* config/rs6000/rs6000.md: Document why a pattern is not
[official-gcc.git] / gcc / cse.c
blob0a7281ecc426b4bbe75c00eabd4b714e6b19faa9
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.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
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"
47 /* The basic idea of common subexpression elimination is to go
48 through the code, keeping a record of expressions that would
49 have the same value at the current scan point, and replacing
50 expressions encountered with the cheapest equivalent expression.
52 It is too complicated to keep track of the different possibilities
53 when control paths merge in this code; so, at each label, we forget all
54 that is known and start fresh. This can be described as processing each
55 extended basic block separately. We have a separate pass to perform
56 global CSE.
58 Note CSE can turn a conditional or computed jump into a nop or
59 an unconditional jump. When this occurs we arrange to run the jump
60 optimizer after CSE to delete the unreachable code.
62 We use two data structures to record the equivalent expressions:
63 a hash table for most expressions, and a vector of "quantity
64 numbers" to record equivalent (pseudo) registers.
66 The use of the special data structure for registers is desirable
67 because it is faster. It is possible because registers references
68 contain a fairly small number, the register number, taken from
69 a contiguously allocated series, and two register references are
70 identical if they have the same number. General expressions
71 do not have any such thing, so the only way to retrieve the
72 information recorded on an expression other than a register
73 is to keep it in a hash table.
75 Registers and "quantity numbers":
77 At the start of each basic block, all of the (hardware and pseudo)
78 registers used in the function are given distinct quantity
79 numbers to indicate their contents. During scan, when the code
80 copies one register into another, we copy the quantity number.
81 When a register is loaded in any other way, we allocate a new
82 quantity number to describe the value generated by this operation.
83 `reg_qty' records what quantity a register is currently thought
84 of as containing.
86 All real quantity numbers are greater than or equal to `max_reg'.
87 If register N has not been assigned a quantity, reg_qty[N] will equal N.
89 Quantity numbers below `max_reg' do not exist and none of the `qty_table'
90 entries should be referenced with an index below `max_reg'.
92 We also maintain a bidirectional chain of registers for each
93 quantity number. The `qty_table` members `first_reg' and `last_reg',
94 and `reg_eqv_table' members `next' and `prev' hold these chains.
96 The first register in a chain is the one whose lifespan is least local.
97 Among equals, it is the one that was seen first.
98 We replace any equivalent register with that one.
100 If two registers have the same quantity number, it must be true that
101 REG expressions with qty_table `mode' must be in the hash table for both
102 registers and must be in the same class.
104 The converse is not true. Since hard registers may be referenced in
105 any mode, two REG expressions might be equivalent in the hash table
106 but not have the same quantity number if the quantity number of one
107 of the registers is not the same mode as those expressions.
109 Constants and quantity numbers
111 When a quantity has a known constant value, that value is stored
112 in the appropriate qty_table `const_rtx'. This is in addition to
113 putting the constant in the hash table as is usual for non-regs.
115 Whether a reg or a constant is preferred is determined by the configuration
116 macro CONST_COSTS and will often depend on the constant value. In any
117 event, expressions containing constants can be simplified, by fold_rtx.
119 When a quantity has a known nearly constant value (such as an address
120 of a stack slot), that value is stored in the appropriate qty_table
121 `const_rtx'.
123 Integer constants don't have a machine mode. However, cse
124 determines the intended machine mode from the destination
125 of the instruction that moves the constant. The machine mode
126 is recorded in the hash table along with the actual RTL
127 constant expression so that different modes are kept separate.
129 Other expressions:
131 To record known equivalences among expressions in general
132 we use a hash table called `table'. It has a fixed number of buckets
133 that contain chains of `struct table_elt' elements for expressions.
134 These chains connect the elements whose expressions have the same
135 hash codes.
137 Other chains through the same elements connect the elements which
138 currently have equivalent values.
140 Register references in an expression are canonicalized before hashing
141 the expression. This is done using `reg_qty' and qty_table `first_reg'.
142 The hash code of a register reference is computed using the quantity
143 number, not the register number.
145 When the value of an expression changes, it is necessary to remove from the
146 hash table not just that expression but all expressions whose values
147 could be different as a result.
149 1. If the value changing is in memory, except in special cases
150 ANYTHING referring to memory could be changed. That is because
151 nobody knows where a pointer does not point.
152 The function `invalidate_memory' removes what is necessary.
154 The special cases are when the address is constant or is
155 a constant plus a fixed register such as the frame pointer
156 or a static chain pointer. When such addresses are stored in,
157 we can tell exactly which other such addresses must be invalidated
158 due to overlap. `invalidate' does this.
159 All expressions that refer to non-constant
160 memory addresses are also invalidated. `invalidate_memory' does this.
162 2. If the value changing is a register, all expressions
163 containing references to that register, and only those,
164 must be removed.
166 Because searching the entire hash table for expressions that contain
167 a register is very slow, we try to figure out when it isn't necessary.
168 Precisely, this is necessary only when expressions have been
169 entered in the hash table using this register, and then the value has
170 changed, and then another expression wants to be added to refer to
171 the register's new value. This sequence of circumstances is rare
172 within any one basic block.
174 The vectors `reg_tick' and `reg_in_table' are used to detect this case.
175 reg_tick[i] is incremented whenever a value is stored in register i.
176 reg_in_table[i] holds -1 if no references to register i have been
177 entered in the table; otherwise, it contains the value reg_tick[i] had
178 when the references were entered. If we want to enter a reference
179 and reg_in_table[i] != reg_tick[i], we must scan and remove old references.
180 Until we want to enter a new entry, the mere fact that the two vectors
181 don't match makes the entries be ignored if anyone tries to match them.
183 Registers themselves are entered in the hash table as well as in
184 the equivalent-register chains. However, the vectors `reg_tick'
185 and `reg_in_table' do not apply to expressions which are simple
186 register references. These expressions are removed from the table
187 immediately when they become invalid, and this can be done even if
188 we do not immediately search for all the expressions that refer to
189 the register.
191 A CLOBBER rtx in an instruction invalidates its operand for further
192 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
193 invalidates everything that resides in memory.
195 Related expressions:
197 Constant expressions that differ only by an additive integer
198 are called related. When a constant expression is put in
199 the table, the related expression with no constant term
200 is also entered. These are made to point at each other
201 so that it is possible to find out if there exists any
202 register equivalent to an expression related to a given expression. */
204 /* One plus largest register number used in this function. */
206 static int max_reg;
208 /* One plus largest instruction UID used in this function at time of
209 cse_main call. */
211 static int max_insn_uid;
213 /* Length of qty_table vector. We know in advance we will not need
214 a quantity number this big. */
216 static int max_qty;
218 /* Next quantity number to be allocated.
219 This is 1 + the largest number needed so far. */
221 static int next_qty;
223 /* Per-qty information tracking.
225 `first_reg' and `last_reg' track the head and tail of the
226 chain of registers which currently contain this quantity.
228 `mode' contains the machine mode of this quantity.
230 `const_rtx' holds the rtx of the constant value of this
231 quantity, if known. A summations of the frame/arg pointer
232 and a constant can also be entered here. When this holds
233 a known value, `const_insn' is the insn which stored the
234 constant value.
236 `comparison_{code,const,qty}' are used to track when a
237 comparison between a quantity and some constant or register has
238 been passed. In such a case, we know the results of the comparison
239 in case we see it again. These members record a comparison that
240 is known to be true. `comparison_code' holds the rtx code of such
241 a comparison, else it is set to UNKNOWN and the other two
242 comparison members are undefined. `comparison_const' holds
243 the constant being compared against, or zero if the comparison
244 is not against a constant. `comparison_qty' holds the quantity
245 being compared against when the result is known. If the comparison
246 is not with a register, `comparison_qty' is -1. */
248 struct qty_table_elem
250 rtx const_rtx;
251 rtx const_insn;
252 rtx comparison_const;
253 int comparison_qty;
254 unsigned int first_reg, last_reg;
255 /* The sizes of these fields should match the sizes of the
256 code and mode fields of struct rtx_def (see rtl.h). */
257 ENUM_BITFIELD(rtx_code) comparison_code : 16;
258 ENUM_BITFIELD(machine_mode) mode : 8;
261 /* The table of all qtys, indexed by qty number. */
262 static struct qty_table_elem *qty_table;
264 #ifdef HAVE_cc0
265 /* For machines that have a CC0, we do not record its value in the hash
266 table since its use is guaranteed to be the insn immediately following
267 its definition and any other insn is presumed to invalidate it.
269 Instead, we store below the value last assigned to CC0. If it should
270 happen to be a constant, it is stored in preference to the actual
271 assigned value. In case it is a constant, we store the mode in which
272 the constant should be interpreted. */
274 static rtx prev_insn_cc0;
275 static enum machine_mode prev_insn_cc0_mode;
277 /* Previous actual insn. 0 if at first insn of basic block. */
279 static rtx prev_insn;
280 #endif
282 /* Insn being scanned. */
284 static rtx this_insn;
286 /* Index by register number, gives the number of the next (or
287 previous) register in the chain of registers sharing the same
288 value.
290 Or -1 if this register is at the end of the chain.
292 If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */
294 /* Per-register equivalence chain. */
295 struct reg_eqv_elem
297 int next, prev;
300 /* The table of all register equivalence chains. */
301 static struct reg_eqv_elem *reg_eqv_table;
303 struct cse_reg_info
305 /* Next in hash chain. */
306 struct cse_reg_info *hash_next;
308 /* The next cse_reg_info structure in the free or used list. */
309 struct cse_reg_info *next;
311 /* Search key */
312 unsigned int regno;
314 /* The quantity number of the register's current contents. */
315 int reg_qty;
317 /* The number of times the register has been altered in the current
318 basic block. */
319 int reg_tick;
321 /* The REG_TICK value at which rtx's containing this register are
322 valid in the hash table. If this does not equal the current
323 reg_tick value, such expressions existing in the hash table are
324 invalid. */
325 int reg_in_table;
327 /* The SUBREG that was set when REG_TICK was last incremented. Set
328 to -1 if the last store was to the whole register, not a subreg. */
329 unsigned int subreg_ticked;
332 /* A free list of cse_reg_info entries. */
333 static struct cse_reg_info *cse_reg_info_free_list;
335 /* A used list of cse_reg_info entries. */
336 static struct cse_reg_info *cse_reg_info_used_list;
337 static struct cse_reg_info *cse_reg_info_used_list_end;
339 /* A mapping from registers to cse_reg_info data structures. */
340 #define REGHASH_SHIFT 7
341 #define REGHASH_SIZE (1 << REGHASH_SHIFT)
342 #define REGHASH_MASK (REGHASH_SIZE - 1)
343 static struct cse_reg_info *reg_hash[REGHASH_SIZE];
345 #define REGHASH_FN(REGNO) \
346 (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK)
348 /* The last lookup we did into the cse_reg_info_tree. This allows us
349 to cache repeated lookups. */
350 static unsigned int cached_regno;
351 static struct cse_reg_info *cached_cse_reg_info;
353 /* A HARD_REG_SET containing all the hard registers for which there is
354 currently a REG expression in the hash table. Note the difference
355 from the above variables, which indicate if the REG is mentioned in some
356 expression in the table. */
358 static HARD_REG_SET hard_regs_in_table;
360 /* CUID of insn that starts the basic block currently being cse-processed. */
362 static int cse_basic_block_start;
364 /* CUID of insn that ends the basic block currently being cse-processed. */
366 static int cse_basic_block_end;
368 /* Vector mapping INSN_UIDs to cuids.
369 The cuids are like uids but increase monotonically always.
370 We use them to see whether a reg is used outside a given basic block. */
372 static int *uid_cuid;
374 /* Highest UID in UID_CUID. */
375 static int max_uid;
377 /* Get the cuid of an insn. */
379 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
381 /* Nonzero if this pass has made changes, and therefore it's
382 worthwhile to run the garbage collector. */
384 static int cse_altered;
386 /* Nonzero if cse has altered conditional jump insns
387 in such a way that jump optimization should be redone. */
389 static int cse_jumps_altered;
391 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
392 REG_LABEL, we have to rerun jump after CSE to put in the note. */
393 static int recorded_label_ref;
395 /* canon_hash stores 1 in do_not_record
396 if it notices a reference to CC0, PC, or some other volatile
397 subexpression. */
399 static int do_not_record;
401 #ifdef LOAD_EXTEND_OP
403 /* Scratch rtl used when looking for load-extended copy of a MEM. */
404 static rtx memory_extend_rtx;
405 #endif
407 /* canon_hash stores 1 in hash_arg_in_memory
408 if it notices a reference to memory within the expression being hashed. */
410 static int hash_arg_in_memory;
412 /* The hash table contains buckets which are chains of `struct table_elt's,
413 each recording one expression's information.
414 That expression is in the `exp' field.
416 The canon_exp field contains a canonical (from the point of view of
417 alias analysis) version of the `exp' field.
419 Those elements with the same hash code are chained in both directions
420 through the `next_same_hash' and `prev_same_hash' fields.
422 Each set of expressions with equivalent values
423 are on a two-way chain through the `next_same_value'
424 and `prev_same_value' fields, and all point with
425 the `first_same_value' field at the first element in
426 that chain. The chain is in order of increasing cost.
427 Each element's cost value is in its `cost' field.
429 The `in_memory' field is nonzero for elements that
430 involve any reference to memory. These elements are removed
431 whenever a write is done to an unidentified location in memory.
432 To be safe, we assume that a memory address is unidentified unless
433 the address is either a symbol constant or a constant plus
434 the frame pointer or argument pointer.
436 The `related_value' field is used to connect related expressions
437 (that differ by adding an integer).
438 The related expressions are chained in a circular fashion.
439 `related_value' is zero for expressions for which this
440 chain is not useful.
442 The `cost' field stores the cost of this element's expression.
443 The `regcost' field stores the value returned by approx_reg_cost for
444 this element's expression.
446 The `is_const' flag is set if the element is a constant (including
447 a fixed address).
449 The `flag' field is used as a temporary during some search routines.
451 The `mode' field is usually the same as GET_MODE (`exp'), but
452 if `exp' is a CONST_INT and has no machine mode then the `mode'
453 field is the mode it was being used as. Each constant is
454 recorded separately for each mode it is used with. */
456 struct table_elt
458 rtx exp;
459 rtx canon_exp;
460 struct table_elt *next_same_hash;
461 struct table_elt *prev_same_hash;
462 struct table_elt *next_same_value;
463 struct table_elt *prev_same_value;
464 struct table_elt *first_same_value;
465 struct table_elt *related_value;
466 int cost;
467 int regcost;
468 /* The size of this field should match the size
469 of the mode field of struct rtx_def (see rtl.h). */
470 ENUM_BITFIELD(machine_mode) mode : 8;
471 char in_memory;
472 char is_const;
473 char flag;
476 /* We don't want a lot of buckets, because we rarely have very many
477 things stored in the hash table, and a lot of buckets slows
478 down a lot of loops that happen frequently. */
479 #define HASH_SHIFT 5
480 #define HASH_SIZE (1 << HASH_SHIFT)
481 #define HASH_MASK (HASH_SIZE - 1)
483 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
484 register (hard registers may require `do_not_record' to be set). */
486 #define HASH(X, M) \
487 ((GET_CODE (X) == REG && REGNO (X) >= FIRST_PSEUDO_REGISTER \
488 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
489 : canon_hash (X, M)) & HASH_MASK)
491 /* Determine whether register number N is considered a fixed register for the
492 purpose of approximating register costs.
493 It is desirable to replace other regs with fixed regs, to reduce need for
494 non-fixed hard regs.
495 A reg wins if it is either the frame pointer or designated as fixed. */
496 #define FIXED_REGNO_P(N) \
497 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
498 || fixed_regs[N] || global_regs[N])
500 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
501 hard registers and pointers into the frame are the cheapest with a cost
502 of 0. Next come pseudos with a cost of one and other hard registers with
503 a cost of 2. Aside from these special cases, call `rtx_cost'. */
505 #define CHEAP_REGNO(N) \
506 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
507 || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \
508 || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \
509 || ((N) < FIRST_PSEUDO_REGISTER \
510 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
512 #define COST(X) (GET_CODE (X) == REG ? 0 : notreg_cost (X, SET))
513 #define COST_IN(X,OUTER) (GET_CODE (X) == REG ? 0 : notreg_cost (X, OUTER))
515 /* Get the info associated with register N. */
517 #define GET_CSE_REG_INFO(N) \
518 (((N) == cached_regno && cached_cse_reg_info) \
519 ? cached_cse_reg_info : get_cse_reg_info ((N)))
521 /* Get the number of times this register has been updated in this
522 basic block. */
524 #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->reg_tick)
526 /* Get the point at which REG was recorded in the table. */
528 #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table)
530 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
531 SUBREG). */
533 #define SUBREG_TICKED(N) ((GET_CSE_REG_INFO (N))->subreg_ticked)
535 /* Get the quantity number for REG. */
537 #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty)
539 /* Determine if the quantity number for register X represents a valid index
540 into the qty_table. */
542 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) != (int) (N))
544 static struct table_elt *table[HASH_SIZE];
546 /* Chain of `struct table_elt's made so far for this function
547 but currently removed from the table. */
549 static struct table_elt *free_element_chain;
551 /* Number of `struct table_elt' structures made so far for this function. */
553 static int n_elements_made;
555 /* Maximum value `n_elements_made' has had so far in this compilation
556 for functions previously processed. */
558 static int max_elements_made;
560 /* Surviving equivalence class when two equivalence classes are merged
561 by recording the effects of a jump in the last insn. Zero if the
562 last insn was not a conditional jump. */
564 static struct table_elt *last_jump_equiv_class;
566 /* Set to the cost of a constant pool reference if one was found for a
567 symbolic constant. If this was found, it means we should try to
568 convert constants into constant pool entries if they don't fit in
569 the insn. */
571 static int constant_pool_entries_cost;
572 static int constant_pool_entries_regcost;
574 /* This data describes a block that will be processed by cse_basic_block. */
576 struct cse_basic_block_data
578 /* Lowest CUID value of insns in block. */
579 int low_cuid;
580 /* Highest CUID value of insns in block. */
581 int high_cuid;
582 /* Total number of SETs in block. */
583 int nsets;
584 /* Last insn in the block. */
585 rtx last;
586 /* Size of current branch path, if any. */
587 int path_size;
588 /* Current branch path, indicating which branches will be taken. */
589 struct branch_path
591 /* The branch insn. */
592 rtx branch;
593 /* Whether it should be taken or not. AROUND is the same as taken
594 except that it is used when the destination label is not preceded
595 by a BARRIER. */
596 enum taken {TAKEN, NOT_TAKEN, AROUND} status;
597 } *path;
600 static bool fixed_base_plus_p (rtx x);
601 static int notreg_cost (rtx, enum rtx_code);
602 static int approx_reg_cost_1 (rtx *, void *);
603 static int approx_reg_cost (rtx);
604 static int preferable (int, int, int, int);
605 static void new_basic_block (void);
606 static void make_new_qty (unsigned int, enum machine_mode);
607 static void make_regs_eqv (unsigned int, unsigned int);
608 static void delete_reg_equiv (unsigned int);
609 static int mention_regs (rtx);
610 static int insert_regs (rtx, struct table_elt *, int);
611 static void remove_from_table (struct table_elt *, unsigned);
612 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
613 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
614 static rtx lookup_as_function (rtx, enum rtx_code);
615 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
616 enum machine_mode);
617 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
618 static void invalidate (rtx, enum machine_mode);
619 static int cse_rtx_varies_p (rtx, int);
620 static void remove_invalid_refs (unsigned int);
621 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
622 enum machine_mode);
623 static void rehash_using_reg (rtx);
624 static void invalidate_memory (void);
625 static void invalidate_for_call (void);
626 static rtx use_related_value (rtx, struct table_elt *);
627 static unsigned canon_hash (rtx, enum machine_mode);
628 static unsigned canon_hash_string (const char *);
629 static unsigned safe_hash (rtx, enum machine_mode);
630 static int exp_equiv_p (rtx, rtx, int, int);
631 static rtx canon_reg (rtx, rtx);
632 static void find_best_addr (rtx, rtx *, enum machine_mode);
633 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
634 enum machine_mode *,
635 enum machine_mode *);
636 static rtx fold_rtx (rtx, rtx);
637 static rtx equiv_constant (rtx);
638 static void record_jump_equiv (rtx, int);
639 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
640 int);
641 static void cse_insn (rtx, rtx);
642 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
643 int, int, int);
644 static int addr_affects_sp_p (rtx);
645 static void invalidate_from_clobbers (rtx);
646 static rtx cse_process_notes (rtx, rtx);
647 static void cse_around_loop (rtx);
648 static void invalidate_skipped_set (rtx, rtx, void *);
649 static void invalidate_skipped_block (rtx);
650 static void cse_check_loop_start (rtx, rtx, void *);
651 static void cse_set_around_loop (rtx, rtx, rtx);
652 static rtx cse_basic_block (rtx, rtx, struct branch_path *, int);
653 static void count_reg_usage (rtx, int *, int);
654 static int check_for_label_ref (rtx *, void *);
655 extern void dump_class (struct table_elt*);
656 static struct cse_reg_info * get_cse_reg_info (unsigned int);
657 static int check_dependence (rtx *, void *);
659 static void flush_hash_table (void);
660 static bool insn_live_p (rtx, int *);
661 static bool set_live_p (rtx, rtx, int *);
662 static bool dead_libcall_p (rtx, int *);
663 static int cse_change_cc_mode (rtx *, void *);
664 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
665 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
667 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
668 virtual regs here because the simplify_*_operation routines are called
669 by integrate.c, which is called before virtual register instantiation. */
671 static bool
672 fixed_base_plus_p (rtx x)
674 switch (GET_CODE (x))
676 case REG:
677 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
678 return true;
679 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
680 return true;
681 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
682 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
683 return true;
684 return false;
686 case PLUS:
687 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
688 return false;
689 return fixed_base_plus_p (XEXP (x, 0));
691 case ADDRESSOF:
692 return true;
694 default:
695 return false;
699 /* Dump the expressions in the equivalence class indicated by CLASSP.
700 This function is used only for debugging. */
701 void
702 dump_class (struct table_elt *classp)
704 struct table_elt *elt;
706 fprintf (stderr, "Equivalence chain for ");
707 print_rtl (stderr, classp->exp);
708 fprintf (stderr, ": \n");
710 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
712 print_rtl (stderr, elt->exp);
713 fprintf (stderr, "\n");
717 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
719 static int
720 approx_reg_cost_1 (rtx *xp, void *data)
722 rtx x = *xp;
723 int *cost_p = data;
725 if (x && GET_CODE (x) == REG)
727 unsigned int regno = REGNO (x);
729 if (! CHEAP_REGNO (regno))
731 if (regno < FIRST_PSEUDO_REGISTER)
733 if (SMALL_REGISTER_CLASSES)
734 return 1;
735 *cost_p += 2;
737 else
738 *cost_p += 1;
742 return 0;
745 /* Return an estimate of the cost of the registers used in an rtx.
746 This is mostly the number of different REG expressions in the rtx;
747 however for some exceptions like fixed registers we use a cost of
748 0. If any other hard register reference occurs, return MAX_COST. */
750 static int
751 approx_reg_cost (rtx x)
753 int cost = 0;
755 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
756 return MAX_COST;
758 return cost;
761 /* Return a negative value if an rtx A, whose costs are given by COST_A
762 and REGCOST_A, is more desirable than an rtx B.
763 Return a positive value if A is less desirable, or 0 if the two are
764 equally good. */
765 static int
766 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
768 /* First, get rid of cases involving expressions that are entirely
769 unwanted. */
770 if (cost_a != cost_b)
772 if (cost_a == MAX_COST)
773 return 1;
774 if (cost_b == MAX_COST)
775 return -1;
778 /* Avoid extending lifetimes of hardregs. */
779 if (regcost_a != regcost_b)
781 if (regcost_a == MAX_COST)
782 return 1;
783 if (regcost_b == MAX_COST)
784 return -1;
787 /* Normal operation costs take precedence. */
788 if (cost_a != cost_b)
789 return cost_a - cost_b;
790 /* Only if these are identical consider effects on register pressure. */
791 if (regcost_a != regcost_b)
792 return regcost_a - regcost_b;
793 return 0;
796 /* Internal function, to compute cost when X is not a register; called
797 from COST macro to keep it simple. */
799 static int
800 notreg_cost (rtx x, enum rtx_code outer)
802 return ((GET_CODE (x) == SUBREG
803 && GET_CODE (SUBREG_REG (x)) == REG
804 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
805 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
806 && (GET_MODE_SIZE (GET_MODE (x))
807 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
808 && subreg_lowpart_p (x)
809 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
810 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
812 : rtx_cost (x, outer) * 2);
816 static struct cse_reg_info *
817 get_cse_reg_info (unsigned int regno)
819 struct cse_reg_info **hash_head = &reg_hash[REGHASH_FN (regno)];
820 struct cse_reg_info *p;
822 for (p = *hash_head; p != NULL; p = p->hash_next)
823 if (p->regno == regno)
824 break;
826 if (p == NULL)
828 /* Get a new cse_reg_info structure. */
829 if (cse_reg_info_free_list)
831 p = cse_reg_info_free_list;
832 cse_reg_info_free_list = p->next;
834 else
835 p = xmalloc (sizeof (struct cse_reg_info));
837 /* Insert into hash table. */
838 p->hash_next = *hash_head;
839 *hash_head = p;
841 /* Initialize it. */
842 p->reg_tick = 1;
843 p->reg_in_table = -1;
844 p->subreg_ticked = -1;
845 p->reg_qty = regno;
846 p->regno = regno;
847 p->next = cse_reg_info_used_list;
848 cse_reg_info_used_list = p;
849 if (!cse_reg_info_used_list_end)
850 cse_reg_info_used_list_end = p;
853 /* Cache this lookup; we tend to be looking up information about the
854 same register several times in a row. */
855 cached_regno = regno;
856 cached_cse_reg_info = p;
858 return p;
861 /* Clear the hash table and initialize each register with its own quantity,
862 for a new basic block. */
864 static void
865 new_basic_block (void)
867 int i;
869 next_qty = max_reg;
871 /* Clear out hash table state for this pass. */
873 memset (reg_hash, 0, sizeof reg_hash);
875 if (cse_reg_info_used_list)
877 cse_reg_info_used_list_end->next = cse_reg_info_free_list;
878 cse_reg_info_free_list = cse_reg_info_used_list;
879 cse_reg_info_used_list = cse_reg_info_used_list_end = 0;
881 cached_cse_reg_info = 0;
883 CLEAR_HARD_REG_SET (hard_regs_in_table);
885 /* The per-quantity values used to be initialized here, but it is
886 much faster to initialize each as it is made in `make_new_qty'. */
888 for (i = 0; i < HASH_SIZE; i++)
890 struct table_elt *first;
892 first = table[i];
893 if (first != NULL)
895 struct table_elt *last = first;
897 table[i] = NULL;
899 while (last->next_same_hash != NULL)
900 last = last->next_same_hash;
902 /* Now relink this hash entire chain into
903 the free element list. */
905 last->next_same_hash = free_element_chain;
906 free_element_chain = first;
910 #ifdef HAVE_cc0
911 prev_insn = 0;
912 prev_insn_cc0 = 0;
913 #endif
916 /* Say that register REG contains a quantity in mode MODE not in any
917 register before and initialize that quantity. */
919 static void
920 make_new_qty (unsigned int reg, enum machine_mode mode)
922 int q;
923 struct qty_table_elem *ent;
924 struct reg_eqv_elem *eqv;
926 if (next_qty >= max_qty)
927 abort ();
929 q = REG_QTY (reg) = next_qty++;
930 ent = &qty_table[q];
931 ent->first_reg = reg;
932 ent->last_reg = reg;
933 ent->mode = mode;
934 ent->const_rtx = ent->const_insn = NULL_RTX;
935 ent->comparison_code = UNKNOWN;
937 eqv = &reg_eqv_table[reg];
938 eqv->next = eqv->prev = -1;
941 /* Make reg NEW equivalent to reg OLD.
942 OLD is not changing; NEW is. */
944 static void
945 make_regs_eqv (unsigned int new, unsigned int old)
947 unsigned int lastr, firstr;
948 int q = REG_QTY (old);
949 struct qty_table_elem *ent;
951 ent = &qty_table[q];
953 /* Nothing should become eqv until it has a "non-invalid" qty number. */
954 if (! REGNO_QTY_VALID_P (old))
955 abort ();
957 REG_QTY (new) = q;
958 firstr = ent->first_reg;
959 lastr = ent->last_reg;
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)]))))))
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;
985 else
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;
1005 /* Remove REG from its equivalence class. */
1007 static void
1008 delete_reg_equiv (unsigned int reg)
1010 struct qty_table_elem *ent;
1011 int q = REG_QTY (reg);
1012 int p, n;
1014 /* If invalid, do nothing. */
1015 if (q == (int) reg)
1016 return;
1018 ent = &qty_table[q];
1020 p = reg_eqv_table[reg].prev;
1021 n = reg_eqv_table[reg].next;
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;
1032 REG_QTY (reg) = reg;
1035 /* Remove any invalid expressions from the hash table
1036 that refer to any of the registers contained in expression X.
