* gfortran.dg/reduction.f90: Add checks with complex arguments.
[official-gcc.git] / gcc / cse.c
blob47708e21cb4d9ec3c519a1ab1f126b586936ad95
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
28 #include "rtl.h"
29 #include "tm_p.h"
30 #include "regs.h"
31 #include "hard-reg-set.h"
32 #include "basic-block.h"
33 #include "flags.h"
34 #include "real.h"
35 #include "insn-config.h"
36 #include "recog.h"
37 #include "function.h"
38 #include "expr.h"
39 #include "toplev.h"
40 #include "output.h"
41 #include "ggc.h"
42 #include "timevar.h"
43 #include "except.h"
44 #include "target.h"
45 #include "params.h"
46 #include "rtlhooks-def.h"
48 /* The basic idea of common subexpression elimination is to go
49 through the code, keeping a record of expressions that would
50 have the same value at the current scan point, and replacing
51 expressions encountered with the cheapest equivalent expression.
53 It is too complicated to keep track of the different possibilities
54 when control paths merge in this code; so, at each label, we forget all
55 that is known and start fresh. This can be described as processing each
56 extended basic block separately. We have a separate pass to perform
57 global CSE.
59 Note CSE can turn a conditional or computed jump into a nop or
60 an unconditional jump. When this occurs we arrange to run the jump
61 optimizer after CSE to delete the unreachable code.
63 We use two data structures to record the equivalent expressions:
64 a hash table for most expressions, and a vector of "quantity
65 numbers" to record equivalent (pseudo) registers.
67 The use of the special data structure for registers is desirable
68 because it is faster. It is possible because registers references
69 contain a fairly small number, the register number, taken from
70 a contiguously allocated series, and two register references are
71 identical if they have the same number. General expressions
72 do not have any such thing, so the only way to retrieve the
73 information recorded on an expression other than a register
74 is to keep it in a hash table.
76 Registers and "quantity numbers":
78 At the start of each basic block, all of the (hardware and pseudo)
79 registers used in the function are given distinct quantity
80 numbers to indicate their contents. During scan, when the code
81 copies one register into another, we copy the quantity number.
82 When a register is loaded in any other way, we allocate a new
83 quantity number to describe the value generated by this operation.
84 `reg_qty' records what quantity a register is currently thought
85 of as containing.
87 All real quantity numbers are greater than or equal to `max_reg'.
88 If register N has not been assigned a quantity, reg_qty[N] will equal N.
90 Quantity numbers below `max_reg' do not exist and none of the `qty_table'
91 entries should be referenced with an index below `max_reg'.
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 #ifdef HAVE_cc0
266 /* For machines that have a CC0, we do not record its value in the hash
267 table since its use is guaranteed to be the insn immediately following
268 its definition and any other insn is presumed to invalidate it.
270 Instead, we store below the value last assigned to CC0. If it should
271 happen to be a constant, it is stored in preference to the actual
272 assigned value. In case it is a constant, we store the mode in which
273 the constant should be interpreted. */
275 static rtx prev_insn_cc0;
276 static enum machine_mode prev_insn_cc0_mode;
278 /* Previous actual insn. 0 if at first insn of basic block. */
280 static rtx prev_insn;
281 #endif
283 /* Insn being scanned. */
285 static rtx this_insn;
287 /* Index by register number, gives the number of the next (or
288 previous) register in the chain of registers sharing the same
289 value.
291 Or -1 if this register is at the end of the chain.
293 If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */
295 /* Per-register equivalence chain. */
296 struct reg_eqv_elem
298 int next, prev;
301 /* The table of all register equivalence chains. */
302 static struct reg_eqv_elem *reg_eqv_table;
304 struct cse_reg_info
306 /* Next in hash chain. */
307 struct cse_reg_info *hash_next;
309 /* The next cse_reg_info structure in the free or used list. */
310 struct cse_reg_info *next;
312 /* Search key */
313 unsigned int regno;
315 /* The quantity number of the register's current contents. */
316 int reg_qty;
318 /* The number of times the register has been altered in the current
319 basic block. */
320 int reg_tick;
322 /* The REG_TICK value at which rtx's containing this register are
323 valid in the hash table. If this does not equal the current
324 reg_tick value, such expressions existing in the hash table are
325 invalid. */
326 int reg_in_table;
328 /* The SUBREG that was set when REG_TICK was last incremented. Set
329 to -1 if the last store was to the whole register, not a subreg. */
330 unsigned int subreg_ticked;
333 /* A free list of cse_reg_info entries. */
334 static struct cse_reg_info *cse_reg_info_free_list;
336 /* A used list of cse_reg_info entries. */
337 static struct cse_reg_info *cse_reg_info_used_list;
338 static struct cse_reg_info *cse_reg_info_used_list_end;
340 /* A mapping from registers to cse_reg_info data structures. */
341 #define REGHASH_SHIFT 7
342 #define REGHASH_SIZE (1 << REGHASH_SHIFT)
343 #define REGHASH_MASK (REGHASH_SIZE - 1)
344 static struct cse_reg_info *reg_hash[REGHASH_SIZE];
346 #define REGHASH_FN(REGNO) \
347 (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK)
349 /* The last lookup we did into the cse_reg_info_tree. This allows us
350 to cache repeated lookups. */
351 static unsigned int cached_regno;
352 static struct cse_reg_info *cached_cse_reg_info;
354 /* A HARD_REG_SET containing all the hard registers for which there is
355 currently a REG expression in the hash table. Note the difference
356 from the above variables, which indicate if the REG is mentioned in some
357 expression in the table. */
359 static HARD_REG_SET hard_regs_in_table;
361 /* CUID of insn that starts the basic block currently being cse-processed. */
363 static int cse_basic_block_start;
365 /* CUID of insn that ends the basic block currently being cse-processed. */
367 static int cse_basic_block_end;
369 /* Vector mapping INSN_UIDs to cuids.
370 The cuids are like uids but increase monotonically always.
371 We use them to see whether a reg is used outside a given basic block. */
373 static int *uid_cuid;
375 /* Highest UID in UID_CUID. */
376 static int max_uid;
378 /* Get the cuid of an insn. */
380 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
382 /* Nonzero if this pass has made changes, and therefore it's
383 worthwhile to run the garbage collector. */
385 static int cse_altered;
387 /* Nonzero if cse has altered conditional jump insns
388 in such a way that jump optimization should be redone. */
390 static int cse_jumps_altered;
392 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
393 REG_LABEL, we have to rerun jump after CSE to put in the note. */
394 static int recorded_label_ref;
396 /* canon_hash stores 1 in do_not_record
397 if it notices a reference to CC0, PC, or some other volatile
398 subexpression. */
400 static int do_not_record;
402 #ifdef LOAD_EXTEND_OP
404 /* Scratch rtl used when looking for load-extended copy of a MEM. */
405 static rtx memory_extend_rtx;
406 #endif
408 /* canon_hash stores 1 in hash_arg_in_memory
409 if it notices a reference to memory within the expression being hashed. */
411 static int hash_arg_in_memory;
413 /* The hash table contains buckets which are chains of `struct table_elt's,
414 each recording one expression's information.
415 That expression is in the `exp' field.
417 The canon_exp field contains a canonical (from the point of view of
418 alias analysis) version of the `exp' field.
420 Those elements with the same hash code are chained in both directions
421 through the `next_same_hash' and `prev_same_hash' fields.
423 Each set of expressions with equivalent values
424 are on a two-way chain through the `next_same_value'
425 and `prev_same_value' fields, and all point with
426 the `first_same_value' field at the first element in
427 that chain. The chain is in order of increasing cost.
428 Each element's cost value is in its `cost' field.
430 The `in_memory' field is nonzero for elements that
431 involve any reference to memory. These elements are removed
432 whenever a write is done to an unidentified location in memory.
433 To be safe, we assume that a memory address is unidentified unless
434 the address is either a symbol constant or a constant plus
435 the frame pointer or argument pointer.
437 The `related_value' field is used to connect related expressions
438 (that differ by adding an integer).
439 The related expressions are chained in a circular fashion.
440 `related_value' is zero for expressions for which this
441 chain is not useful.
443 The `cost' field stores the cost of this element's expression.
444 The `regcost' field stores the value returned by approx_reg_cost for
445 this element's expression.
447 The `is_const' flag is set if the element is a constant (including
448 a fixed address).
450 The `flag' field is used as a temporary during some search routines.
452 The `mode' field is usually the same as GET_MODE (`exp'), but
453 if `exp' is a CONST_INT and has no machine mode then the `mode'
454 field is the mode it was being used as. Each constant is
455 recorded separately for each mode it is used with. */
457 struct table_elt
459 rtx exp;
460 rtx canon_exp;
461 struct table_elt *next_same_hash;
462 struct table_elt *prev_same_hash;
463 struct table_elt *next_same_value;
464 struct table_elt *prev_same_value;
465 struct table_elt *first_same_value;
466 struct table_elt *related_value;
467 int cost;
468 int regcost;
469 /* The size of this field should match the size
470 of the mode field of struct rtx_def (see rtl.h). */
471 ENUM_BITFIELD(machine_mode) mode : 8;
472 char in_memory;
473 char is_const;
474 char flag;
477 /* We don't want a lot of buckets, because we rarely have very many
478 things stored in the hash table, and a lot of buckets slows
479 down a lot of loops that happen frequently. */
480 #define HASH_SHIFT 5
481 #define HASH_SIZE (1 << HASH_SHIFT)
482 #define HASH_MASK (HASH_SIZE - 1)
484 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
485 register (hard registers may require `do_not_record' to be set). */
487 #define HASH(X, M) \
488 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
489 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
490 : canon_hash (X, M)) & HASH_MASK)
492 /* Like HASH, but without side-effects. */
493 #define SAFE_HASH(X, M) \
494 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
495 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
496 : safe_hash (X, M)) & HASH_MASK)
498 /* Determine whether register number N is considered a fixed register for the
499 purpose of approximating register costs.
500 It is desirable to replace other regs with fixed regs, to reduce need for
501 non-fixed hard regs.
502 A reg wins if it is either the frame pointer or designated as fixed. */
503 #define FIXED_REGNO_P(N) \
504 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
505 || fixed_regs[N] || global_regs[N])
507 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
508 hard registers and pointers into the frame are the cheapest with a cost
509 of 0. Next come pseudos with a cost of one and other hard registers with
510 a cost of 2. Aside from these special cases, call `rtx_cost'. */
512 #define CHEAP_REGNO(N) \
513 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
514 || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \
515 || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \
516 || ((N) < FIRST_PSEUDO_REGISTER \
517 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
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) != (int) (N))
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 /* Surviving equivalence class when two equivalence classes are merged
568 by recording the effects of a jump in the last insn. Zero if the
569 last insn was not a conditional jump. */
571 static struct table_elt *last_jump_equiv_class;
573 /* Set to the cost of a constant pool reference if one was found for a
574 symbolic constant. If this was found, it means we should try to
575 convert constants into constant pool entries if they don't fit in
576 the insn. */
578 static int constant_pool_entries_cost;
579 static int constant_pool_entries_regcost;
581 /* This data describes a block that will be processed by cse_basic_block. */
583 struct cse_basic_block_data
585 /* Lowest CUID value of insns in block. */
586 int low_cuid;
587 /* Highest CUID value of insns in block. */
588 int high_cuid;
589 /* Total number of SETs in block. */
590 int nsets;
591 /* Last insn in the block. */
592 rtx last;
593 /* Size of current branch path, if any. */
594 int path_size;
595 /* Current branch path, indicating which branches will be taken. */
596 struct branch_path
598 /* The branch insn. */
599 rtx branch;
600 /* Whether it should be taken or not. AROUND is the same as taken
601 except that it is used when the destination label is not preceded
602 by a BARRIER. */
603 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
604 } *path;
607 static bool fixed_base_plus_p (rtx x);
608 static int notreg_cost (rtx, enum rtx_code);
609 static int approx_reg_cost_1 (rtx *, void *);
610 static int approx_reg_cost (rtx);
611 static int preferable (int, int, int, int);
612 static void new_basic_block (void);
613 static void make_new_qty (unsigned int, enum machine_mode);
614 static void make_regs_eqv (unsigned int, unsigned int);
615 static void delete_reg_equiv (unsigned int);
616 static int mention_regs (rtx);
617 static int insert_regs (rtx, struct table_elt *, int);
618 static void remove_from_table (struct table_elt *, unsigned);
619 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
620 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
621 static rtx lookup_as_function (rtx, enum rtx_code);
622 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
623 enum machine_mode);
624 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
625 static void invalidate (rtx, enum machine_mode);
626 static int cse_rtx_varies_p (rtx, int);
627 static void remove_invalid_refs (unsigned int);
628 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
629 enum machine_mode);
630 static void rehash_using_reg (rtx);
631 static void invalidate_memory (void);
632 static void invalidate_for_call (void);
633 static rtx use_related_value (rtx, struct table_elt *);
635 static inline unsigned canon_hash (rtx, enum machine_mode);
636 static inline unsigned safe_hash (rtx, enum machine_mode);
637 static unsigned hash_rtx_string (const char *);
639 static rtx canon_reg (rtx, rtx);
640 static void find_best_addr (rtx, rtx *, enum machine_mode);
641 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
642 enum machine_mode *,
643 enum machine_mode *);
644 static rtx fold_rtx (rtx, rtx);
645 static rtx equiv_constant (rtx);
646 static void record_jump_equiv (rtx, int);
647 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
648 int);
649 static void cse_insn (rtx, rtx);
650 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
651 int, int, int);
652 static int addr_affects_sp_p (rtx);
653 static void invalidate_from_clobbers (rtx);
654 static rtx cse_process_notes (rtx, rtx);
655 static void cse_around_loop (rtx);
656 static void invalidate_skipped_set (rtx, rtx, void *);
657 static void invalidate_skipped_block (rtx);
658 static void cse_check_loop_start (rtx, rtx, void *);
659 static void cse_set_around_loop (rtx, rtx, rtx);
660 static rtx cse_basic_block (rtx, rtx, struct branch_path *, int);
661 static void count_reg_usage (rtx, int *, int);
662 static int check_for_label_ref (rtx *, void *);
663 extern void dump_class (struct table_elt*);
664 static struct cse_reg_info * get_cse_reg_info (unsigned int);
665 static int check_dependence (rtx *, void *);
667 static void flush_hash_table (void);
668 static bool insn_live_p (rtx, int *);
669 static bool set_live_p (rtx, rtx, int *);
670 static bool dead_libcall_p (rtx, int *);
671 static int cse_change_cc_mode (rtx *, void *);
672 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
673 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
676 #undef RTL_HOOKS_GEN_LOWPART
677 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
679 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
681 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
682 virtual regs here because the simplify_*_operation routines are called
683 by integrate.c, which is called before virtual register instantiation. */
685 static bool
686 fixed_base_plus_p (rtx x)
688 switch (GET_CODE (x))
690 case REG:
691 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
692 return true;
693 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
694 return true;
695 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
696 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
697 return true;
698 return false;
700 case PLUS:
701 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
702 return false;
703 return fixed_base_plus_p (XEXP (x, 0));
705 default:
706 return false;
710 /* Dump the expressions in the equivalence class indicated by CLASSP.
711 This function is used only for debugging. */
712 void
713 dump_class (struct table_elt *classp)
715 struct table_elt *elt;
717 fprintf (stderr, "Equivalence chain for ");
718 print_rtl (stderr, classp->exp);
719 fprintf (stderr, ": \n");
721 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
723 print_rtl (stderr, elt->exp);
724 fprintf (stderr, "\n");
728 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
730 static int
731 approx_reg_cost_1 (rtx *xp, void *data)
733 rtx x = *xp;
734 int *cost_p = data;
736 if (x && REG_P (x))
738 unsigned int regno = REGNO (x);
740 if (! CHEAP_REGNO (regno))
742 if (regno < FIRST_PSEUDO_REGISTER)
744 if (SMALL_REGISTER_CLASSES)
745 return 1;
746 *cost_p += 2;
748 else
749 *cost_p += 1;
753 return 0;
756 /* Return an estimate of the cost of the registers used in an rtx.
757 This is mostly the number of different REG expressions in the rtx;
758 however for some exceptions like fixed registers we use a cost of
759 0. If any other hard register reference occurs, return MAX_COST. */
761 static int
762 approx_reg_cost (rtx x)
764 int cost = 0;
766 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
767 return MAX_COST;
769 return cost;
772 /* Return a negative value if an rtx A, whose costs are given by COST_A
773 and REGCOST_A, is more desirable than an rtx B.
774 Return a positive value if A is less desirable, or 0 if the two are
775 equally good. */
776 static int
777 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
779 /* First, get rid of cases involving expressions that are entirely
780 unwanted. */
781 if (cost_a != cost_b)
783 if (cost_a == MAX_COST)
784 return 1;
785 if (cost_b == MAX_COST)
786 return -1;
789 /* Avoid extending lifetimes of hardregs. */
790 if (regcost_a != regcost_b)
792 if (regcost_a == MAX_COST)
793 return 1;
794 if (regcost_b == MAX_COST)
795 return -1;
798 /* Normal operation costs take precedence. */
799 if (cost_a != cost_b)
800 return cost_a - cost_b;
801 /* Only if these are identical consider effects on register pressure. */
802 if (regcost_a != regcost_b)
803 return regcost_a - regcost_b;
804 return 0;
807 /* Internal function, to compute cost when X is not a register; called
808 from COST macro to keep it simple. */
810 static int
811 notreg_cost (rtx x, enum rtx_code outer)
813 return ((GET_CODE (x) == SUBREG
814 && REG_P (SUBREG_REG (x))
815 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
816 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
817 && (GET_MODE_SIZE (GET_MODE (x))
818 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
819 && subreg_lowpart_p (x)
820 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
821 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
823 : rtx_cost (x, outer) * 2);
827 static struct cse_reg_info *
828 get_cse_reg_info (unsigned int regno)
830 struct cse_reg_info **hash_head = &reg_hash[REGHASH_FN (regno)];
831 struct cse_reg_info *p;
833 for (p = *hash_head; p != NULL; p = p->hash_next)
834 if (p->regno == regno)
835 break;
837 if (p == NULL)
839 /* Get a new cse_reg_info structure. */
840 if (cse_reg_info_free_list)
842 p = cse_reg_info_free_list;
843 cse_reg_info_free_list = p->next;
845 else
846 p = xmalloc (sizeof (struct cse_reg_info));
848 /* Insert into hash table. */
849 p->hash_next = *hash_head;
850 *hash_head = p;
852 /* Initialize it. */
853 p->reg_tick = 1;
854 p->reg_in_table = -1;
855 p->subreg_ticked = -1;
856 p->reg_qty = regno;
857 p->regno = regno;
858 p->next = cse_reg_info_used_list;
859 cse_reg_info_used_list = p;
860 if (!cse_reg_info_used_list_end)
861 cse_reg_info_used_list_end = p;
864 /* Cache this lookup; we tend to be looking up information about the
865 same register several times in a row. */
866 cached_regno = regno;
867 cached_cse_reg_info = p;
869 return p;
872 /* Clear the hash table and initialize each register with its own quantity,
873 for a new basic block. */
875 static void
876 new_basic_block (void)
878 int i;
880 next_qty = max_reg;
882 /* Clear out hash table state for this pass. */
884 memset (reg_hash, 0, sizeof reg_hash);
886 if (cse_reg_info_used_list)
888 cse_reg_info_used_list_end->next = cse_reg_info_free_list;
889 cse_reg_info_free_list = cse_reg_info_used_list;
890 cse_reg_info_used_list = cse_reg_info_used_list_end = 0;
892 cached_cse_reg_info = 0;
894 CLEAR_HARD_REG_SET (hard_regs_in_table);
896 /* The per-quantity values used to be initialized here, but it is
897 much faster to initialize each as it is made in `make_new_qty'. */
899 for (i = 0; i < HASH_SIZE; i++)
901 struct table_elt *first;
903 first = table[i];
904 if (first != NULL)
906 struct table_elt *last = first;
908 table[i] = NULL;
910 while (last->next_same_hash != NULL)
911 last = last->next_same_hash;
913 /* Now relink this hash entire chain into
914 the free element list. */
916 last->next_same_hash = free_element_chain;
917 free_element_chain = first;
921 #ifdef HAVE_cc0
922 prev_insn = 0;
923 prev_insn_cc0 = 0;
924 #endif
927 /* Say that register REG contains a quantity in mode MODE not in any
928 register before and initialize that quantity. */
930 static void
931 make_new_qty (unsigned int reg, enum machine_mode mode)
933 int q;
934 struct qty_table_elem *ent;
935 struct reg_eqv_elem *eqv;
937 if (next_qty >= max_qty)
938 abort ();
940 q = REG_QTY (reg) = next_qty++;
941 ent = &qty_table[q];
942 ent->first_reg = reg;
943 ent->last_reg = reg;
944 ent->mode = mode;
945 ent->const_rtx = ent->const_insn = NULL_RTX;
946 ent->comparison_code = UNKNOWN;
948 eqv = &reg_eqv_table[reg];
949 eqv->next = eqv->prev = -1;
952 /* Make reg NEW equivalent to reg OLD.
953 OLD is not changing; NEW is. */
955 static void
956 make_regs_eqv (unsigned int new, unsigned int old)
958 unsigned int lastr, firstr;
959 int q = REG_QTY (old);
960 struct qty_table_elem *ent;
962 ent = &qty_table[q];
964 /* Nothing should become eqv until it has a "non-invalid" qty number. */
965 if (! REGNO_QTY_VALID_P (old))
966 abort ();
968 REG_QTY (new) = q;
969 firstr = ent->first_reg;
970 lastr = ent->last_reg;
972 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
973 hard regs. Among pseudos, if NEW will live longer than any other reg
974 of the same qty, and that is beyond the current basic block,
975 make it the new canonical replacement for this qty. */
976 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
977 /* Certain fixed registers might be of the class NO_REGS. This means
978 that not only can they not be allocated by the compiler, but
979 they cannot be used in substitutions or canonicalizations
980 either. */
981 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
982 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
983 || (new >= FIRST_PSEUDO_REGISTER
984 && (firstr < FIRST_PSEUDO_REGISTER
985 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
986 || (uid_cuid[REGNO_FIRST_UID (new)]
987 < cse_basic_block_start))
988 && (uid_cuid[REGNO_LAST_UID (new)]
989 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
991 reg_eqv_table[firstr].prev = new;
992 reg_eqv_table[new].next = firstr;
993 reg_eqv_table[new].prev = -1;
994 ent->first_reg = new;
996 else
998 /* If NEW is a hard reg (known to be non-fixed), insert at end.
999 Otherwise, insert before any non-fixed hard regs that are at the
1000 end. Registers of class NO_REGS cannot be used as an
1001 equivalent for anything. */
1002 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1003 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1004 && new >= FIRST_PSEUDO_REGISTER)
1005 lastr = reg_eqv_table[lastr].prev;
1006 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1007 if (reg_eqv_table[lastr].next >= 0)
1008 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1009 else
1010 qty_table[q].last_reg = new;
1011 reg_eqv_table[lastr].next = new;
1012 reg_eqv_table[new].prev = lastr;
1016 /* Remove REG from its equivalence class. */
1018 static void
1019 delete_reg_equiv (unsigned int reg)
1021 struct qty_table_elem *ent;
1022 int q = REG_QTY (reg);
1023 int p, n;
1025 /* If invalid, do nothing. */
1026 if (q == (int) reg)
1027 return;
1029 ent = &qty_table[q];
1031 p = reg_eqv_table[reg].prev;
1032 n = reg_eqv_table[reg].next;
1034 if (n != -1)
1035 reg_eqv_table[n].prev = p;
1036 else
1037 ent->last_reg = p;
1038 if (p != -1)
1039 reg_eqv_table[p].next = n;
1040 else
1041 ent->first_reg = n;
1043 REG_QTY (reg) = reg;
1046 /* Remove any invalid expressions from the hash table
1047 that refer to any of the registers contained in expression X.
1049 Make sure that newly inserted references to those registers
1050 as subexpressions will be considered valid.
1052 mention_regs is not called when a register itself
1053 is being stored in the table.
1055 Return 1 if we have done something that may have changed the hash code
1056 of X. */
1058 static int
1059 mention_regs (rtx x)
1061 enum rtx_code code;
1062 int i, j;
1063 const char *fmt;
1064 int changed = 0;
1066 if (x == 0)
1067 return 0;
1069 code = GET_CODE (x);
1070 if (code == REG)
1072 unsigned int regno = REGNO (x);
1073 unsigned int endregno
1074 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1075 : hard_regno_nregs[regno][GET_MODE (x)]);
1076 unsigned int i;
1078 for (i = regno; i < endregno; i++)
1080 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1081 remove_invalid_refs (i);
1083 REG_IN_TABLE (i) = REG_TICK (i);
1084 SUBREG_TICKED (i) = -1;
1087 return 0;
1090 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1091 pseudo if they don't use overlapping words. We handle only pseudos
1092 here for simplicity. */
1093 if (code == SUBREG && REG_P (SUBREG_REG (x))
1094 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1096 unsigned int i = REGNO (SUBREG_REG (x));
1098 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1100 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1101 the last store to this register really stored into this
1102 subreg, then remove the memory of this subreg.
1103 Otherwise, remove any memory of the entire register and
1104 all its subregs from the table. */
1105 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1106 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1107 remove_invalid_refs (i);
1108 else
1109 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1112 REG_IN_TABLE (i) = REG_TICK (i);
1113 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1114 return 0;
1117 /* If X is a comparison or a COMPARE and either operand is a register
1118 that does not have a quantity, give it one. This is so that a later
1119 call to record_jump_equiv won't cause X to be assigned a different
1120 hash code and not found in the table after that call.
1122 It is not necessary to do this here, since rehash_using_reg can
1123 fix up the table later, but doing this here eliminates the need to
1124 call that expensive function in the most common case where the only
1125 use of the register is in the comparison. */
1127 if (code == COMPARE || COMPARISON_P (x))
1129 if (REG_P (XEXP (x, 0))
1130 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1131 if (insert_regs (XEXP (x, 0), NULL, 0))
1133 rehash_using_reg (XEXP (x, 0));
1134 changed = 1;
1137 if (REG_P (XEXP (x, 1))
1138 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1139 if (insert_regs (XEXP (x, 1), NULL, 0))
1141 rehash_using_reg (XEXP (x, 1));
1142 changed = 1;
1146 fmt = GET_RTX_FORMAT (code);
1147 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1148 if (fmt[i] == 'e')
1149 changed |= mention_regs (XEXP (x, i));
1150 else if (fmt[i] == 'E')
1151 for (j = 0; j < XVECLEN (x, i); j++)
1152 changed |= mention_regs (XVECEXP (x, i, j));
1154 return changed;
1157 /* Update the register quantities for inserting X into the hash table
1158 with a value equivalent to CLASSP.
1159 (If the class does not contain a REG, it is irrelevant.)
1160 If MODIFIED is nonzero, X is a destination; it is being modified.
1161 Note that delete_reg_equiv should be called on a register
1162 before insert_regs is done on that register with MODIFIED != 0.
