* config/sparc/sparc.c (load_pic_register): Emit the appropriate
[official-gcc.git] / gcc / cse.c
blob9e3e4965a388eb5531cf4bd542f0ae9fd551c153
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, 2005 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "rtl.h"
28 #include "tm_p.h"
29 #include "hard-reg-set.h"
30 #include "regs.h"
31 #include "basic-block.h"
32 #include "flags.h"
33 #include "real.h"
34 #include "insn-config.h"
35 #include "recog.h"
36 #include "function.h"
37 #include "expr.h"
38 #include "toplev.h"
39 #include "output.h"
40 #include "ggc.h"
41 #include "timevar.h"
42 #include "except.h"
43 #include "target.h"
44 #include "params.h"
45 #include "rtlhooks-def.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 zero.
87 If register N has not been assigned a quantity, reg_qty[N] will
88 equal -N - 1, which is always negative.
90 Quantity numbers below zero do not exist and none of the `qty_table'
91 entries should be referenced with a negative index.
93 We also maintain a bidirectional chain of registers for each
94 quantity number. The `qty_table` members `first_reg' and `last_reg',
95 and `reg_eqv_table' members `next' and `prev' hold these chains.
97 The first register in a chain is the one whose lifespan is least local.
98 Among equals, it is the one that was seen first.
99 We replace any equivalent register with that one.
101 If two registers have the same quantity number, it must be true that
102 REG expressions with qty_table `mode' must be in the hash table for both
103 registers and must be in the same class.
105 The converse is not true. Since hard registers may be referenced in
106 any mode, two REG expressions might be equivalent in the hash table
107 but not have the same quantity number if the quantity number of one
108 of the registers is not the same mode as those expressions.
110 Constants and quantity numbers
112 When a quantity has a known constant value, that value is stored
113 in the appropriate qty_table `const_rtx'. This is in addition to
114 putting the constant in the hash table as is usual for non-regs.
116 Whether a reg or a constant is preferred is determined by the configuration
117 macro CONST_COSTS and will often depend on the constant value. In any
118 event, expressions containing constants can be simplified, by fold_rtx.
120 When a quantity has a known nearly constant value (such as an address
121 of a stack slot), that value is stored in the appropriate qty_table
122 `const_rtx'.
124 Integer constants don't have a machine mode. However, cse
125 determines the intended machine mode from the destination
126 of the instruction that moves the constant. The machine mode
127 is recorded in the hash table along with the actual RTL
128 constant expression so that different modes are kept separate.
130 Other expressions:
132 To record known equivalences among expressions in general
133 we use a hash table called `table'. It has a fixed number of buckets
134 that contain chains of `struct table_elt' elements for expressions.
135 These chains connect the elements whose expressions have the same
136 hash codes.
138 Other chains through the same elements connect the elements which
139 currently have equivalent values.
141 Register references in an expression are canonicalized before hashing
142 the expression. This is done using `reg_qty' and qty_table `first_reg'.
143 The hash code of a register reference is computed using the quantity
144 number, not the register number.
146 When the value of an expression changes, it is necessary to remove from the
147 hash table not just that expression but all expressions whose values
148 could be different as a result.
150 1. If the value changing is in memory, except in special cases
151 ANYTHING referring to memory could be changed. That is because
152 nobody knows where a pointer does not point.
153 The function `invalidate_memory' removes what is necessary.
155 The special cases are when the address is constant or is
156 a constant plus a fixed register such as the frame pointer
157 or a static chain pointer. When such addresses are stored in,
158 we can tell exactly which other such addresses must be invalidated
159 due to overlap. `invalidate' does this.
160 All expressions that refer to non-constant
161 memory addresses are also invalidated. `invalidate_memory' does this.
163 2. If the value changing is a register, all expressions
164 containing references to that register, and only those,
165 must be removed.
167 Because searching the entire hash table for expressions that contain
168 a register is very slow, we try to figure out when it isn't necessary.
169 Precisely, this is necessary only when expressions have been
170 entered in the hash table using this register, and then the value has
171 changed, and then another expression wants to be added to refer to
172 the register's new value. This sequence of circumstances is rare
173 within any one basic block.
175 The vectors `reg_tick' and `reg_in_table' are used to detect this case.
176 reg_tick[i] is incremented whenever a value is stored in register i.
177 reg_in_table[i] holds -1 if no references to register i have been
178 entered in the table; otherwise, it contains the value reg_tick[i] had
179 when the references were entered. If we want to enter a reference
180 and reg_in_table[i] != reg_tick[i], we must scan and remove old references.
181 Until we want to enter a new entry, the mere fact that the two vectors
182 don't match makes the entries be ignored if anyone tries to match them.
184 Registers themselves are entered in the hash table as well as in
185 the equivalent-register chains. However, the vectors `reg_tick'
186 and `reg_in_table' do not apply to expressions which are simple
187 register references. These expressions are removed from the table
188 immediately when they become invalid, and this can be done even if
189 we do not immediately search for all the expressions that refer to
190 the register.
192 A CLOBBER rtx in an instruction invalidates its operand for further
193 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
194 invalidates everything that resides in memory.
196 Related expressions:
198 Constant expressions that differ only by an additive integer
199 are called related. When a constant expression is put in
200 the table, the related expression with no constant term
201 is also entered. These are made to point at each other
202 so that it is possible to find out if there exists any
203 register equivalent to an expression related to a given expression. */
205 /* One plus largest register number used in this function. */
207 static int max_reg;
209 /* One plus largest instruction UID used in this function at time of
210 cse_main call. */
212 static int max_insn_uid;
214 /* Length of qty_table vector. We know in advance we will not need
215 a quantity number this big. */
217 static int max_qty;
219 /* Next quantity number to be allocated.
220 This is 1 + the largest number needed so far. */
222 static int next_qty;
224 /* Per-qty information tracking.
226 `first_reg' and `last_reg' track the head and tail of the
227 chain of registers which currently contain this quantity.
229 `mode' contains the machine mode of this quantity.
231 `const_rtx' holds the rtx of the constant value of this
232 quantity, if known. A summations of the frame/arg pointer
233 and a constant can also be entered here. When this holds
234 a known value, `const_insn' is the insn which stored the
235 constant value.
237 `comparison_{code,const,qty}' are used to track when a
238 comparison between a quantity and some constant or register has
239 been passed. In such a case, we know the results of the comparison
240 in case we see it again. These members record a comparison that
241 is known to be true. `comparison_code' holds the rtx code of such
242 a comparison, else it is set to UNKNOWN and the other two
243 comparison members are undefined. `comparison_const' holds
244 the constant being compared against, or zero if the comparison
245 is not against a constant. `comparison_qty' holds the quantity
246 being compared against when the result is known. If the comparison
247 is not with a register, `comparison_qty' is -1. */
249 struct qty_table_elem
251 rtx const_rtx;
252 rtx const_insn;
253 rtx comparison_const;
254 int comparison_qty;
255 unsigned int first_reg, last_reg;
256 /* The sizes of these fields should match the sizes of the
257 code and mode fields of struct rtx_def (see rtl.h). */
258 ENUM_BITFIELD(rtx_code) comparison_code : 16;
259 ENUM_BITFIELD(machine_mode) mode : 8;
262 /* The table of all qtys, indexed by qty number. */
263 static struct qty_table_elem *qty_table;
265 /* Structure used to pass arguments via for_each_rtx to function
266 cse_change_cc_mode. */
267 struct change_cc_mode_args
269 rtx insn;
270 rtx newreg;
273 #ifdef HAVE_cc0
274 /* For machines that have a CC0, we do not record its value in the hash
275 table since its use is guaranteed to be the insn immediately following
276 its definition and any other insn is presumed to invalidate it.
278 Instead, we store below the value last assigned to CC0. If it should
279 happen to be a constant, it is stored in preference to the actual
280 assigned value. In case it is a constant, we store the mode in which
281 the constant should be interpreted. */
283 static rtx prev_insn_cc0;
284 static enum machine_mode prev_insn_cc0_mode;
286 /* Previous actual insn. 0 if at first insn of basic block. */
288 static rtx prev_insn;
289 #endif
291 /* Insn being scanned. */
293 static rtx this_insn;
295 /* Index by register number, gives the number of the next (or
296 previous) register in the chain of registers sharing the same
297 value.
299 Or -1 if this register is at the end of the chain.
301 If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */
303 /* Per-register equivalence chain. */
304 struct reg_eqv_elem
306 int next, prev;
309 /* The table of all register equivalence chains. */
310 static struct reg_eqv_elem *reg_eqv_table;
312 struct cse_reg_info
314 /* Next in hash chain. */
315 struct cse_reg_info *hash_next;
317 /* The next cse_reg_info structure in the free or used list. */
318 struct cse_reg_info *next;
320 /* Search key */
321 unsigned int regno;
323 /* The quantity number of the register's current contents. */
324 int reg_qty;
326 /* The number of times the register has been altered in the current
327 basic block. */
328 int reg_tick;
330 /* The REG_TICK value at which rtx's containing this register are
331 valid in the hash table. If this does not equal the current
332 reg_tick value, such expressions existing in the hash table are
333 invalid. */
334 int reg_in_table;
336 /* The SUBREG that was set when REG_TICK was last incremented. Set
337 to -1 if the last store was to the whole register, not a subreg. */
338 unsigned int subreg_ticked;
341 /* A free list of cse_reg_info entries. */
342 static struct cse_reg_info *cse_reg_info_free_list;
344 /* A used list of cse_reg_info entries. */
345 static struct cse_reg_info *cse_reg_info_used_list;
346 static struct cse_reg_info *cse_reg_info_used_list_end;
348 /* A mapping from registers to cse_reg_info data structures. */
349 #define REGHASH_SHIFT 7
350 #define REGHASH_SIZE (1 << REGHASH_SHIFT)
351 #define REGHASH_MASK (REGHASH_SIZE - 1)
352 static struct cse_reg_info *reg_hash[REGHASH_SIZE];
354 #define REGHASH_FN(REGNO) \
355 (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK)
357 /* The last lookup we did into the cse_reg_info_tree. This allows us
358 to cache repeated lookups. */
359 static unsigned int cached_regno;
360 static struct cse_reg_info *cached_cse_reg_info;
362 /* A HARD_REG_SET containing all the hard registers for which there is
363 currently a REG expression in the hash table. Note the difference
364 from the above variables, which indicate if the REG is mentioned in some
365 expression in the table. */
367 static HARD_REG_SET hard_regs_in_table;
369 /* CUID of insn that starts the basic block currently being cse-processed. */
371 static int cse_basic_block_start;
373 /* CUID of insn that ends the basic block currently being cse-processed. */
375 static int cse_basic_block_end;
377 /* Vector mapping INSN_UIDs to cuids.
378 The cuids are like uids but increase monotonically always.
379 We use them to see whether a reg is used outside a given basic block. */
381 static int *uid_cuid;
383 /* Highest UID in UID_CUID. */
384 static int max_uid;
386 /* Get the cuid of an insn. */
388 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
390 /* Nonzero if this pass has made changes, and therefore it's
391 worthwhile to run the garbage collector. */
393 static int cse_altered;
395 /* Nonzero if cse has altered conditional jump insns
396 in such a way that jump optimization should be redone. */
398 static int cse_jumps_altered;
400 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
401 REG_LABEL, we have to rerun jump after CSE to put in the note. */
402 static int recorded_label_ref;
404 /* canon_hash stores 1 in do_not_record
405 if it notices a reference to CC0, PC, or some other volatile
406 subexpression. */
408 static int do_not_record;
410 /* canon_hash stores 1 in hash_arg_in_memory
411 if it notices a reference to memory within the expression being hashed. */
413 static int hash_arg_in_memory;
415 /* The hash table contains buckets which are chains of `struct table_elt's,
416 each recording one expression's information.
417 That expression is in the `exp' field.
419 The canon_exp field contains a canonical (from the point of view of
420 alias analysis) version of the `exp' field.
422 Those elements with the same hash code are chained in both directions
423 through the `next_same_hash' and `prev_same_hash' fields.
425 Each set of expressions with equivalent values
426 are on a two-way chain through the `next_same_value'
427 and `prev_same_value' fields, and all point with
428 the `first_same_value' field at the first element in
429 that chain. The chain is in order of increasing cost.
430 Each element's cost value is in its `cost' field.
432 The `in_memory' field is nonzero for elements that
433 involve any reference to memory. These elements are removed
434 whenever a write is done to an unidentified location in memory.
435 To be safe, we assume that a memory address is unidentified unless
436 the address is either a symbol constant or a constant plus
437 the frame pointer or argument pointer.
439 The `related_value' field is used to connect related expressions
440 (that differ by adding an integer).
441 The related expressions are chained in a circular fashion.
442 `related_value' is zero for expressions for which this
443 chain is not useful.
445 The `cost' field stores the cost of this element's expression.
446 The `regcost' field stores the value returned by approx_reg_cost for
447 this element's expression.
449 The `is_const' flag is set if the element is a constant (including
450 a fixed address).
452 The `flag' field is used as a temporary during some search routines.
454 The `mode' field is usually the same as GET_MODE (`exp'), but
455 if `exp' is a CONST_INT and has no machine mode then the `mode'
456 field is the mode it was being used as. Each constant is
457 recorded separately for each mode it is used with. */
459 struct table_elt
461 rtx exp;
462 rtx canon_exp;
463 struct table_elt *next_same_hash;
464 struct table_elt *prev_same_hash;
465 struct table_elt *next_same_value;
466 struct table_elt *prev_same_value;
467 struct table_elt *first_same_value;
468 struct table_elt *related_value;
469 int cost;
470 int regcost;
471 /* The size of this field should match the size
472 of the mode field of struct rtx_def (see rtl.h). */
473 ENUM_BITFIELD(machine_mode) mode : 8;
474 char in_memory;
475 char is_const;
476 char flag;
479 /* We don't want a lot of buckets, because we rarely have very many
480 things stored in the hash table, and a lot of buckets slows
481 down a lot of loops that happen frequently. */
482 #define HASH_SHIFT 5
483 #define HASH_SIZE (1 << HASH_SHIFT)
484 #define HASH_MASK (HASH_SIZE - 1)
486 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
487 register (hard registers may require `do_not_record' to be set). */
489 #define HASH(X, M) \
490 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
491 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
492 : canon_hash (X, M)) & HASH_MASK)
494 /* Like HASH, but without side-effects. */
495 #define SAFE_HASH(X, M) \
496 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
497 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
498 : safe_hash (X, M)) & HASH_MASK)
500 /* Determine whether register number N is considered a fixed register for the
501 purpose of approximating register costs.
502 It is desirable to replace other regs with fixed regs, to reduce need for
503 non-fixed hard regs.
504 A reg wins if it is either the frame pointer or designated as fixed. */
505 #define FIXED_REGNO_P(N) \
506 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
507 || fixed_regs[N] || global_regs[N])
509 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
510 hard registers and pointers into the frame are the cheapest with a cost
511 of 0. Next come pseudos with a cost of one and other hard registers with
512 a cost of 2. Aside from these special cases, call `rtx_cost'. */
514 #define CHEAP_REGNO(N) \
515 (REGNO_PTR_FRAME_P(N) \
516 || (HARD_REGISTER_NUM_P (N) \
517 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
519 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
520 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
522 /* Get the info associated with register N. */
524 #define GET_CSE_REG_INFO(N) \
525 (((N) == cached_regno && cached_cse_reg_info) \
526 ? cached_cse_reg_info : get_cse_reg_info ((N)))
528 /* Get the number of times this register has been updated in this
529 basic block. */
531 #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->reg_tick)
533 /* Get the point at which REG was recorded in the table. */
535 #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table)
537 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
538 SUBREG). */
540 #define SUBREG_TICKED(N) ((GET_CSE_REG_INFO (N))->subreg_ticked)
542 /* Get the quantity number for REG. */
544 #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty)
546 /* Determine if the quantity number for register X represents a valid index
547 into the qty_table. */
549 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
551 static struct table_elt *table[HASH_SIZE];
553 /* Chain of `struct table_elt's made so far for this function
554 but currently removed from the table. */
556 static struct table_elt *free_element_chain;
558 /* Number of `struct table_elt' structures made so far for this function. */
560 static int n_elements_made;
562 /* Maximum value `n_elements_made' has had so far in this compilation
563 for functions previously processed. */
565 static int max_elements_made;
567 /* Set to the cost of a constant pool reference if one was found for a
568 symbolic constant. If this was found, it means we should try to
569 convert constants into constant pool entries if they don't fit in
570 the insn. */
572 static int constant_pool_entries_cost;
573 static int constant_pool_entries_regcost;
575 /* This data describes a block that will be processed by cse_basic_block. */
577 struct cse_basic_block_data
579 /* Lowest CUID value of insns in block. */
580 int low_cuid;
581 /* Highest CUID value of insns in block. */
582 int high_cuid;
583 /* Total number of SETs in block. */
584 int nsets;
585 /* Last insn in the block. */
586 rtx last;
587 /* Size of current branch path, if any. */
588 int path_size;
589 /* Current branch path, indicating which branches will be taken. */
590 struct branch_path
592 /* The branch insn. */
593 rtx branch;
594 /* Whether it should be taken or not. AROUND is the same as taken
595 except that it is used when the destination label is not preceded
596 by a BARRIER. */
597 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
598 } *path;
601 static bool fixed_base_plus_p (rtx x);
602 static int notreg_cost (rtx, enum rtx_code);
603 static int approx_reg_cost_1 (rtx *, void *);
604 static int approx_reg_cost (rtx);
605 static int preferable (int, int, int, int);
606 static void new_basic_block (void);
607 static void make_new_qty (unsigned int, enum machine_mode);
608 static void make_regs_eqv (unsigned int, unsigned int);
609 static void delete_reg_equiv (unsigned int);
610 static int mention_regs (rtx);
611 static int insert_regs (rtx, struct table_elt *, int);
612 static void remove_from_table (struct table_elt *, unsigned);
613 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
614 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
615 static rtx lookup_as_function (rtx, enum rtx_code);
616 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
617 enum machine_mode);
618 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
619 static void invalidate (rtx, enum machine_mode);
620 static int cse_rtx_varies_p (rtx, int);
621 static void remove_invalid_refs (unsigned int);
622 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
623 enum machine_mode);
624 static void rehash_using_reg (rtx);
625 static void invalidate_memory (void);
626 static void invalidate_for_call (void);
627 static rtx use_related_value (rtx, struct table_elt *);
629 static inline unsigned canon_hash (rtx, enum machine_mode);
630 static inline unsigned safe_hash (rtx, enum machine_mode);
631 static unsigned hash_rtx_string (const char *);
633 static rtx canon_reg (rtx, rtx);
634 static void find_best_addr (rtx, rtx *, enum machine_mode);
635 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
636 enum machine_mode *,
637 enum machine_mode *);
638 static rtx fold_rtx (rtx, rtx);
639 static rtx equiv_constant (rtx);
640 static void record_jump_equiv (rtx, int);
641 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
642 int);
643 static void cse_insn (rtx, rtx);
644 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
645 int, int);
646 static int addr_affects_sp_p (rtx);
647 static void invalidate_from_clobbers (rtx);
648 static rtx cse_process_notes (rtx, rtx);
649 static void invalidate_skipped_set (rtx, rtx, void *);
650 static void invalidate_skipped_block (rtx);
651 static rtx cse_basic_block (rtx, rtx, struct branch_path *);
652 static void count_reg_usage (rtx, int *, int);
653 static int check_for_label_ref (rtx *, void *);
654 extern void dump_class (struct table_elt*);
655 static struct cse_reg_info * get_cse_reg_info (unsigned int);
656 static int check_dependence (rtx *, void *);
658 static void flush_hash_table (void);
659 static bool insn_live_p (rtx, int *);
660 static bool set_live_p (rtx, rtx, int *);
661 static bool dead_libcall_p (rtx, int *);
662 static int cse_change_cc_mode (rtx *, void *);
663 static void cse_change_cc_mode_insn (rtx, rtx);
664 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
665 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
668 #undef RTL_HOOKS_GEN_LOWPART
669 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
671 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
673 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
674 virtual regs here because the simplify_*_operation routines are called
675 by integrate.c, which is called before virtual register instantiation. */
677 static bool
678 fixed_base_plus_p (rtx x)
680 switch (GET_CODE (x))
682 case REG:
683 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
684 return true;
685 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
686 return true;
687 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
688 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
689 return true;
690 return false;
692 case PLUS:
693 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
694 return false;
695 return fixed_base_plus_p (XEXP (x, 0));
697 default:
698 return false;
702 /* Dump the expressions in the equivalence class indicated by CLASSP.
703 This function is used only for debugging. */
704 void
705 dump_class (struct table_elt *classp)
707 struct table_elt *elt;
709 fprintf (stderr, "Equivalence chain for ");
710 print_rtl (stderr, classp->exp);
711 fprintf (stderr, ": \n");
713 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
715 print_rtl (stderr, elt->exp);
716 fprintf (stderr, "\n");
720 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
722 static int
723 approx_reg_cost_1 (rtx *xp, void *data)
725 rtx x = *xp;
726 int *cost_p = data;
728 if (x && REG_P (x))
730 unsigned int regno = REGNO (x);
732 if (! CHEAP_REGNO (regno))
734 if (regno < FIRST_PSEUDO_REGISTER)
736 if (SMALL_REGISTER_CLASSES)
737 return 1;
738 *cost_p += 2;
740 else
741 *cost_p += 1;
745 return 0;
748 /* Return an estimate of the cost of the registers used in an rtx.
749 This is mostly the number of different REG expressions in the rtx;
750 however for some exceptions like fixed registers we use a cost of
751 0. If any other hard register reference occurs, return MAX_COST. */
753 static int
754 approx_reg_cost (rtx x)
756 int cost = 0;
758 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
759 return MAX_COST;
761 return cost;
764 /* Returns a canonical version of X for the address, from the point of view,
765 that all multiplications are represented as MULT instead of the multiply
766 by a power of 2 being represented as ASHIFT. */
768 static rtx
769 canon_for_address (rtx x)
771 enum rtx_code code;
772 enum machine_mode mode;
773 rtx new = 0;
774 int i;
775 const char *fmt;
777 if (!x)
778 return x;
780 code = GET_CODE (x);
781 mode = GET_MODE (x);
783 switch (code)
785 case ASHIFT:
786 if (GET_CODE (XEXP (x, 1)) == CONST_INT
787 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)
788 && INTVAL (XEXP (x, 1)) >= 0)
790 new = canon_for_address (XEXP (x, 0));
791 new = gen_rtx_MULT (mode, new,
792 gen_int_mode ((HOST_WIDE_INT) 1
793 << INTVAL (XEXP (x, 1)),
794 mode));
796 break;
797 default:
798 break;
801 if (new)
802 return new;
804 /* Now recursively process each operand of this operation. */
805 fmt = GET_RTX_FORMAT (code);
806 for (i = 0; i < GET_RTX_LENGTH (code); i++)
807 if (fmt[i] == 'e')
809 new = canon_for_address (XEXP (x, i));
810 XEXP (x, i) = new;
812 return x;
815 /* Return a negative value if an rtx A, whose costs are given by COST_A
816 and REGCOST_A, is more desirable than an rtx B.
817 Return a positive value if A is less desirable, or 0 if the two are
818 equally good. */
819 static int
820 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
822 /* First, get rid of cases involving expressions that are entirely
823 unwanted. */
824 if (cost_a != cost_b)
826 if (cost_a == MAX_COST)
827 return 1;
828 if (cost_b == MAX_COST)
829 return -1;
832 /* Avoid extending lifetimes of hardregs. */
833 if (regcost_a != regcost_b)
835 if (regcost_a == MAX_COST)
836 return 1;
837 if (regcost_b == MAX_COST)
838 return -1;
841 /* Normal operation costs take precedence. */
842 if (cost_a != cost_b)
843 return cost_a - cost_b;
844 /* Only if these are identical consider effects on register pressure. */
845 if (regcost_a != regcost_b)
846 return regcost_a - regcost_b;
847 return 0;
850 /* Internal function, to compute cost when X is not a register; called
851 from COST macro to keep it simple. */
853 static int
854 notreg_cost (rtx x, enum rtx_code outer)
856 return ((GET_CODE (x) == SUBREG
857 && REG_P (SUBREG_REG (x))
858 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
859 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
860 && (GET_MODE_SIZE (GET_MODE (x))
861 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
862 && subreg_lowpart_p (x)
863 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
864 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
866 : rtx_cost (x, outer) * 2);
870 static struct cse_reg_info *
871 get_cse_reg_info (unsigned int regno)
873 struct cse_reg_info **hash_head = &reg_hash[REGHASH_FN (regno)];
874 struct cse_reg_info *p;
876 for (p = *hash_head; p != NULL; p = p->hash_next)
877 if (p->regno == regno)
878 break;
880 if (p == NULL)
882 /* Get a new cse_reg_info structure. */
883 if (cse_reg_info_free_list)
885 p = cse_reg_info_free_list;
886 cse_reg_info_free_list = p->next;
888 else
889 p = xmalloc (sizeof (struct cse_reg_info));
891 /* Insert into hash table. */
892 p->hash_next = *hash_head;
893 *hash_head = p;
895 /* Initialize it. */
896 p->reg_tick = 1;
897 p->reg_in_table = -1;
898 p->subreg_ticked = -1;
899 p->reg_qty = -regno - 1;
900 p->regno = regno;
901 p->next = cse_reg_info_used_list;
902 cse_reg_info_used_list = p;
903 if (!cse_reg_info_used_list_end)
904 cse_reg_info_used_list_end = p;
907 /* Cache this lookup; we tend to be looking up information about the
908 same register several times in a row. */
909 cached_regno = regno;
910 cached_cse_reg_info = p;
912 return p;
915 /* Clear the hash table and initialize each register with its own quantity,
916 for a new basic block. */
918 static void
919 new_basic_block (void)
921 int i;
923 next_qty = 0;
925 /* Clear out hash table state for this pass. */
927 memset (reg_hash, 0, sizeof reg_hash);
929 if (cse_reg_info_used_list)
931 cse_reg_info_used_list_end->next = cse_reg_info_free_list;
932 cse_reg_info_free_list = cse_reg_info_used_list;
933 cse_reg_info_used_list = cse_reg_info_used_list_end = 0;
935 cached_cse_reg_info = 0;
937 CLEAR_HARD_REG_SET (hard_regs_in_table);
939 /* The per-quantity values used to be initialized here, but it is
940 much faster to initialize each as it is made in `make_new_qty'. */
942 for (i = 0; i < HASH_SIZE; i++)
944 struct table_elt *first;
946 first = table[i];
947 if (first != NULL)
949 struct table_elt *last = first;
951 table[i] = NULL;
953 while (last->next_same_hash != NULL)
954 last = last->next_same_hash;
956 /* Now relink this hash entire chain into
957 the free element list. */
959 last->next_same_hash = free_element_chain;
960 free_element_chain = first;
964 #ifdef HAVE_cc0
965 prev_insn = 0;
966 prev_insn_cc0 = 0;
967 #endif
970 /* Say that register REG contains a quantity in mode MODE not in any
971 register before and initialize that quantity. */
973 static void
974 make_new_qty (unsigned int reg, enum machine_mode mode)
976 int q;
977 struct qty_table_elem *ent;
978 struct reg_eqv_elem *eqv;
980 gcc_assert (next_qty < max_qty);
982 q = REG_QTY (reg) = next_qty++;
983 ent = &qty_table[q];
984 ent->first_reg = reg;
985 ent->last_reg = reg;
986 ent->mode = mode;
987 ent->const_rtx = ent->const_insn = NULL_RTX;
988 ent->comparison_code = UNKNOWN;
990 eqv = &reg_eqv_table[reg];
991 eqv->next = eqv->prev = -1;
994 /* Make reg NEW equivalent to reg OLD.