1038 Make sure that newly inserted references to those registers
1039 as subexpressions will be considered valid.
1041 mention_regs is not called when a register itself
1042 is being stored in the table.
1044 Return 1 if we have done something that may have changed the hash code
1045 of X. */
1047 static int
1048 mention_regs (rtx x)
1050 enum rtx_code code;
1051 int i, j;
1052 const char *fmt;
1053 int changed = 0;
1055 if (x == 0)
1056 return 0;
1058 code = GET_CODE (x);
1059 if (code == REG)
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;
1067 for (i = regno; i < endregno; i++)
1069 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1070 remove_invalid_refs (i);
1072 REG_IN_TABLE (i) = REG_TICK (i);
1073 SUBREG_TICKED (i) = -1;
1076 return 0;
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 && GET_CODE (SUBREG_REG (x)) == REG
1083 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1085 unsigned int i = REGNO (SUBREG_REG (x));
1087 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
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));
1101 REG_IN_TABLE (i) = REG_TICK (i);
1102 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1103 return 0;
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.
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. */
1116 if (code == COMPARE || COMPARISON_P (x))
1118 if (GET_CODE (XEXP (x, 0)) == REG
1119 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1120 if (insert_regs (XEXP (x, 0), NULL, 0))
1122 rehash_using_reg (XEXP (x, 0));
1123 changed = 1;
1126 if (GET_CODE (XEXP (x, 1)) == REG
1127 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1128 if (insert_regs (XEXP (x, 1), NULL, 0))
1130 rehash_using_reg (XEXP (x, 1));
1131 changed = 1;
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));
1143 return changed;
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.
1153 Nonzero value means that elements of reg_qty have changed
1154 so X's hash code may be different. */
1156 static int
1157 insert_regs (rtx x, struct table_elt *classp, int modified)
1159 if (GET_CODE (x) == REG)
1161 unsigned int regno = REGNO (x);
1162 int qty_valid;
1164 /* If REGNO is in the equivalence table already but is of the
1165 wrong mode for that equivalence, don't do anything here. */
1167 qty_valid = REGNO_QTY_VALID_P (regno);
1168 if (qty_valid)
1170 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1172 if (ent->mode != GET_MODE (x))
1173 return 0;
1176 if (modified || ! qty_valid)
1178 if (classp)
1179 for (classp = classp->first_same_value;
1180 classp != 0;
1181 classp = classp->next_same_value)
1182 if (GET_CODE (classp->exp) == REG
1183 && GET_MODE (classp->exp) == GET_MODE (x))
1185 make_regs_eqv (regno, REGNO (classp->exp));
1186 return 1;
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;
1206 return 0;
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. */
1215 else if (GET_CODE (x) == SUBREG && GET_CODE (SUBREG_REG (x)) == REG
1216 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1218 insert_regs (SUBREG_REG (x), NULL, 0);
1219 mention_regs (x);
1220 return 1;
1222 else
1223 return mention_regs (x);
1226 /* Look in or update the hash table. */
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. */
1233 static void
1234 remove_from_table (struct table_elt *elt, unsigned int hash)
1236 if (elt == 0)
1237 return;
1239 /* Mark this element as removed. See cse_insn. */
1240 elt->first_same_value = 0;
1242 /* Remove the table element from its equivalence class. */
1245 struct table_elt *prev = elt->prev_same_value;
1246 struct table_elt *next = elt->next_same_value;
1248 if (next)
1249 next->prev_same_value = prev;
1251 if (prev)
1252 prev->next_same_value = next;
1253 else
1255 struct table_elt *newfirst = next;
1256 while (next)
1258 next->first_same_value = newfirst;
1259 next = next->next_same_value;
1264 /* Remove the table element from its hash bucket. */
1267 struct table_elt *prev = elt->prev_same_hash;
1268 struct table_elt *next = elt->next_same_hash;
1270 if (next)
1271 next->prev_same_hash = prev;
1273 if (prev)
1274 prev->next_same_hash = next;
1275 else if (table[hash] == elt)
1276 table[hash] = next;
1277 else
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;
1289 /* Remove the table element from its related-value circular chain. */
1291 if (elt->related_value != 0 && elt->related_value != elt)
1293 struct table_elt *p = elt->related_value;
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;
1302 /* Now add it to the free element chain. */
1303 elt->next_same_hash = free_element_chain;
1304 free_element_chain = elt;
1307 /* Look up X in the hash table and return its table element,
1308 or 0 if X is not in the table.
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.
1313 Here we are satisfied to find an expression whose tree structure
1314 looks like X. */
1316 static struct table_elt *
1317 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1319 struct table_elt *p;
1321 for (p = table[hash]; p; p = p->next_same_hash)
1322 if (mode == p->mode && ((x == p->exp && GET_CODE (x) == REG)
1323 || exp_equiv_p (x, p->exp, GET_CODE (x) != REG, 0)))
1324 return p;
1326 return 0;
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. */
1332 static struct table_elt *
1333 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1335 struct table_elt *p;
1337 if (GET_CODE (x) == REG)
1339 unsigned int regno = REGNO (x);
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 (GET_CODE (p->exp) == REG
1345 && REGNO (p->exp) == regno)
1346 return p;
1348 else
1350 for (p = table[hash]; p; p = p->next_same_hash)
1351 if (mode == p->mode && (x == p->exp || exp_equiv_p (x, p->exp, 0, 0)))
1352 return p;
1355 return 0;
1358 /* Look for an expression equivalent to X and with code CODE.
1359 If one is found, return that expression. */
1361 static rtx
1362 lookup_as_function (rtx x, enum rtx_code code)
1364 struct table_elt *p
1365 = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, GET_MODE (x));
1367 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1368 long as we are narrowing. So if we looked in vain for a mode narrower
1369 than word_mode before, look for word_mode now. */
1370 if (p == 0 && code == CONST_INT
1371 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1373 x = copy_rtx (x);
1374 PUT_MODE (x, word_mode);
1375 p = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, word_mode);
1378 if (p == 0)
1379 return 0;
1381 for (p = p->first_same_value; p; p = p->next_same_value)
1382 if (GET_CODE (p->exp) == code
1383 /* Make sure this is a valid entry in the table. */
1384 && exp_equiv_p (p->exp, p->exp, 1, 0))
1385 return p->exp;
1387 return 0;
1390 /* Insert X in the hash table, assuming HASH is its hash code
1391 and CLASSP is an element of the class it should go in
1392 (or 0 if a new class should be made).
1393 It is inserted at the proper position to keep the class in
1394 the order cheapest first.
1396 MODE is the machine-mode of X, or if X is an integer constant
1397 with VOIDmode then MODE is the mode with which X will be used.
1399 For elements of equal cheapness, the most recent one
1400 goes in front, except that the first element in the list
1401 remains first unless a cheaper element is added. The order of
1402 pseudo-registers does not matter, as canon_reg will be called to
1403 find the cheapest when a register is retrieved from the table.
1405 The in_memory field in the hash table element is set to 0.
1406 The caller must set it nonzero if appropriate.
1408 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1409 and if insert_regs returns a nonzero value
1410 you must then recompute its hash code before calling here.
1412 If necessary, update table showing constant values of quantities. */
1414 #define CHEAPER(X, Y) \
1415 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1417 static struct table_elt *
1418 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1420 struct table_elt *elt;
1422 /* If X is a register and we haven't made a quantity for it,
1423 something is wrong. */
1424 if (GET_CODE (x) == REG && ! REGNO_QTY_VALID_P (REGNO (x)))
1425 abort ();
1427 /* If X is a hard register, show it is being put in the table. */
1428 if (GET_CODE (x) == REG && REGNO (x) < FIRST_PSEUDO_REGISTER)
1430 unsigned int regno = REGNO (x);
1431 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1432 unsigned int i;
1434 for (i = regno; i < endregno; i++)
1435 SET_HARD_REG_BIT (hard_regs_in_table, i);
1438 /* Put an element for X into the right hash bucket. */
1440 elt = free_element_chain;
1441 if (elt)
1442 free_element_chain = elt->next_same_hash;
1443 else
1445 n_elements_made++;
1446 elt = xmalloc (sizeof (struct table_elt));
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)
1461 /* GNU C++ takes advantage of this for `this'
1462 (and other const values). */
1463 || (GET_CODE (x) == REG
1464 && RTX_UNCHANGING_P (x)
1465 && REGNO (x) >= FIRST_PSEUDO_REGISTER)
1466 || fixed_base_plus_p (x));
1468 if (table[hash])
1469 table[hash]->prev_same_hash = elt;
1470 table[hash] = elt;
1472 /* Put it into the proper value-class. */
1473 if (classp)
1475 classp = classp->first_same_value;
1476 if (CHEAPER (elt, classp))
1477 /* Insert at the head of the class. */
1479 struct table_elt *p;
1480 elt->next_same_value = classp;
1481 classp->prev_same_value = elt;
1482 elt->first_same_value = elt;
1484 for (p = classp; p; p = p->next_same_value)
1485 p->first_same_value = elt;
1487 else
1489 /* Insert not at head of the class. */
1490 /* Put it after the last element cheaper than X. */
1491 struct table_elt *p, *next;
1493 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1494 p = next);
1496 /* Put it after P and before NEXT. */
1497 elt->next_same_value = next;
1498 if (next)
1499 next->prev_same_value = elt;
1501 elt->prev_same_value = p;
1502 p->next_same_value = elt;
1503 elt->first_same_value = classp;
1506 else
1507 elt->first_same_value = elt;
1509 /* If this is a constant being set equivalent to a register or a register
1510 being set equivalent to a constant, note the constant equivalence.
1512 If this is a constant, it cannot be equivalent to a different constant,
1513 and a constant is the only thing that can be cheaper than a register. So
1514 we know the register is the head of the class (before the constant was
1515 inserted).
1517 If this is a register that is not already known equivalent to a
1518 constant, we must check the entire class.
1520 If this is a register that is already known equivalent to an insn,
1521 update the qtys `const_insn' to show that `this_insn' is the latest
1522 insn making that quantity equivalent to the constant. */
1524 if (elt->is_const && classp && GET_CODE (classp->exp) == REG
1525 && GET_CODE (x) != REG)
1527 int exp_q = REG_QTY (REGNO (classp->exp));
1528 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1530 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1531 exp_ent->const_insn = this_insn;
1534 else if (GET_CODE (x) == REG
1535 && classp
1536 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1537 && ! elt->is_const)
1539 struct table_elt *p;
1541 for (p = classp; p != 0; p = p->next_same_value)
1543 if (p->is_const && GET_CODE (p->exp) != REG)
1545 int x_q = REG_QTY (REGNO (x));
1546 struct qty_table_elem *x_ent = &qty_table[x_q];
1548 x_ent->const_rtx
1549 = gen_lowpart (GET_MODE (x), p->exp);
1550 x_ent->const_insn = this_insn;
1551 break;
1556 else if (GET_CODE (x) == REG
1557 && qty_table[REG_QTY (REGNO (x))].const_rtx
1558 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1559 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1561 /* If this is a constant with symbolic value,
1562 and it has a term with an explicit integer value,
1563 link it up with related expressions. */
1564 if (GET_CODE (x) == CONST)
1566 rtx subexp = get_related_value (x);
1567 unsigned subhash;
1568 struct table_elt *subelt, *subelt_prev;
1570 if (subexp != 0)
1572 /* Get the integer-free subexpression in the hash table. */
1573 subhash = safe_hash (subexp, mode) & HASH_MASK;
1574 subelt = lookup (subexp, subhash, mode);
1575 if (subelt == 0)
1576 subelt = insert (subexp, NULL, subhash, mode);
1577 /* Initialize SUBELT's circular chain if it has none. */
1578 if (subelt->related_value == 0)
1579 subelt->related_value = subelt;
1580 /* Find the element in the circular chain that precedes SUBELT. */
1581 subelt_prev = subelt;
1582 while (subelt_prev->related_value != subelt)
1583 subelt_prev = subelt_prev->related_value;
1584 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1585 This way the element that follows SUBELT is the oldest one. */
1586 elt->related_value = subelt_prev->related_value;
1587 subelt_prev->related_value = elt;
1591 return elt;
1594 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1595 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1596 the two classes equivalent.
1598 CLASS1 will be the surviving class; CLASS2 should not be used after this
1599 call.
1601 Any invalid entries in CLASS2 will not be copied. */
1603 static void
1604 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1606 struct table_elt *elt, *next, *new;
1608 /* Ensure we start with the head of the classes. */
1609 class1 = class1->first_same_value;
1610 class2 = class2->first_same_value;
1612 /* If they were already equal, forget it. */
1613 if (class1 == class2)
1614 return;
1616 for (elt = class2; elt; elt = next)
1618 unsigned int hash;
1619 rtx exp = elt->exp;
1620 enum machine_mode mode = elt->mode;
1622 next = elt->next_same_value;
1624 /* Remove old entry, make a new one in CLASS1's class.
1625 Don't do this for invalid entries as we cannot find their
1626 hash code (it also isn't necessary). */
1627 if (GET_CODE (exp) == REG || exp_equiv_p (exp, exp, 1, 0))
1629 hash_arg_in_memory = 0;
1630 hash = HASH (exp, mode);
1632 if (GET_CODE (exp) == REG)
1633 delete_reg_equiv (REGNO (exp));
1635 remove_from_table (elt, hash);
1637 if (insert_regs (exp, class1, 0))
1639 rehash_using_reg (exp);
1640 hash = HASH (exp, mode);
1642 new = insert (exp, class1, hash, mode);
1643 new->in_memory = hash_arg_in_memory;
1648 /* Flush the entire hash table. */
1650 static void
1651 flush_hash_table (void)
1653 int i;
1654 struct table_elt *p;
1656 for (i = 0; i < HASH_SIZE; i++)
1657 for (p = table[i]; p; p = table[i])
1659 /* Note that invalidate can remove elements
1660 after P in the current hash chain. */
1661 if (GET_CODE (p->exp) == REG)
1662 invalidate (p->exp, p->mode);
1663 else
1664 remove_from_table (p, i);
1668 /* Function called for each rtx to check whether true dependence exist. */
1669 struct check_dependence_data
1671 enum machine_mode mode;
1672 rtx exp;
1673 rtx addr;
1676 static int
1677 check_dependence (rtx *x, void *data)
1679 struct check_dependence_data *d = (struct check_dependence_data *) data;
1680 if (*x && GET_CODE (*x) == MEM)
1681 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1682 cse_rtx_varies_p);
1683 else
1684 return 0;
1687 /* Remove from the hash table, or mark as invalid, all expressions whose
1688 values could be altered by storing in X. X is a register, a subreg, or
1689 a memory reference with nonvarying address (because, when a memory
1690 reference with a varying address is stored in, all memory references are
1691 removed by invalidate_memory so specific invalidation is superfluous).
1692 FULL_MODE, if not VOIDmode, indicates that this much should be
1693 invalidated instead of just the amount indicated by the mode of X. This
1694 is only used for bitfield stores into memory.
1696 A nonvarying address may be just a register or just a symbol reference,
1697 or it may be either of those plus a numeric offset. */
1699 static void
1700 invalidate (rtx x, enum machine_mode full_mode)
1702 int i;
1703 struct table_elt *p;
1704 rtx addr;
1706 switch (GET_CODE (x))
1708 case REG:
1710 /* If X is a register, dependencies on its contents are recorded
1711 through the qty number mechanism. Just change the qty number of
1712 the register, mark it as invalid for expressions that refer to it,
1713 and remove it itself. */
1714 unsigned int regno = REGNO (x);
1715 unsigned int hash = HASH (x, GET_MODE (x));
1717 /* Remove REGNO from any quantity list it might be on and indicate
1718 that its value might have changed. If it is a pseudo, remove its
1719 entry from the hash table.
1721 For a hard register, we do the first two actions above for any
1722 additional hard registers corresponding to X. Then, if any of these
1723 registers are in the table, we must remove any REG entries that
1724 overlap these registers. */
1726 delete_reg_equiv (regno);
1727 REG_TICK (regno)++;
1728 SUBREG_TICKED (regno) = -1;
1730 if (regno >= FIRST_PSEUDO_REGISTER)
1732 /* Because a register can be referenced in more than one mode,
1733 we might have to remove more than one table entry. */
1734 struct table_elt *elt;
1736 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1737 remove_from_table (elt, hash);
1739 else
1741 HOST_WIDE_INT in_table
1742 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1743 unsigned int endregno
1744 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1745 unsigned int tregno, tendregno, rn;
1746 struct table_elt *p, *next;
1748 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1750 for (rn = regno + 1; rn < endregno; rn++)
1752 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1753 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1754 delete_reg_equiv (rn);
1755 REG_TICK (rn)++;
1756 SUBREG_TICKED (rn) = -1;
1759 if (in_table)
1760 for (hash = 0; hash < HASH_SIZE; hash++)
1761 for (p = table[hash]; p; p = next)
1763 next = p->next_same_hash;
1765 if (GET_CODE (p->exp) != REG
1766 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1767 continue;
1769 tregno = REGNO (p->exp);
1770 tendregno
1771 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1772 if (tendregno > regno && tregno < endregno)
1773 remove_from_table (p, hash);
1777 return;
1779 case SUBREG:
1780 invalidate (SUBREG_REG (x), VOIDmode);
1781 return;
1783 case PARALLEL:
1784 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1785 invalidate (XVECEXP (x, 0, i), VOIDmode);
1786 return;
1788 case EXPR_LIST:
1789 /* This is part of a disjoint return value; extract the location in
1790 question ignoring the offset. */
1791 invalidate (XEXP (x, 0), VOIDmode);
1792 return;
1794 case MEM:
1795 addr = canon_rtx (get_addr (XEXP (x, 0)));
1796 /* Calculate the canonical version of X here so that
1797 true_dependence doesn't generate new RTL for X on each call. */
1798 x = canon_rtx (x);
1800 /* Remove all hash table elements that refer to overlapping pieces of
1801 memory. */
1802 if (full_mode == VOIDmode)
1803 full_mode = GET_MODE (x);
1805 for (i = 0; i < HASH_SIZE; i++)
1807 struct table_elt *next;
1809 for (p = table[i]; p; p = next)
1811 next = p->next_same_hash;
1812 if (p->in_memory)
1814 struct check_dependence_data d;
1816 /* Just canonicalize the expression once;
1817 otherwise each time we call invalidate
1818 true_dependence will canonicalize the
1819 expression again. */
1820 if (!p->canon_exp)
1821 p->canon_exp = canon_rtx (p->exp);
1822 d.exp = x;
1823 d.addr = addr;
1824 d.mode = full_mode;
1825 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1826 remove_from_table (p, i);
1830 return;
1832 default:
1833 abort ();
1837 /* Remove all expressions that refer to register REGNO,
1838 since they are already invalid, and we are about to
1839 mark that register valid again and don't want the old
1840 expressions to reappear as valid. */
1842 static void
1843 remove_invalid_refs (unsigned int regno)
1845 unsigned int i;
1846 struct table_elt *p, *next;
1848 for (i = 0; i < HASH_SIZE; i++)
1849 for (p = table[i]; p; p = next)
1851 next = p->next_same_hash;
1852 if (GET_CODE (p->exp) != REG
1853 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1854 remove_from_table (p, i);
1858 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1859 and mode MODE. */
1860 static void
1861 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1862 enum machine_mode mode)
1864 unsigned int i;
1865 struct table_elt *p, *next;
1866 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1868 for (i = 0; i < HASH_SIZE; i++)
1869 for (p = table[i]; p; p = next)
1871 rtx exp = p->exp;
1872 next = p->next_same_hash;
1874 if (GET_CODE (exp) != REG
1875 && (GET_CODE (exp) != SUBREG
1876 || GET_CODE (SUBREG_REG (exp)) != REG
1877 || REGNO (SUBREG_REG (exp)) != regno
1878 || (((SUBREG_BYTE (exp)
1879 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1880 && SUBREG_BYTE (exp) <= end))
1881 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1882 remove_from_table (p, i);
1886 /* Recompute the hash codes of any valid entries in the hash table that
1887 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1889 This is called when we make a jump equivalence. */
1891 static void
1892 rehash_using_reg (rtx x)
1894 unsigned int i;
1895 struct table_elt *p, *next;
1896 unsigned hash;
1898 if (GET_CODE (x) == SUBREG)
1899 x = SUBREG_REG (x);
1901 /* If X is not a register or if the register is known not to be in any
1902 valid entries in the table, we have no work to do. */
1904 if (GET_CODE (x) != REG
1905 || REG_IN_TABLE (REGNO (x)) < 0
1906 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1907 return;
1909 /* Scan all hash chains looking for valid entries that mention X.
1910 If we find one and it is in the wrong hash chain, move it. We can skip
1911 objects that are registers, since they are handled specially. */
1913 for (i = 0; i < HASH_SIZE; i++)
1914 for (p = table[i]; p; p = next)
1916 next = p->next_same_hash;
1917 if (GET_CODE (p->exp) != REG && reg_mentioned_p (x, p->exp)
1918 && exp_equiv_p (p->exp, p->exp, 1, 0)
1919 && i != (hash = safe_hash (p->exp, p->mode) & HASH_MASK))
1921 if (p->next_same_hash)
1922 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1924 if (p->prev_same_hash)
1925 p->prev_same_hash->next_same_hash = p->next_same_hash;
1926 else
1927 table[i] = p->next_same_hash;
1929 p->next_same_hash = table[hash];
1930 p->prev_same_hash = 0;
1931 if (table[hash])
1932 table[hash]->prev_same_hash = p;
1933 table[hash] = p;
1938 /* Remove from the hash table any expression that is a call-clobbered
1939 register. Also update their TICK values. */
1941 static void
1942 invalidate_for_call (void)
1944 unsigned int regno, endregno;
1945 unsigned int i;
1946 unsigned hash;
1947 struct table_elt *p, *next;
1948 int in_table = 0;
1950 /* Go through all the hard registers. For each that is clobbered in
1951 a CALL_INSN, remove the register from quantity chains and update
1952 reg_tick if defined. Also see if any of these registers is currently
1953 in the table. */
1955 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1956 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
1958 delete_reg_equiv (regno);
1959 if (REG_TICK (regno) >= 0)
1961 REG_TICK (regno)++;
1962 SUBREG_TICKED (regno) = -1;
1965 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
1968 /* In the case where we have no call-clobbered hard registers in the
1969 table, we are done. Otherwise, scan the table and remove any
1970 entry that overlaps a call-clobbered register. */
1972 if (in_table)
1973 for (hash = 0; hash < HASH_SIZE; hash++)
1974 for (p = table[hash]; p; p = next)
1976 next = p->next_same_hash;
1978 if (GET_CODE (p->exp) != REG
1979 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1980 continue;
1982 regno = REGNO (p->exp);
1983 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
1985 for (i = regno; i < endregno; i++)
1986 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
1988 remove_from_table (p, hash);
1989 break;
1994 /* Given an expression X of type CONST,
1995 and ELT which is its table entry (or 0 if it
1996 is not in the hash table),
1997 return an alternate expression for X as a register plus integer.
1998 If none can be found, return 0. */
2000 static rtx
2001 use_related_value (rtx x, struct table_elt *elt)
2003 struct table_elt *relt = 0;
2004 struct table_elt *p, *q;
2005 HOST_WIDE_INT offset;
2007 /* First, is there anything related known?
2008 If we have a table element, we can tell from that.
2009 Otherwise, must look it up. */
2011 if (elt != 0 && elt->related_value != 0)
2012 relt = elt;
2013 else if (elt == 0 && GET_CODE (x) == CONST)
2015 rtx subexp = get_related_value (x);
2016 if (subexp != 0)
2017 relt = lookup (subexp,
2018 safe_hash (subexp, GET_MODE (subexp)) & HASH_MASK,
2019 GET_MODE (subexp));
2022 if (relt == 0)
2023 return 0;
2025 /* Search all related table entries for one that has an
2026 equivalent register. */
2028 p = relt;
2029 while (1)
2031 /* This loop is strange in that it is executed in two different cases.
2032 The first is when X is already in the table. Then it is searching
2033 the RELATED_VALUE list of X's class (RELT). The second case is when
2034 X is not in the table. Then RELT points to a class for the related
2035 value.
2037 Ensure that, whatever case we are in, that we ignore classes that have
2038 the same value as X. */
2040 if (rtx_equal_p (x, p->exp))
2041 q = 0;
2042 else
2043 for (q = p->first_same_value; q; q = q->next_same_value)
2044 if (GET_CODE (q->exp) == REG)
2045 break;
2047 if (q)
2048 break;
2050 p = p->related_value;
2052 /* We went all the way around, so there is nothing to be found.
2053 Alternatively, perhaps RELT was in the table for some other reason
2054 and it has no related values recorded. */
2055 if (p == relt || p == 0)
2056 break;
2059 if (q == 0)
2060 return 0;
2062 offset = (get_integer_term (x) - get_integer_term (p->exp));
2063 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2064 return plus_constant (q->exp, offset);
2067 /* Hash a string. Just add its bytes up. */
2068 static inline unsigned
2069 canon_hash_string (const char *ps)
2071 unsigned hash = 0;
2072 const unsigned char *p = (const unsigned char *) ps;
2074 if (p)
2075 while (*p)
2076 hash += *p++;
2078 return hash;
2081 /* Hash an rtx. We are careful to make sure the value is never negative.
2082 Equivalent registers hash identically.
2083 MODE is used in hashing for CONST_INTs only;
2084 otherwise the mode of X is used.
2086 Store 1 in do_not_record if any subexpression is volatile.
2088 Store 1 in hash_arg_in_memory if X contains a MEM rtx
2089 which does not have the RTX_UNCHANGING_P bit set.
2091 Note that cse_insn knows that the hash code of a MEM expression
2092 is just (int) MEM plus the hash code of the address. */
2094 static unsigned
2095 canon_hash (rtx x, enum machine_mode mode)
2097 int i, j;
2098 unsigned hash = 0;
2099 enum rtx_code code;
2100 const char *fmt;
2102 /* repeat is used to turn tail-recursion into iteration. */
2103 repeat:
2104 if (x == 0)
2105 return hash;
2107 code = GET_CODE (x);
2108 switch (code)
2110 case REG:
2112 unsigned int regno = REGNO (x);
2113 bool record;
2115 /* On some machines, we can't record any non-fixed hard register,
2116 because extending its life will cause reload problems. We
2117 consider ap, fp, sp, gp to be fixed for this purpose.
2119 We also consider CCmode registers to be fixed for this purpose;
2120 failure to do so leads to failure to simplify 0<100 type of
2121 conditionals.
