* c-typeck.c (convert_arguments): Don't check for width changes
[official-gcc.git] / gcc / cse.c
blob6913aef32b514be478147000bad96f82baaa1ef6
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include <setjmp.h>
27 #include "rtl.h"
28 #include "tm_p.h"
29 #include "regs.h"
30 #include "hard-reg-set.h"
31 #include "basic-block.h"
32 #include "flags.h"
33 #include "real.h"
34 #include "insn-config.h"
35 #include "recog.h"
36 #include "function.h"
37 #include "expr.h"
38 #include "toplev.h"
39 #include "output.h"
40 #include "ggc.h"
42 /* The basic idea of common subexpression elimination is to go
43 through the code, keeping a record of expressions that would
44 have the same value at the current scan point, and replacing
45 expressions encountered with the cheapest equivalent expression.
47 It is too complicated to keep track of the different possibilities
48 when control paths merge in this code; so, at each label, we forget all
49 that is known and start fresh. This can be described as processing each
50 extended basic block separately. We have a separate pass to perform
51 global CSE.
53 Note CSE can turn a conditional or computed jump into a nop or
54 an unconditional jump. When this occurs we arrange to run the jump
55 optimizer after CSE to delete the unreachable code.
57 We use two data structures to record the equivalent expressions:
58 a hash table for most expressions, and a vector of "quantity
59 numbers" to record equivalent (pseudo) registers.
61 The use of the special data structure for registers is desirable
62 because it is faster. It is possible because registers references
63 contain a fairly small number, the register number, taken from
64 a contiguously allocated series, and two register references are
65 identical if they have the same number. General expressions
66 do not have any such thing, so the only way to retrieve the
67 information recorded on an expression other than a register
68 is to keep it in a hash table.
70 Registers and "quantity numbers":
72 At the start of each basic block, all of the (hardware and pseudo)
73 registers used in the function are given distinct quantity
74 numbers to indicate their contents. During scan, when the code
75 copies one register into another, we copy the quantity number.
76 When a register is loaded in any other way, we allocate a new
77 quantity number to describe the value generated by this operation.
78 `reg_qty' records what quantity a register is currently thought
79 of as containing.
81 All real quantity numbers are greater than or equal to `max_reg'.
82 If register N has not been assigned a quantity, reg_qty[N] will equal N.
84 Quantity numbers below `max_reg' do not exist and none of the `qty_table'
85 entries should be referenced with an index below `max_reg'.
87 We also maintain a bidirectional chain of registers for each
88 quantity number. The `qty_table` members `first_reg' and `last_reg',
89 and `reg_eqv_table' members `next' and `prev' hold these chains.
91 The first register in a chain is the one whose lifespan is least local.
92 Among equals, it is the one that was seen first.
93 We replace any equivalent register with that one.
95 If two registers have the same quantity number, it must be true that
96 REG expressions with qty_table `mode' must be in the hash table for both
97 registers and must be in the same class.
99 The converse is not true. Since hard registers may be referenced in
100 any mode, two REG expressions might be equivalent in the hash table
101 but not have the same quantity number if the quantity number of one
102 of the registers is not the same mode as those expressions.
104 Constants and quantity numbers
106 When a quantity has a known constant value, that value is stored
107 in the appropriate qty_table `const_rtx'. This is in addition to
108 putting the constant in the hash table as is usual for non-regs.
110 Whether a reg or a constant is preferred is determined by the configuration
111 macro CONST_COSTS and will often depend on the constant value. In any
112 event, expressions containing constants can be simplified, by fold_rtx.
114 When a quantity has a known nearly constant value (such as an address
115 of a stack slot), that value is stored in the appropriate qty_table
116 `const_rtx'.
118 Integer constants don't have a machine mode. However, cse
119 determines the intended machine mode from the destination
120 of the instruction that moves the constant. The machine mode
121 is recorded in the hash table along with the actual RTL
122 constant expression so that different modes are kept separate.
124 Other expressions:
126 To record known equivalences among expressions in general
127 we use a hash table called `table'. It has a fixed number of buckets
128 that contain chains of `struct table_elt' elements for expressions.
129 These chains connect the elements whose expressions have the same
130 hash codes.
132 Other chains through the same elements connect the elements which
133 currently have equivalent values.
135 Register references in an expression are canonicalized before hashing
136 the expression. This is done using `reg_qty' and qty_table `first_reg'.
137 The hash code of a register reference is computed using the quantity
138 number, not the register number.
140 When the value of an expression changes, it is necessary to remove from the
141 hash table not just that expression but all expressions whose values
142 could be different as a result.
144 1. If the value changing is in memory, except in special cases
145 ANYTHING referring to memory could be changed. That is because
146 nobody knows where a pointer does not point.
147 The function `invalidate_memory' removes what is necessary.
149 The special cases are when the address is constant or is
150 a constant plus a fixed register such as the frame pointer
151 or a static chain pointer. When such addresses are stored in,
152 we can tell exactly which other such addresses must be invalidated
153 due to overlap. `invalidate' does this.
154 All expressions that refer to non-constant
155 memory addresses are also invalidated. `invalidate_memory' does this.
157 2. If the value changing is a register, all expressions
158 containing references to that register, and only those,
159 must be removed.
161 Because searching the entire hash table for expressions that contain
162 a register is very slow, we try to figure out when it isn't necessary.
163 Precisely, this is necessary only when expressions have been
164 entered in the hash table using this register, and then the value has
165 changed, and then another expression wants to be added to refer to
166 the register's new value. This sequence of circumstances is rare
167 within any one basic block.
169 The vectors `reg_tick' and `reg_in_table' are used to detect this case.
170 reg_tick[i] is incremented whenever a value is stored in register i.
171 reg_in_table[i] holds -1 if no references to register i have been
172 entered in the table; otherwise, it contains the value reg_tick[i] had
173 when the references were entered. If we want to enter a reference
174 and reg_in_table[i] != reg_tick[i], we must scan and remove old references.
175 Until we want to enter a new entry, the mere fact that the two vectors
176 don't match makes the entries be ignored if anyone tries to match them.
178 Registers themselves are entered in the hash table as well as in
179 the equivalent-register chains. However, the vectors `reg_tick'
180 and `reg_in_table' do not apply to expressions which are simple
181 register references. These expressions are removed from the table
182 immediately when they become invalid, and this can be done even if
183 we do not immediately search for all the expressions that refer to
184 the register.
186 A CLOBBER rtx in an instruction invalidates its operand for further
187 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
188 invalidates everything that resides in memory.
190 Related expressions:
192 Constant expressions that differ only by an additive integer
193 are called related. When a constant expression is put in
194 the table, the related expression with no constant term
195 is also entered. These are made to point at each other
196 so that it is possible to find out if there exists any
197 register equivalent to an expression related to a given expression. */
199 /* One plus largest register number used in this function. */
201 static int max_reg;
203 /* One plus largest instruction UID used in this function at time of
204 cse_main call. */
206 static int max_insn_uid;
208 /* Length of qty_table vector. We know in advance we will not need
209 a quantity number this big. */
211 static int max_qty;
213 /* Next quantity number to be allocated.
214 This is 1 + the largest number needed so far. */
216 static int next_qty;
218 /* Per-qty information tracking.
220 `first_reg' and `last_reg' track the head and tail of the
221 chain of registers which currently contain this quantity.
223 `mode' contains the machine mode of this quantity.
225 `const_rtx' holds the rtx of the constant value of this
226 quantity, if known. A summations of the frame/arg pointer
227 and a constant can also be entered here. When this holds
228 a known value, `const_insn' is the insn which stored the
229 constant value.
231 `comparison_{code,const,qty}' are used to track when a
232 comparison between a quantity and some constant or register has
233 been passed. In such a case, we know the results of the comparison
234 in case we see it again. These members record a comparison that
235 is known to be true. `comparison_code' holds the rtx code of such
236 a comparison, else it is set to UNKNOWN and the other two
237 comparison members are undefined. `comparison_const' holds
238 the constant being compared against, or zero if the comparison
239 is not against a constant. `comparison_qty' holds the quantity
240 being compared against when the result is known. If the comparison
241 is not with a register, `comparison_qty' is -1. */
243 struct qty_table_elem
245 rtx const_rtx;
246 rtx const_insn;
247 rtx comparison_const;
248 int comparison_qty;
249 unsigned int first_reg, last_reg;
250 enum machine_mode mode;
251 enum rtx_code comparison_code;
254 /* The table of all qtys, indexed by qty number. */
255 static struct qty_table_elem *qty_table;
257 #ifdef HAVE_cc0
258 /* For machines that have a CC0, we do not record its value in the hash
259 table since its use is guaranteed to be the insn immediately following
260 its definition and any other insn is presumed to invalidate it.
262 Instead, we store below the value last assigned to CC0. If it should
263 happen to be a constant, it is stored in preference to the actual
264 assigned value. In case it is a constant, we store the mode in which
265 the constant should be interpreted. */
267 static rtx prev_insn_cc0;
268 static enum machine_mode prev_insn_cc0_mode;
269 #endif
271 /* Previous actual insn. 0 if at first insn of basic block. */
273 static rtx prev_insn;
275 /* Insn being scanned. */
277 static rtx this_insn;
279 /* Index by register number, gives the number of the next (or
280 previous) register in the chain of registers sharing the same
281 value.
283 Or -1 if this register is at the end of the chain.
285 If reg_qty[N] == N, reg_eqv_table[N].next is undefined. */
287 /* Per-register equivalence chain. */
288 struct reg_eqv_elem
290 int next, prev;
293 /* The table of all register equivalence chains. */
294 static struct reg_eqv_elem *reg_eqv_table;
296 struct cse_reg_info
298 /* Next in hash chain. */
299 struct cse_reg_info *hash_next;
301 /* The next cse_reg_info structure in the free or used list. */
302 struct cse_reg_info *next;
304 /* Search key */
305 unsigned int regno;
307 /* The quantity number of the register's current contents. */
308 int reg_qty;
310 /* The number of times the register has been altered in the current
311 basic block. */
312 int reg_tick;
314 /* The REG_TICK value at which rtx's containing this register are
315 valid in the hash table. If this does not equal the current
316 reg_tick value, such expressions existing in the hash table are
317 invalid. */
318 int reg_in_table;
321 /* A free list of cse_reg_info entries. */
322 static struct cse_reg_info *cse_reg_info_free_list;
324 /* A used list of cse_reg_info entries. */
325 static struct cse_reg_info *cse_reg_info_used_list;
326 static struct cse_reg_info *cse_reg_info_used_list_end;
328 /* A mapping from registers to cse_reg_info data structures. */
329 #define REGHASH_SHIFT 7
330 #define REGHASH_SIZE (1 << REGHASH_SHIFT)
331 #define REGHASH_MASK (REGHASH_SIZE - 1)
332 static struct cse_reg_info *reg_hash[REGHASH_SIZE];
334 #define REGHASH_FN(REGNO) \
335 (((REGNO) ^ ((REGNO) >> REGHASH_SHIFT)) & REGHASH_MASK)
337 /* The last lookup we did into the cse_reg_info_tree. This allows us
338 to cache repeated lookups. */
339 static unsigned int cached_regno;
340 static struct cse_reg_info *cached_cse_reg_info;
342 /* A HARD_REG_SET containing all the hard registers for which there is
343 currently a REG expression in the hash table. Note the difference
344 from the above variables, which indicate if the REG is mentioned in some
345 expression in the table. */
347 static HARD_REG_SET hard_regs_in_table;
349 /* A HARD_REG_SET containing all the hard registers that are invalidated
350 by a CALL_INSN. */
352 static HARD_REG_SET regs_invalidated_by_call;
354 /* CUID of insn that starts the basic block currently being cse-processed. */
356 static int cse_basic_block_start;
358 /* CUID of insn that ends the basic block currently being cse-processed. */
360 static int cse_basic_block_end;
362 /* Vector mapping INSN_UIDs to cuids.
363 The cuids are like uids but increase monotonically always.
364 We use them to see whether a reg is used outside a given basic block. */
366 static int *uid_cuid;
368 /* Highest UID in UID_CUID. */
369 static int max_uid;
371 /* Get the cuid of an insn. */
373 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
375 /* Nonzero if this pass has made changes, and therefore it's
376 worthwhile to run the garbage collector. */
378 static int cse_altered;
380 /* Nonzero if cse has altered conditional jump insns
381 in such a way that jump optimization should be redone. */
383 static int cse_jumps_altered;
385 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
386 REG_LABEL, we have to rerun jump after CSE to put in the note. */
387 static int recorded_label_ref;
389 /* Says which LABEL_REF was put in the hash table. Used to see if we need
390 to set the above flag. */
391 static rtx new_label_ref;
393 /* canon_hash stores 1 in do_not_record
394 if it notices a reference to CC0, PC, or some other volatile
395 subexpression. */
397 static int do_not_record;
399 #ifdef LOAD_EXTEND_OP
401 /* Scratch rtl used when looking for load-extended copy of a MEM. */
402 static rtx memory_extend_rtx;
403 #endif
405 /* canon_hash stores 1 in hash_arg_in_memory
406 if it notices a reference to memory within the expression being hashed. */
408 static int hash_arg_in_memory;
410 /* The hash table contains buckets which are chains of `struct table_elt's,
411 each recording one expression's information.
412 That expression is in the `exp' field.
414 The canon_exp field contains a canonical (from the point of view of
415 alias analysis) version of the `exp' field.
417 Those elements with the same hash code are chained in both directions
418 through the `next_same_hash' and `prev_same_hash' fields.
420 Each set of expressions with equivalent values
421 are on a two-way chain through the `next_same_value'
422 and `prev_same_value' fields, and all point with
423 the `first_same_value' field at the first element in
424 that chain. The chain is in order of increasing cost.
425 Each element's cost value is in its `cost' field.
427 The `in_memory' field is nonzero for elements that
428 involve any reference to memory. These elements are removed
429 whenever a write is done to an unidentified location in memory.
430 To be safe, we assume that a memory address is unidentified unless
431 the address is either a symbol constant or a constant plus
432 the frame pointer or argument pointer.
434 The `related_value' field is used to connect related expressions
435 (that differ by adding an integer).
436 The related expressions are chained in a circular fashion.
437 `related_value' is zero for expressions for which this
438 chain is not useful.
440 The `cost' field stores the cost of this element's expression.
441 The `regcost' field stores the value returned by approx_reg_cost for
442 this element's expression.
444 The `is_const' flag is set if the element is a constant (including
445 a fixed address).
447 The `flag' field is used as a temporary during some search routines.
449 The `mode' field is usually the same as GET_MODE (`exp'), but
450 if `exp' is a CONST_INT and has no machine mode then the `mode'
451 field is the mode it was being used as. Each constant is
452 recorded separately for each mode it is used with. */
454 struct table_elt
456 rtx exp;
457 rtx canon_exp;
458 struct table_elt *next_same_hash;
459 struct table_elt *prev_same_hash;
460 struct table_elt *next_same_value;
461 struct table_elt *prev_same_value;
462 struct table_elt *first_same_value;
463 struct table_elt *related_value;
464 int cost;
465 int regcost;
466 enum machine_mode mode;
467 char in_memory;
468 char is_const;
469 char flag;
472 /* We don't want a lot of buckets, because we rarely have very many
473 things stored in the hash table, and a lot of buckets slows
474 down a lot of loops that happen frequently. */
475 #define HASH_SHIFT 5
476 #define HASH_SIZE (1 << HASH_SHIFT)
477 #define HASH_MASK (HASH_SIZE - 1)
479 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
480 register (hard registers may require `do_not_record' to be set). */
482 #define HASH(X, M) \
483 ((GET_CODE (X) == REG && REGNO (X) >= FIRST_PSEUDO_REGISTER \
484 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
485 : canon_hash (X, M)) & HASH_MASK)
487 /* Determine whether register number N is considered a fixed register for the
488 purpose of approximating register costs.
489 It is desirable to replace other regs with fixed regs, to reduce need for
490 non-fixed hard regs.
491 A reg wins if it is either the frame pointer or designated as fixed. */
492 #define FIXED_REGNO_P(N) \
493 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
494 || fixed_regs[N] || global_regs[N])
496 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
497 hard registers and pointers into the frame are the cheapest with a cost
498 of 0. Next come pseudos with a cost of one and other hard registers with
499 a cost of 2. Aside from these special cases, call `rtx_cost'. */
501 #define CHEAP_REGNO(N) \
502 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
503 || (N) == STACK_POINTER_REGNUM || (N) == ARG_POINTER_REGNUM \
504 || ((N) >= FIRST_VIRTUAL_REGISTER && (N) <= LAST_VIRTUAL_REGISTER) \
505 || ((N) < FIRST_PSEUDO_REGISTER \
506 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
508 #define COST(X) (GET_CODE (X) == REG ? 0 : notreg_cost (X, SET))
509 #define COST_IN(X,OUTER) (GET_CODE (X) == REG ? 0 : notreg_cost (X, OUTER))
511 /* Get the info associated with register N. */
513 #define GET_CSE_REG_INFO(N) \
514 (((N) == cached_regno && cached_cse_reg_info) \
515 ? cached_cse_reg_info : get_cse_reg_info ((N)))
517 /* Get the number of times this register has been updated in this
518 basic block. */
520 #define REG_TICK(N) ((GET_CSE_REG_INFO (N))->reg_tick)
522 /* Get the point at which REG was recorded in the table. */
524 #define REG_IN_TABLE(N) ((GET_CSE_REG_INFO (N))->reg_in_table)
526 /* Get the quantity number for REG. */
528 #define REG_QTY(N) ((GET_CSE_REG_INFO (N))->reg_qty)
530 /* Determine if the quantity number for register X represents a valid index
531 into the qty_table. */
533 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) != (int) (N))
535 static struct table_elt *table[HASH_SIZE];
537 /* Chain of `struct table_elt's made so far for this function
538 but currently removed from the table. */
540 static struct table_elt *free_element_chain;
542 /* Number of `struct table_elt' structures made so far for this function. */
544 static int n_elements_made;
546 /* Maximum value `n_elements_made' has had so far in this compilation
547 for functions previously processed. */
549 static int max_elements_made;
551 /* Surviving equivalence class when two equivalence classes are merged
552 by recording the effects of a jump in the last insn. Zero if the
553 last insn was not a conditional jump. */
555 static struct table_elt *last_jump_equiv_class;
557 /* Set to the cost of a constant pool reference if one was found for a
558 symbolic constant. If this was found, it means we should try to
559 convert constants into constant pool entries if they don't fit in
560 the insn. */
562 static int constant_pool_entries_cost;
564 /* Define maximum length of a branch path. */
566 #define PATHLENGTH 10
568 /* This data describes a block that will be processed by cse_basic_block. */
570 struct cse_basic_block_data
572 /* Lowest CUID value of insns in block. */
573 int low_cuid;
574 /* Highest CUID value of insns in block. */
575 int high_cuid;
576 /* Total number of SETs in block. */
577 int nsets;
578 /* Last insn in the block. */
579 rtx last;
580 /* Size of current branch path, if any. */
581 int path_size;
582 /* Current branch path, indicating which branches will be taken. */
583 struct branch_path
585 /* The branch insn. */
586 rtx branch;
587 /* Whether it should be taken or not. AROUND is the same as taken
588 except that it is used when the destination label is not preceded
589 by a BARRIER. */
590 enum taken {TAKEN, NOT_TAKEN, AROUND} status;
591 } path[PATHLENGTH];
594 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
595 virtual regs here because the simplify_*_operation routines are called
596 by integrate.c, which is called before virtual register instantiation.
598 ?!? FIXED_BASE_PLUS_P and NONZERO_BASE_PLUS_P need to move into
599 a header file so that their definitions can be shared with the
600 simplification routines in simplify-rtx.c. Until then, do not
601 change these macros without also changing the copy in simplify-rtx.c. */
603 #define FIXED_BASE_PLUS_P(X) \
604 ((X) == frame_pointer_rtx || (X) == hard_frame_pointer_rtx \
605 || ((X) == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])\
606 || (X) == virtual_stack_vars_rtx \
607 || (X) == virtual_incoming_args_rtx \
608 || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \
609 && (XEXP (X, 0) == frame_pointer_rtx \
610 || XEXP (X, 0) == hard_frame_pointer_rtx \
611 || ((X) == arg_pointer_rtx \
612 && fixed_regs[ARG_POINTER_REGNUM]) \
613 || XEXP (X, 0) == virtual_stack_vars_rtx \
614 || XEXP (X, 0) == virtual_incoming_args_rtx)) \
615 || GET_CODE (X) == ADDRESSOF)
617 /* Similar, but also allows reference to the stack pointer.
619 This used to include FIXED_BASE_PLUS_P, however, we can't assume that
620 arg_pointer_rtx by itself is nonzero, because on at least one machine,
621 the i960, the arg pointer is zero when it is unused. */
623 #define NONZERO_BASE_PLUS_P(X) \
624 ((X) == frame_pointer_rtx || (X) == hard_frame_pointer_rtx \
625 || (X) == virtual_stack_vars_rtx \
626 || (X) == virtual_incoming_args_rtx \
627 || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \
628 && (XEXP (X, 0) == frame_pointer_rtx \
629 || XEXP (X, 0) == hard_frame_pointer_rtx \
630 || ((X) == arg_pointer_rtx \
631 && fixed_regs[ARG_POINTER_REGNUM]) \
632 || XEXP (X, 0) == virtual_stack_vars_rtx \
633 || XEXP (X, 0) == virtual_incoming_args_rtx)) \
634 || (X) == stack_pointer_rtx \
635 || (X) == virtual_stack_dynamic_rtx \
636 || (X) == virtual_outgoing_args_rtx \
637 || (GET_CODE (X) == PLUS && GET_CODE (XEXP (X, 1)) == CONST_INT \
638 && (XEXP (X, 0) == stack_pointer_rtx \
639 || XEXP (X, 0) == virtual_stack_dynamic_rtx \
640 || XEXP (X, 0) == virtual_outgoing_args_rtx)) \
641 || GET_CODE (X) == ADDRESSOF)
643 static int notreg_cost PARAMS ((rtx, enum rtx_code));
644 static int approx_reg_cost_1 PARAMS ((rtx *, void *));
645 static int approx_reg_cost PARAMS ((rtx));
646 static int preferrable PARAMS ((int, int, int, int));
647 static void new_basic_block PARAMS ((void));
648 static void make_new_qty PARAMS ((unsigned int, enum machine_mode));
649 static void make_regs_eqv PARAMS ((unsigned int, unsigned int));
650 static void delete_reg_equiv PARAMS ((unsigned int));
651 static int mention_regs PARAMS ((rtx));
652 static int insert_regs PARAMS ((rtx, struct table_elt *, int));
653 static void remove_from_table PARAMS ((struct table_elt *, unsigned));
654 static struct table_elt *lookup PARAMS ((rtx, unsigned, enum machine_mode)),
655 *lookup_for_remove PARAMS ((rtx, unsigned, enum machine_mode));
656 static rtx lookup_as_function PARAMS ((rtx, enum rtx_code));
657 static struct table_elt *insert PARAMS ((rtx, struct table_elt *, unsigned,
658 enum machine_mode));
659 static void merge_equiv_classes PARAMS ((struct table_elt *,
660 struct table_elt *));
661 static void invalidate PARAMS ((rtx, enum machine_mode));
662 static int cse_rtx_varies_p PARAMS ((rtx, int));
663 static void remove_invalid_refs PARAMS ((unsigned int));
664 static void remove_invalid_subreg_refs PARAMS ((unsigned int, unsigned int,
665 enum machine_mode));
666 static void rehash_using_reg PARAMS ((rtx));
667 static void invalidate_memory PARAMS ((void));
668 static void invalidate_for_call PARAMS ((void));
669 static rtx use_related_value PARAMS ((rtx, struct table_elt *));
670 static unsigned canon_hash PARAMS ((rtx, enum machine_mode));
671 static unsigned canon_hash_string PARAMS ((const char *));
672 static unsigned safe_hash PARAMS ((rtx, enum machine_mode));
673 static int exp_equiv_p PARAMS ((rtx, rtx, int, int));
674 static rtx canon_reg PARAMS ((rtx, rtx));
675 static void find_best_addr PARAMS ((rtx, rtx *, enum machine_mode));
676 static enum rtx_code find_comparison_args PARAMS ((enum rtx_code, rtx *, rtx *,
677 enum machine_mode *,
678 enum machine_mode *));
679 static rtx fold_rtx PARAMS ((rtx, rtx));
680 static rtx equiv_constant PARAMS ((rtx));
681 static void record_jump_equiv PARAMS ((rtx, int));
682 static void record_jump_cond PARAMS ((enum rtx_code, enum machine_mode,
683 rtx, rtx, int));
684 static void cse_insn PARAMS ((rtx, rtx));
685 static int addr_affects_sp_p PARAMS ((rtx));
686 static void invalidate_from_clobbers PARAMS ((rtx));
687 static rtx cse_process_notes PARAMS ((rtx, rtx));
688 static void cse_around_loop PARAMS ((rtx));
689 static void invalidate_skipped_set PARAMS ((rtx, rtx, void *));
690 static void invalidate_skipped_block PARAMS ((rtx));
691 static void cse_check_loop_start PARAMS ((rtx, rtx, void *));
692 static void cse_set_around_loop PARAMS ((rtx, rtx, rtx));
693 static rtx cse_basic_block PARAMS ((rtx, rtx, struct branch_path *, int));
694 static void count_reg_usage PARAMS ((rtx, int *, rtx, int));
695 extern void dump_class PARAMS ((struct table_elt*));
696 static struct cse_reg_info * get_cse_reg_info PARAMS ((unsigned int));
697 static int check_dependence PARAMS ((rtx *, void *));
699 static void flush_hash_table PARAMS ((void));
701 /* Dump the expressions in the equivalence class indicated by CLASSP.
702 This function is used only for debugging. */
703 void
704 dump_class (classp)
705 struct table_elt *classp;
707 struct table_elt *elt;
709 fprintf (stderr, "Equivalence chain for ");
710 print_rtl (stderr, classp->exp);
711 fprintf (stderr, ": \n");
713 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
715 print_rtl (stderr, elt->exp);
716 fprintf (stderr, "\n");
720 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
721 static int
722 approx_reg_cost_1 (xp, data)
723 rtx *xp;
724 void *data;
726 rtx x = *xp;
727 regset set = (regset) data;
729 if (x && GET_CODE (x) == REG)
730 SET_REGNO_REG_SET (set, REGNO (x));
731 return 0;
734 /* Return an estimate of the cost of the registers used in an rtx.
735 This is mostly the number of different REG expressions in the rtx;
736 however for some excecptions like fixed registers we use a cost of
737 0. If any other hard register reference occurs, return MAX_COST. */
739 static int
740 approx_reg_cost (x)
741 rtx x;
743 regset_head set;
744 int i;
745 int cost = 0;
746 int hardregs = 0;
748 INIT_REG_SET (&set);
749 for_each_rtx (&x, approx_reg_cost_1, (void *)&set);
751 EXECUTE_IF_SET_IN_REG_SET
752 (&set, 0, i,
754 if (! CHEAP_REGNO (i))
756 if (i < FIRST_PSEUDO_REGISTER)
757 hardregs++;
759 cost += i < FIRST_PSEUDO_REGISTER ? 2 : 1;
763 CLEAR_REG_SET (&set);
764 return hardregs && SMALL_REGISTER_CLASSES ? MAX_COST : cost;
767 /* Return a negative value if an rtx A, whose costs are given by COST_A
768 and REGCOST_A, is more desirable than an rtx B.
769 Return a positive value if A is less desirable, or 0 if the two are
770 equally good. */
771 static int
772 preferrable (cost_a, regcost_a, cost_b, regcost_b)
773 int cost_a, regcost_a, cost_b, regcost_b;
775 /* First, get rid of a cases involving expressions that are entirely
776 unwanted. */
777 if (cost_a != cost_b)
779 if (cost_a == MAX_COST)
780 return 1;
781 if (cost_b == MAX_COST)
782 return -1;
785 /* Avoid extending lifetimes of hardregs. */
786 if (regcost_a != regcost_b)
788 if (regcost_a == MAX_COST)
789 return 1;
790 if (regcost_b == MAX_COST)
791 return -1;
794 /* Normal operation costs take precedence. */
795 if (cost_a != cost_b)
796 return cost_a - cost_b;
797 /* Only if these are identical consider effects on register pressure. */
798 if (regcost_a != regcost_b)
799 return regcost_a - regcost_b;
800 return 0;
803 /* Internal function, to compute cost when X is not a register; called
804 from COST macro to keep it simple. */
806 static int
807 notreg_cost (x, outer)
808 rtx x;
809 enum rtx_code outer;
811 return ((GET_CODE (x) == SUBREG
812 && GET_CODE (SUBREG_REG (x)) == REG
813 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
814 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
815 && (GET_MODE_SIZE (GET_MODE (x))
816 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
817 && subreg_lowpart_p (x)
818 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
819 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
821 : rtx_cost (x, outer) * 2);
824 /* Return an estimate of the cost of computing rtx X.
825 One use is in cse, to decide which expression to keep in the hash table.
826 Another is in rtl generation, to pick the cheapest way to multiply.
827 Other uses like the latter are expected in the future. */
830 rtx_cost (x, outer_code)
831 rtx x;
832 enum rtx_code outer_code ATTRIBUTE_UNUSED;
834 register int i, j;
835 register enum rtx_code code;
836 register const char *fmt;
837 register int total;
839 if (x == 0)
840 return 0;
842 /* Compute the default costs of certain things.
843 Note that RTX_COSTS can override the defaults. */
845 code = GET_CODE (x);
846 switch (code)
848 case MULT:
849 /* Count multiplication by 2**n as a shift,
850 because if we are considering it, we would output it as a shift. */
851 if (GET_CODE (XEXP (x, 1)) == CONST_INT
852 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0)
853 total = 2;
854 else
855 total = COSTS_N_INSNS (5);
856 break;
857 case DIV:
858 case UDIV:
859 case MOD:
860 case UMOD:
861 total = COSTS_N_INSNS (7);
862 break;
863 case USE:
864 /* Used in loop.c and combine.c as a marker. */
865 total = 0;
866 break;
867 default:
868 total = COSTS_N_INSNS (1);
871 switch (code)
873 case REG:
874 return 0;
876 case SUBREG:
877 /* If we can't tie these modes, make this expensive. The larger
878 the mode, the more expensive it is. */
879 if (! MODES_TIEABLE_P (GET_MODE (x), GET_MODE (SUBREG_REG (x))))
880 return COSTS_N_INSNS (2
881 + GET_MODE_SIZE (GET_MODE (x)) / UNITS_PER_WORD);
882 break;
884 #ifdef RTX_COSTS
885 RTX_COSTS (x, code, outer_code);
886 #endif
887 #ifdef CONST_COSTS
888 CONST_COSTS (x, code, outer_code);
889 #endif
891 default:
892 #ifdef DEFAULT_RTX_COSTS
893 DEFAULT_RTX_COSTS (x, code, outer_code);
894 #endif
895 break;
898 /* Sum the costs of the sub-rtx's, plus cost of this operation,
899 which is already in total. */
901 fmt = GET_RTX_FORMAT (code);
902 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
903 if (fmt[i] == 'e')
904 total += rtx_cost (XEXP (x, i), code);
905 else if (fmt[i] == 'E')
906 for (j = 0; j < XVECLEN (x, i); j++)
907 total += rtx_cost (XVECEXP (x, i, j), code);
909 return total;
912 /* Return cost of address expression X.