995 OLD is not changing; NEW is. */
997 static void
998 make_regs_eqv (unsigned int new, unsigned int old)
1000 unsigned int lastr, firstr;
1001 int q = REG_QTY (old);
1002 struct qty_table_elem *ent;
1004 ent = &qty_table[q];
1006 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1007 gcc_assert (REGNO_QTY_VALID_P (old));
1009 REG_QTY (new) = q;
1010 firstr = ent->first_reg;
1011 lastr = ent->last_reg;
1013 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1014 hard regs. Among pseudos, if NEW will live longer than any other reg
1015 of the same qty, and that is beyond the current basic block,
1016 make it the new canonical replacement for this qty. */
1017 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1018 /* Certain fixed registers might be of the class NO_REGS. This means
1019 that not only can they not be allocated by the compiler, but
1020 they cannot be used in substitutions or canonicalizations
1021 either. */
1022 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1023 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1024 || (new >= FIRST_PSEUDO_REGISTER
1025 && (firstr < FIRST_PSEUDO_REGISTER
1026 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1027 || (uid_cuid[REGNO_FIRST_UID (new)]
1028 < cse_basic_block_start))
1029 && (uid_cuid[REGNO_LAST_UID (new)]
1030 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1032 reg_eqv_table[firstr].prev = new;
1033 reg_eqv_table[new].next = firstr;
1034 reg_eqv_table[new].prev = -1;
1035 ent->first_reg = new;
1037 else
1039 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1040 Otherwise, insert before any non-fixed hard regs that are at the
1041 end. Registers of class NO_REGS cannot be used as an
1042 equivalent for anything. */
1043 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1044 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1045 && new >= FIRST_PSEUDO_REGISTER)
1046 lastr = reg_eqv_table[lastr].prev;
1047 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1048 if (reg_eqv_table[lastr].next >= 0)
1049 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1050 else
1051 qty_table[q].last_reg = new;
1052 reg_eqv_table[lastr].next = new;
1053 reg_eqv_table[new].prev = lastr;
1057 /* Remove REG from its equivalence class. */
1059 static void
1060 delete_reg_equiv (unsigned int reg)
1062 struct qty_table_elem *ent;
1063 int q = REG_QTY (reg);
1064 int p, n;
1066 /* If invalid, do nothing. */
1067 if (! REGNO_QTY_VALID_P (reg))
1068 return;
1070 ent = &qty_table[q];
1072 p = reg_eqv_table[reg].prev;
1073 n = reg_eqv_table[reg].next;
1075 if (n != -1)
1076 reg_eqv_table[n].prev = p;
1077 else
1078 ent->last_reg = p;
1079 if (p != -1)
1080 reg_eqv_table[p].next = n;
1081 else
1082 ent->first_reg = n;
1084 REG_QTY (reg) = -reg - 1;
1087 /* Remove any invalid expressions from the hash table
1088 that refer to any of the registers contained in expression X.
1090 Make sure that newly inserted references to those registers
1091 as subexpressions will be considered valid.
1093 mention_regs is not called when a register itself
1094 is being stored in the table.
1096 Return 1 if we have done something that may have changed the hash code
1097 of X. */
1099 static int
1100 mention_regs (rtx x)
1102 enum rtx_code code;
1103 int i, j;
1104 const char *fmt;
1105 int changed = 0;
1107 if (x == 0)
1108 return 0;
1110 code = GET_CODE (x);
1111 if (code == REG)
1113 unsigned int regno = REGNO (x);
1114 unsigned int endregno
1115 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1116 : hard_regno_nregs[regno][GET_MODE (x)]);
1117 unsigned int i;
1119 for (i = regno; i < endregno; i++)
1121 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1122 remove_invalid_refs (i);
1124 REG_IN_TABLE (i) = REG_TICK (i);
1125 SUBREG_TICKED (i) = -1;
1128 return 0;
1131 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1132 pseudo if they don't use overlapping words. We handle only pseudos
1133 here for simplicity. */
1134 if (code == SUBREG && REG_P (SUBREG_REG (x))
1135 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1137 unsigned int i = REGNO (SUBREG_REG (x));
1139 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1141 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1142 the last store to this register really stored into this
1143 subreg, then remove the memory of this subreg.
1144 Otherwise, remove any memory of the entire register and
1145 all its subregs from the table. */
1146 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1147 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1148 remove_invalid_refs (i);
1149 else
1150 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1153 REG_IN_TABLE (i) = REG_TICK (i);
1154 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1155 return 0;
1158 /* If X is a comparison or a COMPARE and either operand is a register
1159 that does not have a quantity, give it one. This is so that a later
1160 call to record_jump_equiv won't cause X to be assigned a different
1161 hash code and not found in the table after that call.
1163 It is not necessary to do this here, since rehash_using_reg can
1164 fix up the table later, but doing this here eliminates the need to
1165 call that expensive function in the most common case where the only
1166 use of the register is in the comparison. */
1168 if (code == COMPARE || COMPARISON_P (x))
1170 if (REG_P (XEXP (x, 0))
1171 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1172 if (insert_regs (XEXP (x, 0), NULL, 0))
1174 rehash_using_reg (XEXP (x, 0));
1175 changed = 1;
1178 if (REG_P (XEXP (x, 1))
1179 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1180 if (insert_regs (XEXP (x, 1), NULL, 0))
1182 rehash_using_reg (XEXP (x, 1));
1183 changed = 1;
1187 fmt = GET_RTX_FORMAT (code);
1188 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1189 if (fmt[i] == 'e')
1190 changed |= mention_regs (XEXP (x, i));
1191 else if (fmt[i] == 'E')
1192 for (j = 0; j < XVECLEN (x, i); j++)
1193 changed |= mention_regs (XVECEXP (x, i, j));
1195 return changed;
1198 /* Update the register quantities for inserting X into the hash table
1199 with a value equivalent to CLASSP.
1200 (If the class does not contain a REG, it is irrelevant.)
1201 If MODIFIED is nonzero, X is a destination; it is being modified.
1202 Note that delete_reg_equiv should be called on a register
1203 before insert_regs is done on that register with MODIFIED != 0.
1205 Nonzero value means that elements of reg_qty have changed
1206 so X's hash code may be different. */
1208 static int
1209 insert_regs (rtx x, struct table_elt *classp, int modified)
1211 if (REG_P (x))
1213 unsigned int regno = REGNO (x);
1214 int qty_valid;
1216 /* If REGNO is in the equivalence table already but is of the
1217 wrong mode for that equivalence, don't do anything here. */
1219 qty_valid = REGNO_QTY_VALID_P (regno);
1220 if (qty_valid)
1222 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1224 if (ent->mode != GET_MODE (x))
1225 return 0;
1228 if (modified || ! qty_valid)
1230 if (classp)
1231 for (classp = classp->first_same_value;
1232 classp != 0;
1233 classp = classp->next_same_value)
1234 if (REG_P (classp->exp)
1235 && GET_MODE (classp->exp) == GET_MODE (x))
1237 make_regs_eqv (regno, REGNO (classp->exp));
1238 return 1;
1241 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1242 than REG_IN_TABLE to find out if there was only a single preceding
1243 invalidation - for the SUBREG - or another one, which would be
1244 for the full register. However, if we find here that REG_TICK
1245 indicates that the register is invalid, it means that it has
1246 been invalidated in a separate operation. The SUBREG might be used
1247 now (then this is a recursive call), or we might use the full REG
1248 now and a SUBREG of it later. So bump up REG_TICK so that
1249 mention_regs will do the right thing. */
1250 if (! modified
1251 && REG_IN_TABLE (regno) >= 0
1252 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1253 REG_TICK (regno)++;
1254 make_new_qty (regno, GET_MODE (x));
1255 return 1;
1258 return 0;
1261 /* If X is a SUBREG, we will likely be inserting the inner register in the
1262 table. If that register doesn't have an assigned quantity number at
1263 this point but does later, the insertion that we will be doing now will
1264 not be accessible because its hash code will have changed. So assign
1265 a quantity number now. */
1267 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1268 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1270 insert_regs (SUBREG_REG (x), NULL, 0);
1271 mention_regs (x);
1272 return 1;
1274 else
1275 return mention_regs (x);
1278 /* Look in or update the hash table. */
1280 /* Remove table element ELT from use in the table.
1281 HASH is its hash code, made using the HASH macro.
1282 It's an argument because often that is known in advance
1283 and we save much time not recomputing it. */
1285 static void
1286 remove_from_table (struct table_elt *elt, unsigned int hash)
1288 if (elt == 0)
1289 return;
1291 /* Mark this element as removed. See cse_insn. */
1292 elt->first_same_value = 0;
1294 /* Remove the table element from its equivalence class. */
1297 struct table_elt *prev = elt->prev_same_value;
1298 struct table_elt *next = elt->next_same_value;
1300 if (next)
1301 next->prev_same_value = prev;
1303 if (prev)
1304 prev->next_same_value = next;
1305 else
1307 struct table_elt *newfirst = next;
1308 while (next)
1310 next->first_same_value = newfirst;
1311 next = next->next_same_value;
1316 /* Remove the table element from its hash bucket. */
1319 struct table_elt *prev = elt->prev_same_hash;
1320 struct table_elt *next = elt->next_same_hash;
1322 if (next)
1323 next->prev_same_hash = prev;
1325 if (prev)
1326 prev->next_same_hash = next;
1327 else if (table[hash] == elt)
1328 table[hash] = next;
1329 else
1331 /* This entry is not in the proper hash bucket. This can happen
1332 when two classes were merged by `merge_equiv_classes'. Search
1333 for the hash bucket that it heads. This happens only very
1334 rarely, so the cost is acceptable. */
1335 for (hash = 0; hash < HASH_SIZE; hash++)
1336 if (table[hash] == elt)
1337 table[hash] = next;
1341 /* Remove the table element from its related-value circular chain. */
1343 if (elt->related_value != 0 && elt->related_value != elt)
1345 struct table_elt *p = elt->related_value;
1347 while (p->related_value != elt)
1348 p = p->related_value;
1349 p->related_value = elt->related_value;
1350 if (p->related_value == p)
1351 p->related_value = 0;
1354 /* Now add it to the free element chain. */
1355 elt->next_same_hash = free_element_chain;
1356 free_element_chain = elt;
1359 /* Look up X in the hash table and return its table element,
1360 or 0 if X is not in the table.
1362 MODE is the machine-mode of X, or if X is an integer constant
1363 with VOIDmode then MODE is the mode with which X will be used.
1365 Here we are satisfied to find an expression whose tree structure
1366 looks like X. */
1368 static struct table_elt *
1369 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1371 struct table_elt *p;
1373 for (p = table[hash]; p; p = p->next_same_hash)
1374 if (mode == p->mode && ((x == p->exp && REG_P (x))
1375 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1376 return p;
1378 return 0;
1381 /* Like `lookup' but don't care whether the table element uses invalid regs.
1382 Also ignore discrepancies in the machine mode of a register. */
1384 static struct table_elt *
1385 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1387 struct table_elt *p;
1389 if (REG_P (x))
1391 unsigned int regno = REGNO (x);
1393 /* Don't check the machine mode when comparing registers;
1394 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1395 for (p = table[hash]; p; p = p->next_same_hash)
1396 if (REG_P (p->exp)
1397 && REGNO (p->exp) == regno)
1398 return p;
1400 else
1402 for (p = table[hash]; p; p = p->next_same_hash)
1403 if (mode == p->mode
1404 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1405 return p;
1408 return 0;
1411 /* Look for an expression equivalent to X and with code CODE.
1412 If one is found, return that expression. */
1414 static rtx
1415 lookup_as_function (rtx x, enum rtx_code code)
1417 struct table_elt *p
1418 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1420 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1421 long as we are narrowing. So if we looked in vain for a mode narrower
1422 than word_mode before, look for word_mode now. */
1423 if (p == 0 && code == CONST_INT
1424 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1426 x = copy_rtx (x);
1427 PUT_MODE (x, word_mode);
1428 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1431 if (p == 0)
1432 return 0;
1434 for (p = p->first_same_value; p; p = p->next_same_value)
1435 if (GET_CODE (p->exp) == code
1436 /* Make sure this is a valid entry in the table. */
1437 && exp_equiv_p (p->exp, p->exp, 1, false))
1438 return p->exp;
1440 return 0;
1443 /* Insert X in the hash table, assuming HASH is its hash code
1444 and CLASSP is an element of the class it should go in
1445 (or 0 if a new class should be made).
1446 It is inserted at the proper position to keep the class in
1447 the order cheapest first.
1449 MODE is the machine-mode of X, or if X is an integer constant
1450 with VOIDmode then MODE is the mode with which X will be used.
1452 For elements of equal cheapness, the most recent one
1453 goes in front, except that the first element in the list
1454 remains first unless a cheaper element is added. The order of
1455 pseudo-registers does not matter, as canon_reg will be called to
1456 find the cheapest when a register is retrieved from the table.
1458 The in_memory field in the hash table element is set to 0.
1459 The caller must set it nonzero if appropriate.
1461 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1462 and if insert_regs returns a nonzero value
1463 you must then recompute its hash code before calling here.
1465 If necessary, update table showing constant values of quantities. */
1467 #define CHEAPER(X, Y) \
1468 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1470 static struct table_elt *
1471 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1473 struct table_elt *elt;
1475 /* If X is a register and we haven't made a quantity for it,
1476 something is wrong. */
1477 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1479 /* If X is a hard register, show it is being put in the table. */
1480 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1482 unsigned int regno = REGNO (x);
1483 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1484 unsigned int i;
1486 for (i = regno; i < endregno; i++)
1487 SET_HARD_REG_BIT (hard_regs_in_table, i);
1490 /* Put an element for X into the right hash bucket. */
1492 elt = free_element_chain;
1493 if (elt)
1494 free_element_chain = elt->next_same_hash;
1495 else
1497 n_elements_made++;
1498 elt = xmalloc (sizeof (struct table_elt));
1501 elt->exp = x;
1502 elt->canon_exp = NULL_RTX;
1503 elt->cost = COST (x);
1504 elt->regcost = approx_reg_cost (x);
1505 elt->next_same_value = 0;
1506 elt->prev_same_value = 0;
1507 elt->next_same_hash = table[hash];
1508 elt->prev_same_hash = 0;
1509 elt->related_value = 0;
1510 elt->in_memory = 0;
1511 elt->mode = mode;
1512 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1514 if (table[hash])
1515 table[hash]->prev_same_hash = elt;
1516 table[hash] = elt;
1518 /* Put it into the proper value-class. */
1519 if (classp)
1521 classp = classp->first_same_value;
1522 if (CHEAPER (elt, classp))
1523 /* Insert at the head of the class. */
1525 struct table_elt *p;
1526 elt->next_same_value = classp;
1527 classp->prev_same_value = elt;
1528 elt->first_same_value = elt;
1530 for (p = classp; p; p = p->next_same_value)
1531 p->first_same_value = elt;
1533 else
1535 /* Insert not at head of the class. */
1536 /* Put it after the last element cheaper than X. */
1537 struct table_elt *p, *next;
1539 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1540 p = next);
1542 /* Put it after P and before NEXT. */
1543 elt->next_same_value = next;
1544 if (next)
1545 next->prev_same_value = elt;
1547 elt->prev_same_value = p;
1548 p->next_same_value = elt;
1549 elt->first_same_value = classp;
1552 else
1553 elt->first_same_value = elt;
1555 /* If this is a constant being set equivalent to a register or a register
1556 being set equivalent to a constant, note the constant equivalence.
1558 If this is a constant, it cannot be equivalent to a different constant,
1559 and a constant is the only thing that can be cheaper than a register. So
1560 we know the register is the head of the class (before the constant was
1561 inserted).
1563 If this is a register that is not already known equivalent to a
1564 constant, we must check the entire class.
1566 If this is a register that is already known equivalent to an insn,
1567 update the qtys `const_insn' to show that `this_insn' is the latest
1568 insn making that quantity equivalent to the constant. */
1570 if (elt->is_const && classp && REG_P (classp->exp)
1571 && !REG_P (x))
1573 int exp_q = REG_QTY (REGNO (classp->exp));
1574 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1576 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1577 exp_ent->const_insn = this_insn;
1580 else if (REG_P (x)
1581 && classp
1582 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1583 && ! elt->is_const)
1585 struct table_elt *p;
1587 for (p = classp; p != 0; p = p->next_same_value)
1589 if (p->is_const && !REG_P (p->exp))
1591 int x_q = REG_QTY (REGNO (x));
1592 struct qty_table_elem *x_ent = &qty_table[x_q];
1594 x_ent->const_rtx
1595 = gen_lowpart (GET_MODE (x), p->exp);
1596 x_ent->const_insn = this_insn;
1597 break;
1602 else if (REG_P (x)
1603 && qty_table[REG_QTY (REGNO (x))].const_rtx
1604 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1605 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1607 /* If this is a constant with symbolic value,
1608 and it has a term with an explicit integer value,
1609 link it up with related expressions. */
1610 if (GET_CODE (x) == CONST)
1612 rtx subexp = get_related_value (x);
1613 unsigned subhash;
1614 struct table_elt *subelt, *subelt_prev;
1616 if (subexp != 0)
1618 /* Get the integer-free subexpression in the hash table. */
1619 subhash = SAFE_HASH (subexp, mode);
1620 subelt = lookup (subexp, subhash, mode);
1621 if (subelt == 0)
1622 subelt = insert (subexp, NULL, subhash, mode);
1623 /* Initialize SUBELT's circular chain if it has none. */
1624 if (subelt->related_value == 0)
1625 subelt->related_value = subelt;
1626 /* Find the element in the circular chain that precedes SUBELT. */
1627 subelt_prev = subelt;
1628 while (subelt_prev->related_value != subelt)
1629 subelt_prev = subelt_prev->related_value;
1630 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1631 This way the element that follows SUBELT is the oldest one. */
1632 elt->related_value = subelt_prev->related_value;
1633 subelt_prev->related_value = elt;
1637 return elt;
1640 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1641 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1642 the two classes equivalent.
1644 CLASS1 will be the surviving class; CLASS2 should not be used after this
1645 call.
1647 Any invalid entries in CLASS2 will not be copied. */
1649 static void
1650 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1652 struct table_elt *elt, *next, *new;
1654 /* Ensure we start with the head of the classes. */
1655 class1 = class1->first_same_value;
1656 class2 = class2->first_same_value;
1658 /* If they were already equal, forget it. */
1659 if (class1 == class2)
1660 return;
1662 for (elt = class2; elt; elt = next)
1664 unsigned int hash;
1665 rtx exp = elt->exp;
1666 enum machine_mode mode = elt->mode;
1668 next = elt->next_same_value;
1670 /* Remove old entry, make a new one in CLASS1's class.
1671 Don't do this for invalid entries as we cannot find their
1672 hash code (it also isn't necessary). */
1673 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1675 bool need_rehash = false;
1677 hash_arg_in_memory = 0;
1678 hash = HASH (exp, mode);
1680 if (REG_P (exp))
1682 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1683 delete_reg_equiv (REGNO (exp));
1686 remove_from_table (elt, hash);
1688 if (insert_regs (exp, class1, 0) || need_rehash)
1690 rehash_using_reg (exp);
1691 hash = HASH (exp, mode);
1693 new = insert (exp, class1, hash, mode);
1694 new->in_memory = hash_arg_in_memory;
1699 /* Flush the entire hash table. */
1701 static void
1702 flush_hash_table (void)
1704 int i;
1705 struct table_elt *p;
1707 for (i = 0; i < HASH_SIZE; i++)
1708 for (p = table[i]; p; p = table[i])
1710 /* Note that invalidate can remove elements
1711 after P in the current hash chain. */
1712 if (REG_P (p->exp))
1713 invalidate (p->exp, p->mode);
1714 else
1715 remove_from_table (p, i);
1719 /* Function called for each rtx to check whether true dependence exist. */
1720 struct check_dependence_data
1722 enum machine_mode mode;
1723 rtx exp;
1724 rtx addr;
1727 static int
1728 check_dependence (rtx *x, void *data)
1730 struct check_dependence_data *d = (struct check_dependence_data *) data;
1731 if (*x && MEM_P (*x))
1732 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1733 cse_rtx_varies_p);
1734 else
1735 return 0;
1738 /* Remove from the hash table, or mark as invalid, all expressions whose
1739 values could be altered by storing in X. X is a register, a subreg, or
1740 a memory reference with nonvarying address (because, when a memory
1741 reference with a varying address is stored in, all memory references are
1742 removed by invalidate_memory so specific invalidation is superfluous).
1743 FULL_MODE, if not VOIDmode, indicates that this much should be
1744 invalidated instead of just the amount indicated by the mode of X. This
1745 is only used for bitfield stores into memory.
1747 A nonvarying address may be just a register or just a symbol reference,
1748 or it may be either of those plus a numeric offset. */
1750 static void
1751 invalidate (rtx x, enum machine_mode full_mode)
1753 int i;
1754 struct table_elt *p;
1755 rtx addr;
1757 switch (GET_CODE (x))
1759 case REG:
1761 /* If X is a register, dependencies on its contents are recorded
1762 through the qty number mechanism. Just change the qty number of
1763 the register, mark it as invalid for expressions that refer to it,
1764 and remove it itself. */
1765 unsigned int regno = REGNO (x);
1766 unsigned int hash = HASH (x, GET_MODE (x));
1768 /* Remove REGNO from any quantity list it might be on and indicate
1769 that its value might have changed. If it is a pseudo, remove its
1770 entry from the hash table.
1772 For a hard register, we do the first two actions above for any
1773 additional hard registers corresponding to X. Then, if any of these
1774 registers are in the table, we must remove any REG entries that
1775 overlap these registers. */
1777 delete_reg_equiv (regno);
1778 REG_TICK (regno)++;
1779 SUBREG_TICKED (regno) = -1;
1781 if (regno >= FIRST_PSEUDO_REGISTER)
1783 /* Because a register can be referenced in more than one mode,
1784 we might have to remove more than one table entry. */
1785 struct table_elt *elt;
1787 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1788 remove_from_table (elt, hash);
1790 else
1792 HOST_WIDE_INT in_table
1793 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1794 unsigned int endregno
1795 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1796 unsigned int tregno, tendregno, rn;
1797 struct table_elt *p, *next;
1799 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1801 for (rn = regno + 1; rn < endregno; rn++)
1803 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1804 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1805 delete_reg_equiv (rn);
1806 REG_TICK (rn)++;
1807 SUBREG_TICKED (rn) = -1;
1810 if (in_table)
1811 for (hash = 0; hash < HASH_SIZE; hash++)
1812 for (p = table[hash]; p; p = next)
1814 next = p->next_same_hash;
1816 if (!REG_P (p->exp)
1817 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1818 continue;
1820 tregno = REGNO (p->exp);
1821 tendregno
1822 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1823 if (tendregno > regno && tregno < endregno)
1824 remove_from_table (p, hash);
1828 return;
1830 case SUBREG:
1831 invalidate (SUBREG_REG (x), VOIDmode);
1832 return;
1834 case PARALLEL:
1835 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1836 invalidate (XVECEXP (x, 0, i), VOIDmode);
1837 return;
1839 case EXPR_LIST:
1840 /* This is part of a disjoint return value; extract the location in
1841 question ignoring the offset. */
1842 invalidate (XEXP (x, 0), VOIDmode);
1843 return;
1845 case MEM:
1846 addr = canon_rtx (get_addr (XEXP (x, 0)));
1847 /* Calculate the canonical version of X here so that
1848 true_dependence doesn't generate new RTL for X on each call. */
1849 x = canon_rtx (x);
1851 /* Remove all hash table elements that refer to overlapping pieces of
1852 memory. */
1853 if (full_mode == VOIDmode)
1854 full_mode = GET_MODE (x);
1856 for (i = 0; i < HASH_SIZE; i++)
1858 struct table_elt *next;
1860 for (p = table[i]; p; p = next)
1862 next = p->next_same_hash;
1863 if (p->in_memory)
1865 struct check_dependence_data d;
1867 /* Just canonicalize the expression once;
1868 otherwise each time we call invalidate
1869 true_dependence will canonicalize the
1870 expression again. */
1871 if (!p->canon_exp)
1872 p->canon_exp = canon_rtx (p->exp);
1873 d.exp = x;
1874 d.addr = addr;
1875 d.mode = full_mode;
1876 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1877 remove_from_table (p, i);
1881 return;
1883 default:
1884 gcc_unreachable ();
1888 /* Remove all expressions that refer to register REGNO,
1889 since they are already invalid, and we are about to
1890 mark that register valid again and don't want the old
1891 expressions to reappear as valid. */
1893 static void
1894 remove_invalid_refs (unsigned int regno)
1896 unsigned int i;
1897 struct table_elt *p, *next;
1899 for (i = 0; i < HASH_SIZE; i++)
1900 for (p = table[i]; p; p = next)
1902 next = p->next_same_hash;
1903 if (!REG_P (p->exp)
1904 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1905 remove_from_table (p, i);
1909 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1910 and mode MODE. */
1911 static void
1912 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1913 enum machine_mode mode)
1915 unsigned int i;
1916 struct table_elt *p, *next;
1917 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1919 for (i = 0; i < HASH_SIZE; i++)
1920 for (p = table[i]; p; p = next)
1922 rtx exp = p->exp;
1923 next = p->next_same_hash;
1925 if (!REG_P (exp)
1926 && (GET_CODE (exp) != SUBREG
1927 || !REG_P (SUBREG_REG (exp))
1928 || REGNO (SUBREG_REG (exp)) != regno
1929 || (((SUBREG_BYTE (exp)
1930 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1931 && SUBREG_BYTE (exp) <= end))
1932 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1933 remove_from_table (p, i);
1937 /* Recompute the hash codes of any valid entries in the hash table that
1938 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1940 This is called when we make a jump equivalence. */
1942 static void
1943 rehash_using_reg (rtx x)
1945 unsigned int i;
1946 struct table_elt *p, *next;
1947 unsigned hash;
1949 if (GET_CODE (x) == SUBREG)
1950 x = SUBREG_REG (x);
1952 /* If X is not a register or if the register is known not to be in any
1953 valid entries in the table, we have no work to do. */
1955 if (!REG_P (x)
1956 || REG_IN_TABLE (REGNO (x)) < 0
1957 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1958 return;
1960 /* Scan all hash chains looking for valid entries that mention X.