2123 On all machines, we can't record any global registers.
2124 Nor should we record any register that is in a small
2125 class, as defined by CLASS_LIKELY_SPILLED_P. */
2127 if (regno >= FIRST_PSEUDO_REGISTER)
2128 record = true;
2129 else if (x == frame_pointer_rtx
2130 || x == hard_frame_pointer_rtx
2131 || x == arg_pointer_rtx
2132 || x == stack_pointer_rtx
2133 || x == pic_offset_table_rtx)
2134 record = true;
2135 else if (global_regs[regno])
2136 record = false;
2137 else if (fixed_regs[regno])
2138 record = true;
2139 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2140 record = true;
2141 else if (SMALL_REGISTER_CLASSES)
2142 record = false;
2143 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2144 record = false;
2145 else
2146 record = true;
2148 if (!record)
2150 do_not_record = 1;
2151 return 0;
2154 hash += ((unsigned) REG << 7) + (unsigned) REG_QTY (regno);
2155 return hash;
2158 /* We handle SUBREG of a REG specially because the underlying
2159 reg changes its hash value with every value change; we don't
2160 want to have to forget unrelated subregs when one subreg changes. */
2161 case SUBREG:
2163 if (GET_CODE (SUBREG_REG (x)) == REG)
2165 hash += (((unsigned) SUBREG << 7)
2166 + REGNO (SUBREG_REG (x))
2167 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2168 return hash;
2170 break;
2173 case CONST_INT:
2175 unsigned HOST_WIDE_INT tem = INTVAL (x);
2176 hash += ((unsigned) CONST_INT << 7) + (unsigned) mode + tem;
2177 return hash;
2180 case CONST_DOUBLE:
2181 /* This is like the general case, except that it only counts
2182 the integers representing the constant. */
2183 hash += (unsigned) code + (unsigned) GET_MODE (x);
2184 if (GET_MODE (x) != VOIDmode)
2185 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2186 else
2187 hash += ((unsigned) CONST_DOUBLE_LOW (x)
2188 + (unsigned) CONST_DOUBLE_HIGH (x));
2189 return hash;
2191 case CONST_VECTOR:
2193 int units;
2194 rtx elt;
2196 units = CONST_VECTOR_NUNITS (x);
2198 for (i = 0; i < units; ++i)
2200 elt = CONST_VECTOR_ELT (x, i);
2201 hash += canon_hash (elt, GET_MODE (elt));
2204 return hash;
2207 /* Assume there is only one rtx object for any given label. */
2208 case LABEL_REF:
2209 hash += ((unsigned) LABEL_REF << 7) + (unsigned long) XEXP (x, 0);
2210 return hash;
2212 case SYMBOL_REF:
2213 hash += ((unsigned) SYMBOL_REF << 7) + (unsigned long) XSTR (x, 0);
2214 return hash;
2216 case MEM:
2217 /* We don't record if marked volatile or if BLKmode since we don't
2218 know the size of the move. */
2219 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2221 do_not_record = 1;
2222 return 0;
2224 if (! RTX_UNCHANGING_P (x) || fixed_base_plus_p (XEXP (x, 0)))
2225 hash_arg_in_memory = 1;
2227 /* Now that we have already found this special case,
2228 might as well speed it up as much as possible. */
2229 hash += (unsigned) MEM;
2230 x = XEXP (x, 0);
2231 goto repeat;
2233 case USE:
2234 /* A USE that mentions non-volatile memory needs special
2235 handling since the MEM may be BLKmode which normally
2236 prevents an entry from being made. Pure calls are
2237 marked by a USE which mentions BLKmode memory. */
2238 if (GET_CODE (XEXP (x, 0)) == MEM
2239 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2241 hash += (unsigned) USE;
2242 x = XEXP (x, 0);
2244 if (! RTX_UNCHANGING_P (x) || fixed_base_plus_p (XEXP (x, 0)))
2245 hash_arg_in_memory = 1;
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;
2253 break;
2255 case PRE_DEC:
2256 case PRE_INC:
2257 case POST_DEC:
2258 case POST_INC:
2259 case PRE_MODIFY:
2260 case POST_MODIFY:
2261 case PC:
2262 case CC0:
2263 case CALL:
2264 case UNSPEC_VOLATILE:
2265 do_not_record = 1;
2266 return 0;
2268 case ASM_OPERANDS:
2269 if (MEM_VOLATILE_P (x))
2271 do_not_record = 1;
2272 return 0;
2274 else
2276 /* We don't want to take the filename and line into account. */
2277 hash += (unsigned) code + (unsigned) GET_MODE (x)
2278 + canon_hash_string (ASM_OPERANDS_TEMPLATE (x))
2279 + canon_hash_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2280 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2282 if (ASM_OPERANDS_INPUT_LENGTH (x))
2284 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2286 hash += (canon_hash (ASM_OPERANDS_INPUT (x, i),
2287 GET_MODE (ASM_OPERANDS_INPUT (x, i)))
2288 + canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT
2289 (x, i)));
2292 hash += canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2293 x = ASM_OPERANDS_INPUT (x, 0);
2294 mode = GET_MODE (x);
2295 goto repeat;
2298 return hash;
2300 break;
2302 default:
2303 break;
2306 i = GET_RTX_LENGTH (code) - 1;
2307 hash += (unsigned) code + (unsigned) GET_MODE (x);
2308 fmt = GET_RTX_FORMAT (code);
2309 for (; i >= 0; i--)
2311 if (fmt[i] == 'e')
2313 rtx tem = XEXP (x, i);
2315 /* If we are about to do the last recursive call
2316 needed at this level, change it into iteration.
2317 This function is called enough to be worth it. */
2318 if (i == 0)
2320 x = tem;
2321 goto repeat;
2323 hash += canon_hash (tem, 0);
2325 else if (fmt[i] == 'E')
2326 for (j = 0; j < XVECLEN (x, i); j++)
2327 hash += canon_hash (XVECEXP (x, i, j), 0);
2328 else if (fmt[i] == 's')
2329 hash += canon_hash_string (XSTR (x, i));
2330 else if (fmt[i] == 'i')
2332 unsigned tem = XINT (x, i);
2333 hash += tem;
2335 else if (fmt[i] == '0' || fmt[i] == 't')
2336 /* Unused. */
2338 else
2339 abort ();
2341 return hash;
2344 /* Like canon_hash but with no side effects. */
2346 static unsigned
2347 safe_hash (rtx x, enum machine_mode mode)
2349 int save_do_not_record = do_not_record;
2350 int save_hash_arg_in_memory = hash_arg_in_memory;
2351 unsigned hash = canon_hash (x, mode);
2352 hash_arg_in_memory = save_hash_arg_in_memory;
2353 do_not_record = save_do_not_record;
2354 return hash;
2357 /* Return 1 iff X and Y would canonicalize into the same thing,
2358 without actually constructing the canonicalization of either one.
2359 If VALIDATE is nonzero,
2360 we assume X is an expression being processed from the rtl
2361 and Y was found in the hash table. We check register refs
2362 in Y for being marked as valid.
2364 If EQUAL_VALUES is nonzero, we allow a register to match a constant value
2365 that is known to be in the register. Ordinarily, we don't allow them
2366 to match, because letting them match would cause unpredictable results
2367 in all the places that search a hash table chain for an equivalent
2368 for a given value. A possible equivalent that has different structure
2369 has its hash code computed from different data. Whether the hash code
2370 is the same as that of the given value is pure luck. */
2372 static int
2373 exp_equiv_p (rtx x, rtx y, int validate, int equal_values)
2375 int i, j;
2376 enum rtx_code code;
2377 const char *fmt;
2379 /* Note: it is incorrect to assume an expression is equivalent to itself
2380 if VALIDATE is nonzero. */
2381 if (x == y && !validate)
2382 return 1;
2383 if (x == 0 || y == 0)
2384 return x == y;
2386 code = GET_CODE (x);
2387 if (code != GET_CODE (y))
2389 if (!equal_values)
2390 return 0;
2392 /* If X is a constant and Y is a register or vice versa, they may be
2393 equivalent. We only have to validate if Y is a register. */
2394 if (CONSTANT_P (x) && GET_CODE (y) == REG
2395 && REGNO_QTY_VALID_P (REGNO (y)))
2397 int y_q = REG_QTY (REGNO (y));
2398 struct qty_table_elem *y_ent = &qty_table[y_q];
2400 if (GET_MODE (y) == y_ent->mode
2401 && rtx_equal_p (x, y_ent->const_rtx)
2402 && (! validate || REG_IN_TABLE (REGNO (y)) == REG_TICK (REGNO (y))))
2403 return 1;
2406 if (CONSTANT_P (y) && code == REG
2407 && REGNO_QTY_VALID_P (REGNO (x)))
2409 int x_q = REG_QTY (REGNO (x));
2410 struct qty_table_elem *x_ent = &qty_table[x_q];
2412 if (GET_MODE (x) == x_ent->mode
2413 && rtx_equal_p (y, x_ent->const_rtx))
2414 return 1;
2417 return 0;
2420 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2421 if (GET_MODE (x) != GET_MODE (y))
2422 return 0;
2424 switch (code)
2426 case PC:
2427 case CC0:
2428 case CONST_INT:
2429 return x == y;
2431 case LABEL_REF:
2432 return XEXP (x, 0) == XEXP (y, 0);
2434 case SYMBOL_REF:
2435 return XSTR (x, 0) == XSTR (y, 0);
2437 case REG:
2439 unsigned int regno = REGNO (y);
2440 unsigned int endregno
2441 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2442 : hard_regno_nregs[regno][GET_MODE (y)]);
2443 unsigned int i;
2445 /* If the quantities are not the same, the expressions are not
2446 equivalent. If there are and we are not to validate, they
2447 are equivalent. Otherwise, ensure all regs are up-to-date. */
2449 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2450 return 0;
2452 if (! validate)
2453 return 1;
2455 for (i = regno; i < endregno; i++)
2456 if (REG_IN_TABLE (i) != REG_TICK (i))
2457 return 0;
2459 return 1;
2462 /* For commutative operations, check both orders. */
2463 case PLUS:
2464 case MULT:
2465 case AND:
2466 case IOR:
2467 case XOR:
2468 case NE:
2469 case EQ:
2470 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0), validate, equal_values)
2471 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2472 validate, equal_values))
2473 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2474 validate, equal_values)
2475 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2476 validate, equal_values)));
2478 case ASM_OPERANDS:
2479 /* We don't use the generic code below because we want to
2480 disregard filename and line numbers. */
2482 /* A volatile asm isn't equivalent to any other. */
2483 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2484 return 0;
2486 if (GET_MODE (x) != GET_MODE (y)
2487 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2488 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2489 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2490 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2491 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2492 return 0;
2494 if (ASM_OPERANDS_INPUT_LENGTH (x))
2496 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2497 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2498 ASM_OPERANDS_INPUT (y, i),
2499 validate, equal_values)
2500 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2501 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2502 return 0;
2505 return 1;
2507 default:
2508 break;
2511 /* Compare the elements. If any pair of corresponding elements
2512 fail to match, return 0 for the whole things. */
2514 fmt = GET_RTX_FORMAT (code);
2515 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2517 switch (fmt[i])
2519 case 'e':
2520 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i), validate, equal_values))
2521 return 0;
2522 break;
2524 case 'E':
2525 if (XVECLEN (x, i) != XVECLEN (y, i))
2526 return 0;
2527 for (j = 0; j < XVECLEN (x, i); j++)
2528 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2529 validate, equal_values))
2530 return 0;
2531 break;
2533 case 's':
2534 if (strcmp (XSTR (x, i), XSTR (y, i)))
2535 return 0;
2536 break;
2538 case 'i':
2539 if (XINT (x, i) != XINT (y, i))
2540 return 0;
2541 break;
2543 case 'w':
2544 if (XWINT (x, i) != XWINT (y, i))
2545 return 0;
2546 break;
2548 case '0':
2549 case 't':
2550 break;
2552 default:
2553 abort ();
2557 return 1;
2560 /* Return 1 if X has a value that can vary even between two
2561 executions of the program. 0 means X can be compared reliably
2562 against certain constants or near-constants. */
2564 static int
2565 cse_rtx_varies_p (rtx x, int from_alias)
2567 /* We need not check for X and the equivalence class being of the same
2568 mode because if X is equivalent to a constant in some mode, it
2569 doesn't vary in any mode. */
2571 if (GET_CODE (x) == REG
2572 && REGNO_QTY_VALID_P (REGNO (x)))
2574 int x_q = REG_QTY (REGNO (x));
2575 struct qty_table_elem *x_ent = &qty_table[x_q];
2577 if (GET_MODE (x) == x_ent->mode
2578 && x_ent->const_rtx != NULL_RTX)
2579 return 0;
2582 if (GET_CODE (x) == PLUS
2583 && GET_CODE (XEXP (x, 1)) == CONST_INT
2584 && GET_CODE (XEXP (x, 0)) == REG
2585 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2587 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2588 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2590 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2591 && x0_ent->const_rtx != NULL_RTX)
2592 return 0;
2595 /* This can happen as the result of virtual register instantiation, if
2596 the initial constant is too large to be a valid address. This gives
2597 us a three instruction sequence, load large offset into a register,
2598 load fp minus a constant into a register, then a MEM which is the
2599 sum of the two `constant' registers. */
2600 if (GET_CODE (x) == PLUS
2601 && GET_CODE (XEXP (x, 0)) == REG
2602 && GET_CODE (XEXP (x, 1)) == REG
2603 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2604 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2606 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2607 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2608 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2609 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2611 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2612 && x0_ent->const_rtx != NULL_RTX
2613 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2614 && x1_ent->const_rtx != NULL_RTX)
2615 return 0;
2618 return rtx_varies_p (x, from_alias);
2621 /* Canonicalize an expression:
2622 replace each register reference inside it
2623 with the "oldest" equivalent register.
2625 If INSN is nonzero and we are replacing a pseudo with a hard register
2626 or vice versa, validate_change is used to ensure that INSN remains valid
2627 after we make our substitution. The calls are made with IN_GROUP nonzero
2628 so apply_change_group must be called upon the outermost return from this
2629 function (unless INSN is zero). The result of apply_change_group can
2630 generally be discarded since the changes we are making are optional. */
2632 static rtx
2633 canon_reg (rtx x, rtx insn)
2635 int i;
2636 enum rtx_code code;
2637 const char *fmt;
2639 if (x == 0)
2640 return x;
2642 code = GET_CODE (x);
2643 switch (code)
2645 case PC:
2646 case CC0:
2647 case CONST:
2648 case CONST_INT:
2649 case CONST_DOUBLE:
2650 case CONST_VECTOR:
2651 case SYMBOL_REF:
2652 case LABEL_REF:
2653 case ADDR_VEC:
2654 case ADDR_DIFF_VEC:
2655 return x;
2657 case REG:
2659 int first;
2660 int q;
2661 struct qty_table_elem *ent;
2663 /* Never replace a hard reg, because hard regs can appear
2664 in more than one machine mode, and we must preserve the mode
2665 of each occurrence. Also, some hard regs appear in
2666 MEMs that are shared and mustn't be altered. Don't try to
2667 replace any reg that maps to a reg of class NO_REGS. */
2668 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2669 || ! REGNO_QTY_VALID_P (REGNO (x)))
2670 return x;
2672 q = REG_QTY (REGNO (x));
2673 ent = &qty_table[q];
2674 first = ent->first_reg;
2675 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2676 : REGNO_REG_CLASS (first) == NO_REGS ? x
2677 : gen_rtx_REG (ent->mode, first));
2680 default:
2681 break;
2684 fmt = GET_RTX_FORMAT (code);
2685 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2687 int j;
2689 if (fmt[i] == 'e')
2691 rtx new = canon_reg (XEXP (x, i), insn);
2692 int insn_code;
2694 /* If replacing pseudo with hard reg or vice versa, ensure the
2695 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2696 if (insn != 0 && new != 0
2697 && GET_CODE (new) == REG && GET_CODE (XEXP (x, i)) == REG
2698 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2699 != (REGNO (XEXP (x, i)) < FIRST_PSEUDO_REGISTER))
2700 || (insn_code = recog_memoized (insn)) < 0
2701 || insn_data[insn_code].n_dups > 0))
2702 validate_change (insn, &XEXP (x, i), new, 1);
2703 else
2704 XEXP (x, i) = new;
2706 else if (fmt[i] == 'E')
2707 for (j = 0; j < XVECLEN (x, i); j++)
2708 XVECEXP (x, i, j) = canon_reg (XVECEXP (x, i, j), insn);
2711 return x;
2714 /* LOC is a location within INSN that is an operand address (the contents of
2715 a MEM). Find the best equivalent address to use that is valid for this
2716 insn.
2718 On most CISC machines, complicated address modes are costly, and rtx_cost
2719 is a good approximation for that cost. However, most RISC machines have
2720 only a few (usually only one) memory reference formats. If an address is
2721 valid at all, it is often just as cheap as any other address. Hence, for
2722 RISC machines, we use `address_cost' to compare the costs of various
2723 addresses. For two addresses of equal cost, choose the one with the
2724 highest `rtx_cost' value as that has the potential of eliminating the
2725 most insns. For equal costs, we choose the first in the equivalence
2726 class. Note that we ignore the fact that pseudo registers are cheaper than
2727 hard registers here because we would also prefer the pseudo registers. */
2729 static void
2730 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2732 struct table_elt *elt;
2733 rtx addr = *loc;
2734 struct table_elt *p;
2735 int found_better = 1;
2736 int save_do_not_record = do_not_record;
2737 int save_hash_arg_in_memory = hash_arg_in_memory;
2738 int addr_volatile;
2739 int regno;
2740 unsigned hash;
2742 /* Do not try to replace constant addresses or addresses of local and
2743 argument slots. These MEM expressions are made only once and inserted
2744 in many instructions, as well as being used to control symbol table
2745 output. It is not safe to clobber them.
2747 There are some uncommon cases where the address is already in a register
2748 for some reason, but we cannot take advantage of that because we have
2749 no easy way to unshare the MEM. In addition, looking up all stack
2750 addresses is costly. */
2751 if ((GET_CODE (addr) == PLUS
2752 && GET_CODE (XEXP (addr, 0)) == REG
2753 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2754 && (regno = REGNO (XEXP (addr, 0)),
2755 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2756 || regno == ARG_POINTER_REGNUM))
2757 || (GET_CODE (addr) == REG
2758 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2759 || regno == HARD_FRAME_POINTER_REGNUM
2760 || regno == ARG_POINTER_REGNUM))
2761 || GET_CODE (addr) == ADDRESSOF
2762 || CONSTANT_ADDRESS_P (addr))
2763 return;
2765 /* If this address is not simply a register, try to fold it. This will
2766 sometimes simplify the expression. Many simplifications
2767 will not be valid, but some, usually applying the associative rule, will
2768 be valid and produce better code. */
2769 if (GET_CODE (addr) != REG)
2771 rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX);
2772 int addr_folded_cost = address_cost (folded, mode);
2773 int addr_cost = address_cost (addr, mode);
2775 if ((addr_folded_cost < addr_cost
2776 || (addr_folded_cost == addr_cost
2777 /* ??? The rtx_cost comparison is left over from an older
2778 version of this code. It is probably no longer helpful. */
2779 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2780 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2781 && validate_change (insn, loc, folded, 0))
2782 addr = folded;
2785 /* If this address is not in the hash table, we can't look for equivalences
2786 of the whole address. Also, ignore if volatile. */
2788 do_not_record = 0;
2789 hash = HASH (addr, Pmode);
2790 addr_volatile = do_not_record;
2791 do_not_record = save_do_not_record;
2792 hash_arg_in_memory = save_hash_arg_in_memory;
2794 if (addr_volatile)
2795 return;
2797 elt = lookup (addr, hash, Pmode);
2799 if (elt)
2801 /* We need to find the best (under the criteria documented above) entry
2802 in the class that is valid. We use the `flag' field to indicate
2803 choices that were invalid and iterate until we can't find a better
2804 one that hasn't already been tried. */
2806 for (p = elt->first_same_value; p; p = p->next_same_value)
2807 p->flag = 0;
2809 while (found_better)
2811 int best_addr_cost = address_cost (*loc, mode);
2812 int best_rtx_cost = (elt->cost + 1) >> 1;
2813 int exp_cost;
2814 struct table_elt *best_elt = elt;
2816 found_better = 0;
2817 for (p = elt->first_same_value; p; p = p->next_same_value)
2818 if (! p->flag)
2820 if ((GET_CODE (p->exp) == REG
2821 || exp_equiv_p (p->exp, p->exp, 1, 0))
2822 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2823 || (exp_cost == best_addr_cost
2824 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2826 found_better = 1;
2827 best_addr_cost = exp_cost;
2828 best_rtx_cost = (p->cost + 1) >> 1;
2829 best_elt = p;
2833 if (found_better)
2835 if (validate_change (insn, loc,
2836 canon_reg (copy_rtx (best_elt->exp),
2837 NULL_RTX), 0))
2838 return;
2839 else
2840 best_elt->flag = 1;
2845 /* If the address is a binary operation with the first operand a register
2846 and the second a constant, do the same as above, but looking for
2847 equivalences of the register. Then try to simplify before checking for
2848 the best address to use. This catches a few cases: First is when we
2849 have REG+const and the register is another REG+const. We can often merge
2850 the constants and eliminate one insn and one register. It may also be
2851 that a machine has a cheap REG+REG+const. Finally, this improves the
2852 code on the Alpha for unaligned byte stores. */
2854 if (flag_expensive_optimizations
2855 && ARITHMETIC_P (*loc)
2856 && GET_CODE (XEXP (*loc, 0)) == REG)
2858 rtx op1 = XEXP (*loc, 1);
2860 do_not_record = 0;
2861 hash = HASH (XEXP (*loc, 0), Pmode);
2862 do_not_record = save_do_not_record;
2863 hash_arg_in_memory = save_hash_arg_in_memory;
2865 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2866 if (elt == 0)
2867 return;
2869 /* We need to find the best (under the criteria documented above) entry
2870 in the class that is valid. We use the `flag' field to indicate
2871 choices that were invalid and iterate until we can't find a better
2872 one that hasn't already been tried. */
2874 for (p = elt->first_same_value; p; p = p->next_same_value)
2875 p->flag = 0;
2877 while (found_better)
2879 int best_addr_cost = address_cost (*loc, mode);
2880 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2881 struct table_elt *best_elt = elt;
2882 rtx best_rtx = *loc;
2883 int count;
2885 /* This is at worst case an O(n^2) algorithm, so limit our search
2886 to the first 32 elements on the list. This avoids trouble
2887 compiling code with very long basic blocks that can easily
2888 call simplify_gen_binary so many times that we run out of
2889 memory. */
2891 found_better = 0;
2892 for (p = elt->first_same_value, count = 0;
2893 p && count < 32;
2894 p = p->next_same_value, count++)
2895 if (! p->flag
2896 && (GET_CODE (p->exp) == REG
2897 || exp_equiv_p (p->exp, p->exp, 1, 0)))
2899 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
2900 p->exp, op1);
2901 int new_cost;
2902 new_cost = address_cost (new, mode);
2904 if (new_cost < best_addr_cost
2905 || (new_cost == best_addr_cost
2906 && (COST (new) + 1) >> 1 > best_rtx_cost))
2908 found_better = 1;
2909 best_addr_cost = new_cost;
2910 best_rtx_cost = (COST (new) + 1) >> 1;
2911 best_elt = p;
2912 best_rtx = new;
2916 if (found_better)
2918 if (validate_change (insn, loc,
2919 canon_reg (copy_rtx (best_rtx),
2920 NULL_RTX), 0))
2921 return;
2922 else
2923 best_elt->flag = 1;
2929 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2930 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2931 what values are being compared.
2933 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2934 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2935 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2936 compared to produce cc0.
2938 The return value is the comparison operator and is either the code of
2939 A or the code corresponding to the inverse of the comparison. */
2941 static enum rtx_code
2942 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2943 enum machine_mode *pmode1, enum machine_mode *pmode2)
2945 rtx arg1, arg2;
2947 arg1 = *parg1, arg2 = *parg2;
2949 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2951 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2953 /* Set nonzero when we find something of interest. */
2954 rtx x = 0;
2955 int reverse_code = 0;
2956 struct table_elt *p = 0;
2958 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2959 On machines with CC0, this is the only case that can occur, since
2960 fold_rtx will return the COMPARE or item being compared with zero
2961 when given CC0. */
2963 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2964 x = arg1;
2966 /* If ARG1 is a comparison operator and CODE is testing for
2967 STORE_FLAG_VALUE, get the inner arguments. */
2969 else if (COMPARISON_P (arg1))
2971 #ifdef FLOAT_STORE_FLAG_VALUE
2972 REAL_VALUE_TYPE fsfv;
2973 #endif
2975 if (code == NE
2976 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2977 && code == LT && STORE_FLAG_VALUE == -1)
2978 #ifdef FLOAT_STORE_FLAG_VALUE
2979 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
2980 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2981 REAL_VALUE_NEGATIVE (fsfv)))
2982 #endif
2984 x = arg1;
2985 else if (code == EQ
2986 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2987 && code == GE && STORE_FLAG_VALUE == -1)
2988 #ifdef FLOAT_STORE_FLAG_VALUE
2989 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
2990 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2991 REAL_VALUE_NEGATIVE (fsfv)))
2992 #endif
2994 x = arg1, reverse_code = 1;
2997 /* ??? We could also check for
2999 (ne (and (eq (...) (const_int 1))) (const_int 0))
3001 and related forms, but let's wait until we see them occurring. */
3003 if (x == 0)
3004 /* Look up ARG1 in the hash table and see if it has an equivalence
3005 that lets us see what is being compared. */
3006 p = lookup (arg1, safe_hash (arg1, GET_MODE (arg1)) & HASH_MASK,
3007 GET_MODE (arg1));
3008 if (p)
3010 p = p->first_same_value;
3012 /* If what we compare is already known to be constant, that is as
3013 good as it gets.
3014 We need to break the loop in this case, because otherwise we
3015 can have an infinite loop when looking at a reg that is known
3016 to be a constant which is the same as a comparison of a reg
3017 against zero which appears later in the insn stream, which in
3018 turn is constant and the same as the comparison of the first reg
3019 against zero... */
3020 if (p->is_const)
3021 break;
3024 for (; p; p = p->next_same_value)
3026 enum machine_mode inner_mode = GET_MODE (p->exp);
3027 #ifdef FLOAT_STORE_FLAG_VALUE
3028 REAL_VALUE_TYPE fsfv;
3029 #endif
3031 /* If the entry isn't valid, skip it. */
3032 if (! exp_equiv_p (p->exp, p->exp, 1, 0))
3033 continue;
3035 if (GET_CODE (p->exp) == COMPARE
3036 /* Another possibility is that this machine has a compare insn
3037 that includes the comparison code. In that case, ARG1 would
3038 be equivalent to a comparison operation that would set ARG1 to
3039 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3040 ORIG_CODE is the actual comparison being done; if it is an EQ,
3041 we must reverse ORIG_CODE. On machine with a negative value
3042 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3043 || ((code == NE
3044 || (code == LT
3045 && GET_MODE_CLASS (inner_mode) == MODE_INT
3046 && (GET_MODE_BITSIZE (inner_mode)
3047 <= HOST_BITS_PER_WIDE_INT)
3048 && (STORE_FLAG_VALUE
3049 & ((HOST_WIDE_INT) 1
3050 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3051 #ifdef FLOAT_STORE_FLAG_VALUE
3052 || (code == LT
3053 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3054 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3055 REAL_VALUE_NEGATIVE (fsfv)))
3056 #endif
3058 && COMPARISON_P (p->exp)))
3060 x = p->exp;
3061 break;
3063 else if ((code == EQ
3064 || (code == GE
3065 && GET_MODE_CLASS (inner_mode) == MODE_INT
3066 && (GET_MODE_BITSIZE (inner_mode)
3067 <= HOST_BITS_PER_WIDE_INT)
3068 && (STORE_FLAG_VALUE
3069 & ((HOST_WIDE_INT) 1
3070 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3071 #ifdef FLOAT_STORE_FLAG_VALUE
3072 || (code == GE
3073 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3074 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3075 REAL_VALUE_NEGATIVE (fsfv)))
3076 #endif
3078 && COMPARISON_P (p->exp))
3080 reverse_code = 1;
3081 x = p->exp;
3082 break;
3085 /* If this non-trapping address, e.g. fp + constant, the
3086 equivalent is a better operand since it may let us predict
3087 the value of the comparison. */
3088 else if (!rtx_addr_can_trap_p (p->exp))
3090 arg1 = p->exp;
3091 continue;
3095 /* If we didn't find a useful equivalence for ARG1, we are done.
3096 Otherwise, set up for the next iteration. */
3097 if (x == 0)
3098 break;
3100 /* If we need to reverse the comparison, make sure that that is
3101 possible -- we can't necessarily infer the value of GE from LT
3102 with floating-point operands. */
3103 if (reverse_code)
3105 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3106 if (reversed == UNKNOWN)
3107 break;
3108 else
3109 code = reversed;
3111 else if (COMPARISON_P (x))
3112 code = GET_CODE (x);
3113 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3116 /* Return our results. Return the modes from before fold_rtx
3117 because fold_rtx might produce const_int, and then it's too late. */
3118 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3119 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3121 return code;
3124 /* If X is a nontrivial arithmetic operation on an argument
3125 for which a constant value can be determined, return
3126 the result of operating on that value, as a constant.
3127 Otherwise, return X, possibly with one or more operands
3128 modified by recursive calls to this function.
3130 If X is a register whose contents are known, we do NOT
3131 return those contents here. equiv_constant is called to
3132 perform that task.
3134 INSN is the insn that we may be modifying. If it is 0, make a copy
3135 of X before modifying it. */
3137 static rtx
3138 fold_rtx (rtx x, rtx insn)
3140 enum rtx_code code;
3141 enum machine_mode mode;
3142 const char *fmt;
3143 int i;
3144 rtx new = 0;
3145 int copied = 0;
3146 int must_swap = 0;
3148 /* Folded equivalents of first two operands of X. */
3149 rtx folded_arg0;
3150 rtx folded_arg1;
3152 /* Constant equivalents of first three operands of X;
3153 0 when no such equivalent is known. */
3154 rtx const_arg0;
3155 rtx const_arg1;
3156 rtx const_arg2;
3158 /* The mode of the first operand of X. We need this for sign and zero
3159 extends. */
3160 enum machine_mode mode_arg0;
3162 if (x == 0)
3163 return x;
3165 mode = GET_MODE (x);
3166 code = GET_CODE (x);
3167 switch (code)
3169 case CONST:
3170 case CONST_INT:
3171 case CONST_DOUBLE:
3172 case CONST_VECTOR:
3173 case SYMBOL_REF:
3174 case LABEL_REF:
3175 case REG:
3176 /* No use simplifying an EXPR_LIST
3177 since they are used only for lists of args
3178 in a function call's REG_EQUAL note. */
3179 case EXPR_LIST:
3180 /* Changing anything inside an ADDRESSOF is incorrect; we don't
3181 want to (e.g.,) make (addressof (const_int 0)) just because
3182 the location is known to be zero. */
3183 case ADDRESSOF:
3184 return x;
3186 #ifdef HAVE_cc0
3187 case CC0:
3188 return prev_insn_cc0;
3189 #endif
3191 case PC:
3192 /* If the next insn is a CODE_LABEL followed by a jump table,
3193 PC's value is a LABEL_REF pointing to that label. That
3194 lets us fold switch statements on the VAX. */
3196 rtx next;
3197 if (insn && tablejump_p (insn, &next, NULL))
3198 return gen_rtx_LABEL_REF (Pmode, next);
3200 break;
3202 case SUBREG:
3203 /* See if we previously assigned a constant value to this SUBREG. */
3204 if ((new = lookup_as_function (x, CONST_INT)) != 0
3205 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3206 return new;
3208 /* If this is a paradoxical SUBREG, we have no idea what value the
3209 extra bits would have. However, if the operand is equivalent
3210 to a SUBREG whose operand is the same as our mode, and all the
3211 modes are within a word, we can just use the inner operand
3212 because these SUBREGs just say how to treat the register.