913 Expect that X is propertly formed address reference. */
916 address_cost (x, mode)
917 rtx x;
918 enum machine_mode mode;
920 /* The ADDRESS_COST macro does not deal with ADDRESSOF nodes. But,
921 during CSE, such nodes are present. Using an ADDRESSOF node which
922 refers to the address of a REG is a good thing because we can then
923 turn (MEM (ADDRESSSOF (REG))) into just plain REG. */
925 if (GET_CODE (x) == ADDRESSOF && REG_P (XEXP ((x), 0)))
926 return -1;
928 /* We may be asked for cost of various unusual addresses, such as operands
929 of push instruction. It is not worthwhile to complicate writing
930 of ADDRESS_COST macro by such cases. */
932 if (!memory_address_p (mode, x))
933 return 1000;
934 #ifdef ADDRESS_COST
935 return ADDRESS_COST (x);
936 #else
937 return rtx_cost (x, MEM);
938 #endif
942 static struct cse_reg_info *
943 get_cse_reg_info (regno)
944 unsigned int regno;
946 struct cse_reg_info **hash_head = &reg_hash[REGHASH_FN (regno)];
947 struct cse_reg_info *p;
949 for (p = *hash_head; p != NULL; p = p->hash_next)
950 if (p->regno == regno)
951 break;
953 if (p == NULL)
955 /* Get a new cse_reg_info structure. */
956 if (cse_reg_info_free_list)
958 p = cse_reg_info_free_list;
959 cse_reg_info_free_list = p->next;
961 else
962 p = (struct cse_reg_info *) xmalloc (sizeof (struct cse_reg_info));
964 /* Insert into hash table. */
965 p->hash_next = *hash_head;
966 *hash_head = p;
968 /* Initialize it. */
969 p->reg_tick = 1;
970 p->reg_in_table = -1;
971 p->reg_qty = regno;
972 p->regno = regno;
973 p->next = cse_reg_info_used_list;
974 cse_reg_info_used_list = p;
975 if (!cse_reg_info_used_list_end)
976 cse_reg_info_used_list_end = p;
979 /* Cache this lookup; we tend to be looking up information about the
980 same register several times in a row. */
981 cached_regno = regno;
982 cached_cse_reg_info = p;
984 return p;
987 /* Clear the hash table and initialize each register with its own quantity,
988 for a new basic block. */
990 static void
991 new_basic_block ()
993 register int i;
995 next_qty = max_reg;
997 /* Clear out hash table state for this pass. */
999 memset ((char *) reg_hash, 0, sizeof reg_hash);
1001 if (cse_reg_info_used_list)
1003 cse_reg_info_used_list_end->next = cse_reg_info_free_list;
1004 cse_reg_info_free_list = cse_reg_info_used_list;
1005 cse_reg_info_used_list = cse_reg_info_used_list_end = 0;
1007 cached_cse_reg_info = 0;
1009 CLEAR_HARD_REG_SET (hard_regs_in_table);
1011 /* The per-quantity values used to be initialized here, but it is
1012 much faster to initialize each as it is made in `make_new_qty'. */
1014 for (i = 0; i < HASH_SIZE; i++)
1016 struct table_elt *first;
1018 first = table[i];
1019 if (first != NULL)
1021 struct table_elt *last = first;
1023 table[i] = NULL;
1025 while (last->next_same_hash != NULL)
1026 last = last->next_same_hash;
1028 /* Now relink this hash entire chain into
1029 the free element list. */
1031 last->next_same_hash = free_element_chain;
1032 free_element_chain = first;
1036 prev_insn = 0;
1038 #ifdef HAVE_cc0
1039 prev_insn_cc0 = 0;
1040 #endif
1043 /* Say that register REG contains a quantity in mode MODE not in any
1044 register before and initialize that quantity. */
1046 static void
1047 make_new_qty (reg, mode)
1048 unsigned int reg;
1049 enum machine_mode mode;
1051 register int q;
1052 register struct qty_table_elem *ent;
1053 register struct reg_eqv_elem *eqv;
1055 if (next_qty >= max_qty)
1056 abort ();
1058 q = REG_QTY (reg) = next_qty++;
1059 ent = &qty_table[q];
1060 ent->first_reg = reg;
1061 ent->last_reg = reg;
1062 ent->mode = mode;
1063 ent->const_rtx = ent->const_insn = NULL_RTX;
1064 ent->comparison_code = UNKNOWN;
1066 eqv = &reg_eqv_table[reg];
1067 eqv->next = eqv->prev = -1;
1070 /* Make reg NEW equivalent to reg OLD.
1071 OLD is not changing; NEW is. */
1073 static void
1074 make_regs_eqv (new, old)
1075 unsigned int new, old;
1077 unsigned int lastr, firstr;
1078 int q = REG_QTY (old);
1079 struct qty_table_elem *ent;
1081 ent = &qty_table[q];
1083 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1084 if (! REGNO_QTY_VALID_P (old))
1085 abort ();
1087 REG_QTY (new) = q;
1088 firstr = ent->first_reg;
1089 lastr = ent->last_reg;
1091 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1092 hard regs. Among pseudos, if NEW will live longer than any other reg
1093 of the same qty, and that is beyond the current basic block,
1094 make it the new canonical replacement for this qty. */
1095 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1096 /* Certain fixed registers might be of the class NO_REGS. This means
1097 that not only can they not be allocated by the compiler, but
1098 they cannot be used in substitutions or canonicalizations
1099 either. */
1100 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1101 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1102 || (new >= FIRST_PSEUDO_REGISTER
1103 && (firstr < FIRST_PSEUDO_REGISTER
1104 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1105 || (uid_cuid[REGNO_FIRST_UID (new)]
1106 < cse_basic_block_start))
1107 && (uid_cuid[REGNO_LAST_UID (new)]
1108 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1110 reg_eqv_table[firstr].prev = new;
1111 reg_eqv_table[new].next = firstr;
1112 reg_eqv_table[new].prev = -1;
1113 ent->first_reg = new;
1115 else
1117 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1118 Otherwise, insert before any non-fixed hard regs that are at the
1119 end. Registers of class NO_REGS cannot be used as an
1120 equivalent for anything. */
1121 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1122 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1123 && new >= FIRST_PSEUDO_REGISTER)
1124 lastr = reg_eqv_table[lastr].prev;
1125 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1126 if (reg_eqv_table[lastr].next >= 0)
1127 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1128 else
1129 qty_table[q].last_reg = new;
1130 reg_eqv_table[lastr].next = new;
1131 reg_eqv_table[new].prev = lastr;
1135 /* Remove REG from its equivalence class. */
1137 static void
1138 delete_reg_equiv (reg)
1139 unsigned int reg;
1141 register struct qty_table_elem *ent;
1142 register int q = REG_QTY (reg);
1143 register int p, n;
1145 /* If invalid, do nothing. */
1146 if (q == (int) reg)
1147 return;
1149 ent = &qty_table[q];
1151 p = reg_eqv_table[reg].prev;
1152 n = reg_eqv_table[reg].next;
1154 if (n != -1)
1155 reg_eqv_table[n].prev = p;
1156 else
1157 ent->last_reg = p;
1158 if (p != -1)
1159 reg_eqv_table[p].next = n;
1160 else
1161 ent->first_reg = n;
1163 REG_QTY (reg) = reg;
1166 /* Remove any invalid expressions from the hash table
1167 that refer to any of the registers contained in expression X.
1169 Make sure that newly inserted references to those registers
1170 as subexpressions will be considered valid.
1172 mention_regs is not called when a register itself
1173 is being stored in the table.
1175 Return 1 if we have done something that may have changed the hash code
1176 of X. */
1178 static int
1179 mention_regs (x)
1180 rtx x;
1182 register enum rtx_code code;
1183 register int i, j;
1184 register const char *fmt;
1185 register int changed = 0;
1187 if (x == 0)
1188 return 0;
1190 code = GET_CODE (x);
1191 if (code == REG)
1193 unsigned int regno = REGNO (x);
1194 unsigned int endregno
1195 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1196 : HARD_REGNO_NREGS (regno, GET_MODE (x)));
1197 unsigned int i;
1199 for (i = regno; i < endregno; i++)
1201 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1202 remove_invalid_refs (i);
1204 REG_IN_TABLE (i) = REG_TICK (i);
1207 return 0;
1210 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1211 pseudo if they don't use overlapping words. We handle only pseudos
1212 here for simplicity. */
1213 if (code == SUBREG && GET_CODE (SUBREG_REG (x)) == REG
1214 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1216 unsigned int i = REGNO (SUBREG_REG (x));
1218 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1220 /* If reg_tick has been incremented more than once since
1221 reg_in_table was last set, that means that the entire
1222 register has been set before, so discard anything memorized
1223 for the entire register, including all SUBREG expressions. */
1224 if (REG_IN_TABLE (i) != REG_TICK (i) - 1)
1225 remove_invalid_refs (i);
1226 else
1227 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1230 REG_IN_TABLE (i) = REG_TICK (i);
1231 return 0;
1234 /* If X is a comparison or a COMPARE and either operand is a register
1235 that does not have a quantity, give it one. This is so that a later
1236 call to record_jump_equiv won't cause X to be assigned a different
1237 hash code and not found in the table after that call.
1239 It is not necessary to do this here, since rehash_using_reg can
1240 fix up the table later, but doing this here eliminates the need to
1241 call that expensive function in the most common case where the only
1242 use of the register is in the comparison. */
1244 if (code == COMPARE || GET_RTX_CLASS (code) == '<')
1246 if (GET_CODE (XEXP (x, 0)) == REG
1247 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1248 if (insert_regs (XEXP (x, 0), NULL_PTR, 0))
1250 rehash_using_reg (XEXP (x, 0));
1251 changed = 1;
1254 if (GET_CODE (XEXP (x, 1)) == REG
1255 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1256 if (insert_regs (XEXP (x, 1), NULL_PTR, 0))
1258 rehash_using_reg (XEXP (x, 1));
1259 changed = 1;
1263 fmt = GET_RTX_FORMAT (code);
1264 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1265 if (fmt[i] == 'e')
1266 changed |= mention_regs (XEXP (x, i));
1267 else if (fmt[i] == 'E')
1268 for (j = 0; j < XVECLEN (x, i); j++)
1269 changed |= mention_regs (XVECEXP (x, i, j));
1271 return changed;
1274 /* Update the register quantities for inserting X into the hash table
1275 with a value equivalent to CLASSP.
1276 (If the class does not contain a REG, it is irrelevant.)
1277 If MODIFIED is nonzero, X is a destination; it is being modified.
1278 Note that delete_reg_equiv should be called on a register
1279 before insert_regs is done on that register with MODIFIED != 0.
1281 Nonzero value means that elements of reg_qty have changed
1282 so X's hash code may be different. */
1284 static int
1285 insert_regs (x, classp, modified)
1286 rtx x;
1287 struct table_elt *classp;
1288 int modified;
1290 if (GET_CODE (x) == REG)
1292 unsigned int regno = REGNO (x);
1293 int qty_valid;
1295 /* If REGNO is in the equivalence table already but is of the
1296 wrong mode for that equivalence, don't do anything here. */
1298 qty_valid = REGNO_QTY_VALID_P (regno);
1299 if (qty_valid)
1301 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1303 if (ent->mode != GET_MODE (x))
1304 return 0;
1307 if (modified || ! qty_valid)
1309 if (classp)
1310 for (classp = classp->first_same_value;
1311 classp != 0;
1312 classp = classp->next_same_value)
1313 if (GET_CODE (classp->exp) == REG
1314 && GET_MODE (classp->exp) == GET_MODE (x))
1316 make_regs_eqv (regno, REGNO (classp->exp));
1317 return 1;
1320 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1321 than REG_IN_TABLE to find out if there was only a single preceding
1322 invalidation - for the SUBREG - or another one, which would be
1323 for the full register. However, if we find here that REG_TICK
1324 indicates that the register is invalid, it means that it has
1325 been invalidated in a separate operation. The SUBREG might be used
1326 now (then this is a recursive call), or we might use the full REG
1327 now and a SUBREG of it later. So bump up REG_TICK so that
1328 mention_regs will do the right thing. */
1329 if (! modified
1330 && REG_IN_TABLE (regno) >= 0
1331 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1332 REG_TICK (regno)++;
1333 make_new_qty (regno, GET_MODE (x));
1334 return 1;
1337 return 0;
1340 /* If X is a SUBREG, we will likely be inserting the inner register in the
1341 table. If that register doesn't have an assigned quantity number at
1342 this point but does later, the insertion that we will be doing now will
1343 not be accessible because its hash code will have changed. So assign
1344 a quantity number now. */
1346 else if (GET_CODE (x) == SUBREG && GET_CODE (SUBREG_REG (x)) == REG
1347 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1349 insert_regs (SUBREG_REG (x), NULL_PTR, 0);
1350 mention_regs (x);
1351 return 1;
1353 else
1354 return mention_regs (x);
1357 /* Look in or update the hash table. */
1359 /* Remove table element ELT from use in the table.
1360 HASH is its hash code, made using the HASH macro.
1361 It's an argument because often that is known in advance
1362 and we save much time not recomputing it. */
1364 static void
1365 remove_from_table (elt, hash)
1366 register struct table_elt *elt;
1367 unsigned hash;
1369 if (elt == 0)
1370 return;
1372 /* Mark this element as removed. See cse_insn. */
1373 elt->first_same_value = 0;
1375 /* Remove the table element from its equivalence class. */
1378 register struct table_elt *prev = elt->prev_same_value;
1379 register struct table_elt *next = elt->next_same_value;
1381 if (next)
1382 next->prev_same_value = prev;
1384 if (prev)
1385 prev->next_same_value = next;
1386 else
1388 register struct table_elt *newfirst = next;
1389 while (next)
1391 next->first_same_value = newfirst;
1392 next = next->next_same_value;
1397 /* Remove the table element from its hash bucket. */
1400 register struct table_elt *prev = elt->prev_same_hash;
1401 register struct table_elt *next = elt->next_same_hash;
1403 if (next)
1404 next->prev_same_hash = prev;
1406 if (prev)
1407 prev->next_same_hash = next;
1408 else if (table[hash] == elt)
1409 table[hash] = next;
1410 else
1412 /* This entry is not in the proper hash bucket. This can happen
1413 when two classes were merged by `merge_equiv_classes'. Search
1414 for the hash bucket that it heads. This happens only very
1415 rarely, so the cost is acceptable. */
1416 for (hash = 0; hash < HASH_SIZE; hash++)
1417 if (table[hash] == elt)
1418 table[hash] = next;
1422 /* Remove the table element from its related-value circular chain. */
1424 if (elt->related_value != 0 && elt->related_value != elt)
1426 register struct table_elt *p = elt->related_value;
1428 while (p->related_value != elt)
1429 p = p->related_value;
1430 p->related_value = elt->related_value;
1431 if (p->related_value == p)
1432 p->related_value = 0;
1435 /* Now add it to the free element chain. */
1436 elt->next_same_hash = free_element_chain;
1437 free_element_chain = elt;
1440 /* Look up X in the hash table and return its table element,
1441 or 0 if X is not in the table.
1443 MODE is the machine-mode of X, or if X is an integer constant
1444 with VOIDmode then MODE is the mode with which X will be used.
1446 Here we are satisfied to find an expression whose tree structure
1447 looks like X. */
1449 static struct table_elt *
1450 lookup (x, hash, mode)
1451 rtx x;
1452 unsigned hash;
1453 enum machine_mode mode;
1455 register struct table_elt *p;
1457 for (p = table[hash]; p; p = p->next_same_hash)
1458 if (mode == p->mode && ((x == p->exp && GET_CODE (x) == REG)
1459 || exp_equiv_p (x, p->exp, GET_CODE (x) != REG, 0)))
1460 return p;
1462 return 0;
1465 /* Like `lookup' but don't care whether the table element uses invalid regs.
1466 Also ignore discrepancies in the machine mode of a register. */
1468 static struct table_elt *
1469 lookup_for_remove (x, hash, mode)
1470 rtx x;
1471 unsigned hash;
1472 enum machine_mode mode;
1474 register struct table_elt *p;
1476 if (GET_CODE (x) == REG)
1478 unsigned int regno = REGNO (x);
1480 /* Don't check the machine mode when comparing registers;
1481 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1482 for (p = table[hash]; p; p = p->next_same_hash)
1483 if (GET_CODE (p->exp) == REG
1484 && REGNO (p->exp) == regno)
1485 return p;
1487 else
1489 for (p = table[hash]; p; p = p->next_same_hash)
1490 if (mode == p->mode && (x == p->exp || exp_equiv_p (x, p->exp, 0, 0)))
1491 return p;
1494 return 0;
1497 /* Look for an expression equivalent to X and with code CODE.
1498 If one is found, return that expression. */
1500 static rtx
1501 lookup_as_function (x, code)
1502 rtx x;
1503 enum rtx_code code;
1505 register struct table_elt *p
1506 = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, GET_MODE (x));
1508 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1509 long as we are narrowing. So if we looked in vain for a mode narrower
1510 than word_mode before, look for word_mode now. */
1511 if (p == 0 && code == CONST_INT
1512 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1514 x = copy_rtx (x);
1515 PUT_MODE (x, word_mode);
1516 p = lookup (x, safe_hash (x, VOIDmode) & HASH_MASK, word_mode);
1519 if (p == 0)
1520 return 0;
1522 for (p = p->first_same_value; p; p = p->next_same_value)
1523 if (GET_CODE (p->exp) == code
1524 /* Make sure this is a valid entry in the table. */
1525 && exp_equiv_p (p->exp, p->exp, 1, 0))
1526 return p->exp;
1528 return 0;
1531 /* Insert X in the hash table, assuming HASH is its hash code
1532 and CLASSP is an element of the class it should go in
1533 (or 0 if a new class should be made).
1534 It is inserted at the proper position to keep the class in
1535 the order cheapest first.
1537 MODE is the machine-mode of X, or if X is an integer constant
1538 with VOIDmode then MODE is the mode with which X will be used.
1540 For elements of equal cheapness, the most recent one
1541 goes in front, except that the first element in the list
1542 remains first unless a cheaper element is added. The order of
1543 pseudo-registers does not matter, as canon_reg will be called to
1544 find the cheapest when a register is retrieved from the table.
1546 The in_memory field in the hash table element is set to 0.
1547 The caller must set it nonzero if appropriate.
1549 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1550 and if insert_regs returns a nonzero value
1551 you must then recompute its hash code before calling here.
1553 If necessary, update table showing constant values of quantities. */
1555 #define CHEAPER(X, Y) \
1556 (preferrable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1558 static struct table_elt *
1559 insert (x, classp, hash, mode)
1560 register rtx x;
1561 register struct table_elt *classp;
1562 unsigned hash;
1563 enum machine_mode mode;
1565 register struct table_elt *elt;
1567 /* If X is a register and we haven't made a quantity for it,
1568 something is wrong. */
1569 if (GET_CODE (x) == REG && ! REGNO_QTY_VALID_P (REGNO (x)))
1570 abort ();
1572 /* If X is a hard register, show it is being put in the table. */
1573 if (GET_CODE (x) == REG && REGNO (x) < FIRST_PSEUDO_REGISTER)
1575 unsigned int regno = REGNO (x);
1576 unsigned int endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (x));
1577 unsigned int i;
1579 for (i = regno; i < endregno; i++)
1580 SET_HARD_REG_BIT (hard_regs_in_table, i);
1583 /* If X is a label, show we recorded it. */
1584 if (GET_CODE (x) == LABEL_REF
1585 || (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS
1586 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF))
1587 new_label_ref = x;
1589 /* Put an element for X into the right hash bucket. */
1591 elt = free_element_chain;
1592 if (elt)
1593 free_element_chain = elt->next_same_hash;
1594 else
1596 n_elements_made++;
1597 elt = (struct table_elt *) xmalloc (sizeof (struct table_elt));
1600 elt->exp = x;
1601 elt->canon_exp = NULL_RTX;
1602 elt->cost = COST (x);
1603 elt->regcost = approx_reg_cost (x);
1604 elt->next_same_value = 0;
1605 elt->prev_same_value = 0;
1606 elt->next_same_hash = table[hash];
1607 elt->prev_same_hash = 0;
1608 elt->related_value = 0;
1609 elt->in_memory = 0;
1610 elt->mode = mode;
1611 elt->is_const = (CONSTANT_P (x)
1612 /* GNU C++ takes advantage of this for `this'
1613 (and other const values). */
1614 || (RTX_UNCHANGING_P (x)
1615 && GET_CODE (x) == REG
1616 && REGNO (x) >= FIRST_PSEUDO_REGISTER)
1617 || FIXED_BASE_PLUS_P (x));
1619 if (table[hash])
1620 table[hash]->prev_same_hash = elt;
1621 table[hash] = elt;
1623 /* Put it into the proper value-class. */
1624 if (classp)
1626 classp = classp->first_same_value;
1627 if (CHEAPER (elt, classp))
1628 /* Insert at the head of the class */
1630 register struct table_elt *p;
1631 elt->next_same_value = classp;
1632 classp->prev_same_value = elt;
1633 elt->first_same_value = elt;
1635 for (p = classp; p; p = p->next_same_value)
1636 p->first_same_value = elt;
1638 else
1640 /* Insert not at head of the class. */
1641 /* Put it after the last element cheaper than X. */
1642 register struct table_elt *p, *next;
1644 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1645 p = next);
1647 /* Put it after P and before NEXT. */
1648 elt->next_same_value = next;
1649 if (next)
1650 next->prev_same_value = elt;
1652 elt->prev_same_value = p;
1653 p->next_same_value = elt;
1654 elt->first_same_value = classp;
1657 else
1658 elt->first_same_value = elt;
1660 /* If this is a constant being set equivalent to a register or a register
1661 being set equivalent to a constant, note the constant equivalence.
1663 If this is a constant, it cannot be equivalent to a different constant,
1664 and a constant is the only thing that can be cheaper than a register. So
1665 we know the register is the head of the class (before the constant was
1666 inserted).
1668 If this is a register that is not already known equivalent to a
1669 constant, we must check the entire class.
1671 If this is a register that is already known equivalent to an insn,
1672 update the qtys `const_insn' to show that `this_insn' is the latest
1673 insn making that quantity equivalent to the constant. */
1675 if (elt->is_const && classp && GET_CODE (classp->exp) == REG
1676 && GET_CODE (x) != REG)
1678 int exp_q = REG_QTY (REGNO (classp->exp));
1679 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1681 exp_ent->const_rtx = gen_lowpart_if_possible (exp_ent->mode, x);
1682 exp_ent->const_insn = this_insn;
1685 else if (GET_CODE (x) == REG
1686 && classp
1687 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1688 && ! elt->is_const)
1690 register struct table_elt *p;
1692 for (p = classp; p != 0; p = p->next_same_value)
1694 if (p->is_const && GET_CODE (p->exp) != REG)
1696 int x_q = REG_QTY (REGNO (x));
1697 struct qty_table_elem *x_ent = &qty_table[x_q];
1699 x_ent->const_rtx
1700 = gen_lowpart_if_possible (GET_MODE (x), p->exp);
1701 x_ent->const_insn = this_insn;
1702 break;
1707 else if (GET_CODE (x) == REG
1708 && qty_table[REG_QTY (REGNO (x))].const_rtx
1709 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1710 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1712 /* If this is a constant with symbolic value,
1713 and it has a term with an explicit integer value,
1714 link it up with related expressions. */
1715 if (GET_CODE (x) == CONST)
1717 rtx subexp = get_related_value (x);
1718 unsigned subhash;
1719 struct table_elt *subelt, *subelt_prev;
1721 if (subexp != 0)
1723 /* Get the integer-free subexpression in the hash table. */
1724 subhash = safe_hash (subexp, mode) & HASH_MASK;
1725 subelt = lookup (subexp, subhash, mode);
1726 if (subelt == 0)
1727 subelt = insert (subexp, NULL_PTR, subhash, mode);
1728 /* Initialize SUBELT's circular chain if it has none. */
1729 if (subelt->related_value == 0)
1730 subelt->related_value = subelt;
1731 /* Find the element in the circular chain that precedes SUBELT. */
1732 subelt_prev = subelt;
1733 while (subelt_prev->related_value != subelt)
1734 subelt_prev = subelt_prev->related_value;
1735 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1736 This way the element that follows SUBELT is the oldest one. */
1737 elt->related_value = subelt_prev->related_value;
1738 subelt_prev->related_value = elt;
1742 return elt;
1745 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1746 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1747 the two classes equivalent.
1749 CLASS1 will be the surviving class; CLASS2 should not be used after this
1750 call.
1752 Any invalid entries in CLASS2 will not be copied. */
1754 static void
1755 merge_equiv_classes (class1, class2)
1756 struct table_elt *class1, *class2;
1758 struct table_elt *elt, *next, *new;
1760 /* Ensure we start with the head of the classes. */
1761 class1 = class1->first_same_value;
1762 class2 = class2->first_same_value;
1764 /* If they were already equal, forget it. */
1765 if (class1 == class2)
1766 return;
1768 for (elt = class2; elt; elt = next)
1770 unsigned int hash;
1771 rtx exp = elt->exp;
1772 enum machine_mode mode = elt->mode;
1774 next = elt->next_same_value;
1776 /* Remove old entry, make a new one in CLASS1's class.
1777 Don't do this for invalid entries as we cannot find their
1778 hash code (it also isn't necessary). */
1779 if (GET_CODE (exp) == REG || exp_equiv_p (exp, exp, 1, 0))
1781 hash_arg_in_memory = 0;
1782 hash = HASH (exp, mode);
1784 if (GET_CODE (exp) == REG)
1785 delete_reg_equiv (REGNO (exp));
1787 remove_from_table (elt, hash);
1789 if (insert_regs (exp, class1, 0))
1791 rehash_using_reg (exp);
1792 hash = HASH (exp, mode);
1794 new = insert (exp, class1, hash, mode);
1795 new->in_memory = hash_arg_in_memory;
1800 /* Flush the entire hash table. */
1802 static void
1803 flush_hash_table ()
1805 int i;
1806 struct table_elt *p;
1808 for (i = 0; i < HASH_SIZE; i++)
1809 for (p = table[i]; p; p = table[i])
1811 /* Note that invalidate can remove elements
1812 after P in the current hash chain. */
1813 if (GET_CODE (p->exp) == REG)
1814 invalidate (p->exp, p->mode);
1815 else
1816 remove_from_table (p, i);
1820 /* Function called for each rtx to check whether true dependence exist. */
1821 struct check_dependence_data
1823 enum machine_mode mode;
1824 rtx exp;
1826 static int
1827 check_dependence (x, data)
1828 rtx *x;
1829 void *data;
1831 struct check_dependence_data *d = (struct check_dependence_data *) data;
1832 if (*x && GET_CODE (*x) == MEM)
1833 return true_dependence (d->exp, d->mode, *x, cse_rtx_varies_p);
1834 else
1835 return 0;
1838 /* Remove from the hash table, or mark as invalid, all expressions whose
1839 values could be altered by storing in X. X is a register, a subreg, or
1840 a memory reference with nonvarying address (because, when a memory
1841 reference with a varying address is stored in, all memory references are
1842 removed by invalidate_memory so specific invalidation is superfluous).
1843 FULL_MODE, if not VOIDmode, indicates that this much should be
1844 invalidated instead of just the amount indicated by the mode of X. This
1845 is only used for bitfield stores into memory.
1847 A nonvarying address may be just a register or just a symbol reference,
1848 or it may be either of those plus a numeric offset. */
1850 static void
1851 invalidate (x, full_mode)
1852 rtx x;
1853 enum machine_mode full_mode;
1855 register int i;
1856 register struct table_elt *p;
1858 switch (GET_CODE (x))
1860 case REG:
1862 /* If X is a register, dependencies on its contents are recorded
1863 through the qty number mechanism. Just change the qty number of
1864 the register, mark it as invalid for expressions that refer to it,
1865 and remove it itself. */
1866 unsigned int regno = REGNO (x);
1867 unsigned int hash = HASH (x, GET_MODE (x));
1869 /* Remove REGNO from any quantity list it might be on and indicate
1870 that its value might have changed. If it is a pseudo, remove its
1871 entry from the hash table.