1164 Nonzero value means that elements of reg_qty have changed
1165 so X's hash code may be different. */
1167 static int
1168 insert_regs (rtx x, struct table_elt *classp, int modified)
1170 if (REG_P (x))
1172 unsigned int regno = REGNO (x);
1173 int qty_valid;
1175 /* If REGNO is in the equivalence table already but is of the
1176 wrong mode for that equivalence, don't do anything here. */
1178 qty_valid = REGNO_QTY_VALID_P (regno);
1179 if (qty_valid)
1181 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1183 if (ent->mode != GET_MODE (x))
1184 return 0;
1187 if (modified || ! qty_valid)
1189 if (classp)
1190 for (classp = classp->first_same_value;
1191 classp != 0;
1192 classp = classp->next_same_value)
1193 if (REG_P (classp->exp)
1194 && GET_MODE (classp->exp) == GET_MODE (x))
1196 make_regs_eqv (regno, REGNO (classp->exp));
1197 return 1;
1200 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1201 than REG_IN_TABLE to find out if there was only a single preceding
1202 invalidation - for the SUBREG - or another one, which would be
1203 for the full register. However, if we find here that REG_TICK
1204 indicates that the register is invalid, it means that it has
1205 been invalidated in a separate operation. The SUBREG might be used
1206 now (then this is a recursive call), or we might use the full REG
1207 now and a SUBREG of it later. So bump up REG_TICK so that
1208 mention_regs will do the right thing. */
1209 if (! modified
1210 && REG_IN_TABLE (regno) >= 0
1211 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1212 REG_TICK (regno)++;
1213 make_new_qty (regno, GET_MODE (x));
1214 return 1;
1217 return 0;
1220 /* If X is a SUBREG, we will likely be inserting the inner register in the
1221 table. If that register doesn't have an assigned quantity number at
1222 this point but does later, the insertion that we will be doing now will
1223 not be accessible because its hash code will have changed. So assign
1224 a quantity number now. */
1226 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1227 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1229 insert_regs (SUBREG_REG (x), NULL, 0);
1230 mention_regs (x);
1231 return 1;
1233 else
1234 return mention_regs (x);
1237 /* Look in or update the hash table. */
1239 /* Remove table element ELT from use in the table.
1240 HASH is its hash code, made using the HASH macro.
1241 It's an argument because often that is known in advance
1242 and we save much time not recomputing it. */
1244 static void
1245 remove_from_table (struct table_elt *elt, unsigned int hash)
1247 if (elt == 0)
1248 return;
1250 /* Mark this element as removed. See cse_insn. */
1251 elt->first_same_value = 0;
1253 /* Remove the table element from its equivalence class. */
1256 struct table_elt *prev = elt->prev_same_value;
1257 struct table_elt *next = elt->next_same_value;
1259 if (next)
1260 next->prev_same_value = prev;
1262 if (prev)
1263 prev->next_same_value = next;
1264 else
1266 struct table_elt *newfirst = next;
1267 while (next)
1269 next->first_same_value = newfirst;
1270 next = next->next_same_value;
1275 /* Remove the table element from its hash bucket. */
1278 struct table_elt *prev = elt->prev_same_hash;
1279 struct table_elt *next = elt->next_same_hash;
1281 if (next)
1282 next->prev_same_hash = prev;
1284 if (prev)
1285 prev->next_same_hash = next;
1286 else if (table[hash] == elt)
1287 table[hash] = next;
1288 else
1290 /* This entry is not in the proper hash bucket. This can happen
1291 when two classes were merged by `merge_equiv_classes'. Search
1292 for the hash bucket that it heads. This happens only very
1293 rarely, so the cost is acceptable. */
1294 for (hash = 0; hash < HASH_SIZE; hash++)
1295 if (table[hash] == elt)
1296 table[hash] = next;
1300 /* Remove the table element from its related-value circular chain. */
1302 if (elt->related_value != 0 && elt->related_value != elt)
1304 struct table_elt *p = elt->related_value;
1306 while (p->related_value != elt)
1307 p = p->related_value;
1308 p->related_value = elt->related_value;
1309 if (p->related_value == p)
1310 p->related_value = 0;
1313 /* Now add it to the free element chain. */
1314 elt->next_same_hash = free_element_chain;
1315 free_element_chain = elt;
1318 /* Look up X in the hash table and return its table element,
1319 or 0 if X is not in the table.
1321 MODE is the machine-mode of X, or if X is an integer constant
1322 with VOIDmode then MODE is the mode with which X will be used.
1324 Here we are satisfied to find an expression whose tree structure
1325 looks like X. */
1327 static struct table_elt *
1328 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1330 struct table_elt *p;
1332 for (p = table[hash]; p; p = p->next_same_hash)
1333 if (mode == p->mode && ((x == p->exp && REG_P (x))
1334 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1335 return p;
1337 return 0;
1340 /* Like `lookup' but don't care whether the table element uses invalid regs.
1341 Also ignore discrepancies in the machine mode of a register. */
1343 static struct table_elt *
1344 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1346 struct table_elt *p;
1348 if (REG_P (x))
1350 unsigned int regno = REGNO (x);
1352 /* Don't check the machine mode when comparing registers;
1353 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1354 for (p = table[hash]; p; p = p->next_same_hash)
1355 if (REG_P (p->exp)
1356 && REGNO (p->exp) == regno)
1357 return p;
1359 else
1361 for (p = table[hash]; p; p = p->next_same_hash)
1362 if (mode == p->mode
1363 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1364 return p;
1367 return 0;
1370 /* Look for an expression equivalent to X and with code CODE.
1371 If one is found, return that expression. */
1373 static rtx
1374 lookup_as_function (rtx x, enum rtx_code code)
1376 struct table_elt *p
1377 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1379 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1380 long as we are narrowing. So if we looked in vain for a mode narrower
1381 than word_mode before, look for word_mode now. */
1382 if (p == 0 && code == CONST_INT
1383 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1385 x = copy_rtx (x);
1386 PUT_MODE (x, word_mode);
1387 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1390 if (p == 0)
1391 return 0;
1393 for (p = p->first_same_value; p; p = p->next_same_value)
1394 if (GET_CODE (p->exp) == code
1395 /* Make sure this is a valid entry in the table. */
1396 && exp_equiv_p (p->exp, p->exp, 1, false))
1397 return p->exp;
1399 return 0;
1402 /* Insert X in the hash table, assuming HASH is its hash code
1403 and CLASSP is an element of the class it should go in
1404 (or 0 if a new class should be made).
1405 It is inserted at the proper position to keep the class in
1406 the order cheapest first.
1408 MODE is the machine-mode of X, or if X is an integer constant
1409 with VOIDmode then MODE is the mode with which X will be used.
1411 For elements of equal cheapness, the most recent one
1412 goes in front, except that the first element in the list
1413 remains first unless a cheaper element is added. The order of
1414 pseudo-registers does not matter, as canon_reg will be called to
1415 find the cheapest when a register is retrieved from the table.
1417 The in_memory field in the hash table element is set to 0.
1418 The caller must set it nonzero if appropriate.
1420 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1421 and if insert_regs returns a nonzero value
1422 you must then recompute its hash code before calling here.
1424 If necessary, update table showing constant values of quantities. */
1426 #define CHEAPER(X, Y) \
1427 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1429 static struct table_elt *
1430 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1432 struct table_elt *elt;
1434 /* If X is a register and we haven't made a quantity for it,
1435 something is wrong. */
1436 if (REG_P (x) && ! REGNO_QTY_VALID_P (REGNO (x)))
1437 abort ();
1439 /* If X is a hard register, show it is being put in the table. */
1440 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1442 unsigned int regno = REGNO (x);
1443 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1444 unsigned int i;
1446 for (i = regno; i < endregno; i++)
1447 SET_HARD_REG_BIT (hard_regs_in_table, i);
1450 /* Put an element for X into the right hash bucket. */
1452 elt = free_element_chain;
1453 if (elt)
1454 free_element_chain = elt->next_same_hash;
1455 else
1457 n_elements_made++;
1458 elt = xmalloc (sizeof (struct table_elt));
1461 elt->exp = x;
1462 elt->canon_exp = NULL_RTX;
1463 elt->cost = COST (x);
1464 elt->regcost = approx_reg_cost (x);
1465 elt->next_same_value = 0;
1466 elt->prev_same_value = 0;
1467 elt->next_same_hash = table[hash];
1468 elt->prev_same_hash = 0;
1469 elt->related_value = 0;
1470 elt->in_memory = 0;
1471 elt->mode = mode;
1472 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1474 if (table[hash])
1475 table[hash]->prev_same_hash = elt;
1476 table[hash] = elt;
1478 /* Put it into the proper value-class. */
1479 if (classp)
1481 classp = classp->first_same_value;
1482 if (CHEAPER (elt, classp))
1483 /* Insert at the head of the class. */
1485 struct table_elt *p;
1486 elt->next_same_value = classp;
1487 classp->prev_same_value = elt;
1488 elt->first_same_value = elt;
1490 for (p = classp; p; p = p->next_same_value)
1491 p->first_same_value = elt;
1493 else
1495 /* Insert not at head of the class. */
1496 /* Put it after the last element cheaper than X. */
1497 struct table_elt *p, *next;
1499 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1500 p = next);
1502 /* Put it after P and before NEXT. */
1503 elt->next_same_value = next;
1504 if (next)
1505 next->prev_same_value = elt;
1507 elt->prev_same_value = p;
1508 p->next_same_value = elt;
1509 elt->first_same_value = classp;
1512 else
1513 elt->first_same_value = elt;
1515 /* If this is a constant being set equivalent to a register or a register
1516 being set equivalent to a constant, note the constant equivalence.
1518 If this is a constant, it cannot be equivalent to a different constant,
1519 and a constant is the only thing that can be cheaper than a register. So
1520 we know the register is the head of the class (before the constant was
1521 inserted).
1523 If this is a register that is not already known equivalent to a
1524 constant, we must check the entire class.
1526 If this is a register that is already known equivalent to an insn,
1527 update the qtys `const_insn' to show that `this_insn' is the latest
1528 insn making that quantity equivalent to the constant. */
1530 if (elt->is_const && classp && REG_P (classp->exp)
1531 && !REG_P (x))
1533 int exp_q = REG_QTY (REGNO (classp->exp));
1534 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1536 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1537 exp_ent->const_insn = this_insn;
1540 else if (REG_P (x)
1541 && classp
1542 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1543 && ! elt->is_const)
1545 struct table_elt *p;
1547 for (p = classp; p != 0; p = p->next_same_value)
1549 if (p->is_const && !REG_P (p->exp))
1551 int x_q = REG_QTY (REGNO (x));
1552 struct qty_table_elem *x_ent = &qty_table[x_q];
1554 x_ent->const_rtx
1555 = gen_lowpart (GET_MODE (x), p->exp);
1556 x_ent->const_insn = this_insn;
1557 break;
1562 else if (REG_P (x)
1563 && qty_table[REG_QTY (REGNO (x))].const_rtx
1564 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1565 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1567 /* If this is a constant with symbolic value,
1568 and it has a term with an explicit integer value,
1569 link it up with related expressions. */
1570 if (GET_CODE (x) == CONST)
1572 rtx subexp = get_related_value (x);
1573 unsigned subhash;
1574 struct table_elt *subelt, *subelt_prev;
1576 if (subexp != 0)
1578 /* Get the integer-free subexpression in the hash table. */
1579 subhash = SAFE_HASH (subexp, mode);
1580 subelt = lookup (subexp, subhash, mode);
1581 if (subelt == 0)
1582 subelt = insert (subexp, NULL, subhash, mode);
1583 /* Initialize SUBELT's circular chain if it has none. */
1584 if (subelt->related_value == 0)
1585 subelt->related_value = subelt;
1586 /* Find the element in the circular chain that precedes SUBELT. */
1587 subelt_prev = subelt;
1588 while (subelt_prev->related_value != subelt)
1589 subelt_prev = subelt_prev->related_value;
1590 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1591 This way the element that follows SUBELT is the oldest one. */
1592 elt->related_value = subelt_prev->related_value;
1593 subelt_prev->related_value = elt;
1597 return elt;
1600 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1601 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1602 the two classes equivalent.
1604 CLASS1 will be the surviving class; CLASS2 should not be used after this
1605 call.
1607 Any invalid entries in CLASS2 will not be copied. */
1609 static void
1610 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1612 struct table_elt *elt, *next, *new;
1614 /* Ensure we start with the head of the classes. */
1615 class1 = class1->first_same_value;
1616 class2 = class2->first_same_value;
1618 /* If they were already equal, forget it. */
1619 if (class1 == class2)
1620 return;
1622 for (elt = class2; elt; elt = next)
1624 unsigned int hash;
1625 rtx exp = elt->exp;
1626 enum machine_mode mode = elt->mode;
1628 next = elt->next_same_value;
1630 /* Remove old entry, make a new one in CLASS1's class.
1631 Don't do this for invalid entries as we cannot find their
1632 hash code (it also isn't necessary). */
1633 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1635 bool need_rehash = false;
1637 hash_arg_in_memory = 0;
1638 hash = HASH (exp, mode);
1640 if (REG_P (exp))
1642 need_rehash = (unsigned) REG_QTY (REGNO (exp)) != REGNO (exp);
1643 delete_reg_equiv (REGNO (exp));
1646 remove_from_table (elt, hash);
1648 if (insert_regs (exp, class1, 0) || need_rehash)
1650 rehash_using_reg (exp);
1651 hash = HASH (exp, mode);
1653 new = insert (exp, class1, hash, mode);
1654 new->in_memory = hash_arg_in_memory;
1659 /* Flush the entire hash table. */
1661 static void
1662 flush_hash_table (void)
1664 int i;
1665 struct table_elt *p;
1667 for (i = 0; i < HASH_SIZE; i++)
1668 for (p = table[i]; p; p = table[i])
1670 /* Note that invalidate can remove elements
1671 after P in the current hash chain. */
1672 if (REG_P (p->exp))
1673 invalidate (p->exp, p->mode);
1674 else
1675 remove_from_table (p, i);
1679 /* Function called for each rtx to check whether true dependence exist. */
1680 struct check_dependence_data
1682 enum machine_mode mode;
1683 rtx exp;
1684 rtx addr;
1687 static int
1688 check_dependence (rtx *x, void *data)
1690 struct check_dependence_data *d = (struct check_dependence_data *) data;
1691 if (*x && MEM_P (*x))
1692 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1693 cse_rtx_varies_p);
1694 else
1695 return 0;
1698 /* Remove from the hash table, or mark as invalid, all expressions whose
1699 values could be altered by storing in X. X is a register, a subreg, or
1700 a memory reference with nonvarying address (because, when a memory
1701 reference with a varying address is stored in, all memory references are
1702 removed by invalidate_memory so specific invalidation is superfluous).
1703 FULL_MODE, if not VOIDmode, indicates that this much should be
1704 invalidated instead of just the amount indicated by the mode of X. This
1705 is only used for bitfield stores into memory.
1707 A nonvarying address may be just a register or just a symbol reference,
1708 or it may be either of those plus a numeric offset. */
1710 static void
1711 invalidate (rtx x, enum machine_mode full_mode)
1713 int i;
1714 struct table_elt *p;
1715 rtx addr;
1717 switch (GET_CODE (x))
1719 case REG:
1721 /* If X is a register, dependencies on its contents are recorded
1722 through the qty number mechanism. Just change the qty number of
1723 the register, mark it as invalid for expressions that refer to it,
1724 and remove it itself. */
1725 unsigned int regno = REGNO (x);
1726 unsigned int hash = HASH (x, GET_MODE (x));
1728 /* Remove REGNO from any quantity list it might be on and indicate
1729 that its value might have changed. If it is a pseudo, remove its
1730 entry from the hash table.
1732 For a hard register, we do the first two actions above for any
1733 additional hard registers corresponding to X. Then, if any of these
1734 registers are in the table, we must remove any REG entries that
1735 overlap these registers. */
1737 delete_reg_equiv (regno);
1738 REG_TICK (regno)++;
1739 SUBREG_TICKED (regno) = -1;
1741 if (regno >= FIRST_PSEUDO_REGISTER)
1743 /* Because a register can be referenced in more than one mode,
1744 we might have to remove more than one table entry. */
1745 struct table_elt *elt;
1747 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1748 remove_from_table (elt, hash);
1750 else
1752 HOST_WIDE_INT in_table
1753 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1754 unsigned int endregno
1755 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1756 unsigned int tregno, tendregno, rn;
1757 struct table_elt *p, *next;
1759 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1761 for (rn = regno + 1; rn < endregno; rn++)
1763 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1764 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1765 delete_reg_equiv (rn);
1766 REG_TICK (rn)++;
1767 SUBREG_TICKED (rn) = -1;
1770 if (in_table)
1771 for (hash = 0; hash < HASH_SIZE; hash++)
1772 for (p = table[hash]; p; p = next)
1774 next = p->next_same_hash;
1776 if (!REG_P (p->exp)
1777 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1778 continue;
1780 tregno = REGNO (p->exp);
1781 tendregno
1782 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1783 if (tendregno > regno && tregno < endregno)
1784 remove_from_table (p, hash);
1788 return;
1790 case SUBREG:
1791 invalidate (SUBREG_REG (x), VOIDmode);
1792 return;
1794 case PARALLEL:
1795 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1796 invalidate (XVECEXP (x, 0, i), VOIDmode);
1797 return;
1799 case EXPR_LIST:
1800 /* This is part of a disjoint return value; extract the location in
1801 question ignoring the offset. */
1802 invalidate (XEXP (x, 0), VOIDmode);
1803 return;
1805 case MEM:
1806 addr = canon_rtx (get_addr (XEXP (x, 0)));
1807 /* Calculate the canonical version of X here so that
1808 true_dependence doesn't generate new RTL for X on each call. */
1809 x = canon_rtx (x);
1811 /* Remove all hash table elements that refer to overlapping pieces of
1812 memory. */
1813 if (full_mode == VOIDmode)
1814 full_mode = GET_MODE (x);
1816 for (i = 0; i < HASH_SIZE; i++)
1818 struct table_elt *next;
1820 for (p = table[i]; p; p = next)
1822 next = p->next_same_hash;
1823 if (p->in_memory)
1825 struct check_dependence_data d;
1827 /* Just canonicalize the expression once;
1828 otherwise each time we call invalidate
1829 true_dependence will canonicalize the
1830 expression again. */
1831 if (!p->canon_exp)
1832 p->canon_exp = canon_rtx (p->exp);
1833 d.exp = x;
1834 d.addr = addr;
1835 d.mode = full_mode;
1836 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1837 remove_from_table (p, i);
1841 return;
1843 default:
1844 abort ();
1848 /* Remove all expressions that refer to register REGNO,
1849 since they are already invalid, and we are about to
1850 mark that register valid again and don't want the old
1851 expressions to reappear as valid. */
1853 static void
1854 remove_invalid_refs (unsigned int regno)
1856 unsigned int i;
1857 struct table_elt *p, *next;
1859 for (i = 0; i < HASH_SIZE; i++)
1860 for (p = table[i]; p; p = next)
1862 next = p->next_same_hash;
1863 if (!REG_P (p->exp)
1864 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1865 remove_from_table (p, i);
1869 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1870 and mode MODE. */
1871 static void
1872 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1873 enum machine_mode mode)
1875 unsigned int i;
1876 struct table_elt *p, *next;
1877 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1879 for (i = 0; i < HASH_SIZE; i++)
1880 for (p = table[i]; p; p = next)
1882 rtx exp = p->exp;
1883 next = p->next_same_hash;
1885 if (!REG_P (exp)
1886 && (GET_CODE (exp) != SUBREG
1887 || !REG_P (SUBREG_REG (exp))
1888 || REGNO (SUBREG_REG (exp)) != regno
1889 || (((SUBREG_BYTE (exp)
1890 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1891 && SUBREG_BYTE (exp) <= end))
1892 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1893 remove_from_table (p, i);
1897 /* Recompute the hash codes of any valid entries in the hash table that
1898 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1900 This is called when we make a jump equivalence. */
1902 static void
1903 rehash_using_reg (rtx x)
1905 unsigned int i;
1906 struct table_elt *p, *next;
1907 unsigned hash;
1909 if (GET_CODE (x) == SUBREG)
1910 x = SUBREG_REG (x);
1912 /* If X is not a register or if the register is known not to be in any
1913 valid entries in the table, we have no work to do. */
1915 if (!REG_P (x)
1916 || REG_IN_TABLE (REGNO (x)) < 0
1917 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1918 return;
1920 /* Scan all hash chains looking for valid entries that mention X.
1921 If we find one and it is in the wrong hash chain, move it. */
1923 for (i = 0; i < HASH_SIZE; i++)
1924 for (p = table[i]; p; p = next)
1926 next = p->next_same_hash;
1927 if (reg_mentioned_p (x, p->exp)
1928 && exp_equiv_p (p->exp, p->exp, 1, false)
1929 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1931 if (p->next_same_hash)
1932 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1934 if (p->prev_same_hash)
1935 p->prev_same_hash->next_same_hash = p->next_same_hash;
1936 else
1937 table[i] = p->next_same_hash;
1939 p->next_same_hash = table[hash];
1940 p->prev_same_hash = 0;
1941 if (table[hash])
1942 table[hash]->prev_same_hash = p;
1943 table[hash] = p;
1948 /* Remove from the hash table any expression that is a call-clobbered
1949 register. Also update their TICK values. */
1951 static void
1952 invalidate_for_call (void)
1954 unsigned int regno, endregno;
1955 unsigned int i;
1956 unsigned hash;
1957 struct table_elt *p, *next;
1958 int in_table = 0;
1960 /* Go through all the hard registers. For each that is clobbered in
1961 a CALL_INSN, remove the register from quantity chains and update
1962 reg_tick if defined. Also see if any of these registers is currently
1963 in the table. */
1965 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1966 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
1968 delete_reg_equiv (regno);
1969 if (REG_TICK (regno) >= 0)
1971 REG_TICK (regno)++;
1972 SUBREG_TICKED (regno) = -1;
1975 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
1978 /* In the case where we have no call-clobbered hard registers in the
1979 table, we are done. Otherwise, scan the table and remove any
1980 entry that overlaps a call-clobbered register. */
1982 if (in_table)
1983 for (hash = 0; hash < HASH_SIZE; hash++)
1984 for (p = table[hash]; p; p = next)
1986 next = p->next_same_hash;
1988 if (!REG_P (p->exp)
1989 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1990 continue;
1992 regno = REGNO (p->exp);
1993 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
1995 for (i = regno; i < endregno; i++)
1996 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
1998 remove_from_table (p, hash);
1999 break;
2004 /* Given an expression X of type CONST,
2005 and ELT which is its table entry (or 0 if it
2006 is not in the hash table),
2007 return an alternate expression for X as a register plus integer.
2008 If none can be found, return 0. */
2010 static rtx
2011 use_related_value (rtx x, struct table_elt *elt)
2013 struct table_elt *relt = 0;
2014 struct table_elt *p, *q;
2015 HOST_WIDE_INT offset;
2017 /* First, is there anything related known?
2018 If we have a table element, we can tell from that.
2019 Otherwise, must look it up. */
2021 if (elt != 0 && elt->related_value != 0)
2022 relt = elt;
2023 else if (elt == 0 && GET_CODE (x) == CONST)
2025 rtx subexp = get_related_value (x);
2026 if (subexp != 0)
2027 relt = lookup (subexp,
2028 SAFE_HASH (subexp, GET_MODE (subexp)),
2029 GET_MODE (subexp));
2032 if (relt == 0)
2033 return 0;
2035 /* Search all related table entries for one that has an
2036 equivalent register. */
2038 p = relt;
2039 while (1)
2041 /* This loop is strange in that it is executed in two different cases.
2042 The first is when X is already in the table. Then it is searching
2043 the RELATED_VALUE list of X's class (RELT). The second case is when
2044 X is not in the table. Then RELT points to a class for the related
2045 value.
2047 Ensure that, whatever case we are in, that we ignore classes that have
2048 the same value as X. */
2050 if (rtx_equal_p (x, p->exp))
2051 q = 0;
2052 else
2053 for (q = p->first_same_value; q; q = q->next_same_value)
2054 if (REG_P (q->exp))
2055 break;
2057 if (q)
2058 break;
2060 p = p->related_value;
2062 /* We went all the way around, so there is nothing to be found.
2063 Alternatively, perhaps RELT was in the table for some other reason
2064 and it has no related values recorded. */
2065 if (p == relt || p == 0)
2066 break;
2069 if (q == 0)
2070 return 0;
2072 offset = (get_integer_term (x) - get_integer_term (p->exp));
2073 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2074 return plus_constant (q->exp, offset);
2077 /* Hash a string. Just add its bytes up. */
2078 static inline unsigned
2079 hash_rtx_string (const char *ps)
2081 unsigned hash = 0;
2082 const unsigned char *p = (const unsigned char *) ps;
2084 if (p)
2085 while (*p)
2086 hash += *p++;
2088 return hash;
2091 /* Hash an rtx. We are careful to make sure the value is never negative.
2092 Equivalent registers hash identically.
2093 MODE is used in hashing for CONST_INTs only;
2094 otherwise the mode of X is used.
2096 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2098 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2099 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2101 Note that cse_insn knows that the hash code of a MEM expression
2102 is just (int) MEM plus the hash code of the address. */
2104 unsigned
2105 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2106 int *hash_arg_in_memory_p, bool have_reg_qty)
2108 int i, j;
2109 unsigned hash = 0;
2110 enum rtx_code code;
2111 const char *fmt;
2113 /* Used to turn recursion into iteration. We can't rely on GCC's
2114 tail-recursion elimination since we need to keep accumulating values
2115 in HASH. */
2116 repeat:
2117 if (x == 0)
2118 return hash;
2120 code = GET_CODE (x);
2121 switch (code)
2123 case REG:
2125 unsigned int regno = REGNO (x);
2127 if (!reload_completed)
2129 /* On some machines, we can't record any non-fixed hard register,
2130 because extending its life will cause reload problems. We
2131 consider ap, fp, sp, gp to be fixed for this purpose.
2133 We also consider CCmode registers to be fixed for this purpose;
2134 failure to do so leads to failure to simplify 0<100 type of
2135 conditionals.
2137 On all machines, we can't record any global registers.
2138 Nor should we record any register that is in a small
2139 class, as defined by CLASS_LIKELY_SPILLED_P. */
2140 bool record;
2142 if (regno >= FIRST_PSEUDO_REGISTER)
2143 record = true;
2144 else if (x == frame_pointer_rtx
2145 || x == hard_frame_pointer_rtx
2146 || x == arg_pointer_rtx
2147 || x == stack_pointer_rtx
2148 || x == pic_offset_table_rtx)
2149 record = true;
2150 else if (global_regs[regno])
2151 record = false;
2152 else if (fixed_regs[regno])
2153 record = true;
2154 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2155 record = true;
2156 else if (SMALL_REGISTER_CLASSES)
2157 record = false;
2158 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2159 record = false;
2160 else
2161 record = true;
2163 if (!record)
2165 *do_not_record_p = 1;
2166 return 0;
2170 hash += ((unsigned int) REG << 7);
2171 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2172 return hash;
2175 /* We handle SUBREG of a REG specially because the underlying
2176 reg changes its hash value with every value change; we don't
2177 want to have to forget unrelated subregs when one subreg changes. */
2178 case SUBREG:
2180 if (REG_P (SUBREG_REG (x)))
2182 hash += (((unsigned int) SUBREG << 7)
2183 + REGNO (SUBREG_REG (x))
2184 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2185 return hash;
2187 break;
2190 case CONST_INT:
2191 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2192 + (unsigned int) INTVAL (x));
2193 return hash;
2195 case CONST_DOUBLE:
2196 /* This is like the general case, except that it only counts
2197 the integers representing the constant. */
2198 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2199 if (GET_MODE (x) != VOIDmode)
2200 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2201 else
2202 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2203 + (unsigned int) CONST_DOUBLE_HIGH (x));
2204 return hash;
2206 case CONST_VECTOR:
2208 int units;
2209 rtx elt;
2211 units = CONST_VECTOR_NUNITS (x);
2213 for (i = 0; i < units; ++i)
2215 elt = CONST_VECTOR_ELT (x, i);
2216 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2217 hash_arg_in_memory_p, have_reg_qty);
2220 return hash;
2223 /* Assume there is only one rtx object for any given label. */
2224 case LABEL_REF:
2225 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2226 differences and differences between each stage's debugging dumps. */
2227 hash += (((unsigned int) LABEL_REF << 7)
2228 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2229 return hash;
2231 case SYMBOL_REF:
2233 /* Don't hash on the symbol's address to avoid bootstrap differences.