1961 If we find one and it is in the wrong hash chain, move it. */
1963 for (i = 0; i < HASH_SIZE; i++)
1964 for (p = table[i]; p; p = next)
1966 next = p->next_same_hash;
1967 if (reg_mentioned_p (x, p->exp)
1968 && exp_equiv_p (p->exp, p->exp, 1, false)
1969 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1971 if (p->next_same_hash)
1972 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1974 if (p->prev_same_hash)
1975 p->prev_same_hash->next_same_hash = p->next_same_hash;
1976 else
1977 table[i] = p->next_same_hash;
1979 p->next_same_hash = table[hash];
1980 p->prev_same_hash = 0;
1981 if (table[hash])
1982 table[hash]->prev_same_hash = p;
1983 table[hash] = p;
1988 /* Remove from the hash table any expression that is a call-clobbered
1989 register. Also update their TICK values. */
1991 static void
1992 invalidate_for_call (void)
1994 unsigned int regno, endregno;
1995 unsigned int i;
1996 unsigned hash;
1997 struct table_elt *p, *next;
1998 int in_table = 0;
2000 /* Go through all the hard registers. For each that is clobbered in
2001 a CALL_INSN, remove the register from quantity chains and update
2002 reg_tick if defined. Also see if any of these registers is currently
2003 in the table. */
2005 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2006 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2008 delete_reg_equiv (regno);
2009 if (REG_TICK (regno) >= 0)
2011 REG_TICK (regno)++;
2012 SUBREG_TICKED (regno) = -1;
2015 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2018 /* In the case where we have no call-clobbered hard registers in the
2019 table, we are done. Otherwise, scan the table and remove any
2020 entry that overlaps a call-clobbered register. */
2022 if (in_table)
2023 for (hash = 0; hash < HASH_SIZE; hash++)
2024 for (p = table[hash]; p; p = next)
2026 next = p->next_same_hash;
2028 if (!REG_P (p->exp)
2029 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2030 continue;
2032 regno = REGNO (p->exp);
2033 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
2035 for (i = regno; i < endregno; i++)
2036 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2038 remove_from_table (p, hash);
2039 break;
2044 /* Given an expression X of type CONST,
2045 and ELT which is its table entry (or 0 if it
2046 is not in the hash table),
2047 return an alternate expression for X as a register plus integer.
2048 If none can be found, return 0. */
2050 static rtx
2051 use_related_value (rtx x, struct table_elt *elt)
2053 struct table_elt *relt = 0;
2054 struct table_elt *p, *q;
2055 HOST_WIDE_INT offset;
2057 /* First, is there anything related known?
2058 If we have a table element, we can tell from that.
2059 Otherwise, must look it up. */
2061 if (elt != 0 && elt->related_value != 0)
2062 relt = elt;
2063 else if (elt == 0 && GET_CODE (x) == CONST)
2065 rtx subexp = get_related_value (x);
2066 if (subexp != 0)
2067 relt = lookup (subexp,
2068 SAFE_HASH (subexp, GET_MODE (subexp)),
2069 GET_MODE (subexp));
2072 if (relt == 0)
2073 return 0;
2075 /* Search all related table entries for one that has an
2076 equivalent register. */
2078 p = relt;
2079 while (1)
2081 /* This loop is strange in that it is executed in two different cases.
2082 The first is when X is already in the table. Then it is searching
2083 the RELATED_VALUE list of X's class (RELT). The second case is when
2084 X is not in the table. Then RELT points to a class for the related
2085 value.
2087 Ensure that, whatever case we are in, that we ignore classes that have
2088 the same value as X. */
2090 if (rtx_equal_p (x, p->exp))
2091 q = 0;
2092 else
2093 for (q = p->first_same_value; q; q = q->next_same_value)
2094 if (REG_P (q->exp))
2095 break;
2097 if (q)
2098 break;
2100 p = p->related_value;
2102 /* We went all the way around, so there is nothing to be found.
2103 Alternatively, perhaps RELT was in the table for some other reason
2104 and it has no related values recorded. */
2105 if (p == relt || p == 0)
2106 break;
2109 if (q == 0)
2110 return 0;
2112 offset = (get_integer_term (x) - get_integer_term (p->exp));
2113 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2114 return plus_constant (q->exp, offset);
2117 /* Hash a string. Just add its bytes up. */
2118 static inline unsigned
2119 hash_rtx_string (const char *ps)
2121 unsigned hash = 0;
2122 const unsigned char *p = (const unsigned char *) ps;
2124 if (p)
2125 while (*p)
2126 hash += *p++;
2128 return hash;
2131 /* Hash an rtx. We are careful to make sure the value is never negative.
2132 Equivalent registers hash identically.
2133 MODE is used in hashing for CONST_INTs only;
2134 otherwise the mode of X is used.
2136 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2138 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2139 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2141 Note that cse_insn knows that the hash code of a MEM expression
2142 is just (int) MEM plus the hash code of the address. */
2144 unsigned
2145 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2146 int *hash_arg_in_memory_p, bool have_reg_qty)
2148 int i, j;
2149 unsigned hash = 0;
2150 enum rtx_code code;
2151 const char *fmt;
2153 /* Used to turn recursion into iteration. We can't rely on GCC's
2154 tail-recursion elimination since we need to keep accumulating values
2155 in HASH. */
2156 repeat:
2157 if (x == 0)
2158 return hash;
2160 code = GET_CODE (x);
2161 switch (code)
2163 case REG:
2165 unsigned int regno = REGNO (x);
2167 if (!reload_completed)
2169 /* On some machines, we can't record any non-fixed hard register,
2170 because extending its life will cause reload problems. We
2171 consider ap, fp, sp, gp to be fixed for this purpose.
2173 We also consider CCmode registers to be fixed for this purpose;
2174 failure to do so leads to failure to simplify 0<100 type of
2175 conditionals.
2177 On all machines, we can't record any global registers.
2178 Nor should we record any register that is in a small
2179 class, as defined by CLASS_LIKELY_SPILLED_P. */
2180 bool record;
2182 if (regno >= FIRST_PSEUDO_REGISTER)
2183 record = true;
2184 else if (x == frame_pointer_rtx
2185 || x == hard_frame_pointer_rtx
2186 || x == arg_pointer_rtx
2187 || x == stack_pointer_rtx
2188 || x == pic_offset_table_rtx)
2189 record = true;
2190 else if (global_regs[regno])
2191 record = false;
2192 else if (fixed_regs[regno])
2193 record = true;
2194 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2195 record = true;
2196 else if (SMALL_REGISTER_CLASSES)
2197 record = false;
2198 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2199 record = false;
2200 else
2201 record = true;
2203 if (!record)
2205 *do_not_record_p = 1;
2206 return 0;
2210 hash += ((unsigned int) REG << 7);
2211 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2212 return hash;
2215 /* We handle SUBREG of a REG specially because the underlying
2216 reg changes its hash value with every value change; we don't
2217 want to have to forget unrelated subregs when one subreg changes. */
2218 case SUBREG:
2220 if (REG_P (SUBREG_REG (x)))
2222 hash += (((unsigned int) SUBREG << 7)
2223 + REGNO (SUBREG_REG (x))
2224 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2225 return hash;
2227 break;
2230 case CONST_INT:
2231 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2232 + (unsigned int) INTVAL (x));
2233 return hash;
2235 case CONST_DOUBLE:
2236 /* This is like the general case, except that it only counts
2237 the integers representing the constant. */
2238 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2239 if (GET_MODE (x) != VOIDmode)
2240 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2241 else
2242 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2243 + (unsigned int) CONST_DOUBLE_HIGH (x));
2244 return hash;
2246 case CONST_VECTOR:
2248 int units;
2249 rtx elt;
2251 units = CONST_VECTOR_NUNITS (x);
2253 for (i = 0; i < units; ++i)
2255 elt = CONST_VECTOR_ELT (x, i);
2256 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2257 hash_arg_in_memory_p, have_reg_qty);
2260 return hash;
2263 /* Assume there is only one rtx object for any given label. */
2264 case LABEL_REF:
2265 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2266 differences and differences between each stage's debugging dumps. */
2267 hash += (((unsigned int) LABEL_REF << 7)
2268 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2269 return hash;
2271 case SYMBOL_REF:
2273 /* Don't hash on the symbol's address to avoid bootstrap differences.
2274 Different hash values may cause expressions to be recorded in
2275 different orders and thus different registers to be used in the
2276 final assembler. This also avoids differences in the dump files
2277 between various stages. */
2278 unsigned int h = 0;
2279 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2281 while (*p)
2282 h += (h << 7) + *p++; /* ??? revisit */
2284 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2285 return hash;
2288 case MEM:
2289 /* We don't record if marked volatile or if BLKmode since we don't
2290 know the size of the move. */
2291 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2293 *do_not_record_p = 1;
2294 return 0;
2296 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2297 *hash_arg_in_memory_p = 1;
2299 /* Now that we have already found this special case,
2300 might as well speed it up as much as possible. */
2301 hash += (unsigned) MEM;
2302 x = XEXP (x, 0);
2303 goto repeat;
2305 case USE:
2306 /* A USE that mentions non-volatile memory needs special
2307 handling since the MEM may be BLKmode which normally
2308 prevents an entry from being made. Pure calls are
2309 marked by a USE which mentions BLKmode memory.
2310 See calls.c:emit_call_1. */
2311 if (MEM_P (XEXP (x, 0))
2312 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2314 hash += (unsigned) USE;
2315 x = XEXP (x, 0);
2317 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2318 *hash_arg_in_memory_p = 1;
2320 /* Now that we have already found this special case,
2321 might as well speed it up as much as possible. */
2322 hash += (unsigned) MEM;
2323 x = XEXP (x, 0);
2324 goto repeat;
2326 break;
2328 case PRE_DEC:
2329 case PRE_INC:
2330 case POST_DEC:
2331 case POST_INC:
2332 case PRE_MODIFY:
2333 case POST_MODIFY:
2334 case PC:
2335 case CC0:
2336 case CALL:
2337 case UNSPEC_VOLATILE:
2338 *do_not_record_p = 1;
2339 return 0;
2341 case ASM_OPERANDS:
2342 if (MEM_VOLATILE_P (x))
2344 *do_not_record_p = 1;
2345 return 0;
2347 else
2349 /* We don't want to take the filename and line into account. */
2350 hash += (unsigned) code + (unsigned) GET_MODE (x)
2351 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2352 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2353 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2355 if (ASM_OPERANDS_INPUT_LENGTH (x))
2357 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2359 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2360 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2361 do_not_record_p, hash_arg_in_memory_p,
2362 have_reg_qty)
2363 + hash_rtx_string
2364 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2367 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2368 x = ASM_OPERANDS_INPUT (x, 0);
2369 mode = GET_MODE (x);
2370 goto repeat;
2373 return hash;
2375 break;
2377 default:
2378 break;
2381 i = GET_RTX_LENGTH (code) - 1;
2382 hash += (unsigned) code + (unsigned) GET_MODE (x);
2383 fmt = GET_RTX_FORMAT (code);
2384 for (; i >= 0; i--)
2386 switch (fmt[i])
2388 case 'e':
2389 /* If we are about to do the last recursive call
2390 needed at this level, change it into iteration.
2391 This function is called enough to be worth it. */
2392 if (i == 0)
2394 x = XEXP (x, i);
2395 goto repeat;
2398 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2399 hash_arg_in_memory_p, have_reg_qty);
2400 break;
2402 case 'E':
2403 for (j = 0; j < XVECLEN (x, i); j++)
2404 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2405 hash_arg_in_memory_p, have_reg_qty);
2406 break;
2408 case 's':
2409 hash += hash_rtx_string (XSTR (x, i));
2410 break;
2412 case 'i':
2413 hash += (unsigned int) XINT (x, i);
2414 break;
2416 case '0': case 't':
2417 /* Unused. */
2418 break;
2420 default:
2421 gcc_unreachable ();
2425 return hash;
2428 /* Hash an rtx X for cse via hash_rtx.
2429 Stores 1 in do_not_record if any subexpression is volatile.
2430 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2431 does not have the RTX_UNCHANGING_P bit set. */
2433 static inline unsigned
2434 canon_hash (rtx x, enum machine_mode mode)
2436 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2439 /* Like canon_hash but with no side effects, i.e. do_not_record
2440 and hash_arg_in_memory are not changed. */
2442 static inline unsigned
2443 safe_hash (rtx x, enum machine_mode mode)
2445 int dummy_do_not_record;
2446 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2449 /* Return 1 iff X and Y would canonicalize into the same thing,
2450 without actually constructing the canonicalization of either one.
2451 If VALIDATE is nonzero,
2452 we assume X is an expression being processed from the rtl
2453 and Y was found in the hash table. We check register refs
2454 in Y for being marked as valid.
2456 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2459 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2461 int i, j;
2462 enum rtx_code code;
2463 const char *fmt;
2465 /* Note: it is incorrect to assume an expression is equivalent to itself
2466 if VALIDATE is nonzero. */
2467 if (x == y && !validate)
2468 return 1;
2470 if (x == 0 || y == 0)
2471 return x == y;
2473 code = GET_CODE (x);
2474 if (code != GET_CODE (y))
2475 return 0;
2477 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2478 if (GET_MODE (x) != GET_MODE (y))
2479 return 0;
2481 switch (code)
2483 case PC:
2484 case CC0:
2485 case CONST_INT:
2486 return x == y;
2488 case LABEL_REF:
2489 return XEXP (x, 0) == XEXP (y, 0);
2491 case SYMBOL_REF:
2492 return XSTR (x, 0) == XSTR (y, 0);
2494 case REG:
2495 if (for_gcse)
2496 return REGNO (x) == REGNO (y);
2497 else
2499 unsigned int regno = REGNO (y);
2500 unsigned int i;
2501 unsigned int endregno
2502 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2503 : hard_regno_nregs[regno][GET_MODE (y)]);
2505 /* If the quantities are not the same, the expressions are not
2506 equivalent. If there are and we are not to validate, they
2507 are equivalent. Otherwise, ensure all regs are up-to-date. */
2509 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2510 return 0;
2512 if (! validate)
2513 return 1;
2515 for (i = regno; i < endregno; i++)
2516 if (REG_IN_TABLE (i) != REG_TICK (i))
2517 return 0;
2519 return 1;
2522 case MEM:
2523 if (for_gcse)
2525 /* Can't merge two expressions in different alias sets, since we
2526 can decide that the expression is transparent in a block when
2527 it isn't, due to it being set with the different alias set. */
2528 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
2529 return 0;
2531 /* A volatile mem should not be considered equivalent to any
2532 other. */
2533 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2534 return 0;
2536 break;
2538 /* For commutative operations, check both orders. */
2539 case PLUS:
2540 case MULT:
2541 case AND:
2542 case IOR:
2543 case XOR:
2544 case NE:
2545 case EQ:
2546 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2547 validate, for_gcse)
2548 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2549 validate, for_gcse))
2550 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2551 validate, for_gcse)
2552 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2553 validate, for_gcse)));
2555 case ASM_OPERANDS:
2556 /* We don't use the generic code below because we want to
2557 disregard filename and line numbers. */
2559 /* A volatile asm isn't equivalent to any other. */
2560 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2561 return 0;
2563 if (GET_MODE (x) != GET_MODE (y)
2564 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2565 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2566 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2567 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2568 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2569 return 0;
2571 if (ASM_OPERANDS_INPUT_LENGTH (x))
2573 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2574 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2575 ASM_OPERANDS_INPUT (y, i),
2576 validate, for_gcse)
2577 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2578 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2579 return 0;
2582 return 1;
2584 default:
2585 break;
2588 /* Compare the elements. If any pair of corresponding elements
2589 fail to match, return 0 for the whole thing. */
2591 fmt = GET_RTX_FORMAT (code);
2592 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2594 switch (fmt[i])
2596 case 'e':
2597 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2598 validate, for_gcse))
2599 return 0;
2600 break;
2602 case 'E':
2603 if (XVECLEN (x, i) != XVECLEN (y, i))
2604 return 0;
2605 for (j = 0; j < XVECLEN (x, i); j++)
2606 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2607 validate, for_gcse))
2608 return 0;
2609 break;
2611 case 's':
2612 if (strcmp (XSTR (x, i), XSTR (y, i)))
2613 return 0;
2614 break;
2616 case 'i':
2617 if (XINT (x, i) != XINT (y, i))
2618 return 0;
2619 break;
2621 case 'w':
2622 if (XWINT (x, i) != XWINT (y, i))
2623 return 0;
2624 break;
2626 case '0':
2627 case 't':
2628 break;
2630 default:
2631 gcc_unreachable ();
2635 return 1;
2638 /* Return 1 if X has a value that can vary even between two
2639 executions of the program. 0 means X can be compared reliably
2640 against certain constants or near-constants. */
2642 static int
2643 cse_rtx_varies_p (rtx x, int from_alias)
2645 /* We need not check for X and the equivalence class being of the same
2646 mode because if X is equivalent to a constant in some mode, it
2647 doesn't vary in any mode. */
2649 if (REG_P (x)
2650 && REGNO_QTY_VALID_P (REGNO (x)))
2652 int x_q = REG_QTY (REGNO (x));
2653 struct qty_table_elem *x_ent = &qty_table[x_q];
2655 if (GET_MODE (x) == x_ent->mode
2656 && x_ent->const_rtx != NULL_RTX)
2657 return 0;
2660 if (GET_CODE (x) == PLUS
2661 && GET_CODE (XEXP (x, 1)) == CONST_INT
2662 && REG_P (XEXP (x, 0))
2663 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2665 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2666 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2668 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2669 && x0_ent->const_rtx != NULL_RTX)
2670 return 0;
2673 /* This can happen as the result of virtual register instantiation, if
2674 the initial constant is too large to be a valid address. This gives
2675 us a three instruction sequence, load large offset into a register,
2676 load fp minus a constant into a register, then a MEM which is the
2677 sum of the two `constant' registers. */
2678 if (GET_CODE (x) == PLUS
2679 && REG_P (XEXP (x, 0))
2680 && REG_P (XEXP (x, 1))
2681 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2682 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2684 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2685 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2686 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2687 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2689 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2690 && x0_ent->const_rtx != NULL_RTX
2691 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2692 && x1_ent->const_rtx != NULL_RTX)
2693 return 0;
2696 return rtx_varies_p (x, from_alias);
2699 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2700 the result if necessary. INSN is as for canon_reg. */
2702 static void
2703 validate_canon_reg (rtx *xloc, rtx insn)
2705 rtx new = canon_reg (*xloc, insn);
2706 int insn_code;
2708 /* If replacing pseudo with hard reg or vice versa, ensure the
2709 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2710 if (insn != 0 && new != 0
2711 && REG_P (new) && REG_P (*xloc)
2712 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2713 != (REGNO (*xloc) < FIRST_PSEUDO_REGISTER))
2714 || GET_MODE (new) != GET_MODE (*xloc)
2715 || (insn_code = recog_memoized (insn)) < 0
2716 || insn_data[insn_code].n_dups > 0))
2717 validate_change (insn, xloc, new, 1);
2718 else
2719 *xloc = new;
2722 /* Canonicalize an expression:
2723 replace each register reference inside it
2724 with the "oldest" equivalent register.
2726 If INSN is nonzero and we are replacing a pseudo with a hard register
2727 or vice versa, validate_change is used to ensure that INSN remains valid
2728 after we make our substitution. The calls are made with IN_GROUP nonzero
2729 so apply_change_group must be called upon the outermost return from this
2730 function (unless INSN is zero). The result of apply_change_group can
2731 generally be discarded since the changes we are making are optional. */
2733 static rtx
2734 canon_reg (rtx x, rtx insn)
2736 int i;
2737 enum rtx_code code;
2738 const char *fmt;
2740 if (x == 0)
2741 return x;
2743 code = GET_CODE (x);
2744 switch (code)
2746 case PC:
2747 case CC0:
2748 case CONST:
2749 case CONST_INT:
2750 case CONST_DOUBLE:
2751 case CONST_VECTOR:
2752 case SYMBOL_REF:
2753 case LABEL_REF:
2754 case ADDR_VEC:
2755 case ADDR_DIFF_VEC:
2756 return x;
2758 case REG:
2760 int first;
2761 int q;
2762 struct qty_table_elem *ent;
2764 /* Never replace a hard reg, because hard regs can appear
2765 in more than one machine mode, and we must preserve the mode
2766 of each occurrence. Also, some hard regs appear in
2767 MEMs that are shared and mustn't be altered. Don't try to
2768 replace any reg that maps to a reg of class NO_REGS. */
2769 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2770 || ! REGNO_QTY_VALID_P (REGNO (x)))
2771 return x;
2773 q = REG_QTY (REGNO (x));
2774 ent = &qty_table[q];
2775 first = ent->first_reg;
2776 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2777 : REGNO_REG_CLASS (first) == NO_REGS ? x
2778 : gen_rtx_REG (ent->mode, first));
2781 default:
2782 break;
2785 fmt = GET_RTX_FORMAT (code);
2786 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2788 int j;
2790 if (fmt[i] == 'e')
2791 validate_canon_reg (&XEXP (x, i), insn);
2792 else if (fmt[i] == 'E')
2793 for (j = 0; j < XVECLEN (x, i); j++)
2794 validate_canon_reg (&XVECEXP (x, i, j), insn);
2797 return x;
2800 /* LOC is a location within INSN that is an operand address (the contents of
2801 a MEM). Find the best equivalent address to use that is valid for this
2802 insn.
2804 On most CISC machines, complicated address modes are costly, and rtx_cost
2805 is a good approximation for that cost. However, most RISC machines have
2806 only a few (usually only one) memory reference formats. If an address is
2807 valid at all, it is often just as cheap as any other address. Hence, for
2808 RISC machines, we use `address_cost' to compare the costs of various
2809 addresses. For two addresses of equal cost, choose the one with the
2810 highest `rtx_cost' value as that has the potential of eliminating the
2811 most insns. For equal costs, we choose the first in the equivalence
2812 class. Note that we ignore the fact that pseudo registers are cheaper than
2813 hard registers here because we would also prefer the pseudo registers. */
2815 static void
2816 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2818 struct table_elt *elt;
2819 rtx addr = *loc;
2820 struct table_elt *p;
2821 int found_better = 1;
2822 int save_do_not_record = do_not_record;
2823 int save_hash_arg_in_memory = hash_arg_in_memory;
2824 int addr_volatile;
2825 int regno;
2826 unsigned hash;
2828 /* Do not try to replace constant addresses or addresses of local and
2829 argument slots. These MEM expressions are made only once and inserted
2830 in many instructions, as well as being used to control symbol table
2831 output. It is not safe to clobber them.
2833 There are some uncommon cases where the address is already in a register
2834 for some reason, but we cannot take advantage of that because we have
2835 no easy way to unshare the MEM. In addition, looking up all stack
2836 addresses is costly. */
2837 if ((GET_CODE (addr) == PLUS
2838 && REG_P (XEXP (addr, 0))
2839 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2840 && (regno = REGNO (XEXP (addr, 0)),
2841 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2842 || regno == ARG_POINTER_REGNUM))
2843 || (REG_P (addr)
2844 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2845 || regno == HARD_FRAME_POINTER_REGNUM
2846 || regno == ARG_POINTER_REGNUM))
2847 || CONSTANT_ADDRESS_P (addr))
2848 return;
2850 /* If this address is not simply a register, try to fold it. This will
2851 sometimes simplify the expression. Many simplifications
2852 will not be valid, but some, usually applying the associative rule, will
2853 be valid and produce better code. */
2854 if (!REG_P (addr))
2856 rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX);
2857 int addr_folded_cost = address_cost (folded, mode);
2858 int addr_cost = address_cost (addr, mode);
2860 if ((addr_folded_cost < addr_cost
2861 || (addr_folded_cost == addr_cost
2862 /* ??? The rtx_cost comparison is left over from an older
2863 version of this code. It is probably no longer helpful. */
2864 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2865 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2866 && validate_change (insn, loc, folded, 0))
2867 addr = folded;
2870 /* If this address is not in the hash table, we can't look for equivalences
2871 of the whole address. Also, ignore if volatile. */
2873 do_not_record = 0;
2874 hash = HASH (addr, Pmode);
2875 addr_volatile = do_not_record;
2876 do_not_record = save_do_not_record;
2877 hash_arg_in_memory = save_hash_arg_in_memory;
2879 if (addr_volatile)
2880 return;
2882 elt = lookup (addr, hash, Pmode);
2884 if (elt)
2886 /* We need to find the best (under the criteria documented above) entry
2887 in the class that is valid. We use the `flag' field to indicate
2888 choices that were invalid and iterate until we can't find a better
2889 one that hasn't already been tried. */
2891 for (p = elt->first_same_value; p; p = p->next_same_value)
2892 p->flag = 0;
2894 while (found_better)
2896 int best_addr_cost = address_cost (*loc, mode);
2897 int best_rtx_cost = (elt->cost + 1) >> 1;
2898 int exp_cost;
2899 struct table_elt *best_elt = elt;
2901 found_better = 0;
2902 for (p = elt->first_same_value; p; p = p->next_same_value)
2903 if (! p->flag)
2905 if ((REG_P (p->exp)
2906 || exp_equiv_p (p->exp, p->exp, 1, false))
2907 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2908 || (exp_cost == best_addr_cost
2909 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2911 found_better = 1;
2912 best_addr_cost = exp_cost;
2913 best_rtx_cost = (p->cost + 1) >> 1;
2914 best_elt = p;
2918 if (found_better)
2920 if (validate_change (insn, loc,
2921 canon_reg (copy_rtx (best_elt->exp),
2922 NULL_RTX), 0))
2923 return;
2924 else
2925 best_elt->flag = 1;
2930 /* If the address is a binary operation with the first operand a register
2931 and the second a constant, do the same as above, but looking for
2932 equivalences of the register. Then try to simplify before checking for
2933 the best address to use. This catches a few cases: First is when we
2934 have REG+const and the register is another REG+const. We can often merge
2935 the constants and eliminate one insn and one register. It may also be
2936 that a machine has a cheap REG+REG+const. Finally, this improves the
2937 code on the Alpha for unaligned byte stores. */
2939 if (flag_expensive_optimizations
2940 && ARITHMETIC_P (*loc)
2941 && REG_P (XEXP (*loc, 0)))
2943 rtx op1 = XEXP (*loc, 1);
2945 do_not_record = 0;
2946 hash = HASH (XEXP (*loc, 0), Pmode);
2947 do_not_record = save_do_not_record;
2948 hash_arg_in_memory = save_hash_arg_in_memory;
2950 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2951 if (elt == 0)
2952 return;
2954 /* We need to find the best (under the criteria documented above) entry
2955 in the class that is valid. We use the `flag' field to indicate
2956 choices that were invalid and iterate until we can't find a better
2957 one that hasn't already been tried. */
2959 for (p = elt->first_same_value; p; p = p->next_same_value)
2960 p->flag = 0;
2962 while (found_better)
2964 int best_addr_cost = address_cost (*loc, mode);
2965 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2966 struct table_elt *best_elt = elt;
2967 rtx best_rtx = *loc;
2968 int count;
2970 /* This is at worst case an O(n^2) algorithm, so limit our search
2971 to the first 32 elements on the list. This avoids trouble
2972 compiling code with very long basic blocks that can easily
2973 call simplify_gen_binary so many times that we run out of
2974 memory. */
2976 found_better = 0;
2977 for (p = elt->first_same_value, count = 0;
2978 p && count < 32;
2979 p = p->next_same_value, count++)
2980 if (! p->flag
2981 && (REG_P (p->exp)
2982 || exp_equiv_p (p->exp, p->exp, 1, false)))
2984 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
2985 p->exp, op1);
2986 int new_cost;
2988 /* Get the canonical version of the address so we can accept
2989 more. */
2990 new = canon_for_address (new);
2992 new_cost = address_cost (new, mode);
2994 if (new_cost < best_addr_cost
2995 || (new_cost == best_addr_cost
2996 && (COST (new) + 1) >> 1 > best_rtx_cost))
2998 found_better = 1;
2999 best_addr_cost = new_cost;
3000 best_rtx_cost = (COST (new) + 1) >> 1;
3001 best_elt = p;
3002 best_rtx = new;
3006 if (found_better)
3008 if (validate_change (insn, loc,
3009 canon_reg (copy_rtx (best_rtx),
3010 NULL_RTX), 0))
3011 return;
3012 else
3013 best_elt->flag = 1;
3019 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3020 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3021 what values are being compared.