1873 For a hard register, we do the first two actions above for any
1874 additional hard registers corresponding to X. Then, if any of these
1875 registers are in the table, we must remove any REG entries that
1876 overlap these registers. */
1878 delete_reg_equiv (regno);
1879 REG_TICK (regno)++;
1881 if (regno >= FIRST_PSEUDO_REGISTER)
1883 /* Because a register can be referenced in more than one mode,
1884 we might have to remove more than one table entry. */
1885 struct table_elt *elt;
1887 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1888 remove_from_table (elt, hash);
1890 else
1892 HOST_WIDE_INT in_table
1893 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1894 unsigned int endregno
1895 = regno + HARD_REGNO_NREGS (regno, GET_MODE (x));
1896 unsigned int tregno, tendregno, rn;
1897 register struct table_elt *p, *next;
1899 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1901 for (rn = regno + 1; rn < endregno; rn++)
1903 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1904 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1905 delete_reg_equiv (rn);
1906 REG_TICK (rn)++;
1909 if (in_table)
1910 for (hash = 0; hash < HASH_SIZE; hash++)
1911 for (p = table[hash]; p; p = next)
1913 next = p->next_same_hash;
1915 if (GET_CODE (p->exp) != REG
1916 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1917 continue;
1919 tregno = REGNO (p->exp);
1920 tendregno
1921 = tregno + HARD_REGNO_NREGS (tregno, GET_MODE (p->exp));
1922 if (tendregno > regno && tregno < endregno)
1923 remove_from_table (p, hash);
1927 return;
1929 case SUBREG:
1930 invalidate (SUBREG_REG (x), VOIDmode);
1931 return;
1933 case PARALLEL:
1934 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1935 invalidate (XVECEXP (x, 0, i), VOIDmode);
1936 return;
1938 case EXPR_LIST:
1939 /* This is part of a disjoint return value; extract the location in
1940 question ignoring the offset. */
1941 invalidate (XEXP (x, 0), VOIDmode);
1942 return;
1944 case MEM:
1945 /* Calculate the canonical version of X here so that
1946 true_dependence doesn't generate new RTL for X on each call. */
1947 x = canon_rtx (x);
1949 /* Remove all hash table elements that refer to overlapping pieces of
1950 memory. */
1951 if (full_mode == VOIDmode)
1952 full_mode = GET_MODE (x);
1954 for (i = 0; i < HASH_SIZE; i++)
1956 register struct table_elt *next;
1958 for (p = table[i]; p; p = next)
1960 next = p->next_same_hash;
1961 if (p->in_memory)
1963 struct check_dependence_data d;
1965 /* Just canonicalize the expression once;
1966 otherwise each time we call invalidate
1967 true_dependence will canonicalize the
1968 expression again. */
1969 if (!p->canon_exp)
1970 p->canon_exp = canon_rtx (p->exp);
1971 d.exp = x;
1972 d.mode = full_mode;
1973 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1974 remove_from_table (p, i);
1978 return;
1980 default:
1981 abort ();
1985 /* Remove all expressions that refer to register REGNO,
1986 since they are already invalid, and we are about to
1987 mark that register valid again and don't want the old
1988 expressions to reappear as valid. */
1990 static void
1991 remove_invalid_refs (regno)
1992 unsigned int regno;
1994 unsigned int i;
1995 struct table_elt *p, *next;
1997 for (i = 0; i < HASH_SIZE; i++)
1998 for (p = table[i]; p; p = next)
2000 next = p->next_same_hash;
2001 if (GET_CODE (p->exp) != REG
2002 && refers_to_regno_p (regno, regno + 1, p->exp, NULL_PTR))
2003 remove_from_table (p, i);
2007 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2008 and mode MODE. */
2009 static void
2010 remove_invalid_subreg_refs (regno, offset, mode)
2011 unsigned int regno;
2012 unsigned int offset;
2013 enum machine_mode mode;
2015 unsigned int i;
2016 struct table_elt *p, *next;
2017 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2019 for (i = 0; i < HASH_SIZE; i++)
2020 for (p = table[i]; p; p = next)
2022 rtx exp = p->exp;
2023 next = p->next_same_hash;
2025 if (GET_CODE (exp) != REG
2026 && (GET_CODE (exp) != SUBREG
2027 || GET_CODE (SUBREG_REG (exp)) != REG
2028 || REGNO (SUBREG_REG (exp)) != regno
2029 || (((SUBREG_BYTE (exp)
2030 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2031 && SUBREG_BYTE (exp) <= end))
2032 && refers_to_regno_p (regno, regno + 1, p->exp, NULL_PTR))
2033 remove_from_table (p, i);
2037 /* Recompute the hash codes of any valid entries in the hash table that
2038 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2040 This is called when we make a jump equivalence. */
2042 static void
2043 rehash_using_reg (x)
2044 rtx x;
2046 unsigned int i;
2047 struct table_elt *p, *next;
2048 unsigned hash;
2050 if (GET_CODE (x) == SUBREG)
2051 x = SUBREG_REG (x);
2053 /* If X is not a register or if the register is known not to be in any
2054 valid entries in the table, we have no work to do. */
2056 if (GET_CODE (x) != REG
2057 || REG_IN_TABLE (REGNO (x)) < 0
2058 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2059 return;
2061 /* Scan all hash chains looking for valid entries that mention X.
2062 If we find one and it is in the wrong hash chain, move it. We can skip
2063 objects that are registers, since they are handled specially. */
2065 for (i = 0; i < HASH_SIZE; i++)
2066 for (p = table[i]; p; p = next)
2068 next = p->next_same_hash;
2069 if (GET_CODE (p->exp) != REG && reg_mentioned_p (x, p->exp)
2070 && exp_equiv_p (p->exp, p->exp, 1, 0)
2071 && i != (hash = safe_hash (p->exp, p->mode) & HASH_MASK))
2073 if (p->next_same_hash)
2074 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2076 if (p->prev_same_hash)
2077 p->prev_same_hash->next_same_hash = p->next_same_hash;
2078 else
2079 table[i] = p->next_same_hash;
2081 p->next_same_hash = table[hash];
2082 p->prev_same_hash = 0;
2083 if (table[hash])
2084 table[hash]->prev_same_hash = p;
2085 table[hash] = p;
2090 /* Remove from the hash table any expression that is a call-clobbered
2091 register. Also update their TICK values. */
2093 static void
2094 invalidate_for_call ()
2096 unsigned int regno, endregno;
2097 unsigned int i;
2098 unsigned hash;
2099 struct table_elt *p, *next;
2100 int in_table = 0;
2102 /* Go through all the hard registers. For each that is clobbered in
2103 a CALL_INSN, remove the register from quantity chains and update
2104 reg_tick if defined. Also see if any of these registers is currently
2105 in the table. */
2107 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2108 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2110 delete_reg_equiv (regno);
2111 if (REG_TICK (regno) >= 0)
2112 REG_TICK (regno)++;
2114 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2117 /* In the case where we have no call-clobbered hard registers in the
2118 table, we are done. Otherwise, scan the table and remove any
2119 entry that overlaps a call-clobbered register. */
2121 if (in_table)
2122 for (hash = 0; hash < HASH_SIZE; hash++)
2123 for (p = table[hash]; p; p = next)
2125 next = p->next_same_hash;
2127 if (GET_CODE (p->exp) != REG
2128 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2129 continue;
2131 regno = REGNO (p->exp);
2132 endregno = regno + HARD_REGNO_NREGS (regno, GET_MODE (p->exp));
2134 for (i = regno; i < endregno; i++)
2135 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2137 remove_from_table (p, hash);
2138 break;
2143 /* Given an expression X of type CONST,
2144 and ELT which is its table entry (or 0 if it
2145 is not in the hash table),
2146 return an alternate expression for X as a register plus integer.
2147 If none can be found, return 0. */
2149 static rtx
2150 use_related_value (x, elt)
2151 rtx x;
2152 struct table_elt *elt;
2154 register struct table_elt *relt = 0;
2155 register struct table_elt *p, *q;
2156 HOST_WIDE_INT offset;
2158 /* First, is there anything related known?
2159 If we have a table element, we can tell from that.
2160 Otherwise, must look it up. */
2162 if (elt != 0 && elt->related_value != 0)
2163 relt = elt;
2164 else if (elt == 0 && GET_CODE (x) == CONST)
2166 rtx subexp = get_related_value (x);
2167 if (subexp != 0)
2168 relt = lookup (subexp,
2169 safe_hash (subexp, GET_MODE (subexp)) & HASH_MASK,
2170 GET_MODE (subexp));
2173 if (relt == 0)
2174 return 0;
2176 /* Search all related table entries for one that has an
2177 equivalent register. */
2179 p = relt;
2180 while (1)
2182 /* This loop is strange in that it is executed in two different cases.
2183 The first is when X is already in the table. Then it is searching
2184 the RELATED_VALUE list of X's class (RELT). The second case is when
2185 X is not in the table. Then RELT points to a class for the related
2186 value.
2188 Ensure that, whatever case we are in, that we ignore classes that have
2189 the same value as X. */
2191 if (rtx_equal_p (x, p->exp))
2192 q = 0;
2193 else
2194 for (q = p->first_same_value; q; q = q->next_same_value)
2195 if (GET_CODE (q->exp) == REG)
2196 break;
2198 if (q)
2199 break;
2201 p = p->related_value;
2203 /* We went all the way around, so there is nothing to be found.
2204 Alternatively, perhaps RELT was in the table for some other reason
2205 and it has no related values recorded. */
2206 if (p == relt || p == 0)
2207 break;
2210 if (q == 0)
2211 return 0;
2213 offset = (get_integer_term (x) - get_integer_term (p->exp));
2214 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2215 return plus_constant (q->exp, offset);
2218 /* Hash a string. Just add its bytes up. */
2219 static inline unsigned
2220 canon_hash_string (ps)
2221 const char *ps;
2223 unsigned hash = 0;
2224 const unsigned char *p = (const unsigned char *)ps;
2226 if (p)
2227 while (*p)
2228 hash += *p++;
2230 return hash;
2233 /* Hash an rtx. We are careful to make sure the value is never negative.
2234 Equivalent registers hash identically.
2235 MODE is used in hashing for CONST_INTs only;
2236 otherwise the mode of X is used.
2238 Store 1 in do_not_record if any subexpression is volatile.
2240 Store 1 in hash_arg_in_memory if X contains a MEM rtx
2241 which does not have the RTX_UNCHANGING_P bit set.
2243 Note that cse_insn knows that the hash code of a MEM expression
2244 is just (int) MEM plus the hash code of the address. */
2246 static unsigned
2247 canon_hash (x, mode)
2248 rtx x;
2249 enum machine_mode mode;
2251 register int i, j;
2252 register unsigned hash = 0;
2253 register enum rtx_code code;
2254 register const char *fmt;
2256 /* repeat is used to turn tail-recursion into iteration. */
2257 repeat:
2258 if (x == 0)
2259 return hash;
2261 code = GET_CODE (x);
2262 switch (code)
2264 case REG:
2266 unsigned int regno = REGNO (x);
2268 /* On some machines, we can't record any non-fixed hard register,
2269 because extending its life will cause reload problems. We
2270 consider ap, fp, and sp to be fixed for this purpose.
2272 We also consider CCmode registers to be fixed for this purpose;
2273 failure to do so leads to failure to simplify 0<100 type of
2274 conditionals.
2276 On all machines, we can't record any global registers. */
2278 if (regno < FIRST_PSEUDO_REGISTER
2279 && (global_regs[regno]
2280 || (SMALL_REGISTER_CLASSES
2281 && ! fixed_regs[regno]
2282 && regno != FRAME_POINTER_REGNUM
2283 && regno != HARD_FRAME_POINTER_REGNUM
2284 && regno != ARG_POINTER_REGNUM
2285 && regno != STACK_POINTER_REGNUM
2286 && GET_MODE_CLASS (GET_MODE (x)) != MODE_CC)))
2288 do_not_record = 1;
2289 return 0;
2292 hash += ((unsigned) REG << 7) + (unsigned) REG_QTY (regno);
2293 return hash;
2296 /* We handle SUBREG of a REG specially because the underlying
2297 reg changes its hash value with every value change; we don't
2298 want to have to forget unrelated subregs when one subreg changes. */
2299 case SUBREG:
2301 if (GET_CODE (SUBREG_REG (x)) == REG)
2303 hash += (((unsigned) SUBREG << 7)
2304 + REGNO (SUBREG_REG (x))
2305 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2306 return hash;
2308 break;
2311 case CONST_INT:
2313 unsigned HOST_WIDE_INT tem = INTVAL (x);
2314 hash += ((unsigned) CONST_INT << 7) + (unsigned) mode + tem;
2315 return hash;
2318 case CONST_DOUBLE:
2319 /* This is like the general case, except that it only counts
2320 the integers representing the constant. */
2321 hash += (unsigned) code + (unsigned) GET_MODE (x);
2322 if (GET_MODE (x) != VOIDmode)
2323 for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
2325 unsigned HOST_WIDE_INT tem = XWINT (x, i);
2326 hash += tem;
2328 else
2329 hash += ((unsigned) CONST_DOUBLE_LOW (x)
2330 + (unsigned) CONST_DOUBLE_HIGH (x));
2331 return hash;
2333 /* Assume there is only one rtx object for any given label. */
2334 case LABEL_REF:
2335 hash += ((unsigned) LABEL_REF << 7) + (unsigned long) XEXP (x, 0);
2336 return hash;
2338 case SYMBOL_REF:
2339 hash += ((unsigned) SYMBOL_REF << 7) + (unsigned long) XSTR (x, 0);
2340 return hash;
2342 case MEM:
2343 /* We don't record if marked volatile or if BLKmode since we don't
2344 know the size of the move. */
2345 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2347 do_not_record = 1;
2348 return 0;
2350 if (! RTX_UNCHANGING_P (x) || FIXED_BASE_PLUS_P (XEXP (x, 0)))
2352 hash_arg_in_memory = 1;
2354 /* Now that we have already found this special case,
2355 might as well speed it up as much as possible. */
2356 hash += (unsigned) MEM;
2357 x = XEXP (x, 0);
2358 goto repeat;
2360 case USE:
2361 /* A USE that mentions non-volatile memory needs special
2362 handling since the MEM may be BLKmode which normally
2363 prevents an entry from being made. Pure calls are
2364 marked by a USE which mentions BLKmode memory. */
2365 if (GET_CODE (XEXP (x, 0)) == MEM
2366 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2368 hash += (unsigned)USE;
2369 x = XEXP (x, 0);
2371 if (! RTX_UNCHANGING_P (x) || FIXED_BASE_PLUS_P (XEXP (x, 0)))
2372 hash_arg_in_memory = 1;
2374 /* Now that we have already found this special case,
2375 might as well speed it up as much as possible. */
2376 hash += (unsigned) MEM;
2377 x = XEXP (x, 0);
2378 goto repeat;
2380 break;
2382 case PRE_DEC:
2383 case PRE_INC:
2384 case POST_DEC:
2385 case POST_INC:
2386 case PRE_MODIFY:
2387 case POST_MODIFY:
2388 case PC:
2389 case CC0:
2390 case CALL:
2391 case UNSPEC_VOLATILE:
2392 do_not_record = 1;
2393 return 0;
2395 case ASM_OPERANDS:
2396 if (MEM_VOLATILE_P (x))
2398 do_not_record = 1;
2399 return 0;
2401 else
2403 /* We don't want to take the filename and line into account. */
2404 hash += (unsigned) code + (unsigned) GET_MODE (x)
2405 + canon_hash_string (ASM_OPERANDS_TEMPLATE (x))
2406 + canon_hash_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2407 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2409 if (ASM_OPERANDS_INPUT_LENGTH (x))
2411 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2413 hash += (canon_hash (ASM_OPERANDS_INPUT (x, i),
2414 GET_MODE (ASM_OPERANDS_INPUT (x, i)))
2415 + canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT
2416 (x, i)));
2419 hash += canon_hash_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2420 x = ASM_OPERANDS_INPUT (x, 0);
2421 mode = GET_MODE (x);
2422 goto repeat;
2425 return hash;
2427 break;
2429 default:
2430 break;
2433 i = GET_RTX_LENGTH (code) - 1;
2434 hash += (unsigned) code + (unsigned) GET_MODE (x);
2435 fmt = GET_RTX_FORMAT (code);
2436 for (; i >= 0; i--)
2438 if (fmt[i] == 'e')
2440 rtx tem = XEXP (x, i);
2442 /* If we are about to do the last recursive call
2443 needed at this level, change it into iteration.
2444 This function is called enough to be worth it. */
2445 if (i == 0)
2447 x = tem;
2448 goto repeat;
2450 hash += canon_hash (tem, 0);
2452 else if (fmt[i] == 'E')
2453 for (j = 0; j < XVECLEN (x, i); j++)
2454 hash += canon_hash (XVECEXP (x, i, j), 0);
2455 else if (fmt[i] == 's')
2456 hash += canon_hash_string (XSTR (x, i));
2457 else if (fmt[i] == 'i')
2459 register unsigned tem = XINT (x, i);
2460 hash += tem;
2462 else if (fmt[i] == '0' || fmt[i] == 't')
2463 /* Unused. */
2465 else
2466 abort ();
2468 return hash;
2471 /* Like canon_hash but with no side effects. */
2473 static unsigned
2474 safe_hash (x, mode)
2475 rtx x;
2476 enum machine_mode mode;
2478 int save_do_not_record = do_not_record;
2479 int save_hash_arg_in_memory = hash_arg_in_memory;
2480 unsigned hash = canon_hash (x, mode);
2481 hash_arg_in_memory = save_hash_arg_in_memory;
2482 do_not_record = save_do_not_record;
2483 return hash;
2486 /* Return 1 iff X and Y would canonicalize into the same thing,
2487 without actually constructing the canonicalization of either one.
2488 If VALIDATE is nonzero,
2489 we assume X is an expression being processed from the rtl
2490 and Y was found in the hash table. We check register refs
2491 in Y for being marked as valid.
2493 If EQUAL_VALUES is nonzero, we allow a register to match a constant value
2494 that is known to be in the register. Ordinarily, we don't allow them
2495 to match, because letting them match would cause unpredictable results
2496 in all the places that search a hash table chain for an equivalent
2497 for a given value. A possible equivalent that has different structure
2498 has its hash code computed from different data. Whether the hash code
2499 is the same as that of the given value is pure luck. */
2501 static int
2502 exp_equiv_p (x, y, validate, equal_values)
2503 rtx x, y;
2504 int validate;
2505 int equal_values;
2507 register int i, j;
2508 register enum rtx_code code;
2509 register const char *fmt;
2511 /* Note: it is incorrect to assume an expression is equivalent to itself
2512 if VALIDATE is nonzero. */
2513 if (x == y && !validate)
2514 return 1;
2515 if (x == 0 || y == 0)
2516 return x == y;
2518 code = GET_CODE (x);
2519 if (code != GET_CODE (y))
2521 if (!equal_values)
2522 return 0;
2524 /* If X is a constant and Y is a register or vice versa, they may be
2525 equivalent. We only have to validate if Y is a register. */
2526 if (CONSTANT_P (x) && GET_CODE (y) == REG
2527 && REGNO_QTY_VALID_P (REGNO (y)))
2529 int y_q = REG_QTY (REGNO (y));
2530 struct qty_table_elem *y_ent = &qty_table[y_q];
2532 if (GET_MODE (y) == y_ent->mode
2533 && rtx_equal_p (x, y_ent->const_rtx)
2534 && (! validate || REG_IN_TABLE (REGNO (y)) == REG_TICK (REGNO (y))))
2535 return 1;
2538 if (CONSTANT_P (y) && code == REG
2539 && REGNO_QTY_VALID_P (REGNO (x)))
2541 int x_q = REG_QTY (REGNO (x));
2542 struct qty_table_elem *x_ent = &qty_table[x_q];
2544 if (GET_MODE (x) == x_ent->mode
2545 && rtx_equal_p (y, x_ent->const_rtx))
2546 return 1;
2549 return 0;
2552 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2553 if (GET_MODE (x) != GET_MODE (y))
2554 return 0;
2556 switch (code)
2558 case PC:
2559 case CC0:
2560 case CONST_INT:
2561 return x == y;
2563 case LABEL_REF:
2564 return XEXP (x, 0) == XEXP (y, 0);
2566 case SYMBOL_REF:
2567 return XSTR (x, 0) == XSTR (y, 0);
2569 case REG:
2571 unsigned int regno = REGNO (y);
2572 unsigned int endregno
2573 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2574 : HARD_REGNO_NREGS (regno, GET_MODE (y)));
2575 unsigned int i;
2577 /* If the quantities are not the same, the expressions are not
2578 equivalent. If there are and we are not to validate, they
2579 are equivalent. Otherwise, ensure all regs are up-to-date. */
2581 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2582 return 0;
2584 if (! validate)
2585 return 1;
2587 for (i = regno; i < endregno; i++)
2588 if (REG_IN_TABLE (i) != REG_TICK (i))
2589 return 0;
2591 return 1;
2594 /* For commutative operations, check both orders. */
2595 case PLUS:
2596 case MULT:
2597 case AND:
2598 case IOR:
2599 case XOR:
2600 case NE:
2601 case EQ:
2602 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0), validate, equal_values)
2603 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2604 validate, equal_values))
2605 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2606 validate, equal_values)
2607 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2608 validate, equal_values)));
2610 case ASM_OPERANDS:
2611 /* We don't use the generic code below because we want to
2612 disregard filename and line numbers. */
2614 /* A volatile asm isn't equivalent to any other. */
2615 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2616 return 0;
2618 if (GET_MODE (x) != GET_MODE (y)
2619 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2620 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2621 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2622 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2623 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2624 return 0;
2626 if (ASM_OPERANDS_INPUT_LENGTH (x))
2628 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2629 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2630 ASM_OPERANDS_INPUT (y, i),
2631 validate, equal_values)
2632 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2633 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2634 return 0;
2637 return 1;
2639 default:
2640 break;
2643 /* Compare the elements. If any pair of corresponding elements
2644 fail to match, return 0 for the whole things. */
2646 fmt = GET_RTX_FORMAT (code);
2647 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2649 switch (fmt[i])
2651 case 'e':
2652 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i), validate, equal_values))
2653 return 0;
2654 break;
2656 case 'E':
2657 if (XVECLEN (x, i) != XVECLEN (y, i))
2658 return 0;
2659 for (j = 0; j < XVECLEN (x, i); j++)
2660 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2661 validate, equal_values))
2662 return 0;
2663 break;
2665 case 's':
2666 if (strcmp (XSTR (x, i), XSTR (y, i)))
2667 return 0;
2668 break;
2670 case 'i':
2671 if (XINT (x, i) != XINT (y, i))
2672 return 0;
2673 break;
2675 case 'w':
2676 if (XWINT (x, i) != XWINT (y, i))
2677 return 0;
2678 break;
2680 case '0':
2681 case 't':
2682 break;
2684 default:
2685 abort ();
2689 return 1;
2692 /* Return 1 if X has a value that can vary even between two
2693 executions of the program. 0 means X can be compared reliably
2694 against certain constants or near-constants. */
2696 static int
2697 cse_rtx_varies_p (x, from_alias)
2698 register rtx x;
2699 int from_alias;
2701 /* We need not check for X and the equivalence class being of the same
2702 mode because if X is equivalent to a constant in some mode, it
2703 doesn't vary in any mode. */
2705 if (GET_CODE (x) == REG
2706 && REGNO_QTY_VALID_P (REGNO (x)))
2708 int x_q = REG_QTY (REGNO (x));
2709 struct qty_table_elem *x_ent = &qty_table[x_q];
2711 if (GET_MODE (x) == x_ent->mode
2712 && x_ent->const_rtx != NULL_RTX)
2713 return 0;
2716 if (GET_CODE (x) == PLUS
2717 && GET_CODE (XEXP (x, 1)) == CONST_INT
2718 && GET_CODE (XEXP (x, 0)) == REG
2719 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2721 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2722 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2724 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2725 && x0_ent->const_rtx != NULL_RTX)
2726 return 0;
2729 /* This can happen as the result of virtual register instantiation, if
2730 the initial constant is too large to be a valid address. This gives
2731 us a three instruction sequence, load large offset into a register,
2732 load fp minus a constant into a register, then a MEM which is the
2733 sum of the two `constant' registers. */
2734 if (GET_CODE (x) == PLUS
2735 && GET_CODE (XEXP (x, 0)) == REG
2736 && GET_CODE (XEXP (x, 1)) == REG
2737 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2738 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2740 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2741 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2742 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2743 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2745 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2746 && x0_ent->const_rtx != NULL_RTX
2747 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2748 && x1_ent->const_rtx != NULL_RTX)
2749 return 0;
2752 return rtx_varies_p (x, from_alias);
2755 /* Canonicalize an expression:
2756 replace each register reference inside it
2757 with the "oldest" equivalent register.
2759 If INSN is non-zero and we are replacing a pseudo with a hard register
2760 or vice versa, validate_change is used to ensure that INSN remains valid
2761 after we make our substitution. The calls are made with IN_GROUP non-zero
2762 so apply_change_group must be called upon the outermost return from this
2763 function (unless INSN is zero). The result of apply_change_group can
2764 generally be discarded since the changes we are making are optional. */
2766 static rtx
2767 canon_reg (x, insn)
2768 rtx x;
2769 rtx insn;
2771 register int i;
2772 register enum rtx_code code;
2773 register const char *fmt;
2775 if (x == 0)
2776 return x;
2778 code = GET_CODE (x);
2779 switch (code)
2781 case PC:
2782 case CC0:
2783 case CONST:
2784 case CONST_INT:
2785 case CONST_DOUBLE:
2786 case SYMBOL_REF:
2787 case LABEL_REF:
2788 case ADDR_VEC:
2789 case ADDR_DIFF_VEC:
2790 return x;
2792 case REG:
2794 register int first;
2795 register int q;
2796 register struct qty_table_elem *ent;
2798 /* Never replace a hard reg, because hard regs can appear
2799 in more than one machine mode, and we must preserve the mode
2800 of each occurrence. Also, some hard regs appear in
2801 MEMs that are shared and mustn't be altered. Don't try to
2802 replace any reg that maps to a reg of class NO_REGS. */
2803 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2804 || ! REGNO_QTY_VALID_P (REGNO (x)))
2805 return x;
2807 q = REG_QTY (REGNO (x));
2808 ent = &qty_table[q];
2809 first = ent->first_reg;
2810 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2811 : REGNO_REG_CLASS (first) == NO_REGS ? x
2812 : gen_rtx_REG (ent->mode, first));
2815 default:
2816 break;
2819 fmt = GET_RTX_FORMAT (code);
2820 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2822 register int j;
2824 if (fmt[i] == 'e')
2826 rtx new = canon_reg (XEXP (x, i), insn);
2827 int insn_code;
2829 /* If replacing pseudo with hard reg or vice versa, ensure the
2830 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2831 if (insn != 0 && new != 0
2832 && GET_CODE (new) == REG && GET_CODE (XEXP (x, i)) == REG
2833 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2834 != (REGNO (XEXP (x, i)) < FIRST_PSEUDO_REGISTER))
2835 || (insn_code = recog_memoized (insn)) < 0
2836 || insn_data[insn_code].n_dups > 0))
2837 validate_change (insn, &XEXP (x, i), new, 1);
2838 else
2839 XEXP (x, i) = new;
2841 else if (fmt[i] == 'E')
2842 for (j = 0; j < XVECLEN (x, i); j++)
2843 XVECEXP (x, i, j) = canon_reg (XVECEXP (x, i, j), insn);
2846 return x;
2849 /* LOC is a location within INSN that is an operand address (the contents of
2850 a MEM). Find the best equivalent address to use that is valid for this
2851 insn.
2853 On most CISC machines, complicated address modes are costly, and rtx_cost
2854 is a good approximation for that cost. However, most RISC machines have
2855 only a few (usually only one) memory reference formats. If an address is
2856 valid at all, it is often just as cheap as any other address. Hence, for
2857 RISC machines, we use the configuration macro `ADDRESS_COST' to compare the
2858 costs of various addresses. For two addresses of equal cost, choose the one
2859 with the highest `rtx_cost' value as that has the potential of eliminating
2860 the most insns. For equal costs, we choose the first in the equivalence
2861 class. Note that we ignore the fact that pseudo registers are cheaper
2862 than hard registers here because we would also prefer the pseudo registers.
2865 static void
2866 find_best_addr (insn, loc, mode)
2867 rtx insn;
2868 rtx *loc;
2869 enum machine_mode mode;
2871 struct table_elt *elt;
2872 rtx addr = *loc;
2873 #ifdef ADDRESS_COST
2874 struct table_elt *p;
2875 int found_better = 1;
2876 #endif
2877 int save_do_not_record = do_not_record;
2878 int save_hash_arg_in_memory = hash_arg_in_memory;
2879 int addr_volatile;
2880 int regno;
2881 unsigned hash;
2883 /* Do not try to replace constant addresses or addresses of local and
2884 argument slots. These MEM expressions are made only once and inserted
2885 in many instructions, as well as being used to control symbol table
2886 output. It is not safe to clobber them.