2234 Different hash values may cause expressions to be recorded in
2235 different orders and thus different registers to be used in the
2236 final assembler. This also avoids differences in the dump files
2237 between various stages. */
2238 unsigned int h = 0;
2239 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2241 while (*p)
2242 h += (h << 7) + *p++; /* ??? revisit */
2244 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2245 return hash;
2248 case MEM:
2249 /* We don't record if marked volatile or if BLKmode since we don't
2250 know the size of the move. */
2251 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2253 *do_not_record_p = 1;
2254 return 0;
2256 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2257 *hash_arg_in_memory_p = 1;
2259 /* Now that we have already found this special case,
2260 might as well speed it up as much as possible. */
2261 hash += (unsigned) MEM;
2262 x = XEXP (x, 0);
2263 goto repeat;
2265 case USE:
2266 /* A USE that mentions non-volatile memory needs special
2267 handling since the MEM may be BLKmode which normally
2268 prevents an entry from being made. Pure calls are
2269 marked by a USE which mentions BLKmode memory.
2270 See calls.c:emit_call_1. */
2271 if (MEM_P (XEXP (x, 0))
2272 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2274 hash += (unsigned) USE;
2275 x = XEXP (x, 0);
2277 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2278 *hash_arg_in_memory_p = 1;
2280 /* Now that we have already found this special case,
2281 might as well speed it up as much as possible. */
2282 hash += (unsigned) MEM;
2283 x = XEXP (x, 0);
2284 goto repeat;
2286 break;
2288 case PRE_DEC:
2289 case PRE_INC:
2290 case POST_DEC:
2291 case POST_INC:
2292 case PRE_MODIFY:
2293 case POST_MODIFY:
2294 case PC:
2295 case CC0:
2296 case CALL:
2297 case UNSPEC_VOLATILE:
2298 *do_not_record_p = 1;
2299 return 0;
2301 case ASM_OPERANDS:
2302 if (MEM_VOLATILE_P (x))
2304 *do_not_record_p = 1;
2305 return 0;
2307 else
2309 /* We don't want to take the filename and line into account. */
2310 hash += (unsigned) code + (unsigned) GET_MODE (x)
2311 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2312 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2313 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2315 if (ASM_OPERANDS_INPUT_LENGTH (x))
2317 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2319 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2320 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2321 do_not_record_p, hash_arg_in_memory_p,
2322 have_reg_qty)
2323 + hash_rtx_string
2324 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2327 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2328 x = ASM_OPERANDS_INPUT (x, 0);
2329 mode = GET_MODE (x);
2330 goto repeat;
2333 return hash;
2335 break;
2337 default:
2338 break;
2341 i = GET_RTX_LENGTH (code) - 1;
2342 hash += (unsigned) code + (unsigned) GET_MODE (x);
2343 fmt = GET_RTX_FORMAT (code);
2344 for (; i >= 0; i--)
2346 if (fmt[i] == 'e')
2348 /* If we are about to do the last recursive call
2349 needed at this level, change it into iteration.
2350 This function is called enough to be worth it. */
2351 if (i == 0)
2353 x = XEXP (x, i);
2354 goto repeat;
2357 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2358 hash_arg_in_memory_p, have_reg_qty);
2361 else if (fmt[i] == 'E')
2362 for (j = 0; j < XVECLEN (x, i); j++)
2364 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2365 hash_arg_in_memory_p, have_reg_qty);
2368 else if (fmt[i] == 's')
2369 hash += hash_rtx_string (XSTR (x, i));
2370 else if (fmt[i] == 'i')
2371 hash += (unsigned int) XINT (x, i);
2372 else if (fmt[i] == '0' || fmt[i] == 't')
2373 /* Unused. */
2375 else
2376 abort ();
2379 return hash;
2382 /* Hash an rtx X for cse via hash_rtx.
2383 Stores 1 in do_not_record if any subexpression is volatile.
2384 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2385 does not have the RTX_UNCHANGING_P bit set. */
2387 static inline unsigned
2388 canon_hash (rtx x, enum machine_mode mode)
2390 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2393 /* Like canon_hash but with no side effects, i.e. do_not_record
2394 and hash_arg_in_memory are not changed. */
2396 static inline unsigned
2397 safe_hash (rtx x, enum machine_mode mode)
2399 int dummy_do_not_record;
2400 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2403 /* Return 1 iff X and Y would canonicalize into the same thing,
2404 without actually constructing the canonicalization of either one.
2405 If VALIDATE is nonzero,
2406 we assume X is an expression being processed from the rtl
2407 and Y was found in the hash table. We check register refs
2408 in Y for being marked as valid.
2410 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2413 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2415 int i, j;
2416 enum rtx_code code;
2417 const char *fmt;
2419 /* Note: it is incorrect to assume an expression is equivalent to itself
2420 if VALIDATE is nonzero. */
2421 if (x == y && !validate)
2422 return 1;
2424 if (x == 0 || y == 0)
2425 return x == y;
2427 code = GET_CODE (x);
2428 if (code != GET_CODE (y))
2429 return 0;
2431 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2432 if (GET_MODE (x) != GET_MODE (y))
2433 return 0;
2435 switch (code)
2437 case PC:
2438 case CC0:
2439 case CONST_INT:
2440 return x == y;
2442 case LABEL_REF:
2443 return XEXP (x, 0) == XEXP (y, 0);
2445 case SYMBOL_REF:
2446 return XSTR (x, 0) == XSTR (y, 0);
2448 case REG:
2449 if (for_gcse)
2450 return REGNO (x) == REGNO (y);
2451 else
2453 unsigned int regno = REGNO (y);
2454 unsigned int i;
2455 unsigned int endregno
2456 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2457 : hard_regno_nregs[regno][GET_MODE (y)]);
2459 /* If the quantities are not the same, the expressions are not
2460 equivalent. If there are and we are not to validate, they
2461 are equivalent. Otherwise, ensure all regs are up-to-date. */
2463 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2464 return 0;
2466 if (! validate)
2467 return 1;
2469 for (i = regno; i < endregno; i++)
2470 if (REG_IN_TABLE (i) != REG_TICK (i))
2471 return 0;
2473 return 1;
2476 case MEM:
2477 if (for_gcse)
2479 /* Can't merge two expressions in different alias sets, since we
2480 can decide that the expression is transparent in a block when
2481 it isn't, due to it being set with the different alias set. */
2482 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
2483 return 0;
2485 /* A volatile mem should not be considered equivalent to any
2486 other. */
2487 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2488 return 0;
2490 break;
2492 /* For commutative operations, check both orders. */
2493 case PLUS:
2494 case MULT:
2495 case AND:
2496 case IOR:
2497 case XOR:
2498 case NE:
2499 case EQ:
2500 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2501 validate, for_gcse)
2502 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2503 validate, for_gcse))
2504 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2505 validate, for_gcse)
2506 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2507 validate, for_gcse)));
2509 case ASM_OPERANDS:
2510 /* We don't use the generic code below because we want to
2511 disregard filename and line numbers. */
2513 /* A volatile asm isn't equivalent to any other. */
2514 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2515 return 0;
2517 if (GET_MODE (x) != GET_MODE (y)
2518 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2519 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2520 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2521 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2522 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2523 return 0;
2525 if (ASM_OPERANDS_INPUT_LENGTH (x))
2527 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2528 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2529 ASM_OPERANDS_INPUT (y, i),
2530 validate, for_gcse)
2531 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2532 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2533 return 0;
2536 return 1;
2538 default:
2539 break;
2542 /* Compare the elements. If any pair of corresponding elements
2543 fail to match, return 0 for the whole thing. */
2545 fmt = GET_RTX_FORMAT (code);
2546 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2548 switch (fmt[i])
2550 case 'e':
2551 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2552 validate, for_gcse))
2553 return 0;
2554 break;
2556 case 'E':
2557 if (XVECLEN (x, i) != XVECLEN (y, i))
2558 return 0;
2559 for (j = 0; j < XVECLEN (x, i); j++)
2560 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2561 validate, for_gcse))
2562 return 0;
2563 break;
2565 case 's':
2566 if (strcmp (XSTR (x, i), XSTR (y, i)))
2567 return 0;
2568 break;
2570 case 'i':
2571 if (XINT (x, i) != XINT (y, i))
2572 return 0;
2573 break;
2575 case 'w':
2576 if (XWINT (x, i) != XWINT (y, i))
2577 return 0;
2578 break;
2580 case '0':
2581 case 't':
2582 break;
2584 default:
2585 abort ();
2589 return 1;
2592 /* Return 1 if X has a value that can vary even between two
2593 executions of the program. 0 means X can be compared reliably
2594 against certain constants or near-constants. */
2596 static int
2597 cse_rtx_varies_p (rtx x, int from_alias)
2599 /* We need not check for X and the equivalence class being of the same
2600 mode because if X is equivalent to a constant in some mode, it
2601 doesn't vary in any mode. */
2603 if (REG_P (x)
2604 && REGNO_QTY_VALID_P (REGNO (x)))
2606 int x_q = REG_QTY (REGNO (x));
2607 struct qty_table_elem *x_ent = &qty_table[x_q];
2609 if (GET_MODE (x) == x_ent->mode
2610 && x_ent->const_rtx != NULL_RTX)
2611 return 0;
2614 if (GET_CODE (x) == PLUS
2615 && GET_CODE (XEXP (x, 1)) == CONST_INT
2616 && REG_P (XEXP (x, 0))
2617 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2619 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2620 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2622 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2623 && x0_ent->const_rtx != NULL_RTX)
2624 return 0;
2627 /* This can happen as the result of virtual register instantiation, if
2628 the initial constant is too large to be a valid address. This gives
2629 us a three instruction sequence, load large offset into a register,
2630 load fp minus a constant into a register, then a MEM which is the
2631 sum of the two `constant' registers. */
2632 if (GET_CODE (x) == PLUS
2633 && REG_P (XEXP (x, 0))
2634 && REG_P (XEXP (x, 1))
2635 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2636 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2638 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2639 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2640 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2641 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2643 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2644 && x0_ent->const_rtx != NULL_RTX
2645 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2646 && x1_ent->const_rtx != NULL_RTX)
2647 return 0;
2650 return rtx_varies_p (x, from_alias);
2653 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2654 the result if necessary. INSN is as for canon_reg. */
2656 static void
2657 validate_canon_reg (rtx *xloc, rtx insn)
2659 rtx new = canon_reg (*xloc, insn);
2660 int insn_code;
2662 /* If replacing pseudo with hard reg or vice versa, ensure the
2663 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2664 if (insn != 0 && new != 0
2665 && REG_P (new) && REG_P (*xloc)
2666 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2667 != (REGNO (*xloc) < FIRST_PSEUDO_REGISTER))
2668 || GET_MODE (new) != GET_MODE (*xloc)
2669 || (insn_code = recog_memoized (insn)) < 0
2670 || insn_data[insn_code].n_dups > 0))
2671 validate_change (insn, xloc, new, 1);
2672 else
2673 *xloc = new;
2676 /* Canonicalize an expression:
2677 replace each register reference inside it
2678 with the "oldest" equivalent register.
2680 If INSN is nonzero and we are replacing a pseudo with a hard register
2681 or vice versa, validate_change is used to ensure that INSN remains valid
2682 after we make our substitution. The calls are made with IN_GROUP nonzero
2683 so apply_change_group must be called upon the outermost return from this
2684 function (unless INSN is zero). The result of apply_change_group can
2685 generally be discarded since the changes we are making are optional. */
2687 static rtx
2688 canon_reg (rtx x, rtx insn)
2690 int i;
2691 enum rtx_code code;
2692 const char *fmt;
2694 if (x == 0)
2695 return x;
2697 code = GET_CODE (x);
2698 switch (code)
2700 case PC:
2701 case CC0:
2702 case CONST:
2703 case CONST_INT:
2704 case CONST_DOUBLE:
2705 case CONST_VECTOR:
2706 case SYMBOL_REF:
2707 case LABEL_REF:
2708 case ADDR_VEC:
2709 case ADDR_DIFF_VEC:
2710 return x;
2712 case REG:
2714 int first;
2715 int q;
2716 struct qty_table_elem *ent;
2718 /* Never replace a hard reg, because hard regs can appear
2719 in more than one machine mode, and we must preserve the mode
2720 of each occurrence. Also, some hard regs appear in
2721 MEMs that are shared and mustn't be altered. Don't try to
2722 replace any reg that maps to a reg of class NO_REGS. */
2723 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2724 || ! REGNO_QTY_VALID_P (REGNO (x)))
2725 return x;
2727 q = REG_QTY (REGNO (x));
2728 ent = &qty_table[q];
2729 first = ent->first_reg;
2730 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2731 : REGNO_REG_CLASS (first) == NO_REGS ? x
2732 : gen_rtx_REG (ent->mode, first));
2735 default:
2736 break;
2739 fmt = GET_RTX_FORMAT (code);
2740 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2742 int j;
2744 if (fmt[i] == 'e')
2745 validate_canon_reg (&XEXP (x, i), insn);
2746 else if (fmt[i] == 'E')
2747 for (j = 0; j < XVECLEN (x, i); j++)
2748 validate_canon_reg (&XVECEXP (x, i, j), insn);
2751 return x;
2754 /* LOC is a location within INSN that is an operand address (the contents of
2755 a MEM). Find the best equivalent address to use that is valid for this
2756 insn.
2758 On most CISC machines, complicated address modes are costly, and rtx_cost
2759 is a good approximation for that cost. However, most RISC machines have
2760 only a few (usually only one) memory reference formats. If an address is
2761 valid at all, it is often just as cheap as any other address. Hence, for
2762 RISC machines, we use `address_cost' to compare the costs of various
2763 addresses. For two addresses of equal cost, choose the one with the
2764 highest `rtx_cost' value as that has the potential of eliminating the
2765 most insns. For equal costs, we choose the first in the equivalence
2766 class. Note that we ignore the fact that pseudo registers are cheaper than
2767 hard registers here because we would also prefer the pseudo registers. */
2769 static void
2770 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2772 struct table_elt *elt;
2773 rtx addr = *loc;
2774 struct table_elt *p;
2775 int found_better = 1;
2776 int save_do_not_record = do_not_record;
2777 int save_hash_arg_in_memory = hash_arg_in_memory;
2778 int addr_volatile;
2779 int regno;
2780 unsigned hash;
2782 /* Do not try to replace constant addresses or addresses of local and
2783 argument slots. These MEM expressions are made only once and inserted
2784 in many instructions, as well as being used to control symbol table
2785 output. It is not safe to clobber them.
2787 There are some uncommon cases where the address is already in a register
2788 for some reason, but we cannot take advantage of that because we have
2789 no easy way to unshare the MEM. In addition, looking up all stack
2790 addresses is costly. */
2791 if ((GET_CODE (addr) == PLUS
2792 && REG_P (XEXP (addr, 0))
2793 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2794 && (regno = REGNO (XEXP (addr, 0)),
2795 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2796 || regno == ARG_POINTER_REGNUM))
2797 || (REG_P (addr)
2798 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2799 || regno == HARD_FRAME_POINTER_REGNUM
2800 || regno == ARG_POINTER_REGNUM))
2801 || CONSTANT_ADDRESS_P (addr))
2802 return;
2804 /* If this address is not simply a register, try to fold it. This will
2805 sometimes simplify the expression. Many simplifications
2806 will not be valid, but some, usually applying the associative rule, will
2807 be valid and produce better code. */
2808 if (!REG_P (addr))
2810 rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX);
2811 int addr_folded_cost = address_cost (folded, mode);
2812 int addr_cost = address_cost (addr, mode);
2814 if ((addr_folded_cost < addr_cost
2815 || (addr_folded_cost == addr_cost
2816 /* ??? The rtx_cost comparison is left over from an older
2817 version of this code. It is probably no longer helpful. */
2818 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2819 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2820 && validate_change (insn, loc, folded, 0))
2821 addr = folded;
2824 /* If this address is not in the hash table, we can't look for equivalences
2825 of the whole address. Also, ignore if volatile. */
2827 do_not_record = 0;
2828 hash = HASH (addr, Pmode);
2829 addr_volatile = do_not_record;
2830 do_not_record = save_do_not_record;
2831 hash_arg_in_memory = save_hash_arg_in_memory;
2833 if (addr_volatile)
2834 return;
2836 elt = lookup (addr, hash, Pmode);
2838 if (elt)
2840 /* We need to find the best (under the criteria documented above) entry
2841 in the class that is valid. We use the `flag' field to indicate
2842 choices that were invalid and iterate until we can't find a better
2843 one that hasn't already been tried. */
2845 for (p = elt->first_same_value; p; p = p->next_same_value)
2846 p->flag = 0;
2848 while (found_better)
2850 int best_addr_cost = address_cost (*loc, mode);
2851 int best_rtx_cost = (elt->cost + 1) >> 1;
2852 int exp_cost;
2853 struct table_elt *best_elt = elt;
2855 found_better = 0;
2856 for (p = elt->first_same_value; p; p = p->next_same_value)
2857 if (! p->flag)
2859 if ((REG_P (p->exp)
2860 || exp_equiv_p (p->exp, p->exp, 1, false))
2861 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2862 || (exp_cost == best_addr_cost
2863 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2865 found_better = 1;
2866 best_addr_cost = exp_cost;
2867 best_rtx_cost = (p->cost + 1) >> 1;
2868 best_elt = p;
2872 if (found_better)
2874 if (validate_change (insn, loc,
2875 canon_reg (copy_rtx (best_elt->exp),
2876 NULL_RTX), 0))
2877 return;
2878 else
2879 best_elt->flag = 1;
2884 /* If the address is a binary operation with the first operand a register
2885 and the second a constant, do the same as above, but looking for
2886 equivalences of the register. Then try to simplify before checking for
2887 the best address to use. This catches a few cases: First is when we
2888 have REG+const and the register is another REG+const. We can often merge
2889 the constants and eliminate one insn and one register. It may also be
2890 that a machine has a cheap REG+REG+const. Finally, this improves the
2891 code on the Alpha for unaligned byte stores. */
2893 if (flag_expensive_optimizations
2894 && ARITHMETIC_P (*loc)
2895 && REG_P (XEXP (*loc, 0)))
2897 rtx op1 = XEXP (*loc, 1);
2899 do_not_record = 0;
2900 hash = HASH (XEXP (*loc, 0), Pmode);
2901 do_not_record = save_do_not_record;
2902 hash_arg_in_memory = save_hash_arg_in_memory;
2904 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2905 if (elt == 0)
2906 return;
2908 /* We need to find the best (under the criteria documented above) entry
2909 in the class that is valid. We use the `flag' field to indicate
2910 choices that were invalid and iterate until we can't find a better
2911 one that hasn't already been tried. */
2913 for (p = elt->first_same_value; p; p = p->next_same_value)
2914 p->flag = 0;
2916 while (found_better)
2918 int best_addr_cost = address_cost (*loc, mode);
2919 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2920 struct table_elt *best_elt = elt;
2921 rtx best_rtx = *loc;
2922 int count;
2924 /* This is at worst case an O(n^2) algorithm, so limit our search
2925 to the first 32 elements on the list. This avoids trouble
2926 compiling code with very long basic blocks that can easily
2927 call simplify_gen_binary so many times that we run out of
2928 memory. */
2930 found_better = 0;
2931 for (p = elt->first_same_value, count = 0;
2932 p && count < 32;
2933 p = p->next_same_value, count++)
2934 if (! p->flag
2935 && (REG_P (p->exp)
2936 || exp_equiv_p (p->exp, p->exp, 1, false)))
2938 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
2939 p->exp, op1);
2940 int new_cost;
2941 new_cost = address_cost (new, mode);
2943 if (new_cost < best_addr_cost
2944 || (new_cost == best_addr_cost
2945 && (COST (new) + 1) >> 1 > best_rtx_cost))
2947 found_better = 1;
2948 best_addr_cost = new_cost;
2949 best_rtx_cost = (COST (new) + 1) >> 1;
2950 best_elt = p;
2951 best_rtx = new;
2955 if (found_better)
2957 if (validate_change (insn, loc,
2958 canon_reg (copy_rtx (best_rtx),
2959 NULL_RTX), 0))
2960 return;
2961 else
2962 best_elt->flag = 1;
2968 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2969 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2970 what values are being compared.
2972 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2973 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2974 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2975 compared to produce cc0.
2977 The return value is the comparison operator and is either the code of
2978 A or the code corresponding to the inverse of the comparison. */
2980 static enum rtx_code
2981 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2982 enum machine_mode *pmode1, enum machine_mode *pmode2)
2984 rtx arg1, arg2;
2986 arg1 = *parg1, arg2 = *parg2;
2988 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2990 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2992 /* Set nonzero when we find something of interest. */
2993 rtx x = 0;
2994 int reverse_code = 0;
2995 struct table_elt *p = 0;
2997 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2998 On machines with CC0, this is the only case that can occur, since
2999 fold_rtx will return the COMPARE or item being compared with zero
3000 when given CC0. */
3002 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3003 x = arg1;
3005 /* If ARG1 is a comparison operator and CODE is testing for
3006 STORE_FLAG_VALUE, get the inner arguments. */
3008 else if (COMPARISON_P (arg1))
3010 #ifdef FLOAT_STORE_FLAG_VALUE
3011 REAL_VALUE_TYPE fsfv;
3012 #endif
3014 if (code == NE
3015 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3016 && code == LT && STORE_FLAG_VALUE == -1)
3017 #ifdef FLOAT_STORE_FLAG_VALUE
3018 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3019 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3020 REAL_VALUE_NEGATIVE (fsfv)))
3021 #endif
3023 x = arg1;
3024 else if (code == EQ
3025 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3026 && code == GE && STORE_FLAG_VALUE == -1)
3027 #ifdef FLOAT_STORE_FLAG_VALUE
3028 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3029 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3030 REAL_VALUE_NEGATIVE (fsfv)))
3031 #endif
3033 x = arg1, reverse_code = 1;
3036 /* ??? We could also check for
3038 (ne (and (eq (...) (const_int 1))) (const_int 0))
3040 and related forms, but let's wait until we see them occurring. */
3042 if (x == 0)
3043 /* Look up ARG1 in the hash table and see if it has an equivalence
3044 that lets us see what is being compared. */
3045 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3046 if (p)
3048 p = p->first_same_value;
3050 /* If what we compare is already known to be constant, that is as
3051 good as it gets.
3052 We need to break the loop in this case, because otherwise we
3053 can have an infinite loop when looking at a reg that is known
3054 to be a constant which is the same as a comparison of a reg
3055 against zero which appears later in the insn stream, which in
3056 turn is constant and the same as the comparison of the first reg
3057 against zero... */
3058 if (p->is_const)
3059 break;
3062 for (; p; p = p->next_same_value)
3064 enum machine_mode inner_mode = GET_MODE (p->exp);
3065 #ifdef FLOAT_STORE_FLAG_VALUE
3066 REAL_VALUE_TYPE fsfv;
3067 #endif
3069 /* If the entry isn't valid, skip it. */
3070 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3071 continue;
3073 if (GET_CODE (p->exp) == COMPARE
3074 /* Another possibility is that this machine has a compare insn
3075 that includes the comparison code. In that case, ARG1 would
3076 be equivalent to a comparison operation that would set ARG1 to
3077 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3078 ORIG_CODE is the actual comparison being done; if it is an EQ,
3079 we must reverse ORIG_CODE. On machine with a negative value
3080 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3081 || ((code == NE
3082 || (code == LT
3083 && GET_MODE_CLASS (inner_mode) == MODE_INT
3084 && (GET_MODE_BITSIZE (inner_mode)
3085 <= HOST_BITS_PER_WIDE_INT)
3086 && (STORE_FLAG_VALUE
3087 & ((HOST_WIDE_INT) 1
3088 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3089 #ifdef FLOAT_STORE_FLAG_VALUE
3090 || (code == LT
3091 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3092 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3093 REAL_VALUE_NEGATIVE (fsfv)))
3094 #endif
3096 && COMPARISON_P (p->exp)))
3098 x = p->exp;
3099 break;
3101 else if ((code == EQ
3102 || (code == GE
3103 && GET_MODE_CLASS (inner_mode) == MODE_INT
3104 && (GET_MODE_BITSIZE (inner_mode)
3105 <= HOST_BITS_PER_WIDE_INT)
3106 && (STORE_FLAG_VALUE
3107 & ((HOST_WIDE_INT) 1
3108 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3109 #ifdef FLOAT_STORE_FLAG_VALUE
3110 || (code == GE
3111 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3112 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3113 REAL_VALUE_NEGATIVE (fsfv)))
3114 #endif
3116 && COMPARISON_P (p->exp))
3118 reverse_code = 1;
3119 x = p->exp;
3120 break;
3123 /* If this non-trapping address, e.g. fp + constant, the
3124 equivalent is a better operand since it may let us predict
3125 the value of the comparison. */
3126 else if (!rtx_addr_can_trap_p (p->exp))
3128 arg1 = p->exp;
3129 continue;
3133 /* If we didn't find a useful equivalence for ARG1, we are done.
3134 Otherwise, set up for the next iteration. */
3135 if (x == 0)
3136 break;
3138 /* If we need to reverse the comparison, make sure that that is
3139 possible -- we can't necessarily infer the value of GE from LT
3140 with floating-point operands. */
3141 if (reverse_code)
3143 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3144 if (reversed == UNKNOWN)
3145 break;
3146 else
3147 code = reversed;
3149 else if (COMPARISON_P (x))
3150 code = GET_CODE (x);
3151 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3154 /* Return our results. Return the modes from before fold_rtx
3155 because fold_rtx might produce const_int, and then it's too late. */
3156 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3157 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3159 return code;
3162 /* If X is a nontrivial arithmetic operation on an argument
3163 for which a constant value can be determined, return
3164 the result of operating on that value, as a constant.
3165 Otherwise, return X, possibly with one or more operands
3166 modified by recursive calls to this function.
3168 If X is a register whose contents are known, we do NOT
3169 return those contents here. equiv_constant is called to
3170 perform that task.
3172 INSN is the insn that we may be modifying. If it is 0, make a copy
3173 of X before modifying it. */
3175 static rtx
3176 fold_rtx (rtx x, rtx insn)
3178 enum rtx_code code;
3179 enum machine_mode mode;
3180 const char *fmt;
3181 int i;
3182 rtx new = 0;
3183 int copied = 0;
3184 int must_swap = 0;
3186 /* Folded equivalents of first two operands of X. */
3187 rtx folded_arg0;
3188 rtx folded_arg1;
3190 /* Constant equivalents of first three operands of X;
3191 0 when no such equivalent is known. */
3192 rtx const_arg0;
3193 rtx const_arg1;
3194 rtx const_arg2;
3196 /* The mode of the first operand of X. We need this for sign and zero
3197 extends. */
3198 enum machine_mode mode_arg0;
3200 if (x == 0)
3201 return x;
3203 mode = GET_MODE (x);
3204 code = GET_CODE (x);
3205 switch (code)
3207 case CONST:
3208 case CONST_INT:
3209 case CONST_DOUBLE:
3210 case CONST_VECTOR:
3211 case SYMBOL_REF:
3212 case LABEL_REF:
3213 case REG:
3214 /* No use simplifying an EXPR_LIST
3215 since they are used only for lists of args
3216 in a function call's REG_EQUAL note. */
3217 case EXPR_LIST:
3218 return x;
3220 #ifdef HAVE_cc0
3221 case CC0:
3222 return prev_insn_cc0;
3223 #endif
3225 case PC:
3226 /* If the next insn is a CODE_LABEL followed by a jump table,
3227 PC's value is a LABEL_REF pointing to that label. That
3228 lets us fold switch statements on the VAX. */
3230 rtx next;
3231 if (insn && tablejump_p (insn, &next, NULL))
3232 return gen_rtx_LABEL_REF (Pmode, next);
3234 break;
3236 case SUBREG:
3237 /* See if we previously assigned a constant value to this SUBREG. */
3238 if ((new = lookup_as_function (x, CONST_INT)) != 0
3239 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3240 return new;
3242 /* If this is a paradoxical SUBREG, we have no idea what value the
3243 extra bits would have. However, if the operand is equivalent
3244 to a SUBREG whose operand is the same as our mode, and all the
3245 modes are within a word, we can just use the inner operand
3246 because these SUBREGs just say how to treat the register.