3023 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3024 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3025 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3026 compared to produce cc0.
3028 The return value is the comparison operator and is either the code of
3029 A or the code corresponding to the inverse of the comparison. */
3031 static enum rtx_code
3032 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3033 enum machine_mode *pmode1, enum machine_mode *pmode2)
3035 rtx arg1, arg2;
3037 arg1 = *parg1, arg2 = *parg2;
3039 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3041 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3043 /* Set nonzero when we find something of interest. */
3044 rtx x = 0;
3045 int reverse_code = 0;
3046 struct table_elt *p = 0;
3048 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3049 On machines with CC0, this is the only case that can occur, since
3050 fold_rtx will return the COMPARE or item being compared with zero
3051 when given CC0. */
3053 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3054 x = arg1;
3056 /* If ARG1 is a comparison operator and CODE is testing for
3057 STORE_FLAG_VALUE, get the inner arguments. */
3059 else if (COMPARISON_P (arg1))
3061 #ifdef FLOAT_STORE_FLAG_VALUE
3062 REAL_VALUE_TYPE fsfv;
3063 #endif
3065 if (code == NE
3066 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3067 && code == LT && STORE_FLAG_VALUE == -1)
3068 #ifdef FLOAT_STORE_FLAG_VALUE
3069 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3070 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3071 REAL_VALUE_NEGATIVE (fsfv)))
3072 #endif
3074 x = arg1;
3075 else if (code == EQ
3076 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3077 && code == GE && STORE_FLAG_VALUE == -1)
3078 #ifdef FLOAT_STORE_FLAG_VALUE
3079 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3080 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3081 REAL_VALUE_NEGATIVE (fsfv)))
3082 #endif
3084 x = arg1, reverse_code = 1;
3087 /* ??? We could also check for
3089 (ne (and (eq (...) (const_int 1))) (const_int 0))
3091 and related forms, but let's wait until we see them occurring. */
3093 if (x == 0)
3094 /* Look up ARG1 in the hash table and see if it has an equivalence
3095 that lets us see what is being compared. */
3096 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3097 if (p)
3099 p = p->first_same_value;
3101 /* If what we compare is already known to be constant, that is as
3102 good as it gets.
3103 We need to break the loop in this case, because otherwise we
3104 can have an infinite loop when looking at a reg that is known
3105 to be a constant which is the same as a comparison of a reg
3106 against zero which appears later in the insn stream, which in
3107 turn is constant and the same as the comparison of the first reg
3108 against zero... */
3109 if (p->is_const)
3110 break;
3113 for (; p; p = p->next_same_value)
3115 enum machine_mode inner_mode = GET_MODE (p->exp);
3116 #ifdef FLOAT_STORE_FLAG_VALUE
3117 REAL_VALUE_TYPE fsfv;
3118 #endif
3120 /* If the entry isn't valid, skip it. */
3121 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3122 continue;
3124 if (GET_CODE (p->exp) == COMPARE
3125 /* Another possibility is that this machine has a compare insn
3126 that includes the comparison code. In that case, ARG1 would
3127 be equivalent to a comparison operation that would set ARG1 to
3128 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3129 ORIG_CODE is the actual comparison being done; if it is an EQ,
3130 we must reverse ORIG_CODE. On machine with a negative value
3131 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3132 || ((code == NE
3133 || (code == LT
3134 && GET_MODE_CLASS (inner_mode) == MODE_INT
3135 && (GET_MODE_BITSIZE (inner_mode)
3136 <= HOST_BITS_PER_WIDE_INT)
3137 && (STORE_FLAG_VALUE
3138 & ((HOST_WIDE_INT) 1
3139 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3140 #ifdef FLOAT_STORE_FLAG_VALUE
3141 || (code == LT
3142 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3143 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3144 REAL_VALUE_NEGATIVE (fsfv)))
3145 #endif
3147 && COMPARISON_P (p->exp)))
3149 x = p->exp;
3150 break;
3152 else if ((code == EQ
3153 || (code == GE
3154 && GET_MODE_CLASS (inner_mode) == MODE_INT
3155 && (GET_MODE_BITSIZE (inner_mode)
3156 <= HOST_BITS_PER_WIDE_INT)
3157 && (STORE_FLAG_VALUE
3158 & ((HOST_WIDE_INT) 1
3159 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3160 #ifdef FLOAT_STORE_FLAG_VALUE
3161 || (code == GE
3162 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3163 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3164 REAL_VALUE_NEGATIVE (fsfv)))
3165 #endif
3167 && COMPARISON_P (p->exp))
3169 reverse_code = 1;
3170 x = p->exp;
3171 break;
3174 /* If this non-trapping address, e.g. fp + constant, the
3175 equivalent is a better operand since it may let us predict
3176 the value of the comparison. */
3177 else if (!rtx_addr_can_trap_p (p->exp))
3179 arg1 = p->exp;
3180 continue;
3184 /* If we didn't find a useful equivalence for ARG1, we are done.
3185 Otherwise, set up for the next iteration. */
3186 if (x == 0)
3187 break;
3189 /* If we need to reverse the comparison, make sure that that is
3190 possible -- we can't necessarily infer the value of GE from LT
3191 with floating-point operands. */
3192 if (reverse_code)
3194 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3195 if (reversed == UNKNOWN)
3196 break;
3197 else
3198 code = reversed;
3200 else if (COMPARISON_P (x))
3201 code = GET_CODE (x);
3202 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3205 /* Return our results. Return the modes from before fold_rtx
3206 because fold_rtx might produce const_int, and then it's too late. */
3207 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3208 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3210 return code;
3213 /* If X is a nontrivial arithmetic operation on an argument
3214 for which a constant value can be determined, return
3215 the result of operating on that value, as a constant.
3216 Otherwise, return X, possibly with one or more operands
3217 modified by recursive calls to this function.
3219 If X is a register whose contents are known, we do NOT
3220 return those contents here. equiv_constant is called to
3221 perform that task.
3223 INSN is the insn that we may be modifying. If it is 0, make a copy
3224 of X before modifying it. */
3226 static rtx
3227 fold_rtx (rtx x, rtx insn)
3229 enum rtx_code code;
3230 enum machine_mode mode;
3231 const char *fmt;
3232 int i;
3233 rtx new = 0;
3234 int copied = 0;
3235 int must_swap = 0;
3237 /* Folded equivalents of first two operands of X. */
3238 rtx folded_arg0;
3239 rtx folded_arg1;
3241 /* Constant equivalents of first three operands of X;
3242 0 when no such equivalent is known. */
3243 rtx const_arg0;
3244 rtx const_arg1;
3245 rtx const_arg2;
3247 /* The mode of the first operand of X. We need this for sign and zero
3248 extends. */
3249 enum machine_mode mode_arg0;
3251 if (x == 0)
3252 return x;
3254 mode = GET_MODE (x);
3255 code = GET_CODE (x);
3256 switch (code)
3258 case CONST:
3259 case CONST_INT:
3260 case CONST_DOUBLE:
3261 case CONST_VECTOR:
3262 case SYMBOL_REF:
3263 case LABEL_REF:
3264 case REG:
3265 /* No use simplifying an EXPR_LIST
3266 since they are used only for lists of args
3267 in a function call's REG_EQUAL note. */
3268 case EXPR_LIST:
3269 return x;
3271 #ifdef HAVE_cc0
3272 case CC0:
3273 return prev_insn_cc0;
3274 #endif
3276 case PC:
3277 /* If the next insn is a CODE_LABEL followed by a jump table,
3278 PC's value is a LABEL_REF pointing to that label. That
3279 lets us fold switch statements on the VAX. */
3281 rtx next;
3282 if (insn && tablejump_p (insn, &next, NULL))
3283 return gen_rtx_LABEL_REF (Pmode, next);
3285 break;
3287 case SUBREG:
3288 /* See if we previously assigned a constant value to this SUBREG. */
3289 if ((new = lookup_as_function (x, CONST_INT)) != 0
3290 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3291 return new;
3293 /* If this is a paradoxical SUBREG, we have no idea what value the
3294 extra bits would have. However, if the operand is equivalent
3295 to a SUBREG whose operand is the same as our mode, and all the
3296 modes are within a word, we can just use the inner operand
3297 because these SUBREGs just say how to treat the register.
3299 Similarly if we find an integer constant. */
3301 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3303 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3304 struct table_elt *elt;
3306 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3307 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3308 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3309 imode)) != 0)
3310 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3312 if (CONSTANT_P (elt->exp)
3313 && GET_MODE (elt->exp) == VOIDmode)
3314 return elt->exp;
3316 if (GET_CODE (elt->exp) == SUBREG
3317 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3318 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3319 return copy_rtx (SUBREG_REG (elt->exp));
3322 return x;
3325 /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG.
3326 We might be able to if the SUBREG is extracting a single word in an
3327 integral mode or extracting the low part. */
3329 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3330 const_arg0 = equiv_constant (folded_arg0);
3331 if (const_arg0)
3332 folded_arg0 = const_arg0;
3334 if (folded_arg0 != SUBREG_REG (x))
3336 new = simplify_subreg (mode, folded_arg0,
3337 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3338 if (new)
3339 return new;
3342 if (REG_P (folded_arg0)
3343 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3345 struct table_elt *elt;
3347 elt = lookup (folded_arg0,
3348 HASH (folded_arg0, GET_MODE (folded_arg0)),
3349 GET_MODE (folded_arg0));
3351 if (elt)
3352 elt = elt->first_same_value;
3354 if (subreg_lowpart_p (x))
3355 /* If this is a narrowing SUBREG and our operand is a REG, see
3356 if we can find an equivalence for REG that is an arithmetic
3357 operation in a wider mode where both operands are paradoxical
3358 SUBREGs from objects of our result mode. In that case, we
3359 couldn-t report an equivalent value for that operation, since we
3360 don't know what the extra bits will be. But we can find an
3361 equivalence for this SUBREG by folding that operation in the
3362 narrow mode. This allows us to fold arithmetic in narrow modes
3363 when the machine only supports word-sized arithmetic.
3365 Also look for a case where we have a SUBREG whose operand
3366 is the same as our result. If both modes are smaller
3367 than a word, we are simply interpreting a register in
3368 different modes and we can use the inner value. */
3370 for (; elt; elt = elt->next_same_value)
3372 enum rtx_code eltcode = GET_CODE (elt->exp);
3374 /* Just check for unary and binary operations. */
3375 if (UNARY_P (elt->exp)
3376 && eltcode != SIGN_EXTEND
3377 && eltcode != ZERO_EXTEND
3378 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3379 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3380 && (GET_MODE_CLASS (mode)
3381 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3383 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3385 if (!REG_P (op0) && ! CONSTANT_P (op0))
3386 op0 = fold_rtx (op0, NULL_RTX);
3388 op0 = equiv_constant (op0);
3389 if (op0)
3390 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3391 op0, mode);
3393 else if (ARITHMETIC_P (elt->exp)
3394 && eltcode != DIV && eltcode != MOD
3395 && eltcode != UDIV && eltcode != UMOD
3396 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3397 && eltcode != ROTATE && eltcode != ROTATERT
3398 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3399 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3400 == mode))
3401 || CONSTANT_P (XEXP (elt->exp, 0)))
3402 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3403 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3404 == mode))
3405 || CONSTANT_P (XEXP (elt->exp, 1))))
3407 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3408 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3410 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3411 op0 = fold_rtx (op0, NULL_RTX);
3413 if (op0)
3414 op0 = equiv_constant (op0);
3416 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3417 op1 = fold_rtx (op1, NULL_RTX);
3419 if (op1)
3420 op1 = equiv_constant (op1);
3422 /* If we are looking for the low SImode part of
3423 (ashift:DI c (const_int 32)), it doesn't work
3424 to compute that in SImode, because a 32-bit shift
3425 in SImode is unpredictable. We know the value is 0. */
3426 if (op0 && op1
3427 && GET_CODE (elt->exp) == ASHIFT
3428 && GET_CODE (op1) == CONST_INT
3429 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3431 if (INTVAL (op1)
3432 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3433 /* If the count fits in the inner mode's width,
3434 but exceeds the outer mode's width,
3435 the value will get truncated to 0
3436 by the subreg. */
3437 new = CONST0_RTX (mode);
3438 else
3439 /* If the count exceeds even the inner mode's width,
3440 don't fold this expression. */
3441 new = 0;
3443 else if (op0 && op1)
3444 new = simplify_binary_operation (GET_CODE (elt->exp), mode, op0, op1);
3447 else if (GET_CODE (elt->exp) == SUBREG
3448 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3449 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3450 <= UNITS_PER_WORD)
3451 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3452 new = copy_rtx (SUBREG_REG (elt->exp));
3454 if (new)
3455 return new;
3457 else
3458 /* A SUBREG resulting from a zero extension may fold to zero if
3459 it extracts higher bits than the ZERO_EXTEND's source bits.
3460 FIXME: if combine tried to, er, combine these instructions,
3461 this transformation may be moved to simplify_subreg. */
3462 for (; elt; elt = elt->next_same_value)
3464 if (GET_CODE (elt->exp) == ZERO_EXTEND
3465 && subreg_lsb (x)
3466 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3467 return CONST0_RTX (mode);
3471 return x;
3473 case NOT:
3474 case NEG:
3475 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3476 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3477 new = lookup_as_function (XEXP (x, 0), code);
3478 if (new)
3479 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3480 break;
3482 case MEM:
3483 /* If we are not actually processing an insn, don't try to find the
3484 best address. Not only don't we care, but we could modify the
3485 MEM in an invalid way since we have no insn to validate against. */
3486 if (insn != 0)
3487 find_best_addr (insn, &XEXP (x, 0), GET_MODE (x));
3490 /* Even if we don't fold in the insn itself,
3491 we can safely do so here, in hopes of getting a constant. */
3492 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3493 rtx base = 0;
3494 HOST_WIDE_INT offset = 0;
3496 if (REG_P (addr)
3497 && REGNO_QTY_VALID_P (REGNO (addr)))
3499 int addr_q = REG_QTY (REGNO (addr));
3500 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3502 if (GET_MODE (addr) == addr_ent->mode
3503 && addr_ent->const_rtx != NULL_RTX)
3504 addr = addr_ent->const_rtx;
3507 /* If address is constant, split it into a base and integer offset. */
3508 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3509 base = addr;
3510 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3511 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3513 base = XEXP (XEXP (addr, 0), 0);
3514 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3516 else if (GET_CODE (addr) == LO_SUM
3517 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3518 base = XEXP (addr, 1);
3520 /* If this is a constant pool reference, we can fold it into its
3521 constant to allow better value tracking. */
3522 if (base && GET_CODE (base) == SYMBOL_REF
3523 && CONSTANT_POOL_ADDRESS_P (base))
3525 rtx constant = get_pool_constant (base);
3526 enum machine_mode const_mode = get_pool_mode (base);
3527 rtx new;
3529 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3531 constant_pool_entries_cost = COST (constant);
3532 constant_pool_entries_regcost = approx_reg_cost (constant);
3535 /* If we are loading the full constant, we have an equivalence. */
3536 if (offset == 0 && mode == const_mode)
3537 return constant;
3539 /* If this actually isn't a constant (weird!), we can't do
3540 anything. Otherwise, handle the two most common cases:
3541 extracting a word from a multi-word constant, and extracting
3542 the low-order bits. Other cases don't seem common enough to
3543 worry about. */
3544 if (! CONSTANT_P (constant))
3545 return x;
3547 if (GET_MODE_CLASS (mode) == MODE_INT
3548 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3549 && offset % UNITS_PER_WORD == 0
3550 && (new = operand_subword (constant,
3551 offset / UNITS_PER_WORD,
3552 0, const_mode)) != 0)
3553 return new;
3555 if (((BYTES_BIG_ENDIAN
3556 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3557 || (! BYTES_BIG_ENDIAN && offset == 0))
3558 && (new = gen_lowpart (mode, constant)) != 0)
3559 return new;
3562 /* If this is a reference to a label at a known position in a jump
3563 table, we also know its value. */
3564 if (base && GET_CODE (base) == LABEL_REF)
3566 rtx label = XEXP (base, 0);
3567 rtx table_insn = NEXT_INSN (label);
3569 if (table_insn && JUMP_P (table_insn)
3570 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3572 rtx table = PATTERN (table_insn);
3574 if (offset >= 0
3575 && (offset / GET_MODE_SIZE (GET_MODE (table))
3576 < XVECLEN (table, 0)))
3577 return XVECEXP (table, 0,
3578 offset / GET_MODE_SIZE (GET_MODE (table)));
3580 if (table_insn && JUMP_P (table_insn)
3581 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3583 rtx table = PATTERN (table_insn);
3585 if (offset >= 0
3586 && (offset / GET_MODE_SIZE (GET_MODE (table))
3587 < XVECLEN (table, 1)))
3589 offset /= GET_MODE_SIZE (GET_MODE (table));
3590 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3591 XEXP (table, 0));
3593 if (GET_MODE (table) != Pmode)
3594 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3596 /* Indicate this is a constant. This isn't a
3597 valid form of CONST, but it will only be used
3598 to fold the next insns and then discarded, so
3599 it should be safe.
3601 Note this expression must be explicitly discarded,
3602 by cse_insn, else it may end up in a REG_EQUAL note
3603 and "escape" to cause problems elsewhere. */
3604 return gen_rtx_CONST (GET_MODE (new), new);
3609 return x;
3612 #ifdef NO_FUNCTION_CSE
3613 case CALL:
3614 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3615 return x;
3616 break;
3617 #endif
3619 case ASM_OPERANDS:
3620 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3621 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3622 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3623 break;
3625 default:
3626 break;
3629 const_arg0 = 0;
3630 const_arg1 = 0;
3631 const_arg2 = 0;
3632 mode_arg0 = VOIDmode;
3634 /* Try folding our operands.
3635 Then see which ones have constant values known. */
3637 fmt = GET_RTX_FORMAT (code);
3638 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3639 if (fmt[i] == 'e')
3641 rtx arg = XEXP (x, i);
3642 rtx folded_arg = arg, const_arg = 0;
3643 enum machine_mode mode_arg = GET_MODE (arg);
3644 rtx cheap_arg, expensive_arg;
3645 rtx replacements[2];
3646 int j;
3647 int old_cost = COST_IN (XEXP (x, i), code);
3649 /* Most arguments are cheap, so handle them specially. */
3650 switch (GET_CODE (arg))
3652 case REG:
3653 /* This is the same as calling equiv_constant; it is duplicated
3654 here for speed. */
3655 if (REGNO_QTY_VALID_P (REGNO (arg)))
3657 int arg_q = REG_QTY (REGNO (arg));
3658 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3660 if (arg_ent->const_rtx != NULL_RTX
3661 && !REG_P (arg_ent->const_rtx)
3662 && GET_CODE (arg_ent->const_rtx) != PLUS)
3663 const_arg
3664 = gen_lowpart (GET_MODE (arg),
3665 arg_ent->const_rtx);
3667 break;
3669 case CONST:
3670 case CONST_INT:
3671 case SYMBOL_REF:
3672 case LABEL_REF:
3673 case CONST_DOUBLE:
3674 case CONST_VECTOR:
3675 const_arg = arg;
3676 break;
3678 #ifdef HAVE_cc0
3679 case CC0:
3680 folded_arg = prev_insn_cc0;
3681 mode_arg = prev_insn_cc0_mode;
3682 const_arg = equiv_constant (folded_arg);
3683 break;
3684 #endif
3686 default:
3687 folded_arg = fold_rtx (arg, insn);
3688 const_arg = equiv_constant (folded_arg);
3691 /* For the first three operands, see if the operand
3692 is constant or equivalent to a constant. */
3693 switch (i)
3695 case 0:
3696 folded_arg0 = folded_arg;
3697 const_arg0 = const_arg;
3698 mode_arg0 = mode_arg;
3699 break;
3700 case 1:
3701 folded_arg1 = folded_arg;
3702 const_arg1 = const_arg;
3703 break;
3704 case 2:
3705 const_arg2 = const_arg;
3706 break;
3709 /* Pick the least expensive of the folded argument and an
3710 equivalent constant argument. */
3711 if (const_arg == 0 || const_arg == folded_arg
3712 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3713 cheap_arg = folded_arg, expensive_arg = const_arg;
3714 else
3715 cheap_arg = const_arg, expensive_arg = folded_arg;
3717 /* Try to replace the operand with the cheapest of the two
3718 possibilities. If it doesn't work and this is either of the first
3719 two operands of a commutative operation, try swapping them.
3720 If THAT fails, try the more expensive, provided it is cheaper
3721 than what is already there. */
3723 if (cheap_arg == XEXP (x, i))
3724 continue;
3726 if (insn == 0 && ! copied)
3728 x = copy_rtx (x);
3729 copied = 1;
3732 /* Order the replacements from cheapest to most expensive. */
3733 replacements[0] = cheap_arg;
3734 replacements[1] = expensive_arg;
3736 for (j = 0; j < 2 && replacements[j]; j++)
3738 int new_cost = COST_IN (replacements[j], code);
3740 /* Stop if what existed before was cheaper. Prefer constants
3741 in the case of a tie. */
3742 if (new_cost > old_cost
3743 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3744 break;
3746 /* It's not safe to substitute the operand of a conversion
3747 operator with a constant, as the conversion's identity
3748 depends upon the mode of it's operand. This optimization
3749 is handled by the call to simplify_unary_operation. */
3750 if (GET_RTX_CLASS (code) == RTX_UNARY
3751 && GET_MODE (replacements[j]) != mode_arg0
3752 && (code == ZERO_EXTEND
3753 || code == SIGN_EXTEND
3754 || code == TRUNCATE
3755 || code == FLOAT_TRUNCATE
3756 || code == FLOAT_EXTEND
3757 || code == FLOAT
3758 || code == FIX
3759 || code == UNSIGNED_FLOAT
3760 || code == UNSIGNED_FIX))
3761 continue;
3763 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3764 break;
3766 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3767 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3769 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3770 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3772 if (apply_change_group ())
3774 /* Swap them back to be invalid so that this loop can
3775 continue and flag them to be swapped back later. */
3776 rtx tem;
3778 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3779 XEXP (x, 1) = tem;
3780 must_swap = 1;
3781 break;
3787 else
3789 if (fmt[i] == 'E')
3790 /* Don't try to fold inside of a vector of expressions.
3791 Doing nothing is harmless. */
3795 /* If a commutative operation, place a constant integer as the second
3796 operand unless the first operand is also a constant integer. Otherwise,
3797 place any constant second unless the first operand is also a constant. */
3799 if (COMMUTATIVE_P (x))
3801 if (must_swap
3802 || swap_commutative_operands_p (const_arg0 ? const_arg0
3803 : XEXP (x, 0),
3804 const_arg1 ? const_arg1
3805 : XEXP (x, 1)))
3807 rtx tem = XEXP (x, 0);
3809 if (insn == 0 && ! copied)
3811 x = copy_rtx (x);
3812 copied = 1;
3815 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3816 validate_change (insn, &XEXP (x, 1), tem, 1);
3817 if (apply_change_group ())
3819 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3820 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3825 /* If X is an arithmetic operation, see if we can simplify it. */
3827 switch (GET_RTX_CLASS (code))
3829 case RTX_UNARY:
3831 int is_const = 0;
3833 /* We can't simplify extension ops unless we know the
3834 original mode. */
3835 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3836 && mode_arg0 == VOIDmode)
3837 break;
3839 /* If we had a CONST, strip it off and put it back later if we
3840 fold. */
3841 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3842 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3844 new = simplify_unary_operation (code, mode,
3845 const_arg0 ? const_arg0 : folded_arg0,
3846 mode_arg0);
3847 /* NEG of PLUS could be converted into MINUS, but that causes
3848 expressions of the form
3849 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3850 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3851 FIXME: those ports should be fixed. */
3852 if (new != 0 && is_const
3853 && GET_CODE (new) == PLUS
3854 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3855 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3856 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3857 new = gen_rtx_CONST (mode, new);
3859 break;
3861 case RTX_COMPARE:
3862 case RTX_COMM_COMPARE:
3863 /* See what items are actually being compared and set FOLDED_ARG[01]
3864 to those values and CODE to the actual comparison code. If any are
3865 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3866 do anything if both operands are already known to be constant. */
3868 /* ??? Vector mode comparisons are not supported yet. */
3869 if (VECTOR_MODE_P (mode))
3870 break;
3872 if (const_arg0 == 0 || const_arg1 == 0)
3874 struct table_elt *p0, *p1;
3875 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3876 enum machine_mode mode_arg1;
3878 #ifdef FLOAT_STORE_FLAG_VALUE
3879 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3881 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3882 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3883 false_rtx = CONST0_RTX (mode);
3885 #endif
3887 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3888 &mode_arg0, &mode_arg1);
3889 const_arg0 = equiv_constant (folded_arg0);
3890 const_arg1 = equiv_constant (folded_arg1);
3892 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3893 what kinds of things are being compared, so we can't do
3894 anything with this comparison. */
3896 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3897 break;
3899 /* If we do not now have two constants being compared, see
3900 if we can nevertheless deduce some things about the
3901 comparison. */
3902 if (const_arg0 == 0 || const_arg1 == 0)
3904 /* Some addresses are known to be nonzero. We don't know
3905 their sign, but equality comparisons are known. */
3906 if (const_arg1 == const0_rtx
3907 && nonzero_address_p (folded_arg0))
3909 if (code == EQ)
3910 return false_rtx;
3911 else if (code == NE)
3912 return true_rtx;
3915 /* See if the two operands are the same. */
3917 if (folded_arg0 == folded_arg1
3918 || (REG_P (folded_arg0)
3919 && REG_P (folded_arg1)
3920 && (REG_QTY (REGNO (folded_arg0))
3921 == REG_QTY (REGNO (folded_arg1))))
3922 || ((p0 = lookup (folded_arg0,
3923 SAFE_HASH (folded_arg0, mode_arg0),
3924 mode_arg0))
3925 && (p1 = lookup (folded_arg1,
3926 SAFE_HASH (folded_arg1, mode_arg0),
3927 mode_arg0))
3928 && p0->first_same_value == p1->first_same_value))
3930 /* Sadly two equal NaNs are not equivalent. */
3931 if (!HONOR_NANS (mode_arg0))
3932 return ((code == EQ || code == LE || code == GE
3933 || code == LEU || code == GEU || code == UNEQ
3934 || code == UNLE || code == UNGE
3935 || code == ORDERED)
3936 ? true_rtx : false_rtx);
3937 /* Take care for the FP compares we can resolve. */
3938 if (code == UNEQ || code == UNLE || code == UNGE)
3939 return true_rtx;
3940 if (code == LTGT || code == LT || code == GT)
3941 return false_rtx;
3944 /* If FOLDED_ARG0 is a register, see if the comparison we are
3945 doing now is either the same as we did before or the reverse
3946 (we only check the reverse if not floating-point). */
3947 else if (REG_P (folded_arg0))
3949 int qty = REG_QTY (REGNO (folded_arg0));
3951 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3953 struct qty_table_elem *ent = &qty_table[qty];
3955 if ((comparison_dominates_p (ent->comparison_code, code)
3956 || (! FLOAT_MODE_P (mode_arg0)
3957 && comparison_dominates_p (ent->comparison_code,
3958 reverse_condition (code))))
3959 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3960 || (const_arg1
3961 && rtx_equal_p (ent->comparison_const,
3962 const_arg1))
3963 || (REG_P (folded_arg1)
3964 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3965 return (comparison_dominates_p (ent->comparison_code, code)
3966 ? true_rtx : false_rtx);
3972 /* If we are comparing against zero, see if the first operand is
3973 equivalent to an IOR with a constant. If so, we may be able to
3974 determine the result of this comparison. */
3976 if (const_arg1 == const0_rtx)
3978 rtx y = lookup_as_function (folded_arg0, IOR);
3979 rtx inner_const;
3981 if (y != 0
3982 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3983 && GET_CODE (inner_const) == CONST_INT
3984 && INTVAL (inner_const) != 0)
3986 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
3987 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
3988 && (INTVAL (inner_const)
3989 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
3990 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3992 #ifdef FLOAT_STORE_FLAG_VALUE
3993 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3995 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3996 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3997 false_rtx = CONST0_RTX (mode);
3999 #endif
4001 switch (code)
4003 case EQ:
4004 return false_rtx;
4005 case NE:
4006 return true_rtx;
4007 case LT: case LE:
4008 if (has_sign)
4009 return true_rtx;
4010 break;
4011 case GT: case GE:
4012 if (has_sign)
4013 return false_rtx;
4014 break;
4015 default:
4016 break;
4022 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4023 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4024 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4026 break;
4028 case RTX_BIN_ARITH:
4029 case RTX_COMM_ARITH:
4030 switch (code)
4032 case PLUS:
4033 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4034 with that LABEL_REF as its second operand. If so, the result is
4035 the first operand of that MINUS. This handles switches with an
4036 ADDR_DIFF_VEC table. */
4037 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4039 rtx y
4040 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4041 : lookup_as_function (folded_arg0, MINUS);
4043 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4044 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4045 return XEXP (y, 0);
4047 /* Now try for a CONST of a MINUS like the above. */
4048 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4049 : lookup_as_function (folded_arg0, CONST))) != 0
4050 && GET_CODE (XEXP (y, 0)) == MINUS
4051 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4052 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4053 return XEXP (XEXP (y, 0), 0);
4056 /* Likewise if the operands are in the other order. */
4057 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4059 rtx y
4060 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4061 : lookup_as_function (folded_arg1, MINUS);
4063 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4064 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4065 return XEXP (y, 0);
4067 /* Now try for a CONST of a MINUS like the above. */
4068 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4069 : lookup_as_function (folded_arg1, CONST))) != 0
4070 && GET_CODE (XEXP (y, 0)) == MINUS
4071 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4072 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4073 return XEXP (XEXP (y, 0), 0);
4076 /* If second operand is a register equivalent to a negative
4077 CONST_INT, see if we can find a register equivalent to the
4078 positive constant. Make a MINUS if so. Don't do this for
4079 a non-negative constant since we might then alternate between
4080 choosing positive and negative constants. Having the positive
4081 constant previously-used is the more common case. Be sure
4082 the resulting constant is non-negative; if const_arg1 were
4083 the smallest negative number this would overflow: depending
4084 on the mode, this would either just be the same value (and
4085 hence not save anything) or be incorrect. */
4086 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4087 && INTVAL (const_arg1) < 0
4088 /* This used to test
4090 -INTVAL (const_arg1) >= 0
4092 But The Sun V5.0 compilers mis-compiled that test. So
4093 instead we test for the problematic value in a more direct
4094 manner and hope the Sun compilers get it correct. */
4095 && INTVAL (const_arg1) !=
4096 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4097 && REG_P (folded_arg1))
4099 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4100 struct table_elt *p
4101 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4103 if (p)
4104 for (p = p->first_same_value; p; p = p->next_same_value)
4105 if (REG_P (p->exp))
4106 return simplify_gen_binary (MINUS, mode, folded_arg0,
4107 canon_reg (p->exp, NULL_RTX));
4109 goto from_plus;
4111 case MINUS:
4112 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4113 If so, produce (PLUS Z C2-C). */
4114 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4116 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4117 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4118 return fold_rtx (plus_constant (copy_rtx (y),
4119 -INTVAL (const_arg1)),
4120 NULL_RTX);
4123 /* Fall through. */
4125 from_plus:
4126 case SMIN: case SMAX: case UMIN: case UMAX:
4127 case IOR: case AND: case XOR:
4128 case MULT:
4129 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4130 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4131 is known to be of similar form, we may be able to replace the
4132 operation with a combined operation. This may eliminate the
4133 intermediate operation if every use is simplified in this way.