2888 There are some uncommon cases where the address is already in a register
2889 for some reason, but we cannot take advantage of that because we have
2890 no easy way to unshare the MEM. In addition, looking up all stack
2891 addresses is costly. */
2892 if ((GET_CODE (addr) == PLUS
2893 && GET_CODE (XEXP (addr, 0)) == REG
2894 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2895 && (regno = REGNO (XEXP (addr, 0)),
2896 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2897 || regno == ARG_POINTER_REGNUM))
2898 || (GET_CODE (addr) == REG
2899 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2900 || regno == HARD_FRAME_POINTER_REGNUM
2901 || regno == ARG_POINTER_REGNUM))
2902 || GET_CODE (addr) == ADDRESSOF
2903 || CONSTANT_ADDRESS_P (addr))
2904 return;
2906 /* If this address is not simply a register, try to fold it. This will
2907 sometimes simplify the expression. Many simplifications
2908 will not be valid, but some, usually applying the associative rule, will
2909 be valid and produce better code. */
2910 if (GET_CODE (addr) != REG)
2912 rtx folded = fold_rtx (copy_rtx (addr), NULL_RTX);
2913 int addr_folded_cost = address_cost (folded, mode);
2914 int addr_cost = address_cost (addr, mode);
2916 if ((addr_folded_cost < addr_cost
2917 || (addr_folded_cost == addr_cost
2918 /* ??? The rtx_cost comparison is left over from an older
2919 version of this code. It is probably no longer helpful. */
2920 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2921 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2922 && validate_change (insn, loc, folded, 0))
2923 addr = folded;
2926 /* If this address is not in the hash table, we can't look for equivalences
2927 of the whole address. Also, ignore if volatile. */
2929 do_not_record = 0;
2930 hash = HASH (addr, Pmode);
2931 addr_volatile = do_not_record;
2932 do_not_record = save_do_not_record;
2933 hash_arg_in_memory = save_hash_arg_in_memory;
2935 if (addr_volatile)
2936 return;
2938 elt = lookup (addr, hash, Pmode);
2940 #ifndef ADDRESS_COST
2941 if (elt)
2943 int our_cost = elt->cost;
2945 /* Find the lowest cost below ours that works. */
2946 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
2947 if (elt->cost < our_cost
2948 && (GET_CODE (elt->exp) == REG
2949 || exp_equiv_p (elt->exp, elt->exp, 1, 0))
2950 && validate_change (insn, loc,
2951 canon_reg (copy_rtx (elt->exp), NULL_RTX), 0))
2952 return;
2954 #else
2956 if (elt)
2958 /* We need to find the best (under the criteria documented above) entry
2959 in the class that is valid. We use the `flag' field to indicate
2960 choices that were invalid and iterate until we can't find a better
2961 one that hasn't already been tried. */
2963 for (p = elt->first_same_value; p; p = p->next_same_value)
2964 p->flag = 0;
2966 while (found_better)
2968 int best_addr_cost = address_cost (*loc, mode);
2969 int best_rtx_cost = (elt->cost + 1) >> 1;
2970 int exp_cost;
2971 struct table_elt *best_elt = elt;
2973 found_better = 0;
2974 for (p = elt->first_same_value; p; p = p->next_same_value)
2975 if (! p->flag)
2977 if ((GET_CODE (p->exp) == REG
2978 || exp_equiv_p (p->exp, p->exp, 1, 0))
2979 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2980 || (exp_cost == best_addr_cost
2981 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2983 found_better = 1;
2984 best_addr_cost = exp_cost;
2985 best_rtx_cost = (p->cost + 1) >> 1;
2986 best_elt = p;
2990 if (found_better)
2992 if (validate_change (insn, loc,
2993 canon_reg (copy_rtx (best_elt->exp),
2994 NULL_RTX), 0))
2995 return;
2996 else
2997 best_elt->flag = 1;
3002 /* If the address is a binary operation with the first operand a register
3003 and the second a constant, do the same as above, but looking for
3004 equivalences of the register. Then try to simplify before checking for
3005 the best address to use. This catches a few cases: First is when we
3006 have REG+const and the register is another REG+const. We can often merge
3007 the constants and eliminate one insn and one register. It may also be
3008 that a machine has a cheap REG+REG+const. Finally, this improves the
3009 code on the Alpha for unaligned byte stores. */
3011 if (flag_expensive_optimizations
3012 && (GET_RTX_CLASS (GET_CODE (*loc)) == '2'
3013 || GET_RTX_CLASS (GET_CODE (*loc)) == 'c')
3014 && GET_CODE (XEXP (*loc, 0)) == REG
3015 && GET_CODE (XEXP (*loc, 1)) == CONST_INT)
3017 rtx c = XEXP (*loc, 1);
3019 do_not_record = 0;
3020 hash = HASH (XEXP (*loc, 0), Pmode);
3021 do_not_record = save_do_not_record;
3022 hash_arg_in_memory = save_hash_arg_in_memory;
3024 elt = lookup (XEXP (*loc, 0), hash, Pmode);
3025 if (elt == 0)
3026 return;
3028 /* We need to find the best (under the criteria documented above) entry
3029 in the class that is valid. We use the `flag' field to indicate
3030 choices that were invalid and iterate until we can't find a better
3031 one that hasn't already been tried. */
3033 for (p = elt->first_same_value; p; p = p->next_same_value)
3034 p->flag = 0;
3036 while (found_better)
3038 int best_addr_cost = address_cost (*loc, mode);
3039 int best_rtx_cost = (COST (*loc) + 1) >> 1;
3040 struct table_elt *best_elt = elt;
3041 rtx best_rtx = *loc;
3042 int count;
3044 /* This is at worst case an O(n^2) algorithm, so limit our search
3045 to the first 32 elements on the list. This avoids trouble
3046 compiling code with very long basic blocks that can easily
3047 call simplify_gen_binary so many times that we run out of
3048 memory. */
3050 found_better = 0;
3051 for (p = elt->first_same_value, count = 0;
3052 p && count < 32;
3053 p = p->next_same_value, count++)
3054 if (! p->flag
3055 && (GET_CODE (p->exp) == REG
3056 || exp_equiv_p (p->exp, p->exp, 1, 0)))
3058 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3059 p->exp, c);
3060 int new_cost;
3061 new_cost = address_cost (new, mode);
3063 if (new_cost < best_addr_cost
3064 || (new_cost == best_addr_cost
3065 && (COST (new) + 1) >> 1 > best_rtx_cost))
3067 found_better = 1;
3068 best_addr_cost = new_cost;
3069 best_rtx_cost = (COST (new) + 1) >> 1;
3070 best_elt = p;
3071 best_rtx = new;
3075 if (found_better)
3077 if (validate_change (insn, loc,
3078 canon_reg (copy_rtx (best_rtx),
3079 NULL_RTX), 0))
3080 return;
3081 else
3082 best_elt->flag = 1;
3086 #endif
3089 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3090 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3091 what values are being compared.
3093 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3094 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3095 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3096 compared to produce cc0.
3098 The return value is the comparison operator and is either the code of
3099 A or the code corresponding to the inverse of the comparison. */
3101 static enum rtx_code
3102 find_comparison_args (code, parg1, parg2, pmode1, pmode2)
3103 enum rtx_code code;
3104 rtx *parg1, *parg2;
3105 enum machine_mode *pmode1, *pmode2;
3107 rtx arg1, arg2;
3109 arg1 = *parg1, arg2 = *parg2;
3111 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3113 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3115 /* Set non-zero when we find something of interest. */
3116 rtx x = 0;
3117 int reverse_code = 0;
3118 struct table_elt *p = 0;
3120 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3121 On machines with CC0, this is the only case that can occur, since
3122 fold_rtx will return the COMPARE or item being compared with zero
3123 when given CC0. */
3125 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3126 x = arg1;
3128 /* If ARG1 is a comparison operator and CODE is testing for
3129 STORE_FLAG_VALUE, get the inner arguments. */
3131 else if (GET_RTX_CLASS (GET_CODE (arg1)) == '<')
3133 if (code == NE
3134 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3135 && code == LT && STORE_FLAG_VALUE == -1)
3136 #ifdef FLOAT_STORE_FLAG_VALUE
3137 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3138 && (REAL_VALUE_NEGATIVE
3139 (FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)))))
3140 #endif
3142 x = arg1;
3143 else if (code == EQ
3144 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3145 && code == GE && STORE_FLAG_VALUE == -1)
3146 #ifdef FLOAT_STORE_FLAG_VALUE
3147 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3148 && (REAL_VALUE_NEGATIVE
3149 (FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)))))
3150 #endif
3152 x = arg1, reverse_code = 1;
3155 /* ??? We could also check for
3157 (ne (and (eq (...) (const_int 1))) (const_int 0))
3159 and related forms, but let's wait until we see them occurring. */
3161 if (x == 0)
3162 /* Look up ARG1 in the hash table and see if it has an equivalence
3163 that lets us see what is being compared. */
3164 p = lookup (arg1, safe_hash (arg1, GET_MODE (arg1)) & HASH_MASK,
3165 GET_MODE (arg1));
3166 if (p)
3168 p = p->first_same_value;
3170 /* If what we compare is already known to be constant, that is as
3171 good as it gets.
3172 We need to break the loop in this case, because otherwise we
3173 can have an infinite loop when looking at a reg that is known
3174 to be a constant which is the same as a comparison of a reg
3175 against zero which appears later in the insn stream, which in
3176 turn is constant and the same as the comparison of the first reg
3177 against zero... */
3178 if (p->is_const)
3179 break;
3182 for (; p; p = p->next_same_value)
3184 enum machine_mode inner_mode = GET_MODE (p->exp);
3186 /* If the entry isn't valid, skip it. */
3187 if (! exp_equiv_p (p->exp, p->exp, 1, 0))
3188 continue;
3190 if (GET_CODE (p->exp) == COMPARE
3191 /* Another possibility is that this machine has a compare insn
3192 that includes the comparison code. In that case, ARG1 would
3193 be equivalent to a comparison operation that would set ARG1 to
3194 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3195 ORIG_CODE is the actual comparison being done; if it is an EQ,
3196 we must reverse ORIG_CODE. On machine with a negative value
3197 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3198 || ((code == NE
3199 || (code == LT
3200 && GET_MODE_CLASS (inner_mode) == MODE_INT
3201 && (GET_MODE_BITSIZE (inner_mode)
3202 <= HOST_BITS_PER_WIDE_INT)
3203 && (STORE_FLAG_VALUE
3204 & ((HOST_WIDE_INT) 1
3205 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3206 #ifdef FLOAT_STORE_FLAG_VALUE
3207 || (code == LT
3208 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3209 && (REAL_VALUE_NEGATIVE
3210 (FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)))))
3211 #endif
3213 && GET_RTX_CLASS (GET_CODE (p->exp)) == '<'))
3215 x = p->exp;
3216 break;
3218 else if ((code == EQ
3219 || (code == GE
3220 && GET_MODE_CLASS (inner_mode) == MODE_INT
3221 && (GET_MODE_BITSIZE (inner_mode)
3222 <= HOST_BITS_PER_WIDE_INT)
3223 && (STORE_FLAG_VALUE
3224 & ((HOST_WIDE_INT) 1
3225 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3226 #ifdef FLOAT_STORE_FLAG_VALUE
3227 || (code == GE
3228 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3229 && (REAL_VALUE_NEGATIVE
3230 (FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)))))
3231 #endif
3233 && GET_RTX_CLASS (GET_CODE (p->exp)) == '<')
3235 reverse_code = 1;
3236 x = p->exp;
3237 break;
3240 /* If this is fp + constant, the equivalent is a better operand since
3241 it may let us predict the value of the comparison. */
3242 else if (NONZERO_BASE_PLUS_P (p->exp))
3244 arg1 = p->exp;
3245 continue;
3249 /* If we didn't find a useful equivalence for ARG1, we are done.
3250 Otherwise, set up for the next iteration. */
3251 if (x == 0)
3252 break;
3254 /* If we need to reverse the comparison, make sure that that is
3255 possible -- we can't necessarily infer the value of GE from LT
3256 with floating-point operands. */
3257 if (reverse_code)
3259 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3260 if (reversed == UNKNOWN)
3261 break;
3262 else code = reversed;
3264 else if (GET_RTX_CLASS (GET_CODE (x)) == '<')
3265 code = GET_CODE (x);
3266 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3269 /* Return our results. Return the modes from before fold_rtx
3270 because fold_rtx might produce const_int, and then it's too late. */
3271 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3272 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3274 return code;
3277 /* If X is a nontrivial arithmetic operation on an argument
3278 for which a constant value can be determined, return
3279 the result of operating on that value, as a constant.
3280 Otherwise, return X, possibly with one or more operands
3281 modified by recursive calls to this function.
3283 If X is a register whose contents are known, we do NOT
3284 return those contents here. equiv_constant is called to
3285 perform that task.
3287 INSN is the insn that we may be modifying. If it is 0, make a copy
3288 of X before modifying it. */
3290 static rtx
3291 fold_rtx (x, insn)
3292 rtx x;
3293 rtx insn;
3295 register enum rtx_code code;
3296 register enum machine_mode mode;
3297 register const char *fmt;
3298 register int i;
3299 rtx new = 0;
3300 int copied = 0;
3301 int must_swap = 0;
3303 /* Folded equivalents of first two operands of X. */
3304 rtx folded_arg0;
3305 rtx folded_arg1;
3307 /* Constant equivalents of first three operands of X;
3308 0 when no such equivalent is known. */
3309 rtx const_arg0;
3310 rtx const_arg1;
3311 rtx const_arg2;
3313 /* The mode of the first operand of X. We need this for sign and zero
3314 extends. */
3315 enum machine_mode mode_arg0;
3317 if (x == 0)
3318 return x;
3320 mode = GET_MODE (x);
3321 code = GET_CODE (x);
3322 switch (code)
3324 case CONST:
3325 case CONST_INT:
3326 case CONST_DOUBLE:
3327 case SYMBOL_REF:
3328 case LABEL_REF:
3329 case REG:
3330 /* No use simplifying an EXPR_LIST
3331 since they are used only for lists of args
3332 in a function call's REG_EQUAL note. */
3333 case EXPR_LIST:
3334 /* Changing anything inside an ADDRESSOF is incorrect; we don't
3335 want to (e.g.,) make (addressof (const_int 0)) just because
3336 the location is known to be zero. */
3337 case ADDRESSOF:
3338 return x;
3340 #ifdef HAVE_cc0
3341 case CC0:
3342 return prev_insn_cc0;
3343 #endif
3345 case PC:
3346 /* If the next insn is a CODE_LABEL followed by a jump table,
3347 PC's value is a LABEL_REF pointing to that label. That
3348 lets us fold switch statements on the Vax. */
3349 if (insn && GET_CODE (insn) == JUMP_INSN)
3351 rtx next = next_nonnote_insn (insn);
3353 if (next && GET_CODE (next) == CODE_LABEL
3354 && NEXT_INSN (next) != 0
3355 && GET_CODE (NEXT_INSN (next)) == JUMP_INSN
3356 && (GET_CODE (PATTERN (NEXT_INSN (next))) == ADDR_VEC
3357 || GET_CODE (PATTERN (NEXT_INSN (next))) == ADDR_DIFF_VEC))
3358 return gen_rtx_LABEL_REF (Pmode, next);
3360 break;
3362 case SUBREG:
3363 /* See if we previously assigned a constant value to this SUBREG. */
3364 if ((new = lookup_as_function (x, CONST_INT)) != 0
3365 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3366 return new;
3368 /* If this is a paradoxical SUBREG, we have no idea what value the
3369 extra bits would have. However, if the operand is equivalent
3370 to a SUBREG whose operand is the same as our mode, and all the
3371 modes are within a word, we can just use the inner operand
3372 because these SUBREGs just say how to treat the register.
3374 Similarly if we find an integer constant. */
3376 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3378 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3379 struct table_elt *elt;
3381 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3382 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3383 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3384 imode)) != 0)
3385 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3387 if (CONSTANT_P (elt->exp)
3388 && GET_MODE (elt->exp) == VOIDmode)
3389 return elt->exp;
3391 if (GET_CODE (elt->exp) == SUBREG
3392 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3393 && exp_equiv_p (elt->exp, elt->exp, 1, 0))
3394 return copy_rtx (SUBREG_REG (elt->exp));
3397 return x;
3400 /* Fold SUBREG_REG. If it changed, see if we can simplify the SUBREG.
3401 We might be able to if the SUBREG is extracting a single word in an
3402 integral mode or extracting the low part. */
3404 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3405 const_arg0 = equiv_constant (folded_arg0);
3406 if (const_arg0)
3407 folded_arg0 = const_arg0;
3409 if (folded_arg0 != SUBREG_REG (x))
3411 new = 0;
3413 if (GET_MODE_CLASS (mode) == MODE_INT
3414 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3415 && GET_MODE (SUBREG_REG (x)) != VOIDmode)
3416 new = operand_subword (folded_arg0,
3417 (SUBREG_BYTE (x) / UNITS_PER_WORD), 0,
3418 GET_MODE (SUBREG_REG (x)));
3419 if (new == 0 && subreg_lowpart_p (x))
3420 new = gen_lowpart_if_possible (mode, folded_arg0);
3421 if (new)
3422 return new;
3425 /* If this is a narrowing SUBREG and our operand is a REG, see if
3426 we can find an equivalence for REG that is an arithmetic operation
3427 in a wider mode where both operands are paradoxical SUBREGs
3428 from objects of our result mode. In that case, we couldn't report
3429 an equivalent value for that operation, since we don't know what the
3430 extra bits will be. But we can find an equivalence for this SUBREG
3431 by folding that operation is the narrow mode. This allows us to
3432 fold arithmetic in narrow modes when the machine only supports
3433 word-sized arithmetic.
3435 Also look for a case where we have a SUBREG whose operand is the
3436 same as our result. If both modes are smaller than a word, we
3437 are simply interpreting a register in different modes and we
3438 can use the inner value. */
3440 if (GET_CODE (folded_arg0) == REG
3441 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0))
3442 && subreg_lowpart_p (x))
3444 struct table_elt *elt;
3446 /* We can use HASH here since we know that canon_hash won't be
3447 called. */
3448 elt = lookup (folded_arg0,
3449 HASH (folded_arg0, GET_MODE (folded_arg0)),
3450 GET_MODE (folded_arg0));
3452 if (elt)
3453 elt = elt->first_same_value;
3455 for (; elt; elt = elt->next_same_value)
3457 enum rtx_code eltcode = GET_CODE (elt->exp);
3459 /* Just check for unary and binary operations. */
3460 if (GET_RTX_CLASS (GET_CODE (elt->exp)) == '1'
3461 && GET_CODE (elt->exp) != SIGN_EXTEND
3462 && GET_CODE (elt->exp) != ZERO_EXTEND
3463 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3464 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode)
3466 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3468 if (GET_CODE (op0) != REG && ! CONSTANT_P (op0))
3469 op0 = fold_rtx (op0, NULL_RTX);
3471 op0 = equiv_constant (op0);
3472 if (op0)
3473 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3474 op0, mode);
3476 else if ((GET_RTX_CLASS (GET_CODE (elt->exp)) == '2'
3477 || GET_RTX_CLASS (GET_CODE (elt->exp)) == 'c')
3478 && eltcode != DIV && eltcode != MOD
3479 && eltcode != UDIV && eltcode != UMOD
3480 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3481 && eltcode != ROTATE && eltcode != ROTATERT
3482 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3483 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3484 == mode))
3485 || CONSTANT_P (XEXP (elt->exp, 0)))
3486 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3487 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3488 == mode))
3489 || CONSTANT_P (XEXP (elt->exp, 1))))
3491 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3492 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3494 if (op0 && GET_CODE (op0) != REG && ! CONSTANT_P (op0))
3495 op0 = fold_rtx (op0, NULL_RTX);
3497 if (op0)
3498 op0 = equiv_constant (op0);
3500 if (op1 && GET_CODE (op1) != REG && ! CONSTANT_P (op1))
3501 op1 = fold_rtx (op1, NULL_RTX);
3503 if (op1)
3504 op1 = equiv_constant (op1);
3506 /* If we are looking for the low SImode part of
3507 (ashift:DI c (const_int 32)), it doesn't work
3508 to compute that in SImode, because a 32-bit shift
3509 in SImode is unpredictable. We know the value is 0. */
3510 if (op0 && op1
3511 && GET_CODE (elt->exp) == ASHIFT
3512 && GET_CODE (op1) == CONST_INT
3513 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3515 if (INTVAL (op1) < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3517 /* If the count fits in the inner mode's width,
3518 but exceeds the outer mode's width,
3519 the value will get truncated to 0
3520 by the subreg. */
3521 new = const0_rtx;
3522 else
3523 /* If the count exceeds even the inner mode's width,
3524 don't fold this expression. */
3525 new = 0;
3527 else if (op0 && op1)
3528 new = simplify_binary_operation (GET_CODE (elt->exp), mode,
3529 op0, op1);
3532 else if (GET_CODE (elt->exp) == SUBREG
3533 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3534 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3535 <= UNITS_PER_WORD)
3536 && exp_equiv_p (elt->exp, elt->exp, 1, 0))
3537 new = copy_rtx (SUBREG_REG (elt->exp));
3539 if (new)
3540 return new;
3544 return x;
3546 case NOT:
3547 case NEG:
3548 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3549 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3550 new = lookup_as_function (XEXP (x, 0), code);
3551 if (new)
3552 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3553 break;
3555 case MEM:
3556 /* If we are not actually processing an insn, don't try to find the
3557 best address. Not only don't we care, but we could modify the
3558 MEM in an invalid way since we have no insn to validate against. */
3559 if (insn != 0)
3560 find_best_addr (insn, &XEXP (x, 0), GET_MODE (x));
3563 /* Even if we don't fold in the insn itself,
3564 we can safely do so here, in hopes of getting a constant. */
3565 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3566 rtx base = 0;
3567 HOST_WIDE_INT offset = 0;
3569 if (GET_CODE (addr) == REG
3570 && REGNO_QTY_VALID_P (REGNO (addr)))
3572 int addr_q = REG_QTY (REGNO (addr));
3573 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3575 if (GET_MODE (addr) == addr_ent->mode
3576 && addr_ent->const_rtx != NULL_RTX)
3577 addr = addr_ent->const_rtx;
3580 /* If address is constant, split it into a base and integer offset. */
3581 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3582 base = addr;
3583 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3584 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3586 base = XEXP (XEXP (addr, 0), 0);
3587 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3589 else if (GET_CODE (addr) == LO_SUM
3590 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3591 base = XEXP (addr, 1);
3592 else if (GET_CODE (addr) == ADDRESSOF)
3593 return change_address (x, VOIDmode, addr);
3595 /* If this is a constant pool reference, we can fold it into its
3596 constant to allow better value tracking. */
3597 if (base && GET_CODE (base) == SYMBOL_REF
3598 && CONSTANT_POOL_ADDRESS_P (base))
3600 rtx constant = get_pool_constant (base);
3601 enum machine_mode const_mode = get_pool_mode (base);
3602 rtx new;
3604 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3605 constant_pool_entries_cost = COST (constant);
3607 /* If we are loading the full constant, we have an equivalence. */
3608 if (offset == 0 && mode == const_mode)
3609 return constant;
3611 /* If this actually isn't a constant (weird!), we can't do
3612 anything. Otherwise, handle the two most common cases:
3613 extracting a word from a multi-word constant, and extracting
3614 the low-order bits. Other cases don't seem common enough to
3615 worry about. */
3616 if (! CONSTANT_P (constant))
3617 return x;
3619 if (GET_MODE_CLASS (mode) == MODE_INT
3620 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3621 && offset % UNITS_PER_WORD == 0
3622 && (new = operand_subword (constant,
3623 offset / UNITS_PER_WORD,
3624 0, const_mode)) != 0)
3625 return new;
3627 if (((BYTES_BIG_ENDIAN
3628 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3629 || (! BYTES_BIG_ENDIAN && offset == 0))
3630 && (new = gen_lowpart_if_possible (mode, constant)) != 0)
3631 return new;
3634 /* If this is a reference to a label at a known position in a jump
3635 table, we also know its value. */
3636 if (base && GET_CODE (base) == LABEL_REF)
3638 rtx label = XEXP (base, 0);
3639 rtx table_insn = NEXT_INSN (label);
3641 if (table_insn && GET_CODE (table_insn) == JUMP_INSN
3642 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3644 rtx table = PATTERN (table_insn);
3646 if (offset >= 0
3647 && (offset / GET_MODE_SIZE (GET_MODE (table))
3648 < XVECLEN (table, 0)))
3649 return XVECEXP (table, 0,
3650 offset / GET_MODE_SIZE (GET_MODE (table)));
3652 if (table_insn && GET_CODE (table_insn) == JUMP_INSN
3653 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3655 rtx table = PATTERN (table_insn);
3657 if (offset >= 0
3658 && (offset / GET_MODE_SIZE (GET_MODE (table))
3659 < XVECLEN (table, 1)))
3661 offset /= GET_MODE_SIZE (GET_MODE (table));
3662 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3663 XEXP (table, 0));
3665 if (GET_MODE (table) != Pmode)
3666 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3668 /* Indicate this is a constant. This isn't a
3669 valid form of CONST, but it will only be used
3670 to fold the next insns and then discarded, so
3671 it should be safe.
3673 Note this expression must be explicitly discarded,
3674 by cse_insn, else it may end up in a REG_EQUAL note
3675 and "escape" to cause problems elsewhere. */
3676 return gen_rtx_CONST (GET_MODE (new), new);
3681 return x;
3684 #ifdef NO_FUNCTION_CSE
3685 case CALL:
3686 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3687 return x;
3688 break;
3689 #endif
3691 case ASM_OPERANDS:
3692 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3693 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3694 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3695 break;
3697 default:
3698 break;
3701 const_arg0 = 0;
3702 const_arg1 = 0;
3703 const_arg2 = 0;
3704 mode_arg0 = VOIDmode;
3706 /* Try folding our operands.
3707 Then see which ones have constant values known. */
3709 fmt = GET_RTX_FORMAT (code);
3710 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3711 if (fmt[i] == 'e')
3713 rtx arg = XEXP (x, i);
3714 rtx folded_arg = arg, const_arg = 0;
3715 enum machine_mode mode_arg = GET_MODE (arg);
3716 rtx cheap_arg, expensive_arg;
3717 rtx replacements[2];
3718 int j;
3720 /* Most arguments are cheap, so handle them specially. */
3721 switch (GET_CODE (arg))
3723 case REG:
3724 /* This is the same as calling equiv_constant; it is duplicated
3725 here for speed. */
3726 if (REGNO_QTY_VALID_P (REGNO (arg)))
3728 int arg_q = REG_QTY (REGNO (arg));
3729 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3731 if (arg_ent->const_rtx != NULL_RTX
3732 && GET_CODE (arg_ent->const_rtx) != REG
3733 && GET_CODE (arg_ent->const_rtx) != PLUS)
3734 const_arg
3735 = gen_lowpart_if_possible (GET_MODE (arg),
3736 arg_ent->const_rtx);
3738 break;
3740 case CONST:
3741 case CONST_INT:
3742 case SYMBOL_REF:
3743 case LABEL_REF:
3744 case CONST_DOUBLE:
3745 const_arg = arg;
3746 break;
3748 #ifdef HAVE_cc0
3749 case CC0:
3750 folded_arg = prev_insn_cc0;
3751 mode_arg = prev_insn_cc0_mode;
3752 const_arg = equiv_constant (folded_arg);
3753 break;
3754 #endif
3756 default:
3757 folded_arg = fold_rtx (arg, insn);
3758 const_arg = equiv_constant (folded_arg);
3761 /* For the first three operands, see if the operand
3762 is constant or equivalent to a constant. */
3763 switch (i)
3765 case 0:
3766 folded_arg0 = folded_arg;
3767 const_arg0 = const_arg;
3768 mode_arg0 = mode_arg;
3769 break;
3770 case 1:
3771 folded_arg1 = folded_arg;
3772 const_arg1 = const_arg;
3773 break;
3774 case 2:
3775 const_arg2 = const_arg;
3776 break;
3779 /* Pick the least expensive of the folded argument and an
3780 equivalent constant argument. */
3781 if (const_arg == 0 || const_arg == folded_arg
3782 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3783 cheap_arg = folded_arg, expensive_arg = const_arg;
3784 else
3785 cheap_arg = const_arg, expensive_arg = folded_arg;
3787 /* Try to replace the operand with the cheapest of the two
3788 possibilities. If it doesn't work and this is either of the first
3789 two operands of a commutative operation, try swapping them.
3790 If THAT fails, try the more expensive, provided it is cheaper
3791 than what is already there. */
3793 if (cheap_arg == XEXP (x, i))
3794 continue;
3796 if (insn == 0 && ! copied)
3798 x = copy_rtx (x);
3799 copied = 1;
3802 /* Order the replacements from cheapest to most expensive. */
3803 replacements[0] = cheap_arg;
3804 replacements[1] = expensive_arg;
3806 for (j = 0; j < 2 && replacements[j]; j++)
3808 int old_cost = COST_IN (XEXP (x, i), code);
3809 int new_cost = COST_IN (replacements[j], code);
3811 /* Stop if what existed before was cheaper. Prefer constants
3812 in the case of a tie. */
3813 if (new_cost > old_cost
3814 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3815 break;
3817 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3818 break;
3820 if (code == NE || code == EQ || GET_RTX_CLASS (code) == 'c'
3821 || code == LTGT || code == UNEQ || code == ORDERED
3822 || code == UNORDERED)
3824 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3825 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3827 if (apply_change_group ())
3829 /* Swap them back to be invalid so that this loop can
3830 continue and flag them to be swapped back later. */
3831 rtx tem;
3833 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3834 XEXP (x, 1) = tem;
3835 must_swap = 1;
3836 break;
3842 else
3844 if (fmt[i] == 'E')
3845 /* Don't try to fold inside of a vector of expressions.
