum machine_mode mode;
7593 enum machine_mode orig_mode;
7594
7595 /* Look for blocks which end with a conditional jump based on a
7596 condition code register. Then look for the instruction which
7597 sets the condition code register. Then look through the
7598 successor blocks for instructions which set the condition
7599 code register to the same value. There are other possible
7600 uses of the condition code register, but these are by far the
7601 most common and the ones which we are most likely to be able
7602 to optimize. */
7603
7604 last_insn = BB_END (bb);
7605 if (!JUMP_P (last_insn))
7606 continue;
7607
7608 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7609 cc_reg = cc_reg_1;
7610 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7611 cc_reg = cc_reg_2;
7612 else
7613 continue;
7614
7615 cc_src_insn = NULL_RTX;
7616 cc_src = NULL_RTX;
7617 for (insn = PREV_INSN (last_insn);
7618 insn && insn != PREV_INSN (BB_HEAD (bb));
7619 insn = PREV_INSN (insn))
7620 {
7621 rtx set;
7622
7623 if (! INSN_P (insn))
7624 continue;
7625 set = single_set (insn);
7626 if (set
7627 && REG_P (SET_DEST (set))
7628 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7629 {
7630 cc_src_insn = insn;
7631 cc_src = SET_SRC (set);
7632 break;
7633 }
7634 else if (reg_set_p (cc_reg, insn))
7635 break;
7636 }
7637
7638 if (! cc_src_insn)
7639 continue;
7640
7641 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7642 continue;
7643
7644 /* Now CC_REG is a condition code register used for a
7645 conditional jump at the end of the block, and CC_SRC, in
7646 CC_SRC_INSN, is the value to which that condition code
7647 register is set, and CC_SRC is still meaningful at the end of
7648 the basic block. */
7649
7650 orig_mode = GET_MODE (cc_src);
7651 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7652 if (mode != VOIDmode)
7653 {
7654 gcc_assert (mode == GET_MODE (cc_src));
7655 if (mode != orig_mode)
7656 {
7657 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7658
7659 /* Change the mode of CC_REG in CC_SRC_INSN to
7660 GET_MODE (NEWREG). */
7661 for_each_rtx (&PATTERN (cc_src_insn), cse_change_cc_mode,
7662 newreg);
7663 for_each_rtx (&REG_NOTES (cc_src_insn), cse_change_cc_mode,
7664 newreg);
7665
7666 /* Do the same in the following insns that use the
7667 current value of CC_REG within BB. */
7668 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7669 NEXT_INSN (last_insn),
7670 newreg);
7671 }
7672 }
7673 }
7674 }