3248 Similarly if we find an integer constant. */
3250 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3252 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3253 struct table_elt *elt;
3255 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3256 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3257 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3258 imode)) != 0)
3259 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3261 if (CONSTANT_P (elt->exp)
3262 && GET_MODE (elt->exp) == VOIDmode)
3263 return elt->exp;
3265 if (GET_CODE (elt->exp) == SUBREG
3266 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3267 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3268 return copy_rtx (SUBREG_REG (elt->exp));
3271 return x;
3274 /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG.
3275 We might be able to if the SUBREG is extracting a single word in an
3276 integral mode or extracting the low part. */
3278 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3279 const_arg0 = equiv_constant (folded_arg0);
3280 if (const_arg0)
3281 folded_arg0 = const_arg0;
3283 if (folded_arg0 != SUBREG_REG (x))
3285 new = simplify_subreg (mode, folded_arg0,
3286 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3287 if (new)
3288 return new;
3291 if (REG_P (folded_arg0)
3292 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3294 struct table_elt *elt;
3296 elt = lookup (folded_arg0,
3297 HASH (folded_arg0, GET_MODE (folded_arg0)),
3298 GET_MODE (folded_arg0));
3300 if (elt)
3301 elt = elt->first_same_value;
3303 if (subreg_lowpart_p (x))
3304 /* If this is a narrowing SUBREG and our operand is a REG, see
3305 if we can find an equivalence for REG that is an arithmetic
3306 operation in a wider mode where both operands are paradoxical
3307 SUBREGs from objects of our result mode. In that case, we
3308 couldn-t report an equivalent value for that operation, since we
3309 don't know what the extra bits will be. But we can find an
3310 equivalence for this SUBREG by folding that operation in the
3311 narrow mode. This allows us to fold arithmetic in narrow modes
3312 when the machine only supports word-sized arithmetic.
3314 Also look for a case where we have a SUBREG whose operand
3315 is the same as our result. If both modes are smaller
3316 than a word, we are simply interpreting a register in
3317 different modes and we can use the inner value. */
3319 for (; elt; elt = elt->next_same_value)
3321 enum rtx_code eltcode = GET_CODE (elt->exp);
3323 /* Just check for unary and binary operations. */
3324 if (UNARY_P (elt->exp)
3325 && eltcode != SIGN_EXTEND
3326 && eltcode != ZERO_EXTEND
3327 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3328 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3329 && (GET_MODE_CLASS (mode)
3330 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3332 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3334 if (!REG_P (op0) && ! CONSTANT_P (op0))
3335 op0 = fold_rtx (op0, NULL_RTX);
3337 op0 = equiv_constant (op0);
3338 if (op0)
3339 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3340 op0, mode);
3342 else if (ARITHMETIC_P (elt->exp)
3343 && eltcode != DIV && eltcode != MOD
3344 && eltcode != UDIV && eltcode != UMOD
3345 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3346 && eltcode != ROTATE && eltcode != ROTATERT
3347 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3348 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3349 == mode))
3350 || CONSTANT_P (XEXP (elt->exp, 0)))
3351 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3352 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3353 == mode))
3354 || CONSTANT_P (XEXP (elt->exp, 1))))
3356 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3357 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3359 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3360 op0 = fold_rtx (op0, NULL_RTX);
3362 if (op0)
3363 op0 = equiv_constant (op0);
3365 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3366 op1 = fold_rtx (op1, NULL_RTX);
3368 if (op1)
3369 op1 = equiv_constant (op1);
3371 /* If we are looking for the low SImode part of
3372 (ashift:DI c (const_int 32)), it doesn't work
3373 to compute that in SImode, because a 32-bit shift
3374 in SImode is unpredictable. We know the value is 0. */
3375 if (op0 && op1
3376 && GET_CODE (elt->exp) == ASHIFT
3377 && GET_CODE (op1) == CONST_INT
3378 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3380 if (INTVAL (op1)
3381 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3382 /* If the count fits in the inner mode's width,
3383 but exceeds the outer mode's width,
3384 the value will get truncated to 0
3385 by the subreg. */
3386 new = CONST0_RTX (mode);
3387 else
3388 /* If the count exceeds even the inner mode's width,
3389 don't fold this expression. */
3390 new = 0;
3392 else if (op0 && op1)
3393 new = simplify_binary_operation (GET_CODE (elt->exp), mode, op0, op1);
3396 else if (GET_CODE (elt->exp) == SUBREG
3397 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3398 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3399 <= UNITS_PER_WORD)
3400 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3401 new = copy_rtx (SUBREG_REG (elt->exp));
3403 if (new)
3404 return new;
3406 else
3407 /* A SUBREG resulting from a zero extension may fold to zero if
3408 it extracts higher bits than the ZERO_EXTEND's source bits.
3409 FIXME: if combine tried to, er, combine these instructions,
3410 this transformation may be moved to simplify_subreg. */
3411 for (; elt; elt = elt->next_same_value)
3413 if (GET_CODE (elt->exp) == ZERO_EXTEND
3414 && subreg_lsb (x)
3415 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3416 return CONST0_RTX (mode);
3420 return x;
3422 case NOT:
3423 case NEG:
3424 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3425 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3426 new = lookup_as_function (XEXP (x, 0), code);
3427 if (new)
3428 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3429 break;
3431 case MEM:
3432 /* If we are not actually processing an insn, don't try to find the
3433 best address. Not only don't we care, but we could modify the
3434 MEM in an invalid way since we have no insn to validate against. */
3435 if (insn != 0)
3436 find_best_addr (insn, &XEXP (x, 0), GET_MODE (x));
3439 /* Even if we don't fold in the insn itself,
3440 we can safely do so here, in hopes of getting a constant. */
3441 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3442 rtx base = 0;
3443 HOST_WIDE_INT offset = 0;
3445 if (REG_P (addr)
3446 && REGNO_QTY_VALID_P (REGNO (addr)))
3448 int addr_q = REG_QTY (REGNO (addr));
3449 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3451 if (GET_MODE (addr) == addr_ent->mode
3452 && addr_ent->const_rtx != NULL_RTX)
3453 addr = addr_ent->const_rtx;
3456 /* If address is constant, split it into a base and integer offset. */
3457 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3458 base = addr;
3459 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3460 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3462 base = XEXP (XEXP (addr, 0), 0);
3463 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3465 else if (GET_CODE (addr) == LO_SUM
3466 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3467 base = XEXP (addr, 1);
3469 /* If this is a constant pool reference, we can fold it into its
3470 constant to allow better value tracking. */
3471 if (base && GET_CODE (base) == SYMBOL_REF
3472 && CONSTANT_POOL_ADDRESS_P (base))
3474 rtx constant = get_pool_constant (base);
3475 enum machine_mode const_mode = get_pool_mode (base);
3476 rtx new;
3478 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3480 constant_pool_entries_cost = COST (constant);
3481 constant_pool_entries_regcost = approx_reg_cost (constant);
3484 /* If we are loading the full constant, we have an equivalence. */
3485 if (offset == 0 && mode == const_mode)
3486 return constant;
3488 /* If this actually isn't a constant (weird!), we can't do
3489 anything. Otherwise, handle the two most common cases:
3490 extracting a word from a multi-word constant, and extracting
3491 the low-order bits. Other cases don't seem common enough to
3492 worry about. */
3493 if (! CONSTANT_P (constant))
3494 return x;
3496 if (GET_MODE_CLASS (mode) == MODE_INT
3497 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3498 && offset % UNITS_PER_WORD == 0
3499 && (new = operand_subword (constant,
3500 offset / UNITS_PER_WORD,
3501 0, const_mode)) != 0)
3502 return new;
3504 if (((BYTES_BIG_ENDIAN
3505 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3506 || (! BYTES_BIG_ENDIAN && offset == 0))
3507 && (new = gen_lowpart (mode, constant)) != 0)
3508 return new;
3511 /* If this is a reference to a label at a known position in a jump
3512 table, we also know its value. */
3513 if (base && GET_CODE (base) == LABEL_REF)
3515 rtx label = XEXP (base, 0);
3516 rtx table_insn = NEXT_INSN (label);
3518 if (table_insn && JUMP_P (table_insn)
3519 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3521 rtx table = PATTERN (table_insn);
3523 if (offset >= 0
3524 && (offset / GET_MODE_SIZE (GET_MODE (table))
3525 < XVECLEN (table, 0)))
3526 return XVECEXP (table, 0,
3527 offset / GET_MODE_SIZE (GET_MODE (table)));
3529 if (table_insn && JUMP_P (table_insn)
3530 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3532 rtx table = PATTERN (table_insn);
3534 if (offset >= 0
3535 && (offset / GET_MODE_SIZE (GET_MODE (table))
3536 < XVECLEN (table, 1)))
3538 offset /= GET_MODE_SIZE (GET_MODE (table));
3539 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3540 XEXP (table, 0));
3542 if (GET_MODE (table) != Pmode)
3543 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3545 /* Indicate this is a constant. This isn't a
3546 valid form of CONST, but it will only be used
3547 to fold the next insns and then discarded, so
3548 it should be safe.
3550 Note this expression must be explicitly discarded,
3551 by cse_insn, else it may end up in a REG_EQUAL note
3552 and "escape" to cause problems elsewhere. */
3553 return gen_rtx_CONST (GET_MODE (new), new);
3558 return x;
3561 #ifdef NO_FUNCTION_CSE
3562 case CALL:
3563 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3564 return x;
3565 break;
3566 #endif
3568 case ASM_OPERANDS:
3569 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3570 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3571 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3572 break;
3574 default:
3575 break;
3578 const_arg0 = 0;
3579 const_arg1 = 0;
3580 const_arg2 = 0;
3581 mode_arg0 = VOIDmode;
3583 /* Try folding our operands.
3584 Then see which ones have constant values known. */
3586 fmt = GET_RTX_FORMAT (code);
3587 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3588 if (fmt[i] == 'e')
3590 rtx arg = XEXP (x, i);
3591 rtx folded_arg = arg, const_arg = 0;
3592 enum machine_mode mode_arg = GET_MODE (arg);
3593 rtx cheap_arg, expensive_arg;
3594 rtx replacements[2];
3595 int j;
3596 int old_cost = COST_IN (XEXP (x, i), code);
3598 /* Most arguments are cheap, so handle them specially. */
3599 switch (GET_CODE (arg))
3601 case REG:
3602 /* This is the same as calling equiv_constant; it is duplicated
3603 here for speed. */
3604 if (REGNO_QTY_VALID_P (REGNO (arg)))
3606 int arg_q = REG_QTY (REGNO (arg));
3607 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3609 if (arg_ent->const_rtx != NULL_RTX
3610 && !REG_P (arg_ent->const_rtx)
3611 && GET_CODE (arg_ent->const_rtx) != PLUS)
3612 const_arg
3613 = gen_lowpart (GET_MODE (arg),
3614 arg_ent->const_rtx);
3616 break;
3618 case CONST:
3619 case CONST_INT:
3620 case SYMBOL_REF:
3621 case LABEL_REF:
3622 case CONST_DOUBLE:
3623 case CONST_VECTOR:
3624 const_arg = arg;
3625 break;
3627 #ifdef HAVE_cc0
3628 case CC0:
3629 folded_arg = prev_insn_cc0;
3630 mode_arg = prev_insn_cc0_mode;
3631 const_arg = equiv_constant (folded_arg);
3632 break;
3633 #endif
3635 default:
3636 folded_arg = fold_rtx (arg, insn);
3637 const_arg = equiv_constant (folded_arg);
3640 /* For the first three operands, see if the operand
3641 is constant or equivalent to a constant. */
3642 switch (i)
3644 case 0:
3645 folded_arg0 = folded_arg;
3646 const_arg0 = const_arg;
3647 mode_arg0 = mode_arg;
3648 break;
3649 case 1:
3650 folded_arg1 = folded_arg;
3651 const_arg1 = const_arg;
3652 break;
3653 case 2:
3654 const_arg2 = const_arg;
3655 break;
3658 /* Pick the least expensive of the folded argument and an
3659 equivalent constant argument. */
3660 if (const_arg == 0 || const_arg == folded_arg
3661 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3662 cheap_arg = folded_arg, expensive_arg = const_arg;
3663 else
3664 cheap_arg = const_arg, expensive_arg = folded_arg;
3666 /* Try to replace the operand with the cheapest of the two
3667 possibilities. If it doesn't work and this is either of the first
3668 two operands of a commutative operation, try swapping them.
3669 If THAT fails, try the more expensive, provided it is cheaper
3670 than what is already there. */
3672 if (cheap_arg == XEXP (x, i))
3673 continue;
3675 if (insn == 0 && ! copied)
3677 x = copy_rtx (x);
3678 copied = 1;
3681 /* Order the replacements from cheapest to most expensive. */
3682 replacements[0] = cheap_arg;
3683 replacements[1] = expensive_arg;
3685 for (j = 0; j < 2 && replacements[j]; j++)
3687 int new_cost = COST_IN (replacements[j], code);
3689 /* Stop if what existed before was cheaper. Prefer constants
3690 in the case of a tie. */
3691 if (new_cost > old_cost
3692 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3693 break;
3695 /* It's not safe to substitute the operand of a conversion
3696 operator with a constant, as the conversion's identity
3697 depends upon the mode of it's operand. This optimization
3698 is handled by the call to simplify_unary_operation. */
3699 if (GET_RTX_CLASS (code) == RTX_UNARY
3700 && GET_MODE (replacements[j]) != mode_arg0
3701 && (code == ZERO_EXTEND
3702 || code == SIGN_EXTEND
3703 || code == TRUNCATE
3704 || code == FLOAT_TRUNCATE
3705 || code == FLOAT_EXTEND
3706 || code == FLOAT
3707 || code == FIX
3708 || code == UNSIGNED_FLOAT
3709 || code == UNSIGNED_FIX))
3710 continue;
3712 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3713 break;
3715 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3716 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3718 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3719 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3721 if (apply_change_group ())
3723 /* Swap them back to be invalid so that this loop can
3724 continue and flag them to be swapped back later. */
3725 rtx tem;
3727 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3728 XEXP (x, 1) = tem;
3729 must_swap = 1;
3730 break;
3736 else
3738 if (fmt[i] == 'E')
3739 /* Don't try to fold inside of a vector of expressions.
3740 Doing nothing is harmless. */
3744 /* If a commutative operation, place a constant integer as the second
3745 operand unless the first operand is also a constant integer. Otherwise,
3746 place any constant second unless the first operand is also a constant. */
3748 if (COMMUTATIVE_P (x))
3750 if (must_swap
3751 || swap_commutative_operands_p (const_arg0 ? const_arg0
3752 : XEXP (x, 0),
3753 const_arg1 ? const_arg1
3754 : XEXP (x, 1)))
3756 rtx tem = XEXP (x, 0);
3758 if (insn == 0 && ! copied)
3760 x = copy_rtx (x);
3761 copied = 1;
3764 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3765 validate_change (insn, &XEXP (x, 1), tem, 1);
3766 if (apply_change_group ())
3768 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3769 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3774 /* If X is an arithmetic operation, see if we can simplify it. */
3776 switch (GET_RTX_CLASS (code))
3778 case RTX_UNARY:
3780 int is_const = 0;
3782 /* We can't simplify extension ops unless we know the
3783 original mode. */
3784 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3785 && mode_arg0 == VOIDmode)
3786 break;
3788 /* If we had a CONST, strip it off and put it back later if we
3789 fold. */
3790 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3791 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3793 new = simplify_unary_operation (code, mode,
3794 const_arg0 ? const_arg0 : folded_arg0,
3795 mode_arg0);
3796 if (new != 0 && is_const)
3797 new = gen_rtx_CONST (mode, new);
3799 break;
3801 case RTX_COMPARE:
3802 case RTX_COMM_COMPARE:
3803 /* See what items are actually being compared and set FOLDED_ARG[01]
3804 to those values and CODE to the actual comparison code. If any are
3805 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3806 do anything if both operands are already known to be constant. */
3808 if (const_arg0 == 0 || const_arg1 == 0)
3810 struct table_elt *p0, *p1;
3811 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3812 enum machine_mode mode_arg1;
3814 #ifdef FLOAT_STORE_FLAG_VALUE
3815 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3817 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3818 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3819 false_rtx = CONST0_RTX (mode);
3821 #endif
3823 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3824 &mode_arg0, &mode_arg1);
3825 const_arg0 = equiv_constant (folded_arg0);
3826 const_arg1 = equiv_constant (folded_arg1);
3828 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3829 what kinds of things are being compared, so we can't do
3830 anything with this comparison. */
3832 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3833 break;
3835 /* If we do not now have two constants being compared, see
3836 if we can nevertheless deduce some things about the
3837 comparison. */
3838 if (const_arg0 == 0 || const_arg1 == 0)
3840 /* Some addresses are known to be nonzero. We don't know
3841 their sign, but equality comparisons are known. */
3842 if (const_arg1 == const0_rtx
3843 && nonzero_address_p (folded_arg0))
3845 if (code == EQ)
3846 return false_rtx;
3847 else if (code == NE)
3848 return true_rtx;
3851 /* See if the two operands are the same. */
3853 if (folded_arg0 == folded_arg1
3854 || (REG_P (folded_arg0)
3855 && REG_P (folded_arg1)
3856 && (REG_QTY (REGNO (folded_arg0))
3857 == REG_QTY (REGNO (folded_arg1))))
3858 || ((p0 = lookup (folded_arg0,
3859 SAFE_HASH (folded_arg0, mode_arg0),
3860 mode_arg0))
3861 && (p1 = lookup (folded_arg1,
3862 SAFE_HASH (folded_arg1, mode_arg0),
3863 mode_arg0))
3864 && p0->first_same_value == p1->first_same_value))
3866 /* Sadly two equal NaNs are not equivalent. */
3867 if (!HONOR_NANS (mode_arg0))
3868 return ((code == EQ || code == LE || code == GE
3869 || code == LEU || code == GEU || code == UNEQ
3870 || code == UNLE || code == UNGE
3871 || code == ORDERED)
3872 ? true_rtx : false_rtx);
3873 /* Take care for the FP compares we can resolve. */
3874 if (code == UNEQ || code == UNLE || code == UNGE)
3875 return true_rtx;
3876 if (code == LTGT || code == LT || code == GT)
3877 return false_rtx;
3880 /* If FOLDED_ARG0 is a register, see if the comparison we are
3881 doing now is either the same as we did before or the reverse
3882 (we only check the reverse if not floating-point). */
3883 else if (REG_P (folded_arg0))
3885 int qty = REG_QTY (REGNO (folded_arg0));
3887 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3889 struct qty_table_elem *ent = &qty_table[qty];
3891 if ((comparison_dominates_p (ent->comparison_code, code)
3892 || (! FLOAT_MODE_P (mode_arg0)
3893 && comparison_dominates_p (ent->comparison_code,
3894 reverse_condition (code))))
3895 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3896 || (const_arg1
3897 && rtx_equal_p (ent->comparison_const,
3898 const_arg1))
3899 || (REG_P (folded_arg1)
3900 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3901 return (comparison_dominates_p (ent->comparison_code, code)
3902 ? true_rtx : false_rtx);
3908 /* If we are comparing against zero, see if the first operand is
3909 equivalent to an IOR with a constant. If so, we may be able to
3910 determine the result of this comparison. */
3912 if (const_arg1 == const0_rtx)
3914 rtx y = lookup_as_function (folded_arg0, IOR);
3915 rtx inner_const;
3917 if (y != 0
3918 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3919 && GET_CODE (inner_const) == CONST_INT
3920 && INTVAL (inner_const) != 0)
3922 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
3923 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
3924 && (INTVAL (inner_const)
3925 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
3926 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3928 #ifdef FLOAT_STORE_FLAG_VALUE
3929 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3931 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3932 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3933 false_rtx = CONST0_RTX (mode);
3935 #endif
3937 switch (code)
3939 case EQ:
3940 return false_rtx;
3941 case NE:
3942 return true_rtx;
3943 case LT: case LE:
3944 if (has_sign)
3945 return true_rtx;
3946 break;
3947 case GT: case GE:
3948 if (has_sign)
3949 return false_rtx;
3950 break;
3951 default:
3952 break;
3958 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3959 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3960 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3962 break;
3964 case RTX_BIN_ARITH:
3965 case RTX_COMM_ARITH:
3966 switch (code)
3968 case PLUS:
3969 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3970 with that LABEL_REF as its second operand. If so, the result is
3971 the first operand of that MINUS. This handles switches with an
3972 ADDR_DIFF_VEC table. */
3973 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3975 rtx y
3976 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3977 : lookup_as_function (folded_arg0, MINUS);
3979 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3980 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3981 return XEXP (y, 0);
3983 /* Now try for a CONST of a MINUS like the above. */
3984 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3985 : lookup_as_function (folded_arg0, CONST))) != 0
3986 && GET_CODE (XEXP (y, 0)) == MINUS
3987 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3988 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3989 return XEXP (XEXP (y, 0), 0);
3992 /* Likewise if the operands are in the other order. */
3993 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3995 rtx y
3996 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3997 : lookup_as_function (folded_arg1, MINUS);
3999 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4000 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4001 return XEXP (y, 0);
4003 /* Now try for a CONST of a MINUS like the above. */
4004 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4005 : lookup_as_function (folded_arg1, CONST))) != 0
4006 && GET_CODE (XEXP (y, 0)) == MINUS
4007 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4008 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4009 return XEXP (XEXP (y, 0), 0);
4012 /* If second operand is a register equivalent to a negative
4013 CONST_INT, see if we can find a register equivalent to the
4014 positive constant. Make a MINUS if so. Don't do this for
4015 a non-negative constant since we might then alternate between
4016 choosing positive and negative constants. Having the positive
4017 constant previously-used is the more common case. Be sure
4018 the resulting constant is non-negative; if const_arg1 were
4019 the smallest negative number this would overflow: depending
4020 on the mode, this would either just be the same value (and
4021 hence not save anything) or be incorrect. */
4022 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4023 && INTVAL (const_arg1) < 0
4024 /* This used to test
4026 -INTVAL (const_arg1) >= 0
4028 But The Sun V5.0 compilers mis-compiled that test. So
4029 instead we test for the problematic value in a more direct
4030 manner and hope the Sun compilers get it correct. */
4031 && INTVAL (const_arg1) !=
4032 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4033 && REG_P (folded_arg1))
4035 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4036 struct table_elt *p
4037 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4039 if (p)
4040 for (p = p->first_same_value; p; p = p->next_same_value)
4041 if (REG_P (p->exp))
4042 return simplify_gen_binary (MINUS, mode, folded_arg0,
4043 canon_reg (p->exp, NULL_RTX));
4045 goto from_plus;
4047 case MINUS:
4048 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4049 If so, produce (PLUS Z C2-C). */
4050 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4052 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4053 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4054 return fold_rtx (plus_constant (copy_rtx (y),
4055 -INTVAL (const_arg1)),
4056 NULL_RTX);
4059 /* Fall through. */
4061 from_plus:
4062 case SMIN: case SMAX: case UMIN: case UMAX:
4063 case IOR: case AND: case XOR:
4064 case MULT:
4065 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4066 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4067 is known to be of similar form, we may be able to replace the
4068 operation with a combined operation. This may eliminate the
4069 intermediate operation if every use is simplified in this way.
4070 Note that the similar optimization done by combine.c only works
4071 if the intermediate operation's result has only one reference. */
4073 if (REG_P (folded_arg0)
4074 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4076 int is_shift
4077 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4078 rtx y = lookup_as_function (folded_arg0, code);
4079 rtx inner_const;
4080 enum rtx_code associate_code;
4081 rtx new_const;
4083 if (y == 0
4084 || 0 == (inner_const
4085 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4086 || GET_CODE (inner_const) != CONST_INT
4087 /* If we have compiled a statement like
4088 "if (x == (x & mask1))", and now are looking at
4089 "x & mask2", we will have a case where the first operand
4090 of Y is the same as our first operand. Unless we detect
4091 this case, an infinite loop will result. */
4092 || XEXP (y, 0) == folded_arg0)
4093 break;
4095 /* Don't associate these operations if they are a PLUS with the
4096 same constant and it is a power of two. These might be doable
4097 with a pre- or post-increment. Similarly for two subtracts of
4098 identical powers of two with post decrement. */
4100 if (code == PLUS && const_arg1 == inner_const
4101 && ((HAVE_PRE_INCREMENT
4102 && exact_log2 (INTVAL (const_arg1)) >= 0)
4103 || (HAVE_POST_INCREMENT
4104 && exact_log2 (INTVAL (const_arg1)) >= 0)
4105 || (HAVE_PRE_DECREMENT
4106 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4107 || (HAVE_POST_DECREMENT
4108 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4109 break;
4111 /* Compute the code used to compose the constants. For example,
4112 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4114 associate_code = (is_shift || code == MINUS ? PLUS : code);
4116 new_const = simplify_binary_operation (associate_code, mode,
4117 const_arg1, inner_const);
4119 if (new_const == 0)
4120 break;
4122 /* If we are associating shift operations, don't let this
4123 produce a shift of the size of the object or larger.
4124 This could occur when we follow a sign-extend by a right
4125 shift on a machine that does a sign-extend as a pair
4126 of shifts. */
4128 if (is_shift && GET_CODE (new_const) == CONST_INT
4129 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4131 /* As an exception, we can turn an ASHIFTRT of this
4132 form into a shift of the number of bits - 1. */
4133 if (code == ASHIFTRT)
4134 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4135 else
4136 break;
4139 y = copy_rtx (XEXP (y, 0));
4141 /* If Y contains our first operand (the most common way this
4142 can happen is if Y is a MEM), we would do into an infinite
4143 loop if we tried to fold it. So don't in that case. */
4145 if (! reg_mentioned_p (folded_arg0, y))
4146 y = fold_rtx (y, insn);
4148 return simplify_gen_binary (code, mode, y, new_const);
4150 break;
4152 case DIV: case UDIV:
4153 /* ??? The associative optimization performed immediately above is
4154 also possible for DIV and UDIV using associate_code of MULT.
4155 However, we would need extra code to verify that the
4156 multiplication does not overflow, that is, there is no overflow
4157 in the calculation of new_const. */
4158 break;
4160 default:
4161 break;
4164 new = simplify_binary_operation (code, mode,
4165 const_arg0 ? const_arg0 : folded_arg0,
4166 const_arg1 ? const_arg1 : folded_arg1);
4167 break;
4169 case RTX_OBJ:
4170 /* (lo_sum (high X) X) is simply X. */
4171 if (code == LO_SUM && const_arg0 != 0
4172 && GET_CODE (const_arg0) == HIGH
4173 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4174 return const_arg1;
4175 break;
4177 case RTX_TERNARY:
4178 case RTX_BITFIELD_OPS:
4179 new = simplify_ternary_operation (code, mode, mode_arg0,
4180 const_arg0 ? const_arg0 : folded_arg0,
4181 const_arg1 ? const_arg1 : folded_arg1,
4182 const_arg2 ? const_arg2 : XEXP (x, 2));
4183 break;
4185 default:
4186 break;
4189 return new ? new : x;
4192 /* Return a constant value currently equivalent to X.
4193 Return 0 if we don't know one. */
4195 static rtx
4196 equiv_constant (rtx x)
4198 if (REG_P (x)
4199 && REGNO_QTY_VALID_P (REGNO (x)))
4201 int x_q = REG_QTY (REGNO (x));
4202 struct qty_table_elem *x_ent = &qty_table[x_q];
4204 if (x_ent->const_rtx)
4205 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4208 if (x == 0 || CONSTANT_P (x))
4209 return x;
4211 /* If X is a MEM, try to fold it outside the context of any insn to see if
4212 it might be equivalent to a constant. That handles the case where it
4213 is a constant-pool reference. Then try to look it up in the hash table
4214 in case it is something whose value we have seen before. */
4216 if (MEM_P (x))
4218 struct table_elt *elt;
4220 x = fold_rtx (x, NULL_RTX);
4221 if (CONSTANT_P (x))
4222 return x;
4224 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4225 if (elt == 0)
4226 return 0;
4228 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4229 if (elt->is_const && CONSTANT_P (elt->exp))
4230 return elt->exp;
4233 return 0;
4236 /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point
4237 number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
4238 least-significant part of X.