3214 Similarly if we find an integer constant. */
3216 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3218 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3219 struct table_elt *elt;
3221 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3222 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3223 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3224 imode)) != 0)
3225 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3227 if (CONSTANT_P (elt->exp)
3228 && GET_MODE (elt->exp) == VOIDmode)
3229 return elt->exp;
3231 if (GET_CODE (elt->exp) == SUBREG
3232 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3233 && exp_equiv_p (elt->exp, elt->exp, 1, 0))
3234 return copy_rtx (SUBREG_REG (elt->exp));
3237 return x;
3240 /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG.
3241 We might be able to if the SUBREG is extracting a single word in an
3242 integral mode or extracting the low part. */
3244 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3245 const_arg0 = equiv_constant (folded_arg0);
3246 if (const_arg0)
3247 folded_arg0 = const_arg0;
3249 if (folded_arg0 != SUBREG_REG (x))
3251 new = simplify_subreg (mode, folded_arg0,
3252 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3253 if (new)
3254 return new;
3257 if (GET_CODE (folded_arg0) == REG
3258 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3260 struct table_elt *elt;
3262 /* We can use HASH here since we know that canon_hash won't be
3263 called. */
3264 elt = lookup (folded_arg0,
3265 HASH (folded_arg0, GET_MODE (folded_arg0)),
3266 GET_MODE (folded_arg0));
3268 if (elt)
3269 elt = elt->first_same_value;
3271 if (subreg_lowpart_p (x))
3272 /* If this is a narrowing SUBREG and our operand is a REG, see
3273 if we can find an equivalence for REG that is an arithmetic
3274 operation in a wider mode where both operands are paradoxical
3275 SUBREGs from objects of our result mode. In that case, we
3276 couldn-t report an equivalent value for that operation, since we
3277 don't know what the extra bits will be. But we can find an
3278 equivalence for this SUBREG by folding that operation in the
3279 narrow mode. This allows us to fold arithmetic in narrow modes
3280 when the machine only supports word-sized arithmetic.
3282 Also look for a case where we have a SUBREG whose operand
3283 is the same as our result. If both modes are smaller
3284 than a word, we are simply interpreting a register in
3285 different modes and we can use the inner value. */
3287 for (; elt; elt = elt->next_same_value)
3289 enum rtx_code eltcode = GET_CODE (elt->exp);
3291 /* Just check for unary and binary operations. */
3292 if (UNARY_P (elt->exp)
3293 && eltcode != SIGN_EXTEND
3294 && eltcode != ZERO_EXTEND
3295 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3296 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3297 && (GET_MODE_CLASS (mode)
3298 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3300 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3302 if (GET_CODE (op0) != REG && ! CONSTANT_P (op0))
3303 op0 = fold_rtx (op0, NULL_RTX);
3305 op0 = equiv_constant (op0);
3306 if (op0)
3307 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3308 op0, mode);
3310 else if (ARITHMETIC_P (elt->exp)
3311 && eltcode != DIV && eltcode != MOD
3312 && eltcode != UDIV && eltcode != UMOD
3313 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3314 && eltcode != ROTATE && eltcode != ROTATERT
3315 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3316 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3317 == mode))
3318 || CONSTANT_P (XEXP (elt->exp, 0)))
3319 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3320 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3321 == mode))
3322 || CONSTANT_P (XEXP (elt->exp, 1))))
3324 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3325 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3327 if (op0 && GET_CODE (op0) != REG && ! CONSTANT_P (op0))
3328 op0 = fold_rtx (op0, NULL_RTX);
3330 if (op0)
3331 op0 = equiv_constant (op0);
3333 if (op1 && GET_CODE (op1) != REG && ! CONSTANT_P (op1))
3334 op1 = fold_rtx (op1, NULL_RTX);
3336 if (op1)
3337 op1 = equiv_constant (op1);
3339 /* If we are looking for the low SImode part of
3340 (ashift:DI c (const_int 32)), it doesn't work
3341 to compute that in SImode, because a 32-bit shift
3342 in SImode is unpredictable. We know the value is 0. */
3343 if (op0 && op1
3344 && GET_CODE (elt->exp) == ASHIFT
3345 && GET_CODE (op1) == CONST_INT
3346 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3348 if (INTVAL (op1)
3349 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3350 /* If the count fits in the inner mode's width,
3351 but exceeds the outer mode's width,
3352 the value will get truncated to 0
3353 by the subreg. */
3354 new = CONST0_RTX (mode);
3355 else
3356 /* If the count exceeds even the inner mode's width,
3357 don't fold this expression. */
3358 new = 0;
3360 else if (op0 && op1)
3361 new = simplify_binary_operation (GET_CODE (elt->exp), mode, op0, op1);
3364 else if (GET_CODE (elt->exp) == SUBREG
3365 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3366 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3367 <= UNITS_PER_WORD)
3368 && exp_equiv_p (elt->exp, elt->exp, 1, 0))
3369 new = copy_rtx (SUBREG_REG (elt->exp));
3371 if (new)
3372 return new;
3374 else
3375 /* A SUBREG resulting from a zero extension may fold to zero if
3376 it extracts higher bits than the ZERO_EXTEND's source bits.
3377 FIXME: if combine tried to, er, combine these instructions,
3378 this transformation may be moved to simplify_subreg. */
3379 for (; elt; elt = elt->next_same_value)
3381 if (GET_CODE (elt->exp) == ZERO_EXTEND
3382 && subreg_lsb (x)
3383 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3384 return CONST0_RTX (mode);
3388 return x;
3390 case NOT:
3391 case NEG:
3392 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3393 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3394 new = lookup_as_function (XEXP (x, 0), code);
3395 if (new)
3396 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3397 break;
3399 case MEM:
3400 /* If we are not actually processing an insn, don't try to find the
3401 best address. Not only don't we care, but we could modify the
3402 MEM in an invalid way since we have no insn to validate against. */
3403 if (insn != 0)
3404 find_best_addr (insn, &XEXP (x, 0), GET_MODE (x));
3407 /* Even if we don't fold in the insn itself,
3408 we can safely do so here, in hopes of getting a constant. */
3409 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3410 rtx base = 0;
3411 HOST_WIDE_INT offset = 0;
3413 if (GET_CODE (addr) == REG
3414 && REGNO_QTY_VALID_P (REGNO (addr)))
3416 int addr_q = REG_QTY (REGNO (addr));
3417 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3419 if (GET_MODE (addr) == addr_ent->mode
3420 && addr_ent->const_rtx != NULL_RTX)
3421 addr = addr_ent->const_rtx;
3424 /* If address is constant, split it into a base and integer offset. */
3425 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3426 base = addr;
3427 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3428 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3430 base = XEXP (XEXP (addr, 0), 0);
3431 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3433 else if (GET_CODE (addr) == LO_SUM
3434 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3435 base = XEXP (addr, 1);
3436 else if (GET_CODE (addr) == ADDRESSOF)
3437 return change_address (x, VOIDmode, addr);
3439 /* If this is a constant pool reference, we can fold it into its
3440 constant to allow better value tracking. */
3441 if (base && GET_CODE (base) == SYMBOL_REF
3442 && CONSTANT_POOL_ADDRESS_P (base))
3444 rtx constant = get_pool_constant (base);
3445 enum machine_mode const_mode = get_pool_mode (base);
3446 rtx new;
3448 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3450 constant_pool_entries_cost = COST (constant);
3451 constant_pool_entries_regcost = approx_reg_cost (constant);
3454 /* If we are loading the full constant, we have an equivalence. */
3455 if (offset == 0 && mode == const_mode)
3456 return constant;
3458 /* If this actually isn't a constant (weird!), we can't do
3459 anything. Otherwise, handle the two most common cases:
3460 extracting a word from a multi-word constant, and extracting
3461 the low-order bits. Other cases don't seem common enough to
3462 worry about. */
3463 if (! CONSTANT_P (constant))
3464 return x;
3466 if (GET_MODE_CLASS (mode) == MODE_INT
3467 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3468 && offset % UNITS_PER_WORD == 0
3469 && (new = operand_subword (constant,
3470 offset / UNITS_PER_WORD,
3471 0, const_mode)) != 0)
3472 return new;
3474 if (((BYTES_BIG_ENDIAN
3475 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3476 || (! BYTES_BIG_ENDIAN && offset == 0))
3477 && (new = gen_lowpart (mode, constant)) != 0)
3478 return new;
3481 /* If this is a reference to a label at a known position in a jump
3482 table, we also know its value. */
3483 if (base && GET_CODE (base) == LABEL_REF)
3485 rtx label = XEXP (base, 0);
3486 rtx table_insn = NEXT_INSN (label);
3488 if (table_insn && GET_CODE (table_insn) == JUMP_INSN
3489 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3491 rtx table = PATTERN (table_insn);
3493 if (offset >= 0
3494 && (offset / GET_MODE_SIZE (GET_MODE (table))
3495 < XVECLEN (table, 0)))
3496 return XVECEXP (table, 0,
3497 offset / GET_MODE_SIZE (GET_MODE (table)));
3499 if (table_insn && GET_CODE (table_insn) == JUMP_INSN
3500 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3502 rtx table = PATTERN (table_insn);
3504 if (offset >= 0
3505 && (offset / GET_MODE_SIZE (GET_MODE (table))
3506 < XVECLEN (table, 1)))
3508 offset /= GET_MODE_SIZE (GET_MODE (table));
3509 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3510 XEXP (table, 0));
3512 if (GET_MODE (table) != Pmode)
3513 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3515 /* Indicate this is a constant. This isn't a
3516 valid form of CONST, but it will only be used
3517 to fold the next insns and then discarded, so
3518 it should be safe.
3520 Note this expression must be explicitly discarded,
3521 by cse_insn, else it may end up in a REG_EQUAL note
3522 and "escape" to cause problems elsewhere. */
3523 return gen_rtx_CONST (GET_MODE (new), new);
3528 return x;
3531 #ifdef NO_FUNCTION_CSE
3532 case CALL:
3533 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3534 return x;
3535 break;
3536 #endif
3538 case ASM_OPERANDS:
3539 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3540 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3541 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3542 break;
3544 default:
3545 break;
3548 const_arg0 = 0;
3549 const_arg1 = 0;
3550 const_arg2 = 0;
3551 mode_arg0 = VOIDmode;
3553 /* Try folding our operands.
3554 Then see which ones have constant values known. */
3556 fmt = GET_RTX_FORMAT (code);
3557 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3558 if (fmt[i] == 'e')
3560 rtx arg = XEXP (x, i);
3561 rtx folded_arg = arg, const_arg = 0;
3562 enum machine_mode mode_arg = GET_MODE (arg);
3563 rtx cheap_arg, expensive_arg;
3564 rtx replacements[2];
3565 int j;
3566 int old_cost = COST_IN (XEXP (x, i), code);
3568 /* Most arguments are cheap, so handle them specially. */
3569 switch (GET_CODE (arg))
3571 case REG:
3572 /* This is the same as calling equiv_constant; it is duplicated
3573 here for speed. */
3574 if (REGNO_QTY_VALID_P (REGNO (arg)))
3576 int arg_q = REG_QTY (REGNO (arg));
3577 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3579 if (arg_ent->const_rtx != NULL_RTX
3580 && GET_CODE (arg_ent->const_rtx) != REG
3581 && GET_CODE (arg_ent->const_rtx) != PLUS)
3582 const_arg
3583 = gen_lowpart (GET_MODE (arg),
3584 arg_ent->const_rtx);
3586 break;
3588 case CONST:
3589 case CONST_INT:
3590 case SYMBOL_REF:
3591 case LABEL_REF:
3592 case CONST_DOUBLE:
3593 case CONST_VECTOR:
3594 const_arg = arg;
3595 break;
3597 #ifdef HAVE_cc0
3598 case CC0:
3599 folded_arg = prev_insn_cc0;
3600 mode_arg = prev_insn_cc0_mode;
3601 const_arg = equiv_constant (folded_arg);
3602 break;
3603 #endif
3605 default:
3606 folded_arg = fold_rtx (arg, insn);
3607 const_arg = equiv_constant (folded_arg);
3610 /* For the first three operands, see if the operand
3611 is constant or equivalent to a constant. */
3612 switch (i)
3614 case 0:
3615 folded_arg0 = folded_arg;
3616 const_arg0 = const_arg;
3617 mode_arg0 = mode_arg;
3618 break;
3619 case 1:
3620 folded_arg1 = folded_arg;
3621 const_arg1 = const_arg;
3622 break;
3623 case 2:
3624 const_arg2 = const_arg;
3625 break;
3628 /* Pick the least expensive of the folded argument and an
3629 equivalent constant argument. */
3630 if (const_arg == 0 || const_arg == folded_arg
3631 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3632 cheap_arg = folded_arg, expensive_arg = const_arg;
3633 else
3634 cheap_arg = const_arg, expensive_arg = folded_arg;
3636 /* Try to replace the operand with the cheapest of the two
3637 possibilities. If it doesn't work and this is either of the first
3638 two operands of a commutative operation, try swapping them.
3639 If THAT fails, try the more expensive, provided it is cheaper
3640 than what is already there. */
3642 if (cheap_arg == XEXP (x, i))
3643 continue;
3645 if (insn == 0 && ! copied)
3647 x = copy_rtx (x);
3648 copied = 1;
3651 /* Order the replacements from cheapest to most expensive. */
3652 replacements[0] = cheap_arg;
3653 replacements[1] = expensive_arg;
3655 for (j = 0; j < 2 && replacements[j]; j++)
3657 int new_cost = COST_IN (replacements[j], code);
3659 /* Stop if what existed before was cheaper. Prefer constants
3660 in the case of a tie. */
3661 if (new_cost > old_cost
3662 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3663 break;
3665 /* It's not safe to substitute the operand of a conversion
3666 operator with a constant, as the conversion's identity
3667 depends upon the mode of it's operand. This optimization
3668 is handled by the call to simplify_unary_operation. */
3669 if (GET_RTX_CLASS (code) == RTX_UNARY
3670 && GET_MODE (replacements[j]) != mode_arg0
3671 && (code == ZERO_EXTEND
3672 || code == SIGN_EXTEND
3673 || code == TRUNCATE
3674 || code == FLOAT_TRUNCATE
3675 || code == FLOAT_EXTEND
3676 || code == FLOAT
3677 || code == FIX
3678 || code == UNSIGNED_FLOAT
3679 || code == UNSIGNED_FIX))
3680 continue;
3682 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3683 break;
3685 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3686 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3688 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3689 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3691 if (apply_change_group ())
3693 /* Swap them back to be invalid so that this loop can
3694 continue and flag them to be swapped back later. */
3695 rtx tem;
3697 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3698 XEXP (x, 1) = tem;
3699 must_swap = 1;
3700 break;
3706 else
3708 if (fmt[i] == 'E')
3709 /* Don't try to fold inside of a vector of expressions.
3710 Doing nothing is harmless. */
3714 /* If a commutative operation, place a constant integer as the second
3715 operand unless the first operand is also a constant integer. Otherwise,
3716 place any constant second unless the first operand is also a constant. */
3718 if (COMMUTATIVE_P (x))
3720 if (must_swap
3721 || swap_commutative_operands_p (const_arg0 ? const_arg0
3722 : XEXP (x, 0),
3723 const_arg1 ? const_arg1
3724 : XEXP (x, 1)))
3726 rtx tem = XEXP (x, 0);
3728 if (insn == 0 && ! copied)
3730 x = copy_rtx (x);
3731 copied = 1;
3734 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3735 validate_change (insn, &XEXP (x, 1), tem, 1);
3736 if (apply_change_group ())
3738 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3739 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3744 /* If X is an arithmetic operation, see if we can simplify it. */
3746 switch (GET_RTX_CLASS (code))
3748 case RTX_UNARY:
3750 int is_const = 0;
3752 /* We can't simplify extension ops unless we know the
3753 original mode. */
3754 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3755 && mode_arg0 == VOIDmode)
3756 break;
3758 /* If we had a CONST, strip it off and put it back later if we
3759 fold. */
3760 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3761 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3763 new = simplify_unary_operation (code, mode,
3764 const_arg0 ? const_arg0 : folded_arg0,
3765 mode_arg0);
3766 if (new != 0 && is_const)
3767 new = gen_rtx_CONST (mode, new);
3769 break;
3771 case RTX_COMPARE:
3772 case RTX_COMM_COMPARE:
3773 /* See what items are actually being compared and set FOLDED_ARG[01]
3774 to those values and CODE to the actual comparison code. If any are
3775 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3776 do anything if both operands are already known to be constant. */
3778 if (const_arg0 == 0 || const_arg1 == 0)
3780 struct table_elt *p0, *p1;
3781 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3782 enum machine_mode mode_arg1;
3784 #ifdef FLOAT_STORE_FLAG_VALUE
3785 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3787 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3788 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3789 false_rtx = CONST0_RTX (mode);
3791 #endif
3793 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3794 &mode_arg0, &mode_arg1);
3795 const_arg0 = equiv_constant (folded_arg0);
3796 const_arg1 = equiv_constant (folded_arg1);
3798 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3799 what kinds of things are being compared, so we can't do
3800 anything with this comparison. */
3802 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3803 break;
3805 /* If we do not now have two constants being compared, see
3806 if we can nevertheless deduce some things about the
3807 comparison. */
3808 if (const_arg0 == 0 || const_arg1 == 0)
3810 /* Some addresses are known to be nonzero. We don't know
3811 their sign, but equality comparisons are known. */
3812 if (const_arg1 == const0_rtx
3813 && nonzero_address_p (folded_arg0))
3815 if (code == EQ)
3816 return false_rtx;
3817 else if (code == NE)
3818 return true_rtx;
3821 /* See if the two operands are the same. */
3823 if (folded_arg0 == folded_arg1
3824 || (GET_CODE (folded_arg0) == REG
3825 && GET_CODE (folded_arg1) == REG
3826 && (REG_QTY (REGNO (folded_arg0))
3827 == REG_QTY (REGNO (folded_arg1))))
3828 || ((p0 = lookup (folded_arg0,
3829 (safe_hash (folded_arg0, mode_arg0)
3830 & HASH_MASK), mode_arg0))
3831 && (p1 = lookup (folded_arg1,
3832 (safe_hash (folded_arg1, mode_arg0)
3833 & HASH_MASK), mode_arg0))
3834 && p0->first_same_value == p1->first_same_value))
3836 /* Sadly two equal NaNs are not equivalent. */
3837 if (!HONOR_NANS (mode_arg0))
3838 return ((code == EQ || code == LE || code == GE
3839 || code == LEU || code == GEU || code == UNEQ
3840 || code == UNLE || code == UNGE
3841 || code == ORDERED)
3842 ? true_rtx : false_rtx);
3843 /* Take care for the FP compares we can resolve. */
3844 if (code == UNEQ || code == UNLE || code == UNGE)
3845 return true_rtx;
3846 if (code == LTGT || code == LT || code == GT)
3847 return false_rtx;
3850 /* If FOLDED_ARG0 is a register, see if the comparison we are
3851 doing now is either the same as we did before or the reverse
3852 (we only check the reverse if not floating-point). */
3853 else if (GET_CODE (folded_arg0) == REG)
3855 int qty = REG_QTY (REGNO (folded_arg0));
3857 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3859 struct qty_table_elem *ent = &qty_table[qty];
3861 if ((comparison_dominates_p (ent->comparison_code, code)
3862 || (! FLOAT_MODE_P (mode_arg0)
3863 && comparison_dominates_p (ent->comparison_code,
3864 reverse_condition (code))))
3865 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3866 || (const_arg1
3867 && rtx_equal_p (ent->comparison_const,
3868 const_arg1))
3869 || (GET_CODE (folded_arg1) == REG
3870 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3871 return (comparison_dominates_p (ent->comparison_code, code)
3872 ? true_rtx : false_rtx);
3878 /* If we are comparing against zero, see if the first operand is
3879 equivalent to an IOR with a constant. If so, we may be able to
3880 determine the result of this comparison. */
3882 if (const_arg1 == const0_rtx)
3884 rtx y = lookup_as_function (folded_arg0, IOR);
3885 rtx inner_const;
3887 if (y != 0
3888 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3889 && GET_CODE (inner_const) == CONST_INT
3890 && INTVAL (inner_const) != 0)
3892 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
3893 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
3894 && (INTVAL (inner_const)
3895 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
3896 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3898 #ifdef FLOAT_STORE_FLAG_VALUE
3899 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3901 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3902 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3903 false_rtx = CONST0_RTX (mode);
3905 #endif
3907 switch (code)
3909 case EQ:
3910 return false_rtx;
3911 case NE:
3912 return true_rtx;
3913 case LT: case LE:
3914 if (has_sign)
3915 return true_rtx;
3916 break;
3917 case GT: case GE:
3918 if (has_sign)
3919 return false_rtx;
3920 break;
3921 default:
3922 break;
3927 new = simplify_relational_operation (code, mode,
3928 (mode_arg0 != VOIDmode
3929 ? mode_arg0
3930 : (GET_MODE (const_arg0
3931 ? const_arg0
3932 : folded_arg0)
3933 != VOIDmode)
3934 ? GET_MODE (const_arg0
3935 ? const_arg0
3936 : folded_arg0)
3937 : GET_MODE (const_arg1
3938 ? const_arg1
3939 : folded_arg1)),
3940 const_arg0 ? const_arg0 : folded_arg0,
3941 const_arg1 ? const_arg1 : folded_arg1);
3942 break;
3944 case RTX_BIN_ARITH:
3945 case RTX_COMM_ARITH:
3946 switch (code)
3948 case PLUS:
3949 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3950 with that LABEL_REF as its second operand. If so, the result is
3951 the first operand of that MINUS. This handles switches with an
3952 ADDR_DIFF_VEC table. */
3953 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3955 rtx y
3956 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3957 : lookup_as_function (folded_arg0, MINUS);
3959 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3960 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3961 return XEXP (y, 0);
3963 /* Now try for a CONST of a MINUS like the above. */
3964 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3965 : lookup_as_function (folded_arg0, CONST))) != 0
3966 && GET_CODE (XEXP (y, 0)) == MINUS
3967 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3968 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3969 return XEXP (XEXP (y, 0), 0);
3972 /* Likewise if the operands are in the other order. */
3973 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3975 rtx y
3976 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3977 : lookup_as_function (folded_arg1, MINUS);
3979 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3980 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3981 return XEXP (y, 0);
3983 /* Now try for a CONST of a MINUS like the above. */
3984 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3985 : lookup_as_function (folded_arg1, CONST))) != 0
3986 && GET_CODE (XEXP (y, 0)) == MINUS
3987 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3988 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3989 return XEXP (XEXP (y, 0), 0);
3992 /* If second operand is a register equivalent to a negative
3993 CONST_INT, see if we can find a register equivalent to the
3994 positive constant. Make a MINUS if so. Don't do this for
3995 a non-negative constant since we might then alternate between
3996 choosing positive and negative constants. Having the positive
3997 constant previously-used is the more common case. Be sure
3998 the resulting constant is non-negative; if const_arg1 were
3999 the smallest negative number this would overflow: depending
4000 on the mode, this would either just be the same value (and
4001 hence not save anything) or be incorrect. */
4002 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4003 && INTVAL (const_arg1) < 0
4004 /* This used to test
4006 -INTVAL (const_arg1) >= 0
4008 But The Sun V5.0 compilers mis-compiled that test. So
4009 instead we test for the problematic value in a more direct
4010 manner and hope the Sun compilers get it correct. */
4011 && INTVAL (const_arg1) !=
4012 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4013 && GET_CODE (folded_arg1) == REG)
4015 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4016 struct table_elt *p
4017 = lookup (new_const, safe_hash (new_const, mode) & HASH_MASK,
4018 mode);
4020 if (p)
4021 for (p = p->first_same_value; p; p = p->next_same_value)
4022 if (GET_CODE (p->exp) == REG)
4023 return simplify_gen_binary (MINUS, mode, folded_arg0,
4024 canon_reg (p->exp, NULL_RTX));
4026 goto from_plus;
4028 case MINUS:
4029 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4030 If so, produce (PLUS Z C2-C). */
4031 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4033 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4034 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4035 return fold_rtx (plus_constant (copy_rtx (y),
4036 -INTVAL (const_arg1)),
4037 NULL_RTX);
4040 /* Fall through. */
4042 from_plus:
4043 case SMIN: case SMAX: case UMIN: case UMAX:
4044 case IOR: case AND: case XOR:
4045 case MULT:
4046 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4047 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4048 is known to be of similar form, we may be able to replace the
4049 operation with a combined operation. This may eliminate the
4050 intermediate operation if every use is simplified in this way.
4051 Note that the similar optimization done by combine.c only works
4052 if the intermediate operation's result has only one reference. */
4054 if (GET_CODE (folded_arg0) == REG
4055 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4057 int is_shift
4058 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4059 rtx y = lookup_as_function (folded_arg0, code);
4060 rtx inner_const;
4061 enum rtx_code associate_code;
4062 rtx new_const;
4064 if (y == 0
4065 || 0 == (inner_const
4066 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4067 || GET_CODE (inner_const) != CONST_INT
4068 /* If we have compiled a statement like
4069 "if (x == (x & mask1))", and now are looking at
4070 "x & mask2", we will have a case where the first operand
4071 of Y is the same as our first operand. Unless we detect
4072 this case, an infinite loop will result. */
4073 || XEXP (y, 0) == folded_arg0)
4074 break;
4076 /* Don't associate these operations if they are a PLUS with the
4077 same constant and it is a power of two. These might be doable
4078 with a pre- or post-increment. Similarly for two subtracts of
4079 identical powers of two with post decrement. */
4081 if (code == PLUS && const_arg1 == inner_const
4082 && ((HAVE_PRE_INCREMENT
4083 && exact_log2 (INTVAL (const_arg1)) >= 0)
4084 || (HAVE_POST_INCREMENT
4085 && exact_log2 (INTVAL (const_arg1)) >= 0)
4086 || (HAVE_PRE_DECREMENT
4087 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4088 || (HAVE_POST_DECREMENT
4089 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4090 break;
4092 /* Compute the code used to compose the constants. For example,
4093 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4095 associate_code = (is_shift || code == MINUS ? PLUS : code);
4097 new_const = simplify_binary_operation (associate_code, mode,
4098 const_arg1, inner_const);
4100 if (new_const == 0)
4101 break;
4103 /* If we are associating shift operations, don't let this
4104 produce a shift of the size of the object or larger.
4105 This could occur when we follow a sign-extend by a right
4106 shift on a machine that does a sign-extend as a pair
4107 of shifts. */
4109 if (is_shift && GET_CODE (new_const) == CONST_INT
4110 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4112 /* As an exception, we can turn an ASHIFTRT of this
4113 form into a shift of the number of bits - 1. */
4114 if (code == ASHIFTRT)
4115 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4116 else
4117 break;
4120 y = copy_rtx (XEXP (y, 0));
4122 /* If Y contains our first operand (the most common way this
4123 can happen is if Y is a MEM), we would do into an infinite
4124 loop if we tried to fold it. So don't in that case. */
4126 if (! reg_mentioned_p (folded_arg0, y))
4127 y = fold_rtx (y, insn);
4129 return simplify_gen_binary (code, mode, y, new_const);
4131 break;
4133 case DIV: case UDIV:
4134 /* ??? The associative optimization performed immediately above is
4135 also possible for DIV and UDIV using associate_code of MULT.
4136 However, we would need extra code to verify that the
4137 multiplication does not overflow, that is, there is no overflow
4138 in the calculation of new_const. */
4139 break;
4141 default:
4142 break;
4145 new = simplify_binary_operation (code, mode,
4146 const_arg0 ? const_arg0 : folded_arg0,
4147 const_arg1 ? const_arg1 : folded_arg1);
4148 break;
4150 case RTX_OBJ:
4151 /* (lo_sum (high X) X) is simply X. */
4152 if (code == LO_SUM && const_arg0 != 0
4153 && GET_CODE (const_arg0) == HIGH
4154 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4155 return const_arg1;
4156 break;
4158 case RTX_TERNARY:
4159 case RTX_BITFIELD_OPS:
4160 new = simplify_ternary_operation (code, mode, mode_arg0,
4161 const_arg0 ? const_arg0 : folded_arg0,
4162 const_arg1 ? const_arg1 : folded_arg1,
4163 const_arg2 ? const_arg2 : XEXP (x, 2));
4164 break;
4166 case RTX_EXTRA:
4167 /* Eliminate CONSTANT_P_RTX if its constant. */
4168 if (code == CONSTANT_P_RTX)
4170 if (const_arg0)
4171 return const1_rtx;
4172 if (optimize == 0 || !flag_gcse)
4173 return const0_rtx;
4175 break;
4177 default:
4178 break;
4181 return new ? new : x;
4184 /* Return a constant value currently equivalent to X.