3846 Doing nothing is harmless. */
3850 /* If a commutative operation, place a constant integer as the second
3851 operand unless the first operand is also a constant integer. Otherwise,
3852 place any constant second unless the first operand is also a constant. */
3854 if (code == EQ || code == NE || GET_RTX_CLASS (code) == 'c'
3855 || code == LTGT || code == UNEQ || code == ORDERED
3856 || code == UNORDERED)
3858 if (must_swap || (const_arg0
3859 && (const_arg1 == 0
3860 || (GET_CODE (const_arg0) == CONST_INT
3861 && GET_CODE (const_arg1) != CONST_INT))))
3863 register rtx tem = XEXP (x, 0);
3865 if (insn == 0 && ! copied)
3867 x = copy_rtx (x);
3868 copied = 1;
3871 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3872 validate_change (insn, &XEXP (x, 1), tem, 1);
3873 if (apply_change_group ())
3875 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3876 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3881 /* If X is an arithmetic operation, see if we can simplify it. */
3883 switch (GET_RTX_CLASS (code))
3885 case '1':
3887 int is_const = 0;
3889 /* We can't simplify extension ops unless we know the
3890 original mode. */
3891 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3892 && mode_arg0 == VOIDmode)
3893 break;
3895 /* If we had a CONST, strip it off and put it back later if we
3896 fold. */
3897 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3898 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3900 new = simplify_unary_operation (code, mode,
3901 const_arg0 ? const_arg0 : folded_arg0,
3902 mode_arg0);
3903 if (new != 0 && is_const)
3904 new = gen_rtx_CONST (mode, new);
3906 break;
3908 case '<':
3909 /* See what items are actually being compared and set FOLDED_ARG[01]
3910 to those values and CODE to the actual comparison code. If any are
3911 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3912 do anything if both operands are already known to be constant. */
3914 if (const_arg0 == 0 || const_arg1 == 0)
3916 struct table_elt *p0, *p1;
3917 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3918 enum machine_mode mode_arg1;
3920 #ifdef FLOAT_STORE_FLAG_VALUE
3921 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3923 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3924 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3925 false_rtx = CONST0_RTX (mode);
3927 #endif
3929 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3930 &mode_arg0, &mode_arg1);
3931 const_arg0 = equiv_constant (folded_arg0);
3932 const_arg1 = equiv_constant (folded_arg1);
3934 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3935 what kinds of things are being compared, so we can't do
3936 anything with this comparison. */
3938 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3939 break;
3941 /* If we do not now have two constants being compared, see
3942 if we can nevertheless deduce some things about the
3943 comparison. */
3944 if (const_arg0 == 0 || const_arg1 == 0)
3946 /* Is FOLDED_ARG0 frame-pointer plus a constant? Or
3947 non-explicit constant? These aren't zero, but we
3948 don't know their sign. */
3949 if (const_arg1 == const0_rtx
3950 && (NONZERO_BASE_PLUS_P (folded_arg0)
3951 #if 0 /* Sad to say, on sysvr4, #pragma weak can make a symbol address
3952 come out as 0. */
3953 || GET_CODE (folded_arg0) == SYMBOL_REF
3954 #endif
3955 || GET_CODE (folded_arg0) == LABEL_REF
3956 || GET_CODE (folded_arg0) == CONST))
3958 if (code == EQ)
3959 return false_rtx;
3960 else if (code == NE)
3961 return true_rtx;
3964 /* See if the two operands are the same. */
3966 if (folded_arg0 == folded_arg1
3967 || (GET_CODE (folded_arg0) == REG
3968 && GET_CODE (folded_arg1) == REG
3969 && (REG_QTY (REGNO (folded_arg0))
3970 == REG_QTY (REGNO (folded_arg1))))
3971 || ((p0 = lookup (folded_arg0,
3972 (safe_hash (folded_arg0, mode_arg0)
3973 & HASH_MASK), mode_arg0))
3974 && (p1 = lookup (folded_arg1,
3975 (safe_hash (folded_arg1, mode_arg0)
3976 & HASH_MASK), mode_arg0))
3977 && p0->first_same_value == p1->first_same_value))
3979 /* Sadly two equal NaNs are not equivalent. */
3980 if (TARGET_FLOAT_FORMAT != IEEE_FLOAT_FORMAT
3981 || ! FLOAT_MODE_P (mode_arg0)
3982 || flag_unsafe_math_optimizations)
3983 return ((code == EQ || code == LE || code == GE
3984 || code == LEU || code == GEU || code == UNEQ
3985 || code == UNLE || code == UNGE || code == ORDERED)
3986 ? true_rtx : false_rtx);
3987 /* Take care for the FP compares we can resolve. */
3988 if (code == UNEQ || code == UNLE || code == UNGE)
3989 return true_rtx;
3990 if (code == LTGT || code == LT || code == GT)
3991 return false_rtx;
3994 /* If FOLDED_ARG0 is a register, see if the comparison we are
3995 doing now is either the same as we did before or the reverse
3996 (we only check the reverse if not floating-point). */
3997 else if (GET_CODE (folded_arg0) == REG)
3999 int qty = REG_QTY (REGNO (folded_arg0));
4001 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4003 struct qty_table_elem *ent = &qty_table[qty];
4005 if ((comparison_dominates_p (ent->comparison_code, code)
4006 || (! FLOAT_MODE_P (mode_arg0)
4007 && comparison_dominates_p (ent->comparison_code,
4008 reverse_condition (code))))
4009 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4010 || (const_arg1
4011 && rtx_equal_p (ent->comparison_const,
4012 const_arg1))
4013 || (GET_CODE (folded_arg1) == REG
4014 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4015 return (comparison_dominates_p (ent->comparison_code, code)
4016 ? true_rtx : false_rtx);
4022 /* If we are comparing against zero, see if the first operand is
4023 equivalent to an IOR with a constant. If so, we may be able to
4024 determine the result of this comparison. */
4026 if (const_arg1 == const0_rtx)
4028 rtx y = lookup_as_function (folded_arg0, IOR);
4029 rtx inner_const;
4031 if (y != 0
4032 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4033 && GET_CODE (inner_const) == CONST_INT
4034 && INTVAL (inner_const) != 0)
4036 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4037 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4038 && (INTVAL (inner_const)
4039 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4040 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4042 #ifdef FLOAT_STORE_FLAG_VALUE
4043 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
4045 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4046 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4047 false_rtx = CONST0_RTX (mode);
4049 #endif
4051 switch (code)
4053 case EQ:
4054 return false_rtx;
4055 case NE:
4056 return true_rtx;
4057 case LT: case LE:
4058 if (has_sign)
4059 return true_rtx;
4060 break;
4061 case GT: case GE:
4062 if (has_sign)
4063 return false_rtx;
4064 break;
4065 default:
4066 break;
4071 new = simplify_relational_operation (code,
4072 (mode_arg0 != VOIDmode
4073 ? mode_arg0
4074 : (GET_MODE (const_arg0
4075 ? const_arg0
4076 : folded_arg0)
4077 != VOIDmode)
4078 ? GET_MODE (const_arg0
4079 ? const_arg0
4080 : folded_arg0)
4081 : GET_MODE (const_arg1
4082 ? const_arg1
4083 : folded_arg1)),
4084 const_arg0 ? const_arg0 : folded_arg0,
4085 const_arg1 ? const_arg1 : folded_arg1);
4086 #ifdef FLOAT_STORE_FLAG_VALUE
4087 if (new != 0 && GET_MODE_CLASS (mode) == MODE_FLOAT)
4089 if (new == const0_rtx)
4090 new = CONST0_RTX (mode);
4091 else
4092 new = (CONST_DOUBLE_FROM_REAL_VALUE
4093 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4095 #endif
4096 break;
4098 case '2':
4099 case 'c':
4100 switch (code)
4102 case PLUS:
4103 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4104 with that LABEL_REF as its second operand. If so, the result is
4105 the first operand of that MINUS. This handles switches with an
4106 ADDR_DIFF_VEC table. */
4107 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4109 rtx y
4110 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4111 : lookup_as_function (folded_arg0, MINUS);
4113 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4114 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4115 return XEXP (y, 0);
4117 /* Now try for a CONST of a MINUS like the above. */
4118 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4119 : lookup_as_function (folded_arg0, CONST))) != 0
4120 && GET_CODE (XEXP (y, 0)) == MINUS
4121 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4122 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4123 return XEXP (XEXP (y, 0), 0);
4126 /* Likewise if the operands are in the other order. */
4127 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4129 rtx y
4130 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4131 : lookup_as_function (folded_arg1, MINUS);
4133 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4134 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4135 return XEXP (y, 0);
4137 /* Now try for a CONST of a MINUS like the above. */
4138 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4139 : lookup_as_function (folded_arg1, CONST))) != 0
4140 && GET_CODE (XEXP (y, 0)) == MINUS
4141 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4142 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4143 return XEXP (XEXP (y, 0), 0);
4146 /* If second operand is a register equivalent to a negative
4147 CONST_INT, see if we can find a register equivalent to the
4148 positive constant. Make a MINUS if so. Don't do this for
4149 a non-negative constant since we might then alternate between
4150 chosing positive and negative constants. Having the positive
4151 constant previously-used is the more common case. Be sure
4152 the resulting constant is non-negative; if const_arg1 were
4153 the smallest negative number this would overflow: depending
4154 on the mode, this would either just be the same value (and
4155 hence not save anything) or be incorrect. */
4156 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4157 && INTVAL (const_arg1) < 0
4158 /* This used to test
4160 -INTVAL (const_arg1) >= 0
4162 But The Sun V5.0 compilers mis-compiled that test. So
4163 instead we test for the problematic value in a more direct
4164 manner and hope the Sun compilers get it correct. */
4165 && INTVAL (const_arg1) !=
4166 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4167 && GET_CODE (folded_arg1) == REG)
4169 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4170 struct table_elt *p
4171 = lookup (new_const, safe_hash (new_const, mode) & HASH_MASK,
4172 mode);
4174 if (p)
4175 for (p = p->first_same_value; p; p = p->next_same_value)
4176 if (GET_CODE (p->exp) == REG)
4177 return simplify_gen_binary (MINUS, mode, folded_arg0,
4178 canon_reg (p->exp, NULL_RTX));
4180 goto from_plus;
4182 case MINUS:
4183 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4184 If so, produce (PLUS Z C2-C). */
4185 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4187 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4188 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4189 return fold_rtx (plus_constant (copy_rtx (y),
4190 -INTVAL (const_arg1)),
4191 NULL_RTX);
4194 /* Fall through. */
4196 from_plus:
4197 case SMIN: case SMAX: case UMIN: case UMAX:
4198 case IOR: case AND: case XOR:
4199 case MULT: case DIV: case UDIV:
4200 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4201 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4202 is known to be of similar form, we may be able to replace the
4203 operation with a combined operation. This may eliminate the
4204 intermediate operation if every use is simplified in this way.
4205 Note that the similar optimization done by combine.c only works
4206 if the intermediate operation's result has only one reference. */
4208 if (GET_CODE (folded_arg0) == REG
4209 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4211 int is_shift
4212 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4213 rtx y = lookup_as_function (folded_arg0, code);
4214 rtx inner_const;
4215 enum rtx_code associate_code;
4216 rtx new_const;
4218 if (y == 0
4219 || 0 == (inner_const
4220 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4221 || GET_CODE (inner_const) != CONST_INT
4222 /* If we have compiled a statement like
4223 "if (x == (x & mask1))", and now are looking at
4224 "x & mask2", we will have a case where the first operand
4225 of Y is the same as our first operand. Unless we detect
4226 this case, an infinite loop will result. */
4227 || XEXP (y, 0) == folded_arg0)
4228 break;
4230 /* Don't associate these operations if they are a PLUS with the
4231 same constant and it is a power of two. These might be doable
4232 with a pre- or post-increment. Similarly for two subtracts of
4233 identical powers of two with post decrement. */
4235 if (code == PLUS && INTVAL (const_arg1) == INTVAL (inner_const)
4236 && ((HAVE_PRE_INCREMENT
4237 && exact_log2 (INTVAL (const_arg1)) >= 0)
4238 || (HAVE_POST_INCREMENT
4239 && exact_log2 (INTVAL (const_arg1)) >= 0)
4240 || (HAVE_PRE_DECREMENT
4241 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4242 || (HAVE_POST_DECREMENT
4243 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4244 break;
4246 /* Compute the code used to compose the constants. For example,
4247 A/C1/C2 is A/(C1 * C2), so if CODE == DIV, we want MULT. */
4249 associate_code
4250 = (code == MULT || code == DIV || code == UDIV ? MULT
4251 : is_shift || code == PLUS || code == MINUS ? PLUS : code);
4253 new_const = simplify_binary_operation (associate_code, mode,
4254 const_arg1, inner_const);
4256 if (new_const == 0)
4257 break;
4259 /* If we are associating shift operations, don't let this
4260 produce a shift of the size of the object or larger.
4261 This could occur when we follow a sign-extend by a right
4262 shift on a machine that does a sign-extend as a pair
4263 of shifts. */
4265 if (is_shift && GET_CODE (new_const) == CONST_INT
4266 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4268 /* As an exception, we can turn an ASHIFTRT of this
4269 form into a shift of the number of bits - 1. */
4270 if (code == ASHIFTRT)
4271 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4272 else
4273 break;
4276 y = copy_rtx (XEXP (y, 0));
4278 /* If Y contains our first operand (the most common way this
4279 can happen is if Y is a MEM), we would do into an infinite
4280 loop if we tried to fold it. So don't in that case. */
4282 if (! reg_mentioned_p (folded_arg0, y))
4283 y = fold_rtx (y, insn);
4285 return simplify_gen_binary (code, mode, y, new_const);
4287 break;
4289 default:
4290 break;
4293 new = simplify_binary_operation (code, mode,
4294 const_arg0 ? const_arg0 : folded_arg0,
4295 const_arg1 ? const_arg1 : folded_arg1);
4296 break;
4298 case 'o':
4299 /* (lo_sum (high X) X) is simply X. */
4300 if (code == LO_SUM && const_arg0 != 0
4301 && GET_CODE (const_arg0) == HIGH
4302 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4303 return const_arg1;
4304 break;
4306 case '3':
4307 case 'b':
4308 new = simplify_ternary_operation (code, mode, mode_arg0,
4309 const_arg0 ? const_arg0 : folded_arg0,
4310 const_arg1 ? const_arg1 : folded_arg1,
4311 const_arg2 ? const_arg2 : XEXP (x, 2));
4312 break;
4314 case 'x':
4315 /* Always eliminate CONSTANT_P_RTX at this stage. */
4316 if (code == CONSTANT_P_RTX)
4317 return (const_arg0 ? const1_rtx : const0_rtx);
4318 break;
4321 return new ? new : x;
4324 /* Return a constant value currently equivalent to X.
4325 Return 0 if we don't know one. */
4327 static rtx
4328 equiv_constant (x)
4329 rtx x;
4331 if (GET_CODE (x) == REG
4332 && REGNO_QTY_VALID_P (REGNO (x)))
4334 int x_q = REG_QTY (REGNO (x));
4335 struct qty_table_elem *x_ent = &qty_table[x_q];
4337 if (x_ent->const_rtx)
4338 x = gen_lowpart_if_possible (GET_MODE (x), x_ent->const_rtx);
4341 if (x == 0 || CONSTANT_P (x))
4342 return x;
4344 /* If X is a MEM, try to fold it outside the context of any insn to see if
4345 it might be equivalent to a constant. That handles the case where it
4346 is a constant-pool reference. Then try to look it up in the hash table
4347 in case it is something whose value we have seen before. */
4349 if (GET_CODE (x) == MEM)
4351 struct table_elt *elt;
4353 x = fold_rtx (x, NULL_RTX);
4354 if (CONSTANT_P (x))
4355 return x;
4357 elt = lookup (x, safe_hash (x, GET_MODE (x)) & HASH_MASK, GET_MODE (x));
4358 if (elt == 0)
4359 return 0;
4361 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4362 if (elt->is_const && CONSTANT_P (elt->exp))
4363 return elt->exp;
4366 return 0;
4369 /* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a fixed-point
4370 number, return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
4371 least-significant part of X.
4372 MODE specifies how big a part of X to return.
4374 If the requested operation cannot be done, 0 is returned.
4376 This is similar to gen_lowpart in emit-rtl.c. */
4379 gen_lowpart_if_possible (mode, x)
4380 enum machine_mode mode;
4381 register rtx x;
4383 rtx result = gen_lowpart_common (mode, x);
4385 if (result)
4386 return result;
4387 else if (GET_CODE (x) == MEM)
4389 /* This is the only other case we handle. */
4390 register int offset = 0;
4391 rtx new;
4393 if (WORDS_BIG_ENDIAN)
4394 offset = (MAX (GET_MODE_SIZE (GET_MODE (x)), UNITS_PER_WORD)
4395 - MAX (GET_MODE_SIZE (mode), UNITS_PER_WORD));
4396 if (BYTES_BIG_ENDIAN)
4398 /* Adjust the address so that the address-after-the-data is
4399 unchanged. */
4400 offset -= (MIN (UNITS_PER_WORD, GET_MODE_SIZE (mode))
4401 - MIN (UNITS_PER_WORD, GET_MODE_SIZE (GET_MODE (x))));
4403 new = gen_rtx_MEM (mode, plus_constant (XEXP (x, 0), offset));
4404 if (! memory_address_p (mode, XEXP (new, 0)))
4405 return 0;
4406 MEM_COPY_ATTRIBUTES (new, x);
4407 return new;
4409 else
4410 return 0;
4413 /* Given INSN, a jump insn, TAKEN indicates if we are following the "taken"
4414 branch. It will be zero if not.
4416 In certain cases, this can cause us to add an equivalence. For example,
4417 if we are following the taken case of
4418 if (i == 2)
4419 we can add the fact that `i' and '2' are now equivalent.
4421 In any case, we can record that this comparison was passed. If the same
4422 comparison is seen later, we will know its value. */
4424 static void
4425 record_jump_equiv (insn, taken)
4426 rtx insn;
4427 int taken;
4429 int cond_known_true;
4430 rtx op0, op1;
4431 rtx set;
4432 enum machine_mode mode, mode0, mode1;
4433 int reversed_nonequality = 0;
4434 enum rtx_code code;
4436 /* Ensure this is the right kind of insn. */
4437 if (! any_condjump_p (insn))
4438 return;
4439 set = pc_set (insn);
4441 /* See if this jump condition is known true or false. */
4442 if (taken)
4443 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4444 else
4445 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4447 /* Get the type of comparison being done and the operands being compared.
4448 If we had to reverse a non-equality condition, record that fact so we
4449 know that it isn't valid for floating-point. */
4450 code = GET_CODE (XEXP (SET_SRC (set), 0));
4451 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4452 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4454 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4455 if (! cond_known_true)
4457 code = reversed_comparison_code_parts (code, op0, op1, insn);
4459 /* Don't remember if we can't find the inverse. */
4460 if (code == UNKNOWN)
4461 return;
4464 /* The mode is the mode of the non-constant. */
4465 mode = mode0;
4466 if (mode1 != VOIDmode)
4467 mode = mode1;
4469 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4472 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4473 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4474 Make any useful entries we can with that information. Called from
4475 above function and called recursively. */
4477 static void
4478 record_jump_cond (code, mode, op0, op1, reversed_nonequality)
4479 enum rtx_code code;
4480 enum machine_mode mode;
4481 rtx op0, op1;
4482 int reversed_nonequality;
4484 unsigned op0_hash, op1_hash;
4485 int op0_in_memory, op1_in_memory;
4486 struct table_elt *op0_elt, *op1_elt;
4488 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4489 we know that they are also equal in the smaller mode (this is also
4490 true for all smaller modes whether or not there is a SUBREG, but
4491 is not worth testing for with no SUBREG). */
4493 /* Note that GET_MODE (op0) may not equal MODE. */
4494 if (code == EQ && GET_CODE (op0) == SUBREG
4495 && (GET_MODE_SIZE (GET_MODE (op0))
4496 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4498 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4499 rtx tem = gen_lowpart_if_possible (inner_mode, op1);
4501 record_jump_cond (code, mode, SUBREG_REG (op0),
4502 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4503 reversed_nonequality);
4506 if (code == EQ && GET_CODE (op1) == SUBREG
4507 && (GET_MODE_SIZE (GET_MODE (op1))
4508 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4510 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4511 rtx tem = gen_lowpart_if_possible (inner_mode, op0);
4513 record_jump_cond (code, mode, SUBREG_REG (op1),
4514 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4515 reversed_nonequality);
4518 /* Similarly, if this is an NE comparison, and either is a SUBREG
4519 making a smaller mode, we know the whole thing is also NE. */
4521 /* Note that GET_MODE (op0) may not equal MODE;
4522 if we test MODE instead, we can get an infinite recursion
4523 alternating between two modes each wider than MODE. */
4525 if (code == NE && GET_CODE (op0) == SUBREG
4526 && subreg_lowpart_p (op0)
4527 && (GET_MODE_SIZE (GET_MODE (op0))
4528 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4530 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4531 rtx tem = gen_lowpart_if_possible (inner_mode, op1);
4533 record_jump_cond (code, mode, SUBREG_REG (op0),
4534 tem ? tem : gen_rtx_SUBREG (inner_mode, op1, 0),
4535 reversed_nonequality);
4538 if (code == NE && GET_CODE (op1) == SUBREG
4539 && subreg_lowpart_p (op1)
4540 && (GET_MODE_SIZE (GET_MODE (op1))
4541 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4543 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4544 rtx tem = gen_lowpart_if_possible (inner_mode, op0);
4546 record_jump_cond (code, mode, SUBREG_REG (op1),
4547 tem ? tem : gen_rtx_SUBREG (inner_mode, op0, 0),
4548 reversed_nonequality);
4551 /* Hash both operands. */
4553 do_not_record = 0;
4554 hash_arg_in_memory = 0;
4555 op0_hash = HASH (op0, mode);
4556 op0_in_memory = hash_arg_in_memory;
4558 if (do_not_record)
4559 return;
4561 do_not_record = 0;
4562 hash_arg_in_memory = 0;
4563 op1_hash = HASH (op1, mode);
4564 op1_in_memory = hash_arg_in_memory;
4566 if (do_not_record)
4567 return;
4569 /* Look up both operands. */
4570 op0_elt = lookup (op0, op0_hash, mode);
4571 op1_elt = lookup (op1, op1_hash, mode);
4573 /* If both operands are already equivalent or if they are not in the
4574 table but are identical, do nothing. */
4575 if ((op0_elt != 0 && op1_elt != 0
4576 && op0_elt->first_same_value == op1_elt->first_same_value)
4577 || op0 == op1 || rtx_equal_p (op0, op1))
4578 return;
4580 /* If we aren't setting two things equal all we can do is save this
4581 comparison. Similarly if this is floating-point. In the latter
4582 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4583 If we record the equality, we might inadvertently delete code
4584 whose intent was to change -0 to +0. */
4586 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4588 struct qty_table_elem *ent;
4589 int qty;
4591 /* If we reversed a floating-point comparison, if OP0 is not a
4592 register, or if OP1 is neither a register or constant, we can't
4593 do anything. */
4595 if (GET_CODE (op1) != REG)
4596 op1 = equiv_constant (op1);
4598 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4599 || GET_CODE (op0) != REG || op1 == 0)
4600 return;
4602 /* Put OP0 in the hash table if it isn't already. This gives it a
4603 new quantity number. */
4604 if (op0_elt == 0)
4606 if (insert_regs (op0, NULL_PTR, 0))
4608 rehash_using_reg (op0);
4609 op0_hash = HASH (op0, mode);
4611 /* If OP0 is contained in OP1, this changes its hash code
4612 as well. Faster to rehash than to check, except
4613 for the simple case of a constant. */
4614 if (! CONSTANT_P (op1))
4615 op1_hash = HASH (op1,mode);
4618 op0_elt = insert (op0, NULL_PTR, op0_hash, mode);
4619 op0_elt->in_memory = op0_in_memory;
4622 qty = REG_QTY (REGNO (op0));
4623 ent = &qty_table[qty];
4625 ent->comparison_code = code;
4626 if (GET_CODE (op1) == REG)
4628 /* Look it up again--in case op0 and op1 are the same. */
4629 op1_elt = lookup (op1, op1_hash, mode);
4631 /* Put OP1 in the hash table so it gets a new quantity number. */
4632 if (op1_elt == 0)
4634 if (insert_regs (op1, NULL_PTR, 0))
4636 rehash_using_reg (op1);
4637 op1_hash = HASH (op1, mode);
4640 op1_elt = insert (op1, NULL_PTR, op1_hash, mode);
4641 op1_elt->in_memory = op1_in_memory;
4644 ent->comparison_const = NULL_RTX;
4645 ent->comparison_qty = REG_QTY (REGNO (op1));
4647 else
4649 ent->comparison_const = op1;
4650 ent->comparison_qty = -1;
4653 return;
4656 /* If either side is still missing an equivalence, make it now,
4657 then merge the equivalences. */
4659 if (op0_elt == 0)
4661 if (insert_regs (op0, NULL_PTR, 0))
4663 rehash_using_reg (op0);
4664 op0_hash = HASH (op0, mode);
4667 op0_elt = insert (op0, NULL_PTR, op0_hash, mode);
4668 op0_elt->in_memory = op0_in_memory;
4671 if (op1_elt == 0)
4673 if (insert_regs (op1, NULL_PTR, 0))
4675 rehash_using_reg (op1);
4676 op1_hash = HASH (op1, mode);
4679 op1_elt = insert (op1, NULL_PTR, op1_hash, mode);
4680 op1_elt->in_memory = op1_in_memory;
4683 merge_equiv_classes (op0_elt, op1_elt);
4684 last_jump_equiv_class = op0_elt;
4687 /* CSE processing for one instruction.
4688 First simplify sources and addresses of all assignments
4689 in the instruction, using previously-computed equivalents values.
4690 Then install the new sources and destinations in the table
4691 of available values.
4693 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4694 the insn. It means that INSN is inside libcall block. In this
4695 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4697 /* Data on one SET contained in the instruction. */
4699 struct set
4701 /* The SET rtx itself. */
4702 rtx rtl;
4703 /* The SET_SRC of the rtx (the original value, if it is changing). */
4704 rtx src;
4705 /* The hash-table element for the SET_SRC of the SET. */
4706 struct table_elt *src_elt;
4707 /* Hash value for the SET_SRC. */
4708 unsigned src_hash;
4709 /* Hash value for the SET_DEST. */
4710 unsigned dest_hash;
4711 /* The SET_DEST, with SUBREG, etc., stripped. */
4712 rtx inner_dest;
4713 /* Nonzero if the SET_SRC is in memory. */
4714 char src_in_memory;
4715 /* Nonzero if the SET_SRC contains something
4716 whose value cannot be predicted and understood. */
4717 char src_volatile;
4718 /* Original machine mode, in case it becomes a CONST_INT. */
4719 enum machine_mode mode;
4720 /* A constant equivalent for SET_SRC, if any. */
4721 rtx src_const;
4722 /* Original SET_SRC value used for libcall notes. */
4723 rtx orig_src;
4724 /* Hash value of constant equivalent for SET_SRC. */
4725 unsigned src_const_hash;
4726 /* Table entry for constant equivalent for SET_SRC, if any. */
4727 struct table_elt *src_const_elt;
4730 static void
4731 cse_insn (insn, libcall_insn)
4732 rtx insn;
4733 rtx libcall_insn;
4735 register rtx x = PATTERN (insn);
4736 register int i;
4737 rtx tem;
4738 register int n_sets = 0;
4740 #ifdef HAVE_cc0
4741 /* Records what this insn does to set CC0. */
4742 rtx this_insn_cc0 = 0;
4743 enum machine_mode this_insn_cc0_mode = VOIDmode;
4744 #endif
4746 rtx src_eqv = 0;
4747 struct table_elt *src_eqv_elt = 0;
4748 int src_eqv_volatile = 0;
4749 int src_eqv_in_memory = 0;
4750 unsigned src_eqv_hash = 0;
4752 struct set *sets = (struct set *) NULL_PTR;
4754 this_insn = insn;
4756 /* Find all the SETs and CLOBBERs in this instruction.
4757 Record all the SETs in the array `set' and count them.
4758 Also determine whether there is a CLOBBER that invalidates
4759 all memory references, or all references at varying addresses. */
4761 if (GET_CODE (insn) == CALL_INSN)
4763 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4765 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4766 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4767 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4771 if (GET_CODE (x) == SET)
4773 sets = (struct set *) alloca (sizeof (struct set));
4774 sets[0].rtl = x;
4776 /* Ignore SETs that are unconditional jumps.
4777 They never need cse processing, so this does not hurt.
4778 The reason is not efficiency but rather
4779 so that we can test at the end for instructions
4780 that have been simplified to unconditional jumps
4781 and not be misled by unchanged instructions
4782 that were unconditional jumps to begin with. */
4783 if (SET_DEST (x) == pc_rtx
4784 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4787 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4788 The hard function value register is used only once, to copy to
4789 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4790 Ensure we invalidate the destination register. On the 80386 no
4791 other code would invalidate it since it is a fixed_reg.
4792 We need not check the return of apply_change_group; see canon_reg. */
4794 else if (GET_CODE (SET_SRC (x)) == CALL)
4796 canon_reg (SET_SRC (x), insn);
4797 apply_change_group ();
4798 fold_rtx (SET_SRC (x), insn);
4799 invalidate (SET_DEST (x), VOIDmode);
4801 else
4802 n_sets = 1;
4804 else if (GET_CODE (x) == PARALLEL)
4806 register int lim = XVECLEN (x, 0);
4808 sets = (struct set *) alloca (lim * sizeof (struct set));
4810 /* Find all regs explicitly clobbered in this insn,
4811 and ensure they are not replaced with any other regs
4812 elsewhere in this insn.
4813 When a reg that is clobbered is also used for input,
4814 we should presume that that is for a reason,
4815 and we should not substitute some other register
4816 which is not supposed to be clobbered.
4817 Therefore, this loop cannot be merged into the one below
4818 because a CALL may precede a CLOBBER and refer to the
4819 value clobbered. We must not let a canonicalization do
4820 anything in that case. */
4821 for (i = 0; i < lim; i++)
4823 register rtx y = XVECEXP (x, 0, i);
4824 if (GET_CODE (y) == CLOBBER)
4826 rtx clobbered = XEXP (y, 0);
4828 if (GET_CODE (clobbered) == REG
4829 || GET_CODE (clobbered) == SUBREG)
4830 invalidate (clobbered, VOIDmode);
4831 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4832 || GET_CODE (clobbered) == ZERO_EXTRACT)
4833 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4837 for (i = 0; i < lim; i++)
4839 register rtx y = XVECEXP (x, 0, i);
4840 if (GET_CODE (y) == SET)
4842 /* As above, we ignore unconditional jumps and call-insns and
4843 ignore the result of apply_change_group. */
4844 if (GET_CODE (SET_SRC (y)) == CALL)
4846 canon_reg (SET_SRC (y), insn);
4847 apply_change_group ();
4848 fold_rtx (SET_SRC (y), insn);
4849 invalidate (SET_DEST (y), VOIDmode);
4851 else if (SET_DEST (y) == pc_rtx
4852 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4854 else
4855 sets[n_sets++].rtl = y;
4857 else if (GET_CODE (y) == CLOBBER)
4859 /* If we clobber memory, canon the address.
4860 This does nothing when a register is clobbered
4861 because we have already invalidated the reg. */
4862 if (GET_CODE (XEXP (y, 0)) == MEM)
4863 canon_reg (XEXP (y, 0), NULL_RTX);
4865 else if (GET_CODE (y) == USE
4866 && ! (GET_CODE (XEXP (y, 0)) == REG
4867 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4868 canon_reg (y, NULL_RTX);
4869 else if (GET_CODE (y) == CALL)
4871 /* The result of apply_change_group can be ignored; see
4872 canon_reg. */
4873 canon_reg (y, insn);
4874 apply_change_group ();
4875 fold_rtx (y, insn);
4879 else if (GET_CODE (x) == CLOBBER)
4881 if (GET_CODE (XEXP (x, 0)) == MEM)
4882 canon_reg (XEXP (x, 0), NULL_RTX);
4885 /* Canonicalize a USE of a pseudo register or memory location. */
4886 else if (GET_CODE (x) == USE
4887 && ! (GET_CODE (XEXP (x, 0)) == REG
4888 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4889 canon_reg (XEXP (x, 0), NULL_RTX);
4890 else if (GET_CODE (x) == CALL)
4892 /* The result of apply_change_group can be ignored; see canon_reg. */
4893 canon_reg (x, insn);
4894 apply_change_group ();
4895 fold_rtx (x, insn);
4898 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4899 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4900 is handled specially for this case, and if it isn't set, then there will
4901 be no equivalence for the destination. */
4902 if (n_sets == 1 && REG_NOTES (insn) != 0
4903 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4904 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4905 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4906 src_eqv = canon_reg (XEXP (tem, 0), NULL_RTX);
4908 /* Canonicalize sources and addresses of destinations.
4909 We do this in a separate pass to avoid problems when a MATCH_DUP is
4910 present in the insn pattern. In that case, we want to ensure that
4911 we don't break the duplicate nature of the pattern. So we will replace
4912 both operands at the same time. Otherwise, we would fail to find an
4913 equivalent substitution in the loop calling validate_change below.