4239 MODE specifies how big a part of X to return.
4241 If the requested operation cannot be done, 0 is returned.
4243 This is similar to gen_lowpart_general in emit-rtl.c. */
4246 gen_lowpart_if_possible (enum machine_mode mode, rtx x)
4248 rtx result = gen_lowpart_common (mode, x);
4250 if (result)
4251 return result;
4252 else if (MEM_P (x))
4254 /* This is the only other case we handle. */
4255 int offset = 0;
4256 rtx new;
4258 if (WORDS_BIG_ENDIAN)
4259 offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD)
4260 - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD));
4261 if (BYTES_BIG_ENDIAN)
4262 /* Adjust the address so that the address-after-the-data is
4263 unchanged. */
4264 offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode))
4265 - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x))));
4267 new = adjust_address_nv (x, mode, offset);
4268 if (! memory_address_p (mode, XEXP (new, 0)))
4269 return 0;
4271 return new;
4273 else
4274 return 0;
4277 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4278 branch. It will be zero if not.
4280 In certain cases, this can cause us to add an equivalence. For example,
4281 if we are following the taken case of
4282 if (i == 2)
4283 we can add the fact that `i' and '2' are now equivalent.
4285 In any case, we can record that this comparison was passed. If the same
4286 comparison is seen later, we will know its value. */
4288 static void
4289 record_jump_equiv (rtx insn, int taken)
4291 int cond_known_true;
4292 rtx op0, op1;
4293 rtx set;
4294 enum machine_mode mode, mode0, mode1;
4295 int reversed_nonequality = 0;
4296 enum rtx_code code;
4298 /* Ensure this is the right kind of insn. */
4299 if (! any_condjump_p (insn))
4300 return;
4301 set = pc_set (insn);
4303 /* See if this jump condition is known true or false. */
4304 if (taken)
4305 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4306 else
4307 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4309 /* Get the type of comparison being done and the operands being compared.
4310 If we had to reverse a non-equality condition, record that fact so we
4311 know that it isn't valid for floating-point. */
4312 code = GET_CODE (XEXP (SET_SRC (set), 0));
4313 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4314 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4316 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4317 if (! cond_known_true)
4319 code = reversed_comparison_code_parts (code, op0, op1, insn);
4321 /* Don't remember if we can't find the inverse. */
4322 if (code == UNKNOWN)
4323 return;
4326 /* The mode is the mode of the non-constant. */
4327 mode = mode0;
4328 if (mode1 != VOIDmode)
4329 mode = mode1;
4331 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4334 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4335 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4336 Make any useful entries we can with that information. Called from
4337 above function and called recursively. */
4339 static void
4340 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4341 rtx op1, int reversed_nonequality)
4343 unsigned op0_hash, op1_hash;
4344 int op0_in_memory, op1_in_memory;
4345 struct table_elt *op0_elt, *op1_elt;
4347 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4348 we know that they are also equal in the smaller mode (this is also
4349 true for all smaller modes whether or not there is a SUBREG, but
4350 is not worth testing for with no SUBREG). */
4352 /* Note that GET_MODE (op0) may not equal MODE. */
4353 if (code == EQ && GET_CODE (op0) == SUBREG
4354 && (GET_MODE_SIZE (GET_MODE (op0))
4355 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4357 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4358 rtx tem = gen_lowpart (inner_mode, op1);
4360 record_jump_cond (code, mode, SUBREG_REG (op0),
4361 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4362 reversed_nonequality);
4365 if (code == EQ && GET_CODE (op1) == SUBREG
4366 && (GET_MODE_SIZE (GET_MODE (op1))
4367 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4369 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4370 rtx tem = gen_lowpart (inner_mode, op0);
4372 record_jump_cond (code, mode, SUBREG_REG (op1),
4373 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4374 reversed_nonequality);
4377 /* Similarly, if this is an NE comparison, and either is a SUBREG
4378 making a smaller mode, we know the whole thing is also NE. */
4380 /* Note that GET_MODE (op0) may not equal MODE;
4381 if we test MODE instead, we can get an infinite recursion
4382 alternating between two modes each wider than MODE. */
4384 if (code == NE && GET_CODE (op0) == SUBREG
4385 && subreg_lowpart_p (op0)
4386 && (GET_MODE_SIZE (GET_MODE (op0))
4387 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4389 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4390 rtx tem = gen_lowpart (inner_mode, op1);
4392 record_jump_cond (code, mode, SUBREG_REG (op0),
4393 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4394 reversed_nonequality);
4397 if (code == NE && GET_CODE (op1) == SUBREG
4398 && subreg_lowpart_p (op1)
4399 && (GET_MODE_SIZE (GET_MODE (op1))
4400 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4402 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4403 rtx tem = gen_lowpart (inner_mode, op0);
4405 record_jump_cond (code, mode, SUBREG_REG (op1),
4406 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4407 reversed_nonequality);
4410 /* Hash both operands. */
4412 do_not_record = 0;
4413 hash_arg_in_memory = 0;
4414 op0_hash = HASH (op0, mode);
4415 op0_in_memory = hash_arg_in_memory;
4417 if (do_not_record)
4418 return;
4420 do_not_record = 0;
4421 hash_arg_in_memory = 0;
4422 op1_hash = HASH (op1, mode);
4423 op1_in_memory = hash_arg_in_memory;
4425 if (do_not_record)
4426 return;
4428 /* Look up both operands. */
4429 op0_elt = lookup (op0, op0_hash, mode);
4430 op1_elt = lookup (op1, op1_hash, mode);
4432 /* If both operands are already equivalent or if they are not in the
4433 table but are identical, do nothing. */
4434 if ((op0_elt != 0 && op1_elt != 0
4435 && op0_elt->first_same_value == op1_elt->first_same_value)
4436 || op0 == op1 || rtx_equal_p (op0, op1))
4437 return;
4439 /* If we aren't setting two things equal all we can do is save this
4440 comparison. Similarly if this is floating-point. In the latter
4441 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4442 If we record the equality, we might inadvertently delete code
4443 whose intent was to change -0 to +0. */
4445 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4447 struct qty_table_elem *ent;
4448 int qty;
4450 /* If we reversed a floating-point comparison, if OP0 is not a
4451 register, or if OP1 is neither a register or constant, we can't
4452 do anything. */
4454 if (!REG_P (op1))
4455 op1 = equiv_constant (op1);
4457 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4458 || !REG_P (op0) || op1 == 0)
4459 return;
4461 /* Put OP0 in the hash table if it isn't already. This gives it a
4462 new quantity number. */
4463 if (op0_elt == 0)
4465 if (insert_regs (op0, NULL, 0))
4467 rehash_using_reg (op0);
4468 op0_hash = HASH (op0, mode);
4470 /* If OP0 is contained in OP1, this changes its hash code
4471 as well. Faster to rehash than to check, except
4472 for the simple case of a constant. */
4473 if (! CONSTANT_P (op1))
4474 op1_hash = HASH (op1,mode);
4477 op0_elt = insert (op0, NULL, op0_hash, mode);
4478 op0_elt->in_memory = op0_in_memory;
4481 qty = REG_QTY (REGNO (op0));
4482 ent = &qty_table[qty];
4484 ent->comparison_code = code;
4485 if (REG_P (op1))
4487 /* Look it up again--in case op0 and op1 are the same. */
4488 op1_elt = lookup (op1, op1_hash, mode);
4490 /* Put OP1 in the hash table so it gets a new quantity number. */
4491 if (op1_elt == 0)
4493 if (insert_regs (op1, NULL, 0))
4495 rehash_using_reg (op1);
4496 op1_hash = HASH (op1, mode);
4499 op1_elt = insert (op1, NULL, op1_hash, mode);
4500 op1_elt->in_memory = op1_in_memory;
4503 ent->comparison_const = NULL_RTX;
4504 ent->comparison_qty = REG_QTY (REGNO (op1));
4506 else
4508 ent->comparison_const = op1;
4509 ent->comparison_qty = -1;
4512 return;
4515 /* If either side is still missing an equivalence, make it now,
4516 then merge the equivalences. */
4518 if (op0_elt == 0)
4520 if (insert_regs (op0, NULL, 0))
4522 rehash_using_reg (op0);
4523 op0_hash = HASH (op0, mode);
4526 op0_elt = insert (op0, NULL, op0_hash, mode);
4527 op0_elt->in_memory = op0_in_memory;
4530 if (op1_elt == 0)
4532 if (insert_regs (op1, NULL, 0))
4534 rehash_using_reg (op1);
4535 op1_hash = HASH (op1, mode);
4538 op1_elt = insert (op1, NULL, op1_hash, mode);
4539 op1_elt->in_memory = op1_in_memory;
4542 merge_equiv_classes (op0_elt, op1_elt);
4543 last_jump_equiv_class = op0_elt;
4546 /* CSE processing for one instruction.
4547 First simplify sources and addresses of all assignments
4548 in the instruction, using previously-computed equivalents values.
4549 Then install the new sources and destinations in the table
4550 of available values.
4552 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4553 the insn. It means that INSN is inside libcall block. In this
4554 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4556 /* Data on one SET contained in the instruction. */
4558 struct set
4560 /* The SET rtx itself. */
4561 rtx rtl;
4562 /* The SET_SRC of the rtx (the original value, if it is changing). */
4563 rtx src;
4564 /* The hash-table element for the SET_SRC of the SET. */
4565 struct table_elt *src_elt;
4566 /* Hash value for the SET_SRC. */
4567 unsigned src_hash;
4568 /* Hash value for the SET_DEST. */
4569 unsigned dest_hash;
4570 /* The SET_DEST, with SUBREG, etc., stripped. */
4571 rtx inner_dest;
4572 /* Nonzero if the SET_SRC is in memory. */
4573 char src_in_memory;
4574 /* Nonzero if the SET_SRC contains something
4575 whose value cannot be predicted and understood. */
4576 char src_volatile;
4577 /* Original machine mode, in case it becomes a CONST_INT.
4578 The size of this field should match the size of the mode
4579 field of struct rtx_def (see rtl.h). */
4580 ENUM_BITFIELD(machine_mode) mode : 8;
4581 /* A constant equivalent for SET_SRC, if any. */
4582 rtx src_const;
4583 /* Original SET_SRC value used for libcall notes. */
4584 rtx orig_src;
4585 /* Hash value of constant equivalent for SET_SRC. */
4586 unsigned src_const_hash;
4587 /* Table entry for constant equivalent for SET_SRC, if any. */
4588 struct table_elt *src_const_elt;
4591 static void
4592 cse_insn (rtx insn, rtx libcall_insn)
4594 rtx x = PATTERN (insn);
4595 int i;
4596 rtx tem;
4597 int n_sets = 0;
4599 #ifdef HAVE_cc0
4600 /* Records what this insn does to set CC0. */
4601 rtx this_insn_cc0 = 0;
4602 enum machine_mode this_insn_cc0_mode = VOIDmode;
4603 #endif
4605 rtx src_eqv = 0;
4606 struct table_elt *src_eqv_elt = 0;
4607 int src_eqv_volatile = 0;
4608 int src_eqv_in_memory = 0;
4609 unsigned src_eqv_hash = 0;
4611 struct set *sets = (struct set *) 0;
4613 this_insn = insn;
4615 /* Find all the SETs and CLOBBERs in this instruction.
4616 Record all the SETs in the array `set' and count them.
4617 Also determine whether there is a CLOBBER that invalidates
4618 all memory references, or all references at varying addresses. */
4620 if (CALL_P (insn))
4622 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4624 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4625 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4626 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4630 if (GET_CODE (x) == SET)
4632 sets = alloca (sizeof (struct set));
4633 sets[0].rtl = x;
4635 /* Ignore SETs that are unconditional jumps.
4636 They never need cse processing, so this does not hurt.
4637 The reason is not efficiency but rather
4638 so that we can test at the end for instructions
4639 that have been simplified to unconditional jumps
4640 and not be misled by unchanged instructions
4641 that were unconditional jumps to begin with. */
4642 if (SET_DEST (x) == pc_rtx
4643 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4646 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4647 The hard function value register is used only once, to copy to
4648 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4649 Ensure we invalidate the destination register. On the 80386 no
4650 other code would invalidate it since it is a fixed_reg.
4651 We need not check the return of apply_change_group; see canon_reg. */
4653 else if (GET_CODE (SET_SRC (x)) == CALL)
4655 canon_reg (SET_SRC (x), insn);
4656 apply_change_group ();
4657 fold_rtx (SET_SRC (x), insn);
4658 invalidate (SET_DEST (x), VOIDmode);
4660 else
4661 n_sets = 1;
4663 else if (GET_CODE (x) == PARALLEL)
4665 int lim = XVECLEN (x, 0);
4667 sets = alloca (lim * sizeof (struct set));
4669 /* Find all regs explicitly clobbered in this insn,
4670 and ensure they are not replaced with any other regs
4671 elsewhere in this insn.
4672 When a reg that is clobbered is also used for input,
4673 we should presume that that is for a reason,
4674 and we should not substitute some other register
4675 which is not supposed to be clobbered.
4676 Therefore, this loop cannot be merged into the one below
4677 because a CALL may precede a CLOBBER and refer to the
4678 value clobbered. We must not let a canonicalization do
4679 anything in that case. */
4680 for (i = 0; i < lim; i++)
4682 rtx y = XVECEXP (x, 0, i);
4683 if (GET_CODE (y) == CLOBBER)
4685 rtx clobbered = XEXP (y, 0);
4687 if (REG_P (clobbered)
4688 || GET_CODE (clobbered) == SUBREG)
4689 invalidate (clobbered, VOIDmode);
4690 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4691 || GET_CODE (clobbered) == ZERO_EXTRACT)
4692 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4696 for (i = 0; i < lim; i++)
4698 rtx y = XVECEXP (x, 0, i);
4699 if (GET_CODE (y) == SET)
4701 /* As above, we ignore unconditional jumps and call-insns and
4702 ignore the result of apply_change_group. */
4703 if (GET_CODE (SET_SRC (y)) == CALL)
4705 canon_reg (SET_SRC (y), insn);
4706 apply_change_group ();
4707 fold_rtx (SET_SRC (y), insn);
4708 invalidate (SET_DEST (y), VOIDmode);
4710 else if (SET_DEST (y) == pc_rtx
4711 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4713 else
4714 sets[n_sets++].rtl = y;
4716 else if (GET_CODE (y) == CLOBBER)
4718 /* If we clobber memory, canon the address.
4719 This does nothing when a register is clobbered
4720 because we have already invalidated the reg. */
4721 if (MEM_P (XEXP (y, 0)))
4722 canon_reg (XEXP (y, 0), NULL_RTX);
4724 else if (GET_CODE (y) == USE
4725 && ! (REG_P (XEXP (y, 0))
4726 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4727 canon_reg (y, NULL_RTX);
4728 else if (GET_CODE (y) == CALL)
4730 /* The result of apply_change_group can be ignored; see
4731 canon_reg. */
4732 canon_reg (y, insn);
4733 apply_change_group ();
4734 fold_rtx (y, insn);
4738 else if (GET_CODE (x) == CLOBBER)
4740 if (MEM_P (XEXP (x, 0)))
4741 canon_reg (XEXP (x, 0), NULL_RTX);
4744 /* Canonicalize a USE of a pseudo register or memory location. */
4745 else if (GET_CODE (x) == USE
4746 && ! (REG_P (XEXP (x, 0))
4747 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4748 canon_reg (XEXP (x, 0), NULL_RTX);
4749 else if (GET_CODE (x) == CALL)
4751 /* The result of apply_change_group can be ignored; see canon_reg. */
4752 canon_reg (x, insn);
4753 apply_change_group ();
4754 fold_rtx (x, insn);
4757 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4758 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4759 is handled specially for this case, and if it isn't set, then there will
4760 be no equivalence for the destination. */
4761 if (n_sets == 1 && REG_NOTES (insn) != 0
4762 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4763 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4764 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4766 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4767 XEXP (tem, 0) = src_eqv;
4770 /* Canonicalize sources and addresses of destinations.
4771 We do this in a separate pass to avoid problems when a MATCH_DUP is
4772 present in the insn pattern. In that case, we want to ensure that
4773 we don't break the duplicate nature of the pattern. So we will replace
4774 both operands at the same time. Otherwise, we would fail to find an
4775 equivalent substitution in the loop calling validate_change below.
4777 We used to suppress canonicalization of DEST if it appears in SRC,
4778 but we don't do this any more. */
4780 for (i = 0; i < n_sets; i++)
4782 rtx dest = SET_DEST (sets[i].rtl);
4783 rtx src = SET_SRC (sets[i].rtl);
4784 rtx new = canon_reg (src, insn);
4785 int insn_code;
4787 sets[i].orig_src = src;
4788 if ((REG_P (new) && REG_P (src)
4789 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4790 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4791 || (insn_code = recog_memoized (insn)) < 0
4792 || insn_data[insn_code].n_dups > 0)
4793 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4794 else
4795 SET_SRC (sets[i].rtl) = new;
4797 if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT)
4799 validate_change (insn, &XEXP (dest, 1),
4800 canon_reg (XEXP (dest, 1), insn), 1);
4801 validate_change (insn, &XEXP (dest, 2),
4802 canon_reg (XEXP (dest, 2), insn), 1);
4805 while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART
4806 || GET_CODE (dest) == ZERO_EXTRACT
4807 || GET_CODE (dest) == SIGN_EXTRACT)
4808 dest = XEXP (dest, 0);
4810 if (MEM_P (dest))
4811 canon_reg (dest, insn);
4814 /* Now that we have done all the replacements, we can apply the change
4815 group and see if they all work. Note that this will cause some
4816 canonicalizations that would have worked individually not to be applied
4817 because some other canonicalization didn't work, but this should not
4818 occur often.
4820 The result of apply_change_group can be ignored; see canon_reg. */
4822 apply_change_group ();
4824 /* Set sets[i].src_elt to the class each source belongs to.
4825 Detect assignments from or to volatile things
4826 and set set[i] to zero so they will be ignored
4827 in the rest of this function.
4829 Nothing in this loop changes the hash table or the register chains. */
4831 for (i = 0; i < n_sets; i++)
4833 rtx src, dest;
4834 rtx src_folded;
4835 struct table_elt *elt = 0, *p;
4836 enum machine_mode mode;
4837 rtx src_eqv_here;
4838 rtx src_const = 0;
4839 rtx src_related = 0;
4840 struct table_elt *src_const_elt = 0;
4841 int src_cost = MAX_COST;
4842 int src_eqv_cost = MAX_COST;
4843 int src_folded_cost = MAX_COST;
4844 int src_related_cost = MAX_COST;
4845 int src_elt_cost = MAX_COST;
4846 int src_regcost = MAX_COST;
4847 int src_eqv_regcost = MAX_COST;
4848 int src_folded_regcost = MAX_COST;
4849 int src_related_regcost = MAX_COST;
4850 int src_elt_regcost = MAX_COST;
4851 /* Set nonzero if we need to call force_const_mem on with the
4852 contents of src_folded before using it. */
4853 int src_folded_force_flag = 0;
4855 dest = SET_DEST (sets[i].rtl);
4856 src = SET_SRC (sets[i].rtl);
4858 /* If SRC is a constant that has no machine mode,
4859 hash it with the destination's machine mode.
4860 This way we can keep different modes separate. */
4862 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4863 sets[i].mode = mode;
4865 if (src_eqv)
4867 enum machine_mode eqvmode = mode;
4868 if (GET_CODE (dest) == STRICT_LOW_PART)
4869 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4870 do_not_record = 0;
4871 hash_arg_in_memory = 0;
4872 src_eqv_hash = HASH (src_eqv, eqvmode);
4874 /* Find the equivalence class for the equivalent expression. */
4876 if (!do_not_record)
4877 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4879 src_eqv_volatile = do_not_record;
4880 src_eqv_in_memory = hash_arg_in_memory;
4883 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4884 value of the INNER register, not the destination. So it is not
4885 a valid substitution for the source. But save it for later. */
4886 if (GET_CODE (dest) == STRICT_LOW_PART)
4887 src_eqv_here = 0;
4888 else
4889 src_eqv_here = src_eqv;
4891 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4892 simplified result, which may not necessarily be valid. */
4893 src_folded = fold_rtx (src, insn);
4895 #if 0
4896 /* ??? This caused bad code to be generated for the m68k port with -O2.
4897 Suppose src is (CONST_INT -1), and that after truncation src_folded
4898 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4899 At the end we will add src and src_const to the same equivalence
4900 class. We now have 3 and -1 on the same equivalence class. This
4901 causes later instructions to be mis-optimized. */
4902 /* If storing a constant in a bitfield, pre-truncate the constant
4903 so we will be able to record it later. */
4904 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
4905 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
4907 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4909 if (GET_CODE (src) == CONST_INT
4910 && GET_CODE (width) == CONST_INT
4911 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4912 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4913 src_folded
4914 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4915 << INTVAL (width)) - 1));
4917 #endif
4919 /* Compute SRC's hash code, and also notice if it
4920 should not be recorded at all. In that case,
4921 prevent any further processing of this assignment. */
4922 do_not_record = 0;
4923 hash_arg_in_memory = 0;
4925 sets[i].src = src;
4926 sets[i].src_hash = HASH (src, mode);
4927 sets[i].src_volatile = do_not_record;
4928 sets[i].src_in_memory = hash_arg_in_memory;
4930 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4931 a pseudo, do not record SRC. Using SRC as a replacement for
4932 anything else will be incorrect in that situation. Note that
4933 this usually occurs only for stack slots, in which case all the
4934 RTL would be referring to SRC, so we don't lose any optimization
4935 opportunities by not having SRC in the hash table. */
4937 if (MEM_P (src)
4938 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4939 && REG_P (dest)
4940 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4941 sets[i].src_volatile = 1;
4943 #if 0
4944 /* It is no longer clear why we used to do this, but it doesn't
4945 appear to still be needed. So let's try without it since this
4946 code hurts cse'ing widened ops. */
4947 /* If source is a paradoxical subreg (such as QI treated as an SI),
4948 treat it as volatile. It may do the work of an SI in one context
4949 where the extra bits are not being used, but cannot replace an SI
4950 in general. */
4951 if (GET_CODE (src) == SUBREG
4952 && (GET_MODE_SIZE (GET_MODE (src))
4953 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
4954 sets[i].src_volatile = 1;
4955 #endif
4957 /* Locate all possible equivalent forms for SRC. Try to replace
4958 SRC in the insn with each cheaper equivalent.
4960 We have the following types of equivalents: SRC itself, a folded
4961 version, a value given in a REG_EQUAL note, or a value related
4962 to a constant.
4964 Each of these equivalents may be part of an additional class
4965 of equivalents (if more than one is in the table, they must be in
4966 the same class; we check for this).
4968 If the source is volatile, we don't do any table lookups.
4970 We note any constant equivalent for possible later use in a
4971 REG_NOTE. */
4973 if (!sets[i].src_volatile)
4974 elt = lookup (src, sets[i].src_hash, mode);
4976 sets[i].src_elt = elt;
4978 if (elt && src_eqv_here && src_eqv_elt)
4980 if (elt->first_same_value != src_eqv_elt->first_same_value)
4982 /* The REG_EQUAL is indicating that two formerly distinct
4983 classes are now equivalent. So merge them. */
4984 merge_equiv_classes (elt, src_eqv_elt);
4985 src_eqv_hash = HASH (src_eqv, elt->mode);
4986 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4989 src_eqv_here = 0;
4992 else if (src_eqv_elt)
4993 elt = src_eqv_elt;
4995 /* Try to find a constant somewhere and record it in `src_const'.
4996 Record its table element, if any, in `src_const_elt'. Look in
4997 any known equivalences first. (If the constant is not in the
4998 table, also set `sets[i].src_const_hash'). */
4999 if (elt)
5000 for (p = elt->first_same_value; p; p = p->next_same_value)
5001 if (p->is_const)
5003 src_const = p->exp;
5004 src_const_elt = elt;
5005 break;
5008 if (src_const == 0
5009 && (CONSTANT_P (src_folded)
5010 /* Consider (minus (label_ref L1) (label_ref L2)) as
5011 "constant" here so we will record it. This allows us
5012 to fold switch statements when an ADDR_DIFF_VEC is used. */
5013 || (GET_CODE (src_folded) == MINUS
5014 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5015 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5016 src_const = src_folded, src_const_elt = elt;
5017 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5018 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5020 /* If we don't know if the constant is in the table, get its
5021 hash code and look it up. */
5022 if (src_const && src_const_elt == 0)
5024 sets[i].src_const_hash = HASH (src_const, mode);
5025 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5028 sets[i].src_const = src_const;
5029 sets[i].src_const_elt = src_const_elt;
5031 /* If the constant and our source are both in the table, mark them as
5032 equivalent. Otherwise, if a constant is in the table but the source
5033 isn't, set ELT to it. */
5034 if (src_const_elt && elt
5035 && src_const_elt->first_same_value != elt->first_same_value)
5036 merge_equiv_classes (elt, src_const_elt);
5037 else if (src_const_elt && elt == 0)
5038 elt = src_const_elt;
5040 /* See if there is a register linearly related to a constant
5041 equivalent of SRC. */
5042 if (src_const
5043 && (GET_CODE (src_const) == CONST
5044 || (src_const_elt && src_const_elt->related_value != 0)))
5046 src_related = use_related_value (src_const, src_const_elt);
5047 if (src_related)
5049 struct table_elt *src_related_elt
5050 = lookup (src_related, HASH (src_related, mode), mode);
5051 if (src_related_elt && elt)
5053 if (elt->first_same_value
5054 != src_related_elt->first_same_value)
5055 /* This can occur when we previously saw a CONST
5056 involving a SYMBOL_REF and then see the SYMBOL_REF
5057 twice. Merge the involved classes. */
5058 merge_equiv_classes (elt, src_related_elt);
5060 src_related = 0;
5061 src_related_elt = 0;
5063 else if (src_related_elt && elt == 0)
5064 elt = src_related_elt;
5068 /* See if we have a CONST_INT that is already in a register in a
5069 wider mode. */
5071 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5072 && GET_MODE_CLASS (mode) == MODE_INT
5073 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5075 enum machine_mode wider_mode;
5077 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5078 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5079 && src_related == 0;
5080 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5082 struct table_elt *const_elt
5083 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5085 if (const_elt == 0)
5086 continue;
5088 for (const_elt = const_elt->first_same_value;
5089 const_elt; const_elt = const_elt->next_same_value)
5090 if (REG_P (const_elt->exp))
5092 src_related = gen_lowpart (mode,
5093 const_elt->exp);
5094 break;
5099 /* Another possibility is that we have an AND with a constant in
5100 a mode narrower than a word. If so, it might have been generated
5101 as part of an "if" which would narrow the AND. If we already
5102 have done the AND in a wider mode, we can use a SUBREG of that
5103 value. */
5105 if (flag_expensive_optimizations && ! src_related
5106 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5107 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5109 enum machine_mode tmode;
5110 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5112 for (tmode = GET_MODE_WIDER_MODE (mode);
5113 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5114 tmode = GET_MODE_WIDER_MODE (tmode))
5116 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5117 struct table_elt *larger_elt;
5119 if (inner)
5121 PUT_MODE (new_and, tmode);
5122 XEXP (new_and, 0) = inner;
5123 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5124 if (larger_elt == 0)
5125 continue;
5127 for (larger_elt = larger_elt->first_same_value;
5128 larger_elt; larger_elt = larger_elt->next_same_value)
5129 if (REG_P (larger_elt->exp))
5131 src_related
5132 = gen_lowpart (mode, larger_elt->exp);
5133 break;
5136 if (src_related)
5137 break;
5142 #ifdef LOAD_EXTEND_OP
5143 /* See if a MEM has already been loaded with a widening operation;
5144 if it has, we can use a subreg of that. Many CISC machines
5145 also have such operations, but this is only likely to be
5146 beneficial on these machines. */
5148 if (flag_expensive_optimizations && src_related == 0
5149 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5150 && GET_MODE_CLASS (mode) == MODE_INT
5151 && MEM_P (src) && ! do_not_record
5152 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5154 enum machine_mode tmode;
5156 /* Set what we are trying to extend and the operation it might
5157 have been extended with. */
5158 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5159 XEXP (memory_extend_rtx, 0) = src;
5161 for (tmode = GET_MODE_WIDER_MODE (mode);
5162 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5163 tmode = GET_MODE_WIDER_MODE (tmode))
5165 struct table_elt *larger_elt;
5167 PUT_MODE (memory_extend_rtx, tmode);
5168 larger_elt = lookup (memory_extend_rtx,
5169 HASH (memory_extend_rtx, tmode), tmode);
5170 if (larger_elt == 0)
5171 continue;
5173 for (larger_elt = larger_elt->first_same_value;
5174 larger_elt; larger_elt = larger_elt->next_same_value)
5175 if (REG_P (larger_elt->exp))
5177 src_related = gen_lowpart (mode,
5178 larger_elt->exp);
5179 break;
5182 if (src_related)
5183 break;
5186 #endif /* LOAD_EXTEND_OP */
5188 if (src == src_folded)
5189 src_folded = 0;
5191 /* At this point, ELT, if nonzero, points to a class of expressions
5192 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5193 and SRC_RELATED, if nonzero, each contain additional equivalent
5194 expressions. Prune these latter expressions by deleting expressions
5195 already in the equivalence class.