4185 Return 0 if we don't know one. */
4187 static rtx
4188 equiv_constant (rtx x)
4190 if (GET_CODE (x) == REG
4191 && REGNO_QTY_VALID_P (REGNO (x)))
4193 int x_q = REG_QTY (REGNO (x));
4194 struct qty_table_elem *x_ent = &qty_table[x_q];
4196 if (x_ent->const_rtx)
4197 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4200 if (x == 0 || CONSTANT_P (x))
4201 return x;
4203 /* If X is a MEM, try to fold it outside the context of any insn to see if
4204 it might be equivalent to a constant. That handles the case where it
4205 is a constant-pool reference. Then try to look it up in the hash table
4206 in case it is something whose value we have seen before. */
4208 if (GET_CODE (x) == MEM)
4210 struct table_elt *elt;
4212 x = fold_rtx (x, NULL_RTX);
4213 if (CONSTANT_P (x))
4214 return x;
4216 elt = lookup (x, safe_hash (x, GET_MODE (x)) & HASH_MASK, GET_MODE (x));
4217 if (elt == 0)
4218 return 0;
4220 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4221 if (elt->is_const && CONSTANT_P (elt->exp))
4222 return elt->exp;
4225 return 0;
4228 /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point
4229 number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
4230 least-significant part of X.
4231 MODE specifies how big a part of X to return.
4233 If the requested operation cannot be done, 0 is returned.
4235 This is similar to gen_lowpart_general in emit-rtl.c. */
4238 gen_lowpart_if_possible (enum machine_mode mode, rtx x)
4240 rtx result = gen_lowpart_common (mode, x);
4242 if (result)
4243 return result;
4244 else if (GET_CODE (x) == MEM)
4246 /* This is the only other case we handle. */
4247 int offset = 0;
4248 rtx new;
4250 if (WORDS_BIG_ENDIAN)
4251 offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD)
4252 - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD));
4253 if (BYTES_BIG_ENDIAN)
4254 /* Adjust the address so that the address-after-the-data is
4255 unchanged. */
4256 offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode))
4257 - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x))));
4259 new = adjust_address_nv (x, mode, offset);
4260 if (! memory_address_p (mode, XEXP (new, 0)))
4261 return 0;
4263 return new;
4265 else
4266 return 0;
4269 /* Given INSN, a jump insn, TAKEN indicates if we are following the "taken"
4270 branch. It will be zero if not.
4272 In certain cases, this can cause us to add an equivalence. For example,
4273 if we are following the taken case of
4274 if (i == 2)
4275 we can add the fact that `i' and '2' are now equivalent.
4277 In any case, we can record that this comparison was passed. If the same
4278 comparison is seen later, we will know its value. */
4280 static void
4281 record_jump_equiv (rtx insn, int taken)
4283 int cond_known_true;
4284 rtx op0, op1;
4285 rtx set;
4286 enum machine_mode mode, mode0, mode1;
4287 int reversed_nonequality = 0;
4288 enum rtx_code code;
4290 /* Ensure this is the right kind of insn. */
4291 if (! any_condjump_p (insn))
4292 return;
4293 set = pc_set (insn);
4295 /* See if this jump condition is known true or false. */
4296 if (taken)
4297 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4298 else
4299 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4301 /* Get the type of comparison being done and the operands being compared.
4302 If we had to reverse a non-equality condition, record that fact so we
4303 know that it isn't valid for floating-point. */
4304 code = GET_CODE (XEXP (SET_SRC (set), 0));
4305 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4306 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4308 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4309 if (! cond_known_true)
4311 code = reversed_comparison_code_parts (code, op0, op1, insn);
4313 /* Don't remember if we can't find the inverse. */
4314 if (code == UNKNOWN)
4315 return;
4318 /* The mode is the mode of the non-constant. */
4319 mode = mode0;
4320 if (mode1 != VOIDmode)
4321 mode = mode1;
4323 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4326 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4327 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4328 Make any useful entries we can with that information. Called from
4329 above function and called recursively. */
4331 static void
4332 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4333 rtx op1, int reversed_nonequality)
4335 unsigned op0_hash, op1_hash;
4336 int op0_in_memory, op1_in_memory;
4337 struct table_elt *op0_elt, *op1_elt;
4339 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4340 we know that they are also equal in the smaller mode (this is also
4341 true for all smaller modes whether or not there is a SUBREG, but
4342 is not worth testing for with no SUBREG). */
4344 /* Note that GET_MODE (op0) may not equal MODE. */
4345 if (code == EQ && GET_CODE (op0) == SUBREG
4346 && (GET_MODE_SIZE (GET_MODE (op0))
4347 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4349 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4350 rtx tem = gen_lowpart (inner_mode, op1);
4352 record_jump_cond (code, mode, SUBREG_REG (op0),
4353 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4354 reversed_nonequality);
4357 if (code == EQ && GET_CODE (op1) == SUBREG
4358 && (GET_MODE_SIZE (GET_MODE (op1))
4359 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4361 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4362 rtx tem = gen_lowpart (inner_mode, op0);
4364 record_jump_cond (code, mode, SUBREG_REG (op1),
4365 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4366 reversed_nonequality);
4369 /* Similarly, if this is an NE comparison, and either is a SUBREG
4370 making a smaller mode, we know the whole thing is also NE. */
4372 /* Note that GET_MODE (op0) may not equal MODE;
4373 if we test MODE instead, we can get an infinite recursion
4374 alternating between two modes each wider than MODE. */
4376 if (code == NE && GET_CODE (op0) == SUBREG
4377 && subreg_lowpart_p (op0)
4378 && (GET_MODE_SIZE (GET_MODE (op0))
4379 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4381 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4382 rtx tem = gen_lowpart (inner_mode, op1);
4384 record_jump_cond (code, mode, SUBREG_REG (op0),
4385 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4386 reversed_nonequality);
4389 if (code == NE && GET_CODE (op1) == SUBREG
4390 && subreg_lowpart_p (op1)
4391 && (GET_MODE_SIZE (GET_MODE (op1))
4392 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4394 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4395 rtx tem = gen_lowpart (inner_mode, op0);
4397 record_jump_cond (code, mode, SUBREG_REG (op1),
4398 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4399 reversed_nonequality);
4402 /* Hash both operands. */
4404 do_not_record = 0;
4405 hash_arg_in_memory = 0;
4406 op0_hash = HASH (op0, mode);
4407 op0_in_memory = hash_arg_in_memory;
4409 if (do_not_record)
4410 return;
4412 do_not_record = 0;
4413 hash_arg_in_memory = 0;
4414 op1_hash = HASH (op1, mode);
4415 op1_in_memory = hash_arg_in_memory;
4417 if (do_not_record)
4418 return;
4420 /* Look up both operands. */
4421 op0_elt = lookup (op0, op0_hash, mode);
4422 op1_elt = lookup (op1, op1_hash, mode);
4424 /* If both operands are already equivalent or if they are not in the
4425 table but are identical, do nothing. */
4426 if ((op0_elt != 0 && op1_elt != 0
4427 && op0_elt->first_same_value == op1_elt->first_same_value)
4428 || op0 == op1 || rtx_equal_p (op0, op1))
4429 return;
4431 /* If we aren't setting two things equal all we can do is save this
4432 comparison. Similarly if this is floating-point. In the latter
4433 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4434 If we record the equality, we might inadvertently delete code
4435 whose intent was to change -0 to +0. */
4437 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4439 struct qty_table_elem *ent;
4440 int qty;
4442 /* If we reversed a floating-point comparison, if OP0 is not a
4443 register, or if OP1 is neither a register or constant, we can't
4444 do anything. */
4446 if (GET_CODE (op1) != REG)
4447 op1 = equiv_constant (op1);
4449 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4450 || GET_CODE (op0) != REG || op1 == 0)
4451 return;
4453 /* Put OP0 in the hash table if it isn't already. This gives it a
4454 new quantity number. */
4455 if (op0_elt == 0)
4457 if (insert_regs (op0, NULL, 0))
4459 rehash_using_reg (op0);
4460 op0_hash = HASH (op0, mode);
4462 /* If OP0 is contained in OP1, this changes its hash code
4463 as well. Faster to rehash than to check, except
4464 for the simple case of a constant. */
4465 if (! CONSTANT_P (op1))
4466 op1_hash = HASH (op1,mode);
4469 op0_elt = insert (op0, NULL, op0_hash, mode);
4470 op0_elt->in_memory = op0_in_memory;
4473 qty = REG_QTY (REGNO (op0));
4474 ent = &qty_table[qty];
4476 ent->comparison_code = code;
4477 if (GET_CODE (op1) == REG)
4479 /* Look it up again--in case op0 and op1 are the same. */
4480 op1_elt = lookup (op1, op1_hash, mode);
4482 /* Put OP1 in the hash table so it gets a new quantity number. */
4483 if (op1_elt == 0)
4485 if (insert_regs (op1, NULL, 0))
4487 rehash_using_reg (op1);
4488 op1_hash = HASH (op1, mode);
4491 op1_elt = insert (op1, NULL, op1_hash, mode);
4492 op1_elt->in_memory = op1_in_memory;
4495 ent->comparison_const = NULL_RTX;
4496 ent->comparison_qty = REG_QTY (REGNO (op1));
4498 else
4500 ent->comparison_const = op1;
4501 ent->comparison_qty = -1;
4504 return;
4507 /* If either side is still missing an equivalence, make it now,
4508 then merge the equivalences. */
4510 if (op0_elt == 0)
4512 if (insert_regs (op0, NULL, 0))
4514 rehash_using_reg (op0);
4515 op0_hash = HASH (op0, mode);
4518 op0_elt = insert (op0, NULL, op0_hash, mode);
4519 op0_elt->in_memory = op0_in_memory;
4522 if (op1_elt == 0)
4524 if (insert_regs (op1, NULL, 0))
4526 rehash_using_reg (op1);
4527 op1_hash = HASH (op1, mode);
4530 op1_elt = insert (op1, NULL, op1_hash, mode);
4531 op1_elt->in_memory = op1_in_memory;
4534 merge_equiv_classes (op0_elt, op1_elt);
4535 last_jump_equiv_class = op0_elt;
4538 /* CSE processing for one instruction.
4539 First simplify sources and addresses of all assignments
4540 in the instruction, using previously-computed equivalents values.
4541 Then install the new sources and destinations in the table
4542 of available values.
4544 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4545 the insn. It means that INSN is inside libcall block. In this
4546 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4548 /* Data on one SET contained in the instruction. */
4550 struct set
4552 /* The SET rtx itself. */
4553 rtx rtl;
4554 /* The SET_SRC of the rtx (the original value, if it is changing). */
4555 rtx src;
4556 /* The hash-table element for the SET_SRC of the SET. */
4557 struct table_elt *src_elt;
4558 /* Hash value for the SET_SRC. */
4559 unsigned src_hash;
4560 /* Hash value for the SET_DEST. */
4561 unsigned dest_hash;
4562 /* The SET_DEST, with SUBREG, etc., stripped. */
4563 rtx inner_dest;
4564 /* Nonzero if the SET_SRC is in memory. */
4565 char src_in_memory;
4566 /* Nonzero if the SET_SRC contains something
4567 whose value cannot be predicted and understood. */
4568 char src_volatile;
4569 /* Original machine mode, in case it becomes a CONST_INT.
4570 The size of this field should match the size of the mode
4571 field of struct rtx_def (see rtl.h). */
4572 ENUM_BITFIELD(machine_mode) mode : 8;
4573 /* A constant equivalent for SET_SRC, if any. */
4574 rtx src_const;
4575 /* Original SET_SRC value used for libcall notes. */
4576 rtx orig_src;
4577 /* Hash value of constant equivalent for SET_SRC. */
4578 unsigned src_const_hash;
4579 /* Table entry for constant equivalent for SET_SRC, if any. */
4580 struct table_elt *src_const_elt;
4583 static void
4584 cse_insn (rtx insn, rtx libcall_insn)
4586 rtx x = PATTERN (insn);
4587 int i;
4588 rtx tem;
4589 int n_sets = 0;
4591 #ifdef HAVE_cc0
4592 /* Records what this insn does to set CC0. */
4593 rtx this_insn_cc0 = 0;
4594 enum machine_mode this_insn_cc0_mode = VOIDmode;
4595 #endif
4597 rtx src_eqv = 0;
4598 struct table_elt *src_eqv_elt = 0;
4599 int src_eqv_volatile = 0;
4600 int src_eqv_in_memory = 0;
4601 unsigned src_eqv_hash = 0;
4603 struct set *sets = (struct set *) 0;
4605 this_insn = insn;
4607 /* Find all the SETs and CLOBBERs in this instruction.
4608 Record all the SETs in the array `set' and count them.
4609 Also determine whether there is a CLOBBER that invalidates
4610 all memory references, or all references at varying addresses. */
4612 if (GET_CODE (insn) == CALL_INSN)
4614 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4616 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4617 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4618 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4622 if (GET_CODE (x) == SET)
4624 sets = alloca (sizeof (struct set));
4625 sets[0].rtl = x;
4627 /* Ignore SETs that are unconditional jumps.
4628 They never need cse processing, so this does not hurt.
4629 The reason is not efficiency but rather
4630 so that we can test at the end for instructions
4631 that have been simplified to unconditional jumps
4632 and not be misled by unchanged instructions
4633 that were unconditional jumps to begin with. */
4634 if (SET_DEST (x) == pc_rtx
4635 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4638 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4639 The hard function value register is used only once, to copy to
4640 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4641 Ensure we invalidate the destination register. On the 80386 no
4642 other code would invalidate it since it is a fixed_reg.
4643 We need not check the return of apply_change_group; see canon_reg. */
4645 else if (GET_CODE (SET_SRC (x)) == CALL)
4647 canon_reg (SET_SRC (x), insn);
4648 apply_change_group ();
4649 fold_rtx (SET_SRC (x), insn);
4650 invalidate (SET_DEST (x), VOIDmode);
4652 else
4653 n_sets = 1;
4655 else if (GET_CODE (x) == PARALLEL)
4657 int lim = XVECLEN (x, 0);
4659 sets = alloca (lim * sizeof (struct set));
4661 /* Find all regs explicitly clobbered in this insn,
4662 and ensure they are not replaced with any other regs
4663 elsewhere in this insn.
4664 When a reg that is clobbered is also used for input,
4665 we should presume that that is for a reason,
4666 and we should not substitute some other register
4667 which is not supposed to be clobbered.
4668 Therefore, this loop cannot be merged into the one below
4669 because a CALL may precede a CLOBBER and refer to the
4670 value clobbered. We must not let a canonicalization do
4671 anything in that case. */
4672 for (i = 0; i < lim; i++)
4674 rtx y = XVECEXP (x, 0, i);
4675 if (GET_CODE (y) == CLOBBER)
4677 rtx clobbered = XEXP (y, 0);
4679 if (GET_CODE (clobbered) == REG
4680 || GET_CODE (clobbered) == SUBREG)
4681 invalidate (clobbered, VOIDmode);
4682 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4683 || GET_CODE (clobbered) == ZERO_EXTRACT)
4684 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4688 for (i = 0; i < lim; i++)
4690 rtx y = XVECEXP (x, 0, i);
4691 if (GET_CODE (y) == SET)
4693 /* As above, we ignore unconditional jumps and call-insns and
4694 ignore the result of apply_change_group. */
4695 if (GET_CODE (SET_SRC (y)) == CALL)
4697 canon_reg (SET_SRC (y), insn);
4698 apply_change_group ();
4699 fold_rtx (SET_SRC (y), insn);
4700 invalidate (SET_DEST (y), VOIDmode);
4702 else if (SET_DEST (y) == pc_rtx
4703 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4705 else
4706 sets[n_sets++].rtl = y;
4708 else if (GET_CODE (y) == CLOBBER)
4710 /* If we clobber memory, canon the address.
4711 This does nothing when a register is clobbered
4712 because we have already invalidated the reg. */
4713 if (GET_CODE (XEXP (y, 0)) == MEM)
4714 canon_reg (XEXP (y, 0), NULL_RTX);
4716 else if (GET_CODE (y) == USE
4717 && ! (GET_CODE (XEXP (y, 0)) == REG
4718 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4719 canon_reg (y, NULL_RTX);
4720 else if (GET_CODE (y) == CALL)
4722 /* The result of apply_change_group can be ignored; see
4723 canon_reg. */
4724 canon_reg (y, insn);
4725 apply_change_group ();
4726 fold_rtx (y, insn);
4730 else if (GET_CODE (x) == CLOBBER)
4732 if (GET_CODE (XEXP (x, 0)) == MEM)
4733 canon_reg (XEXP (x, 0), NULL_RTX);
4736 /* Canonicalize a USE of a pseudo register or memory location. */
4737 else if (GET_CODE (x) == USE
4738 && ! (GET_CODE (XEXP (x, 0)) == REG
4739 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4740 canon_reg (XEXP (x, 0), NULL_RTX);
4741 else if (GET_CODE (x) == CALL)
4743 /* The result of apply_change_group can be ignored; see canon_reg. */
4744 canon_reg (x, insn);
4745 apply_change_group ();
4746 fold_rtx (x, insn);
4749 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4750 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4751 is handled specially for this case, and if it isn't set, then there will
4752 be no equivalence for the destination. */
4753 if (n_sets == 1 && REG_NOTES (insn) != 0
4754 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4755 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4756 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4758 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4759 XEXP (tem, 0) = src_eqv;
4762 /* Canonicalize sources and addresses of destinations.
4763 We do this in a separate pass to avoid problems when a MATCH_DUP is
4764 present in the insn pattern. In that case, we want to ensure that
4765 we don't break the duplicate nature of the pattern. So we will replace
4766 both operands at the same time. Otherwise, we would fail to find an
4767 equivalent substitution in the loop calling validate_change below.
4769 We used to suppress canonicalization of DEST if it appears in SRC,
4770 but we don't do this any more. */
4772 for (i = 0; i < n_sets; i++)
4774 rtx dest = SET_DEST (sets[i].rtl);
4775 rtx src = SET_SRC (sets[i].rtl);
4776 rtx new = canon_reg (src, insn);
4777 int insn_code;
4779 sets[i].orig_src = src;
4780 if ((GET_CODE (new) == REG && GET_CODE (src) == REG
4781 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4782 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4783 || (insn_code = recog_memoized (insn)) < 0
4784 || insn_data[insn_code].n_dups > 0)
4785 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4786 else
4787 SET_SRC (sets[i].rtl) = new;
4789 if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT)
4791 validate_change (insn, &XEXP (dest, 1),
4792 canon_reg (XEXP (dest, 1), insn), 1);
4793 validate_change (insn, &XEXP (dest, 2),
4794 canon_reg (XEXP (dest, 2), insn), 1);
4797 while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART
4798 || GET_CODE (dest) == ZERO_EXTRACT
4799 || GET_CODE (dest) == SIGN_EXTRACT)
4800 dest = XEXP (dest, 0);
4802 if (GET_CODE (dest) == MEM)
4803 canon_reg (dest, insn);
4806 /* Now that we have done all the replacements, we can apply the change
4807 group and see if they all work. Note that this will cause some
4808 canonicalizations that would have worked individually not to be applied
4809 because some other canonicalization didn't work, but this should not
4810 occur often.
4812 The result of apply_change_group can be ignored; see canon_reg. */
4814 apply_change_group ();
4816 /* Set sets[i].src_elt to the class each source belongs to.
4817 Detect assignments from or to volatile things
4818 and set set[i] to zero so they will be ignored
4819 in the rest of this function.
4821 Nothing in this loop changes the hash table or the register chains. */
4823 for (i = 0; i < n_sets; i++)
4825 rtx src, dest;
4826 rtx src_folded;
4827 struct table_elt *elt = 0, *p;
4828 enum machine_mode mode;
4829 rtx src_eqv_here;
4830 rtx src_const = 0;
4831 rtx src_related = 0;
4832 struct table_elt *src_const_elt = 0;
4833 int src_cost = MAX_COST;
4834 int src_eqv_cost = MAX_COST;
4835 int src_folded_cost = MAX_COST;
4836 int src_related_cost = MAX_COST;
4837 int src_elt_cost = MAX_COST;
4838 int src_regcost = MAX_COST;
4839 int src_eqv_regcost = MAX_COST;
4840 int src_folded_regcost = MAX_COST;
4841 int src_related_regcost = MAX_COST;
4842 int src_elt_regcost = MAX_COST;
4843 /* Set nonzero if we need to call force_const_mem on with the
4844 contents of src_folded before using it. */
4845 int src_folded_force_flag = 0;
4847 dest = SET_DEST (sets[i].rtl);
4848 src = SET_SRC (sets[i].rtl);
4850 /* If SRC is a constant that has no machine mode,
4851 hash it with the destination's machine mode.
4852 This way we can keep different modes separate. */
4854 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4855 sets[i].mode = mode;
4857 if (src_eqv)
4859 enum machine_mode eqvmode = mode;
4860 if (GET_CODE (dest) == STRICT_LOW_PART)
4861 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4862 do_not_record = 0;
4863 hash_arg_in_memory = 0;
4864 src_eqv_hash = HASH (src_eqv, eqvmode);
4866 /* Find the equivalence class for the equivalent expression. */
4868 if (!do_not_record)
4869 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4871 src_eqv_volatile = do_not_record;
4872 src_eqv_in_memory = hash_arg_in_memory;
4875 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4876 value of the INNER register, not the destination. So it is not
4877 a valid substitution for the source. But save it for later. */
4878 if (GET_CODE (dest) == STRICT_LOW_PART)
4879 src_eqv_here = 0;
4880 else
4881 src_eqv_here = src_eqv;
4883 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4884 simplified result, which may not necessarily be valid. */
4885 src_folded = fold_rtx (src, insn);
4887 #if 0
4888 /* ??? This caused bad code to be generated for the m68k port with -O2.
4889 Suppose src is (CONST_INT -1), and that after truncation src_folded
4890 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4891 At the end we will add src and src_const to the same equivalence
4892 class. We now have 3 and -1 on the same equivalence class. This
4893 causes later instructions to be mis-optimized. */
4894 /* If storing a constant in a bitfield, pre-truncate the constant
4895 so we will be able to record it later. */
4896 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
4897 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
4899 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4901 if (GET_CODE (src) == CONST_INT
4902 && GET_CODE (width) == CONST_INT
4903 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4904 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4905 src_folded
4906 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4907 << INTVAL (width)) - 1));
4909 #endif
4911 /* Compute SRC's hash code, and also notice if it
4912 should not be recorded at all. In that case,
4913 prevent any further processing of this assignment. */
4914 do_not_record = 0;
4915 hash_arg_in_memory = 0;
4917 sets[i].src = src;
4918 sets[i].src_hash = HASH (src, mode);
4919 sets[i].src_volatile = do_not_record;
4920 sets[i].src_in_memory = hash_arg_in_memory;
4922 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4923 a pseudo, do not record SRC. Using SRC as a replacement for
4924 anything else will be incorrect in that situation. Note that
4925 this usually occurs only for stack slots, in which case all the
4926 RTL would be referring to SRC, so we don't lose any optimization
4927 opportunities by not having SRC in the hash table. */
4929 if (GET_CODE (src) == MEM
4930 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4931 && GET_CODE (dest) == REG
4932 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4933 sets[i].src_volatile = 1;
4935 #if 0
4936 /* It is no longer clear why we used to do this, but it doesn't
4937 appear to still be needed. So let's try without it since this
4938 code hurts cse'ing widened ops. */
4939 /* If source is a paradoxical subreg (such as QI treated as an SI),
4940 treat it as volatile. It may do the work of an SI in one context
4941 where the extra bits are not being used, but cannot replace an SI
4942 in general. */
4943 if (GET_CODE (src) == SUBREG
4944 && (GET_MODE_SIZE (GET_MODE (src))
4945 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
4946 sets[i].src_volatile = 1;
4947 #endif
4949 /* Locate all possible equivalent forms for SRC. Try to replace
4950 SRC in the insn with each cheaper equivalent.
4952 We have the following types of equivalents: SRC itself, a folded
4953 version, a value given in a REG_EQUAL note, or a value related
4954 to a constant.
4956 Each of these equivalents may be part of an additional class
4957 of equivalents (if more than one is in the table, they must be in
4958 the same class; we check for this).
4960 If the source is volatile, we don't do any table lookups.
4962 We note any constant equivalent for possible later use in a
4963 REG_NOTE. */
4965 if (!sets[i].src_volatile)
4966 elt = lookup (src, sets[i].src_hash, mode);
4968 sets[i].src_elt = elt;
4970 if (elt && src_eqv_here && src_eqv_elt)
4972 if (elt->first_same_value != src_eqv_elt->first_same_value)
4974 /* The REG_EQUAL is indicating that two formerly distinct
4975 classes are now equivalent. So merge them. */
4976 merge_equiv_classes (elt, src_eqv_elt);
4977 src_eqv_hash = HASH (src_eqv, elt->mode);
4978 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4981 src_eqv_here = 0;
4984 else if (src_eqv_elt)
4985 elt = src_eqv_elt;
4987 /* Try to find a constant somewhere and record it in `src_const'.
4988 Record its table element, if any, in `src_const_elt'. Look in
4989 any known equivalences first. (If the constant is not in the
4990 table, also set `sets[i].src_const_hash'). */
4991 if (elt)
4992 for (p = elt->first_same_value; p; p = p->next_same_value)
4993 if (p->is_const)
4995 src_const = p->exp;
4996 src_const_elt = elt;
4997 break;
5000 if (src_const == 0
5001 && (CONSTANT_P (src_folded)
5002 /* Consider (minus (label_ref L1) (label_ref L2)) as
5003 "constant" here so we will record it. This allows us
5004 to fold switch statements when an ADDR_DIFF_VEC is used. */
5005 || (GET_CODE (src_folded) == MINUS
5006 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5007 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5008 src_const = src_folded, src_const_elt = elt;
5009 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5010 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5012 /* If we don't know if the constant is in the table, get its
5013 hash code and look it up. */
5014 if (src_const && src_const_elt == 0)
5016 sets[i].src_const_hash = HASH (src_const, mode);
5017 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5020 sets[i].src_const = src_const;
5021 sets[i].src_const_elt = src_const_elt;
5023 /* If the constant and our source are both in the table, mark them as
5024 equivalent. Otherwise, if a constant is in the table but the source
5025 isn't, set ELT to it. */
5026 if (src_const_elt && elt
5027 && src_const_elt->first_same_value != elt->first_same_value)
5028 merge_equiv_classes (elt, src_const_elt);
5029 else if (src_const_elt && elt == 0)
5030 elt = src_const_elt;
5032 /* See if there is a register linearly related to a constant
5033 equivalent of SRC. */
5034 if (src_const
5035 && (GET_CODE (src_const) == CONST
5036 || (src_const_elt && src_const_elt->related_value != 0)))
5038 src_related = use_related_value (src_const, src_const_elt);
5039 if (src_related)
5041 struct table_elt *src_related_elt
5042 = lookup (src_related, HASH (src_related, mode), mode);
5043 if (src_related_elt && elt)
5045 if (elt->first_same_value
5046 != src_related_elt->first_same_value)
5047 /* This can occur when we previously saw a CONST
5048 involving a SYMBOL_REF and then see the SYMBOL_REF
5049 twice. Merge the involved classes. */
5050 merge_equiv_classes (elt, src_related_elt);
5052 src_related = 0;
5053 src_related_elt = 0;
5055 else if (src_related_elt && elt == 0)
5056 elt = src_related_elt;
5060 /* See if we have a CONST_INT that is already in a register in a
5061 wider mode. */
5063 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5064 && GET_MODE_CLASS (mode) == MODE_INT
5065 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5067 enum machine_mode wider_mode;
5069 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5070 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5071 && src_related == 0;
5072 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5074 struct table_elt *const_elt
5075 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5077 if (const_elt == 0)
5078 continue;
5080 for (const_elt = const_elt->first_same_value;
5081 const_elt; const_elt = const_elt->next_same_value)
5082 if (GET_CODE (const_elt->exp) == REG)
5084 src_related = gen_lowpart (mode,
5085 const_elt->exp);
5086 break;
5091 /* Another possibility is that we have an AND with a constant in
5092 a mode narrower than a word. If so, it might have been generated
5093 as part of an "if" which would narrow the AND. If we already
5094 have done the AND in a wider mode, we can use a SUBREG of that
5095 value. */
5097 if (flag_expensive_optimizations && ! src_related
5098 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5099 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5101 enum machine_mode tmode;
5102 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5104 for (tmode = GET_MODE_WIDER_MODE (mode);
5105 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5106 tmode = GET_MODE_WIDER_MODE (tmode))
5108 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5109 struct table_elt *larger_elt;
5111 if (inner)
5113 PUT_MODE (new_and, tmode);
5114 XEXP (new_and, 0) = inner;
5115 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5116 if (larger_elt == 0)
5117 continue;
5119 for (larger_elt = larger_elt->first_same_value;
5120 larger_elt; larger_elt = larger_elt->next_same_value)
5121 if (GET_CODE (larger_elt->exp) == REG)
5123 src_related
5124 = gen_lowpart (mode, larger_elt->exp);
5125 break;
5128 if (src_related)
5129 break;
5134 #ifdef LOAD_EXTEND_OP
5135 /* See if a MEM has already been loaded with a widening operation;
5136 if it has, we can use a subreg of that. Many CISC machines
5137 also have such operations, but this is only likely to be
5138 beneficial on these machines. */
5140 if (flag_expensive_optimizations && src_related == 0
5141 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5142 && GET_MODE_CLASS (mode) == MODE_INT
5143 && GET_CODE (src) == MEM && ! do_not_record
5144 && LOAD_EXTEND_OP (mode) != NIL)
5146 enum machine_mode tmode;
5148 /* Set what we are trying to extend and the operation it might
5149 have been extended with. */
5150 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5151 XEXP (memory_extend_rtx, 0) = src;
5153 for (tmode = GET_MODE_WIDER_MODE (mode);
5154 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5155 tmode = GET_MODE_WIDER_MODE (tmode))
5157 struct table_elt *larger_elt;
5159 PUT_MODE (memory_extend_rtx, tmode);
5160 larger_elt = lookup (memory_extend_rtx,
5161 HASH (memory_extend_rtx, tmode), tmode);
5162 if (larger_elt == 0)
5163 continue;
5165 for (larger_elt = larger_elt->first_same_value;
5166 larger_elt; larger_elt = larger_elt->next_same_value)
5167 if (GET_CODE (larger_elt->exp) == REG)
5169 src_related = gen_lowpart (mode,
5170 larger_elt->exp);
5171 break;
5174 if (src_related)
5175 break;
5178 #endif /* LOAD_EXTEND_OP */
5180 if (src == src_folded)
5181 src_folded = 0;
5183 /* At this point, ELT, if nonzero, points to a class of expressions
5184 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5185 and SRC_RELATED, if nonzero, each contain additional equivalent
5186 expressions. Prune these latter expressions by deleting expressions
5187 already in the equivalence class.