4134 Note that the similar optimization done by combine.c only works
4135 if the intermediate operation's result has only one reference. */
4137 if (REG_P (folded_arg0)
4138 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4140 int is_shift
4141 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4142 rtx y = lookup_as_function (folded_arg0, code);
4143 rtx inner_const;
4144 enum rtx_code associate_code;
4145 rtx new_const;
4147 if (y == 0
4148 || 0 == (inner_const
4149 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4150 || GET_CODE (inner_const) != CONST_INT
4151 /* If we have compiled a statement like
4152 "if (x == (x & mask1))", and now are looking at
4153 "x & mask2", we will have a case where the first operand
4154 of Y is the same as our first operand. Unless we detect
4155 this case, an infinite loop will result. */
4156 || XEXP (y, 0) == folded_arg0)
4157 break;
4159 /* Don't associate these operations if they are a PLUS with the
4160 same constant and it is a power of two. These might be doable
4161 with a pre- or post-increment. Similarly for two subtracts of
4162 identical powers of two with post decrement. */
4164 if (code == PLUS && const_arg1 == inner_const
4165 && ((HAVE_PRE_INCREMENT
4166 && exact_log2 (INTVAL (const_arg1)) >= 0)
4167 || (HAVE_POST_INCREMENT
4168 && exact_log2 (INTVAL (const_arg1)) >= 0)
4169 || (HAVE_PRE_DECREMENT
4170 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4171 || (HAVE_POST_DECREMENT
4172 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4173 break;
4175 /* Compute the code used to compose the constants. For example,
4176 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4178 associate_code = (is_shift || code == MINUS ? PLUS : code);
4180 new_const = simplify_binary_operation (associate_code, mode,
4181 const_arg1, inner_const);
4183 if (new_const == 0)
4184 break;
4186 /* If we are associating shift operations, don't let this
4187 produce a shift of the size of the object or larger.
4188 This could occur when we follow a sign-extend by a right
4189 shift on a machine that does a sign-extend as a pair
4190 of shifts. */
4192 if (is_shift && GET_CODE (new_const) == CONST_INT
4193 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4195 /* As an exception, we can turn an ASHIFTRT of this
4196 form into a shift of the number of bits - 1. */
4197 if (code == ASHIFTRT)
4198 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4199 else
4200 break;
4203 y = copy_rtx (XEXP (y, 0));
4205 /* If Y contains our first operand (the most common way this
4206 can happen is if Y is a MEM), we would do into an infinite
4207 loop if we tried to fold it. So don't in that case. */
4209 if (! reg_mentioned_p (folded_arg0, y))
4210 y = fold_rtx (y, insn);
4212 return simplify_gen_binary (code, mode, y, new_const);
4214 break;
4216 case DIV: case UDIV:
4217 /* ??? The associative optimization performed immediately above is
4218 also possible for DIV and UDIV using associate_code of MULT.
4219 However, we would need extra code to verify that the
4220 multiplication does not overflow, that is, there is no overflow
4221 in the calculation of new_const. */
4222 break;
4224 default:
4225 break;
4228 new = simplify_binary_operation (code, mode,
4229 const_arg0 ? const_arg0 : folded_arg0,
4230 const_arg1 ? const_arg1 : folded_arg1);
4231 break;
4233 case RTX_OBJ:
4234 /* (lo_sum (high X) X) is simply X. */
4235 if (code == LO_SUM && const_arg0 != 0
4236 && GET_CODE (const_arg0) == HIGH
4237 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4238 return const_arg1;
4239 break;
4241 case RTX_TERNARY:
4242 case RTX_BITFIELD_OPS:
4243 new = simplify_ternary_operation (code, mode, mode_arg0,
4244 const_arg0 ? const_arg0 : folded_arg0,
4245 const_arg1 ? const_arg1 : folded_arg1,
4246 const_arg2 ? const_arg2 : XEXP (x, 2));
4247 break;
4249 default:
4250 break;
4253 return new ? new : x;
4256 /* Return a constant value currently equivalent to X.
4257 Return 0 if we don't know one. */
4259 static rtx
4260 equiv_constant (rtx x)
4262 if (REG_P (x)
4263 && REGNO_QTY_VALID_P (REGNO (x)))
4265 int x_q = REG_QTY (REGNO (x));
4266 struct qty_table_elem *x_ent = &qty_table[x_q];
4268 if (x_ent->const_rtx)
4269 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4272 if (x == 0 || CONSTANT_P (x))
4273 return x;
4275 /* If X is a MEM, try to fold it outside the context of any insn to see if
4276 it might be equivalent to a constant. That handles the case where it
4277 is a constant-pool reference. Then try to look it up in the hash table
4278 in case it is something whose value we have seen before. */
4280 if (MEM_P (x))
4282 struct table_elt *elt;
4284 x = fold_rtx (x, NULL_RTX);
4285 if (CONSTANT_P (x))
4286 return x;
4288 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4289 if (elt == 0)
4290 return 0;
4292 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4293 if (elt->is_const && CONSTANT_P (elt->exp))
4294 return elt->exp;
4297 return 0;
4300 /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point
4301 number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
4302 least-significant part of X.
4303 MODE specifies how big a part of X to return.
4305 If the requested operation cannot be done, 0 is returned.
4307 This is similar to gen_lowpart_general in emit-rtl.c. */
4310 gen_lowpart_if_possible (enum machine_mode mode, rtx x)
4312 rtx result = gen_lowpart_common (mode, x);
4314 if (result)
4315 return result;
4316 else if (MEM_P (x))
4318 /* This is the only other case we handle. */
4319 int offset = 0;
4320 rtx new;
4322 if (WORDS_BIG_ENDIAN)
4323 offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD)
4324 - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD));
4325 if (BYTES_BIG_ENDIAN)
4326 /* Adjust the address so that the address-after-the-data is
4327 unchanged. */
4328 offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode))
4329 - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x))));
4331 new = adjust_address_nv (x, mode, offset);
4332 if (! memory_address_p (mode, XEXP (new, 0)))
4333 return 0;
4335 return new;
4337 else
4338 return 0;
4341 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4342 branch. It will be zero if not.
4344 In certain cases, this can cause us to add an equivalence. For example,
4345 if we are following the taken case of
4346 if (i == 2)
4347 we can add the fact that `i' and '2' are now equivalent.
4349 In any case, we can record that this comparison was passed. If the same
4350 comparison is seen later, we will know its value. */
4352 static void
4353 record_jump_equiv (rtx insn, int taken)
4355 int cond_known_true;
4356 rtx op0, op1;
4357 rtx set;
4358 enum machine_mode mode, mode0, mode1;
4359 int reversed_nonequality = 0;
4360 enum rtx_code code;
4362 /* Ensure this is the right kind of insn. */
4363 if (! any_condjump_p (insn))
4364 return;
4365 set = pc_set (insn);
4367 /* See if this jump condition is known true or false. */
4368 if (taken)
4369 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4370 else
4371 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4373 /* Get the type of comparison being done and the operands being compared.
4374 If we had to reverse a non-equality condition, record that fact so we
4375 know that it isn't valid for floating-point. */
4376 code = GET_CODE (XEXP (SET_SRC (set), 0));
4377 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4378 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4380 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4381 if (! cond_known_true)
4383 code = reversed_comparison_code_parts (code, op0, op1, insn);
4385 /* Don't remember if we can't find the inverse. */
4386 if (code == UNKNOWN)
4387 return;
4390 /* The mode is the mode of the non-constant. */
4391 mode = mode0;
4392 if (mode1 != VOIDmode)
4393 mode = mode1;
4395 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4398 /* Yet another form of subreg creation. In this case, we want something in
4399 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4401 static rtx
4402 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4404 enum machine_mode op_mode = GET_MODE (op);
4405 if (op_mode == mode || op_mode == VOIDmode)
4406 return op;
4407 return lowpart_subreg (mode, op, op_mode);
4410 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4411 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4412 Make any useful entries we can with that information. Called from
4413 above function and called recursively. */
4415 static void
4416 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4417 rtx op1, int reversed_nonequality)
4419 unsigned op0_hash, op1_hash;
4420 int op0_in_memory, op1_in_memory;
4421 struct table_elt *op0_elt, *op1_elt;
4423 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4424 we know that they are also equal in the smaller mode (this is also
4425 true for all smaller modes whether or not there is a SUBREG, but
4426 is not worth testing for with no SUBREG). */
4428 /* Note that GET_MODE (op0) may not equal MODE. */
4429 if (code == EQ && GET_CODE (op0) == SUBREG
4430 && (GET_MODE_SIZE (GET_MODE (op0))
4431 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4433 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4434 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4435 if (tem)
4436 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4437 reversed_nonequality);
4440 if (code == EQ && GET_CODE (op1) == SUBREG
4441 && (GET_MODE_SIZE (GET_MODE (op1))
4442 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4444 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4445 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4446 if (tem)
4447 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4448 reversed_nonequality);
4451 /* Similarly, if this is an NE comparison, and either is a SUBREG
4452 making a smaller mode, we know the whole thing is also NE. */
4454 /* Note that GET_MODE (op0) may not equal MODE;
4455 if we test MODE instead, we can get an infinite recursion
4456 alternating between two modes each wider than MODE. */
4458 if (code == NE && GET_CODE (op0) == SUBREG
4459 && subreg_lowpart_p (op0)
4460 && (GET_MODE_SIZE (GET_MODE (op0))
4461 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4463 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4464 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4465 if (tem)
4466 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4467 reversed_nonequality);
4470 if (code == NE && GET_CODE (op1) == SUBREG
4471 && subreg_lowpart_p (op1)
4472 && (GET_MODE_SIZE (GET_MODE (op1))
4473 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4475 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4476 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4477 if (tem)
4478 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4479 reversed_nonequality);
4482 /* Hash both operands. */
4484 do_not_record = 0;
4485 hash_arg_in_memory = 0;
4486 op0_hash = HASH (op0, mode);
4487 op0_in_memory = hash_arg_in_memory;
4489 if (do_not_record)
4490 return;
4492 do_not_record = 0;
4493 hash_arg_in_memory = 0;
4494 op1_hash = HASH (op1, mode);
4495 op1_in_memory = hash_arg_in_memory;
4497 if (do_not_record)
4498 return;
4500 /* Look up both operands. */
4501 op0_elt = lookup (op0, op0_hash, mode);
4502 op1_elt = lookup (op1, op1_hash, mode);
4504 /* If both operands are already equivalent or if they are not in the
4505 table but are identical, do nothing. */
4506 if ((op0_elt != 0 && op1_elt != 0
4507 && op0_elt->first_same_value == op1_elt->first_same_value)
4508 || op0 == op1 || rtx_equal_p (op0, op1))
4509 return;
4511 /* If we aren't setting two things equal all we can do is save this
4512 comparison. Similarly if this is floating-point. In the latter
4513 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4514 If we record the equality, we might inadvertently delete code
4515 whose intent was to change -0 to +0. */
4517 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4519 struct qty_table_elem *ent;
4520 int qty;
4522 /* If we reversed a floating-point comparison, if OP0 is not a
4523 register, or if OP1 is neither a register or constant, we can't
4524 do anything. */
4526 if (!REG_P (op1))
4527 op1 = equiv_constant (op1);
4529 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4530 || !REG_P (op0) || op1 == 0)
4531 return;
4533 /* Put OP0 in the hash table if it isn't already. This gives it a
4534 new quantity number. */
4535 if (op0_elt == 0)
4537 if (insert_regs (op0, NULL, 0))
4539 rehash_using_reg (op0);
4540 op0_hash = HASH (op0, mode);
4542 /* If OP0 is contained in OP1, this changes its hash code
4543 as well. Faster to rehash than to check, except
4544 for the simple case of a constant. */
4545 if (! CONSTANT_P (op1))
4546 op1_hash = HASH (op1,mode);
4549 op0_elt = insert (op0, NULL, op0_hash, mode);
4550 op0_elt->in_memory = op0_in_memory;
4553 qty = REG_QTY (REGNO (op0));
4554 ent = &qty_table[qty];
4556 ent->comparison_code = code;
4557 if (REG_P (op1))
4559 /* Look it up again--in case op0 and op1 are the same. */
4560 op1_elt = lookup (op1, op1_hash, mode);
4562 /* Put OP1 in the hash table so it gets a new quantity number. */
4563 if (op1_elt == 0)
4565 if (insert_regs (op1, NULL, 0))
4567 rehash_using_reg (op1);
4568 op1_hash = HASH (op1, mode);
4571 op1_elt = insert (op1, NULL, op1_hash, mode);
4572 op1_elt->in_memory = op1_in_memory;
4575 ent->comparison_const = NULL_RTX;
4576 ent->comparison_qty = REG_QTY (REGNO (op1));
4578 else
4580 ent->comparison_const = op1;
4581 ent->comparison_qty = -1;
4584 return;
4587 /* If either side is still missing an equivalence, make it now,
4588 then merge the equivalences. */
4590 if (op0_elt == 0)
4592 if (insert_regs (op0, NULL, 0))
4594 rehash_using_reg (op0);
4595 op0_hash = HASH (op0, mode);
4598 op0_elt = insert (op0, NULL, op0_hash, mode);
4599 op0_elt->in_memory = op0_in_memory;
4602 if (op1_elt == 0)
4604 if (insert_regs (op1, NULL, 0))
4606 rehash_using_reg (op1);
4607 op1_hash = HASH (op1, mode);
4610 op1_elt = insert (op1, NULL, op1_hash, mode);
4611 op1_elt->in_memory = op1_in_memory;
4614 merge_equiv_classes (op0_elt, op1_elt);
4617 /* CSE processing for one instruction.
4618 First simplify sources and addresses of all assignments
4619 in the instruction, using previously-computed equivalents values.
4620 Then install the new sources and destinations in the table
4621 of available values.
4623 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4624 the insn. It means that INSN is inside libcall block. In this
4625 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4627 /* Data on one SET contained in the instruction. */
4629 struct set
4631 /* The SET rtx itself. */
4632 rtx rtl;
4633 /* The SET_SRC of the rtx (the original value, if it is changing). */
4634 rtx src;
4635 /* The hash-table element for the SET_SRC of the SET. */
4636 struct table_elt *src_elt;
4637 /* Hash value for the SET_SRC. */
4638 unsigned src_hash;
4639 /* Hash value for the SET_DEST. */
4640 unsigned dest_hash;
4641 /* The SET_DEST, with SUBREG, etc., stripped. */
4642 rtx inner_dest;
4643 /* Nonzero if the SET_SRC is in memory. */
4644 char src_in_memory;
4645 /* Nonzero if the SET_SRC contains something
4646 whose value cannot be predicted and understood. */
4647 char src_volatile;
4648 /* Original machine mode, in case it becomes a CONST_INT.
4649 The size of this field should match the size of the mode
4650 field of struct rtx_def (see rtl.h). */
4651 ENUM_BITFIELD(machine_mode) mode : 8;
4652 /* A constant equivalent for SET_SRC, if any. */
4653 rtx src_const;
4654 /* Original SET_SRC value used for libcall notes. */
4655 rtx orig_src;
4656 /* Hash value of constant equivalent for SET_SRC. */
4657 unsigned src_const_hash;
4658 /* Table entry for constant equivalent for SET_SRC, if any. */
4659 struct table_elt *src_const_elt;
4662 static void
4663 cse_insn (rtx insn, rtx libcall_insn)
4665 rtx x = PATTERN (insn);
4666 int i;
4667 rtx tem;
4668 int n_sets = 0;
4670 #ifdef HAVE_cc0
4671 /* Records what this insn does to set CC0. */
4672 rtx this_insn_cc0 = 0;
4673 enum machine_mode this_insn_cc0_mode = VOIDmode;
4674 #endif
4676 rtx src_eqv = 0;
4677 struct table_elt *src_eqv_elt = 0;
4678 int src_eqv_volatile = 0;
4679 int src_eqv_in_memory = 0;
4680 unsigned src_eqv_hash = 0;
4682 struct set *sets = (struct set *) 0;
4684 this_insn = insn;
4686 /* Find all the SETs and CLOBBERs in this instruction.
4687 Record all the SETs in the array `set' and count them.
4688 Also determine whether there is a CLOBBER that invalidates
4689 all memory references, or all references at varying addresses. */
4691 if (CALL_P (insn))
4693 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4695 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4696 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4697 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4701 if (GET_CODE (x) == SET)
4703 sets = alloca (sizeof (struct set));
4704 sets[0].rtl = x;
4706 /* Ignore SETs that are unconditional jumps.
4707 They never need cse processing, so this does not hurt.
4708 The reason is not efficiency but rather
4709 so that we can test at the end for instructions
4710 that have been simplified to unconditional jumps
4711 and not be misled by unchanged instructions
4712 that were unconditional jumps to begin with. */
4713 if (SET_DEST (x) == pc_rtx
4714 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4717 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4718 The hard function value register is used only once, to copy to
4719 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4720 Ensure we invalidate the destination register. On the 80386 no
4721 other code would invalidate it since it is a fixed_reg.
4722 We need not check the return of apply_change_group; see canon_reg. */
4724 else if (GET_CODE (SET_SRC (x)) == CALL)
4726 canon_reg (SET_SRC (x), insn);
4727 apply_change_group ();
4728 fold_rtx (SET_SRC (x), insn);
4729 invalidate (SET_DEST (x), VOIDmode);
4731 else
4732 n_sets = 1;
4734 else if (GET_CODE (x) == PARALLEL)
4736 int lim = XVECLEN (x, 0);
4738 sets = alloca (lim * sizeof (struct set));
4740 /* Find all regs explicitly clobbered in this insn,
4741 and ensure they are not replaced with any other regs
4742 elsewhere in this insn.
4743 When a reg that is clobbered is also used for input,
4744 we should presume that that is for a reason,
4745 and we should not substitute some other register
4746 which is not supposed to be clobbered.
4747 Therefore, this loop cannot be merged into the one below
4748 because a CALL may precede a CLOBBER and refer to the
4749 value clobbered. We must not let a canonicalization do
4750 anything in that case. */
4751 for (i = 0; i < lim; i++)
4753 rtx y = XVECEXP (x, 0, i);
4754 if (GET_CODE (y) == CLOBBER)
4756 rtx clobbered = XEXP (y, 0);
4758 if (REG_P (clobbered)
4759 || GET_CODE (clobbered) == SUBREG)
4760 invalidate (clobbered, VOIDmode);
4761 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4762 || GET_CODE (clobbered) == ZERO_EXTRACT)
4763 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4767 for (i = 0; i < lim; i++)
4769 rtx y = XVECEXP (x, 0, i);
4770 if (GET_CODE (y) == SET)
4772 /* As above, we ignore unconditional jumps and call-insns and
4773 ignore the result of apply_change_group. */
4774 if (GET_CODE (SET_SRC (y)) == CALL)
4776 canon_reg (SET_SRC (y), insn);
4777 apply_change_group ();
4778 fold_rtx (SET_SRC (y), insn);
4779 invalidate (SET_DEST (y), VOIDmode);
4781 else if (SET_DEST (y) == pc_rtx
4782 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4784 else
4785 sets[n_sets++].rtl = y;
4787 else if (GET_CODE (y) == CLOBBER)
4789 /* If we clobber memory, canon the address.
4790 This does nothing when a register is clobbered
4791 because we have already invalidated the reg. */
4792 if (MEM_P (XEXP (y, 0)))
4793 canon_reg (XEXP (y, 0), NULL_RTX);
4795 else if (GET_CODE (y) == USE
4796 && ! (REG_P (XEXP (y, 0))
4797 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4798 canon_reg (y, NULL_RTX);
4799 else if (GET_CODE (y) == CALL)
4801 /* The result of apply_change_group can be ignored; see
4802 canon_reg. */
4803 canon_reg (y, insn);
4804 apply_change_group ();
4805 fold_rtx (y, insn);
4809 else if (GET_CODE (x) == CLOBBER)
4811 if (MEM_P (XEXP (x, 0)))
4812 canon_reg (XEXP (x, 0), NULL_RTX);
4815 /* Canonicalize a USE of a pseudo register or memory location. */
4816 else if (GET_CODE (x) == USE
4817 && ! (REG_P (XEXP (x, 0))
4818 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4819 canon_reg (XEXP (x, 0), NULL_RTX);
4820 else if (GET_CODE (x) == CALL)
4822 /* The result of apply_change_group can be ignored; see canon_reg. */
4823 canon_reg (x, insn);
4824 apply_change_group ();
4825 fold_rtx (x, insn);
4828 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4829 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4830 is handled specially for this case, and if it isn't set, then there will
4831 be no equivalence for the destination. */
4832 if (n_sets == 1 && REG_NOTES (insn) != 0
4833 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4834 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4835 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4837 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4838 XEXP (tem, 0) = src_eqv;
4841 /* Canonicalize sources and addresses of destinations.
4842 We do this in a separate pass to avoid problems when a MATCH_DUP is
4843 present in the insn pattern. In that case, we want to ensure that
4844 we don't break the duplicate nature of the pattern. So we will replace
4845 both operands at the same time. Otherwise, we would fail to find an
4846 equivalent substitution in the loop calling validate_change below.
4848 We used to suppress canonicalization of DEST if it appears in SRC,
4849 but we don't do this any more. */
4851 for (i = 0; i < n_sets; i++)
4853 rtx dest = SET_DEST (sets[i].rtl);
4854 rtx src = SET_SRC (sets[i].rtl);
4855 rtx new = canon_reg (src, insn);
4856 int insn_code;
4858 sets[i].orig_src = src;
4859 if ((REG_P (new) && REG_P (src)
4860 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4861 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4862 || (insn_code = recog_memoized (insn)) < 0
4863 || insn_data[insn_code].n_dups > 0)
4864 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4865 else
4866 SET_SRC (sets[i].rtl) = new;
4868 if (GET_CODE (dest) == ZERO_EXTRACT)
4870 validate_change (insn, &XEXP (dest, 1),
4871 canon_reg (XEXP (dest, 1), insn), 1);
4872 validate_change (insn, &XEXP (dest, 2),
4873 canon_reg (XEXP (dest, 2), insn), 1);
4876 while (GET_CODE (dest) == SUBREG
4877 || GET_CODE (dest) == ZERO_EXTRACT
4878 || GET_CODE (dest) == STRICT_LOW_PART)
4879 dest = XEXP (dest, 0);
4881 if (MEM_P (dest))
4882 canon_reg (dest, insn);
4885 /* Now that we have done all the replacements, we can apply the change
4886 group and see if they all work. Note that this will cause some
4887 canonicalizations that would have worked individually not to be applied
4888 because some other canonicalization didn't work, but this should not
4889 occur often.
4891 The result of apply_change_group can be ignored; see canon_reg. */
4893 apply_change_group ();
4895 /* Set sets[i].src_elt to the class each source belongs to.
4896 Detect assignments from or to volatile things
4897 and set set[i] to zero so they will be ignored
4898 in the rest of this function.