5197 Check for an equivalent identical to the destination. If found,
5198 this is the preferred equivalent since it will likely lead to
5199 elimination of the insn. Indicate this by placing it in
5200 `src_related'. */
5202 if (elt)
5203 elt = elt->first_same_value;
5204 for (p = elt; p; p = p->next_same_value)
5206 enum rtx_code code = GET_CODE (p->exp);
5208 /* If the expression is not valid, ignore it. Then we do not
5209 have to check for validity below. In most cases, we can use
5210 `rtx_equal_p', since canonicalization has already been done. */
5211 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5212 continue;
5214 /* Also skip paradoxical subregs, unless that's what we're
5215 looking for. */
5216 if (code == SUBREG
5217 && (GET_MODE_SIZE (GET_MODE (p->exp))
5218 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5219 && ! (src != 0
5220 && GET_CODE (src) == SUBREG
5221 && GET_MODE (src) == GET_MODE (p->exp)
5222 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5223 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5224 continue;
5226 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5227 src = 0;
5228 else if (src_folded && GET_CODE (src_folded) == code
5229 && rtx_equal_p (src_folded, p->exp))
5230 src_folded = 0;
5231 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5232 && rtx_equal_p (src_eqv_here, p->exp))
5233 src_eqv_here = 0;
5234 else if (src_related && GET_CODE (src_related) == code
5235 && rtx_equal_p (src_related, p->exp))
5236 src_related = 0;
5238 /* This is the same as the destination of the insns, we want
5239 to prefer it. Copy it to src_related. The code below will
5240 then give it a negative cost. */
5241 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5242 src_related = dest;
5245 /* Find the cheapest valid equivalent, trying all the available
5246 possibilities. Prefer items not in the hash table to ones
5247 that are when they are equal cost. Note that we can never
5248 worsen an insn as the current contents will also succeed.
5249 If we find an equivalent identical to the destination, use it as best,
5250 since this insn will probably be eliminated in that case. */
5251 if (src)
5253 if (rtx_equal_p (src, dest))
5254 src_cost = src_regcost = -1;
5255 else
5257 src_cost = COST (src);
5258 src_regcost = approx_reg_cost (src);
5262 if (src_eqv_here)
5264 if (rtx_equal_p (src_eqv_here, dest))
5265 src_eqv_cost = src_eqv_regcost = -1;
5266 else
5268 src_eqv_cost = COST (src_eqv_here);
5269 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5273 if (src_folded)
5275 if (rtx_equal_p (src_folded, dest))
5276 src_folded_cost = src_folded_regcost = -1;
5277 else
5279 src_folded_cost = COST (src_folded);
5280 src_folded_regcost = approx_reg_cost (src_folded);
5284 if (src_related)
5286 if (rtx_equal_p (src_related, dest))
5287 src_related_cost = src_related_regcost = -1;
5288 else
5290 src_related_cost = COST (src_related);
5291 src_related_regcost = approx_reg_cost (src_related);
5295 /* If this was an indirect jump insn, a known label will really be
5296 cheaper even though it looks more expensive. */
5297 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5298 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5300 /* Terminate loop when replacement made. This must terminate since
5301 the current contents will be tested and will always be valid. */
5302 while (1)
5304 rtx trial;
5306 /* Skip invalid entries. */
5307 while (elt && !REG_P (elt->exp)
5308 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5309 elt = elt->next_same_value;
5311 /* A paradoxical subreg would be bad here: it'll be the right
5312 size, but later may be adjusted so that the upper bits aren't
5313 what we want. So reject it. */
5314 if (elt != 0
5315 && GET_CODE (elt->exp) == SUBREG
5316 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5317 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5318 /* It is okay, though, if the rtx we're trying to match
5319 will ignore any of the bits we can't predict. */
5320 && ! (src != 0
5321 && GET_CODE (src) == SUBREG
5322 && GET_MODE (src) == GET_MODE (elt->exp)
5323 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5324 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5326 elt = elt->next_same_value;
5327 continue;
5330 if (elt)
5332 src_elt_cost = elt->cost;
5333 src_elt_regcost = elt->regcost;
5336 /* Find cheapest and skip it for the next time. For items
5337 of equal cost, use this order:
5338 src_folded, src, src_eqv, src_related and hash table entry. */
5339 if (src_folded
5340 && preferable (src_folded_cost, src_folded_regcost,
5341 src_cost, src_regcost) <= 0
5342 && preferable (src_folded_cost, src_folded_regcost,
5343 src_eqv_cost, src_eqv_regcost) <= 0
5344 && preferable (src_folded_cost, src_folded_regcost,
5345 src_related_cost, src_related_regcost) <= 0
5346 && preferable (src_folded_cost, src_folded_regcost,
5347 src_elt_cost, src_elt_regcost) <= 0)
5349 trial = src_folded, src_folded_cost = MAX_COST;
5350 if (src_folded_force_flag)
5352 rtx forced = force_const_mem (mode, trial);
5353 if (forced)
5354 trial = forced;
5357 else if (src
5358 && preferable (src_cost, src_regcost,
5359 src_eqv_cost, src_eqv_regcost) <= 0
5360 && preferable (src_cost, src_regcost,
5361 src_related_cost, src_related_regcost) <= 0
5362 && preferable (src_cost, src_regcost,
5363 src_elt_cost, src_elt_regcost) <= 0)
5364 trial = src, src_cost = MAX_COST;
5365 else if (src_eqv_here
5366 && preferable (src_eqv_cost, src_eqv_regcost,
5367 src_related_cost, src_related_regcost) <= 0
5368 && preferable (src_eqv_cost, src_eqv_regcost,
5369 src_elt_cost, src_elt_regcost) <= 0)
5370 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5371 else if (src_related
5372 && preferable (src_related_cost, src_related_regcost,
5373 src_elt_cost, src_elt_regcost) <= 0)
5374 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5375 else
5377 trial = copy_rtx (elt->exp);
5378 elt = elt->next_same_value;
5379 src_elt_cost = MAX_COST;
5382 /* We don't normally have an insn matching (set (pc) (pc)), so
5383 check for this separately here. We will delete such an
5384 insn below.
5386 For other cases such as a table jump or conditional jump
5387 where we know the ultimate target, go ahead and replace the
5388 operand. While that may not make a valid insn, we will
5389 reemit the jump below (and also insert any necessary
5390 barriers). */
5391 if (n_sets == 1 && dest == pc_rtx
5392 && (trial == pc_rtx
5393 || (GET_CODE (trial) == LABEL_REF
5394 && ! condjump_p (insn))))
5396 SET_SRC (sets[i].rtl) = trial;
5397 cse_jumps_altered = 1;
5398 break;
5401 /* Look for a substitution that makes a valid insn. */
5402 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5404 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5406 /* If we just made a substitution inside a libcall, then we
5407 need to make the same substitution in any notes attached
5408 to the RETVAL insn. */
5409 if (libcall_insn
5410 && (REG_P (sets[i].orig_src)
5411 || GET_CODE (sets[i].orig_src) == SUBREG
5412 || MEM_P (sets[i].orig_src)))
5414 rtx note = find_reg_equal_equiv_note (libcall_insn);
5415 if (note != 0)
5416 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5417 sets[i].orig_src,
5418 copy_rtx (new));
5421 /* The result of apply_change_group can be ignored; see
5422 canon_reg. */
5424 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5425 apply_change_group ();
5426 break;
5429 /* If we previously found constant pool entries for
5430 constants and this is a constant, try making a
5431 pool entry. Put it in src_folded unless we already have done
5432 this since that is where it likely came from. */
5434 else if (constant_pool_entries_cost
5435 && CONSTANT_P (trial)
5436 /* Reject cases that will abort in decode_rtx_const.
5437 On the alpha when simplifying a switch, we get
5438 (const (truncate (minus (label_ref) (label_ref)))). */
5439 && ! (GET_CODE (trial) == CONST
5440 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5441 /* Likewise on IA-64, except without the truncate. */
5442 && ! (GET_CODE (trial) == CONST
5443 && GET_CODE (XEXP (trial, 0)) == MINUS
5444 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5445 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5446 && (src_folded == 0
5447 || (!MEM_P (src_folded)
5448 && ! src_folded_force_flag))
5449 && GET_MODE_CLASS (mode) != MODE_CC
5450 && mode != VOIDmode)
5452 src_folded_force_flag = 1;
5453 src_folded = trial;
5454 src_folded_cost = constant_pool_entries_cost;
5455 src_folded_regcost = constant_pool_entries_regcost;
5459 src = SET_SRC (sets[i].rtl);
5461 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5462 However, there is an important exception: If both are registers
5463 that are not the head of their equivalence class, replace SET_SRC
5464 with the head of the class. If we do not do this, we will have
5465 both registers live over a portion of the basic block. This way,
5466 their lifetimes will likely abut instead of overlapping. */
5467 if (REG_P (dest)
5468 && REGNO_QTY_VALID_P (REGNO (dest)))
5470 int dest_q = REG_QTY (REGNO (dest));
5471 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5473 if (dest_ent->mode == GET_MODE (dest)
5474 && dest_ent->first_reg != REGNO (dest)
5475 && REG_P (src) && REGNO (src) == REGNO (dest)
5476 /* Don't do this if the original insn had a hard reg as
5477 SET_SRC or SET_DEST. */
5478 && (!REG_P (sets[i].src)
5479 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5480 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5481 /* We can't call canon_reg here because it won't do anything if
5482 SRC is a hard register. */
5484 int src_q = REG_QTY (REGNO (src));
5485 struct qty_table_elem *src_ent = &qty_table[src_q];
5486 int first = src_ent->first_reg;
5487 rtx new_src
5488 = (first >= FIRST_PSEUDO_REGISTER
5489 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5491 /* We must use validate-change even for this, because this
5492 might be a special no-op instruction, suitable only to
5493 tag notes onto. */
5494 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5496 src = new_src;
5497 /* If we had a constant that is cheaper than what we are now
5498 setting SRC to, use that constant. We ignored it when we
5499 thought we could make this into a no-op. */
5500 if (src_const && COST (src_const) < COST (src)
5501 && validate_change (insn, &SET_SRC (sets[i].rtl),
5502 src_const, 0))
5503 src = src_const;
5508 /* If we made a change, recompute SRC values. */
5509 if (src != sets[i].src)
5511 cse_altered = 1;
5512 do_not_record = 0;
5513 hash_arg_in_memory = 0;
5514 sets[i].src = src;
5515 sets[i].src_hash = HASH (src, mode);
5516 sets[i].src_volatile = do_not_record;
5517 sets[i].src_in_memory = hash_arg_in_memory;
5518 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5521 /* If this is a single SET, we are setting a register, and we have an
5522 equivalent constant, we want to add a REG_NOTE. We don't want
5523 to write a REG_EQUAL note for a constant pseudo since verifying that
5524 that pseudo hasn't been eliminated is a pain. Such a note also
5525 won't help anything.
5527 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5528 which can be created for a reference to a compile time computable
5529 entry in a jump table. */
5531 if (n_sets == 1 && src_const && REG_P (dest)
5532 && !REG_P (src_const)
5533 && ! (GET_CODE (src_const) == CONST
5534 && GET_CODE (XEXP (src_const, 0)) == MINUS
5535 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5536 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5538 /* We only want a REG_EQUAL note if src_const != src. */
5539 if (! rtx_equal_p (src, src_const))
5541 /* Make sure that the rtx is not shared. */
5542 src_const = copy_rtx (src_const);
5544 /* Record the actual constant value in a REG_EQUAL note,
5545 making a new one if one does not already exist. */
5546 set_unique_reg_note (insn, REG_EQUAL, src_const);
5550 /* Now deal with the destination. */
5551 do_not_record = 0;
5553 /* Look within any SIGN_EXTRACT or ZERO_EXTRACT
5554 to the MEM or REG within it. */
5555 while (GET_CODE (dest) == SIGN_EXTRACT
5556 || GET_CODE (dest) == ZERO_EXTRACT
5557 || GET_CODE (dest) == SUBREG
5558 || GET_CODE (dest) == STRICT_LOW_PART)
5559 dest = XEXP (dest, 0);
5561 sets[i].inner_dest = dest;
5563 if (MEM_P (dest))
5565 #ifdef PUSH_ROUNDING
5566 /* Stack pushes invalidate the stack pointer. */
5567 rtx addr = XEXP (dest, 0);
5568 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5569 && XEXP (addr, 0) == stack_pointer_rtx)
5570 invalidate (stack_pointer_rtx, Pmode);
5571 #endif
5572 dest = fold_rtx (dest, insn);
5575 /* Compute the hash code of the destination now,
5576 before the effects of this instruction are recorded,
5577 since the register values used in the address computation
5578 are those before this instruction. */
5579 sets[i].dest_hash = HASH (dest, mode);
5581 /* Don't enter a bit-field in the hash table
5582 because the value in it after the store
5583 may not equal what was stored, due to truncation. */
5585 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5586 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
5588 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5590 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5591 && GET_CODE (width) == CONST_INT
5592 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5593 && ! (INTVAL (src_const)
5594 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5595 /* Exception: if the value is constant,
5596 and it won't be truncated, record it. */
5598 else
5600 /* This is chosen so that the destination will be invalidated
5601 but no new value will be recorded.
5602 We must invalidate because sometimes constant
5603 values can be recorded for bitfields. */
5604 sets[i].src_elt = 0;
5605 sets[i].src_volatile = 1;
5606 src_eqv = 0;
5607 src_eqv_elt = 0;
5611 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5612 the insn. */
5613 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5615 /* One less use of the label this insn used to jump to. */
5616 delete_insn (insn);
5617 cse_jumps_altered = 1;
5618 /* No more processing for this set. */
5619 sets[i].rtl = 0;
5622 /* If this SET is now setting PC to a label, we know it used to
5623 be a conditional or computed branch. */
5624 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF)
5626 /* Now emit a BARRIER after the unconditional jump. */
5627 if (NEXT_INSN (insn) == 0
5628 || !BARRIER_P (NEXT_INSN (insn)))
5629 emit_barrier_after (insn);
5631 /* We reemit the jump in as many cases as possible just in
5632 case the form of an unconditional jump is significantly
5633 different than a computed jump or conditional jump.
5635 If this insn has multiple sets, then reemitting the
5636 jump is nontrivial. So instead we just force rerecognition
5637 and hope for the best. */
5638 if (n_sets == 1)
5640 rtx new, note;
5642 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5643 JUMP_LABEL (new) = XEXP (src, 0);
5644 LABEL_NUSES (XEXP (src, 0))++;
5646 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5647 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5648 if (note)
5650 XEXP (note, 1) = NULL_RTX;
5651 REG_NOTES (new) = note;
5654 delete_insn (insn);
5655 insn = new;
5657 /* Now emit a BARRIER after the unconditional jump. */
5658 if (NEXT_INSN (insn) == 0
5659 || !BARRIER_P (NEXT_INSN (insn)))
5660 emit_barrier_after (insn);
5662 else
5663 INSN_CODE (insn) = -1;
5665 /* Do not bother deleting any unreachable code,
5666 let jump/flow do that. */
5668 cse_jumps_altered = 1;
5669 sets[i].rtl = 0;
5672 /* If destination is volatile, invalidate it and then do no further
5673 processing for this assignment. */
5675 else if (do_not_record)
5677 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5678 invalidate (dest, VOIDmode);
5679 else if (MEM_P (dest))
5681 /* Outgoing arguments for a libcall don't
5682 affect any recorded expressions. */
5683 if (! libcall_insn || insn == libcall_insn)
5684 invalidate (dest, VOIDmode);
5686 else if (GET_CODE (dest) == STRICT_LOW_PART
5687 || GET_CODE (dest) == ZERO_EXTRACT)
5688 invalidate (XEXP (dest, 0), GET_MODE (dest));
5689 sets[i].rtl = 0;
5692 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5693 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5695 #ifdef HAVE_cc0
5696 /* If setting CC0, record what it was set to, or a constant, if it
5697 is equivalent to a constant. If it is being set to a floating-point
5698 value, make a COMPARE with the appropriate constant of 0. If we
5699 don't do this, later code can interpret this as a test against
5700 const0_rtx, which can cause problems if we try to put it into an
5701 insn as a floating-point operand. */
5702 if (dest == cc0_rtx)
5704 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5705 this_insn_cc0_mode = mode;
5706 if (FLOAT_MODE_P (mode))
5707 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5708 CONST0_RTX (mode));
5710 #endif
5713 /* Now enter all non-volatile source expressions in the hash table
5714 if they are not already present.
5715 Record their equivalence classes in src_elt.
5716 This way we can insert the corresponding destinations into
5717 the same classes even if the actual sources are no longer in them
5718 (having been invalidated). */
5720 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5721 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5723 struct table_elt *elt;
5724 struct table_elt *classp = sets[0].src_elt;
5725 rtx dest = SET_DEST (sets[0].rtl);
5726 enum machine_mode eqvmode = GET_MODE (dest);
5728 if (GET_CODE (dest) == STRICT_LOW_PART)
5730 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5731 classp = 0;
5733 if (insert_regs (src_eqv, classp, 0))
5735 rehash_using_reg (src_eqv);
5736 src_eqv_hash = HASH (src_eqv, eqvmode);
5738 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5739 elt->in_memory = src_eqv_in_memory;
5740 src_eqv_elt = elt;
5742 /* Check to see if src_eqv_elt is the same as a set source which
5743 does not yet have an elt, and if so set the elt of the set source
5744 to src_eqv_elt. */
5745 for (i = 0; i < n_sets; i++)
5746 if (sets[i].rtl && sets[i].src_elt == 0
5747 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5748 sets[i].src_elt = src_eqv_elt;
5751 for (i = 0; i < n_sets; i++)
5752 if (sets[i].rtl && ! sets[i].src_volatile
5753 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5755 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5757 /* REG_EQUAL in setting a STRICT_LOW_PART
5758 gives an equivalent for the entire destination register,
5759 not just for the subreg being stored in now.
5760 This is a more interesting equivalence, so we arrange later
5761 to treat the entire reg as the destination. */
5762 sets[i].src_elt = src_eqv_elt;
5763 sets[i].src_hash = src_eqv_hash;
5765 else
5767 /* Insert source and constant equivalent into hash table, if not
5768 already present. */
5769 struct table_elt *classp = src_eqv_elt;
5770 rtx src = sets[i].src;
5771 rtx dest = SET_DEST (sets[i].rtl);
5772 enum machine_mode mode
5773 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5775 /* It's possible that we have a source value known to be
5776 constant but don't have a REG_EQUAL note on the insn.
5777 Lack of a note will mean src_eqv_elt will be NULL. This
5778 can happen where we've generated a SUBREG to access a
5779 CONST_INT that is already in a register in a wider mode.
5780 Ensure that the source expression is put in the proper
5781 constant class. */
5782 if (!classp)
5783 classp = sets[i].src_const_elt;
5785 if (sets[i].src_elt == 0)
5787 /* Don't put a hard register source into the table if this is
5788 the last insn of a libcall. In this case, we only need
5789 to put src_eqv_elt in src_elt. */
5790 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5792 struct table_elt *elt;
5794 /* Note that these insert_regs calls cannot remove
5795 any of the src_elt's, because they would have failed to
5796 match if not still valid. */
5797 if (insert_regs (src, classp, 0))
5799 rehash_using_reg (src);
5800 sets[i].src_hash = HASH (src, mode);
5802 elt = insert (src, classp, sets[i].src_hash, mode);
5803 elt->in_memory = sets[i].src_in_memory;
5804 sets[i].src_elt = classp = elt;
5806 else
5807 sets[i].src_elt = classp;
5809 if (sets[i].src_const && sets[i].src_const_elt == 0
5810 && src != sets[i].src_const
5811 && ! rtx_equal_p (sets[i].src_const, src))
5812 sets[i].src_elt = insert (sets[i].src_const, classp,
5813 sets[i].src_const_hash, mode);
5816 else if (sets[i].src_elt == 0)
5817 /* If we did not insert the source into the hash table (e.g., it was
5818 volatile), note the equivalence class for the REG_EQUAL value, if any,
5819 so that the destination goes into that class. */
5820 sets[i].src_elt = src_eqv_elt;
5822 invalidate_from_clobbers (x);
5824 /* Some registers are invalidated by subroutine calls. Memory is
5825 invalidated by non-constant calls. */
5827 if (CALL_P (insn))
5829 if (! CONST_OR_PURE_CALL_P (insn))
5830 invalidate_memory ();
5831 invalidate_for_call ();
5834 /* Now invalidate everything set by this instruction.
5835 If a SUBREG or other funny destination is being set,
5836 sets[i].rtl is still nonzero, so here we invalidate the reg
5837 a part of which is being set. */
5839 for (i = 0; i < n_sets; i++)
5840 if (sets[i].rtl)
5842 /* We can't use the inner dest, because the mode associated with
5843 a ZERO_EXTRACT is significant. */
5844 rtx dest = SET_DEST (sets[i].rtl);
5846 /* Needed for registers to remove the register from its
5847 previous quantity's chain.
5848 Needed for memory if this is a nonvarying address, unless
5849 we have just done an invalidate_memory that covers even those. */
5850 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5851 invalidate (dest, VOIDmode);
5852 else if (MEM_P (dest))
5854 /* Outgoing arguments for a libcall don't
5855 affect any recorded expressions. */
5856 if (! libcall_insn || insn == libcall_insn)
5857 invalidate (dest, VOIDmode);
5859 else if (GET_CODE (dest) == STRICT_LOW_PART
5860 || GET_CODE (dest) == ZERO_EXTRACT)
5861 invalidate (XEXP (dest, 0), GET_MODE (dest));
5864 /* A volatile ASM invalidates everything. */
5865 if (NONJUMP_INSN_P (insn)
5866 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5867 && MEM_VOLATILE_P (PATTERN (insn)))
5868 flush_hash_table ();
5870 /* Make sure registers mentioned in destinations
5871 are safe for use in an expression to be inserted.
5872 This removes from the hash table
5873 any invalid entry that refers to one of these registers.
5875 We don't care about the return value from mention_regs because
5876 we are going to hash the SET_DEST values unconditionally. */
5878 for (i = 0; i < n_sets; i++)
5880 if (sets[i].rtl)
5882 rtx x = SET_DEST (sets[i].rtl);
5884 if (!REG_P (x))
5885 mention_regs (x);
5886 else
5888 /* We used to rely on all references to a register becoming
5889 inaccessible when a register changes to a new quantity,
5890 since that changes the hash code. However, that is not
5891 safe, since after HASH_SIZE new quantities we get a
5892 hash 'collision' of a register with its own invalid
5893 entries. And since SUBREGs have been changed not to
5894 change their hash code with the hash code of the register,
5895 it wouldn't work any longer at all. So we have to check
5896 for any invalid references lying around now.
5897 This code is similar to the REG case in mention_regs,
5898 but it knows that reg_tick has been incremented, and
5899 it leaves reg_in_table as -1 . */
5900 unsigned int regno = REGNO (x);
5901 unsigned int endregno
5902 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
5903 : hard_regno_nregs[regno][GET_MODE (x)]);
5904 unsigned int i;
5906 for (i = regno; i < endregno; i++)
5908 if (REG_IN_TABLE (i) >= 0)
5910 remove_invalid_refs (i);
5911 REG_IN_TABLE (i) = -1;
5918 /* We may have just removed some of the src_elt's from the hash table.
5919 So replace each one with the current head of the same class. */
5921 for (i = 0; i < n_sets; i++)
5922 if (sets[i].rtl)
5924 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5925 /* If elt was removed, find current head of same class,
5926 or 0 if nothing remains of that class. */
5928 struct table_elt *elt = sets[i].src_elt;
5930 while (elt && elt->prev_same_value)
5931 elt = elt->prev_same_value;
5933 while (elt && elt->first_same_value == 0)
5934 elt = elt->next_same_value;
5935 sets[i].src_elt = elt ? elt->first_same_value : 0;
5939 /* Now insert the destinations into their equivalence classes. */
5941 for (i = 0; i < n_sets; i++)
5942 if (sets[i].rtl)
5944 rtx dest = SET_DEST (sets[i].rtl);
5945 struct table_elt *elt;
5947 /* Don't record value if we are not supposed to risk allocating
5948 floating-point values in registers that might be wider than
5949 memory. */
5950 if ((flag_float_store
5951 && MEM_P (dest)
5952 && FLOAT_MODE_P (GET_MODE (dest)))
5953 /* Don't record BLKmode values, because we don't know the
5954 size of it, and can't be sure that other BLKmode values
5955 have the same or smaller size. */
5956 || GET_MODE (dest) == BLKmode
5957 /* Don't record values of destinations set inside a libcall block
5958 since we might delete the libcall. Things should have been set
5959 up so we won't want to reuse such a value, but we play it safe
5960 here. */
5961 || libcall_insn
5962 /* If we didn't put a REG_EQUAL value or a source into the hash
5963 table, there is no point is recording DEST. */
5964 || sets[i].src_elt == 0
5965 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5966 or SIGN_EXTEND, don't record DEST since it can cause
5967 some tracking to be wrong.
5969 ??? Think about this more later. */
5970 || (GET_CODE (dest) == SUBREG
5971 && (GET_MODE_SIZE (GET_MODE (dest))
5972 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5973 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5974 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5975 continue;
5977 /* STRICT_LOW_PART isn't part of the value BEING set,
5978 and neither is the SUBREG inside it.
5979 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5980 if (GET_CODE (dest) == STRICT_LOW_PART)
5981 dest = SUBREG_REG (XEXP (dest, 0));
5983 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5984 /* Registers must also be inserted into chains for quantities. */
5985 if (insert_regs (dest, sets[i].src_elt, 1))
5987 /* If `insert_regs' changes something, the hash code must be
5988 recalculated. */
5989 rehash_using_reg (dest);
5990 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5993 elt = insert (dest, sets[i].src_elt,
5994 sets[i].dest_hash, GET_MODE (dest));
5996 elt->in_memory = (MEM_P (sets[i].inner_dest)
5997 && !MEM_READONLY_P (sets[i].inner_dest));
5999 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6000 narrower than M2, and both M1 and M2 are the same number of words,
6001 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6002 make that equivalence as well.
6004 However, BAR may have equivalences for which gen_lowpart
6005 will produce a simpler value than gen_lowpart applied to
6006 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6007 BAR's equivalences. If we don't get a simplified form, make
6008 the SUBREG. It will not be used in an equivalence, but will
6009 cause two similar assignments to be detected.