5189 Check for an equivalent identical to the destination. If found,
5190 this is the preferred equivalent since it will likely lead to
5191 elimination of the insn. Indicate this by placing it in
5192 `src_related'. */
5194 if (elt)
5195 elt = elt->first_same_value;
5196 for (p = elt; p; p = p->next_same_value)
5198 enum rtx_code code = GET_CODE (p->exp);
5200 /* If the expression is not valid, ignore it. Then we do not
5201 have to check for validity below. In most cases, we can use
5202 `rtx_equal_p', since canonicalization has already been done. */
5203 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, 0))
5204 continue;
5206 /* Also skip paradoxical subregs, unless that's what we're
5207 looking for. */
5208 if (code == SUBREG
5209 && (GET_MODE_SIZE (GET_MODE (p->exp))
5210 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5211 && ! (src != 0
5212 && GET_CODE (src) == SUBREG
5213 && GET_MODE (src) == GET_MODE (p->exp)
5214 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5215 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5216 continue;
5218 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5219 src = 0;
5220 else if (src_folded && GET_CODE (src_folded) == code
5221 && rtx_equal_p (src_folded, p->exp))
5222 src_folded = 0;
5223 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5224 && rtx_equal_p (src_eqv_here, p->exp))
5225 src_eqv_here = 0;
5226 else if (src_related && GET_CODE (src_related) == code
5227 && rtx_equal_p (src_related, p->exp))
5228 src_related = 0;
5230 /* This is the same as the destination of the insns, we want
5231 to prefer it. Copy it to src_related. The code below will
5232 then give it a negative cost. */
5233 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5234 src_related = dest;
5237 /* Find the cheapest valid equivalent, trying all the available
5238 possibilities. Prefer items not in the hash table to ones
5239 that are when they are equal cost. Note that we can never
5240 worsen an insn as the current contents will also succeed.
5241 If we find an equivalent identical to the destination, use it as best,
5242 since this insn will probably be eliminated in that case. */
5243 if (src)
5245 if (rtx_equal_p (src, dest))
5246 src_cost = src_regcost = -1;
5247 else
5249 src_cost = COST (src);
5250 src_regcost = approx_reg_cost (src);
5254 if (src_eqv_here)
5256 if (rtx_equal_p (src_eqv_here, dest))
5257 src_eqv_cost = src_eqv_regcost = -1;
5258 else
5260 src_eqv_cost = COST (src_eqv_here);
5261 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5265 if (src_folded)
5267 if (rtx_equal_p (src_folded, dest))
5268 src_folded_cost = src_folded_regcost = -1;
5269 else
5271 src_folded_cost = COST (src_folded);
5272 src_folded_regcost = approx_reg_cost (src_folded);
5276 if (src_related)
5278 if (rtx_equal_p (src_related, dest))
5279 src_related_cost = src_related_regcost = -1;
5280 else
5282 src_related_cost = COST (src_related);
5283 src_related_regcost = approx_reg_cost (src_related);
5287 /* If this was an indirect jump insn, a known label will really be
5288 cheaper even though it looks more expensive. */
5289 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5290 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5292 /* Terminate loop when replacement made. This must terminate since
5293 the current contents will be tested and will always be valid. */
5294 while (1)
5296 rtx trial;
5298 /* Skip invalid entries. */
5299 while (elt && GET_CODE (elt->exp) != REG
5300 && ! exp_equiv_p (elt->exp, elt->exp, 1, 0))
5301 elt = elt->next_same_value;
5303 /* A paradoxical subreg would be bad here: it'll be the right
5304 size, but later may be adjusted so that the upper bits aren't
5305 what we want. So reject it. */
5306 if (elt != 0
5307 && GET_CODE (elt->exp) == SUBREG
5308 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5309 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5310 /* It is okay, though, if the rtx we're trying to match
5311 will ignore any of the bits we can't predict. */
5312 && ! (src != 0
5313 && GET_CODE (src) == SUBREG
5314 && GET_MODE (src) == GET_MODE (elt->exp)
5315 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5316 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5318 elt = elt->next_same_value;
5319 continue;
5322 if (elt)
5324 src_elt_cost = elt->cost;
5325 src_elt_regcost = elt->regcost;
5328 /* Find cheapest and skip it for the next time. For items
5329 of equal cost, use this order:
5330 src_folded, src, src_eqv, src_related and hash table entry. */
5331 if (src_folded
5332 && preferable (src_folded_cost, src_folded_regcost,
5333 src_cost, src_regcost) <= 0
5334 && preferable (src_folded_cost, src_folded_regcost,
5335 src_eqv_cost, src_eqv_regcost) <= 0
5336 && preferable (src_folded_cost, src_folded_regcost,
5337 src_related_cost, src_related_regcost) <= 0
5338 && preferable (src_folded_cost, src_folded_regcost,
5339 src_elt_cost, src_elt_regcost) <= 0)
5341 trial = src_folded, src_folded_cost = MAX_COST;
5342 if (src_folded_force_flag)
5344 rtx forced = force_const_mem (mode, trial);
5345 if (forced)
5346 trial = forced;
5349 else if (src
5350 && preferable (src_cost, src_regcost,
5351 src_eqv_cost, src_eqv_regcost) <= 0
5352 && preferable (src_cost, src_regcost,
5353 src_related_cost, src_related_regcost) <= 0
5354 && preferable (src_cost, src_regcost,
5355 src_elt_cost, src_elt_regcost) <= 0)
5356 trial = src, src_cost = MAX_COST;
5357 else if (src_eqv_here
5358 && preferable (src_eqv_cost, src_eqv_regcost,
5359 src_related_cost, src_related_regcost) <= 0
5360 && preferable (src_eqv_cost, src_eqv_regcost,
5361 src_elt_cost, src_elt_regcost) <= 0)
5362 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5363 else if (src_related
5364 && preferable (src_related_cost, src_related_regcost,
5365 src_elt_cost, src_elt_regcost) <= 0)
5366 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5367 else
5369 trial = copy_rtx (elt->exp);
5370 elt = elt->next_same_value;
5371 src_elt_cost = MAX_COST;
5374 /* We don't normally have an insn matching (set (pc) (pc)), so
5375 check for this separately here. We will delete such an
5376 insn below.
5378 For other cases such as a table jump or conditional jump
5379 where we know the ultimate target, go ahead and replace the
5380 operand. While that may not make a valid insn, we will
5381 reemit the jump below (and also insert any necessary
5382 barriers). */
5383 if (n_sets == 1 && dest == pc_rtx
5384 && (trial == pc_rtx
5385 || (GET_CODE (trial) == LABEL_REF
5386 && ! condjump_p (insn))))
5388 SET_SRC (sets[i].rtl) = trial;
5389 cse_jumps_altered = 1;
5390 break;
5393 /* Look for a substitution that makes a valid insn. */
5394 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5396 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5398 /* If we just made a substitution inside a libcall, then we
5399 need to make the same substitution in any notes attached
5400 to the RETVAL insn. */
5401 if (libcall_insn
5402 && (GET_CODE (sets[i].orig_src) == REG
5403 || GET_CODE (sets[i].orig_src) == SUBREG
5404 || GET_CODE (sets[i].orig_src) == MEM))
5406 rtx note = find_reg_equal_equiv_note (libcall_insn);
5407 if (note != 0)
5408 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5409 sets[i].orig_src,
5410 copy_rtx (new));
5413 /* The result of apply_change_group can be ignored; see
5414 canon_reg. */
5416 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5417 apply_change_group ();
5418 break;
5421 /* If we previously found constant pool entries for
5422 constants and this is a constant, try making a
5423 pool entry. Put it in src_folded unless we already have done
5424 this since that is where it likely came from. */
5426 else if (constant_pool_entries_cost
5427 && CONSTANT_P (trial)
5428 /* Reject cases that will abort in decode_rtx_const.
5429 On the alpha when simplifying a switch, we get
5430 (const (truncate (minus (label_ref) (label_ref)))). */
5431 && ! (GET_CODE (trial) == CONST
5432 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5433 /* Likewise on IA-64, except without the truncate. */
5434 && ! (GET_CODE (trial) == CONST
5435 && GET_CODE (XEXP (trial, 0)) == MINUS
5436 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5437 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5438 && (src_folded == 0
5439 || (GET_CODE (src_folded) != MEM
5440 && ! src_folded_force_flag))
5441 && GET_MODE_CLASS (mode) != MODE_CC
5442 && mode != VOIDmode)
5444 src_folded_force_flag = 1;
5445 src_folded = trial;
5446 src_folded_cost = constant_pool_entries_cost;
5447 src_folded_regcost = constant_pool_entries_regcost;
5451 src = SET_SRC (sets[i].rtl);
5453 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5454 However, there is an important exception: If both are registers
5455 that are not the head of their equivalence class, replace SET_SRC
5456 with the head of the class. If we do not do this, we will have
5457 both registers live over a portion of the basic block. This way,
5458 their lifetimes will likely abut instead of overlapping. */
5459 if (GET_CODE (dest) == REG
5460 && REGNO_QTY_VALID_P (REGNO (dest)))
5462 int dest_q = REG_QTY (REGNO (dest));
5463 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5465 if (dest_ent->mode == GET_MODE (dest)
5466 && dest_ent->first_reg != REGNO (dest)
5467 && GET_CODE (src) == REG && REGNO (src) == REGNO (dest)
5468 /* Don't do this if the original insn had a hard reg as
5469 SET_SRC or SET_DEST. */
5470 && (GET_CODE (sets[i].src) != REG
5471 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5472 && (GET_CODE (dest) != REG || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5473 /* We can't call canon_reg here because it won't do anything if
5474 SRC is a hard register. */
5476 int src_q = REG_QTY (REGNO (src));
5477 struct qty_table_elem *src_ent = &qty_table[src_q];
5478 int first = src_ent->first_reg;
5479 rtx new_src
5480 = (first >= FIRST_PSEUDO_REGISTER
5481 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5483 /* We must use validate-change even for this, because this
5484 might be a special no-op instruction, suitable only to
5485 tag notes onto. */
5486 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5488 src = new_src;
5489 /* If we had a constant that is cheaper than what we are now
5490 setting SRC to, use that constant. We ignored it when we
5491 thought we could make this into a no-op. */
5492 if (src_const && COST (src_const) < COST (src)
5493 && validate_change (insn, &SET_SRC (sets[i].rtl),
5494 src_const, 0))
5495 src = src_const;
5500 /* If we made a change, recompute SRC values. */
5501 if (src != sets[i].src)
5503 cse_altered = 1;
5504 do_not_record = 0;
5505 hash_arg_in_memory = 0;
5506 sets[i].src = src;
5507 sets[i].src_hash = HASH (src, mode);
5508 sets[i].src_volatile = do_not_record;
5509 sets[i].src_in_memory = hash_arg_in_memory;
5510 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5513 /* If this is a single SET, we are setting a register, and we have an
5514 equivalent constant, we want to add a REG_NOTE. We don't want
5515 to write a REG_EQUAL note for a constant pseudo since verifying that
5516 that pseudo hasn't been eliminated is a pain. Such a note also
5517 won't help anything.
5519 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5520 which can be created for a reference to a compile time computable
5521 entry in a jump table. */
5523 if (n_sets == 1 && src_const && GET_CODE (dest) == REG
5524 && GET_CODE (src_const) != REG
5525 && ! (GET_CODE (src_const) == CONST
5526 && GET_CODE (XEXP (src_const, 0)) == MINUS
5527 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5528 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5530 /* We only want a REG_EQUAL note if src_const != src. */
5531 if (! rtx_equal_p (src, src_const))
5533 /* Make sure that the rtx is not shared. */
5534 src_const = copy_rtx (src_const);
5536 /* Record the actual constant value in a REG_EQUAL note,
5537 making a new one if one does not already exist. */
5538 set_unique_reg_note (insn, REG_EQUAL, src_const);
5542 /* Now deal with the destination. */
5543 do_not_record = 0;
5545 /* Look within any SIGN_EXTRACT or ZERO_EXTRACT
5546 to the MEM or REG within it. */
5547 while (GET_CODE (dest) == SIGN_EXTRACT
5548 || GET_CODE (dest) == ZERO_EXTRACT
5549 || GET_CODE (dest) == SUBREG
5550 || GET_CODE (dest) == STRICT_LOW_PART)
5551 dest = XEXP (dest, 0);
5553 sets[i].inner_dest = dest;
5555 if (GET_CODE (dest) == MEM)
5557 #ifdef PUSH_ROUNDING
5558 /* Stack pushes invalidate the stack pointer. */
5559 rtx addr = XEXP (dest, 0);
5560 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5561 && XEXP (addr, 0) == stack_pointer_rtx)
5562 invalidate (stack_pointer_rtx, Pmode);
5563 #endif
5564 dest = fold_rtx (dest, insn);
5567 /* Compute the hash code of the destination now,
5568 before the effects of this instruction are recorded,
5569 since the register values used in the address computation
5570 are those before this instruction. */
5571 sets[i].dest_hash = HASH (dest, mode);
5573 /* Don't enter a bit-field in the hash table
5574 because the value in it after the store
5575 may not equal what was stored, due to truncation. */
5577 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5578 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
5580 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5582 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5583 && GET_CODE (width) == CONST_INT
5584 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5585 && ! (INTVAL (src_const)
5586 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5587 /* Exception: if the value is constant,
5588 and it won't be truncated, record it. */
5590 else
5592 /* This is chosen so that the destination will be invalidated
5593 but no new value will be recorded.
5594 We must invalidate because sometimes constant
5595 values can be recorded for bitfields. */
5596 sets[i].src_elt = 0;
5597 sets[i].src_volatile = 1;
5598 src_eqv = 0;
5599 src_eqv_elt = 0;
5603 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5604 the insn. */
5605 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5607 /* One less use of the label this insn used to jump to. */
5608 delete_insn (insn);
5609 cse_jumps_altered = 1;
5610 /* No more processing for this set. */
5611 sets[i].rtl = 0;
5614 /* If this SET is now setting PC to a label, we know it used to
5615 be a conditional or computed branch. */
5616 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF)
5618 /* Now emit a BARRIER after the unconditional jump. */
5619 if (NEXT_INSN (insn) == 0
5620 || GET_CODE (NEXT_INSN (insn)) != BARRIER)
5621 emit_barrier_after (insn);
5623 /* We reemit the jump in as many cases as possible just in
5624 case the form of an unconditional jump is significantly
5625 different than a computed jump or conditional jump.
5627 If this insn has multiple sets, then reemitting the
5628 jump is nontrivial. So instead we just force rerecognition
5629 and hope for the best. */
5630 if (n_sets == 1)
5632 rtx new, note;
5634 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5635 JUMP_LABEL (new) = XEXP (src, 0);
5636 LABEL_NUSES (XEXP (src, 0))++;
5638 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5639 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5640 if (note)
5642 XEXP (note, 1) = NULL_RTX;
5643 REG_NOTES (new) = note;
5646 delete_insn (insn);
5647 insn = new;
5649 /* Now emit a BARRIER after the unconditional jump. */
5650 if (NEXT_INSN (insn) == 0
5651 || GET_CODE (NEXT_INSN (insn)) != BARRIER)
5652 emit_barrier_after (insn);
5654 else
5655 INSN_CODE (insn) = -1;
5657 never_reached_warning (insn, NULL);
5659 /* Do not bother deleting any unreachable code,
5660 let jump/flow do that. */
5662 cse_jumps_altered = 1;
5663 sets[i].rtl = 0;
5666 /* If destination is volatile, invalidate it and then do no further
5667 processing for this assignment. */
5669 else if (do_not_record)
5671 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
5672 invalidate (dest, VOIDmode);
5673 else if (GET_CODE (dest) == MEM)
5675 /* Outgoing arguments for a libcall don't
5676 affect any recorded expressions. */
5677 if (! libcall_insn || insn == libcall_insn)
5678 invalidate (dest, VOIDmode);
5680 else if (GET_CODE (dest) == STRICT_LOW_PART
5681 || GET_CODE (dest) == ZERO_EXTRACT)
5682 invalidate (XEXP (dest, 0), GET_MODE (dest));
5683 sets[i].rtl = 0;
5686 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5687 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5689 #ifdef HAVE_cc0
5690 /* If setting CC0, record what it was set to, or a constant, if it
5691 is equivalent to a constant. If it is being set to a floating-point
5692 value, make a COMPARE with the appropriate constant of 0. If we
5693 don't do this, later code can interpret this as a test against
5694 const0_rtx, which can cause problems if we try to put it into an
5695 insn as a floating-point operand. */
5696 if (dest == cc0_rtx)
5698 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5699 this_insn_cc0_mode = mode;
5700 if (FLOAT_MODE_P (mode))
5701 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5702 CONST0_RTX (mode));
5704 #endif
5707 /* Now enter all non-volatile source expressions in the hash table
5708 if they are not already present.
5709 Record their equivalence classes in src_elt.
5710 This way we can insert the corresponding destinations into
5711 the same classes even if the actual sources are no longer in them
5712 (having been invalidated). */
5714 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5715 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5717 struct table_elt *elt;
5718 struct table_elt *classp = sets[0].src_elt;
5719 rtx dest = SET_DEST (sets[0].rtl);
5720 enum machine_mode eqvmode = GET_MODE (dest);
5722 if (GET_CODE (dest) == STRICT_LOW_PART)
5724 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5725 classp = 0;
5727 if (insert_regs (src_eqv, classp, 0))
5729 rehash_using_reg (src_eqv);
5730 src_eqv_hash = HASH (src_eqv, eqvmode);
5732 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5733 elt->in_memory = src_eqv_in_memory;
5734 src_eqv_elt = elt;
5736 /* Check to see if src_eqv_elt is the same as a set source which
5737 does not yet have an elt, and if so set the elt of the set source
5738 to src_eqv_elt. */
5739 for (i = 0; i < n_sets; i++)
5740 if (sets[i].rtl && sets[i].src_elt == 0
5741 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5742 sets[i].src_elt = src_eqv_elt;
5745 for (i = 0; i < n_sets; i++)
5746 if (sets[i].rtl && ! sets[i].src_volatile
5747 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5749 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5751 /* REG_EQUAL in setting a STRICT_LOW_PART
5752 gives an equivalent for the entire destination register,
5753 not just for the subreg being stored in now.
5754 This is a more interesting equivalence, so we arrange later
5755 to treat the entire reg as the destination. */
5756 sets[i].src_elt = src_eqv_elt;
5757 sets[i].src_hash = src_eqv_hash;
5759 else
5761 /* Insert source and constant equivalent into hash table, if not
5762 already present. */
5763 struct table_elt *classp = src_eqv_elt;
5764 rtx src = sets[i].src;
5765 rtx dest = SET_DEST (sets[i].rtl);
5766 enum machine_mode mode
5767 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5769 /* It's possible that we have a source value known to be
5770 constant but don't have a REG_EQUAL note on the insn.
5771 Lack of a note will mean src_eqv_elt will be NULL. This
5772 can happen where we've generated a SUBREG to access a
5773 CONST_INT that is already in a register in a wider mode.
5774 Ensure that the source expression is put in the proper
5775 constant class. */
5776 if (!classp)
5777 classp = sets[i].src_const_elt;
5779 if (sets[i].src_elt == 0)
5781 /* Don't put a hard register source into the table if this is
5782 the last insn of a libcall. In this case, we only need
5783 to put src_eqv_elt in src_elt. */
5784 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5786 struct table_elt *elt;
5788 /* Note that these insert_regs calls cannot remove
5789 any of the src_elt's, because they would have failed to
5790 match if not still valid. */
5791 if (insert_regs (src, classp, 0))
5793 rehash_using_reg (src);
5794 sets[i].src_hash = HASH (src, mode);
5796 elt = insert (src, classp, sets[i].src_hash, mode);
5797 elt->in_memory = sets[i].src_in_memory;
5798 sets[i].src_elt = classp = elt;
5800 else
5801 sets[i].src_elt = classp;
5803 if (sets[i].src_const && sets[i].src_const_elt == 0
5804 && src != sets[i].src_const
5805 && ! rtx_equal_p (sets[i].src_const, src))
5806 sets[i].src_elt = insert (sets[i].src_const, classp,
5807 sets[i].src_const_hash, mode);
5810 else if (sets[i].src_elt == 0)
5811 /* If we did not insert the source into the hash table (e.g., it was
5812 volatile), note the equivalence class for the REG_EQUAL value, if any,
5813 so that the destination goes into that class. */
5814 sets[i].src_elt = src_eqv_elt;
5816 invalidate_from_clobbers (x);
5818 /* Some registers are invalidated by subroutine calls. Memory is
5819 invalidated by non-constant calls. */
5821 if (GET_CODE (insn) == CALL_INSN)
5823 if (! CONST_OR_PURE_CALL_P (insn))
5824 invalidate_memory ();
5825 invalidate_for_call ();
5828 /* Now invalidate everything set by this instruction.
5829 If a SUBREG or other funny destination is being set,
5830 sets[i].rtl is still nonzero, so here we invalidate the reg
5831 a part of which is being set. */
5833 for (i = 0; i < n_sets; i++)
5834 if (sets[i].rtl)
5836 /* We can't use the inner dest, because the mode associated with
5837 a ZERO_EXTRACT is significant. */
5838 rtx dest = SET_DEST (sets[i].rtl);
5840 /* Needed for registers to remove the register from its
5841 previous quantity's chain.
5842 Needed for memory if this is a nonvarying address, unless
5843 we have just done an invalidate_memory that covers even those. */
5844 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
5845 invalidate (dest, VOIDmode);
5846 else if (GET_CODE (dest) == MEM)
5848 /* Outgoing arguments for a libcall don't
5849 affect any recorded expressions. */
5850 if (! libcall_insn || insn == libcall_insn)
5851 invalidate (dest, VOIDmode);
5853 else if (GET_CODE (dest) == STRICT_LOW_PART
5854 || GET_CODE (dest) == ZERO_EXTRACT)
5855 invalidate (XEXP (dest, 0), GET_MODE (dest));
5858 /* A volatile ASM invalidates everything. */
5859 if (GET_CODE (insn) == INSN
5860 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5861 && MEM_VOLATILE_P (PATTERN (insn)))
5862 flush_hash_table ();
5864 /* Make sure registers mentioned in destinations
5865 are safe for use in an expression to be inserted.
5866 This removes from the hash table
5867 any invalid entry that refers to one of these registers.
5869 We don't care about the return value from mention_regs because
5870 we are going to hash the SET_DEST values unconditionally. */
5872 for (i = 0; i < n_sets; i++)
5874 if (sets[i].rtl)
5876 rtx x = SET_DEST (sets[i].rtl);
5878 if (GET_CODE (x) != REG)
5879 mention_regs (x);
5880 else
5882 /* We used to rely on all references to a register becoming
5883 inaccessible when a register changes to a new quantity,
5884 since that changes the hash code. However, that is not
5885 safe, since after HASH_SIZE new quantities we get a
5886 hash 'collision' of a register with its own invalid
5887 entries. And since SUBREGs have been changed not to
5888 change their hash code with the hash code of the register,
5889 it wouldn't work any longer at all. So we have to check
5890 for any invalid references lying around now.
5891 This code is similar to the REG case in mention_regs,
5892 but it knows that reg_tick has been incremented, and
5893 it leaves reg_in_table as -1 . */
5894 unsigned int regno = REGNO (x);
5895 unsigned int endregno
5896 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
5897 : hard_regno_nregs[regno][GET_MODE (x)]);
5898 unsigned int i;
5900 for (i = regno; i < endregno; i++)
5902 if (REG_IN_TABLE (i) >= 0)
5904 remove_invalid_refs (i);
5905 REG_IN_TABLE (i) = -1;
5912 /* We may have just removed some of the src_elt's from the hash table.
5913 So replace each one with the current head of the same class. */
5915 for (i = 0; i < n_sets; i++)
5916 if (sets[i].rtl)
5918 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5919 /* If elt was removed, find current head of same class,
5920 or 0 if nothing remains of that class. */
5922 struct table_elt *elt = sets[i].src_elt;
5924 while (elt && elt->prev_same_value)
5925 elt = elt->prev_same_value;
5927 while (elt && elt->first_same_value == 0)
5928 elt = elt->next_same_value;
5929 sets[i].src_elt = elt ? elt->first_same_value : 0;
5933 /* Now insert the destinations into their equivalence classes. */
5935 for (i = 0; i < n_sets; i++)
5936 if (sets[i].rtl)
5938 rtx dest = SET_DEST (sets[i].rtl);
5939 rtx inner_dest = sets[i].inner_dest;
5940 struct table_elt *elt;
5942 /* Don't record value if we are not supposed to risk allocating
5943 floating-point values in registers that might be wider than
5944 memory. */
5945 if ((flag_float_store
5946 && GET_CODE (dest) == MEM
5947 && FLOAT_MODE_P (GET_MODE (dest)))
5948 /* Don't record BLKmode values, because we don't know the
5949 size of it, and can't be sure that other BLKmode values
5950 have the same or smaller size. */
5951 || GET_MODE (dest) == BLKmode
5952 /* Don't record values of destinations set inside a libcall block
5953 since we might delete the libcall. Things should have been set
5954 up so we won't want to reuse such a value, but we play it safe
5955 here. */
5956 || libcall_insn
5957 /* If we didn't put a REG_EQUAL value or a source into the hash
5958 table, there is no point is recording DEST. */
5959 || sets[i].src_elt == 0
5960 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5961 or SIGN_EXTEND, don't record DEST since it can cause
5962 some tracking to be wrong.
5964 ??? Think about this more later. */
5965 || (GET_CODE (dest) == SUBREG
5966 && (GET_MODE_SIZE (GET_MODE (dest))
5967 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5968 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5969 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5970 continue;
5972 /* STRICT_LOW_PART isn't part of the value BEING set,
5973 and neither is the SUBREG inside it.
5974 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5975 if (GET_CODE (dest) == STRICT_LOW_PART)
5976 dest = SUBREG_REG (XEXP (dest, 0));
5978 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
5979 /* Registers must also be inserted into chains for quantities. */
5980 if (insert_regs (dest, sets[i].src_elt, 1))
5982 /* If `insert_regs' changes something, the hash code must be
5983 recalculated. */
5984 rehash_using_reg (dest);
5985 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5988 if (GET_CODE (inner_dest) == MEM
5989 && GET_CODE (XEXP (inner_dest, 0)) == ADDRESSOF)
5990 /* Given (SET (MEM (ADDRESSOF (X))) Y) we don't want to say
5991 that (MEM (ADDRESSOF (X))) is equivalent to Y.
5992 Consider the case in which the address of the MEM is
5993 passed to a function, which alters the MEM. Then, if we
5994 later use Y instead of the MEM we'll miss the update. */
5995 elt = insert (dest, 0, sets[i].dest_hash, GET_MODE (dest));
5996 else
5997 elt = insert (dest, sets[i].src_elt,
5998 sets[i].dest_hash, GET_MODE (dest));
6000 elt->in_memory = (GET_CODE (sets[i].inner_dest) == MEM
6001 && (! RTX_UNCHANGING_P (sets[i].inner_dest)
6002 || fixed_base_plus_p (XEXP (sets[i].inner_dest,
6003 0))));
6005 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6006 narrower than M2, and both M1 and M2 are the same number of words,
6007 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6008 make that equivalence as well.