4915 We used to suppress canonicalization of DEST if it appears in SRC,
4916 but we don't do this any more. */
4918 for (i = 0; i < n_sets; i++)
4920 rtx dest = SET_DEST (sets[i].rtl);
4921 rtx src = SET_SRC (sets[i].rtl);
4922 rtx new = canon_reg (src, insn);
4923 int insn_code;
4925 sets[i].orig_src = src;
4926 if ((GET_CODE (new) == REG && GET_CODE (src) == REG
4927 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4928 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4929 || (insn_code = recog_memoized (insn)) < 0
4930 || insn_data[insn_code].n_dups > 0)
4931 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4932 else
4933 SET_SRC (sets[i].rtl) = new;
4935 if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT)
4937 validate_change (insn, &XEXP (dest, 1),
4938 canon_reg (XEXP (dest, 1), insn), 1);
4939 validate_change (insn, &XEXP (dest, 2),
4940 canon_reg (XEXP (dest, 2), insn), 1);
4943 while (GET_CODE (dest) == SUBREG || GET_CODE (dest) == STRICT_LOW_PART
4944 || GET_CODE (dest) == ZERO_EXTRACT
4945 || GET_CODE (dest) == SIGN_EXTRACT)
4946 dest = XEXP (dest, 0);
4948 if (GET_CODE (dest) == MEM)
4949 canon_reg (dest, insn);
4952 /* Now that we have done all the replacements, we can apply the change
4953 group and see if they all work. Note that this will cause some
4954 canonicalizations that would have worked individually not to be applied
4955 because some other canonicalization didn't work, but this should not
4956 occur often.
4958 The result of apply_change_group can be ignored; see canon_reg. */
4960 apply_change_group ();
4962 /* Set sets[i].src_elt to the class each source belongs to.
4963 Detect assignments from or to volatile things
4964 and set set[i] to zero so they will be ignored
4965 in the rest of this function.
4967 Nothing in this loop changes the hash table or the register chains. */
4969 for (i = 0; i < n_sets; i++)
4971 register rtx src, dest;
4972 register rtx src_folded;
4973 register struct table_elt *elt = 0, *p;
4974 enum machine_mode mode;
4975 rtx src_eqv_here;
4976 rtx src_const = 0;
4977 rtx src_related = 0;
4978 struct table_elt *src_const_elt = 0;
4979 int src_cost = MAX_COST;
4980 int src_eqv_cost = MAX_COST;
4981 int src_folded_cost = MAX_COST;
4982 int src_related_cost = MAX_COST;
4983 int src_elt_cost = MAX_COST;
4984 int src_regcost = MAX_COST;
4985 int src_eqv_regcost = MAX_COST;
4986 int src_folded_regcost = MAX_COST;
4987 int src_related_regcost = MAX_COST;
4988 int src_elt_regcost = MAX_COST;
4989 /* Set non-zero if we need to call force_const_mem on with the
4990 contents of src_folded before using it. */
4991 int src_folded_force_flag = 0;
4993 dest = SET_DEST (sets[i].rtl);
4994 src = SET_SRC (sets[i].rtl);
4996 /* If SRC is a constant that has no machine mode,
4997 hash it with the destination's machine mode.
4998 This way we can keep different modes separate. */
5000 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5001 sets[i].mode = mode;
5003 if (src_eqv)
5005 enum machine_mode eqvmode = mode;
5006 if (GET_CODE (dest) == STRICT_LOW_PART)
5007 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5008 do_not_record = 0;
5009 hash_arg_in_memory = 0;
5010 src_eqv = fold_rtx (src_eqv, insn);
5011 src_eqv_hash = HASH (src_eqv, eqvmode);
5013 /* Find the equivalence class for the equivalent expression. */
5015 if (!do_not_record)
5016 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
5018 src_eqv_volatile = do_not_record;
5019 src_eqv_in_memory = hash_arg_in_memory;
5022 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
5023 value of the INNER register, not the destination. So it is not
5024 a valid substitution for the source. But save it for later. */
5025 if (GET_CODE (dest) == STRICT_LOW_PART)
5026 src_eqv_here = 0;
5027 else
5028 src_eqv_here = src_eqv;
5030 /* Simplify and foldable subexpressions in SRC. Then get the fully-
5031 simplified result, which may not necessarily be valid. */
5032 src_folded = fold_rtx (src, insn);
5034 #if 0
5035 /* ??? This caused bad code to be generated for the m68k port with -O2.
5036 Suppose src is (CONST_INT -1), and that after truncation src_folded
5037 is (CONST_INT 3). Suppose src_folded is then used for src_const.
5038 At the end we will add src and src_const to the same equivalence
5039 class. We now have 3 and -1 on the same equivalence class. This
5040 causes later instructions to be mis-optimized. */
5041 /* If storing a constant in a bitfield, pre-truncate the constant
5042 so we will be able to record it later. */
5043 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5044 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
5046 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5048 if (GET_CODE (src) == CONST_INT
5049 && GET_CODE (width) == CONST_INT
5050 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5051 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5052 src_folded
5053 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5054 << INTVAL (width)) - 1));
5056 #endif
5058 /* Compute SRC's hash code, and also notice if it
5059 should not be recorded at all. In that case,
5060 prevent any further processing of this assignment. */
5061 do_not_record = 0;
5062 hash_arg_in_memory = 0;
5064 sets[i].src = src;
5065 sets[i].src_hash = HASH (src, mode);
5066 sets[i].src_volatile = do_not_record;
5067 sets[i].src_in_memory = hash_arg_in_memory;
5069 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5070 a pseudo that is set more than once, do not record SRC. Using
5071 SRC as a replacement for anything else will be incorrect in that
5072 situation. Note that this usually occurs only for stack slots,
5073 in which case all the RTL would be referring to SRC, so we don't
5074 lose any optimization opportunities by not having SRC in the
5075 hash table. */
5077 if (GET_CODE (src) == MEM
5078 && find_reg_note (insn, REG_EQUIV, src) != 0
5079 && GET_CODE (dest) == REG
5080 && REGNO (dest) >= FIRST_PSEUDO_REGISTER
5081 && REG_N_SETS (REGNO (dest)) != 1)
5082 sets[i].src_volatile = 1;
5084 #if 0
5085 /* It is no longer clear why we used to do this, but it doesn't
5086 appear to still be needed. So let's try without it since this
5087 code hurts cse'ing widened ops. */
5088 /* If source is a perverse subreg (such as QI treated as an SI),
5089 treat it as volatile. It may do the work of an SI in one context
5090 where the extra bits are not being used, but cannot replace an SI
5091 in general. */
5092 if (GET_CODE (src) == SUBREG
5093 && (GET_MODE_SIZE (GET_MODE (src))
5094 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5095 sets[i].src_volatile = 1;
5096 #endif
5098 /* Locate all possible equivalent forms for SRC. Try to replace
5099 SRC in the insn with each cheaper equivalent.
5101 We have the following types of equivalents: SRC itself, a folded
5102 version, a value given in a REG_EQUAL note, or a value related
5103 to a constant.
5105 Each of these equivalents may be part of an additional class
5106 of equivalents (if more than one is in the table, they must be in
5107 the same class; we check for this).
5109 If the source is volatile, we don't do any table lookups.
5111 We note any constant equivalent for possible later use in a
5112 REG_NOTE. */
5114 if (!sets[i].src_volatile)
5115 elt = lookup (src, sets[i].src_hash, mode);
5117 sets[i].src_elt = elt;
5119 if (elt && src_eqv_here && src_eqv_elt)
5121 if (elt->first_same_value != src_eqv_elt->first_same_value)
5123 /* The REG_EQUAL is indicating that two formerly distinct
5124 classes are now equivalent. So merge them. */
5125 merge_equiv_classes (elt, src_eqv_elt);
5126 src_eqv_hash = HASH (src_eqv, elt->mode);
5127 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5130 src_eqv_here = 0;
5133 else if (src_eqv_elt)
5134 elt = src_eqv_elt;
5136 /* Try to find a constant somewhere and record it in `src_const'.
5137 Record its table element, if any, in `src_const_elt'. Look in
5138 any known equivalences first. (If the constant is not in the
5139 table, also set `sets[i].src_const_hash'). */
5140 if (elt)
5141 for (p = elt->first_same_value; p; p = p->next_same_value)
5142 if (p->is_const)
5144 src_const = p->exp;
5145 src_const_elt = elt;
5146 break;
5149 if (src_const == 0
5150 && (CONSTANT_P (src_folded)
5151 /* Consider (minus (label_ref L1) (label_ref L2)) as
5152 "constant" here so we will record it. This allows us
5153 to fold switch statements when an ADDR_DIFF_VEC is used. */
5154 || (GET_CODE (src_folded) == MINUS
5155 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5156 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5157 src_const = src_folded, src_const_elt = elt;
5158 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5159 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5161 /* If we don't know if the constant is in the table, get its
5162 hash code and look it up. */
5163 if (src_const && src_const_elt == 0)
5165 sets[i].src_const_hash = HASH (src_const, mode);
5166 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5169 sets[i].src_const = src_const;
5170 sets[i].src_const_elt = src_const_elt;
5172 /* If the constant and our source are both in the table, mark them as
5173 equivalent. Otherwise, if a constant is in the table but the source
5174 isn't, set ELT to it. */
5175 if (src_const_elt && elt
5176 && src_const_elt->first_same_value != elt->first_same_value)
5177 merge_equiv_classes (elt, src_const_elt);
5178 else if (src_const_elt && elt == 0)
5179 elt = src_const_elt;
5181 /* See if there is a register linearly related to a constant
5182 equivalent of SRC. */
5183 if (src_const
5184 && (GET_CODE (src_const) == CONST
5185 || (src_const_elt && src_const_elt->related_value != 0)))
5187 src_related = use_related_value (src_const, src_const_elt);
5188 if (src_related)
5190 struct table_elt *src_related_elt
5191 = lookup (src_related, HASH (src_related, mode), mode);
5192 if (src_related_elt && elt)
5194 if (elt->first_same_value
5195 != src_related_elt->first_same_value)
5196 /* This can occur when we previously saw a CONST
5197 involving a SYMBOL_REF and then see the SYMBOL_REF
5198 twice. Merge the involved classes. */
5199 merge_equiv_classes (elt, src_related_elt);
5201 src_related = 0;
5202 src_related_elt = 0;
5204 else if (src_related_elt && elt == 0)
5205 elt = src_related_elt;
5209 /* See if we have a CONST_INT that is already in a register in a
5210 wider mode. */
5212 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5213 && GET_MODE_CLASS (mode) == MODE_INT
5214 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5216 enum machine_mode wider_mode;
5218 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5219 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5220 && src_related == 0;
5221 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5223 struct table_elt *const_elt
5224 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5226 if (const_elt == 0)
5227 continue;
5229 for (const_elt = const_elt->first_same_value;
5230 const_elt; const_elt = const_elt->next_same_value)
5231 if (GET_CODE (const_elt->exp) == REG)
5233 src_related = gen_lowpart_if_possible (mode,
5234 const_elt->exp);
5235 break;
5240 /* Another possibility is that we have an AND with a constant in
5241 a mode narrower than a word. If so, it might have been generated
5242 as part of an "if" which would narrow the AND. If we already
5243 have done the AND in a wider mode, we can use a SUBREG of that
5244 value. */
5246 if (flag_expensive_optimizations && ! src_related
5247 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5248 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5250 enum machine_mode tmode;
5251 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5253 for (tmode = GET_MODE_WIDER_MODE (mode);
5254 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5255 tmode = GET_MODE_WIDER_MODE (tmode))
5257 rtx inner = gen_lowpart_if_possible (tmode, XEXP (src, 0));
5258 struct table_elt *larger_elt;
5260 if (inner)
5262 PUT_MODE (new_and, tmode);
5263 XEXP (new_and, 0) = inner;
5264 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5265 if (larger_elt == 0)
5266 continue;
5268 for (larger_elt = larger_elt->first_same_value;
5269 larger_elt; larger_elt = larger_elt->next_same_value)
5270 if (GET_CODE (larger_elt->exp) == REG)
5272 src_related
5273 = gen_lowpart_if_possible (mode, larger_elt->exp);
5274 break;
5277 if (src_related)
5278 break;
5283 #ifdef LOAD_EXTEND_OP
5284 /* See if a MEM has already been loaded with a widening operation;
5285 if it has, we can use a subreg of that. Many CISC machines
5286 also have such operations, but this is only likely to be
5287 beneficial these machines. */
5289 if (flag_expensive_optimizations && src_related == 0
5290 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5291 && GET_MODE_CLASS (mode) == MODE_INT
5292 && GET_CODE (src) == MEM && ! do_not_record
5293 && LOAD_EXTEND_OP (mode) != NIL)
5295 enum machine_mode tmode;
5297 /* Set what we are trying to extend and the operation it might
5298 have been extended with. */
5299 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5300 XEXP (memory_extend_rtx, 0) = src;
5302 for (tmode = GET_MODE_WIDER_MODE (mode);
5303 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5304 tmode = GET_MODE_WIDER_MODE (tmode))
5306 struct table_elt *larger_elt;
5308 PUT_MODE (memory_extend_rtx, tmode);
5309 larger_elt = lookup (memory_extend_rtx,
5310 HASH (memory_extend_rtx, tmode), tmode);
5311 if (larger_elt == 0)
5312 continue;
5314 for (larger_elt = larger_elt->first_same_value;
5315 larger_elt; larger_elt = larger_elt->next_same_value)
5316 if (GET_CODE (larger_elt->exp) == REG)
5318 src_related = gen_lowpart_if_possible (mode,
5319 larger_elt->exp);
5320 break;
5323 if (src_related)
5324 break;
5327 #endif /* LOAD_EXTEND_OP */
5329 if (src == src_folded)
5330 src_folded = 0;
5332 /* At this point, ELT, if non-zero, points to a class of expressions
5333 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5334 and SRC_RELATED, if non-zero, each contain additional equivalent
5335 expressions. Prune these latter expressions by deleting expressions
5336 already in the equivalence class.
5338 Check for an equivalent identical to the destination. If found,
5339 this is the preferred equivalent since it will likely lead to
5340 elimination of the insn. Indicate this by placing it in
5341 `src_related'. */
5343 if (elt)
5344 elt = elt->first_same_value;
5345 for (p = elt; p; p = p->next_same_value)
5347 enum rtx_code code = GET_CODE (p->exp);
5349 /* If the expression is not valid, ignore it. Then we do not
5350 have to check for validity below. In most cases, we can use
5351 `rtx_equal_p', since canonicalization has already been done. */
5352 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, 0))
5353 continue;
5355 /* Also skip paradoxical subregs, unless that's what we're
5356 looking for. */
5357 if (code == SUBREG
5358 && (GET_MODE_SIZE (GET_MODE (p->exp))
5359 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5360 && ! (src != 0
5361 && GET_CODE (src) == SUBREG
5362 && GET_MODE (src) == GET_MODE (p->exp)
5363 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5364 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5365 continue;
5367 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5368 src = 0;
5369 else if (src_folded && GET_CODE (src_folded) == code
5370 && rtx_equal_p (src_folded, p->exp))
5371 src_folded = 0;
5372 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5373 && rtx_equal_p (src_eqv_here, p->exp))
5374 src_eqv_here = 0;
5375 else if (src_related && GET_CODE (src_related) == code
5376 && rtx_equal_p (src_related, p->exp))
5377 src_related = 0;
5379 /* This is the same as the destination of the insns, we want
5380 to prefer it. Copy it to src_related. The code below will
5381 then give it a negative cost. */
5382 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5383 src_related = dest;
5386 /* Find the cheapest valid equivalent, trying all the available
5387 possibilities. Prefer items not in the hash table to ones
5388 that are when they are equal cost. Note that we can never
5389 worsen an insn as the current contents will also succeed.
5390 If we find an equivalent identical to the destination, use it as best,
5391 since this insn will probably be eliminated in that case. */
5392 if (src)
5394 if (rtx_equal_p (src, dest))
5395 src_cost = src_regcost = -1;
5396 else
5398 src_cost = COST (src);
5399 src_regcost = approx_reg_cost (src);
5403 if (src_eqv_here)
5405 if (rtx_equal_p (src_eqv_here, dest))
5406 src_eqv_cost = src_eqv_regcost = -1;
5407 else
5409 src_eqv_cost = COST (src_eqv_here);
5410 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5414 if (src_folded)
5416 if (rtx_equal_p (src_folded, dest))
5417 src_folded_cost = src_folded_regcost = -1;
5418 else
5420 src_folded_cost = COST (src_folded);
5421 src_folded_regcost = approx_reg_cost (src_folded);
5425 if (src_related)
5427 if (rtx_equal_p (src_related, dest))
5428 src_related_cost = src_related_regcost = -1;
5429 else
5431 src_related_cost = COST (src_related);
5432 src_related_regcost = approx_reg_cost (src_related);
5436 /* If this was an indirect jump insn, a known label will really be
5437 cheaper even though it looks more expensive. */
5438 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5439 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5441 /* Terminate loop when replacement made. This must terminate since
5442 the current contents will be tested and will always be valid. */
5443 while (1)
5445 rtx trial;
5447 /* Skip invalid entries. */
5448 while (elt && GET_CODE (elt->exp) != REG
5449 && ! exp_equiv_p (elt->exp, elt->exp, 1, 0))
5450 elt = elt->next_same_value;
5452 /* A paradoxical subreg would be bad here: it'll be the right
5453 size, but later may be adjusted so that the upper bits aren't
5454 what we want. So reject it. */
5455 if (elt != 0
5456 && GET_CODE (elt->exp) == SUBREG
5457 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5458 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5459 /* It is okay, though, if the rtx we're trying to match
5460 will ignore any of the bits we can't predict. */
5461 && ! (src != 0
5462 && GET_CODE (src) == SUBREG
5463 && GET_MODE (src) == GET_MODE (elt->exp)
5464 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5465 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5467 elt = elt->next_same_value;
5468 continue;
5471 if (elt)
5473 src_elt_cost = elt->cost;
5474 src_elt_regcost = elt->regcost;
5477 /* Find cheapest and skip it for the next time. For items
5478 of equal cost, use this order:
5479 src_folded, src, src_eqv, src_related and hash table entry. */
5480 if (src_folded
5481 && preferrable (src_folded_cost, src_folded_regcost,
5482 src_cost, src_regcost) <= 0
5483 && preferrable (src_folded_cost, src_folded_regcost,
5484 src_eqv_cost, src_eqv_regcost) <= 0
5485 && preferrable (src_folded_cost, src_folded_regcost,
5486 src_related_cost, src_related_regcost) <= 0
5487 && preferrable (src_folded_cost, src_folded_regcost,
5488 src_elt_cost, src_elt_regcost) <= 0)
5490 trial = src_folded, src_folded_cost = MAX_COST;
5491 if (src_folded_force_flag)
5492 trial = force_const_mem (mode, trial);
5494 else if (src
5495 && preferrable (src_cost, src_regcost,
5496 src_eqv_cost, src_eqv_regcost) <= 0
5497 && preferrable (src_cost, src_regcost,
5498 src_related_cost, src_related_regcost) <= 0
5499 && preferrable (src_cost, src_regcost,
5500 src_elt_cost, src_elt_regcost) <= 0)
5501 trial = src, src_cost = MAX_COST;
5502 else if (src_eqv_here
5503 && preferrable (src_eqv_cost, src_eqv_regcost,
5504 src_related_cost, src_related_regcost) <= 0
5505 && preferrable (src_eqv_cost, src_eqv_regcost,
5506 src_elt_cost, src_elt_regcost) <= 0)
5507 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5508 else if (src_related
5509 && preferrable (src_related_cost, src_related_regcost,
5510 src_elt_cost, src_elt_regcost) <= 0)
5511 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5512 else
5514 trial = copy_rtx (elt->exp);
5515 elt = elt->next_same_value;
5516 src_elt_cost = MAX_COST;
5519 /* We don't normally have an insn matching (set (pc) (pc)), so
5520 check for this separately here. We will delete such an
5521 insn below.
5523 Tablejump insns contain a USE of the table, so simply replacing
5524 the operand with the constant won't match. This is simply an
5525 unconditional branch, however, and is therefore valid. Just
5526 insert the substitution here and we will delete and re-emit
5527 the insn later. */
5529 if (n_sets == 1 && dest == pc_rtx
5530 && (trial == pc_rtx
5531 || (GET_CODE (trial) == LABEL_REF
5532 && ! condjump_p (insn))))
5534 if (trial == pc_rtx)
5536 SET_SRC (sets[i].rtl) = trial;
5537 cse_jumps_altered = 1;
5538 break;
5541 PATTERN (insn) = gen_jump (XEXP (trial, 0));
5542 INSN_CODE (insn) = -1;
5544 if (NEXT_INSN (insn) != 0
5545 && GET_CODE (NEXT_INSN (insn)) != BARRIER)
5546 emit_barrier_after (insn);
5548 cse_jumps_altered = 1;
5549 break;
5552 /* Look for a substitution that makes a valid insn. */
5553 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5555 /* If we just made a substitution inside a libcall, then we
5556 need to make the same substitution in any notes attached
5557 to the RETVAL insn. */
5558 if (libcall_insn
5559 && (GET_CODE (sets[i].orig_src) == REG
5560 || GET_CODE (sets[i].orig_src) == SUBREG
5561 || GET_CODE (sets[i].orig_src) == MEM))
5562 replace_rtx (REG_NOTES (libcall_insn), sets[i].orig_src,
5563 canon_reg (SET_SRC (sets[i].rtl), insn));
5565 /* The result of apply_change_group can be ignored; see
5566 canon_reg. */
5568 validate_change (insn, &SET_SRC (sets[i].rtl),
5569 canon_reg (SET_SRC (sets[i].rtl), insn),
5571 apply_change_group ();
5572 break;
5575 /* If we previously found constant pool entries for
5576 constants and this is a constant, try making a
5577 pool entry. Put it in src_folded unless we already have done
5578 this since that is where it likely came from. */
5580 else if (constant_pool_entries_cost
5581 && CONSTANT_P (trial)
5582 /* Reject cases that will abort in decode_rtx_const.
5583 On the alpha when simplifying a switch, we get
5584 (const (truncate (minus (label_ref) (label_ref)))). */
5585 && ! (GET_CODE (trial) == CONST
5586 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5587 /* Likewise on IA-64, except without the truncate. */
5588 && ! (GET_CODE (trial) == CONST
5589 && GET_CODE (XEXP (trial, 0)) == MINUS
5590 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5591 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5592 && (src_folded == 0
5593 || (GET_CODE (src_folded) != MEM
5594 && ! src_folded_force_flag))
5595 && GET_MODE_CLASS (mode) != MODE_CC
5596 && mode != VOIDmode)
5598 src_folded_force_flag = 1;
5599 src_folded = trial;
5600 src_folded_cost = constant_pool_entries_cost;
5604 src = SET_SRC (sets[i].rtl);
5606 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5607 However, there is an important exception: If both are registers
5608 that are not the head of their equivalence class, replace SET_SRC
5609 with the head of the class. If we do not do this, we will have
5610 both registers live over a portion of the basic block. This way,
5611 their lifetimes will likely abut instead of overlapping. */
5612 if (GET_CODE (dest) == REG
5613 && REGNO_QTY_VALID_P (REGNO (dest)))
5615 int dest_q = REG_QTY (REGNO (dest));
5616 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5618 if (dest_ent->mode == GET_MODE (dest)
5619 && dest_ent->first_reg != REGNO (dest)
5620 && GET_CODE (src) == REG && REGNO (src) == REGNO (dest)
5621 /* Don't do this if the original insn had a hard reg as
5622 SET_SRC or SET_DEST. */
5623 && (GET_CODE (sets[i].src) != REG
5624 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5625 && (GET_CODE (dest) != REG || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5626 /* We can't call canon_reg here because it won't do anything if
5627 SRC is a hard register. */
5629 int src_q = REG_QTY (REGNO (src));
5630 struct qty_table_elem *src_ent = &qty_table[src_q];
5631 int first = src_ent->first_reg;
5632 rtx new_src
5633 = (first >= FIRST_PSEUDO_REGISTER
5634 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5636 /* We must use validate-change even for this, because this
5637 might be a special no-op instruction, suitable only to
5638 tag notes onto. */
5639 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5641 src = new_src;
5642 /* If we had a constant that is cheaper than what we are now
5643 setting SRC to, use that constant. We ignored it when we
5644 thought we could make this into a no-op. */
5645 if (src_const && COST (src_const) < COST (src)
5646 && validate_change (insn, &SET_SRC (sets[i].rtl),
5647 src_const, 0))
5648 src = src_const;
5653 /* If we made a change, recompute SRC values. */
5654 if (src != sets[i].src)
5656 cse_altered = 1;
5657 do_not_record = 0;
5658 hash_arg_in_memory = 0;
5659 sets[i].src = src;
5660 sets[i].src_hash = HASH (src, mode);
5661 sets[i].src_volatile = do_not_record;
5662 sets[i].src_in_memory = hash_arg_in_memory;
5663 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5666 /* If this is a single SET, we are setting a register, and we have an
5667 equivalent constant, we want to add a REG_NOTE. We don't want
5668 to write a REG_EQUAL note for a constant pseudo since verifying that
5669 that pseudo hasn't been eliminated is a pain. Such a note also
5670 won't help anything.
5672 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5673 which can be created for a reference to a compile time computable
5674 entry in a jump table. */
5676 if (n_sets == 1 && src_const && GET_CODE (dest) == REG
5677 && GET_CODE (src_const) != REG
5678 && ! (GET_CODE (src_const) == CONST
5679 && GET_CODE (XEXP (src_const, 0)) == MINUS
5680 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5681 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5683 tem = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5685 /* Make sure that the rtx is not shared with any other insn. */
5686 src_const = copy_rtx (src_const);
5688 /* Record the actual constant value in a REG_EQUAL note, making
5689 a new one if one does not already exist. */
5690 if (tem)
5691 XEXP (tem, 0) = src_const;
5692 else
5693 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL,
5694 src_const, REG_NOTES (insn));
5696 /* If storing a constant value in a register that
5697 previously held the constant value 0,
5698 record this fact with a REG_WAS_0 note on this insn.
5700 Note that the *register* is required to have previously held 0,
5701 not just any register in the quantity and we must point to the
5702 insn that set that register to zero.
5704 Rather than track each register individually, we just see if
5705 the last set for this quantity was for this register. */
5707 if (REGNO_QTY_VALID_P (REGNO (dest)))
5709 int dest_q = REG_QTY (REGNO (dest));
5710 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5712 if (dest_ent->const_rtx == const0_rtx)
5714 /* See if we previously had a REG_WAS_0 note. */
5715 rtx note = find_reg_note (insn, REG_WAS_0, NULL_RTX);
5716 rtx const_insn = dest_ent->const_insn;
5718 if ((tem = single_set (const_insn)) != 0
5719 && rtx_equal_p (SET_DEST (tem), dest))
5721 if (note)
5722 XEXP (note, 0) = const_insn;
5723 else
5724 REG_NOTES (insn)
5725 = gen_rtx_INSN_LIST (REG_WAS_0, const_insn,
5726 REG_NOTES (insn));
5732 /* Now deal with the destination. */
5733 do_not_record = 0;
5735 /* Look within any SIGN_EXTRACT or ZERO_EXTRACT
5736 to the MEM or REG within it. */
5737 while (GET_CODE (dest) == SIGN_EXTRACT
5738 || GET_CODE (dest) == ZERO_EXTRACT
5739 || GET_CODE (dest) == SUBREG
5740 || GET_CODE (dest) == STRICT_LOW_PART)
5741 dest = XEXP (dest, 0);
5743 sets[i].inner_dest = dest;
5745 if (GET_CODE (dest) == MEM)
5747 #ifdef PUSH_ROUNDING
5748 /* Stack pushes invalidate the stack pointer. */
5749 rtx addr = XEXP (dest, 0);
5750 if (GET_RTX_CLASS (GET_CODE (addr)) == 'a'
5751 && XEXP (addr, 0) == stack_pointer_rtx)
5752 invalidate (stack_pointer_rtx, Pmode);
5753 #endif
5754 dest = fold_rtx (dest, insn);
5757 /* Compute the hash code of the destination now,
5758 before the effects of this instruction are recorded,
5759 since the register values used in the address computation
5760 are those before this instruction. */
5761 sets[i].dest_hash = HASH (dest, mode);
5763 /* Don't enter a bit-field in the hash table
5764 because the value in it after the store
5765 may not equal what was stored, due to truncation. */
5767 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5768 || GET_CODE (SET_DEST (sets[i].rtl)) == SIGN_EXTRACT)
5770 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5772 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5773 && GET_CODE (width) == CONST_INT
5774 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5775 && ! (INTVAL (src_const)
5776 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5777 /* Exception: if the value is constant,
5778 and it won't be truncated, record it. */
5780 else
5782 /* This is chosen so that the destination will be invalidated
5783 but no new value will be recorded.
5784 We must invalidate because sometimes constant
5785 values can be recorded for bitfields. */
5786 sets[i].src_elt = 0;
5787 sets[i].src_volatile = 1;
5788 src_eqv = 0;
5789 src_eqv_elt = 0;
5793 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5794 the insn. */
5795 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5797 /* One less use of the label this insn used to jump to. */
5798 if (JUMP_LABEL (insn) != 0)
5799 --LABEL_NUSES (JUMP_LABEL (insn));
5800 PUT_CODE (insn, NOTE);
5801 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
5802 NOTE_SOURCE_FILE (insn) = 0;
5803 cse_jumps_altered = 1;
5804 /* No more processing for this set. */
5805 sets[i].rtl = 0;
5808 /* If this SET is now setting PC to a label, we know it used to
5809 be a conditional or computed branch. So we see if we can follow
5810 it. If it was a computed branch, delete it and re-emit. */
5811 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF)
5813 /* If this is not in the format for a simple branch and
5814 we are the only SET in it, re-emit it. */
5815 if (! simplejump_p (insn) && n_sets == 1)
5817 rtx new = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5818 JUMP_LABEL (new) = XEXP (src, 0);
5819 LABEL_NUSES (XEXP (src, 0))++;
5820 insn = new;
5822 else
5823 /* Otherwise, force rerecognition, since it probably had
5824 a different pattern before.
5825 This shouldn't really be necessary, since whatever
5826 changed the source value above should have done this.