6011 Note the loop below will find SUBREG_REG (DEST) since we have
6012 already entered SRC and DEST of the SET in the table. */
6014 if (GET_CODE (dest) == SUBREG
6015 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6016 / UNITS_PER_WORD)
6017 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6018 && (GET_MODE_SIZE (GET_MODE (dest))
6019 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6020 && sets[i].src_elt != 0)
6022 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6023 struct table_elt *elt, *classp = 0;
6025 for (elt = sets[i].src_elt->first_same_value; elt;
6026 elt = elt->next_same_value)
6028 rtx new_src = 0;
6029 unsigned src_hash;
6030 struct table_elt *src_elt;
6031 int byte = 0;
6033 /* Ignore invalid entries. */
6034 if (!REG_P (elt->exp)
6035 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6036 continue;
6038 /* We may have already been playing subreg games. If the
6039 mode is already correct for the destination, use it. */
6040 if (GET_MODE (elt->exp) == new_mode)
6041 new_src = elt->exp;
6042 else
6044 /* Calculate big endian correction for the SUBREG_BYTE.
6045 We have already checked that M1 (GET_MODE (dest))
6046 is not narrower than M2 (new_mode). */
6047 if (BYTES_BIG_ENDIAN)
6048 byte = (GET_MODE_SIZE (GET_MODE (dest))
6049 - GET_MODE_SIZE (new_mode));
6051 new_src = simplify_gen_subreg (new_mode, elt->exp,
6052 GET_MODE (dest), byte);
6055 /* The call to simplify_gen_subreg fails if the value
6056 is VOIDmode, yet we can't do any simplification, e.g.
6057 for EXPR_LISTs denoting function call results.
6058 It is invalid to construct a SUBREG with a VOIDmode
6059 SUBREG_REG, hence a zero new_src means we can't do
6060 this substitution. */
6061 if (! new_src)
6062 continue;
6064 src_hash = HASH (new_src, new_mode);
6065 src_elt = lookup (new_src, src_hash, new_mode);
6067 /* Put the new source in the hash table is if isn't
6068 already. */
6069 if (src_elt == 0)
6071 if (insert_regs (new_src, classp, 0))
6073 rehash_using_reg (new_src);
6074 src_hash = HASH (new_src, new_mode);
6076 src_elt = insert (new_src, classp, src_hash, new_mode);
6077 src_elt->in_memory = elt->in_memory;
6079 else if (classp && classp != src_elt->first_same_value)
6080 /* Show that two things that we've seen before are
6081 actually the same. */
6082 merge_equiv_classes (src_elt, classp);
6084 classp = src_elt->first_same_value;
6085 /* Ignore invalid entries. */
6086 while (classp
6087 && !REG_P (classp->exp)
6088 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6089 classp = classp->next_same_value;
6094 /* Special handling for (set REG0 REG1) where REG0 is the
6095 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6096 be used in the sequel, so (if easily done) change this insn to
6097 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6098 that computed their value. Then REG1 will become a dead store
6099 and won't cloud the situation for later optimizations.
6101 Do not make this change if REG1 is a hard register, because it will
6102 then be used in the sequel and we may be changing a two-operand insn
6103 into a three-operand insn.
6105 Also do not do this if we are operating on a copy of INSN.
6107 Also don't do this if INSN ends a libcall; this would cause an unrelated
6108 register to be set in the middle of a libcall, and we then get bad code
6109 if the libcall is deleted. */
6111 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6112 && NEXT_INSN (PREV_INSN (insn)) == insn
6113 && REG_P (SET_SRC (sets[0].rtl))
6114 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6115 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6117 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6118 struct qty_table_elem *src_ent = &qty_table[src_q];
6120 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6121 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6123 rtx prev = insn;
6124 /* Scan for the previous nonnote insn, but stop at a basic
6125 block boundary. */
6128 prev = PREV_INSN (prev);
6130 while (prev && NOTE_P (prev)
6131 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6133 /* Do not swap the registers around if the previous instruction
6134 attaches a REG_EQUIV note to REG1.
6136 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6137 from the pseudo that originally shadowed an incoming argument
6138 to another register. Some uses of REG_EQUIV might rely on it
6139 being attached to REG1 rather than REG2.
6141 This section previously turned the REG_EQUIV into a REG_EQUAL
6142 note. We cannot do that because REG_EQUIV may provide an
6143 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6145 if (prev != 0 && NONJUMP_INSN_P (prev)
6146 && GET_CODE (PATTERN (prev)) == SET
6147 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6148 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6150 rtx dest = SET_DEST (sets[0].rtl);
6151 rtx src = SET_SRC (sets[0].rtl);
6152 rtx note;
6154 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6155 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6156 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6157 apply_change_group ();
6159 /* If INSN has a REG_EQUAL note, and this note mentions
6160 REG0, then we must delete it, because the value in
6161 REG0 has changed. If the note's value is REG1, we must
6162 also delete it because that is now this insn's dest. */
6163 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6164 if (note != 0
6165 && (reg_mentioned_p (dest, XEXP (note, 0))
6166 || rtx_equal_p (src, XEXP (note, 0))))
6167 remove_note (insn, note);
6172 /* If this is a conditional jump insn, record any known equivalences due to
6173 the condition being tested. */
6175 last_jump_equiv_class = 0;
6176 if (JUMP_P (insn)
6177 && n_sets == 1 && GET_CODE (x) == SET
6178 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6179 record_jump_equiv (insn, 0);
6181 #ifdef HAVE_cc0
6182 /* If the previous insn set CC0 and this insn no longer references CC0,
6183 delete the previous insn. Here we use the fact that nothing expects CC0
6184 to be valid over an insn, which is true until the final pass. */
6185 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6186 && (tem = single_set (prev_insn)) != 0
6187 && SET_DEST (tem) == cc0_rtx
6188 && ! reg_mentioned_p (cc0_rtx, x))
6189 delete_insn (prev_insn);
6191 prev_insn_cc0 = this_insn_cc0;
6192 prev_insn_cc0_mode = this_insn_cc0_mode;
6193 prev_insn = insn;
6194 #endif
6197 /* Remove from the hash table all expressions that reference memory. */
6199 static void
6200 invalidate_memory (void)
6202 int i;
6203 struct table_elt *p, *next;
6205 for (i = 0; i < HASH_SIZE; i++)
6206 for (p = table[i]; p; p = next)
6208 next = p->next_same_hash;
6209 if (p->in_memory)
6210 remove_from_table (p, i);
6214 /* If ADDR is an address that implicitly affects the stack pointer, return
6215 1 and update the register tables to show the effect. Else, return 0. */
6217 static int
6218 addr_affects_sp_p (rtx addr)
6220 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6221 && REG_P (XEXP (addr, 0))
6222 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6224 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6226 REG_TICK (STACK_POINTER_REGNUM)++;
6227 /* Is it possible to use a subreg of SP? */
6228 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6231 /* This should be *very* rare. */
6232 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6233 invalidate (stack_pointer_rtx, VOIDmode);
6235 return 1;
6238 return 0;
6241 /* Perform invalidation on the basis of everything about an insn
6242 except for invalidating the actual places that are SET in it.
6243 This includes the places CLOBBERed, and anything that might
6244 alias with something that is SET or CLOBBERed.
6246 X is the pattern of the insn. */
6248 static void
6249 invalidate_from_clobbers (rtx x)
6251 if (GET_CODE (x) == CLOBBER)
6253 rtx ref = XEXP (x, 0);
6254 if (ref)
6256 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6257 || MEM_P (ref))
6258 invalidate (ref, VOIDmode);
6259 else if (GET_CODE (ref) == STRICT_LOW_PART
6260 || GET_CODE (ref) == ZERO_EXTRACT)
6261 invalidate (XEXP (ref, 0), GET_MODE (ref));
6264 else if (GET_CODE (x) == PARALLEL)
6266 int i;
6267 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6269 rtx y = XVECEXP (x, 0, i);
6270 if (GET_CODE (y) == CLOBBER)
6272 rtx ref = XEXP (y, 0);
6273 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6274 || MEM_P (ref))
6275 invalidate (ref, VOIDmode);
6276 else if (GET_CODE (ref) == STRICT_LOW_PART
6277 || GET_CODE (ref) == ZERO_EXTRACT)
6278 invalidate (XEXP (ref, 0), GET_MODE (ref));
6284 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6285 and replace any registers in them with either an equivalent constant
6286 or the canonical form of the register. If we are inside an address,
6287 only do this if the address remains valid.
6289 OBJECT is 0 except when within a MEM in which case it is the MEM.
6291 Return the replacement for X. */
6293 static rtx
6294 cse_process_notes (rtx x, rtx object)
6296 enum rtx_code code = GET_CODE (x);
6297 const char *fmt = GET_RTX_FORMAT (code);
6298 int i;
6300 switch (code)
6302 case CONST_INT:
6303 case CONST:
6304 case SYMBOL_REF:
6305 case LABEL_REF:
6306 case CONST_DOUBLE:
6307 case CONST_VECTOR:
6308 case PC:
6309 case CC0:
6310 case LO_SUM:
6311 return x;
6313 case MEM:
6314 validate_change (x, &XEXP (x, 0),
6315 cse_process_notes (XEXP (x, 0), x), 0);
6316 return x;
6318 case EXPR_LIST:
6319 case INSN_LIST:
6320 if (REG_NOTE_KIND (x) == REG_EQUAL)
6321 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6322 if (XEXP (x, 1))
6323 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6324 return x;
6326 case SIGN_EXTEND:
6327 case ZERO_EXTEND:
6328 case SUBREG:
6330 rtx new = cse_process_notes (XEXP (x, 0), object);
6331 /* We don't substitute VOIDmode constants into these rtx,
6332 since they would impede folding. */
6333 if (GET_MODE (new) != VOIDmode)
6334 validate_change (object, &XEXP (x, 0), new, 0);
6335 return x;
6338 case REG:
6339 i = REG_QTY (REGNO (x));
6341 /* Return a constant or a constant register. */
6342 if (REGNO_QTY_VALID_P (REGNO (x)))
6344 struct qty_table_elem *ent = &qty_table[i];
6346 if (ent->const_rtx != NULL_RTX
6347 && (CONSTANT_P (ent->const_rtx)
6348 || REG_P (ent->const_rtx)))
6350 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6351 if (new)
6352 return new;
6356 /* Otherwise, canonicalize this register. */
6357 return canon_reg (x, NULL_RTX);
6359 default:
6360 break;
6363 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6364 if (fmt[i] == 'e')
6365 validate_change (object, &XEXP (x, i),
6366 cse_process_notes (XEXP (x, i), object), 0);
6368 return x;
6371 /* Find common subexpressions between the end test of a loop and the beginning
6372 of the loop. LOOP_START is the CODE_LABEL at the start of a loop.
6374 Often we have a loop where an expression in the exit test is used
6375 in the body of the loop. For example "while (*p) *q++ = *p++;".
6376 Because of the way we duplicate the loop exit test in front of the loop,
6377 however, we don't detect that common subexpression. This will be caught
6378 when global cse is implemented, but this is a quite common case.
6380 This function handles the most common cases of these common expressions.
6381 It is called after we have processed the basic block ending with the
6382 NOTE_INSN_LOOP_END note that ends a loop and the previous JUMP_INSN
6383 jumps to a label used only once. */
6385 static void
6386 cse_around_loop (rtx loop_start)
6388 rtx insn;
6389 int i;
6390 struct table_elt *p;
6392 /* If the jump at the end of the loop doesn't go to the start, we don't
6393 do anything. */
6394 for (insn = PREV_INSN (loop_start);
6395 insn && (NOTE_P (insn) && NOTE_LINE_NUMBER (insn) >= 0);
6396 insn = PREV_INSN (insn))
6399 if (insn == 0
6400 || !NOTE_P (insn)
6401 || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG)
6402 return;
6404 /* If the last insn of the loop (the end test) was an NE comparison,
6405 we will interpret it as an EQ comparison, since we fell through
6406 the loop. Any equivalences resulting from that comparison are
6407 therefore not valid and must be invalidated. */
6408 if (last_jump_equiv_class)
6409 for (p = last_jump_equiv_class->first_same_value; p;
6410 p = p->next_same_value)
6412 if (MEM_P (p->exp) || REG_P (p->exp)
6413 || (GET_CODE (p->exp) == SUBREG
6414 && REG_P (SUBREG_REG (p->exp))))
6415 invalidate (p->exp, VOIDmode);
6416 else if (GET_CODE (p->exp) == STRICT_LOW_PART
6417 || GET_CODE (p->exp) == ZERO_EXTRACT)
6418 invalidate (XEXP (p->exp, 0), GET_MODE (p->exp));
6421 /* Process insns starting after LOOP_START until we hit a CALL_INSN or
6422 a CODE_LABEL (we could handle a CALL_INSN, but it isn't worth it).
6424 The only thing we do with SET_DEST is invalidate entries, so we
6425 can safely process each SET in order. It is slightly less efficient
6426 to do so, but we only want to handle the most common cases.
6428 The gen_move_insn call in cse_set_around_loop may create new pseudos.
6429 These pseudos won't have valid entries in any of the tables indexed
6430 by register number, such as reg_qty. We avoid out-of-range array
6431 accesses by not processing any instructions created after cse started. */
6433 for (insn = NEXT_INSN (loop_start);
6434 !CALL_P (insn) && !LABEL_P (insn)
6435 && INSN_UID (insn) < max_insn_uid
6436 && ! (NOTE_P (insn)
6437 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END);
6438 insn = NEXT_INSN (insn))
6440 if (INSN_P (insn)
6441 && (GET_CODE (PATTERN (insn)) == SET
6442 || GET_CODE (PATTERN (insn)) == CLOBBER))
6443 cse_set_around_loop (PATTERN (insn), insn, loop_start);
6444 else if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == PARALLEL)
6445 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6446 if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == SET
6447 || GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER)
6448 cse_set_around_loop (XVECEXP (PATTERN (insn), 0, i), insn,
6449 loop_start);
6453 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6454 since they are done elsewhere. This function is called via note_stores. */
6456 static void
6457 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6459 enum rtx_code code = GET_CODE (dest);
6461 if (code == MEM
6462 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6463 /* There are times when an address can appear varying and be a PLUS
6464 during this scan when it would be a fixed address were we to know
6465 the proper equivalences. So invalidate all memory if there is
6466 a BLKmode or nonscalar memory reference or a reference to a
6467 variable address. */
6468 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6469 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6471 invalidate_memory ();
6472 return;
6475 if (GET_CODE (set) == CLOBBER
6476 || CC0_P (dest)
6477 || dest == pc_rtx)
6478 return;
6480 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6481 invalidate (XEXP (dest, 0), GET_MODE (dest));
6482 else if (code == REG || code == SUBREG || code == MEM)
6483 invalidate (dest, VOIDmode);
6486 /* Invalidate all insns from START up to the end of the function or the
6487 next label. This called when we wish to CSE around a block that is
6488 conditionally executed. */
6490 static void
6491 invalidate_skipped_block (rtx start)
6493 rtx insn;
6495 for (insn = start; insn && !LABEL_P (insn);
6496 insn = NEXT_INSN (insn))
6498 if (! INSN_P (insn))
6499 continue;
6501 if (CALL_P (insn))
6503 if (! CONST_OR_PURE_CALL_P (insn))
6504 invalidate_memory ();
6505 invalidate_for_call ();
6508 invalidate_from_clobbers (PATTERN (insn));
6509 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6513 /* If modifying X will modify the value in *DATA (which is really an
6514 `rtx *'), indicate that fact by setting the pointed to value to
6515 NULL_RTX. */
6517 static void
6518 cse_check_loop_start (rtx x, rtx set ATTRIBUTE_UNUSED, void *data)
6520 rtx *cse_check_loop_start_value = (rtx *) data;
6522 if (*cse_check_loop_start_value == NULL_RTX
6523 || GET_CODE (x) == CC0 || GET_CODE (x) == PC)
6524 return;
6526 if ((MEM_P (x) && MEM_P (*cse_check_loop_start_value))
6527 || reg_overlap_mentioned_p (x, *cse_check_loop_start_value))
6528 *cse_check_loop_start_value = NULL_RTX;
6531 /* X is a SET or CLOBBER contained in INSN that was found near the start of
6532 a loop that starts with the label at LOOP_START.
6534 If X is a SET, we see if its SET_SRC is currently in our hash table.
6535 If so, we see if it has a value equal to some register used only in the
6536 loop exit code (as marked by jump.c).
6538 If those two conditions are true, we search backwards from the start of
6539 the loop to see if that same value was loaded into a register that still
6540 retains its value at the start of the loop.
6542 If so, we insert an insn after the load to copy the destination of that
6543 load into the equivalent register and (try to) replace our SET_SRC with that
6544 register.
6546 In any event, we invalidate whatever this SET or CLOBBER modifies. */
6548 static void
6549 cse_set_around_loop (rtx x, rtx insn, rtx loop_start)
6551 struct table_elt *src_elt;
6553 /* If this is a SET, see if we can replace SET_SRC, but ignore SETs that
6554 are setting PC or CC0 or whose SET_SRC is already a register. */
6555 if (GET_CODE (x) == SET
6556 && GET_CODE (SET_DEST (x)) != PC && GET_CODE (SET_DEST (x)) != CC0
6557 && !REG_P (SET_SRC (x)))
6559 src_elt = lookup (SET_SRC (x),
6560 HASH (SET_SRC (x), GET_MODE (SET_DEST (x))),
6561 GET_MODE (SET_DEST (x)));
6563 if (src_elt)
6564 for (src_elt = src_elt->first_same_value; src_elt;
6565 src_elt = src_elt->next_same_value)
6566 if (REG_P (src_elt->exp) && REG_LOOP_TEST_P (src_elt->exp)
6567 && COST (src_elt->exp) < COST (SET_SRC (x)))
6569 rtx p, set;
6571 /* Look for an insn in front of LOOP_START that sets
6572 something in the desired mode to SET_SRC (x) before we hit
6573 a label or CALL_INSN. */
6575 for (p = prev_nonnote_insn (loop_start);
6576 p && !CALL_P (p)
6577 && !LABEL_P (p);
6578 p = prev_nonnote_insn (p))
6579 if ((set = single_set (p)) != 0
6580 && REG_P (SET_DEST (set))
6581 && GET_MODE (SET_DEST (set)) == src_elt->mode
6582 && rtx_equal_p (SET_SRC (set), SET_SRC (x)))
6584 /* We now have to ensure that nothing between P
6585 and LOOP_START modified anything referenced in
6586 SET_SRC (x). We know that nothing within the loop
6587 can modify it, or we would have invalidated it in
6588 the hash table. */
6589 rtx q;
6590 rtx cse_check_loop_start_value = SET_SRC (x);
6591 for (q = p; q != loop_start; q = NEXT_INSN (q))
6592 if (INSN_P (q))
6593 note_stores (PATTERN (q),
6594 cse_check_loop_start,
6595 &cse_check_loop_start_value);
6597 /* If nothing was changed and we can replace our
6598 SET_SRC, add an insn after P to copy its destination
6599 to what we will be replacing SET_SRC with. */
6600 if (cse_check_loop_start_value
6601 && single_set (p)
6602 && !can_throw_internal (insn)
6603 && validate_change (insn, &SET_SRC (x),
6604 src_elt->exp, 0))
6606 /* If this creates new pseudos, this is unsafe,
6607 because the regno of new pseudo is unsuitable
6608 to index into reg_qty when cse_insn processes
6609 the new insn. Therefore, if a new pseudo was
6610 created, discard this optimization. */
6611 int nregs = max_reg_num ();
6612 rtx move
6613 = gen_move_insn (src_elt->exp, SET_DEST (set));
6614 if (nregs != max_reg_num ())
6616 if (! validate_change (insn, &SET_SRC (x),
6617 SET_SRC (set), 0))
6618 abort ();
6620 else
6622 if (CONSTANT_P (SET_SRC (set))
6623 && ! find_reg_equal_equiv_note (insn))
6624 set_unique_reg_note (insn, REG_EQUAL,
6625 SET_SRC (set));
6626 if (control_flow_insn_p (p))
6627 /* p can cause a control flow transfer so it
6628 is the last insn of a basic block. We can't
6629 therefore use emit_insn_after. */
6630 emit_insn_before (move, next_nonnote_insn (p));
6631 else
6632 emit_insn_after (move, p);
6635 break;
6640 /* Deal with the destination of X affecting the stack pointer. */
6641 addr_affects_sp_p (SET_DEST (x));
6643 /* See comment on similar code in cse_insn for explanation of these
6644 tests. */
6645 if (REG_P (SET_DEST (x)) || GET_CODE (SET_DEST (x)) == SUBREG
6646 || MEM_P (SET_DEST (x)))
6647 invalidate (SET_DEST (x), VOIDmode);
6648 else if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
6649 || GET_CODE (SET_DEST (x)) == ZERO_EXTRACT)
6650 invalidate (XEXP (SET_DEST (x), 0), GET_MODE (SET_DEST (x)));
6653 /* Find the end of INSN's basic block and return its range,
6654 the total number of SETs in all the insns of the block, the last insn of the
6655 block, and the branch path.
6657 The branch path indicates which branches should be followed. If a nonzero
6658 path size is specified, the block should be rescanned and a different set
6659 of branches will be taken. The branch path is only used if
6660 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6662 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6663 used to describe the block. It is filled in with the information about
6664 the current block. The incoming structure's branch path, if any, is used
6665 to construct the output branch path. */
6667 static void
6668 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6669 int follow_jumps, int after_loop, int skip_blocks)
6671 rtx p = insn, q;
6672 int nsets = 0;
6673 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6674 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6675 int path_size = data->path_size;
6676 int path_entry = 0;
6677 int i;
6679 /* Update the previous branch path, if any. If the last branch was
6680 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6681 If it was previously PATH_NOT_TAKEN,
6682 shorten the path by one and look at the previous branch. We know that
6683 at least one branch must have been taken if PATH_SIZE is nonzero. */
6684 while (path_size > 0)
6686 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6688 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6689 break;
6691 else
6692 path_size--;
6695 /* If the first instruction is marked with QImode, that means we've
6696 already processed this block. Our caller will look at DATA->LAST
6697 to figure out where to go next. We want to return the next block
6698 in the instruction stream, not some branched-to block somewhere
6699 else. We accomplish this by pretending our called forbid us to
6700 follow jumps, or skip blocks. */
6701 if (GET_MODE (insn) == QImode)
6702 follow_jumps = skip_blocks = 0;
6704 /* Scan to end of this basic block. */
6705 while (p && !LABEL_P (p))
6707 /* Don't cse out the end of a loop. This makes a difference
6708 only for the unusual loops that always execute at least once;
6709 all other loops have labels there so we will stop in any case.
6710 Cse'ing out the end of the loop is dangerous because it
6711 might cause an invariant expression inside the loop
6712 to be reused after the end of the loop. This would make it
6713 hard to move the expression out of the loop in loop.c,
6714 especially if it is one of several equivalent expressions
6715 and loop.c would like to eliminate it.
6717 If we are running after loop.c has finished, we can ignore
6718 the NOTE_INSN_LOOP_END. */
6720 if (! after_loop && NOTE_P (p)
6721 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
6722 break;
6724 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6725 the regs restored by the longjmp come from
6726 a later time than the setjmp. */
6727 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6728 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6729 break;
6731 /* A PARALLEL can have lots of SETs in it,
6732 especially if it is really an ASM_OPERANDS. */
6733 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6734 nsets += XVECLEN (PATTERN (p), 0);
6735 else if (!NOTE_P (p))
6736 nsets += 1;
6738 /* Ignore insns made by CSE; they cannot affect the boundaries of
6739 the basic block. */
6741 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6742 high_cuid = INSN_CUID (p);
6743 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6744 low_cuid = INSN_CUID (p);
6746 /* See if this insn is in our branch path. If it is and we are to
6747 take it, do so. */
6748 if (path_entry < path_size && data->path[path_entry].branch == p)
6750 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6751 p = JUMP_LABEL (p);
6753 /* Point to next entry in path, if any. */
6754 path_entry++;
6757 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6758 was specified, we haven't reached our maximum path length, there are
6759 insns following the target of the jump, this is the only use of the
6760 jump label, and the target label is preceded by a BARRIER.
6762 Alternatively, we can follow the jump if it branches around a
6763 block of code and there are no other branches into the block.
6764 In this case invalidate_skipped_block will be called to invalidate any
6765 registers set in the block when following the jump. */
6767 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6768 && JUMP_P (p)
6769 && GET_CODE (PATTERN (p)) == SET
6770 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6771 && JUMP_LABEL (p) != 0
6772 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6773 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6775 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6776 if ((!NOTE_P (q)
6777 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6778 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6779 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6780 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6781 break;
6783 /* If we ran into a BARRIER, this code is an extension of the
6784 basic block when the branch is taken. */
6785 if (follow_jumps && q != 0 && BARRIER_P (q))
6787 /* Don't allow ourself to keep walking around an
6788 always-executed loop. */
6789 if (next_real_insn (q) == next)
6791 p = NEXT_INSN (p);
6792 continue;
6795 /* Similarly, don't put a branch in our path more than once. */
6796 for (i = 0; i < path_entry; i++)
6797 if (data->path[i].branch == p)
6798 break;
6800 if (i != path_entry)
6801 break;
6803 data->path[path_entry].branch = p;
6804 data->path[path_entry++].status = PATH_TAKEN;
6806 /* This branch now ends our path. It was possible that we
6807 didn't see this branch the last time around (when the
6808 insn in front of the target was a JUMP_INSN that was
6809 turned into a no-op). */
6810 path_size = path_entry;
6812 p = JUMP_LABEL (p);
6813 /* Mark block so we won't scan it again later. */
6814 PUT_MODE (NEXT_INSN (p), QImode);
6816 /* Detect a branch around a block of code. */
6817 else if (skip_blocks && q != 0 && !LABEL_P (q))
6819 rtx tmp;
6821 if (next_real_insn (q) == next)
6823 p = NEXT_INSN (p);
6824 continue;
6827 for (i = 0; i < path_entry; i++)
6828 if (data->path[i].branch == p)
6829 break;
6831 if (i != path_entry)
6832 break;
6834 /* This is no_labels_between_p (p, q) with an added check for
6835 reaching the end of a function (in case Q precedes P). */
6836 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6837 if (LABEL_P (tmp))
6838 break;
6840 if (tmp == q)
6842 data->path[path_entry].branch = p;
6843 data->path[path_entry++].status = PATH_AROUND;
6845 path_size = path_entry;
6847 p = JUMP_LABEL (p);
6848 /* Mark block so we won't scan it again later. */
6849 PUT_MODE (NEXT_INSN (p), QImode);
6853 p = NEXT_INSN (p);
6856 data->low_cuid = low_cuid;
6857 data->high_cuid = high_cuid;
6858 data->nsets = nsets;
6859 data->last = p;
6861 /* If all jumps in the path are not taken, set our path length to zero
6862 so a rescan won't be done. */
6863 for (i = path_size - 1; i >= 0; i--)
6864 if (data->path[i].status != PATH_NOT_TAKEN)
6865 break;
6867 if (i == -1)
6868 data->path_size = 0;
6869 else
6870 data->path_size = path_size;
6872 /* End the current branch path. */
6873 data->path[path_size].branch = 0;
6876 /* Perform cse on the instructions of a function.
6877 F is the first instruction.
6878 NREGS is one plus the highest pseudo-reg number used in the instruction.
6880 AFTER_LOOP is 1 if this is the cse call done after loop optimization
6881 (only if -frerun-cse-after-loop).