6010 However, BAR may have equivalences for which gen_lowpart
6011 will produce a simpler value than gen_lowpart applied to
6012 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6013 BAR's equivalences. If we don't get a simplified form, make
6014 the SUBREG. It will not be used in an equivalence, but will
6015 cause two similar assignments to be detected.
6017 Note the loop below will find SUBREG_REG (DEST) since we have
6018 already entered SRC and DEST of the SET in the table. */
6020 if (GET_CODE (dest) == SUBREG
6021 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6022 / UNITS_PER_WORD)
6023 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6024 && (GET_MODE_SIZE (GET_MODE (dest))
6025 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6026 && sets[i].src_elt != 0)
6028 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6029 struct table_elt *elt, *classp = 0;
6031 for (elt = sets[i].src_elt->first_same_value; elt;
6032 elt = elt->next_same_value)
6034 rtx new_src = 0;
6035 unsigned src_hash;
6036 struct table_elt *src_elt;
6037 int byte = 0;
6039 /* Ignore invalid entries. */
6040 if (GET_CODE (elt->exp) != REG
6041 && ! exp_equiv_p (elt->exp, elt->exp, 1, 0))
6042 continue;
6044 /* We may have already been playing subreg games. If the
6045 mode is already correct for the destination, use it. */
6046 if (GET_MODE (elt->exp) == new_mode)
6047 new_src = elt->exp;
6048 else
6050 /* Calculate big endian correction for the SUBREG_BYTE.
6051 We have already checked that M1 (GET_MODE (dest))
6052 is not narrower than M2 (new_mode). */
6053 if (BYTES_BIG_ENDIAN)
6054 byte = (GET_MODE_SIZE (GET_MODE (dest))
6055 - GET_MODE_SIZE (new_mode));
6057 new_src = simplify_gen_subreg (new_mode, elt->exp,
6058 GET_MODE (dest), byte);
6061 /* The call to simplify_gen_subreg fails if the value
6062 is VOIDmode, yet we can't do any simplification, e.g.
6063 for EXPR_LISTs denoting function call results.
6064 It is invalid to construct a SUBREG with a VOIDmode
6065 SUBREG_REG, hence a zero new_src means we can't do
6066 this substitution. */
6067 if (! new_src)
6068 continue;
6070 src_hash = HASH (new_src, new_mode);
6071 src_elt = lookup (new_src, src_hash, new_mode);
6073 /* Put the new source in the hash table is if isn't
6074 already. */
6075 if (src_elt == 0)
6077 if (insert_regs (new_src, classp, 0))
6079 rehash_using_reg (new_src);
6080 src_hash = HASH (new_src, new_mode);
6082 src_elt = insert (new_src, classp, src_hash, new_mode);
6083 src_elt->in_memory = elt->in_memory;
6085 else if (classp && classp != src_elt->first_same_value)
6086 /* Show that two things that we've seen before are
6087 actually the same. */
6088 merge_equiv_classes (src_elt, classp);
6090 classp = src_elt->first_same_value;
6091 /* Ignore invalid entries. */
6092 while (classp
6093 && GET_CODE (classp->exp) != REG
6094 && ! exp_equiv_p (classp->exp, classp->exp, 1, 0))
6095 classp = classp->next_same_value;
6100 /* Special handling for (set REG0 REG1) where REG0 is the
6101 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6102 be used in the sequel, so (if easily done) change this insn to
6103 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6104 that computed their value. Then REG1 will become a dead store
6105 and won't cloud the situation for later optimizations.
6107 Do not make this change if REG1 is a hard register, because it will
6108 then be used in the sequel and we may be changing a two-operand insn
6109 into a three-operand insn.
6111 Also do not do this if we are operating on a copy of INSN.
6113 Also don't do this if INSN ends a libcall; this would cause an unrelated
6114 register to be set in the middle of a libcall, and we then get bad code
6115 if the libcall is deleted. */
6117 if (n_sets == 1 && sets[0].rtl && GET_CODE (SET_DEST (sets[0].rtl)) == REG
6118 && NEXT_INSN (PREV_INSN (insn)) == insn
6119 && GET_CODE (SET_SRC (sets[0].rtl)) == REG
6120 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6121 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6123 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6124 struct qty_table_elem *src_ent = &qty_table[src_q];
6126 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6127 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6129 rtx prev = insn;
6130 /* Scan for the previous nonnote insn, but stop at a basic
6131 block boundary. */
6134 prev = PREV_INSN (prev);
6136 while (prev && GET_CODE (prev) == NOTE
6137 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6139 /* Do not swap the registers around if the previous instruction
6140 attaches a REG_EQUIV note to REG1.
6142 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6143 from the pseudo that originally shadowed an incoming argument
6144 to another register. Some uses of REG_EQUIV might rely on it
6145 being attached to REG1 rather than REG2.
6147 This section previously turned the REG_EQUIV into a REG_EQUAL
6148 note. We cannot do that because REG_EQUIV may provide an
6149 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6151 if (prev != 0 && GET_CODE (prev) == INSN
6152 && GET_CODE (PATTERN (prev)) == SET
6153 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6154 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6156 rtx dest = SET_DEST (sets[0].rtl);
6157 rtx src = SET_SRC (sets[0].rtl);
6158 rtx note;
6160 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6161 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6162 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6163 apply_change_group ();
6165 /* If INSN has a REG_EQUAL note, and this note mentions
6166 REG0, then we must delete it, because the value in
6167 REG0 has changed. If the note's value is REG1, we must
6168 also delete it because that is now this insn's dest. */
6169 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6170 if (note != 0
6171 && (reg_mentioned_p (dest, XEXP (note, 0))
6172 || rtx_equal_p (src, XEXP (note, 0))))
6173 remove_note (insn, note);
6178 /* If this is a conditional jump insn, record any known equivalences due to
6179 the condition being tested. */
6181 last_jump_equiv_class = 0;
6182 if (GET_CODE (insn) == JUMP_INSN
6183 && n_sets == 1 && GET_CODE (x) == SET
6184 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6185 record_jump_equiv (insn, 0);
6187 #ifdef HAVE_cc0
6188 /* If the previous insn set CC0 and this insn no longer references CC0,
6189 delete the previous insn. Here we use the fact that nothing expects CC0
6190 to be valid over an insn, which is true until the final pass. */
6191 if (prev_insn && GET_CODE (prev_insn) == INSN
6192 && (tem = single_set (prev_insn)) != 0
6193 && SET_DEST (tem) == cc0_rtx
6194 && ! reg_mentioned_p (cc0_rtx, x))
6195 delete_insn (prev_insn);
6197 prev_insn_cc0 = this_insn_cc0;
6198 prev_insn_cc0_mode = this_insn_cc0_mode;
6199 prev_insn = insn;
6200 #endif
6203 /* Remove from the hash table all expressions that reference memory. */
6205 static void
6206 invalidate_memory (void)
6208 int i;
6209 struct table_elt *p, *next;
6211 for (i = 0; i < HASH_SIZE; i++)
6212 for (p = table[i]; p; p = next)
6214 next = p->next_same_hash;
6215 if (p->in_memory)
6216 remove_from_table (p, i);
6220 /* If ADDR is an address that implicitly affects the stack pointer, return
6221 1 and update the register tables to show the effect. Else, return 0. */
6223 static int
6224 addr_affects_sp_p (rtx addr)
6226 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6227 && GET_CODE (XEXP (addr, 0)) == REG
6228 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6230 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6232 REG_TICK (STACK_POINTER_REGNUM)++;
6233 /* Is it possible to use a subreg of SP? */
6234 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6237 /* This should be *very* rare. */
6238 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6239 invalidate (stack_pointer_rtx, VOIDmode);
6241 return 1;
6244 return 0;
6247 /* Perform invalidation on the basis of everything about an insn
6248 except for invalidating the actual places that are SET in it.
6249 This includes the places CLOBBERed, and anything that might
6250 alias with something that is SET or CLOBBERed.
6252 X is the pattern of the insn. */
6254 static void
6255 invalidate_from_clobbers (rtx x)
6257 if (GET_CODE (x) == CLOBBER)
6259 rtx ref = XEXP (x, 0);
6260 if (ref)
6262 if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG
6263 || GET_CODE (ref) == MEM)
6264 invalidate (ref, VOIDmode);
6265 else if (GET_CODE (ref) == STRICT_LOW_PART
6266 || GET_CODE (ref) == ZERO_EXTRACT)
6267 invalidate (XEXP (ref, 0), GET_MODE (ref));
6270 else if (GET_CODE (x) == PARALLEL)
6272 int i;
6273 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6275 rtx y = XVECEXP (x, 0, i);
6276 if (GET_CODE (y) == CLOBBER)
6278 rtx ref = XEXP (y, 0);
6279 if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG
6280 || GET_CODE (ref) == MEM)
6281 invalidate (ref, VOIDmode);
6282 else if (GET_CODE (ref) == STRICT_LOW_PART
6283 || GET_CODE (ref) == ZERO_EXTRACT)
6284 invalidate (XEXP (ref, 0), GET_MODE (ref));
6290 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6291 and replace any registers in them with either an equivalent constant
6292 or the canonical form of the register. If we are inside an address,
6293 only do this if the address remains valid.
6295 OBJECT is 0 except when within a MEM in which case it is the MEM.
6297 Return the replacement for X. */
6299 static rtx
6300 cse_process_notes (rtx x, rtx object)
6302 enum rtx_code code = GET_CODE (x);
6303 const char *fmt = GET_RTX_FORMAT (code);
6304 int i;
6306 switch (code)
6308 case CONST_INT:
6309 case CONST:
6310 case SYMBOL_REF:
6311 case LABEL_REF:
6312 case CONST_DOUBLE:
6313 case CONST_VECTOR:
6314 case PC:
6315 case CC0:
6316 case LO_SUM:
6317 return x;
6319 case MEM:
6320 validate_change (x, &XEXP (x, 0),
6321 cse_process_notes (XEXP (x, 0), x), 0);
6322 return x;
6324 case EXPR_LIST:
6325 case INSN_LIST:
6326 if (REG_NOTE_KIND (x) == REG_EQUAL)
6327 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6328 if (XEXP (x, 1))
6329 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6330 return x;
6332 case SIGN_EXTEND:
6333 case ZERO_EXTEND:
6334 case SUBREG:
6336 rtx new = cse_process_notes (XEXP (x, 0), object);
6337 /* We don't substitute VOIDmode constants into these rtx,
6338 since they would impede folding. */
6339 if (GET_MODE (new) != VOIDmode)
6340 validate_change (object, &XEXP (x, 0), new, 0);
6341 return x;
6344 case REG:
6345 i = REG_QTY (REGNO (x));
6347 /* Return a constant or a constant register. */
6348 if (REGNO_QTY_VALID_P (REGNO (x)))
6350 struct qty_table_elem *ent = &qty_table[i];
6352 if (ent->const_rtx != NULL_RTX
6353 && (CONSTANT_P (ent->const_rtx)
6354 || GET_CODE (ent->const_rtx) == REG))
6356 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6357 if (new)
6358 return new;
6362 /* Otherwise, canonicalize this register. */
6363 return canon_reg (x, NULL_RTX);
6365 default:
6366 break;
6369 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6370 if (fmt[i] == 'e')
6371 validate_change (object, &XEXP (x, i),
6372 cse_process_notes (XEXP (x, i), object), 0);
6374 return x;
6377 /* Find common subexpressions between the end test of a loop and the beginning
6378 of the loop. LOOP_START is the CODE_LABEL at the start of a loop.
6380 Often we have a loop where an expression in the exit test is used
6381 in the body of the loop. For example "while (*p) *q++ = *p++;".
6382 Because of the way we duplicate the loop exit test in front of the loop,
6383 however, we don't detect that common subexpression. This will be caught
6384 when global cse is implemented, but this is a quite common case.
6386 This function handles the most common cases of these common expressions.
6387 It is called after we have processed the basic block ending with the
6388 NOTE_INSN_LOOP_END note that ends a loop and the previous JUMP_INSN
6389 jumps to a label used only once. */
6391 static void
6392 cse_around_loop (rtx loop_start)
6394 rtx insn;
6395 int i;
6396 struct table_elt *p;
6398 /* If the jump at the end of the loop doesn't go to the start, we don't
6399 do anything. */
6400 for (insn = PREV_INSN (loop_start);
6401 insn && (GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) >= 0);
6402 insn = PREV_INSN (insn))
6405 if (insn == 0
6406 || GET_CODE (insn) != NOTE
6407 || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG)
6408 return;
6410 /* If the last insn of the loop (the end test) was an NE comparison,
6411 we will interpret it as an EQ comparison, since we fell through
6412 the loop. Any equivalences resulting from that comparison are
6413 therefore not valid and must be invalidated. */
6414 if (last_jump_equiv_class)
6415 for (p = last_jump_equiv_class->first_same_value; p;
6416 p = p->next_same_value)
6418 if (GET_CODE (p->exp) == MEM || GET_CODE (p->exp) == REG
6419 || (GET_CODE (p->exp) == SUBREG
6420 && GET_CODE (SUBREG_REG (p->exp)) == REG))
6421 invalidate (p->exp, VOIDmode);
6422 else if (GET_CODE (p->exp) == STRICT_LOW_PART
6423 || GET_CODE (p->exp) == ZERO_EXTRACT)
6424 invalidate (XEXP (p->exp, 0), GET_MODE (p->exp));
6427 /* Process insns starting after LOOP_START until we hit a CALL_INSN or
6428 a CODE_LABEL (we could handle a CALL_INSN, but it isn't worth it).
6430 The only thing we do with SET_DEST is invalidate entries, so we
6431 can safely process each SET in order. It is slightly less efficient
6432 to do so, but we only want to handle the most common cases.
6434 The gen_move_insn call in cse_set_around_loop may create new pseudos.
6435 These pseudos won't have valid entries in any of the tables indexed
6436 by register number, such as reg_qty. We avoid out-of-range array
6437 accesses by not processing any instructions created after cse started. */
6439 for (insn = NEXT_INSN (loop_start);
6440 GET_CODE (insn) != CALL_INSN && GET_CODE (insn) != CODE_LABEL
6441 && INSN_UID (insn) < max_insn_uid
6442 && ! (GET_CODE (insn) == NOTE
6443 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END);
6444 insn = NEXT_INSN (insn))
6446 if (INSN_P (insn)
6447 && (GET_CODE (PATTERN (insn)) == SET
6448 || GET_CODE (PATTERN (insn)) == CLOBBER))
6449 cse_set_around_loop (PATTERN (insn), insn, loop_start);
6450 else if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == PARALLEL)
6451 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6452 if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == SET
6453 || GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER)
6454 cse_set_around_loop (XVECEXP (PATTERN (insn), 0, i), insn,
6455 loop_start);
6459 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6460 since they are done elsewhere. This function is called via note_stores. */
6462 static void
6463 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6465 enum rtx_code code = GET_CODE (dest);
6467 if (code == MEM
6468 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6469 /* There are times when an address can appear varying and be a PLUS
6470 during this scan when it would be a fixed address were we to know
6471 the proper equivalences. So invalidate all memory if there is
6472 a BLKmode or nonscalar memory reference or a reference to a
6473 variable address. */
6474 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6475 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6477 invalidate_memory ();
6478 return;
6481 if (GET_CODE (set) == CLOBBER
6482 || CC0_P (dest)
6483 || dest == pc_rtx)
6484 return;
6486 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6487 invalidate (XEXP (dest, 0), GET_MODE (dest));
6488 else if (code == REG || code == SUBREG || code == MEM)
6489 invalidate (dest, VOIDmode);
6492 /* Invalidate all insns from START up to the end of the function or the
6493 next label. This called when we wish to CSE around a block that is
6494 conditionally executed. */
6496 static void
6497 invalidate_skipped_block (rtx start)
6499 rtx insn;
6501 for (insn = start; insn && GET_CODE (insn) != CODE_LABEL;
6502 insn = NEXT_INSN (insn))
6504 if (! INSN_P (insn))
6505 continue;
6507 if (GET_CODE (insn) == CALL_INSN)
6509 if (! CONST_OR_PURE_CALL_P (insn))
6510 invalidate_memory ();
6511 invalidate_for_call ();
6514 invalidate_from_clobbers (PATTERN (insn));
6515 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6519 /* If modifying X will modify the value in *DATA (which is really an
6520 `rtx *'), indicate that fact by setting the pointed to value to
6521 NULL_RTX. */
6523 static void
6524 cse_check_loop_start (rtx x, rtx set ATTRIBUTE_UNUSED, void *data)
6526 rtx *cse_check_loop_start_value = (rtx *) data;
6528 if (*cse_check_loop_start_value == NULL_RTX
6529 || GET_CODE (x) == CC0 || GET_CODE (x) == PC)
6530 return;
6532 if ((GET_CODE (x) == MEM && GET_CODE (*cse_check_loop_start_value) == MEM)
6533 || reg_overlap_mentioned_p (x, *cse_check_loop_start_value))
6534 *cse_check_loop_start_value = NULL_RTX;
6537 /* X is a SET or CLOBBER contained in INSN that was found near the start of
6538 a loop that starts with the label at LOOP_START.
6540 If X is a SET, we see if its SET_SRC is currently in our hash table.
6541 If so, we see if it has a value equal to some register used only in the
6542 loop exit code (as marked by jump.c).
6544 If those two conditions are true, we search backwards from the start of
6545 the loop to see if that same value was loaded into a register that still
6546 retains its value at the start of the loop.
6548 If so, we insert an insn after the load to copy the destination of that
6549 load into the equivalent register and (try to) replace our SET_SRC with that
6550 register.
6552 In any event, we invalidate whatever this SET or CLOBBER modifies. */
6554 static void
6555 cse_set_around_loop (rtx x, rtx insn, rtx loop_start)
6557 struct table_elt *src_elt;
6559 /* If this is a SET, see if we can replace SET_SRC, but ignore SETs that
6560 are setting PC or CC0 or whose SET_SRC is already a register. */
6561 if (GET_CODE (x) == SET
6562 && GET_CODE (SET_DEST (x)) != PC && GET_CODE (SET_DEST (x)) != CC0
6563 && GET_CODE (SET_SRC (x)) != REG)
6565 src_elt = lookup (SET_SRC (x),
6566 HASH (SET_SRC (x), GET_MODE (SET_DEST (x))),
6567 GET_MODE (SET_DEST (x)));
6569 if (src_elt)
6570 for (src_elt = src_elt->first_same_value; src_elt;
6571 src_elt = src_elt->next_same_value)
6572 if (GET_CODE (src_elt->exp) == REG && REG_LOOP_TEST_P (src_elt->exp)
6573 && COST (src_elt->exp) < COST (SET_SRC (x)))
6575 rtx p, set;
6577 /* Look for an insn in front of LOOP_START that sets
6578 something in the desired mode to SET_SRC (x) before we hit
6579 a label or CALL_INSN. */
6581 for (p = prev_nonnote_insn (loop_start);
6582 p && GET_CODE (p) != CALL_INSN
6583 && GET_CODE (p) != CODE_LABEL;
6584 p = prev_nonnote_insn (p))
6585 if ((set = single_set (p)) != 0
6586 && GET_CODE (SET_DEST (set)) == REG
6587 && GET_MODE (SET_DEST (set)) == src_elt->mode
6588 && rtx_equal_p (SET_SRC (set), SET_SRC (x)))
6590 /* We now have to ensure that nothing between P
6591 and LOOP_START modified anything referenced in
6592 SET_SRC (x). We know that nothing within the loop
6593 can modify it, or we would have invalidated it in
6594 the hash table. */
6595 rtx q;
6596 rtx cse_check_loop_start_value = SET_SRC (x);
6597 for (q = p; q != loop_start; q = NEXT_INSN (q))
6598 if (INSN_P (q))
6599 note_stores (PATTERN (q),
6600 cse_check_loop_start,
6601 &cse_check_loop_start_value);
6603 /* If nothing was changed and we can replace our
6604 SET_SRC, add an insn after P to copy its destination
6605 to what we will be replacing SET_SRC with. */
6606 if (cse_check_loop_start_value
6607 && single_set (p)
6608 && !can_throw_internal (insn)
6609 && validate_change (insn, &SET_SRC (x),
6610 src_elt->exp, 0))
6612 /* If this creates new pseudos, this is unsafe,
6613 because the regno of new pseudo is unsuitable
6614 to index into reg_qty when cse_insn processes
6615 the new insn. Therefore, if a new pseudo was
6616 created, discard this optimization. */
6617 int nregs = max_reg_num ();
6618 rtx move
6619 = gen_move_insn (src_elt->exp, SET_DEST (set));
6620 if (nregs != max_reg_num ())
6622 if (! validate_change (insn, &SET_SRC (x),
6623 SET_SRC (set), 0))
6624 abort ();
6626 else
6628 if (CONSTANT_P (SET_SRC (set))
6629 && ! find_reg_equal_equiv_note (insn))
6630 set_unique_reg_note (insn, REG_EQUAL,
6631 SET_SRC (set));
6632 if (control_flow_insn_p (p))
6633 /* p can cause a control flow transfer so it
6634 is the last insn of a basic block. We can't
6635 therefore use emit_insn_after. */
6636 emit_insn_before (move, next_nonnote_insn (p));
6637 else
6638 emit_insn_after (move, p);
6641 break;
6646 /* Deal with the destination of X affecting the stack pointer. */
6647 addr_affects_sp_p (SET_DEST (x));
6649 /* See comment on similar code in cse_insn for explanation of these
6650 tests. */
6651 if (GET_CODE (SET_DEST (x)) == REG || GET_CODE (SET_DEST (x)) == SUBREG
6652 || GET_CODE (SET_DEST (x)) == MEM)
6653 invalidate (SET_DEST (x), VOIDmode);
6654 else if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
6655 || GET_CODE (SET_DEST (x)) == ZERO_EXTRACT)
6656 invalidate (XEXP (SET_DEST (x), 0), GET_MODE (SET_DEST (x)));
6659 /* Find the end of INSN's basic block and return its range,
6660 the total number of SETs in all the insns of the block, the last insn of the
6661 block, and the branch path.
6663 The branch path indicates which branches should be followed. If a nonzero
6664 path size is specified, the block should be rescanned and a different set
6665 of branches will be taken. The branch path is only used if
6666 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6668 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6669 used to describe the block. It is filled in with the information about
6670 the current block. The incoming structure's branch path, if any, is used
6671 to construct the output branch path. */
6673 static void
6674 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6675 int follow_jumps, int after_loop, int skip_blocks)
6677 rtx p = insn, q;
6678 int nsets = 0;
6679 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6680 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6681 int path_size = data->path_size;
6682 int path_entry = 0;
6683 int i;
6685 /* Update the previous branch path, if any. If the last branch was
6686 previously TAKEN, mark it NOT_TAKEN. If it was previously NOT_TAKEN,
6687 shorten the path by one and look at the previous branch. We know that
6688 at least one branch must have been taken if PATH_SIZE is nonzero. */
6689 while (path_size > 0)
6691 if (data->path[path_size - 1].status != NOT_TAKEN)
6693 data->path[path_size - 1].status = NOT_TAKEN;
6694 break;
6696 else
6697 path_size--;
6700 /* If the first instruction is marked with QImode, that means we've
6701 already processed this block. Our caller will look at DATA->LAST
6702 to figure out where to go next. We want to return the next block
6703 in the instruction stream, not some branched-to block somewhere
6704 else. We accomplish this by pretending our called forbid us to
6705 follow jumps, or skip blocks. */
6706 if (GET_MODE (insn) == QImode)
6707 follow_jumps = skip_blocks = 0;
6709 /* Scan to end of this basic block. */
6710 while (p && GET_CODE (p) != CODE_LABEL)
6712 /* Don't cse out the end of a loop. This makes a difference
6713 only for the unusual loops that always execute at least once;
6714 all other loops have labels there so we will stop in any case.
6715 Cse'ing out the end of the loop is dangerous because it
6716 might cause an invariant expression inside the loop
6717 to be reused after the end of the loop. This would make it
6718 hard to move the expression out of the loop in loop.c,
6719 especially if it is one of several equivalent expressions
6720 and loop.c would like to eliminate it.
6722 If we are running after loop.c has finished, we can ignore
6723 the NOTE_INSN_LOOP_END. */
6725 if (! after_loop && GET_CODE (p) == NOTE
6726 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
6727 break;
6729 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6730 the regs restored by the longjmp come from
6731 a later time than the setjmp. */
6732 if (PREV_INSN (p) && GET_CODE (PREV_INSN (p)) == CALL_INSN
6733 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6734 break;
6736 /* A PARALLEL can have lots of SETs in it,
6737 especially if it is really an ASM_OPERANDS. */
6738 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6739 nsets += XVECLEN (PATTERN (p), 0);
6740 else if (GET_CODE (p) != NOTE)
6741 nsets += 1;
6743 /* Ignore insns made by CSE; they cannot affect the boundaries of
6744 the basic block. */
6746 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6747 high_cuid = INSN_CUID (p);
6748 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6749 low_cuid = INSN_CUID (p);
6751 /* See if this insn is in our branch path. If it is and we are to
6752 take it, do so. */
6753 if (path_entry < path_size && data->path[path_entry].branch == p)
6755 if (data->path[path_entry].status != NOT_TAKEN)
6756 p = JUMP_LABEL (p);
6758 /* Point to next entry in path, if any. */
6759 path_entry++;
6762 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6763 was specified, we haven't reached our maximum path length, there are
6764 insns following the target of the jump, this is the only use of the
6765 jump label, and the target label is preceded by a BARRIER.
6767 Alternatively, we can follow the jump if it branches around a
6768 block of code and there are no other branches into the block.
6769 In this case invalidate_skipped_block will be called to invalidate any
6770 registers set in the block when following the jump. */
6772 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6773 && GET_CODE (p) == JUMP_INSN
6774 && GET_CODE (PATTERN (p)) == SET
6775 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6776 && JUMP_LABEL (p) != 0
6777 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6778 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6780 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6781 if ((GET_CODE (q) != NOTE
6782 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6783 || (PREV_INSN (q) && GET_CODE (PREV_INSN (q)) == CALL_INSN
6784 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6785 && (GET_CODE (q) != CODE_LABEL || LABEL_NUSES (q) != 0))
6786 break;
6788 /* If we ran into a BARRIER, this code is an extension of the
6789 basic block when the branch is taken. */
6790 if (follow_jumps && q != 0 && GET_CODE (q) == BARRIER)
6792 /* Don't allow ourself to keep walking around an
6793 always-executed loop. */
6794 if (next_real_insn (q) == next)
6796 p = NEXT_INSN (p);
6797 continue;
6800 /* Similarly, don't put a branch in our path more than once. */
6801 for (i = 0; i < path_entry; i++)
6802 if (data->path[i].branch == p)
6803 break;
6805 if (i != path_entry)
6806 break;
6808 data->path[path_entry].branch = p;
6809 data->path[path_entry++].status = TAKEN;
6811 /* This branch now ends our path. It was possible that we
6812 didn't see this branch the last time around (when the
6813 insn in front of the target was a JUMP_INSN that was
6814 turned into a no-op). */
6815 path_size = path_entry;
6817 p = JUMP_LABEL (p);
6818 /* Mark block so we won't scan it again later. */
6819 PUT_MODE (NEXT_INSN (p), QImode);
6821 /* Detect a branch around a block of code. */
6822 else if (skip_blocks && q != 0 && GET_CODE (q) != CODE_LABEL)
6824 rtx tmp;
6826 if (next_real_insn (q) == next)
6828 p = NEXT_INSN (p);
6829 continue;
6832 for (i = 0; i < path_entry; i++)
6833 if (data->path[i].branch == p)
6834 break;
6836 if (i != path_entry)
6837 break;
6839 /* This is no_labels_between_p (p, q) with an added check for
6840 reaching the end of a function (in case Q precedes P). */
6841 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6842 if (GET_CODE (tmp) == CODE_LABEL)
6843 break;
6845 if (tmp == q)
6847 data->path[path_entry].branch = p;
6848 data->path[path_entry++].status = AROUND;
6850 path_size = path_entry;
6852 p = JUMP_LABEL (p);
6853 /* Mark block so we won't scan it again later. */
6854 PUT_MODE (NEXT_INSN (p), QImode);
6858 p = NEXT_INSN (p);
6861 data->low_cuid = low_cuid;
6862 data->high_cuid = high_cuid;
6863 data->nsets = nsets;
6864 data->last = p;
6866 /* If all jumps in the path are not taken, set our path length to zero
6867 so a rescan won't be done. */
6868 for (i = path_size - 1; i >= 0; i--)
6869 if (data->path[i].status != NOT_TAKEN)
6870 break;
6872 if (i == -1)
6873 data->path_size = 0;
6874 else
6875 data->path_size = path_size;
6877 /* End the current branch path. */
6878 data->path[path_size].branch = 0;
6881 /* Perform cse on the instructions of a function.