5827 Until the right place is found, might as well do this here. */
5828 INSN_CODE (insn) = -1;
5830 never_reached_warning (insn);
5832 /* Now emit a BARRIER after the unconditional jump. Do not bother
5833 deleting any unreachable code, let jump/flow do that. */
5834 if (NEXT_INSN (insn) != 0
5835 && GET_CODE (NEXT_INSN (insn)) != BARRIER)
5836 emit_barrier_after (insn);
5838 cse_jumps_altered = 1;
5839 sets[i].rtl = 0;
5842 /* If destination is volatile, invalidate it and then do no further
5843 processing for this assignment. */
5845 else if (do_not_record)
5847 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
5848 invalidate (dest, VOIDmode);
5849 else if (GET_CODE (dest) == MEM)
5851 /* Outgoing arguments for a libcall don't
5852 affect any recorded expressions. */
5853 if (! libcall_insn || insn == libcall_insn)
5854 invalidate (dest, VOIDmode);
5856 else if (GET_CODE (dest) == STRICT_LOW_PART
5857 || GET_CODE (dest) == ZERO_EXTRACT)
5858 invalidate (XEXP (dest, 0), GET_MODE (dest));
5859 sets[i].rtl = 0;
5862 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5863 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5865 #ifdef HAVE_cc0
5866 /* If setting CC0, record what it was set to, or a constant, if it
5867 is equivalent to a constant. If it is being set to a floating-point
5868 value, make a COMPARE with the appropriate constant of 0. If we
5869 don't do this, later code can interpret this as a test against
5870 const0_rtx, which can cause problems if we try to put it into an
5871 insn as a floating-point operand. */
5872 if (dest == cc0_rtx)
5874 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5875 this_insn_cc0_mode = mode;
5876 if (FLOAT_MODE_P (mode))
5877 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5878 CONST0_RTX (mode));
5880 #endif
5883 /* Now enter all non-volatile source expressions in the hash table
5884 if they are not already present.
5885 Record their equivalence classes in src_elt.
5886 This way we can insert the corresponding destinations into
5887 the same classes even if the actual sources are no longer in them
5888 (having been invalidated). */
5890 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5891 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5893 register struct table_elt *elt;
5894 register struct table_elt *classp = sets[0].src_elt;
5895 rtx dest = SET_DEST (sets[0].rtl);
5896 enum machine_mode eqvmode = GET_MODE (dest);
5898 if (GET_CODE (dest) == STRICT_LOW_PART)
5900 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5901 classp = 0;
5903 if (insert_regs (src_eqv, classp, 0))
5905 rehash_using_reg (src_eqv);
5906 src_eqv_hash = HASH (src_eqv, eqvmode);
5908 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5909 elt->in_memory = src_eqv_in_memory;
5910 src_eqv_elt = elt;
5912 /* Check to see if src_eqv_elt is the same as a set source which
5913 does not yet have an elt, and if so set the elt of the set source
5914 to src_eqv_elt. */
5915 for (i = 0; i < n_sets; i++)
5916 if (sets[i].rtl && sets[i].src_elt == 0
5917 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5918 sets[i].src_elt = src_eqv_elt;
5921 for (i = 0; i < n_sets; i++)
5922 if (sets[i].rtl && ! sets[i].src_volatile
5923 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5925 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5927 /* REG_EQUAL in setting a STRICT_LOW_PART
5928 gives an equivalent for the entire destination register,
5929 not just for the subreg being stored in now.
5930 This is a more interesting equivalence, so we arrange later
5931 to treat the entire reg as the destination. */
5932 sets[i].src_elt = src_eqv_elt;
5933 sets[i].src_hash = src_eqv_hash;
5935 else
5937 /* Insert source and constant equivalent into hash table, if not
5938 already present. */
5939 register struct table_elt *classp = src_eqv_elt;
5940 register rtx src = sets[i].src;
5941 register rtx dest = SET_DEST (sets[i].rtl);
5942 enum machine_mode mode
5943 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5945 if (sets[i].src_elt == 0)
5947 /* Don't put a hard register source into the table if this is
5948 the last insn of a libcall. In this case, we only need
5949 to put src_eqv_elt in src_elt. */
5950 if (GET_CODE (src) != REG
5951 || REGNO (src) >= FIRST_PSEUDO_REGISTER
5952 || ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5954 register struct table_elt *elt;
5956 /* Note that these insert_regs calls cannot remove
5957 any of the src_elt's, because they would have failed to
5958 match if not still valid. */
5959 if (insert_regs (src, classp, 0))
5961 rehash_using_reg (src);
5962 sets[i].src_hash = HASH (src, mode);
5964 elt = insert (src, classp, sets[i].src_hash, mode);
5965 elt->in_memory = sets[i].src_in_memory;
5966 sets[i].src_elt = classp = elt;
5968 else
5969 sets[i].src_elt = classp;
5971 if (sets[i].src_const && sets[i].src_const_elt == 0
5972 && src != sets[i].src_const
5973 && ! rtx_equal_p (sets[i].src_const, src))
5974 sets[i].src_elt = insert (sets[i].src_const, classp,
5975 sets[i].src_const_hash, mode);
5978 else if (sets[i].src_elt == 0)
5979 /* If we did not insert the source into the hash table (e.g., it was
5980 volatile), note the equivalence class for the REG_EQUAL value, if any,
5981 so that the destination goes into that class. */
5982 sets[i].src_elt = src_eqv_elt;
5984 invalidate_from_clobbers (x);
5986 /* Some registers are invalidated by subroutine calls. Memory is
5987 invalidated by non-constant calls. */
5989 if (GET_CODE (insn) == CALL_INSN)
5991 if (! CONST_CALL_P (insn))
5992 invalidate_memory ();
5993 invalidate_for_call ();
5996 /* Now invalidate everything set by this instruction.
5997 If a SUBREG or other funny destination is being set,
5998 sets[i].rtl is still nonzero, so here we invalidate the reg
5999 a part of which is being set. */
6001 for (i = 0; i < n_sets; i++)
6002 if (sets[i].rtl)
6004 /* We can't use the inner dest, because the mode associated with
6005 a ZERO_EXTRACT is significant. */
6006 register rtx dest = SET_DEST (sets[i].rtl);
6008 /* Needed for registers to remove the register from its
6009 previous quantity's chain.
6010 Needed for memory if this is a nonvarying address, unless
6011 we have just done an invalidate_memory that covers even those. */
6012 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
6013 invalidate (dest, VOIDmode);
6014 else if (GET_CODE (dest) == MEM)
6016 /* Outgoing arguments for a libcall don't
6017 affect any recorded expressions. */
6018 if (! libcall_insn || insn == libcall_insn)
6019 invalidate (dest, VOIDmode);
6021 else if (GET_CODE (dest) == STRICT_LOW_PART
6022 || GET_CODE (dest) == ZERO_EXTRACT)
6023 invalidate (XEXP (dest, 0), GET_MODE (dest));
6026 /* A volatile ASM invalidates everything. */
6027 if (GET_CODE (insn) == INSN
6028 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
6029 && MEM_VOLATILE_P (PATTERN (insn)))
6030 flush_hash_table ();
6032 /* Make sure registers mentioned in destinations
6033 are safe for use in an expression to be inserted.
6034 This removes from the hash table
6035 any invalid entry that refers to one of these registers.
6037 We don't care about the return value from mention_regs because
6038 we are going to hash the SET_DEST values unconditionally. */
6040 for (i = 0; i < n_sets; i++)
6042 if (sets[i].rtl)
6044 rtx x = SET_DEST (sets[i].rtl);
6046 if (GET_CODE (x) != REG)
6047 mention_regs (x);
6048 else
6050 /* We used to rely on all references to a register becoming
6051 inaccessible when a register changes to a new quantity,
6052 since that changes the hash code. However, that is not
6053 safe, since after HASH_SIZE new quantities we get a
6054 hash 'collision' of a register with its own invalid
6055 entries. And since SUBREGs have been changed not to
6056 change their hash code with the hash code of the register,
6057 it wouldn't work any longer at all. So we have to check
6058 for any invalid references lying around now.
6059 This code is similar to the REG case in mention_regs,
6060 but it knows that reg_tick has been incremented, and
6061 it leaves reg_in_table as -1 . */
6062 unsigned int regno = REGNO (x);
6063 unsigned int endregno
6064 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
6065 : HARD_REGNO_NREGS (regno, GET_MODE (x)));
6066 unsigned int i;
6068 for (i = regno; i < endregno; i++)
6070 if (REG_IN_TABLE (i) >= 0)
6072 remove_invalid_refs (i);
6073 REG_IN_TABLE (i) = -1;
6080 /* We may have just removed some of the src_elt's from the hash table.
6081 So replace each one with the current head of the same class. */
6083 for (i = 0; i < n_sets; i++)
6084 if (sets[i].rtl)
6086 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6087 /* If elt was removed, find current head of same class,
6088 or 0 if nothing remains of that class. */
6090 register struct table_elt *elt = sets[i].src_elt;
6092 while (elt && elt->prev_same_value)
6093 elt = elt->prev_same_value;
6095 while (elt && elt->first_same_value == 0)
6096 elt = elt->next_same_value;
6097 sets[i].src_elt = elt ? elt->first_same_value : 0;
6101 /* Now insert the destinations into their equivalence classes. */
6103 for (i = 0; i < n_sets; i++)
6104 if (sets[i].rtl)
6106 register rtx dest = SET_DEST (sets[i].rtl);
6107 rtx inner_dest = sets[i].inner_dest;
6108 register struct table_elt *elt;
6110 /* Don't record value if we are not supposed to risk allocating
6111 floating-point values in registers that might be wider than
6112 memory. */
6113 if ((flag_float_store
6114 && GET_CODE (dest) == MEM
6115 && FLOAT_MODE_P (GET_MODE (dest)))
6116 /* Don't record BLKmode values, because we don't know the
6117 size of it, and can't be sure that other BLKmode values
6118 have the same or smaller size. */
6119 || GET_MODE (dest) == BLKmode
6120 /* Don't record values of destinations set inside a libcall block
6121 since we might delete the libcall. Things should have been set
6122 up so we won't want to reuse such a value, but we play it safe
6123 here. */
6124 || libcall_insn
6125 /* If we didn't put a REG_EQUAL value or a source into the hash
6126 table, there is no point is recording DEST. */
6127 || sets[i].src_elt == 0
6128 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6129 or SIGN_EXTEND, don't record DEST since it can cause
6130 some tracking to be wrong.
6132 ??? Think about this more later. */
6133 || (GET_CODE (dest) == SUBREG
6134 && (GET_MODE_SIZE (GET_MODE (dest))
6135 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6136 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6137 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6138 continue;
6140 /* STRICT_LOW_PART isn't part of the value BEING set,
6141 and neither is the SUBREG inside it.
6142 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6143 if (GET_CODE (dest) == STRICT_LOW_PART)
6144 dest = SUBREG_REG (XEXP (dest, 0));
6146 if (GET_CODE (dest) == REG || GET_CODE (dest) == SUBREG)
6147 /* Registers must also be inserted into chains for quantities. */
6148 if (insert_regs (dest, sets[i].src_elt, 1))
6150 /* If `insert_regs' changes something, the hash code must be
6151 recalculated. */
6152 rehash_using_reg (dest);
6153 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6156 if (GET_CODE (inner_dest) == MEM
6157 && GET_CODE (XEXP (inner_dest, 0)) == ADDRESSOF)
6158 /* Given (SET (MEM (ADDRESSOF (X))) Y) we don't want to say
6159 that (MEM (ADDRESSOF (X))) is equivalent to Y.
6160 Consider the case in which the address of the MEM is
6161 passed to a function, which alters the MEM. Then, if we
6162 later use Y instead of the MEM we'll miss the update. */
6163 elt = insert (dest, 0, sets[i].dest_hash, GET_MODE (dest));
6164 else
6165 elt = insert (dest, sets[i].src_elt,
6166 sets[i].dest_hash, GET_MODE (dest));
6168 elt->in_memory = (GET_CODE (sets[i].inner_dest) == MEM
6169 && (! RTX_UNCHANGING_P (sets[i].inner_dest)
6170 || FIXED_BASE_PLUS_P (XEXP (sets[i].inner_dest,
6171 0))));
6173 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6174 narrower than M2, and both M1 and M2 are the same number of words,
6175 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6176 make that equivalence as well.
6178 However, BAR may have equivalences for which gen_lowpart_if_possible
6179 will produce a simpler value than gen_lowpart_if_possible applied to
6180 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6181 BAR's equivalences. If we don't get a simplified form, make
6182 the SUBREG. It will not be used in an equivalence, but will
6183 cause two similar assignments to be detected.
6185 Note the loop below will find SUBREG_REG (DEST) since we have
6186 already entered SRC and DEST of the SET in the table. */
6188 if (GET_CODE (dest) == SUBREG
6189 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6190 / UNITS_PER_WORD)
6191 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6192 && (GET_MODE_SIZE (GET_MODE (dest))
6193 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6194 && sets[i].src_elt != 0)
6196 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6197 struct table_elt *elt, *classp = 0;
6199 for (elt = sets[i].src_elt->first_same_value; elt;
6200 elt = elt->next_same_value)
6202 rtx new_src = 0;
6203 unsigned src_hash;
6204 struct table_elt *src_elt;
6206 /* Ignore invalid entries. */
6207 if (GET_CODE (elt->exp) != REG
6208 && ! exp_equiv_p (elt->exp, elt->exp, 1, 0))
6209 continue;
6211 new_src = gen_lowpart_if_possible (new_mode, elt->exp);
6212 if (new_src == 0)
6213 new_src = gen_rtx_SUBREG (new_mode, elt->exp, 0);
6215 src_hash = HASH (new_src, new_mode);
6216 src_elt = lookup (new_src, src_hash, new_mode);
6218 /* Put the new source in the hash table is if isn't
6219 already. */
6220 if (src_elt == 0)
6222 if (insert_regs (new_src, classp, 0))
6224 rehash_using_reg (new_src);
6225 src_hash = HASH (new_src, new_mode);
6227 src_elt = insert (new_src, classp, src_hash, new_mode);
6228 src_elt->in_memory = elt->in_memory;
6230 else if (classp && classp != src_elt->first_same_value)
6231 /* Show that two things that we've seen before are
6232 actually the same. */
6233 merge_equiv_classes (src_elt, classp);
6235 classp = src_elt->first_same_value;
6236 /* Ignore invalid entries. */
6237 while (classp
6238 && GET_CODE (classp->exp) != REG
6239 && ! exp_equiv_p (classp->exp, classp->exp, 1, 0))
6240 classp = classp->next_same_value;
6245 /* Special handling for (set REG0 REG1) where REG0 is the
6246 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6247 be used in the sequel, so (if easily done) change this insn to
6248 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6249 that computed their value. Then REG1 will become a dead store
6250 and won't cloud the situation for later optimizations.
6252 Do not make this change if REG1 is a hard register, because it will
6253 then be used in the sequel and we may be changing a two-operand insn
6254 into a three-operand insn.
6256 Also do not do this if we are operating on a copy of INSN.
6258 Also don't do this if INSN ends a libcall; this would cause an unrelated
6259 register to be set in the middle of a libcall, and we then get bad code
6260 if the libcall is deleted. */
6262 if (n_sets == 1 && sets[0].rtl && GET_CODE (SET_DEST (sets[0].rtl)) == REG
6263 && NEXT_INSN (PREV_INSN (insn)) == insn
6264 && GET_CODE (SET_SRC (sets[0].rtl)) == REG
6265 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6266 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6268 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6269 struct qty_table_elem *src_ent = &qty_table[src_q];
6271 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6272 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6274 rtx prev = prev_nonnote_insn (insn);
6276 /* Do not swap the registers around if the previous instruction
6277 attaches a REG_EQUIV note to REG1.
6279 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6280 from the pseudo that originally shadowed an incoming argument
6281 to another register. Some uses of REG_EQUIV might rely on it
6282 being attached to REG1 rather than REG2.
6284 This section previously turned the REG_EQUIV into a REG_EQUAL
6285 note. We cannot do that because REG_EQUIV may provide an
6286 uninitialised stack slot when REG_PARM_STACK_SPACE is used. */
6288 if (prev != 0 && GET_CODE (prev) == INSN
6289 && GET_CODE (PATTERN (prev)) == SET
6290 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6291 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6293 rtx dest = SET_DEST (sets[0].rtl);
6294 rtx src = SET_SRC (sets[0].rtl);
6295 rtx note;
6297 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6298 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6299 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6300 apply_change_group ();
6302 /* If there was a REG_WAS_0 note on PREV, remove it. Move
6303 any REG_WAS_0 note on INSN to PREV. */
6304 note = find_reg_note (prev, REG_WAS_0, NULL_RTX);
6305 if (note)
6306 remove_note (prev, note);
6308 note = find_reg_note (insn, REG_WAS_0, NULL_RTX);
6309 if (note)
6311 remove_note (insn, note);
6312 XEXP (note, 1) = REG_NOTES (prev);
6313 REG_NOTES (prev) = note;
6316 /* If INSN has a REG_EQUAL note, and this note mentions
6317 REG0, then we must delete it, because the value in
6318 REG0 has changed. If the note's value is REG1, we must
6319 also delete it because that is now this insn's dest. */
6320 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6321 if (note != 0
6322 && (reg_mentioned_p (dest, XEXP (note, 0))
6323 || rtx_equal_p (src, XEXP (note, 0))))
6324 remove_note (insn, note);
6329 /* If this is a conditional jump insn, record any known equivalences due to
6330 the condition being tested. */
6332 last_jump_equiv_class = 0;
6333 if (GET_CODE (insn) == JUMP_INSN
6334 && n_sets == 1 && GET_CODE (x) == SET
6335 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6336 record_jump_equiv (insn, 0);
6338 #ifdef HAVE_cc0
6339 /* If the previous insn set CC0 and this insn no longer references CC0,
6340 delete the previous insn. Here we use the fact that nothing expects CC0
6341 to be valid over an insn, which is true until the final pass. */
6342 if (prev_insn && GET_CODE (prev_insn) == INSN
6343 && (tem = single_set (prev_insn)) != 0
6344 && SET_DEST (tem) == cc0_rtx
6345 && ! reg_mentioned_p (cc0_rtx, x))
6347 PUT_CODE (prev_insn, NOTE);
6348 NOTE_LINE_NUMBER (prev_insn) = NOTE_INSN_DELETED;
6349 NOTE_SOURCE_FILE (prev_insn) = 0;
6352 prev_insn_cc0 = this_insn_cc0;
6353 prev_insn_cc0_mode = this_insn_cc0_mode;
6354 #endif
6356 prev_insn = insn;
6359 /* Remove from the hash table all expressions that reference memory. */
6361 static void
6362 invalidate_memory ()
6364 register int i;
6365 register struct table_elt *p, *next;
6367 for (i = 0; i < HASH_SIZE; i++)
6368 for (p = table[i]; p; p = next)
6370 next = p->next_same_hash;
6371 if (p->in_memory)
6372 remove_from_table (p, i);
6376 /* If ADDR is an address that implicitly affects the stack pointer, return
6377 1 and update the register tables to show the effect. Else, return 0. */
6379 static int
6380 addr_affects_sp_p (addr)
6381 register rtx addr;
6383 if (GET_RTX_CLASS (GET_CODE (addr)) == 'a'
6384 && GET_CODE (XEXP (addr, 0)) == REG
6385 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6387 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6388 REG_TICK (STACK_POINTER_REGNUM)++;
6390 /* This should be *very* rare. */
6391 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6392 invalidate (stack_pointer_rtx, VOIDmode);
6394 return 1;
6397 return 0;
6400 /* Perform invalidation on the basis of everything about an insn
6401 except for invalidating the actual places that are SET in it.
6402 This includes the places CLOBBERed, and anything that might
6403 alias with something that is SET or CLOBBERed.
6405 X is the pattern of the insn. */
6407 static void
6408 invalidate_from_clobbers (x)
6409 rtx x;
6411 if (GET_CODE (x) == CLOBBER)
6413 rtx ref = XEXP (x, 0);
6414 if (ref)
6416 if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG
6417 || GET_CODE (ref) == MEM)
6418 invalidate (ref, VOIDmode);
6419 else if (GET_CODE (ref) == STRICT_LOW_PART
6420 || GET_CODE (ref) == ZERO_EXTRACT)
6421 invalidate (XEXP (ref, 0), GET_MODE (ref));
6424 else if (GET_CODE (x) == PARALLEL)
6426 register int i;
6427 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6429 register rtx y = XVECEXP (x, 0, i);
6430 if (GET_CODE (y) == CLOBBER)
6432 rtx ref = XEXP (y, 0);
6433 if (GET_CODE (ref) == REG || GET_CODE (ref) == SUBREG
6434 || GET_CODE (ref) == MEM)
6435 invalidate (ref, VOIDmode);
6436 else if (GET_CODE (ref) == STRICT_LOW_PART
6437 || GET_CODE (ref) == ZERO_EXTRACT)
6438 invalidate (XEXP (ref, 0), GET_MODE (ref));
6444 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6445 and replace any registers in them with either an equivalent constant
6446 or the canonical form of the register. If we are inside an address,
6447 only do this if the address remains valid.
6449 OBJECT is 0 except when within a MEM in which case it is the MEM.
6451 Return the replacement for X. */
6453 static rtx
6454 cse_process_notes (x, object)
6455 rtx x;
6456 rtx object;
6458 enum rtx_code code = GET_CODE (x);
6459 const char *fmt = GET_RTX_FORMAT (code);
6460 int i;
6462 switch (code)
6464 case CONST_INT:
6465 case CONST:
6466 case SYMBOL_REF:
6467 case LABEL_REF:
6468 case CONST_DOUBLE:
6469 case PC:
6470 case CC0:
6471 case LO_SUM:
6472 return x;
6474 case MEM:
6475 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), x);
6476 return x;
6478 case EXPR_LIST:
6479 case INSN_LIST:
6480 if (REG_NOTE_KIND (x) == REG_EQUAL)
6481 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6482 if (XEXP (x, 1))
6483 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6484 return x;
6486 case SIGN_EXTEND:
6487 case ZERO_EXTEND:
6488 case SUBREG:
6490 rtx new = cse_process_notes (XEXP (x, 0), object);
6491 /* We don't substitute VOIDmode constants into these rtx,
6492 since they would impede folding. */
6493 if (GET_MODE (new) != VOIDmode)
6494 validate_change (object, &XEXP (x, 0), new, 0);
6495 return x;
6498 case REG:
6499 i = REG_QTY (REGNO (x));
6501 /* Return a constant or a constant register. */
6502 if (REGNO_QTY_VALID_P (REGNO (x)))
6504 struct qty_table_elem *ent = &qty_table[i];
6506 if (ent->const_rtx != NULL_RTX
6507 && (CONSTANT_P (ent->const_rtx)
6508 || GET_CODE (ent->const_rtx) == REG))
6510 rtx new = gen_lowpart_if_possible (GET_MODE (x), ent->const_rtx);
6511 if (new)
6512 return new;
6516 /* Otherwise, canonicalize this register. */
6517 return canon_reg (x, NULL_RTX);
6519 default:
6520 break;
6523 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6524 if (fmt[i] == 'e')
6525 validate_change (object, &XEXP (x, i),
6526 cse_process_notes (XEXP (x, i), object), 0);
6528 return x;
6531 /* Find common subexpressions between the end test of a loop and the beginning
6532 of the loop. LOOP_START is the CODE_LABEL at the start of a loop.
6534 Often we have a loop where an expression in the exit test is used
6535 in the body of the loop. For example "while (*p) *q++ = *p++;".
6536 Because of the way we duplicate the loop exit test in front of the loop,
6537 however, we don't detect that common subexpression. This will be caught
6538 when global cse is implemented, but this is a quite common case.
6540 This function handles the most common cases of these common expressions.
6541 It is called after we have processed the basic block ending with the
6542 NOTE_INSN_LOOP_END note that ends a loop and the previous JUMP_INSN
6543 jumps to a label used only once. */
6545 static void
6546 cse_around_loop (loop_start)
6547 rtx loop_start;
6549 rtx insn;
6550 int i;
6551 struct table_elt *p;
6553 /* If the jump at the end of the loop doesn't go to the start, we don't
6554 do anything. */
6555 for (insn = PREV_INSN (loop_start);
6556 insn && (GET_CODE (insn) == NOTE && NOTE_LINE_NUMBER (insn) >= 0);
6557 insn = PREV_INSN (insn))
6560 if (insn == 0
6561 || GET_CODE (insn) != NOTE
6562 || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG)
6563 return;
6565 /* If the last insn of the loop (the end test) was an NE comparison,
6566 we will interpret it as an EQ comparison, since we fell through
6567 the loop. Any equivalences resulting from that comparison are
6568 therefore not valid and must be invalidated. */
6569 if (last_jump_equiv_class)
6570 for (p = last_jump_equiv_class->first_same_value; p;
6571 p = p->next_same_value)
6573 if (GET_CODE (p->exp) == MEM || GET_CODE (p->exp) == REG
6574 || (GET_CODE (p->exp) == SUBREG
6575 && GET_CODE (SUBREG_REG (p->exp)) == REG))
6576 invalidate (p->exp, VOIDmode);
6577 else if (GET_CODE (p->exp) == STRICT_LOW_PART
6578 || GET_CODE (p->exp) == ZERO_EXTRACT)
6579 invalidate (XEXP (p->exp, 0), GET_MODE (p->exp));
6582 /* Process insns starting after LOOP_START until we hit a CALL_INSN or
6583 a CODE_LABEL (we could handle a CALL_INSN, but it isn't worth it).
6585 The only thing we do with SET_DEST is invalidate entries, so we
6586 can safely process each SET in order. It is slightly less efficient
6587 to do so, but we only want to handle the most common cases.
6589 The gen_move_insn call in cse_set_around_loop may create new pseudos.
6590 These pseudos won't have valid entries in any of the tables indexed
6591 by register number, such as reg_qty. We avoid out-of-range array
6592 accesses by not processing any instructions created after cse started. */
6594 for (insn = NEXT_INSN (loop_start);
6595 GET_CODE (insn) != CALL_INSN && GET_CODE (insn) != CODE_LABEL
6596 && INSN_UID (insn) < max_insn_uid
6597 && ! (GET_CODE (insn) == NOTE
6598 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END);
6599 insn = NEXT_INSN (insn))
6601 if (INSN_P (insn)
6602 && (GET_CODE (PATTERN (insn)) == SET
6603 || GET_CODE (PATTERN (insn)) == CLOBBER))
6604 cse_set_around_loop (PATTERN (insn), insn, loop_start);
6605 else if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == PARALLEL)
6606 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6607 if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == SET
6608 || GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER)
6609 cse_set_around_loop (XVECEXP (PATTERN (insn), 0, i), insn,
6610 loop_start);
6614 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6615 since they are done elsewhere. This function is called via note_stores. */
6617 static void
6618 invalidate_skipped_set (dest, set, data)
6619 rtx set;
6620 rtx dest;
6621 void *data ATTRIBUTE_UNUSED;
6623 enum rtx_code code = GET_CODE (dest);
6625 if (code == MEM
6626 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6627 /* There are times when an address can appear varying and be a PLUS
6628 during this scan when it would be a fixed address were we to know
6629 the proper equivalences. So invalidate all memory if there is
6630 a BLKmode or nonscalar memory reference or a reference to a
6631 variable address. */
6632 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6633 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6635 invalidate_memory ();
6636 return;
6639 if (GET_CODE (set) == CLOBBER
6640 #ifdef HAVE_cc0
6641 || dest == cc0_rtx
6642 #endif
6643 || dest == pc_rtx)
6644 return;
6646 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6647 invalidate (XEXP (dest, 0), GET_MODE (dest));
6648 else if (code == REG || code == SUBREG || code == MEM)
6649 invalidate (dest, VOIDmode);
6652 /* Invalidate all insns from START up to the end of the function or the
6653 next label. This called when we wish to CSE around a block that is
6654 conditionally executed. */
6656 static void
6657 invalidate_skipped_block (start)
6658 rtx start;
6660 rtx insn;
6662 for (insn = start; insn && GET_CODE (insn) != CODE_LABEL;
6663 insn = NEXT_INSN (insn))
6665 if (! INSN_P (insn))
6666 continue;
6668 if (GET_CODE (insn) == CALL_INSN)
6670 if (! CONST_CALL_P (insn))
6671 invalidate_memory ();
6672 invalidate_for_call ();
6675 invalidate_from_clobbers (PATTERN (insn));
6676 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6680 /* If modifying X will modify the value in *DATA (which is really an
6681 `rtx *'), indicate that fact by setting the pointed to value to
6682 NULL_RTX. */
6684 static void
6685 cse_check_loop_start (x, set, data)
6686 rtx x;
6687 rtx set ATTRIBUTE_UNUSED;
6688 void *data;
6690 rtx *cse_check_loop_start_value = (rtx *) data;
6692 if (*cse_check_loop_start_value == NULL_RTX
6693 || GET_CODE (x) == CC0 || GET_CODE (x) == PC)
6694 return;
6696 if ((GET_CODE (x) == MEM && GET_CODE (*cse_check_loop_start_value) == MEM)
6697 || reg_overlap_mentioned_p (x, *cse_check_loop_start_value))
6698 *cse_check_loop_start_value = NULL_RTX;
6701 /* X is a SET or CLOBBER contained in INSN that was found near the start of
6702 a loop that starts with the label at LOOP_START.
6704 If X is a SET, we see if its SET_SRC is currently in our hash table.
6705 If so, we see if it has a value equal to some register used only in the
6706 loop exit code (as marked by jump.c).
6708 If those two conditions are true, we search backwards from the start of
6709 the loop to see if that same value was loaded into a register that still
6710 retains its value at the start of the loop.
6712 If so, we insert an insn after the load to copy the destination of that
6713 load into the equivalent register and (try to) replace our SET_SRC with that
6714 register.