6883 Returns 1 if jump_optimize should be redone due to simplifications
6884 in conditional jump instructions. */
6887 cse_main (rtx f, int nregs, int after_loop, FILE *file)
6889 struct cse_basic_block_data val;
6890 rtx insn = f;
6891 int i;
6893 val.path = xmalloc (sizeof (struct branch_path)
6894 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6896 cse_jumps_altered = 0;
6897 recorded_label_ref = 0;
6898 constant_pool_entries_cost = 0;
6899 constant_pool_entries_regcost = 0;
6900 val.path_size = 0;
6901 rtl_hooks = cse_rtl_hooks;
6903 init_recog ();
6904 init_alias_analysis ();
6906 max_reg = nregs;
6908 max_insn_uid = get_max_uid ();
6910 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6912 #ifdef LOAD_EXTEND_OP
6914 /* Allocate scratch rtl here. cse_insn will fill in the memory reference
6915 and change the code and mode as appropriate. */
6916 memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX);
6917 #endif
6919 /* Reset the counter indicating how many elements have been made
6920 thus far. */
6921 n_elements_made = 0;
6923 /* Find the largest uid. */
6925 max_uid = get_max_uid ();
6926 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6928 /* Compute the mapping from uids to cuids.
6929 CUIDs are numbers assigned to insns, like uids,
6930 except that cuids increase monotonically through the code.
6931 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6932 between two insns is not affected by -g. */
6934 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6936 if (!NOTE_P (insn)
6937 || NOTE_LINE_NUMBER (insn) < 0)
6938 INSN_CUID (insn) = ++i;
6939 else
6940 /* Give a line number note the same cuid as preceding insn. */
6941 INSN_CUID (insn) = i;
6944 ggc_push_context ();
6946 /* Loop over basic blocks.
6947 Compute the maximum number of qty's needed for each basic block
6948 (which is 2 for each SET). */
6949 insn = f;
6950 while (insn)
6952 cse_altered = 0;
6953 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps, after_loop,
6954 flag_cse_skip_blocks);
6956 /* If this basic block was already processed or has no sets, skip it. */
6957 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6959 PUT_MODE (insn, VOIDmode);
6960 insn = (val.last ? NEXT_INSN (val.last) : 0);
6961 val.path_size = 0;
6962 continue;
6965 cse_basic_block_start = val.low_cuid;
6966 cse_basic_block_end = val.high_cuid;
6967 max_qty = val.nsets * 2;
6969 if (file)
6970 fnotice (file, ";; Processing block from %d to %d, %d sets.\n",
6971 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6972 val.nsets);
6974 /* Make MAX_QTY bigger to give us room to optimize
6975 past the end of this basic block, if that should prove useful. */
6976 if (max_qty < 500)
6977 max_qty = 500;
6979 max_qty += max_reg;
6981 /* If this basic block is being extended by following certain jumps,
6982 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6983 Otherwise, we start after this basic block. */
6984 if (val.path_size > 0)
6985 cse_basic_block (insn, val.last, val.path, 0);
6986 else
6988 int old_cse_jumps_altered = cse_jumps_altered;
6989 rtx temp;
6991 /* When cse changes a conditional jump to an unconditional
6992 jump, we want to reprocess the block, since it will give
6993 us a new branch path to investigate. */
6994 cse_jumps_altered = 0;
6995 temp = cse_basic_block (insn, val.last, val.path, ! after_loop);
6996 if (cse_jumps_altered == 0
6997 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6998 insn = temp;
7000 cse_jumps_altered |= old_cse_jumps_altered;
7003 if (cse_altered)
7004 ggc_collect ();
7006 #ifdef USE_C_ALLOCA
7007 alloca (0);
7008 #endif
7011 ggc_pop_context ();
7013 if (max_elements_made < n_elements_made)
7014 max_elements_made = n_elements_made;
7016 /* Clean up. */
7017 end_alias_analysis ();
7018 free (uid_cuid);
7019 free (reg_eqv_table);
7020 free (val.path);
7021 rtl_hooks = general_rtl_hooks;
7023 return cse_jumps_altered || recorded_label_ref;
7026 /* Process a single basic block. FROM and TO and the limits of the basic
7027 block. NEXT_BRANCH points to the branch path when following jumps or
7028 a null path when not following jumps.
7030 AROUND_LOOP is nonzero if we are to try to cse around to the start of a
7031 loop. This is true when we are being called for the last time on a
7032 block and this CSE pass is before loop.c. */
7034 static rtx
7035 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch,
7036 int around_loop)
7038 rtx insn;
7039 int to_usage = 0;
7040 rtx libcall_insn = NULL_RTX;
7041 int num_insns = 0;
7042 int no_conflict = 0;
7044 /* This array is undefined before max_reg, so only allocate
7045 the space actually needed and adjust the start. */
7047 qty_table = xmalloc ((max_qty - max_reg) * sizeof (struct qty_table_elem));
7048 qty_table -= max_reg;
7050 new_basic_block ();
7052 /* TO might be a label. If so, protect it from being deleted. */
7053 if (to != 0 && LABEL_P (to))
7054 ++LABEL_NUSES (to);
7056 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7058 enum rtx_code code = GET_CODE (insn);
7060 /* If we have processed 1,000 insns, flush the hash table to
7061 avoid extreme quadratic behavior. We must not include NOTEs
7062 in the count since there may be more of them when generating
7063 debugging information. If we clear the table at different
7064 times, code generated with -g -O might be different than code
7065 generated with -O but not -g.
7067 ??? This is a real kludge and needs to be done some other way.
7068 Perhaps for 2.9. */
7069 if (code != NOTE && num_insns++ > 1000)
7071 flush_hash_table ();
7072 num_insns = 0;
7075 /* See if this is a branch that is part of the path. If so, and it is
7076 to be taken, do so. */
7077 if (next_branch->branch == insn)
7079 enum taken status = next_branch++->status;
7080 if (status != PATH_NOT_TAKEN)
7082 if (status == PATH_TAKEN)
7083 record_jump_equiv (insn, 1);
7084 else
7085 invalidate_skipped_block (NEXT_INSN (insn));
7087 /* Set the last insn as the jump insn; it doesn't affect cc0.
7088 Then follow this branch. */
7089 #ifdef HAVE_cc0
7090 prev_insn_cc0 = 0;
7091 prev_insn = insn;
7092 #endif
7093 insn = JUMP_LABEL (insn);
7094 continue;
7098 if (GET_MODE (insn) == QImode)
7099 PUT_MODE (insn, VOIDmode);
7101 if (GET_RTX_CLASS (code) == RTX_INSN)
7103 rtx p;
7105 /* Process notes first so we have all notes in canonical forms when
7106 looking for duplicate operations. */
7108 if (REG_NOTES (insn))
7109 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7111 /* Track when we are inside in LIBCALL block. Inside such a block,
7112 we do not want to record destinations. The last insn of a
7113 LIBCALL block is not considered to be part of the block, since
7114 its destination is the result of the block and hence should be
7115 recorded. */
7117 if (REG_NOTES (insn) != 0)
7119 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7120 libcall_insn = XEXP (p, 0);
7121 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7123 /* Keep libcall_insn for the last SET insn of a no-conflict
7124 block to prevent changing the destination. */
7125 if (! no_conflict)
7126 libcall_insn = 0;
7127 else
7128 no_conflict = -1;
7130 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
7131 no_conflict = 1;
7134 cse_insn (insn, libcall_insn);
7136 if (no_conflict == -1)
7138 libcall_insn = 0;
7139 no_conflict = 0;
7142 /* If we haven't already found an insn where we added a LABEL_REF,
7143 check this one. */
7144 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
7145 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
7146 (void *) insn))
7147 recorded_label_ref = 1;
7150 /* If INSN is now an unconditional jump, skip to the end of our
7151 basic block by pretending that we just did the last insn in the
7152 basic block. If we are jumping to the end of our block, show
7153 that we can have one usage of TO. */
7155 if (any_uncondjump_p (insn))
7157 if (to == 0)
7159 free (qty_table + max_reg);
7160 return 0;
7163 if (JUMP_LABEL (insn) == to)
7164 to_usage = 1;
7166 /* Maybe TO was deleted because the jump is unconditional.
7167 If so, there is nothing left in this basic block. */
7168 /* ??? Perhaps it would be smarter to set TO
7169 to whatever follows this insn,
7170 and pretend the basic block had always ended here. */
7171 if (INSN_DELETED_P (to))
7172 break;
7174 insn = PREV_INSN (to);
7177 /* See if it is ok to keep on going past the label
7178 which used to end our basic block. Remember that we incremented
7179 the count of that label, so we decrement it here. If we made
7180 a jump unconditional, TO_USAGE will be one; in that case, we don't
7181 want to count the use in that jump. */
7183 if (to != 0 && NEXT_INSN (insn) == to
7184 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7186 struct cse_basic_block_data val;
7187 rtx prev;
7189 insn = NEXT_INSN (to);
7191 /* If TO was the last insn in the function, we are done. */
7192 if (insn == 0)
7194 free (qty_table + max_reg);
7195 return 0;
7198 /* If TO was preceded by a BARRIER we are done with this block
7199 because it has no continuation. */
7200 prev = prev_nonnote_insn (to);
7201 if (prev && BARRIER_P (prev))
7203 free (qty_table + max_reg);
7204 return insn;
7207 /* Find the end of the following block. Note that we won't be
7208 following branches in this case. */
7209 to_usage = 0;
7210 val.path_size = 0;
7211 val.path = xmalloc (sizeof (struct branch_path)
7212 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7213 cse_end_of_basic_block (insn, &val, 0, 0, 0);
7214 free (val.path);
7216 /* If the tables we allocated have enough space left
7217 to handle all the SETs in the next basic block,
7218 continue through it. Otherwise, return,
7219 and that block will be scanned individually. */
7220 if (val.nsets * 2 + next_qty > max_qty)
7221 break;
7223 cse_basic_block_start = val.low_cuid;
7224 cse_basic_block_end = val.high_cuid;
7225 to = val.last;
7227 /* Prevent TO from being deleted if it is a label. */
7228 if (to != 0 && LABEL_P (to))
7229 ++LABEL_NUSES (to);
7231 /* Back up so we process the first insn in the extension. */
7232 insn = PREV_INSN (insn);
7236 if (next_qty > max_qty)
7237 abort ();
7239 /* If we are running before loop.c, we stopped on a NOTE_INSN_LOOP_END, and
7240 the previous insn is the only insn that branches to the head of a loop,
7241 we can cse into the loop. Don't do this if we changed the jump
7242 structure of a loop unless we aren't going to be following jumps. */
7244 insn = prev_nonnote_insn (to);
7245 if ((cse_jumps_altered == 0
7246 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7247 && around_loop && to != 0
7248 && NOTE_P (to) && NOTE_LINE_NUMBER (to) == NOTE_INSN_LOOP_END
7249 && JUMP_P (insn)
7250 && JUMP_LABEL (insn) != 0
7251 && LABEL_NUSES (JUMP_LABEL (insn)) == 1)
7252 cse_around_loop (JUMP_LABEL (insn));
7254 free (qty_table + max_reg);
7256 return to ? NEXT_INSN (to) : 0;
7259 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7260 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7262 static int
7263 check_for_label_ref (rtx *rtl, void *data)
7265 rtx insn = (rtx) data;
7267 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7268 we must rerun jump since it needs to place the note. If this is a
7269 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7270 since no REG_LABEL will be added. */
7271 return (GET_CODE (*rtl) == LABEL_REF
7272 && ! LABEL_REF_NONLOCAL_P (*rtl)
7273 && LABEL_P (XEXP (*rtl, 0))
7274 && INSN_UID (XEXP (*rtl, 0)) != 0
7275 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7278 /* Count the number of times registers are used (not set) in X.
7279 COUNTS is an array in which we accumulate the count, INCR is how much
7280 we count each register usage. */
7282 static void
7283 count_reg_usage (rtx x, int *counts, int incr)
7285 enum rtx_code code;
7286 rtx note;
7287 const char *fmt;
7288 int i, j;
7290 if (x == 0)
7291 return;
7293 switch (code = GET_CODE (x))
7295 case REG:
7296 counts[REGNO (x)] += incr;
7297 return;
7299 case PC:
7300 case CC0:
7301 case CONST:
7302 case CONST_INT:
7303 case CONST_DOUBLE:
7304 case CONST_VECTOR:
7305 case SYMBOL_REF:
7306 case LABEL_REF:
7307 return;
7309 case CLOBBER:
7310 /* If we are clobbering a MEM, mark any registers inside the address
7311 as being used. */
7312 if (MEM_P (XEXP (x, 0)))
7313 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, incr);
7314 return;
7316 case SET:
7317 /* Unless we are setting a REG, count everything in SET_DEST. */
7318 if (!REG_P (SET_DEST (x)))
7319 count_reg_usage (SET_DEST (x), counts, incr);
7320 count_reg_usage (SET_SRC (x), counts, incr);
7321 return;
7323 case CALL_INSN:
7324 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, incr);
7325 /* Fall through. */
7327 case INSN:
7328 case JUMP_INSN:
7329 count_reg_usage (PATTERN (x), counts, incr);
7331 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7332 use them. */
7334 note = find_reg_equal_equiv_note (x);
7335 if (note)
7337 rtx eqv = XEXP (note, 0);
7339 if (GET_CODE (eqv) == EXPR_LIST)
7340 /* This REG_EQUAL note describes the result of a function call.
7341 Process all the arguments. */
7344 count_reg_usage (XEXP (eqv, 0), counts, incr);
7345 eqv = XEXP (eqv, 1);
7347 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7348 else
7349 count_reg_usage (eqv, counts, incr);
7351 return;
7353 case EXPR_LIST:
7354 if (REG_NOTE_KIND (x) == REG_EQUAL
7355 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7356 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7357 involving registers in the address. */
7358 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7359 count_reg_usage (XEXP (x, 0), counts, incr);
7361 count_reg_usage (XEXP (x, 1), counts, incr);
7362 return;
7364 case ASM_OPERANDS:
7365 /* Iterate over just the inputs, not the constraints as well. */
7366 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7367 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, incr);
7368 return;
7370 case INSN_LIST:
7371 abort ();
7373 default:
7374 break;
7377 fmt = GET_RTX_FORMAT (code);
7378 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7380 if (fmt[i] == 'e')
7381 count_reg_usage (XEXP (x, i), counts, incr);
7382 else if (fmt[i] == 'E')
7383 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7384 count_reg_usage (XVECEXP (x, i, j), counts, incr);
7388 /* Return true if set is live. */
7389 static bool
7390 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7391 int *counts)
7393 #ifdef HAVE_cc0
7394 rtx tem;
7395 #endif
7397 if (set_noop_p (set))
7400 #ifdef HAVE_cc0
7401 else if (GET_CODE (SET_DEST (set)) == CC0
7402 && !side_effects_p (SET_SRC (set))
7403 && ((tem = next_nonnote_insn (insn)) == 0
7404 || !INSN_P (tem)
7405 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7406 return false;
7407 #endif
7408 else if (!REG_P (SET_DEST (set))
7409 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7410 || counts[REGNO (SET_DEST (set))] != 0
7411 || side_effects_p (SET_SRC (set)))
7412 return true;
7413 return false;
7416 /* Return true if insn is live. */
7418 static bool
7419 insn_live_p (rtx insn, int *counts)
7421 int i;
7422 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7423 return true;
7424 else if (GET_CODE (PATTERN (insn)) == SET)
7425 return set_live_p (PATTERN (insn), insn, counts);
7426 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7428 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7430 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7432 if (GET_CODE (elt) == SET)
7434 if (set_live_p (elt, insn, counts))
7435 return true;
7437 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7438 return true;
7440 return false;
7442 else
7443 return true;
7446 /* Return true if libcall is dead as a whole. */
7448 static bool
7449 dead_libcall_p (rtx insn, int *counts)
7451 rtx note, set, new;
7453 /* See if there's a REG_EQUAL note on this insn and try to
7454 replace the source with the REG_EQUAL expression.
7456 We assume that insns with REG_RETVALs can only be reg->reg
7457 copies at this point. */
7458 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7459 if (!note)
7460 return false;
7462 set = single_set (insn);
7463 if (!set)
7464 return false;
7466 new = simplify_rtx (XEXP (note, 0));
7467 if (!new)
7468 new = XEXP (note, 0);
7470 /* While changing insn, we must update the counts accordingly. */
7471 count_reg_usage (insn, counts, -1);
7473 if (validate_change (insn, &SET_SRC (set), new, 0))
7475 count_reg_usage (insn, counts, 1);
7476 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7477 remove_note (insn, note);
7478 return true;
7481 if (CONSTANT_P (new))
7483 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7484 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7486 count_reg_usage (insn, counts, 1);
7487 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7488 remove_note (insn, note);
7489 return true;
7493 count_reg_usage (insn, counts, 1);
7494 return false;
7497 /* Scan all the insns and delete any that are dead; i.e., they store a register
7498 that is never used or they copy a register to itself.
7500 This is used to remove insns made obviously dead by cse, loop or other
7501 optimizations. It improves the heuristics in loop since it won't try to
7502 move dead invariants out of loops or make givs for dead quantities. The
7503 remaining passes of the compilation are also sped up. */
7506 delete_trivially_dead_insns (rtx insns, int nreg)
7508 int *counts;
7509 rtx insn, prev;
7510 int in_libcall = 0, dead_libcall = 0;
7511 int ndead = 0, nlastdead, niterations = 0;
7513 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7514 /* First count the number of times each register is used. */
7515 counts = xcalloc (nreg, sizeof (int));
7516 for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn))
7517 count_reg_usage (insn, counts, 1);
7521 nlastdead = ndead;
7522 niterations++;
7523 /* Go from the last insn to the first and delete insns that only set unused
7524 registers or copy a register to itself. As we delete an insn, remove
7525 usage counts for registers it uses.
7527 The first jump optimization pass may leave a real insn as the last
7528 insn in the function. We must not skip that insn or we may end
7529 up deleting code that is not really dead. */
7530 insn = get_last_insn ();
7531 if (! INSN_P (insn))
7532 insn = prev_real_insn (insn);
7534 for (; insn; insn = prev)
7536 int live_insn = 0;
7538 prev = prev_real_insn (insn);
7540 /* Don't delete any insns that are part of a libcall block unless
7541 we can delete the whole libcall block.
7543 Flow or loop might get confused if we did that. Remember
7544 that we are scanning backwards. */
7545 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7547 in_libcall = 1;
7548 live_insn = 1;
7549 dead_libcall = dead_libcall_p (insn, counts);
7551 else if (in_libcall)
7552 live_insn = ! dead_libcall;
7553 else
7554 live_insn = insn_live_p (insn, counts);
7556 /* If this is a dead insn, delete it and show registers in it aren't
7557 being used. */
7559 if (! live_insn)
7561 count_reg_usage (insn, counts, -1);
7562 delete_insn_and_edges (insn);
7563 ndead++;
7566 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7568 in_libcall = 0;
7569 dead_libcall = 0;
7573 while (ndead != nlastdead);
7575 if (dump_file && ndead)
7576 fprintf (dump_file, "Deleted %i trivially dead insns; %i iterations\n",
7577 ndead, niterations);
7578 /* Clean up. */
7579 free (counts);
7580 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7581 return ndead;
7584 /* This function is called via for_each_rtx. The argument, NEWREG, is
7585 a condition code register with the desired mode. If we are looking
7586 at the same register in a different mode, replace it with
7587 NEWREG. */
7589 static int
7590 cse_change_cc_mode (rtx *loc, void *data)
7592 rtx newreg = (rtx) data;
7594 if (*loc
7595 && REG_P (*loc)
7596 && REGNO (*loc) == REGNO (newreg)
7597 && GET_MODE (*loc) != GET_MODE (newreg))
7599 *loc = newreg;
7600 return -1;
7602 return 0;
7605 /* Change the mode of any reference to the register REGNO (NEWREG) to
7606 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7607 any instruction which modifies NEWREG. */
7609 static void
7610 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7612 rtx insn;
7614 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7616 if (! INSN_P (insn))
7617 continue;
7619 if (reg_set_p (newreg, insn))
7620 return;
7622 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, newreg);
7623 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, newreg);
7627 /* BB is a basic block which finishes with CC_REG as a condition code
7628 register which is set to CC_SRC. Look through the successors of BB
7629 to find blocks which have a single predecessor (i.e., this one),
7630 and look through those blocks for an assignment to CC_REG which is
7631 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7632 permitted to change the mode of CC_SRC to a compatible mode. This
7633 returns VOIDmode if no equivalent assignments were found.
7634 Otherwise it returns the mode which CC_SRC should wind up with.
7636 The main complexity in this function is handling the mode issues.
7637 We may have more than one duplicate which we can eliminate, and we
7638 try to find a mode which will work for multiple duplicates. */
7640 static enum machine_mode
7641 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7643 bool found_equiv;
7644 enum machine_mode mode;
7645 unsigned int insn_count;
7646 edge e;
7647 rtx insns[2];
7648 enum machine_mode modes[2];
7649 rtx last_insns[2];
7650 unsigned int i;
7651 rtx newreg;
7653 /* We expect to have two successors. Look at both before picking
7654 the final mode for the comparison. If we have more successors
7655 (i.e., some sort of table jump, although that seems unlikely),
7656 then we require all beyond the first two to use the same
7657 mode. */
7659 found_equiv = false;
7660 mode = GET_MODE (cc_src);
7661 insn_count = 0;
7662 for (e = bb->succ; e; e = e->succ_next)
7664 rtx insn;
7665 rtx end;
7667 if (e->flags & EDGE_COMPLEX)
7668 continue;
7670 if (! e->dest->pred
7671 || e->dest->pred->pred_next
7672 || e->dest == EXIT_BLOCK_PTR)
7673 continue;
7675 end = NEXT_INSN (BB_END (e->dest));
7676 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7678 rtx set;
7680 if (! INSN_P (insn))
7681 continue;
7683 /* If CC_SRC is modified, we have to stop looking for
7684 something which uses it. */
7685 if (modified_in_p (cc_src, insn))
7686 break;
7688 /* Check whether INSN sets CC_REG to CC_SRC. */
7689 set = single_set (insn);
7690 if (set
7691 && REG_P (SET_DEST (set))
7692 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7694 bool found;
7695 enum machine_mode set_mode;
7696 enum machine_mode comp_mode;
7698 found = false;
7699 set_mode = GET_MODE (SET_SRC (set));
7700 comp_mode = set_mode;
7701 if (rtx_equal_p (cc_src, SET_SRC (set)))
7702 found = true;
7703 else if (GET_CODE (cc_src) == COMPARE
7704 && GET_CODE (SET_SRC (set)) == COMPARE
7705 && mode != set_mode
7706 && rtx_equal_p (XEXP (cc_src, 0),
7707 XEXP (SET_SRC (set), 0))
7708 && rtx_equal_p (XEXP (cc_src, 1),
7709 XEXP (SET_SRC (set), 1)))
7712 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7713 if (comp_mode != VOIDmode
7714 && (can_change_mode || comp_mode == mode))
7715 found = true;
7718 if (found)
7720 found_equiv = true;
7721 if (insn_count < ARRAY_SIZE (insns))
7723 insns[insn_count] = insn;
7724 modes[insn_count] = set_mode;
7725 last_insns[insn_count] = end;
7726 ++insn_count;
7728 if (mode != comp_mode)
7730 if (! can_change_mode)
7731 abort ();
7732 mode = comp_mode;
7733 PUT_MODE (cc_src, mode);
7736 else
7738 if (set_mode != mode)
7740 /* We found a matching expression in the
7741 wrong mode, but we don't have room to
7742 store it in the array. Punt. This case
7743 should be rare. */
7744 break;
7746 /* INSN sets CC_REG to a value equal to CC_SRC
7747 with the right mode. We can simply delete
7748 it. */
7749 delete_insn (insn);
7752 /* We found an instruction to delete. Keep looking,
7753 in the hopes of finding a three-way jump. */
7754 continue;
7757 /* We found an instruction which sets the condition
7758 code, so don't look any farther. */
7759 break;
7762 /* If INSN sets CC_REG in some other way, don't look any
7763 farther. */
7764 if (reg_set_p (cc_reg, insn))
7765 break;
7768 /* If we fell off the bottom of the block, we can keep looking
7769 through successors. We pass CAN_CHANGE_MODE as false because
7770 we aren't prepared to handle compatibility between the
7771 further blocks and this block. */
7772 if (insn == end)
7774 enum machine_mode submode;
7776 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7777 if (submode != VOIDmode)
7779 if (submode != mode)
7780 abort ();
7781 found_equiv = true;
7782 can_change_mode = false;
7787 if (! found_equiv)
7788 return VOIDmode;
7790 /* Now INSN_COUNT is the number of instructions we found which set
7791 CC_REG to a value equivalent to CC_SRC. The instructions are in
7792 INSNS. The modes used by those instructions are in MODES. */
7794 newreg = NULL_RTX;
7795 for (i = 0; i < insn_count; ++i)
7797 if (modes[i] != mode)
7799 /* We need to change the mode of CC_REG in INSNS[i] and
7800 subsequent instructions. */
7801 if (! newreg)
7803 if (GET_MODE (cc_reg) == mode)
7804 newreg = cc_reg;
7805 else
7806 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7808 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7809 newreg);
7812 delete_insn (insns[i]);
7815 return mode;
7818 /* If we have a fixed condition code register (or two), walk through
7819 the instructions and try to eliminate duplicate assignments. */
7821 void
7822 cse_condition_code_reg (void)
7824 unsigned int cc_regno_1;
7825 unsigned int cc_regno_2;
7826 rtx cc_reg_1;
7827 rtx cc_reg_2;
7828 basic_block bb;
7830 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7831 return;
7833 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7834 if (cc_regno_2 != INVALID_REGNUM)
7835 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7836 else
7837 cc_reg_2 = NULL_RTX;
7839 FOR_EACH_BB (bb)
7841 rtx last_insn;
7842 rtx cc_reg;
7843 rtx insn;
7844 rtx cc_src_insn;
7845 rtx cc_src;
7846 enum machine_mode mode;
7847 enum machine_mode orig_mode;
7849 /* Look for blocks which end with a conditional jump based on a
7850 condition code register. Then look for the instruction which
7851 sets the condition code register. Then look through the
7852 successor blocks for instructions which set the condition
7853 code register to the same value. There are other possible
7854 uses of the condition code register, but these are by far the
7855 most common and the ones which we are most likely to be able
7856 to optimize. */
7858 last_insn = BB_END (bb);
7859 if (!JUMP_P (last_insn))
7860 continue;
7862 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7863 cc_reg = cc_reg_1;
7864 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7865 cc_reg = cc_reg_2;
7866 else
7867 continue;
7869 cc_src_insn = NULL_RTX;
7870 cc_src = NULL_RTX;
7871 for (insn = PREV_INSN (last_insn);
7872 insn && insn != PREV_INSN (BB_HEAD (bb));
7873 insn = PREV_INSN (insn))
7875 rtx set;
7877 if (! INSN_P (insn))
7878 continue;
7879 set = single_set (insn);
7880 if (set
7881 && REG_P (SET_DEST (set))
7882 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7884 cc_src_insn = insn;
7885 cc_src = SET_SRC (set);
7886 break;
7888 else if (reg_set_p (cc_reg, insn))
7889 break;
7892 if (! cc_src_insn)
7893 continue;
7895 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7896 continue;
7898 /* Now CC_REG is a condition code register used for a
7899 conditional jump at the end of the block, and CC_SRC, in
7900 CC_SRC_INSN, is the value to which that condition code
7901 register is set, and CC_SRC is still meaningful at the end of
7902 the basic block. */
7904 orig_mode = GET_MODE (cc_src);
7905 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7906 if (mode != VOIDmode)
7908 if (mode != GET_MODE (cc_src))
7909 abort ();
7910 if (mode != orig_mode)
7912 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7914 /* Change the mode of CC_REG in CC_SRC_INSN to
7915 GET_MODE (NEWREG). */
7916 for_each_rtx (&PATTERN (cc_src_insn), cse_change_cc_mode,
7917 newreg);
7918 for_each_rtx (&REG_NOTES (cc_src_insn), cse_change_cc_mode,
7919 newreg);
7921 /* Do the same in the following insns that use the
7922 current value of CC_REG within BB. */
7923 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7924 NEXT_INSN (last_insn),
7925 newreg);