4900 Nothing in this loop changes the hash table or the register chains. */
4902 for (i = 0; i < n_sets; i++)
4904 rtx src, dest;
4905 rtx src_folded;
4906 struct table_elt *elt = 0, *p;
4907 enum machine_mode mode;
4908 rtx src_eqv_here;
4909 rtx src_const = 0;
4910 rtx src_related = 0;
4911 struct table_elt *src_const_elt = 0;
4912 int src_cost = MAX_COST;
4913 int src_eqv_cost = MAX_COST;
4914 int src_folded_cost = MAX_COST;
4915 int src_related_cost = MAX_COST;
4916 int src_elt_cost = MAX_COST;
4917 int src_regcost = MAX_COST;
4918 int src_eqv_regcost = MAX_COST;
4919 int src_folded_regcost = MAX_COST;
4920 int src_related_regcost = MAX_COST;
4921 int src_elt_regcost = MAX_COST;
4922 /* Set nonzero if we need to call force_const_mem on with the
4923 contents of src_folded before using it. */
4924 int src_folded_force_flag = 0;
4926 dest = SET_DEST (sets[i].rtl);
4927 src = SET_SRC (sets[i].rtl);
4929 /* If SRC is a constant that has no machine mode,
4930 hash it with the destination's machine mode.
4931 This way we can keep different modes separate. */
4933 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4934 sets[i].mode = mode;
4936 if (src_eqv)
4938 enum machine_mode eqvmode = mode;
4939 if (GET_CODE (dest) == STRICT_LOW_PART)
4940 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4941 do_not_record = 0;
4942 hash_arg_in_memory = 0;
4943 src_eqv_hash = HASH (src_eqv, eqvmode);
4945 /* Find the equivalence class for the equivalent expression. */
4947 if (!do_not_record)
4948 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4950 src_eqv_volatile = do_not_record;
4951 src_eqv_in_memory = hash_arg_in_memory;
4954 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4955 value of the INNER register, not the destination. So it is not
4956 a valid substitution for the source. But save it for later. */
4957 if (GET_CODE (dest) == STRICT_LOW_PART)
4958 src_eqv_here = 0;
4959 else
4960 src_eqv_here = src_eqv;
4962 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4963 simplified result, which may not necessarily be valid. */
4964 src_folded = fold_rtx (src, insn);
4966 #if 0
4967 /* ??? This caused bad code to be generated for the m68k port with -O2.
4968 Suppose src is (CONST_INT -1), and that after truncation src_folded
4969 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4970 At the end we will add src and src_const to the same equivalence
4971 class. We now have 3 and -1 on the same equivalence class. This
4972 causes later instructions to be mis-optimized. */
4973 /* If storing a constant in a bitfield, pre-truncate the constant
4974 so we will be able to record it later. */
4975 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4977 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4979 if (GET_CODE (src) == CONST_INT
4980 && GET_CODE (width) == CONST_INT
4981 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4982 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4983 src_folded
4984 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4985 << INTVAL (width)) - 1));
4987 #endif
4989 /* Compute SRC's hash code, and also notice if it
4990 should not be recorded at all. In that case,
4991 prevent any further processing of this assignment. */
4992 do_not_record = 0;
4993 hash_arg_in_memory = 0;
4995 sets[i].src = src;
4996 sets[i].src_hash = HASH (src, mode);
4997 sets[i].src_volatile = do_not_record;
4998 sets[i].src_in_memory = hash_arg_in_memory;
5000 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5001 a pseudo, do not record SRC. Using SRC as a replacement for
5002 anything else will be incorrect in that situation. Note that
5003 this usually occurs only for stack slots, in which case all the
5004 RTL would be referring to SRC, so we don't lose any optimization
5005 opportunities by not having SRC in the hash table. */
5007 if (MEM_P (src)
5008 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5009 && REG_P (dest)
5010 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5011 sets[i].src_volatile = 1;
5013 #if 0
5014 /* It is no longer clear why we used to do this, but it doesn't
5015 appear to still be needed. So let's try without it since this
5016 code hurts cse'ing widened ops. */
5017 /* If source is a paradoxical subreg (such as QI treated as an SI),
5018 treat it as volatile. It may do the work of an SI in one context
5019 where the extra bits are not being used, but cannot replace an SI
5020 in general. */
5021 if (GET_CODE (src) == SUBREG
5022 && (GET_MODE_SIZE (GET_MODE (src))
5023 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5024 sets[i].src_volatile = 1;
5025 #endif
5027 /* Locate all possible equivalent forms for SRC. Try to replace
5028 SRC in the insn with each cheaper equivalent.
5030 We have the following types of equivalents: SRC itself, a folded
5031 version, a value given in a REG_EQUAL note, or a value related
5032 to a constant.
5034 Each of these equivalents may be part of an additional class
5035 of equivalents (if more than one is in the table, they must be in
5036 the same class; we check for this).
5038 If the source is volatile, we don't do any table lookups.
5040 We note any constant equivalent for possible later use in a
5041 REG_NOTE. */
5043 if (!sets[i].src_volatile)
5044 elt = lookup (src, sets[i].src_hash, mode);
5046 sets[i].src_elt = elt;
5048 if (elt && src_eqv_here && src_eqv_elt)
5050 if (elt->first_same_value != src_eqv_elt->first_same_value)
5052 /* The REG_EQUAL is indicating that two formerly distinct
5053 classes are now equivalent. So merge them. */
5054 merge_equiv_classes (elt, src_eqv_elt);
5055 src_eqv_hash = HASH (src_eqv, elt->mode);
5056 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5059 src_eqv_here = 0;
5062 else if (src_eqv_elt)
5063 elt = src_eqv_elt;
5065 /* Try to find a constant somewhere and record it in `src_const'.
5066 Record its table element, if any, in `src_const_elt'. Look in
5067 any known equivalences first. (If the constant is not in the
5068 table, also set `sets[i].src_const_hash'). */
5069 if (elt)
5070 for (p = elt->first_same_value; p; p = p->next_same_value)
5071 if (p->is_const)
5073 src_const = p->exp;
5074 src_const_elt = elt;
5075 break;
5078 if (src_const == 0
5079 && (CONSTANT_P (src_folded)
5080 /* Consider (minus (label_ref L1) (label_ref L2)) as
5081 "constant" here so we will record it. This allows us
5082 to fold switch statements when an ADDR_DIFF_VEC is used. */
5083 || (GET_CODE (src_folded) == MINUS
5084 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5085 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5086 src_const = src_folded, src_const_elt = elt;
5087 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5088 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5090 /* If we don't know if the constant is in the table, get its
5091 hash code and look it up. */
5092 if (src_const && src_const_elt == 0)
5094 sets[i].src_const_hash = HASH (src_const, mode);
5095 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5098 sets[i].src_const = src_const;
5099 sets[i].src_const_elt = src_const_elt;
5101 /* If the constant and our source are both in the table, mark them as
5102 equivalent. Otherwise, if a constant is in the table but the source
5103 isn't, set ELT to it. */
5104 if (src_const_elt && elt
5105 && src_const_elt->first_same_value != elt->first_same_value)
5106 merge_equiv_classes (elt, src_const_elt);
5107 else if (src_const_elt && elt == 0)
5108 elt = src_const_elt;
5110 /* See if there is a register linearly related to a constant
5111 equivalent of SRC. */
5112 if (src_const
5113 && (GET_CODE (src_const) == CONST
5114 || (src_const_elt && src_const_elt->related_value != 0)))
5116 src_related = use_related_value (src_const, src_const_elt);
5117 if (src_related)
5119 struct table_elt *src_related_elt
5120 = lookup (src_related, HASH (src_related, mode), mode);
5121 if (src_related_elt && elt)
5123 if (elt->first_same_value
5124 != src_related_elt->first_same_value)
5125 /* This can occur when we previously saw a CONST
5126 involving a SYMBOL_REF and then see the SYMBOL_REF
5127 twice. Merge the involved classes. */
5128 merge_equiv_classes (elt, src_related_elt);
5130 src_related = 0;
5131 src_related_elt = 0;
5133 else if (src_related_elt && elt == 0)
5134 elt = src_related_elt;
5138 /* See if we have a CONST_INT that is already in a register in a
5139 wider mode. */
5141 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5142 && GET_MODE_CLASS (mode) == MODE_INT
5143 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5145 enum machine_mode wider_mode;
5147 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5148 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5149 && src_related == 0;
5150 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5152 struct table_elt *const_elt
5153 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5155 if (const_elt == 0)
5156 continue;
5158 for (const_elt = const_elt->first_same_value;
5159 const_elt; const_elt = const_elt->next_same_value)
5160 if (REG_P (const_elt->exp))
5162 src_related = gen_lowpart (mode,
5163 const_elt->exp);
5164 break;
5169 /* Another possibility is that we have an AND with a constant in
5170 a mode narrower than a word. If so, it might have been generated
5171 as part of an "if" which would narrow the AND. If we already
5172 have done the AND in a wider mode, we can use a SUBREG of that
5173 value. */
5175 if (flag_expensive_optimizations && ! src_related
5176 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5177 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5179 enum machine_mode tmode;
5180 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5182 for (tmode = GET_MODE_WIDER_MODE (mode);
5183 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5184 tmode = GET_MODE_WIDER_MODE (tmode))
5186 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5187 struct table_elt *larger_elt;
5189 if (inner)
5191 PUT_MODE (new_and, tmode);
5192 XEXP (new_and, 0) = inner;
5193 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5194 if (larger_elt == 0)
5195 continue;
5197 for (larger_elt = larger_elt->first_same_value;
5198 larger_elt; larger_elt = larger_elt->next_same_value)
5199 if (REG_P (larger_elt->exp))
5201 src_related
5202 = gen_lowpart (mode, larger_elt->exp);
5203 break;
5206 if (src_related)
5207 break;
5212 #ifdef LOAD_EXTEND_OP
5213 /* See if a MEM has already been loaded with a widening operation;
5214 if it has, we can use a subreg of that. Many CISC machines
5215 also have such operations, but this is only likely to be
5216 beneficial on these machines. */
5218 if (flag_expensive_optimizations && src_related == 0
5219 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5220 && GET_MODE_CLASS (mode) == MODE_INT
5221 && MEM_P (src) && ! do_not_record
5222 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5224 struct rtx_def memory_extend_buf;
5225 rtx memory_extend_rtx = &memory_extend_buf;
5226 enum machine_mode tmode;
5228 /* Set what we are trying to extend and the operation it might
5229 have been extended with. */
5230 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5231 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5232 XEXP (memory_extend_rtx, 0) = src;
5234 for (tmode = GET_MODE_WIDER_MODE (mode);
5235 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5236 tmode = GET_MODE_WIDER_MODE (tmode))
5238 struct table_elt *larger_elt;
5240 PUT_MODE (memory_extend_rtx, tmode);
5241 larger_elt = lookup (memory_extend_rtx,
5242 HASH (memory_extend_rtx, tmode), tmode);
5243 if (larger_elt == 0)
5244 continue;
5246 for (larger_elt = larger_elt->first_same_value;
5247 larger_elt; larger_elt = larger_elt->next_same_value)
5248 if (REG_P (larger_elt->exp))
5250 src_related = gen_lowpart (mode,
5251 larger_elt->exp);
5252 break;
5255 if (src_related)
5256 break;
5259 #endif /* LOAD_EXTEND_OP */
5261 if (src == src_folded)
5262 src_folded = 0;
5264 /* At this point, ELT, if nonzero, points to a class of expressions
5265 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5266 and SRC_RELATED, if nonzero, each contain additional equivalent
5267 expressions. Prune these latter expressions by deleting expressions
5268 already in the equivalence class.
5270 Check for an equivalent identical to the destination. If found,
5271 this is the preferred equivalent since it will likely lead to
5272 elimination of the insn. Indicate this by placing it in
5273 `src_related'. */
5275 if (elt)
5276 elt = elt->first_same_value;
5277 for (p = elt; p; p = p->next_same_value)
5279 enum rtx_code code = GET_CODE (p->exp);
5281 /* If the expression is not valid, ignore it. Then we do not
5282 have to check for validity below. In most cases, we can use
5283 `rtx_equal_p', since canonicalization has already been done. */
5284 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5285 continue;
5287 /* Also skip paradoxical subregs, unless that's what we're
5288 looking for. */
5289 if (code == SUBREG
5290 && (GET_MODE_SIZE (GET_MODE (p->exp))
5291 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5292 && ! (src != 0
5293 && GET_CODE (src) == SUBREG
5294 && GET_MODE (src) == GET_MODE (p->exp)
5295 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5296 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5297 continue;
5299 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5300 src = 0;
5301 else if (src_folded && GET_CODE (src_folded) == code
5302 && rtx_equal_p (src_folded, p->exp))
5303 src_folded = 0;
5304 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5305 && rtx_equal_p (src_eqv_here, p->exp))
5306 src_eqv_here = 0;
5307 else if (src_related && GET_CODE (src_related) == code
5308 && rtx_equal_p (src_related, p->exp))
5309 src_related = 0;
5311 /* This is the same as the destination of the insns, we want
5312 to prefer it. Copy it to src_related. The code below will
5313 then give it a negative cost. */
5314 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5315 src_related = dest;
5318 /* Find the cheapest valid equivalent, trying all the available
5319 possibilities. Prefer items not in the hash table to ones
5320 that are when they are equal cost. Note that we can never
5321 worsen an insn as the current contents will also succeed.
5322 If we find an equivalent identical to the destination, use it as best,
5323 since this insn will probably be eliminated in that case. */
5324 if (src)
5326 if (rtx_equal_p (src, dest))
5327 src_cost = src_regcost = -1;
5328 else
5330 src_cost = COST (src);
5331 src_regcost = approx_reg_cost (src);
5335 if (src_eqv_here)
5337 if (rtx_equal_p (src_eqv_here, dest))
5338 src_eqv_cost = src_eqv_regcost = -1;
5339 else
5341 src_eqv_cost = COST (src_eqv_here);
5342 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5346 if (src_folded)
5348 if (rtx_equal_p (src_folded, dest))
5349 src_folded_cost = src_folded_regcost = -1;
5350 else
5352 src_folded_cost = COST (src_folded);
5353 src_folded_regcost = approx_reg_cost (src_folded);
5357 if (src_related)
5359 if (rtx_equal_p (src_related, dest))
5360 src_related_cost = src_related_regcost = -1;
5361 else
5363 src_related_cost = COST (src_related);
5364 src_related_regcost = approx_reg_cost (src_related);
5368 /* If this was an indirect jump insn, a known label will really be
5369 cheaper even though it looks more expensive. */
5370 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5371 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5373 /* Terminate loop when replacement made. This must terminate since
5374 the current contents will be tested and will always be valid. */
5375 while (1)
5377 rtx trial;
5379 /* Skip invalid entries. */
5380 while (elt && !REG_P (elt->exp)
5381 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5382 elt = elt->next_same_value;
5384 /* A paradoxical subreg would be bad here: it'll be the right
5385 size, but later may be adjusted so that the upper bits aren't
5386 what we want. So reject it. */
5387 if (elt != 0
5388 && GET_CODE (elt->exp) == SUBREG
5389 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5390 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5391 /* It is okay, though, if the rtx we're trying to match
5392 will ignore any of the bits we can't predict. */
5393 && ! (src != 0
5394 && GET_CODE (src) == SUBREG
5395 && GET_MODE (src) == GET_MODE (elt->exp)
5396 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5397 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5399 elt = elt->next_same_value;
5400 continue;
5403 if (elt)
5405 src_elt_cost = elt->cost;
5406 src_elt_regcost = elt->regcost;
5409 /* Find cheapest and skip it for the next time. For items
5410 of equal cost, use this order:
5411 src_folded, src, src_eqv, src_related and hash table entry. */
5412 if (src_folded
5413 && preferable (src_folded_cost, src_folded_regcost,
5414 src_cost, src_regcost) <= 0
5415 && preferable (src_folded_cost, src_folded_regcost,
5416 src_eqv_cost, src_eqv_regcost) <= 0
5417 && preferable (src_folded_cost, src_folded_regcost,
5418 src_related_cost, src_related_regcost) <= 0
5419 && preferable (src_folded_cost, src_folded_regcost,
5420 src_elt_cost, src_elt_regcost) <= 0)
5422 trial = src_folded, src_folded_cost = MAX_COST;
5423 if (src_folded_force_flag)
5425 rtx forced = force_const_mem (mode, trial);
5426 if (forced)
5427 trial = forced;
5430 else if (src
5431 && preferable (src_cost, src_regcost,
5432 src_eqv_cost, src_eqv_regcost) <= 0
5433 && preferable (src_cost, src_regcost,
5434 src_related_cost, src_related_regcost) <= 0
5435 && preferable (src_cost, src_regcost,
5436 src_elt_cost, src_elt_regcost) <= 0)
5437 trial = src, src_cost = MAX_COST;
5438 else if (src_eqv_here
5439 && preferable (src_eqv_cost, src_eqv_regcost,
5440 src_related_cost, src_related_regcost) <= 0
5441 && preferable (src_eqv_cost, src_eqv_regcost,
5442 src_elt_cost, src_elt_regcost) <= 0)
5443 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5444 else if (src_related
5445 && preferable (src_related_cost, src_related_regcost,
5446 src_elt_cost, src_elt_regcost) <= 0)
5447 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5448 else
5450 trial = copy_rtx (elt->exp);
5451 elt = elt->next_same_value;
5452 src_elt_cost = MAX_COST;
5455 /* We don't normally have an insn matching (set (pc) (pc)), so
5456 check for this separately here. We will delete such an
5457 insn below.
5459 For other cases such as a table jump or conditional jump
5460 where we know the ultimate target, go ahead and replace the
5461 operand. While that may not make a valid insn, we will
5462 reemit the jump below (and also insert any necessary
5463 barriers). */
5464 if (n_sets == 1 && dest == pc_rtx
5465 && (trial == pc_rtx
5466 || (GET_CODE (trial) == LABEL_REF
5467 && ! condjump_p (insn))))
5469 /* Don't substitute non-local labels, this confuses CFG. */
5470 if (GET_CODE (trial) == LABEL_REF
5471 && LABEL_REF_NONLOCAL_P (trial))
5472 continue;
5474 SET_SRC (sets[i].rtl) = trial;
5475 cse_jumps_altered = 1;
5476 break;
5479 /* Look for a substitution that makes a valid insn. */
5480 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5482 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5484 /* If we just made a substitution inside a libcall, then we
5485 need to make the same substitution in any notes attached
5486 to the RETVAL insn. */
5487 if (libcall_insn
5488 && (REG_P (sets[i].orig_src)
5489 || GET_CODE (sets[i].orig_src) == SUBREG
5490 || MEM_P (sets[i].orig_src)))
5492 rtx note = find_reg_equal_equiv_note (libcall_insn);
5493 if (note != 0)
5494 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5495 sets[i].orig_src,
5496 copy_rtx (new));
5499 /* The result of apply_change_group can be ignored; see
5500 canon_reg. */
5502 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5503 apply_change_group ();
5504 break;
5507 /* If we previously found constant pool entries for
5508 constants and this is a constant, try making a
5509 pool entry. Put it in src_folded unless we already have done
5510 this since that is where it likely came from. */
5512 else if (constant_pool_entries_cost
5513 && CONSTANT_P (trial)
5514 /* Reject cases that will abort in decode_rtx_const.
5515 On the alpha when simplifying a switch, we get
5516 (const (truncate (minus (label_ref) (label_ref)))). */
5517 && ! (GET_CODE (trial) == CONST
5518 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5519 /* Likewise on IA-64, except without the truncate. */
5520 && ! (GET_CODE (trial) == CONST
5521 && GET_CODE (XEXP (trial, 0)) == MINUS
5522 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5523 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5524 && (src_folded == 0
5525 || (!MEM_P (src_folded)
5526 && ! src_folded_force_flag))
5527 && GET_MODE_CLASS (mode) != MODE_CC
5528 && mode != VOIDmode)
5530 src_folded_force_flag = 1;
5531 src_folded = trial;
5532 src_folded_cost = constant_pool_entries_cost;
5533 src_folded_regcost = constant_pool_entries_regcost;
5537 src = SET_SRC (sets[i].rtl);
5539 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5540 However, there is an important exception: If both are registers
5541 that are not the head of their equivalence class, replace SET_SRC
5542 with the head of the class. If we do not do this, we will have
5543 both registers live over a portion of the basic block. This way,
5544 their lifetimes will likely abut instead of overlapping. */
5545 if (REG_P (dest)
5546 && REGNO_QTY_VALID_P (REGNO (dest)))
5548 int dest_q = REG_QTY (REGNO (dest));
5549 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5551 if (dest_ent->mode == GET_MODE (dest)
5552 && dest_ent->first_reg != REGNO (dest)
5553 && REG_P (src) && REGNO (src) == REGNO (dest)
5554 /* Don't do this if the original insn had a hard reg as
5555 SET_SRC or SET_DEST. */
5556 && (!REG_P (sets[i].src)
5557 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5558 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5559 /* We can't call canon_reg here because it won't do anything if
5560 SRC is a hard register. */
5562 int src_q = REG_QTY (REGNO (src));
5563 struct qty_table_elem *src_ent = &qty_table[src_q];
5564 int first = src_ent->first_reg;
5565 rtx new_src
5566 = (first >= FIRST_PSEUDO_REGISTER
5567 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5569 /* We must use validate-change even for this, because this
5570 might be a special no-op instruction, suitable only to
5571 tag notes onto. */
5572 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5574 src = new_src;
5575 /* If we had a constant that is cheaper than what we are now
5576 setting SRC to, use that constant. We ignored it when we
5577 thought we could make this into a no-op. */
5578 if (src_const && COST (src_const) < COST (src)
5579 && validate_change (insn, &SET_SRC (sets[i].rtl),
5580 src_const, 0))
5581 src = src_const;
5586 /* If we made a change, recompute SRC values. */
5587 if (src != sets[i].src)
5589 cse_altered = 1;
5590 do_not_record = 0;
5591 hash_arg_in_memory = 0;
5592 sets[i].src = src;
5593 sets[i].src_hash = HASH (src, mode);
5594 sets[i].src_volatile = do_not_record;
5595 sets[i].src_in_memory = hash_arg_in_memory;
5596 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5599 /* If this is a single SET, we are setting a register, and we have an
5600 equivalent constant, we want to add a REG_NOTE. We don't want
5601 to write a REG_EQUAL note for a constant pseudo since verifying that
5602 that pseudo hasn't been eliminated is a pain. Such a note also
5603 won't help anything.
5605 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5606 which can be created for a reference to a compile time computable
5607 entry in a jump table. */
5609 if (n_sets == 1 && src_const && REG_P (dest)
5610 && !REG_P (src_const)
5611 && ! (GET_CODE (src_const) == CONST
5612 && GET_CODE (XEXP (src_const, 0)) == MINUS
5613 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5614 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5616 /* We only want a REG_EQUAL note if src_const != src. */
5617 if (! rtx_equal_p (src, src_const))
5619 /* Make sure that the rtx is not shared. */
5620 src_const = copy_rtx (src_const);
5622 /* Record the actual constant value in a REG_EQUAL note,
5623 making a new one if one does not already exist. */
5624 set_unique_reg_note (insn, REG_EQUAL, src_const);
5628 /* Now deal with the destination. */
5629 do_not_record = 0;
5631 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5632 while (GET_CODE (dest) == SUBREG
5633 || GET_CODE (dest) == ZERO_EXTRACT
5634 || GET_CODE (dest) == STRICT_LOW_PART)
5635 dest = XEXP (dest, 0);
5637 sets[i].inner_dest = dest;
5639 if (MEM_P (dest))
5641 #ifdef PUSH_ROUNDING
5642 /* Stack pushes invalidate the stack pointer. */
5643 rtx addr = XEXP (dest, 0);
5644 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5645 && XEXP (addr, 0) == stack_pointer_rtx)
5646 invalidate (stack_pointer_rtx, Pmode);
5647 #endif
5648 dest = fold_rtx (dest, insn);
5651 /* Compute the hash code of the destination now,
5652 before the effects of this instruction are recorded,
5653 since the register values used in the address computation
5654 are those before this instruction. */
5655 sets[i].dest_hash = HASH (dest, mode);
5657 /* Don't enter a bit-field in the hash table
5658 because the value in it after the store
5659 may not equal what was stored, due to truncation. */
5661 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5663 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5665 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5666 && GET_CODE (width) == CONST_INT
5667 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5668 && ! (INTVAL (src_const)
5669 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5670 /* Exception: if the value is constant,
5671 and it won't be truncated, record it. */
5673 else
5675 /* This is chosen so that the destination will be invalidated
5676 but no new value will be recorded.
5677 We must invalidate because sometimes constant
5678 values can be recorded for bitfields. */
5679 sets[i].src_elt = 0;
5680 sets[i].src_volatile = 1;
5681 src_eqv = 0;
5682 src_eqv_elt = 0;
5686 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5687 the insn. */
5688 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5690 /* One less use of the label this insn used to jump to. */
5691 delete_insn (insn);
5692 cse_jumps_altered = 1;
5693 /* No more processing for this set. */
5694 sets[i].rtl = 0;
5697 /* If this SET is now setting PC to a label, we know it used to
5698 be a conditional or computed branch. */
5699 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5700 && !LABEL_REF_NONLOCAL_P (src))
5702 /* Now emit a BARRIER after the unconditional jump. */
5703 if (NEXT_INSN (insn) == 0
5704 || !BARRIER_P (NEXT_INSN (insn)))
5705 emit_barrier_after (insn);
5707 /* We reemit the jump in as many cases as possible just in
5708 case the form of an unconditional jump is significantly
5709 different than a computed jump or conditional jump.
5711 If this insn has multiple sets, then reemitting the
5712 jump is nontrivial. So instead we just force rerecognition
5713 and hope for the best. */
5714 if (n_sets == 1)
5716 rtx new, note;
5718 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5719 JUMP_LABEL (new) = XEXP (src, 0);
5720 LABEL_NUSES (XEXP (src, 0))++;
5722 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5723 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5724 if (note)
5726 XEXP (note, 1) = NULL_RTX;
5727 REG_NOTES (new) = note;
5730 delete_insn (insn);
5731 insn = new;
5733 /* Now emit a BARRIER after the unconditional jump. */
5734 if (NEXT_INSN (insn) == 0
5735 || !BARRIER_P (NEXT_INSN (insn)))
5736 emit_barrier_after (insn);
5738 else
5739 INSN_CODE (insn) = -1;
5741 /* Do not bother deleting any unreachable code,
5742 let jump/flow do that. */
5744 cse_jumps_altered = 1;
5745 sets[i].rtl = 0;
5748 /* If destination is volatile, invalidate it and then do no further
5749 processing for this assignment. */
5751 else if (do_not_record)
5753 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5754 invalidate (dest, VOIDmode);
5755 else if (MEM_P (dest))
5756 invalidate (dest, VOIDmode);
5757 else if (GET_CODE (dest) == STRICT_LOW_PART
5758 || GET_CODE (dest) == ZERO_EXTRACT)
5759 invalidate (XEXP (dest, 0), GET_MODE (dest));
5760 sets[i].rtl = 0;
5763 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5764 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5766 #ifdef HAVE_cc0
5767 /* If setting CC0, record what it was set to, or a constant, if it
5768 is equivalent to a constant. If it is being set to a floating-point
5769 value, make a COMPARE with the appropriate constant of 0. If we
5770 don't do this, later code can interpret this as a test against
5771 const0_rtx, which can cause problems if we try to put it into an
5772 insn as a floating-point operand. */
5773 if (dest == cc0_rtx)
5775 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5776 this_insn_cc0_mode = mode;
5777 if (FLOAT_MODE_P (mode))
5778 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5779 CONST0_RTX (mode));
5781 #endif
5784 /* Now enter all non-volatile source expressions in the hash table
5785 if they are not already present.