6716 In any event, we invalidate whatever this SET or CLOBBER modifies. */
6718 static void
6719 cse_set_around_loop (x, insn, loop_start)
6720 rtx x;
6721 rtx insn;
6722 rtx loop_start;
6724 struct table_elt *src_elt;
6726 /* If this is a SET, see if we can replace SET_SRC, but ignore SETs that
6727 are setting PC or CC0 or whose SET_SRC is already a register. */
6728 if (GET_CODE (x) == SET
6729 && GET_CODE (SET_DEST (x)) != PC && GET_CODE (SET_DEST (x)) != CC0
6730 && GET_CODE (SET_SRC (x)) != REG)
6732 src_elt = lookup (SET_SRC (x),
6733 HASH (SET_SRC (x), GET_MODE (SET_DEST (x))),
6734 GET_MODE (SET_DEST (x)));
6736 if (src_elt)
6737 for (src_elt = src_elt->first_same_value; src_elt;
6738 src_elt = src_elt->next_same_value)
6739 if (GET_CODE (src_elt->exp) == REG && REG_LOOP_TEST_P (src_elt->exp)
6740 && COST (src_elt->exp) < COST (SET_SRC (x)))
6742 rtx p, set;
6744 /* Look for an insn in front of LOOP_START that sets
6745 something in the desired mode to SET_SRC (x) before we hit
6746 a label or CALL_INSN. */
6748 for (p = prev_nonnote_insn (loop_start);
6749 p && GET_CODE (p) != CALL_INSN
6750 && GET_CODE (p) != CODE_LABEL;
6751 p = prev_nonnote_insn (p))
6752 if ((set = single_set (p)) != 0
6753 && GET_CODE (SET_DEST (set)) == REG
6754 && GET_MODE (SET_DEST (set)) == src_elt->mode
6755 && rtx_equal_p (SET_SRC (set), SET_SRC (x)))
6757 /* We now have to ensure that nothing between P
6758 and LOOP_START modified anything referenced in
6759 SET_SRC (x). We know that nothing within the loop
6760 can modify it, or we would have invalidated it in
6761 the hash table. */
6762 rtx q;
6763 rtx cse_check_loop_start_value = SET_SRC (x);
6764 for (q = p; q != loop_start; q = NEXT_INSN (q))
6765 if (INSN_P (q))
6766 note_stores (PATTERN (q),
6767 cse_check_loop_start,
6768 &cse_check_loop_start_value);
6770 /* If nothing was changed and we can replace our
6771 SET_SRC, add an insn after P to copy its destination
6772 to what we will be replacing SET_SRC with. */
6773 if (cse_check_loop_start_value
6774 && validate_change (insn, &SET_SRC (x),
6775 src_elt->exp, 0))
6777 /* If this creates new pseudos, this is unsafe,
6778 because the regno of new pseudo is unsuitable
6779 to index into reg_qty when cse_insn processes
6780 the new insn. Therefore, if a new pseudo was
6781 created, discard this optimization. */
6782 int nregs = max_reg_num ();
6783 rtx move
6784 = gen_move_insn (src_elt->exp, SET_DEST (set));
6785 if (nregs != max_reg_num ())
6787 if (! validate_change (insn, &SET_SRC (x),
6788 SET_SRC (set), 0))
6789 abort ();
6791 else
6792 emit_insn_after (move, p);
6794 break;
6799 /* Deal with the destination of X affecting the stack pointer. */
6800 addr_affects_sp_p (SET_DEST (x));
6802 /* See comment on similar code in cse_insn for explanation of these
6803 tests. */
6804 if (GET_CODE (SET_DEST (x)) == REG || GET_CODE (SET_DEST (x)) == SUBREG
6805 || GET_CODE (SET_DEST (x)) == MEM)
6806 invalidate (SET_DEST (x), VOIDmode);
6807 else if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
6808 || GET_CODE (SET_DEST (x)) == ZERO_EXTRACT)
6809 invalidate (XEXP (SET_DEST (x), 0), GET_MODE (SET_DEST (x)));
6812 /* Find the end of INSN's basic block and return its range,
6813 the total number of SETs in all the insns of the block, the last insn of the
6814 block, and the branch path.
6816 The branch path indicates which branches should be followed. If a non-zero
6817 path size is specified, the block should be rescanned and a different set
6818 of branches will be taken. The branch path is only used if
6819 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is non-zero.
6821 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6822 used to describe the block. It is filled in with the information about
6823 the current block. The incoming structure's branch path, if any, is used
6824 to construct the output branch path. */
6826 void
6827 cse_end_of_basic_block (insn, data, follow_jumps, after_loop, skip_blocks)
6828 rtx insn;
6829 struct cse_basic_block_data *data;
6830 int follow_jumps;
6831 int after_loop;
6832 int skip_blocks;
6834 rtx p = insn, q;
6835 int nsets = 0;
6836 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6837 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6838 int path_size = data->path_size;
6839 int path_entry = 0;
6840 int i;
6842 /* Update the previous branch path, if any. If the last branch was
6843 previously TAKEN, mark it NOT_TAKEN. If it was previously NOT_TAKEN,
6844 shorten the path by one and look at the previous branch. We know that
6845 at least one branch must have been taken if PATH_SIZE is non-zero. */
6846 while (path_size > 0)
6848 if (data->path[path_size - 1].status != NOT_TAKEN)
6850 data->path[path_size - 1].status = NOT_TAKEN;
6851 break;
6853 else
6854 path_size--;
6857 /* If the first instruction is marked with QImode, that means we've
6858 already processed this block. Our caller will look at DATA->LAST
6859 to figure out where to go next. We want to return the next block
6860 in the instruction stream, not some branched-to block somewhere
6861 else. We accomplish this by pretending our called forbid us to
6862 follow jumps, or skip blocks. */
6863 if (GET_MODE (insn) == QImode)
6864 follow_jumps = skip_blocks = 0;
6866 /* Scan to end of this basic block. */
6867 while (p && GET_CODE (p) != CODE_LABEL)
6869 /* Don't cse out the end of a loop. This makes a difference
6870 only for the unusual loops that always execute at least once;
6871 all other loops have labels there so we will stop in any case.
6872 Cse'ing out the end of the loop is dangerous because it
6873 might cause an invariant expression inside the loop
6874 to be reused after the end of the loop. This would make it
6875 hard to move the expression out of the loop in loop.c,
6876 especially if it is one of several equivalent expressions
6877 and loop.c would like to eliminate it.
6879 If we are running after loop.c has finished, we can ignore
6880 the NOTE_INSN_LOOP_END. */
6882 if (! after_loop && GET_CODE (p) == NOTE
6883 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
6884 break;
6886 /* Don't cse over a call to setjmp; on some machines (eg vax)
6887 the regs restored by the longjmp come from
6888 a later time than the setjmp. */
6889 if (GET_CODE (p) == NOTE
6890 && NOTE_LINE_NUMBER (p) == NOTE_INSN_SETJMP)
6891 break;
6893 /* A PARALLEL can have lots of SETs in it,
6894 especially if it is really an ASM_OPERANDS. */
6895 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6896 nsets += XVECLEN (PATTERN (p), 0);
6897 else if (GET_CODE (p) != NOTE)
6898 nsets += 1;
6900 /* Ignore insns made by CSE; they cannot affect the boundaries of
6901 the basic block. */
6903 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6904 high_cuid = INSN_CUID (p);
6905 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6906 low_cuid = INSN_CUID (p);
6908 /* See if this insn is in our branch path. If it is and we are to
6909 take it, do so. */
6910 if (path_entry < path_size && data->path[path_entry].branch == p)
6912 if (data->path[path_entry].status != NOT_TAKEN)
6913 p = JUMP_LABEL (p);
6915 /* Point to next entry in path, if any. */
6916 path_entry++;
6919 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6920 was specified, we haven't reached our maximum path length, there are
6921 insns following the target of the jump, this is the only use of the
6922 jump label, and the target label is preceded by a BARRIER.
6924 Alternatively, we can follow the jump if it branches around a
6925 block of code and there are no other branches into the block.
6926 In this case invalidate_skipped_block will be called to invalidate any
6927 registers set in the block when following the jump. */
6929 else if ((follow_jumps || skip_blocks) && path_size < PATHLENGTH - 1
6930 && GET_CODE (p) == JUMP_INSN
6931 && GET_CODE (PATTERN (p)) == SET
6932 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6933 && JUMP_LABEL (p) != 0
6934 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6935 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6937 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6938 if ((GET_CODE (q) != NOTE
6939 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6940 || NOTE_LINE_NUMBER (q) == NOTE_INSN_SETJMP)
6941 && (GET_CODE (q) != CODE_LABEL || LABEL_NUSES (q) != 0))
6942 break;
6944 /* If we ran into a BARRIER, this code is an extension of the
6945 basic block when the branch is taken. */
6946 if (follow_jumps && q != 0 && GET_CODE (q) == BARRIER)
6948 /* Don't allow ourself to keep walking around an
6949 always-executed loop. */
6950 if (next_real_insn (q) == next)
6952 p = NEXT_INSN (p);
6953 continue;
6956 /* Similarly, don't put a branch in our path more than once. */
6957 for (i = 0; i < path_entry; i++)
6958 if (data->path[i].branch == p)
6959 break;
6961 if (i != path_entry)
6962 break;
6964 data->path[path_entry].branch = p;
6965 data->path[path_entry++].status = TAKEN;
6967 /* This branch now ends our path. It was possible that we
6968 didn't see this branch the last time around (when the
6969 insn in front of the target was a JUMP_INSN that was
6970 turned into a no-op). */
6971 path_size = path_entry;
6973 p = JUMP_LABEL (p);
6974 /* Mark block so we won't scan it again later. */
6975 PUT_MODE (NEXT_INSN (p), QImode);
6977 /* Detect a branch around a block of code. */
6978 else if (skip_blocks && q != 0 && GET_CODE (q) != CODE_LABEL)
6980 register rtx tmp;
6982 if (next_real_insn (q) == next)
6984 p = NEXT_INSN (p);
6985 continue;
6988 for (i = 0; i < path_entry; i++)
6989 if (data->path[i].branch == p)
6990 break;
6992 if (i != path_entry)
6993 break;
6995 /* This is no_labels_between_p (p, q) with an added check for
6996 reaching the end of a function (in case Q precedes P). */
6997 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6998 if (GET_CODE (tmp) == CODE_LABEL)
6999 break;
7001 if (tmp == q)
7003 data->path[path_entry].branch = p;
7004 data->path[path_entry++].status = AROUND;
7006 path_size = path_entry;
7008 p = JUMP_LABEL (p);
7009 /* Mark block so we won't scan it again later. */
7010 PUT_MODE (NEXT_INSN (p), QImode);
7014 p = NEXT_INSN (p);
7017 data->low_cuid = low_cuid;
7018 data->high_cuid = high_cuid;
7019 data->nsets = nsets;
7020 data->last = p;
7022 /* If all jumps in the path are not taken, set our path length to zero
7023 so a rescan won't be done. */
7024 for (i = path_size - 1; i >= 0; i--)
7025 if (data->path[i].status != NOT_TAKEN)
7026 break;
7028 if (i == -1)
7029 data->path_size = 0;
7030 else
7031 data->path_size = path_size;
7033 /* End the current branch path. */
7034 data->path[path_size].branch = 0;
7037 /* Perform cse on the instructions of a function.
7038 F is the first instruction.
7039 NREGS is one plus the highest pseudo-reg number used in the instruction.
7041 AFTER_LOOP is 1 if this is the cse call done after loop optimization
7042 (only if -frerun-cse-after-loop).
7044 Returns 1 if jump_optimize should be redone due to simplifications
7045 in conditional jump instructions. */
7048 cse_main (f, nregs, after_loop, file)
7049 rtx f;
7050 int nregs;
7051 int after_loop;
7052 FILE *file;
7054 struct cse_basic_block_data val;
7055 register rtx insn = f;
7056 register int i;
7058 cse_jumps_altered = 0;
7059 recorded_label_ref = 0;
7060 constant_pool_entries_cost = 0;
7061 val.path_size = 0;
7063 init_recog ();
7064 init_alias_analysis ();
7066 max_reg = nregs;
7068 max_insn_uid = get_max_uid ();
7070 reg_eqv_table = (struct reg_eqv_elem *)
7071 xmalloc (nregs * sizeof (struct reg_eqv_elem));
7073 #ifdef LOAD_EXTEND_OP
7075 /* Allocate scratch rtl here. cse_insn will fill in the memory reference
7076 and change the code and mode as appropriate. */
7077 memory_extend_rtx = gen_rtx_ZERO_EXTEND (VOIDmode, NULL_RTX);
7078 #endif
7080 /* Reset the counter indicating how many elements have been made
7081 thus far. */
7082 n_elements_made = 0;
7084 /* Find the largest uid. */
7086 max_uid = get_max_uid ();
7087 uid_cuid = (int *) xcalloc (max_uid + 1, sizeof (int));
7089 /* Compute the mapping from uids to cuids.
7090 CUIDs are numbers assigned to insns, like uids,
7091 except that cuids increase monotonically through the code.
7092 Don't assign cuids to line-number NOTEs, so that the distance in cuids
7093 between two insns is not affected by -g. */
7095 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
7097 if (GET_CODE (insn) != NOTE
7098 || NOTE_LINE_NUMBER (insn) < 0)
7099 INSN_CUID (insn) = ++i;
7100 else
7101 /* Give a line number note the same cuid as preceding insn. */
7102 INSN_CUID (insn) = i;
7105 /* Initialize which registers are clobbered by calls. */
7107 CLEAR_HARD_REG_SET (regs_invalidated_by_call);
7109 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
7110 if ((call_used_regs[i]
7111 /* Used to check !fixed_regs[i] here, but that isn't safe;
7112 fixed regs are still call-clobbered, and sched can get
7113 confused if they can "live across calls".
7115 The frame pointer is always preserved across calls. The arg
7116 pointer is if it is fixed. The stack pointer usually is, unless
7117 RETURN_POPS_ARGS, in which case an explicit CLOBBER
7118 will be present. If we are generating PIC code, the PIC offset
7119 table register is preserved across calls. */
7121 && i != STACK_POINTER_REGNUM
7122 && i != FRAME_POINTER_REGNUM
7123 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
7124 && i != HARD_FRAME_POINTER_REGNUM
7125 #endif
7126 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
7127 && ! (i == ARG_POINTER_REGNUM && fixed_regs[i])
7128 #endif
7129 #if !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
7130 && ! (i == PIC_OFFSET_TABLE_REGNUM && flag_pic)
7131 #endif
7133 || global_regs[i])
7134 SET_HARD_REG_BIT (regs_invalidated_by_call, i);
7136 ggc_push_context ();
7138 /* Loop over basic blocks.
7139 Compute the maximum number of qty's needed for each basic block
7140 (which is 2 for each SET). */
7141 insn = f;
7142 while (insn)
7144 cse_altered = 0;
7145 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps, after_loop,
7146 flag_cse_skip_blocks);
7148 /* If this basic block was already processed or has no sets, skip it. */
7149 if (val.nsets == 0 || GET_MODE (insn) == QImode)
7151 PUT_MODE (insn, VOIDmode);
7152 insn = (val.last ? NEXT_INSN (val.last) : 0);
7153 val.path_size = 0;
7154 continue;
7157 cse_basic_block_start = val.low_cuid;
7158 cse_basic_block_end = val.high_cuid;
7159 max_qty = val.nsets * 2;
7161 if (file)
7162 fnotice (file, ";; Processing block from %d to %d, %d sets.\n",
7163 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
7164 val.nsets);
7166 /* Make MAX_QTY bigger to give us room to optimize
7167 past the end of this basic block, if that should prove useful. */
7168 if (max_qty < 500)
7169 max_qty = 500;
7171 max_qty += max_reg;
7173 /* If this basic block is being extended by following certain jumps,
7174 (see `cse_end_of_basic_block'), we reprocess the code from the start.
7175 Otherwise, we start after this basic block. */
7176 if (val.path_size > 0)
7177 cse_basic_block (insn, val.last, val.path, 0);
7178 else
7180 int old_cse_jumps_altered = cse_jumps_altered;
7181 rtx temp;
7183 /* When cse changes a conditional jump to an unconditional
7184 jump, we want to reprocess the block, since it will give
7185 us a new branch path to investigate. */
7186 cse_jumps_altered = 0;
7187 temp = cse_basic_block (insn, val.last, val.path, ! after_loop);
7188 if (cse_jumps_altered == 0
7189 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7190 insn = temp;
7192 cse_jumps_altered |= old_cse_jumps_altered;
7195 if (cse_altered)
7196 ggc_collect ();
7198 #ifdef USE_C_ALLOCA
7199 alloca (0);
7200 #endif
7203 ggc_pop_context ();
7205 if (max_elements_made < n_elements_made)
7206 max_elements_made = n_elements_made;
7208 /* Clean up. */
7209 end_alias_analysis ();
7210 free (uid_cuid);
7211 free (reg_eqv_table);
7213 return cse_jumps_altered || recorded_label_ref;
7216 /* Process a single basic block. FROM and TO and the limits of the basic
7217 block. NEXT_BRANCH points to the branch path when following jumps or
7218 a null path when not following jumps.
7220 AROUND_LOOP is non-zero if we are to try to cse around to the start of a
7221 loop. This is true when we are being called for the last time on a
7222 block and this CSE pass is before loop.c. */
7224 static rtx
7225 cse_basic_block (from, to, next_branch, around_loop)
7226 register rtx from, to;
7227 struct branch_path *next_branch;
7228 int around_loop;
7230 register rtx insn;
7231 int to_usage = 0;
7232 rtx libcall_insn = NULL_RTX;
7233 int num_insns = 0;
7235 /* This array is undefined before max_reg, so only allocate
7236 the space actually needed and adjust the start. */
7238 qty_table
7239 = (struct qty_table_elem *) xmalloc ((max_qty - max_reg)
7240 * sizeof (struct qty_table_elem));
7241 qty_table -= max_reg;
7243 new_basic_block ();
7245 /* TO might be a label. If so, protect it from being deleted. */
7246 if (to != 0 && GET_CODE (to) == CODE_LABEL)
7247 ++LABEL_NUSES (to);
7249 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7251 register enum rtx_code code = GET_CODE (insn);
7253 /* If we have processed 1,000 insns, flush the hash table to
7254 avoid extreme quadratic behavior. We must not include NOTEs
7255 in the count since there may be more of them when generating
7256 debugging information. If we clear the table at different
7257 times, code generated with -g -O might be different than code
7258 generated with -O but not -g.
7260 ??? This is a real kludge and needs to be done some other way.
7261 Perhaps for 2.9. */
7262 if (code != NOTE && num_insns++ > 1000)
7264 flush_hash_table ();
7265 num_insns = 0;
7268 /* See if this is a branch that is part of the path. If so, and it is
7269 to be taken, do so. */
7270 if (next_branch->branch == insn)
7272 enum taken status = next_branch++->status;
7273 if (status != NOT_TAKEN)
7275 if (status == TAKEN)
7276 record_jump_equiv (insn, 1);
7277 else
7278 invalidate_skipped_block (NEXT_INSN (insn));
7280 /* Set the last insn as the jump insn; it doesn't affect cc0.
7281 Then follow this branch. */
7282 #ifdef HAVE_cc0
7283 prev_insn_cc0 = 0;
7284 #endif
7285 prev_insn = insn;
7286 insn = JUMP_LABEL (insn);
7287 continue;
7291 if (GET_MODE (insn) == QImode)
7292 PUT_MODE (insn, VOIDmode);
7294 if (GET_RTX_CLASS (code) == 'i')
7296 rtx p;
7298 /* Process notes first so we have all notes in canonical forms when
7299 looking for duplicate operations. */
7301 if (REG_NOTES (insn))
7302 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7304 /* Track when we are inside in LIBCALL block. Inside such a block,
7305 we do not want to record destinations. The last insn of a
7306 LIBCALL block is not considered to be part of the block, since
7307 its destination is the result of the block and hence should be
7308 recorded. */
7310 if (REG_NOTES (insn) != 0)
7312 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7313 libcall_insn = XEXP (p, 0);
7314 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7315 libcall_insn = 0;
7318 new_label_ref = 0;
7319 cse_insn (insn, libcall_insn);
7321 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL
7322 note for it, we must rerun jump since it needs to place the
7323 note. If this is a LABEL_REF for a CODE_LABEL that isn't in
7324 the insn chain, don't do this since no REG_LABEL will be added. */
7325 if (new_label_ref != 0 && INSN_UID (XEXP (new_label_ref, 0)) != 0
7326 && reg_mentioned_p (new_label_ref, PATTERN (insn))
7327 && ! find_reg_note (insn, REG_LABEL, XEXP (new_label_ref, 0)))
7328 recorded_label_ref = 1;
7331 /* If INSN is now an unconditional jump, skip to the end of our
7332 basic block by pretending that we just did the last insn in the
7333 basic block. If we are jumping to the end of our block, show
7334 that we can have one usage of TO. */
7336 if (any_uncondjump_p (insn))
7338 if (to == 0)
7340 free (qty_table + max_reg);
7341 return 0;
7344 if (JUMP_LABEL (insn) == to)
7345 to_usage = 1;
7347 /* Maybe TO was deleted because the jump is unconditional.
7348 If so, there is nothing left in this basic block. */
7349 /* ??? Perhaps it would be smarter to set TO
7350 to whatever follows this insn,
7351 and pretend the basic block had always ended here. */
7352 if (INSN_DELETED_P (to))
7353 break;
7355 insn = PREV_INSN (to);
7358 /* See if it is ok to keep on going past the label
7359 which used to end our basic block. Remember that we incremented
7360 the count of that label, so we decrement it here. If we made
7361 a jump unconditional, TO_USAGE will be one; in that case, we don't
7362 want to count the use in that jump. */
7364 if (to != 0 && NEXT_INSN (insn) == to
7365 && GET_CODE (to) == CODE_LABEL && --LABEL_NUSES (to) == to_usage)
7367 struct cse_basic_block_data val;
7368 rtx prev;
7370 insn = NEXT_INSN (to);
7372 /* If TO was the last insn in the function, we are done. */
7373 if (insn == 0)
7375 free (qty_table + max_reg);
7376 return 0;
7379 /* If TO was preceded by a BARRIER we are done with this block
7380 because it has no continuation. */
7381 prev = prev_nonnote_insn (to);
7382 if (prev && GET_CODE (prev) == BARRIER)
7384 free (qty_table + max_reg);
7385 return insn;
7388 /* Find the end of the following block. Note that we won't be
7389 following branches in this case. */
7390 to_usage = 0;
7391 val.path_size = 0;
7392 cse_end_of_basic_block (insn, &val, 0, 0, 0);
7394 /* If the tables we allocated have enough space left
7395 to handle all the SETs in the next basic block,
7396 continue through it. Otherwise, return,
7397 and that block will be scanned individually. */
7398 if (val.nsets * 2 + next_qty > max_qty)
7399 break;
7401 cse_basic_block_start = val.low_cuid;
7402 cse_basic_block_end = val.high_cuid;
7403 to = val.last;
7405 /* Prevent TO from being deleted if it is a label. */
7406 if (to != 0 && GET_CODE (to) == CODE_LABEL)
7407 ++LABEL_NUSES (to);
7409 /* Back up so we process the first insn in the extension. */
7410 insn = PREV_INSN (insn);
7414 if (next_qty > max_qty)
7415 abort ();
7417 /* If we are running before loop.c, we stopped on a NOTE_INSN_LOOP_END, and
7418 the previous insn is the only insn that branches to the head of a loop,
7419 we can cse into the loop. Don't do this if we changed the jump
7420 structure of a loop unless we aren't going to be following jumps. */
7422 if ((cse_jumps_altered == 0
7423 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7424 && around_loop && to != 0
7425 && GET_CODE (to) == NOTE && NOTE_LINE_NUMBER (to) == NOTE_INSN_LOOP_END
7426 && GET_CODE (PREV_INSN (to)) == JUMP_INSN
7427 && JUMP_LABEL (PREV_INSN (to)) != 0
7428 && LABEL_NUSES (JUMP_LABEL (PREV_INSN (to))) == 1)
7429 cse_around_loop (JUMP_LABEL (PREV_INSN (to)));
7431 free (qty_table + max_reg);
7433 return to ? NEXT_INSN (to) : 0;
7436 /* Count the number of times registers are used (not set) in X.
7437 COUNTS is an array in which we accumulate the count, INCR is how much
7438 we count each register usage.
7440 Don't count a usage of DEST, which is the SET_DEST of a SET which
7441 contains X in its SET_SRC. This is because such a SET does not
7442 modify the liveness of DEST. */
7444 static void
7445 count_reg_usage (x, counts, dest, incr)
7446 rtx x;
7447 int *counts;
7448 rtx dest;
7449 int incr;
7451 enum rtx_code code;
7452 const char *fmt;
7453 int i, j;
7455 if (x == 0)
7456 return;
7458 switch (code = GET_CODE (x))
7460 case REG:
7461 if (x != dest)
7462 counts[REGNO (x)] += incr;
7463 return;
7465 case PC:
7466 case CC0:
7467 case CONST:
7468 case CONST_INT:
7469 case CONST_DOUBLE:
7470 case SYMBOL_REF:
7471 case LABEL_REF:
7472 return;
7474 case CLOBBER:
7475 /* If we are clobbering a MEM, mark any registers inside the address
7476 as being used. */
7477 if (GET_CODE (XEXP (x, 0)) == MEM)
7478 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7479 return;
7481 case SET:
7482 /* Unless we are setting a REG, count everything in SET_DEST. */
7483 if (GET_CODE (SET_DEST (x)) != REG)
7484 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7486 /* If SRC has side-effects, then we can't delete this insn, so the
7487 usage of SET_DEST inside SRC counts.
7489 ??? Strictly-speaking, we might be preserving this insn
7490 because some other SET has side-effects, but that's hard
7491 to do and can't happen now. */
7492 count_reg_usage (SET_SRC (x), counts,
7493 side_effects_p (SET_SRC (x)) ? NULL_RTX : SET_DEST (x),
7494 incr);
7495 return;
7497 case CALL_INSN:
7498 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, NULL_RTX, incr);
7499 /* Fall through. */
7501 case INSN:
7502 case JUMP_INSN:
7503 count_reg_usage (PATTERN (x), counts, NULL_RTX, incr);
7505 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7506 use them. */
7508 count_reg_usage (REG_NOTES (x), counts, NULL_RTX, incr);
7509 return;
7511 case EXPR_LIST:
7512 case INSN_LIST:
7513 if (REG_NOTE_KIND (x) == REG_EQUAL
7514 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE))
7515 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7516 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7517 return;
7519 default:
7520 break;
7523 fmt = GET_RTX_FORMAT (code);
7524 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7526 if (fmt[i] == 'e')
7527 count_reg_usage (XEXP (x, i), counts, dest, incr);
7528 else if (fmt[i] == 'E')
7529 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7530 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7534 /* Scan all the insns and delete any that are dead; i.e., they store a register
7535 that is never used or they copy a register to itself.
7537 This is used to remove insns made obviously dead by cse, loop or other
7538 optimizations. It improves the heuristics in loop since it won't try to
7539 move dead invariants out of loops or make givs for dead quantities. The
7540 remaining passes of the compilation are also sped up. */
7542 void
7543 delete_trivially_dead_insns (insns, nreg)
7544 rtx insns;
7545 int nreg;
7547 int *counts;
7548 rtx insn, prev;
7549 #ifdef HAVE_cc0
7550 rtx tem;
7551 #endif
7552 int i;
7553 int in_libcall = 0, dead_libcall = 0;
7555 /* First count the number of times each register is used. */
7556 counts = (int *) xcalloc (nreg, sizeof (int));
7557 for (insn = next_real_insn (insns); insn; insn = next_real_insn (insn))
7558 count_reg_usage (insn, counts, NULL_RTX, 1);
7560 /* Go from the last insn to the first and delete insns that only set unused
7561 registers or copy a register to itself. As we delete an insn, remove
7562 usage counts for registers it uses.
7564 The first jump optimization pass may leave a real insn as the last
7565 insn in the function. We must not skip that insn or we may end
7566 up deleting code that is not really dead. */
7567 insn = get_last_insn ();
7568 if (! INSN_P (insn))
7569 insn = prev_real_insn (insn);
7571 for (; insn; insn = prev)
7573 int live_insn = 0;
7574 rtx note;
7576 prev = prev_real_insn (insn);
7578 /* Don't delete any insns that are part of a libcall block unless
7579 we can delete the whole libcall block.
7581 Flow or loop might get confused if we did that. Remember
7582 that we are scanning backwards. */
7583 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7585 in_libcall = 1;
7586 live_insn = 1;
7587 dead_libcall = 0;
7589 /* See if there's a REG_EQUAL note on this insn and try to
7590 replace the source with the REG_EQUAL expression.
7592 We assume that insns with REG_RETVALs can only be reg->reg
7593 copies at this point. */
7594 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7595 if (note)
7597 rtx set = single_set (insn);
7598 rtx new = simplify_rtx (XEXP (note, 0));
7600 if (!new)
7601 new = XEXP (note, 0);
7603 if (set && validate_change (insn, &SET_SRC (set), new, 0))
7605 remove_note (insn,
7606 find_reg_note (insn, REG_RETVAL, NULL_RTX));
7607 dead_libcall = 1;
7611 else if (in_libcall)
7612 live_insn = ! dead_libcall;
7613 else if (GET_CODE (PATTERN (insn)) == SET)
7615 if (set_noop_p (PATTERN (insn)))
7618 #ifdef HAVE_cc0
7619 else if (GET_CODE (SET_DEST (PATTERN (insn))) == CC0
7620 && ! side_effects_p (SET_SRC (PATTERN (insn)))
7621 && ((tem = next_nonnote_insn (insn)) == 0
7622 || ! INSN_P (tem)
7623 || ! reg_referenced_p (cc0_rtx, PATTERN (tem))))
7625 #endif
7626 else if (GET_CODE (SET_DEST (PATTERN (insn))) != REG
7627 || REGNO (SET_DEST (PATTERN (insn))) < FIRST_PSEUDO_REGISTER
7628 || counts[REGNO (SET_DEST (PATTERN (insn)))] != 0
7629 || side_effects_p (SET_SRC (PATTERN (insn)))
7630 /* An ADDRESSOF expression can turn into a use of the
7631 internal arg pointer, so always consider the
7632 internal arg pointer live. If it is truly dead,
7633 flow will delete the initializing insn. */
7634 || (SET_DEST (PATTERN (insn))
7635 == current_function_internal_arg_pointer))
7636 live_insn = 1;
7638 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7639 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7641 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7643 if (GET_CODE (elt) == SET)
7645 if (set_noop_p (elt))
7648 #ifdef HAVE_cc0
7649 else if (GET_CODE (SET_DEST (elt)) == CC0
7650 && ! side_effects_p (SET_SRC (elt))
7651 && ((tem = next_nonnote_insn (insn)) == 0
7652 || ! INSN_P (tem)
7653 || ! reg_referenced_p (cc0_rtx, PATTERN (tem))))
7655 #endif
7656 else if (GET_CODE (SET_DEST (elt)) != REG
7657 || REGNO (SET_DEST (elt)) < FIRST_PSEUDO_REGISTER
7658 || counts[REGNO (SET_DEST (elt))] != 0
7659 || side_effects_p (SET_SRC (elt))
7660 /* An ADDRESSOF expression can turn into a use of the
7661 internal arg pointer, so always consider the
7662 internal arg pointer live. If it is truly dead,
7663 flow will delete the initializing insn. */
7664 || (SET_DEST (elt)
7665 == current_function_internal_arg_pointer))
7666 live_insn = 1;
7668 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7669 live_insn = 1;
7671 else
7672 live_insn = 1;
7674 /* If this is a dead insn, delete it and show registers in it aren't
7675 being used. */
7677 if (! live_insn)
7679 count_reg_usage (insn, counts, NULL_RTX, -1);
7680 delete_insn (insn);
7683 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7685 in_libcall = 0;
7686 dead_libcall = 0;
7690 /* Clean up. */
7691 free (counts);