6882 F is the first instruction.
6883 NREGS is one plus the highest pseudo-reg number used in the instruction.
6885 AFTER_LOOP is 1 if this is the cse call done after loop optimization
6886 (only if -frerun-cse-after-loop).
6888 Returns 1 if jump_optimize should be redone due to simplifications
6889 in conditional jump instructions. */
6892 cse_main (rtx f, int nregs, int after_loop, FILE *file)
6894 struct cse_basic_block_data val;
6895 rtx insn = f;
6896 int i;
6898 val.path = xmalloc (sizeof (struct branch_path)
6899 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6901 cse_jumps_altered = 0;
6902 recorded_label_ref = 0;
6903 constant_pool_entries_cost = 0;
6904 constant_pool_entries_regcost = 0;
6905 val.path_size = 0;
6906 gen_lowpart = gen_lowpart_if_possible;
6908 init_recog ();
6909 init_alias_analysis ();
6911 max_reg = nregs;
6913 max_insn_uid = get_max_uid ();
6915 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6917 #ifdef LOAD_EXTEND_OP
6919 /* Allocate scratch rtl here. cse_insn will fill in the memory reference
6920 and change the code and mode as appropriate. */
6921 memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX);
6922 #endif
6924 /* Reset the counter indicating how many elements have been made
6925 thus far. */
6926 n_elements_made = 0;
6928 /* Find the largest uid. */
6930 max_uid = get_max_uid ();
6931 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6933 /* Compute the mapping from uids to cuids.
6934 CUIDs are numbers assigned to insns, like uids,
6935 except that cuids increase monotonically through the code.
6936 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6937 between two insns is not affected by -g. */
6939 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6941 if (GET_CODE (insn) != NOTE
6942 || NOTE_LINE_NUMBER (insn) < 0)
6943 INSN_CUID (insn) = ++i;
6944 else
6945 /* Give a line number note the same cuid as preceding insn. */
6946 INSN_CUID (insn) = i;
6949 ggc_push_context ();
6951 /* Loop over basic blocks.
6952 Compute the maximum number of qty's needed for each basic block
6953 (which is 2 for each SET). */
6954 insn = f;
6955 while (insn)
6957 cse_altered = 0;
6958 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps, after_loop,
6959 flag_cse_skip_blocks);
6961 /* If this basic block was already processed or has no sets, skip it. */
6962 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6964 PUT_MODE (insn, VOIDmode);
6965 insn = (val.last ? NEXT_INSN (val.last) : 0);
6966 val.path_size = 0;
6967 continue;
6970 cse_basic_block_start = val.low_cuid;
6971 cse_basic_block_end = val.high_cuid;
6972 max_qty = val.nsets * 2;
6974 if (file)
6975 fnotice (file, ";; Processing block from %d to %d, %d sets.\n",
6976 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6977 val.nsets);
6979 /* Make MAX_QTY bigger to give us room to optimize
6980 past the end of this basic block, if that should prove useful. */
6981 if (max_qty < 500)
6982 max_qty = 500;
6984 max_qty += max_reg;
6986 /* If this basic block is being extended by following certain jumps,
6987 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6988 Otherwise, we start after this basic block. */
6989 if (val.path_size > 0)
6990 cse_basic_block (insn, val.last, val.path, 0);
6991 else
6993 int old_cse_jumps_altered = cse_jumps_altered;
6994 rtx temp;
6996 /* When cse changes a conditional jump to an unconditional
6997 jump, we want to reprocess the block, since it will give
6998 us a new branch path to investigate. */
6999 cse_jumps_altered = 0;
7000 temp = cse_basic_block (insn, val.last, val.path, ! after_loop);
7001 if (cse_jumps_altered == 0
7002 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7003 insn = temp;
7005 cse_jumps_altered |= old_cse_jumps_altered;
7008 if (cse_altered)
7009 ggc_collect ();
7011 #ifdef USE_C_ALLOCA
7012 alloca (0);
7013 #endif
7016 ggc_pop_context ();
7018 if (max_elements_made < n_elements_made)
7019 max_elements_made = n_elements_made;
7021 /* Clean up. */
7022 end_alias_analysis ();
7023 free (uid_cuid);
7024 free (reg_eqv_table);
7025 free (val.path);
7026 gen_lowpart = gen_lowpart_general;
7028 return cse_jumps_altered || recorded_label_ref;
7031 /* Process a single basic block. FROM and TO and the limits of the basic
7032 block. NEXT_BRANCH points to the branch path when following jumps or
7033 a null path when not following jumps.
7035 AROUND_LOOP is nonzero if we are to try to cse around to the start of a
7036 loop. This is true when we are being called for the last time on a
7037 block and this CSE pass is before loop.c. */
7039 static rtx
7040 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch,
7041 int around_loop)
7043 rtx insn;
7044 int to_usage = 0;
7045 rtx libcall_insn = NULL_RTX;
7046 int num_insns = 0;
7047 int no_conflict = 0;
7049 /* This array is undefined before max_reg, so only allocate
7050 the space actually needed and adjust the start. */
7052 qty_table = xmalloc ((max_qty - max_reg) * sizeof (struct qty_table_elem));
7053 qty_table -= max_reg;
7055 new_basic_block ();
7057 /* TO might be a label. If so, protect it from being deleted. */
7058 if (to != 0 && GET_CODE (to) == CODE_LABEL)
7059 ++LABEL_NUSES (to);
7061 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7063 enum rtx_code code = GET_CODE (insn);
7065 /* If we have processed 1,000 insns, flush the hash table to
7066 avoid extreme quadratic behavior. We must not include NOTEs
7067 in the count since there may be more of them when generating
7068 debugging information. If we clear the table at different
7069 times, code generated with -g -O might be different than code
7070 generated with -O but not -g.
7072 ??? This is a real kludge and needs to be done some other way.
7073 Perhaps for 2.9. */
7074 if (code != NOTE && num_insns++ > 1000)
7076 flush_hash_table ();
7077 num_insns = 0;
7080 /* See if this is a branch that is part of the path. If so, and it is
7081 to be taken, do so. */
7082 if (next_branch->branch == insn)
7084 enum taken status = next_branch++->status;
7085 if (status != NOT_TAKEN)
7087 if (status == TAKEN)
7088 record_jump_equiv (insn, 1);
7089 else
7090 invalidate_skipped_block (NEXT_INSN (insn));
7092 /* Set the last insn as the jump insn; it doesn't affect cc0.
7093 Then follow this branch. */
7094 #ifdef HAVE_cc0
7095 prev_insn_cc0 = 0;
7096 prev_insn = insn;
7097 #endif
7098 insn = JUMP_LABEL (insn);
7099 continue;
7103 if (GET_MODE (insn) == QImode)
7104 PUT_MODE (insn, VOIDmode);
7106 if (GET_RTX_CLASS (code) == RTX_INSN)
7108 rtx p;
7110 /* Process notes first so we have all notes in canonical forms when
7111 looking for duplicate operations. */
7113 if (REG_NOTES (insn))
7114 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7116 /* Track when we are inside in LIBCALL block. Inside such a block,
7117 we do not want to record destinations. The last insn of a
7118 LIBCALL block is not considered to be part of the block, since
7119 its destination is the result of the block and hence should be
7120 recorded. */
7122 if (REG_NOTES (insn) != 0)
7124 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7125 libcall_insn = XEXP (p, 0);
7126 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7128 /* Keep libcall_insn for the last SET insn of a no-conflict
7129 block to prevent changing the destination. */
7130 if (! no_conflict)
7131 libcall_insn = 0;
7132 else
7133 no_conflict = -1;
7135 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
7136 no_conflict = 1;
7139 cse_insn (insn, libcall_insn);
7141 if (no_conflict == -1)
7143 libcall_insn = 0;
7144 no_conflict = 0;
7147 /* If we haven't already found an insn where we added a LABEL_REF,
7148 check this one. */
7149 if (GET_CODE (insn) == INSN && ! recorded_label_ref
7150 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
7151 (void *) insn))
7152 recorded_label_ref = 1;
7155 /* If INSN is now an unconditional jump, skip to the end of our
7156 basic block by pretending that we just did the last insn in the
7157 basic block. If we are jumping to the end of our block, show
7158 that we can have one usage of TO. */
7160 if (any_uncondjump_p (insn))
7162 if (to == 0)
7164 free (qty_table + max_reg);
7165 return 0;
7168 if (JUMP_LABEL (insn) == to)
7169 to_usage = 1;
7171 /* Maybe TO was deleted because the jump is unconditional.
7172 If so, there is nothing left in this basic block. */
7173 /* ??? Perhaps it would be smarter to set TO
7174 to whatever follows this insn,
7175 and pretend the basic block had always ended here. */
7176 if (INSN_DELETED_P (to))
7177 break;
7179 insn = PREV_INSN (to);
7182 /* See if it is ok to keep on going past the label
7183 which used to end our basic block. Remember that we incremented
7184 the count of that label, so we decrement it here. If we made
7185 a jump unconditional, TO_USAGE will be one; in that case, we don't
7186 want to count the use in that jump. */
7188 if (to != 0 && NEXT_INSN (insn) == to
7189 && GET_CODE (to) == CODE_LABEL && --LABEL_NUSES (to) == to_usage)
7191 struct cse_basic_block_data val;
7192 rtx prev;
7194 insn = NEXT_INSN (to);
7196 /* If TO was the last insn in the function, we are done. */
7197 if (insn == 0)
7199 free (qty_table + max_reg);
7200 return 0;
7203 /* If TO was preceded by a BARRIER we are done with this block
7204 because it has no continuation. */
7205 prev = prev_nonnote_insn (to);
7206 if (prev && GET_CODE (prev) == BARRIER)
7208 free (qty_table + max_reg);
7209 return insn;
7212 /* Find the end of the following block. Note that we won't be
7213 following branches in this case. */
7214 to_usage = 0;
7215 val.path_size = 0;
7216 val.path = xmalloc (sizeof (struct branch_path)
7217 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7218 cse_end_of_basic_block (insn, &val, 0, 0, 0);
7219 free (val.path);
7221 /* If the tables we allocated have enough space left
7222 to handle all the SETs in the next basic block,
7223 continue through it. Otherwise, return,
7224 and that block will be scanned individually. */
7225 if (val.nsets * 2 + next_qty > max_qty)
7226 break;
7228 cse_basic_block_start = val.low_cuid;
7229 cse_basic_block_end = val.high_cuid;
7230 to = val.last;
7232 /* Prevent TO from being deleted if it is a label. */
7233 if (to != 0 && GET_CODE (to) == CODE_LABEL)
7234 ++LABEL_NUSES (to);
7236 /* Back up so we process the first insn in the extension. */
7237 insn = PREV_INSN (insn);
7241 if (next_qty > max_qty)
7242 abort ();
7244 /* If we are running before loop.c, we stopped on a NOTE_INSN_LOOP_END, and
7245 the previous insn is the only insn that branches to the head of a loop,
7246 we can cse into the loop. Don't do this if we changed the jump
7247 structure of a loop unless we aren't going to be following jumps. */
7249 insn = prev_nonnote_insn (to);
7250 if ((cse_jumps_altered == 0
7251 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7252 && around_loop && to != 0
7253 && GET_CODE (to) == NOTE && NOTE_LINE_NUMBER (to) == NOTE_INSN_LOOP_END
7254 && GET_CODE (insn) == JUMP_INSN
7255 && JUMP_LABEL (insn) != 0
7256 && LABEL_NUSES (JUMP_LABEL (insn)) == 1)
7257 cse_around_loop (JUMP_LABEL (insn));
7259 free (qty_table + max_reg);
7261 return to ? NEXT_INSN (to) : 0;
7264 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7265 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7267 static int
7268 check_for_label_ref (rtx *rtl, void *data)
7270 rtx insn = (rtx) data;
7272 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7273 we must rerun jump since it needs to place the note. If this is a
7274 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7275 since no REG_LABEL will be added. */
7276 return (GET_CODE (*rtl) == LABEL_REF
7277 && ! LABEL_REF_NONLOCAL_P (*rtl)
7278 && LABEL_P (XEXP (*rtl, 0))
7279 && INSN_UID (XEXP (*rtl, 0)) != 0
7280 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7283 /* Count the number of times registers are used (not set) in X.
7284 COUNTS is an array in which we accumulate the count, INCR is how much
7285 we count each register usage. */
7287 static void
7288 count_reg_usage (rtx x, int *counts, int incr)
7290 enum rtx_code code;
7291 rtx note;
7292 const char *fmt;
7293 int i, j;
7295 if (x == 0)
7296 return;
7298 switch (code = GET_CODE (x))
7300 case REG:
7301 counts[REGNO (x)] += incr;
7302 return;
7304 case PC:
7305 case CC0:
7306 case CONST:
7307 case CONST_INT:
7308 case CONST_DOUBLE:
7309 case CONST_VECTOR:
7310 case SYMBOL_REF:
7311 case LABEL_REF:
7312 return;
7314 case CLOBBER:
7315 /* If we are clobbering a MEM, mark any registers inside the address
7316 as being used. */
7317 if (GET_CODE (XEXP (x, 0)) == MEM)
7318 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, incr);
7319 return;
7321 case SET:
7322 /* Unless we are setting a REG, count everything in SET_DEST. */
7323 if (GET_CODE (SET_DEST (x)) != REG)
7324 count_reg_usage (SET_DEST (x), counts, incr);
7325 count_reg_usage (SET_SRC (x), counts, incr);
7326 return;
7328 case CALL_INSN:
7329 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, incr);
7330 /* Fall through. */
7332 case INSN:
7333 case JUMP_INSN:
7334 count_reg_usage (PATTERN (x), counts, incr);
7336 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7337 use them. */
7339 note = find_reg_equal_equiv_note (x);
7340 if (note)
7342 rtx eqv = XEXP (note, 0);
7344 if (GET_CODE (eqv) == EXPR_LIST)
7345 /* This REG_EQUAL note describes the result of a function call.
7346 Process all the arguments. */
7349 count_reg_usage (XEXP (eqv, 0), counts, incr);
7350 eqv = XEXP (eqv, 1);
7352 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7353 else
7354 count_reg_usage (eqv, counts, incr);
7356 return;
7358 case EXPR_LIST:
7359 if (REG_NOTE_KIND (x) == REG_EQUAL
7360 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7361 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7362 involving registers in the address. */
7363 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7364 count_reg_usage (XEXP (x, 0), counts, incr);
7366 count_reg_usage (XEXP (x, 1), counts, incr);
7367 return;
7369 case ASM_OPERANDS:
7370 /* Iterate over just the inputs, not the constraints as well. */
7371 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7372 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, incr);
7373 return;
7375 case INSN_LIST:
7376 abort ();
7378 default:
7379 break;
7382 fmt = GET_RTX_FORMAT (code);
7383 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7385 if (fmt[i] == 'e')
7386 count_reg_usage (XEXP (x, i), counts, incr);
7387 else if (fmt[i] == 'E')
7388 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7389 count_reg_usage (XVECEXP (x, i, j), counts, incr);
7393 /* Return true if set is live. */
7394 static bool
7395 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7396 int *counts)
7398 #ifdef HAVE_cc0
7399 rtx tem;
7400 #endif
7402 if (set_noop_p (set))
7405 #ifdef HAVE_cc0
7406 else if (GET_CODE (SET_DEST (set)) == CC0
7407 && !side_effects_p (SET_SRC (set))
7408 && ((tem = next_nonnote_insn (insn)) == 0
7409 || !INSN_P (tem)
7410 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7411 return false;
7412 #endif
7413 else if (GET_CODE (SET_DEST (set)) != REG
7414 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7415 || counts[REGNO (SET_DEST (set))] != 0
7416 || side_effects_p (SET_SRC (set))
7417 /* An ADDRESSOF expression can turn into a use of the
7418 internal arg pointer, so always consider the
7419 internal arg pointer live. If it is truly dead,
7420 flow will delete the initializing insn. */
7421 || (SET_DEST (set) == current_function_internal_arg_pointer))
7422 return true;
7423 return false;
7426 /* Return true if insn is live. */
7428 static bool
7429 insn_live_p (rtx insn, int *counts)
7431 int i;
7432 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7433 return true;
7434 else if (GET_CODE (PATTERN (insn)) == SET)
7435 return set_live_p (PATTERN (insn), insn, counts);
7436 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7438 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7440 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7442 if (GET_CODE (elt) == SET)
7444 if (set_live_p (elt, insn, counts))
7445 return true;
7447 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7448 return true;
7450 return false;
7452 else
7453 return true;
7456 /* Return true if libcall is dead as a whole. */
7458 static bool
7459 dead_libcall_p (rtx insn, int *counts)
7461 rtx note, set, new;
7463 /* See if there's a REG_EQUAL note on this insn and try to
7464 replace the source with the REG_EQUAL expression.
7466 We assume that insns with REG_RETVALs can only be reg->reg
7467 copies at this point. */
7468 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7469 if (!note)
7470 return false;
7472 set = single_set (insn);
7473 if (!set)
7474 return false;
7476 new = simplify_rtx (XEXP (note, 0));
7477 if (!new)
7478 new = XEXP (note, 0);
7480 /* While changing insn, we must update the counts accordingly. */
7481 count_reg_usage (insn, counts, -1);
7483 if (validate_change (insn, &SET_SRC (set), new, 0))
7485 count_reg_usage (insn, counts, 1);
7486 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7487 remove_note (insn, note);
7488 return true;
7491 if (CONSTANT_P (new))
7493 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7494 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7496 count_reg_usage (insn, counts, 1);
7497 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7498 remove_note (insn, note);
7499 return true;
7503 count_reg_usage (insn, counts, 1);
7504 return false;
7507 /* Scan all the insns and delete any that are dead; i.e., they store a register
7508 that is never used or they copy a register to itself.
7510 This is used to remove insns made obviously dead by cse, loop or other
7511 optimizations. It improves the heuristics in loop since it won't try to
7512 move dead invariants out of loops or make givs for dead quantities. The
7513 remaining passes of the compilation are also sped up. */
7516 delete_trivially_dead_insns (rtx insns, int nreg)
7518 int *counts;
7519 rtx insn, prev;
7520 int in_libcall = 0, dead_libcall = 0;
7521 int ndead = 0, nlastdead, niterations = 0;
7523 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7524 /* First count the number of times each register is used. */
7525 counts = xcalloc (nreg, sizeof (int));
7526 for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn))
7527 count_reg_usage (insn, counts, 1);
7531 nlastdead = ndead;
7532 niterations++;
7533 /* Go from the last insn to the first and delete insns that only set unused
7534 registers or copy a register to itself. As we delete an insn, remove
7535 usage counts for registers it uses.
7537 The first jump optimization pass may leave a real insn as the last
7538 insn in the function. We must not skip that insn or we may end
7539 up deleting code that is not really dead. */
7540 insn = get_last_insn ();
7541 if (! INSN_P (insn))
7542 insn = prev_real_insn (insn);
7544 for (; insn; insn = prev)
7546 int live_insn = 0;
7548 prev = prev_real_insn (insn);
7550 /* Don't delete any insns that are part of a libcall block unless
7551 we can delete the whole libcall block.
7553 Flow or loop might get confused if we did that. Remember
7554 that we are scanning backwards. */
7555 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7557 in_libcall = 1;
7558 live_insn = 1;
7559 dead_libcall = dead_libcall_p (insn, counts);
7561 else if (in_libcall)
7562 live_insn = ! dead_libcall;
7563 else
7564 live_insn = insn_live_p (insn, counts);
7566 /* If this is a dead insn, delete it and show registers in it aren't
7567 being used. */
7569 if (! live_insn)
7571 count_reg_usage (insn, counts, -1);
7572 delete_insn_and_edges (insn);
7573 ndead++;
7576 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7578 in_libcall = 0;
7579 dead_libcall = 0;
7583 while (ndead != nlastdead);
7585 if (dump_file && ndead)
7586 fprintf (dump_file, "Deleted %i trivially dead insns; %i iterations\n",
7587 ndead, niterations);
7588 /* Clean up. */
7589 free (counts);
7590 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7591 return ndead;
7594 /* This function is called via for_each_rtx. The argument, NEWREG, is
7595 a condition code register with the desired mode. If we are looking
7596 at the same register in a different mode, replace it with
7597 NEWREG. */
7599 static int
7600 cse_change_cc_mode (rtx *loc, void *data)
7602 rtx newreg = (rtx) data;
7604 if (*loc
7605 && GET_CODE (*loc) == REG
7606 && REGNO (*loc) == REGNO (newreg)
7607 && GET_MODE (*loc) != GET_MODE (newreg))
7609 *loc = newreg;
7610 return -1;
7612 return 0;
7615 /* Change the mode of any reference to the register REGNO (NEWREG) to
7616 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7617 any instruction which modifies NEWREG. */
7619 static void
7620 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7622 rtx insn;
7624 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7626 if (! INSN_P (insn))
7627 continue;
7629 if (reg_set_p (newreg, insn))
7630 return;
7632 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, newreg);
7633 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, newreg);
7637 /* BB is a basic block which finishes with CC_REG as a condition code
7638 register which is set to CC_SRC. Look through the successors of BB
7639 to find blocks which have a single predecessor (i.e., this one),
7640 and look through those blocks for an assignment to CC_REG which is
7641 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7642 permitted to change the mode of CC_SRC to a compatible mode. This
7643 returns VOIDmode if no equivalent assignments were found.
7644 Otherwise it returns the mode which CC_SRC should wind up with.
7646 The main complexity in this function is handling the mode issues.
7647 We may have more than one duplicate which we can eliminate, and we
7648 try to find a mode which will work for multiple duplicates. */
7650 static enum machine_mode
7651 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7653 bool found_equiv;
7654 enum machine_mode mode;
7655 unsigned int insn_count;
7656 edge e;
7657 rtx insns[2];
7658 enum machine_mode modes[2];
7659 rtx last_insns[2];
7660 unsigned int i;
7661 rtx newreg;
7663 /* We expect to have two successors. Look at both before picking
7664 the final mode for the comparison. If we have more successors
7665 (i.e., some sort of table jump, although that seems unlikely),
7666 then we require all beyond the first two to use the same
7667 mode. */
7669 found_equiv = false;
7670 mode = GET_MODE (cc_src);
7671 insn_count = 0;
7672 for (e = bb->succ; e; e = e->succ_next)
7674 rtx insn;
7675 rtx end;
7677 if (e->flags & EDGE_COMPLEX)
7678 continue;
7680 if (! e->dest->pred
7681 || e->dest->pred->pred_next
7682 || e->dest == EXIT_BLOCK_PTR)
7683 continue;
7685 end = NEXT_INSN (BB_END (e->dest));
7686 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7688 rtx set;
7690 if (! INSN_P (insn))
7691 continue;
7693 /* If CC_SRC is modified, we have to stop looking for
7694 something which uses it. */
7695 if (modified_in_p (cc_src, insn))
7696 break;
7698 /* Check whether INSN sets CC_REG to CC_SRC. */
7699 set = single_set (insn);
7700 if (set
7701 && GET_CODE (SET_DEST (set)) == REG
7702 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7704 bool found;
7705 enum machine_mode set_mode;
7706 enum machine_mode comp_mode;
7708 found = false;
7709 set_mode = GET_MODE (SET_SRC (set));
7710 comp_mode = set_mode;
7711 if (rtx_equal_p (cc_src, SET_SRC (set)))
7712 found = true;
7713 else if (GET_CODE (cc_src) == COMPARE
7714 && GET_CODE (SET_SRC (set)) == COMPARE
7715 && mode != set_mode
7716 && rtx_equal_p (XEXP (cc_src, 0),
7717 XEXP (SET_SRC (set), 0))
7718 && rtx_equal_p (XEXP (cc_src, 1),
7719 XEXP (SET_SRC (set), 1)))
7722 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7723 if (comp_mode != VOIDmode
7724 && (can_change_mode || comp_mode == mode))
7725 found = true;
7728 if (found)
7730 found_equiv = true;
7731 if (insn_count < ARRAY_SIZE (insns))
7733 insns[insn_count] = insn;
7734 modes[insn_count] = set_mode;
7735 last_insns[insn_count] = end;
7736 ++insn_count;
7738 if (mode != comp_mode)
7740 if (! can_change_mode)
7741 abort ();
7742 mode = comp_mode;
7743 PUT_MODE (cc_src, mode);
7746 else
7748 if (set_mode != mode)
7750 /* We found a matching expression in the
7751 wrong mode, but we don't have room to
7752 store it in the array. Punt. This case
7753 should be rare. */
7754 break;
7756 /* INSN sets CC_REG to a value equal to CC_SRC
7757 with the right mode. We can simply delete
7758 it. */
7759 delete_insn (insn);
7762 /* We found an instruction to delete. Keep looking,
7763 in the hopes of finding a three-way jump. */
7764 continue;
7767 /* We found an instruction which sets the condition
7768 code, so don't look any farther. */
7769 break;
7772 /* If INSN sets CC_REG in some other way, don't look any
7773 farther. */
7774 if (reg_set_p (cc_reg, insn))
7775 break;
7778 /* If we fell off the bottom of the block, we can keep looking
7779 through successors. We pass CAN_CHANGE_MODE as false because
7780 we aren't prepared to handle compatibility between the
7781 further blocks and this block. */
7782 if (insn == end)
7784 enum machine_mode submode;
7786 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7787 if (submode != VOIDmode)
7789 if (submode != mode)
7790 abort ();
7791 found_equiv = true;
7792 can_change_mode = false;
7797 if (! found_equiv)
7798 return VOIDmode;
7800 /* Now INSN_COUNT is the number of instructions we found which set
7801 CC_REG to a value equivalent to CC_SRC. The instructions are in
7802 INSNS. The modes used by those instructions are in MODES. */
7804 newreg = NULL_RTX;
7805 for (i = 0; i < insn_count; ++i)
7807 if (modes[i] != mode)
7809 /* We need to change the mode of CC_REG in INSNS[i] and
7810 subsequent instructions. */
7811 if (! newreg)
7813 if (GET_MODE (cc_reg) == mode)
7814 newreg = cc_reg;
7815 else
7816 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7818 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7819 newreg);
7822 delete_insn (insns[i]);
7825 return mode;
7828 /* If we have a fixed condition code register (or two), walk through
7829 the instructions and try to eliminate duplicate assignments. */
7831 void
7832 cse_condition_code_reg (void)
7834 unsigned int cc_regno_1;
7835 unsigned int cc_regno_2;
7836 rtx cc_reg_1;
7837 rtx cc_reg_2;
7838 basic_block bb;
7840 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7841 return;
7843 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7844 if (cc_regno_2 != INVALID_REGNUM)
7845 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7846 else
7847 cc_reg_2 = NULL_RTX;
7849 FOR_EACH_BB (bb)
7851 rtx last_insn;
7852 rtx cc_reg;
7853 rtx insn;
7854 rtx cc_src_insn;
7855 rtx cc_src;
7856 enum machine_mode mode;
7857 enum machine_mode orig_mode;
7859 /* Look for blocks which end with a conditional jump based on a
7860 condition code register. Then look for the instruction which
7861 sets the condition code register. Then look through the
7862 successor blocks for instructions which set the condition
7863 code register to the same value. There are other possible
7864 uses of the condition code register, but these are by far the
7865 most common and the ones which we are most likely to be able
7866 to optimize. */
7868 last_insn = BB_END (bb);
7869 if (GET_CODE (last_insn) != JUMP_INSN)
7870 continue;
7872 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7873 cc_reg = cc_reg_1;
7874 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7875 cc_reg = cc_reg_2;
7876 else
7877 continue;
7879 cc_src_insn = NULL_RTX;
7880 cc_src = NULL_RTX;
7881 for (insn = PREV_INSN (last_insn);
7882 insn && insn != PREV_INSN (BB_HEAD (bb));
7883 insn = PREV_INSN (insn))
7885 rtx set;
7887 if (! INSN_P (insn))
7888 continue;
7889 set = single_set (insn);
7890 if (set
7891 && GET_CODE (SET_DEST (set)) == REG
7892 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7894 cc_src_insn = insn;
7895 cc_src = SET_SRC (set);
7896 break;
7898 else if (reg_set_p (cc_reg, insn))
7899 break;
7902 if (! cc_src_insn)
7903 continue;
7905 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7906 continue;
7908 /* Now CC_REG is a condition code register used for a
7909 conditional jump at the end of the block, and CC_SRC, in
7910 CC_SRC_INSN, is the value to which that condition code
7911 register is set, and CC_SRC is still meaningful at the end of
7912 the basic block. */
7914 orig_mode = GET_MODE (cc_src);
7915 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7916 if (mode != VOIDmode)
7918 if (mode != GET_MODE (cc_src))
7919 abort ();
7920 if (mode != orig_mode)
7922 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7924 /* Change the mode of CC_REG in CC_SRC_INSN to
7925 GET_MODE (NEWREG). */
7926 for_each_rtx (&PATTERN (cc_src_insn), cse_change_cc_mode,
7927 newreg);
7928 for_each_rtx (&REG_NOTES (cc_src_insn), cse_change_cc_mode,
7929 newreg);
7931 /* Do the same in the following insns that use the
7932 current value of CC_REG within BB. */
7933 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7934 NEXT_INSN (last_insn),
7935 newreg);