5786 Record their equivalence classes in src_elt.
5787 This way we can insert the corresponding destinations into
5788 the same classes even if the actual sources are no longer in them
5789 (having been invalidated). */
5791 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5792 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5794 struct table_elt *elt;
5795 struct table_elt *classp = sets[0].src_elt;
5796 rtx dest = SET_DEST (sets[0].rtl);
5797 enum machine_mode eqvmode = GET_MODE (dest);
5799 if (GET_CODE (dest) == STRICT_LOW_PART)
5801 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5802 classp = 0;
5804 if (insert_regs (src_eqv, classp, 0))
5806 rehash_using_reg (src_eqv);
5807 src_eqv_hash = HASH (src_eqv, eqvmode);
5809 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5810 elt->in_memory = src_eqv_in_memory;
5811 src_eqv_elt = elt;
5813 /* Check to see if src_eqv_elt is the same as a set source which
5814 does not yet have an elt, and if so set the elt of the set source
5815 to src_eqv_elt. */
5816 for (i = 0; i < n_sets; i++)
5817 if (sets[i].rtl && sets[i].src_elt == 0
5818 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5819 sets[i].src_elt = src_eqv_elt;
5822 for (i = 0; i < n_sets; i++)
5823 if (sets[i].rtl && ! sets[i].src_volatile
5824 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5826 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5828 /* REG_EQUAL in setting a STRICT_LOW_PART
5829 gives an equivalent for the entire destination register,
5830 not just for the subreg being stored in now.
5831 This is a more interesting equivalence, so we arrange later
5832 to treat the entire reg as the destination. */
5833 sets[i].src_elt = src_eqv_elt;
5834 sets[i].src_hash = src_eqv_hash;
5836 else
5838 /* Insert source and constant equivalent into hash table, if not
5839 already present. */
5840 struct table_elt *classp = src_eqv_elt;
5841 rtx src = sets[i].src;
5842 rtx dest = SET_DEST (sets[i].rtl);
5843 enum machine_mode mode
5844 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5846 /* It's possible that we have a source value known to be
5847 constant but don't have a REG_EQUAL note on the insn.
5848 Lack of a note will mean src_eqv_elt will be NULL. This
5849 can happen where we've generated a SUBREG to access a
5850 CONST_INT that is already in a register in a wider mode.
5851 Ensure that the source expression is put in the proper
5852 constant class. */
5853 if (!classp)
5854 classp = sets[i].src_const_elt;
5856 if (sets[i].src_elt == 0)
5858 /* Don't put a hard register source into the table if this is
5859 the last insn of a libcall. In this case, we only need
5860 to put src_eqv_elt in src_elt. */
5861 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5863 struct table_elt *elt;
5865 /* Note that these insert_regs calls cannot remove
5866 any of the src_elt's, because they would have failed to
5867 match if not still valid. */
5868 if (insert_regs (src, classp, 0))
5870 rehash_using_reg (src);
5871 sets[i].src_hash = HASH (src, mode);
5873 elt = insert (src, classp, sets[i].src_hash, mode);
5874 elt->in_memory = sets[i].src_in_memory;
5875 sets[i].src_elt = classp = elt;
5877 else
5878 sets[i].src_elt = classp;
5880 if (sets[i].src_const && sets[i].src_const_elt == 0
5881 && src != sets[i].src_const
5882 && ! rtx_equal_p (sets[i].src_const, src))
5883 sets[i].src_elt = insert (sets[i].src_const, classp,
5884 sets[i].src_const_hash, mode);
5887 else if (sets[i].src_elt == 0)
5888 /* If we did not insert the source into the hash table (e.g., it was
5889 volatile), note the equivalence class for the REG_EQUAL value, if any,
5890 so that the destination goes into that class. */
5891 sets[i].src_elt = src_eqv_elt;
5893 invalidate_from_clobbers (x);
5895 /* Some registers are invalidated by subroutine calls. Memory is
5896 invalidated by non-constant calls. */
5898 if (CALL_P (insn))
5900 if (! CONST_OR_PURE_CALL_P (insn))
5901 invalidate_memory ();
5902 invalidate_for_call ();
5905 /* Now invalidate everything set by this instruction.
5906 If a SUBREG or other funny destination is being set,
5907 sets[i].rtl is still nonzero, so here we invalidate the reg
5908 a part of which is being set. */
5910 for (i = 0; i < n_sets; i++)
5911 if (sets[i].rtl)
5913 /* We can't use the inner dest, because the mode associated with
5914 a ZERO_EXTRACT is significant. */
5915 rtx dest = SET_DEST (sets[i].rtl);
5917 /* Needed for registers to remove the register from its
5918 previous quantity's chain.
5919 Needed for memory if this is a nonvarying address, unless
5920 we have just done an invalidate_memory that covers even those. */
5921 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5922 invalidate (dest, VOIDmode);
5923 else if (MEM_P (dest))
5924 invalidate (dest, VOIDmode);
5925 else if (GET_CODE (dest) == STRICT_LOW_PART
5926 || GET_CODE (dest) == ZERO_EXTRACT)
5927 invalidate (XEXP (dest, 0), GET_MODE (dest));
5930 /* A volatile ASM invalidates everything. */
5931 if (NONJUMP_INSN_P (insn)
5932 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5933 && MEM_VOLATILE_P (PATTERN (insn)))
5934 flush_hash_table ();
5936 /* Make sure registers mentioned in destinations
5937 are safe for use in an expression to be inserted.
5938 This removes from the hash table
5939 any invalid entry that refers to one of these registers.
5941 We don't care about the return value from mention_regs because
5942 we are going to hash the SET_DEST values unconditionally. */
5944 for (i = 0; i < n_sets; i++)
5946 if (sets[i].rtl)
5948 rtx x = SET_DEST (sets[i].rtl);
5950 if (!REG_P (x))
5951 mention_regs (x);
5952 else
5954 /* We used to rely on all references to a register becoming
5955 inaccessible when a register changes to a new quantity,
5956 since that changes the hash code. However, that is not
5957 safe, since after HASH_SIZE new quantities we get a
5958 hash 'collision' of a register with its own invalid
5959 entries. And since SUBREGs have been changed not to
5960 change their hash code with the hash code of the register,
5961 it wouldn't work any longer at all. So we have to check
5962 for any invalid references lying around now.
5963 This code is similar to the REG case in mention_regs,
5964 but it knows that reg_tick has been incremented, and
5965 it leaves reg_in_table as -1 . */
5966 unsigned int regno = REGNO (x);
5967 unsigned int endregno
5968 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
5969 : hard_regno_nregs[regno][GET_MODE (x)]);
5970 unsigned int i;
5972 for (i = regno; i < endregno; i++)
5974 if (REG_IN_TABLE (i) >= 0)
5976 remove_invalid_refs (i);
5977 REG_IN_TABLE (i) = -1;
5984 /* We may have just removed some of the src_elt's from the hash table.
5985 So replace each one with the current head of the same class. */
5987 for (i = 0; i < n_sets; i++)
5988 if (sets[i].rtl)
5990 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5991 /* If elt was removed, find current head of same class,
5992 or 0 if nothing remains of that class. */
5994 struct table_elt *elt = sets[i].src_elt;
5996 while (elt && elt->prev_same_value)
5997 elt = elt->prev_same_value;
5999 while (elt && elt->first_same_value == 0)
6000 elt = elt->next_same_value;
6001 sets[i].src_elt = elt ? elt->first_same_value : 0;
6005 /* Now insert the destinations into their equivalence classes. */
6007 for (i = 0; i < n_sets; i++)
6008 if (sets[i].rtl)
6010 rtx dest = SET_DEST (sets[i].rtl);
6011 struct table_elt *elt;
6013 /* Don't record value if we are not supposed to risk allocating
6014 floating-point values in registers that might be wider than
6015 memory. */
6016 if ((flag_float_store
6017 && MEM_P (dest)
6018 && FLOAT_MODE_P (GET_MODE (dest)))
6019 /* Don't record BLKmode values, because we don't know the
6020 size of it, and can't be sure that other BLKmode values
6021 have the same or smaller size. */
6022 || GET_MODE (dest) == BLKmode
6023 /* Don't record values of destinations set inside a libcall block
6024 since we might delete the libcall. Things should have been set
6025 up so we won't want to reuse such a value, but we play it safe
6026 here. */
6027 || libcall_insn
6028 /* If we didn't put a REG_EQUAL value or a source into the hash
6029 table, there is no point is recording DEST. */
6030 || sets[i].src_elt == 0
6031 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6032 or SIGN_EXTEND, don't record DEST since it can cause
6033 some tracking to be wrong.
6035 ??? Think about this more later. */
6036 || (GET_CODE (dest) == SUBREG
6037 && (GET_MODE_SIZE (GET_MODE (dest))
6038 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6039 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6040 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6041 continue;
6043 /* STRICT_LOW_PART isn't part of the value BEING set,
6044 and neither is the SUBREG inside it.
6045 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6046 if (GET_CODE (dest) == STRICT_LOW_PART)
6047 dest = SUBREG_REG (XEXP (dest, 0));
6049 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6050 /* Registers must also be inserted into chains for quantities. */
6051 if (insert_regs (dest, sets[i].src_elt, 1))
6053 /* If `insert_regs' changes something, the hash code must be
6054 recalculated. */
6055 rehash_using_reg (dest);
6056 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6059 elt = insert (dest, sets[i].src_elt,
6060 sets[i].dest_hash, GET_MODE (dest));
6062 elt->in_memory = (MEM_P (sets[i].inner_dest)
6063 && !MEM_READONLY_P (sets[i].inner_dest));
6065 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6066 narrower than M2, and both M1 and M2 are the same number of words,
6067 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6068 make that equivalence as well.
6070 However, BAR may have equivalences for which gen_lowpart
6071 will produce a simpler value than gen_lowpart applied to
6072 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6073 BAR's equivalences. If we don't get a simplified form, make
6074 the SUBREG. It will not be used in an equivalence, but will
6075 cause two similar assignments to be detected.
6077 Note the loop below will find SUBREG_REG (DEST) since we have
6078 already entered SRC and DEST of the SET in the table. */
6080 if (GET_CODE (dest) == SUBREG
6081 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6082 / UNITS_PER_WORD)
6083 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6084 && (GET_MODE_SIZE (GET_MODE (dest))
6085 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6086 && sets[i].src_elt != 0)
6088 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6089 struct table_elt *elt, *classp = 0;
6091 for (elt = sets[i].src_elt->first_same_value; elt;
6092 elt = elt->next_same_value)
6094 rtx new_src = 0;
6095 unsigned src_hash;
6096 struct table_elt *src_elt;
6097 int byte = 0;
6099 /* Ignore invalid entries. */
6100 if (!REG_P (elt->exp)
6101 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6102 continue;
6104 /* We may have already been playing subreg games. If the
6105 mode is already correct for the destination, use it. */
6106 if (GET_MODE (elt->exp) == new_mode)
6107 new_src = elt->exp;
6108 else
6110 /* Calculate big endian correction for the SUBREG_BYTE.
6111 We have already checked that M1 (GET_MODE (dest))
6112 is not narrower than M2 (new_mode). */
6113 if (BYTES_BIG_ENDIAN)
6114 byte = (GET_MODE_SIZE (GET_MODE (dest))
6115 - GET_MODE_SIZE (new_mode));
6117 new_src = simplify_gen_subreg (new_mode, elt->exp,
6118 GET_MODE (dest), byte);
6121 /* The call to simplify_gen_subreg fails if the value
6122 is VOIDmode, yet we can't do any simplification, e.g.
6123 for EXPR_LISTs denoting function call results.
6124 It is invalid to construct a SUBREG with a VOIDmode
6125 SUBREG_REG, hence a zero new_src means we can't do
6126 this substitution. */
6127 if (! new_src)
6128 continue;
6130 src_hash = HASH (new_src, new_mode);
6131 src_elt = lookup (new_src, src_hash, new_mode);
6133 /* Put the new source in the hash table is if isn't
6134 already. */
6135 if (src_elt == 0)
6137 if (insert_regs (new_src, classp, 0))
6139 rehash_using_reg (new_src);
6140 src_hash = HASH (new_src, new_mode);
6142 src_elt = insert (new_src, classp, src_hash, new_mode);
6143 src_elt->in_memory = elt->in_memory;
6145 else if (classp && classp != src_elt->first_same_value)
6146 /* Show that two things that we've seen before are
6147 actually the same. */
6148 merge_equiv_classes (src_elt, classp);
6150 classp = src_elt->first_same_value;
6151 /* Ignore invalid entries. */
6152 while (classp
6153 && !REG_P (classp->exp)
6154 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6155 classp = classp->next_same_value;
6160 /* Special handling for (set REG0 REG1) where REG0 is the
6161 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6162 be used in the sequel, so (if easily done) change this insn to
6163 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6164 that computed their value. Then REG1 will become a dead store
6165 and won't cloud the situation for later optimizations.
6167 Do not make this change if REG1 is a hard register, because it will
6168 then be used in the sequel and we may be changing a two-operand insn
6169 into a three-operand insn.
6171 Also do not do this if we are operating on a copy of INSN.
6173 Also don't do this if INSN ends a libcall; this would cause an unrelated
6174 register to be set in the middle of a libcall, and we then get bad code
6175 if the libcall is deleted. */
6177 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6178 && NEXT_INSN (PREV_INSN (insn)) == insn
6179 && REG_P (SET_SRC (sets[0].rtl))
6180 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6181 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6183 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6184 struct qty_table_elem *src_ent = &qty_table[src_q];
6186 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6187 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6189 rtx prev = insn;
6190 /* Scan for the previous nonnote insn, but stop at a basic
6191 block boundary. */
6194 prev = PREV_INSN (prev);
6196 while (prev && NOTE_P (prev)
6197 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6199 /* Do not swap the registers around if the previous instruction
6200 attaches a REG_EQUIV note to REG1.
6202 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6203 from the pseudo that originally shadowed an incoming argument
6204 to another register. Some uses of REG_EQUIV might rely on it
6205 being attached to REG1 rather than REG2.
6207 This section previously turned the REG_EQUIV into a REG_EQUAL
6208 note. We cannot do that because REG_EQUIV may provide an
6209 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6211 if (prev != 0 && NONJUMP_INSN_P (prev)
6212 && GET_CODE (PATTERN (prev)) == SET
6213 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6214 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6216 rtx dest = SET_DEST (sets[0].rtl);
6217 rtx src = SET_SRC (sets[0].rtl);
6218 rtx note;
6220 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6221 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6222 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6223 apply_change_group ();
6225 /* If INSN has a REG_EQUAL note, and this note mentions
6226 REG0, then we must delete it, because the value in
6227 REG0 has changed. If the note's value is REG1, we must
6228 also delete it because that is now this insn's dest. */
6229 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6230 if (note != 0
6231 && (reg_mentioned_p (dest, XEXP (note, 0))
6232 || rtx_equal_p (src, XEXP (note, 0))))
6233 remove_note (insn, note);
6238 /* If this is a conditional jump insn, record any known equivalences due to
6239 the condition being tested. */
6241 if (JUMP_P (insn)
6242 && n_sets == 1 && GET_CODE (x) == SET
6243 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6244 record_jump_equiv (insn, 0);
6246 #ifdef HAVE_cc0
6247 /* If the previous insn set CC0 and this insn no longer references CC0,
6248 delete the previous insn. Here we use the fact that nothing expects CC0
6249 to be valid over an insn, which is true until the final pass. */
6250 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6251 && (tem = single_set (prev_insn)) != 0
6252 && SET_DEST (tem) == cc0_rtx
6253 && ! reg_mentioned_p (cc0_rtx, x))
6254 delete_insn (prev_insn);
6256 prev_insn_cc0 = this_insn_cc0;
6257 prev_insn_cc0_mode = this_insn_cc0_mode;
6258 prev_insn = insn;
6259 #endif
6262 /* Remove from the hash table all expressions that reference memory. */
6264 static void
6265 invalidate_memory (void)
6267 int i;
6268 struct table_elt *p, *next;
6270 for (i = 0; i < HASH_SIZE; i++)
6271 for (p = table[i]; p; p = next)
6273 next = p->next_same_hash;
6274 if (p->in_memory)
6275 remove_from_table (p, i);
6279 /* If ADDR is an address that implicitly affects the stack pointer, return
6280 1 and update the register tables to show the effect. Else, return 0. */
6282 static int
6283 addr_affects_sp_p (rtx addr)
6285 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6286 && REG_P (XEXP (addr, 0))
6287 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6289 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6291 REG_TICK (STACK_POINTER_REGNUM)++;
6292 /* Is it possible to use a subreg of SP? */
6293 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6296 /* This should be *very* rare. */
6297 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6298 invalidate (stack_pointer_rtx, VOIDmode);
6300 return 1;
6303 return 0;
6306 /* Perform invalidation on the basis of everything about an insn
6307 except for invalidating the actual places that are SET in it.
6308 This includes the places CLOBBERed, and anything that might
6309 alias with something that is SET or CLOBBERed.
6311 X is the pattern of the insn. */
6313 static void
6314 invalidate_from_clobbers (rtx x)
6316 if (GET_CODE (x) == CLOBBER)
6318 rtx ref = XEXP (x, 0);
6319 if (ref)
6321 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6322 || MEM_P (ref))
6323 invalidate (ref, VOIDmode);
6324 else if (GET_CODE (ref) == STRICT_LOW_PART
6325 || GET_CODE (ref) == ZERO_EXTRACT)
6326 invalidate (XEXP (ref, 0), GET_MODE (ref));
6329 else if (GET_CODE (x) == PARALLEL)
6331 int i;
6332 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6334 rtx y = XVECEXP (x, 0, i);
6335 if (GET_CODE (y) == CLOBBER)
6337 rtx ref = XEXP (y, 0);
6338 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6339 || MEM_P (ref))
6340 invalidate (ref, VOIDmode);
6341 else if (GET_CODE (ref) == STRICT_LOW_PART
6342 || GET_CODE (ref) == ZERO_EXTRACT)
6343 invalidate (XEXP (ref, 0), GET_MODE (ref));
6349 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6350 and replace any registers in them with either an equivalent constant
6351 or the canonical form of the register. If we are inside an address,
6352 only do this if the address remains valid.
6354 OBJECT is 0 except when within a MEM in which case it is the MEM.
6356 Return the replacement for X. */
6358 static rtx
6359 cse_process_notes (rtx x, rtx object)
6361 enum rtx_code code = GET_CODE (x);
6362 const char *fmt = GET_RTX_FORMAT (code);
6363 int i;
6365 switch (code)
6367 case CONST_INT:
6368 case CONST:
6369 case SYMBOL_REF:
6370 case LABEL_REF:
6371 case CONST_DOUBLE:
6372 case CONST_VECTOR:
6373 case PC:
6374 case CC0:
6375 case LO_SUM:
6376 return x;
6378 case MEM:
6379 validate_change (x, &XEXP (x, 0),
6380 cse_process_notes (XEXP (x, 0), x), 0);
6381 return x;
6383 case EXPR_LIST:
6384 case INSN_LIST:
6385 if (REG_NOTE_KIND (x) == REG_EQUAL)
6386 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6387 if (XEXP (x, 1))
6388 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6389 return x;
6391 case SIGN_EXTEND:
6392 case ZERO_EXTEND:
6393 case SUBREG:
6395 rtx new = cse_process_notes (XEXP (x, 0), object);
6396 /* We don't substitute VOIDmode constants into these rtx,
6397 since they would impede folding. */
6398 if (GET_MODE (new) != VOIDmode)
6399 validate_change (object, &XEXP (x, 0), new, 0);
6400 return x;
6403 case REG:
6404 i = REG_QTY (REGNO (x));
6406 /* Return a constant or a constant register. */
6407 if (REGNO_QTY_VALID_P (REGNO (x)))
6409 struct qty_table_elem *ent = &qty_table[i];
6411 if (ent->const_rtx != NULL_RTX
6412 && (CONSTANT_P (ent->const_rtx)
6413 || REG_P (ent->const_rtx)))
6415 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6416 if (new)
6417 return new;
6421 /* Otherwise, canonicalize this register. */
6422 return canon_reg (x, NULL_RTX);
6424 default:
6425 break;
6428 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6429 if (fmt[i] == 'e')
6430 validate_change (object, &XEXP (x, i),
6431 cse_process_notes (XEXP (x, i), object), 0);
6433 return x;
6436 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6437 since they are done elsewhere. This function is called via note_stores. */
6439 static void
6440 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6442 enum rtx_code code = GET_CODE (dest);
6444 if (code == MEM
6445 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6446 /* There are times when an address can appear varying and be a PLUS
6447 during this scan when it would be a fixed address were we to know
6448 the proper equivalences. So invalidate all memory if there is
6449 a BLKmode or nonscalar memory reference or a reference to a
6450 variable address. */
6451 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6452 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6454 invalidate_memory ();
6455 return;
6458 if (GET_CODE (set) == CLOBBER
6459 || CC0_P (dest)
6460 || dest == pc_rtx)
6461 return;
6463 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6464 invalidate (XEXP (dest, 0), GET_MODE (dest));
6465 else if (code == REG || code == SUBREG || code == MEM)
6466 invalidate (dest, VOIDmode);
6469 /* Invalidate all insns from START up to the end of the function or the
6470 next label. This called when we wish to CSE around a block that is
6471 conditionally executed. */
6473 static void
6474 invalidate_skipped_block (rtx start)
6476 rtx insn;
6478 for (insn = start; insn && !LABEL_P (insn);
6479 insn = NEXT_INSN (insn))
6481 if (! INSN_P (insn))
6482 continue;
6484 if (CALL_P (insn))
6486 if (! CONST_OR_PURE_CALL_P (insn))
6487 invalidate_memory ();
6488 invalidate_for_call ();
6491 invalidate_from_clobbers (PATTERN (insn));
6492 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6496 /* Find the end of INSN's basic block and return its range,
6497 the total number of SETs in all the insns of the block, the last insn of the
6498 block, and the branch path.
6500 The branch path indicates which branches should be followed. If a nonzero
6501 path size is specified, the block should be rescanned and a different set
6502 of branches will be taken. The branch path is only used if
6503 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6505 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6506 used to describe the block. It is filled in with the information about
6507 the current block. The incoming structure's branch path, if any, is used
6508 to construct the output branch path. */
6510 static void
6511 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6512 int follow_jumps, int skip_blocks)
6514 rtx p = insn, q;
6515 int nsets = 0;
6516 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6517 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6518 int path_size = data->path_size;
6519 int path_entry = 0;
6520 int i;
6522 /* Update the previous branch path, if any. If the last branch was
6523 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6524 If it was previously PATH_NOT_TAKEN,
6525 shorten the path by one and look at the previous branch. We know that
6526 at least one branch must have been taken if PATH_SIZE is nonzero. */
6527 while (path_size > 0)
6529 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6531 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6532 break;
6534 else
6535 path_size--;
6538 /* If the first instruction is marked with QImode, that means we've
6539 already processed this block. Our caller will look at DATA->LAST
6540 to figure out where to go next. We want to return the next block
6541 in the instruction stream, not some branched-to block somewhere
6542 else. We accomplish this by pretending our called forbid us to
6543 follow jumps, or skip blocks. */
6544 if (GET_MODE (insn) == QImode)
6545 follow_jumps = skip_blocks = 0;
6547 /* Scan to end of this basic block. */
6548 while (p && !LABEL_P (p))
6550 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6551 the regs restored by the longjmp come from
6552 a later time than the setjmp. */
6553 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6554 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6555 break;
6557 /* A PARALLEL can have lots of SETs in it,
6558 especially if it is really an ASM_OPERANDS. */
6559 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6560 nsets += XVECLEN (PATTERN (p), 0);
6561 else if (!NOTE_P (p))
6562 nsets += 1;
6564 /* Ignore insns made by CSE; they cannot affect the boundaries of
6565 the basic block. */
6567 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6568 high_cuid = INSN_CUID (p);
6569 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6570 low_cuid = INSN_CUID (p);
6572 /* See if this insn is in our branch path. If it is and we are to
6573 take it, do so. */
6574 if (path_entry < path_size && data->path[path_entry].branch == p)
6576 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6577 p = JUMP_LABEL (p);
6579 /* Point to next entry in path, if any. */
6580 path_entry++;
6583 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6584 was specified, we haven't reached our maximum path length, there are
6585 insns following the target of the jump, this is the only use of the
6586 jump label, and the target label is preceded by a BARRIER.
6588 Alternatively, we can follow the jump if it branches around a
6589 block of code and there are no other branches into the block.
6590 In this case invalidate_skipped_block will be called to invalidate any
6591 registers set in the block when following the jump. */
6593 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6594 && JUMP_P (p)
6595 && GET_CODE (PATTERN (p)) == SET
6596 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6597 && JUMP_LABEL (p) != 0
6598 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6599 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6601 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6602 if ((!NOTE_P (q)
6603 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6604 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6605 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6606 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6607 break;
6609 /* If we ran into a BARRIER, this code is an extension of the
6610 basic block when the branch is taken. */
6611 if (follow_jumps && q != 0 && BARRIER_P (q))
6613 /* Don't allow ourself to keep walking around an
6614 always-executed loop. */
6615 if (next_real_insn (q) == next)
6617 p = NEXT_INSN (p);
6618 continue;
6621 /* Similarly, don't put a branch in our path more than once. */
6622 for (i = 0; i < path_entry; i++)
6623 if (data->path[i].branch == p)
6624 break;
6626 if (i != path_entry)
6627 break;
6629 data->path[path_entry].branch = p;
6630 data->path[path_entry++].status = PATH_TAKEN;
6632 /* This branch now ends our path. It was possible that we
6633 didn't see this branch the last time around (when the
6634 insn in front of the target was a JUMP_INSN that was
6635 turned into a no-op). */
6636 path_size = path_entry;
6638 p = JUMP_LABEL (p);
6639 /* Mark block so we won't scan it again later. */
6640 PUT_MODE (NEXT_INSN (p), QImode);
6642 /* Detect a branch around a block of code. */
6643 else if (skip_blocks && q != 0 && !LABEL_P (q))
6645 rtx tmp;
6647 if (next_real_insn (q) == next)
6649 p = NEXT_INSN (p);
6650 continue;
6653 for (i = 0; i < path_entry; i++)
6654 if (data->path[i].branch == p)
6655 break;
6657 if (i != path_entry)
6658 break;
6660 /* This is no_labels_between_p (p, q) with an added check for
6661 reaching the end of a function (in case Q precedes P). */
6662 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6663 if (LABEL_P (tmp))
6664 break;
6666 if (tmp == q)
6668 data->path[path_entry].branch = p;
6669 data->path[path_entry++].status = PATH_AROUND;
6671 path_size = path_entry;
6673 p = JUMP_LABEL (p);
6674 /* Mark block so we won't scan it again later. */
6675 PUT_MODE (NEXT_INSN (p), QImode);
6679 p = NEXT_INSN (p);
6682 data->low_cuid = low_cuid;
6683 data->high_cuid = high_cuid;
6684 data->nsets = nsets;
6685 data->last = p;
6687 /* If all jumps in the path are not taken, set our path length to zero
6688 so a rescan won't be done. */
6689 for (i = path_size - 1; i >= 0; i--)
6690 if (data->path[i].status != PATH_NOT_TAKEN)
6691 break;
6693 if (i == -1)
6694 data->path_size = 0;
6695 else
6696 data->path_size = path_size;
6698 /* End the current branch path. */
6699 data->path[path_size].branch = 0;
6702 /* Perform cse on the instructions of a function.
6703 F is the first instruction.
6704 NREGS is one plus the highest pseudo-reg number used in the instruction.
6706 Returns 1 if jump_optimize should be redone due to simplifications
6707 in conditional jump instructions. */
6710 cse_main (rtx f, int nregs, FILE *file)
6712 struct cse_basic_block_data val;
6713 rtx insn = f;
6714 int i;
6716 val.path = xmalloc (sizeof (struct branch_path)
6717 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6719 cse_jumps_altered = 0;
6720 recorded_label_ref = 0;
6721 constant_pool_entries_cost = 0;
6722 constant_pool_entries_regcost = 0;
6723 val.path_size = 0;
6724 rtl_hooks = cse_rtl_hooks;
6726 init_recog ();
6727 init_alias_analysis ();
6729 max_reg = nregs;
6731 max_insn_uid = get_max_uid ();
6733 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6735 /* Reset the counter indicating how many elements have been made
6736 thus far. */
6737 n_elements_made = 0;
6739 /* Find the largest uid. */
6741 max_uid = get_max_uid ();
6742 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6744 /* Compute the mapping from uids to cuids.
6745 CUIDs are numbers assigned to insns, like uids,
6746 except that cuids increase monotonically through the code.
6747 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6748 between two insns is not affected by -g. */
6750 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6752 if (!NOTE_P (insn)
6753 || NOTE_LINE_NUMBER (insn) < 0)
6754 INSN_CUID (insn) = ++i;
6755 else
6756 /* Give a line number note the same cuid as preceding insn. */
6757 INSN_CUID (insn) = i;
6760 /* Loop over basic blocks.
6761 Compute the maximum number of qty's needed for each basic block
6762 (which is 2 for each SET). */
6763 insn = f;
6764 while (insn)
6766 cse_altered = 0;
6767 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6768 flag_cse_skip_blocks);
6770 /* If this basic block was already processed or has no sets, skip it. */
6771 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6773 PUT_MODE (insn, VOIDmode);
6774 insn = (val.last ? NEXT_INSN (val.last) : 0);
6775 val.path_size = 0;
6776 continue;
6779 cse_basic_block_start = val.low_cuid;
6780 cse_basic_block_end = val.high_cuid;
6781 max_qty = val.nsets * 2;
6783 if (file)
6784 fnotice (file, ";; Processing block from %d to %d, %d sets.\n",
6785 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6786 val.nsets);
6788 /* Make MAX_QTY bigger to give us room to optimize
6789 past the end of this basic block, if that should prove useful. */
6790 if (max_qty < 500)
6791 max_qty = 500;
6793 /* If this basic block is being extended by following certain jumps,
6794 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6795 Otherwise, we start after this basic block. */
6796 if (val.path_size > 0)
6797 cse_basic_block (insn, val.last, val.path);
6798 else
6800 int old_cse_jumps_altered = cse_jumps_altered;
6801 rtx temp;
6803 /* When cse changes a conditional jump to an unconditional
6804 jump, we want to reprocess the block, since it will give
6805 us a new branch path to investigate. */
6806 cse_jumps_altered = 0;
6807 temp = cse_basic_block (insn, val.last, val.path);
6808 if (cse_jumps_altered == 0
6809 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6810 insn = temp;
6812 cse_jumps_altered |= old_cse_jumps_altered;
6815 if (cse_altered)
6816 ggc_collect ();
6818 #ifdef USE_C_ALLOCA
6819 alloca (0);
6820 #endif
6823 if (max_elements_made < n_elements_made)
6824 max_elements_made = n_elements_made;
6826 /* Clean up. */
6827 end_alias_analysis ();
6828 free (uid_cuid);
6829 free (reg_eqv_table);
6830 free (val.path);
6831 rtl_hooks = general_rtl_hooks;
6833 return cse_jumps_altered || recorded_label_ref;
6836 /* Process a single basic block. FROM and TO and the limits of the basic
6837 block. NEXT_BRANCH points to the branch path when following jumps or
6838 a null path when not following jumps. */
6840 static rtx
6841 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
6843 rtx insn;
6844 int to_usage = 0;
6845 rtx libcall_insn = NULL_RTX;
6846 int num_insns = 0;
6847 int no_conflict = 0;
6849 /* Allocate the space needed by qty_table. */
6850 qty_table = xmalloc (max_qty * sizeof (struct qty_table_elem));
6852 new_basic_block ();
6854 /* TO might be a label. If so, protect it from being deleted. */
6855 if (to != 0 && LABEL_P (to))
6856 ++LABEL_NUSES (to);
6858 for (insn = from; insn != to; insn = NEXT_INSN (insn))
6860 enum rtx_code code = GET_CODE (insn);
6862 /* If we have processed 1,000 insns, flush the hash table to
6863 avoid extreme quadratic behavior. We must not include NOTEs
6864 in the count since there may be more of them when generating
6865 debugging information. If we clear the table at different
6866 times, code generated with -g -O might be different than code
6867 generated with -O but not -g.
6869 ??? This is a real kludge and needs to be done some other way.
6870 Perhaps for 2.9. */
6871 if (code != NOTE && num_insns++ > 1000)
6873 flush_hash_table ();
6874 num_insns = 0;
6877 /* See if this is a branch that is part of the path. If so, and it is
6878 to be taken, do so. */
6879 if (next_branch->branch == insn)
6881 enum taken status = next_branch++->status;
6882 if (status != PATH_NOT_TAKEN)
6884 if (status == PATH_TAKEN)
6885 record_jump_equiv (insn, 1);
6886 else
6887 invalidate_skipped_block (NEXT_INSN (insn));
6889 /* Set the last insn as the jump insn; it doesn't affect cc0.
6890 Then follow this branch. */
6891 #ifdef HAVE_cc0
6892 prev_insn_cc0 = 0;
6893 prev_insn = insn;
6894 #endif
6895 insn = JUMP_LABEL (insn);
6896 continue;
6900 if (GET_MODE (insn) == QImode)
6901 PUT_MODE (insn, VOIDmode);
6903 if (GET_RTX_CLASS (code) == RTX_INSN)
6905 rtx p;
6907 /* Process notes first so we have all notes in canonical forms when
6908 looking for duplicate operations. */
6910 if (REG_NOTES (insn))
6911 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
6913 /* Track when we are inside in LIBCALL block. Inside such a block,
6914 we do not want to record destinations. The last insn of a
6915 LIBCALL block is not considered to be part of the block, since
6916 its destination is the result of the block and hence should be
6917 recorded. */
6919 if (REG_NOTES (insn) != 0)
6921 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
6922 libcall_insn = XEXP (p, 0);
6923 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
6925 /* Keep libcall_insn for the last SET insn of a no-conflict
6926 block to prevent changing the destination. */
6927 if (! no_conflict)
6928 libcall_insn = 0;
6929 else
6930 no_conflict = -1;
6932 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
6933 no_conflict = 1;
6936 cse_insn (insn, libcall_insn);
6938 if (no_conflict == -1)
6940 libcall_insn = 0;
6941 no_conflict = 0;
6944 /* If we haven't already found an insn where we added a LABEL_REF,
6945 check this one. */
6946 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
6947 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6948 (void *) insn))
6949 recorded_label_ref = 1;
6952 /* If INSN is now an unconditional jump, skip to the end of our
6953 basic block by pretending that we just did the last insn in the
6954 basic block. If we are jumping to the end of our block, show
6955 that we can have one usage of TO. */
6957 if (any_uncondjump_p (insn))
6959 if (to == 0)
6961 free (qty_table);
6962 return 0;
6965 if (JUMP_LABEL (insn) == to)
6966 to_usage = 1;
6968 /* Maybe TO was deleted because the jump is unconditional.
6969 If so, there is nothing left in this basic block. */
6970 /* ??? Perhaps it would be smarter to set TO
6971 to whatever follows this insn,
6972 and pretend the basic block had always ended here. */
6973 if (INSN_DELETED_P (to))
6974 break;
6976 insn = PREV_INSN (to);
6979 /* See if it is ok to keep on going past the label
6980 which used to end our basic block. Remember that we incremented
6981 the count of that label, so we decrement it here. If we made
6982 a jump unconditional, TO_USAGE will be one; in that case, we don't
6983 want to count the use in that jump. */
6985 if (to != 0 && NEXT_INSN (insn) == to
6986 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
6988 struct cse_basic_block_data val;
6989 rtx prev;
6991 insn = NEXT_INSN (to);
6993 /* If TO was the last insn in the function, we are done. */
6994 if (insn == 0)
6996 free (qty_table);
6997 return 0;
7000 /* If TO was preceded by a BARRIER we are done with this block
7001 because it has no continuation. */
7002 prev = prev_nonnote_insn (to);
7003 if (prev && BARRIER_P (prev))
7005 free (qty_table);
7006 return insn;
7009 /* Find the end of the following block. Note that we won't be
7010 following branches in this case. */
7011 to_usage = 0;
7012 val.path_size = 0;
7013 val.path = xmalloc (sizeof (struct branch_path)
7014 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7015 cse_end_of_basic_block (insn, &val, 0, 0);
7016 free (val.path);
7018 /* If the tables we allocated have enough space left
7019 to handle all the SETs in the next basic block,
7020 continue through it. Otherwise, return,
7021 and that block will be scanned individually. */
7022 if (val.nsets * 2 + next_qty > max_qty)
7023 break;
7025 cse_basic_block_start = val.low_cuid;
7026 cse_basic_block_end = val.high_cuid;
7027 to = val.last;
7029 /* Prevent TO from being deleted if it is a label. */
7030 if (to != 0 && LABEL_P (to))
7031 ++LABEL_NUSES (to);
7033 /* Back up so we process the first insn in the extension. */
7034 insn = PREV_INSN (insn);
7038 gcc_assert (next_qty <= max_qty);
7040 free (qty_table);
7042 return to ? NEXT_INSN (to) : 0;
7045 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7046 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7048 static int
7049 check_for_label_ref (rtx *rtl, void *data)
7051 rtx insn = (rtx) data;
7053 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7054 we must rerun jump since it needs to place the note. If this is a
7055 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7056 since no REG_LABEL will be added. */
7057 return (GET_CODE (*rtl) == LABEL_REF
7058 && ! LABEL_REF_NONLOCAL_P (*rtl)
7059 && LABEL_P (XEXP (*rtl, 0))
7060 && INSN_UID (XEXP (*rtl, 0)) != 0
7061 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7064 /* Count the number of times registers are used (not set) in X.
7065 COUNTS is an array in which we accumulate the count, INCR is how much
7066 we count each register usage. */
7068 static void
7069 count_reg_usage (rtx x, int *counts, int incr)
7071 enum rtx_code code;
7072 rtx note;
7073 const char *fmt;
7074 int i, j;
7076 if (x == 0)
7077 return;
7079 switch (code = GET_CODE (x))
7081 case REG:
7082 counts[REGNO (x)] += incr;
7083 return;
7085 case PC:
7086 case CC0:
7087 case CONST:
7088 case CONST_INT:
7089 case CONST_DOUBLE:
7090 case CONST_VECTOR:
7091 case SYMBOL_REF:
7092 case LABEL_REF:
7093 return;
7095 case CLOBBER:
7096 /* If we are clobbering a MEM, mark any registers inside the address
7097 as being used. */
7098 if (MEM_P (XEXP (x, 0)))
7099 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, incr);
7100 return;
7102 case SET:
7103 /* Unless we are setting a REG, count everything in SET_DEST. */
7104 if (!REG_P (SET_DEST (x)))
7105 count_reg_usage (SET_DEST (x), counts, incr);
7106 count_reg_usage (SET_SRC (x), counts, incr);
7107 return;
7109 case CALL_INSN:
7110 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, incr);
7111 /* Fall through. */
7113 case INSN:
7114 case JUMP_INSN:
7115 count_reg_usage (PATTERN (x), counts, incr);
7117 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7118 use them. */
7120 note = find_reg_equal_equiv_note (x);
7121 if (note)
7123 rtx eqv = XEXP (note, 0);
7125 if (GET_CODE (eqv) == EXPR_LIST)
7126 /* This REG_EQUAL note describes the result of a function call.
7127 Process all the arguments. */
7130 count_reg_usage (XEXP (eqv, 0), counts, incr);
7131 eqv = XEXP (eqv, 1);
7133 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7134 else
7135 count_reg_usage (eqv, counts, incr);
7137 return;
7139 case EXPR_LIST:
7140 if (REG_NOTE_KIND (x) == REG_EQUAL
7141 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7142 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7143 involving registers in the address. */
7144 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7145 count_reg_usage (XEXP (x, 0), counts, incr);
7147 count_reg_usage (XEXP (x, 1), counts, incr);
7148 return;
7150 case ASM_OPERANDS:
7151 /* Iterate over just the inputs, not the constraints as well. */
7152 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7153 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, incr);
7154 return;
7156 case INSN_LIST:
7157 gcc_unreachable ();
7159 default:
7160 break;
7163 fmt = GET_RTX_FORMAT (code);
7164 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7166 if (fmt[i] == 'e')
7167 count_reg_usage (XEXP (x, i), counts, incr);
7168 else if (fmt[i] == 'E')
7169 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7170 count_reg_usage (XVECEXP (x, i, j), counts, incr);
7174 /* Return true if set is live. */
7175 static bool
7176 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7177 int *counts)
7179 #ifdef HAVE_cc0
7180 rtx tem;
7181 #endif
7183 if (set_noop_p (set))
7186 #ifdef HAVE_cc0
7187 else if (GET_CODE (SET_DEST (set)) == CC0
7188 && !side_effects_p (SET_SRC (set))
7189 && ((tem = next_nonnote_insn (insn)) == 0
7190 || !INSN_P (tem)
7191 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7192 return false;
7193 #endif
7194 else if (!REG_P (SET_DEST (set))
7195 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7196 || counts[REGNO (SET_DEST (set))] != 0
7197 || side_effects_p (SET_SRC (set)))
7198 return true;
7199 return false;
7202 /* Return true if insn is live. */
7204 static bool
7205 insn_live_p (rtx insn, int *counts)
7207 int i;
7208 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7209 return true;
7210 else if (GET_CODE (PATTERN (insn)) == SET)
7211 return set_live_p (PATTERN (insn), insn, counts);
7212 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7214 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7216 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7218 if (GET_CODE (elt) == SET)
7220 if (set_live_p (elt, insn, counts))
7221 return true;
7223 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7224 return true;
7226 return false;
7228 else
7229 return true;
7232 /* Return true if libcall is dead as a whole. */
7234 static bool
7235 dead_libcall_p (rtx insn, int *counts)
7237 rtx note, set, new;
7239 /* See if there's a REG_EQUAL note on this insn and try to
7240 replace the source with the REG_EQUAL expression.
7242 We assume that insns with REG_RETVALs can only be reg->reg
7243 copies at this point. */
7244 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7245 if (!note)
7246 return false;
7248 set = single_set (insn);
7249 if (!set)
7250 return false;
7252 new = simplify_rtx (XEXP (note, 0));
7253 if (!new)
7254 new = XEXP (note, 0);
7256 /* While changing insn, we must update the counts accordingly. */
7257 count_reg_usage (insn, counts, -1);
7259 if (validate_change (insn, &SET_SRC (set), new, 0))
7261 count_reg_usage (insn, counts, 1);
7262 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7263 remove_note (insn, note);
7264 return true;
7267 if (CONSTANT_P (new))
7269 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7270 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7272 count_reg_usage (insn, counts, 1);
7273 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7274 remove_note (insn, note);
7275 return true;
7279 count_reg_usage (insn, counts, 1);
7280 return false;
7283 /* Scan all the insns and delete any that are dead; i.e., they store a register
7284 that is never used or they copy a register to itself.
7286 This is used to remove insns made obviously dead by cse, loop or other
7287 optimizations. It improves the heuristics in loop since it won't try to
7288 move dead invariants out of loops or make givs for dead quantities. The
7289 remaining passes of the compilation are also sped up. */
7292 delete_trivially_dead_insns (rtx insns, int nreg)
7294 int *counts;
7295 rtx insn, prev;
7296 int in_libcall = 0, dead_libcall = 0;
7297 int ndead = 0, nlastdead, niterations = 0;
7299 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7300 /* First count the number of times each register is used. */
7301 counts = xcalloc (nreg, sizeof (int));
7302 for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn))
7303 count_reg_usage (insn, counts, 1);
7307 nlastdead = ndead;
7308 niterations++;
7309 /* Go from the last insn to the first and delete insns that only set unused
7310 registers or copy a register to itself. As we delete an insn, remove
7311 usage counts for registers it uses.
7313 The first jump optimization pass may leave a real insn as the last
7314 insn in the function. We must not skip that insn or we may end
7315 up deleting code that is not really dead. */
7316 insn = get_last_insn ();
7317 if (! INSN_P (insn))
7318 insn = prev_real_insn (insn);
7320 for (; insn; insn = prev)
7322 int live_insn = 0;
7324 prev = prev_real_insn (insn);
7326 /* Don't delete any insns that are part of a libcall block unless
7327 we can delete the whole libcall block.
7329 Flow or loop might get confused if we did that. Remember
7330 that we are scanning backwards. */
7331 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7333 in_libcall = 1;
7334 live_insn = 1;
7335 dead_libcall = dead_libcall_p (insn, counts);
7337 else if (in_libcall)
7338 live_insn = ! dead_libcall;
7339 else
7340 live_insn = insn_live_p (insn, counts);
7342 /* If this is a dead insn, delete it and show registers in it aren't
7343 being used. */
7345 if (! live_insn)
7347 count_reg_usage (insn, counts, -1);
7348 delete_insn_and_edges (insn);
7349 ndead++;
7352 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7354 in_libcall = 0;
7355 dead_libcall = 0;
7359 while (ndead != nlastdead);
7361 if (dump_file && ndead)
7362 fprintf (dump_file, "Deleted %i trivially dead insns; %i iterations\n",
7363 ndead, niterations);
7364 /* Clean up. */
7365 free (counts);
7366 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7367 return ndead;
7370 /* This function is called via for_each_rtx. The argument, NEWREG, is
7371 a condition code register with the desired mode. If we are looking
7372 at the same register in a different mode, replace it with
7373 NEWREG. */
7375 static int
7376 cse_change_cc_mode (rtx *loc, void *data)
7378 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7380 if (*loc
7381 && REG_P (*loc)
7382 && REGNO (*loc) == REGNO (args->newreg)
7383 && GET_MODE (*loc) != GET_MODE (args->newreg))
7385 validate_change (args->insn, loc, args->newreg, 1);
7387 return -1;
7389 return 0;
7392 /* Change the mode of any reference to the register REGNO (NEWREG) to
7393 GET_MODE (NEWREG) in INSN. */
7395 static void
7396 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7398 struct change_cc_mode_args args;
7399 int success;
7401 if (!INSN_P (insn))
7402 return;
7404 args.insn = insn;
7405 args.newreg = newreg;
7407 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7408 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7410 /* If the following assertion was triggered, there is most probably
7411 something wrong with the cc_modes_compatible back end function.
7412 CC modes only can be considered compatible if the insn - with the mode
7413 replaced by any of the compatible modes - can still be recognized. */
7414 success = apply_change_group ();
7415 gcc_assert (success);
7418 /* Change the mode of any reference to the register REGNO (NEWREG) to
7419 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7420 any instruction which modifies NEWREG. */
7422 static void
7423 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7425 rtx insn;
7427 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7429 if (! INSN_P (insn))
7430 continue;
7432 if (reg_set_p (newreg, insn))
7433 return;
7435 cse_change_cc_mode_insn (insn, newreg);
7439 /* BB is a basic block which finishes with CC_REG as a condition code
7440 register which is set to CC_SRC. Look through the successors of BB
7441 to find blocks which have a single predecessor (i.e., this one),
7442 and look through those blocks for an assignment to CC_REG which is
7443 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7444 permitted to change the mode of CC_SRC to a compatible mode. This
7445 returns VOIDmode if no equivalent assignments were found.
7446 Otherwise it returns the mode which CC_SRC should wind up with.
7448 The main complexity in this function is handling the mode issues.
7449 We may have more than one duplicate which we can eliminate, and we
7450 try to find a mode which will work for multiple duplicates. */
7452 static enum machine_mode
7453 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7455 bool found_equiv;
7456 enum machine_mode mode;
7457 unsigned int insn_count;
7458 edge e;
7459 rtx insns[2];
7460 enum machine_mode modes[2];
7461 rtx last_insns[2];
7462 unsigned int i;
7463 rtx newreg;
7464 edge_iterator ei;
7466 /* We expect to have two successors. Look at both before picking
7467 the final mode for the comparison. If we have more successors
7468 (i.e., some sort of table jump, although that seems unlikely),
7469 then we require all beyond the first two to use the same
7470 mode. */
7472 found_equiv = false;
7473 mode = GET_MODE (cc_src);
7474 insn_count = 0;
7475 FOR_EACH_EDGE (e, ei, bb->succs)
7477 rtx insn;
7478 rtx end;
7480 if (e->flags & EDGE_COMPLEX)
7481 continue;
7483 if (EDGE_COUNT (e->dest->preds) != 1
7484 || e->dest == EXIT_BLOCK_PTR)
7485 continue;
7487 end = NEXT_INSN (BB_END (e->dest));
7488 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7490 rtx set;
7492 if (! INSN_P (insn))
7493 continue;
7495 /* If CC_SRC is modified, we have to stop looking for
7496 something which uses it. */
7497 if (modified_in_p (cc_src, insn))
7498 break;
7500 /* Check whether INSN sets CC_REG to CC_SRC. */
7501 set = single_set (insn);
7502 if (set
7503 && REG_P (SET_DEST (set))
7504 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7506 bool found;
7507 enum machine_mode set_mode;
7508 enum machine_mode comp_mode;
7510 found = false;
7511 set_mode = GET_MODE (SET_SRC (set));
7512 comp_mode = set_mode;
7513 if (rtx_equal_p (cc_src, SET_SRC (set)))
7514 found = true;
7515 else if (GET_CODE (cc_src) == COMPARE
7516 && GET_CODE (SET_SRC (set)) == COMPARE
7517 && mode != set_mode
7518 && rtx_equal_p (XEXP (cc_src, 0),
7519 XEXP (SET_SRC (set), 0))
7520 && rtx_equal_p (XEXP (cc_src, 1),
7521 XEXP (SET_SRC (set), 1)))
7524 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7525 if (comp_mode != VOIDmode
7526 && (can_change_mode || comp_mode == mode))
7527 found = true;
7530 if (found)
7532 found_equiv = true;
7533 if (insn_count < ARRAY_SIZE (insns))
7535 insns[insn_count] = insn;
7536 modes[insn_count] = set_mode;
7537 last_insns[insn_count] = end;
7538 ++insn_count;
7540 if (mode != comp_mode)
7542 gcc_assert (can_change_mode);
7543 mode = comp_mode;
7545 /* The modified insn will be re-recognized later. */
7546 PUT_MODE (cc_src, mode);
7549 else
7551 if (set_mode != mode)
7553 /* We found a matching expression in the
7554 wrong mode, but we don't have room to
7555 store it in the array. Punt. This case
7556 should be rare. */
7557 break;
7559 /* INSN sets CC_REG to a value equal to CC_SRC
7560 with the right mode. We can simply delete
7561 it. */
7562 delete_insn (insn);
7565 /* We found an instruction to delete. Keep looking,
7566 in the hopes of finding a three-way jump. */
7567 continue;
7570 /* We found an instruction which sets the condition
7571 code, so don't look any farther. */
7572 break;
7575 /* If INSN sets CC_REG in some other way, don't look any
7576 farther. */
7577 if (reg_set_p (cc_reg, insn))
7578 break;
7581 /* If we fell off the bottom of the block, we can keep looking
7582 through successors. We pass CAN_CHANGE_MODE as false because
7583 we aren't prepared to handle compatibility between the
7584 further blocks and this block. */
7585 if (insn == end)
7587 enum machine_mode submode;
7589 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7590 if (submode != VOIDmode)
7592 gcc_assert (submode == mode);
7593 found_equiv = true;
7594 can_change_mode = false;
7599 if (! found_equiv)
7600 return VOIDmode;
7602 /* Now INSN_COUNT is the number of instructions we found which set
7603 CC_REG to a value equivalent to CC_SRC. The instructions are in
7604 INSNS. The modes used by those instructions are in MODES. */
7606 newreg = NULL_RTX;
7607 for (i = 0; i < insn_count; ++i)
7609 if (modes[i] != mode)
7611 /* We need to change the mode of CC_REG in INSNS[i] and
7612 subsequent instructions. */
7613 if (! newreg)
7615 if (GET_MODE (cc_reg) == mode)
7616 newreg = cc_reg;
7617 else
7618 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7620 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7621 newreg);
7624 delete_insn (insns[i]);
7627 return mode;
7630 /* If we have a fixed condition code register (or two), walk through
7631 the instructions and try to eliminate duplicate assignments. */
7633 void
7634 cse_condition_code_reg (void)
7636 unsigned int cc_regno_1;
7637 unsigned int cc_regno_2;
7638 rtx cc_reg_1;
7639 rtx cc_reg_2;
7640 basic_block bb;
7642 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7643 return;
7645 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7646 if (cc_regno_2 != INVALID_REGNUM)
7647 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7648 else
7649 cc_reg_2 = NULL_RTX;
7651 FOR_EACH_BB (bb)
7653 rtx last_insn;
7654 rtx cc_reg;
7655 rtx insn;
7656 rtx cc_src_insn;
7657 rtx cc_src;
7658 enum machine_mode mode;
7659 enum machine_mode orig_mode;
7661 /* Look for blocks which end with a conditional jump based on a
7662 condition code register. Then look for the instruction which
7663 sets the condition code register. Then look through the
7664 successor blocks for instructions which set the condition
7665 code register to the same value. There are other possible
7666 uses of the condition code register, but these are by far the
7667 most common and the ones which we are most likely to be able
7668 to optimize. */
7670 last_insn = BB_END (bb);
7671 if (!JUMP_P (last_insn))
7672 continue;
7674 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7675 cc_reg = cc_reg_1;
7676 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7677 cc_reg = cc_reg_2;
7678 else
7679 continue;
7681 cc_src_insn = NULL_RTX;
7682 cc_src = NULL_RTX;
7683 for (insn = PREV_INSN (last_insn);
7684 insn && insn != PREV_INSN (BB_HEAD (bb));
7685 insn = PREV_INSN (insn))
7687 rtx set;
7689 if (! INSN_P (insn))
7690 continue;
7691 set = single_set (insn);
7692 if (set
7693 && REG_P (SET_DEST (set))
7694 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7696 cc_src_insn = insn;
7697 cc_src = SET_SRC (set);
7698 break;
7700 else if (reg_set_p (cc_reg, insn))
7701 break;
7704 if (! cc_src_insn)
7705 continue;
7707 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7708 continue;
7710 /* Now CC_REG is a condition code register used for a
7711 conditional jump at the end of the block, and CC_SRC, in
7712 CC_SRC_INSN, is the value to which that condition code
7713 register is set, and CC_SRC is still meaningful at the end of
7714 the basic block. */
7716 orig_mode = GET_MODE (cc_src);
7717 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7718 if (mode != VOIDmode)
7720 gcc_assert (mode == GET_MODE (cc_src));
7721 if (mode != orig_mode)
7723 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7725 cse_change_cc_mode_insn (cc_src_insn, newreg);
7727 /* Do the same in the following insns that use the
7728 current value of CC_REG within BB. */
7729 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7730 NEXT_INSN (last_insn),
7731 newreg);