Mark as release
[official-gcc.git] / gcc / cse.c
blob25257e92fb69ae966c0da8302d3afe1538fc950e
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2007 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 /* stdio.h must precede rtl.h for FFS. */
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "regs.h"
30 #include "basic-block.h"
31 #include "flags.h"
32 #include "real.h"
33 #include "insn-config.h"
34 #include "recog.h"
35 #include "function.h"
36 #include "expr.h"
37 #include "toplev.h"
38 #include "output.h"
39 #include "ggc.h"
40 #include "timevar.h"
41 #include "except.h"
42 #include "target.h"
43 #include "params.h"
44 #include "rtlhooks-def.h"
45 #include "tree-pass.h"
47 /* The basic idea of common subexpression elimination is to go
48 through the code, keeping a record of expressions that would
49 have the same value at the current scan point, and replacing
50 expressions encountered with the cheapest equivalent expression.
52 It is too complicated to keep track of the different possibilities
53 when control paths merge in this code; so, at each label, we forget all
54 that is known and start fresh. This can be described as processing each
55 extended basic block separately. We have a separate pass to perform
56 global CSE.
58 Note CSE can turn a conditional or computed jump into a nop or
59 an unconditional jump. When this occurs we arrange to run the jump
60 optimizer after CSE to delete the unreachable code.
62 We use two data structures to record the equivalent expressions:
63 a hash table for most expressions, and a vector of "quantity
64 numbers" to record equivalent (pseudo) registers.
66 The use of the special data structure for registers is desirable
67 because it is faster. It is possible because registers references
68 contain a fairly small number, the register number, taken from
69 a contiguously allocated series, and two register references are
70 identical if they have the same number. General expressions
71 do not have any such thing, so the only way to retrieve the
72 information recorded on an expression other than a register
73 is to keep it in a hash table.
75 Registers and "quantity numbers":
77 At the start of each basic block, all of the (hardware and pseudo)
78 registers used in the function are given distinct quantity
79 numbers to indicate their contents. During scan, when the code
80 copies one register into another, we copy the quantity number.
81 When a register is loaded in any other way, we allocate a new
82 quantity number to describe the value generated by this operation.
83 `REG_QTY (N)' records what quantity register N is currently thought
84 of as containing.
86 All real quantity numbers are greater than or equal to zero.
87 If register N has not been assigned a quantity, `REG_QTY (N)' will
88 equal -N - 1, which is always negative.
90 Quantity numbers below zero do not exist and none of the `qty_table'
91 entries should be referenced with a negative index.
93 We also maintain a bidirectional chain of registers for each
94 quantity number. The `qty_table` members `first_reg' and `last_reg',
95 and `reg_eqv_table' members `next' and `prev' hold these chains.
97 The first register in a chain is the one whose lifespan is least local.
98 Among equals, it is the one that was seen first.
99 We replace any equivalent register with that one.
101 If two registers have the same quantity number, it must be true that
102 REG expressions with qty_table `mode' must be in the hash table for both
103 registers and must be in the same class.
105 The converse is not true. Since hard registers may be referenced in
106 any mode, two REG expressions might be equivalent in the hash table
107 but not have the same quantity number if the quantity number of one
108 of the registers is not the same mode as those expressions.
110 Constants and quantity numbers
112 When a quantity has a known constant value, that value is stored
113 in the appropriate qty_table `const_rtx'. This is in addition to
114 putting the constant in the hash table as is usual for non-regs.
116 Whether a reg or a constant is preferred is determined by the configuration
117 macro CONST_COSTS and will often depend on the constant value. In any
118 event, expressions containing constants can be simplified, by fold_rtx.
120 When a quantity has a known nearly constant value (such as an address
121 of a stack slot), that value is stored in the appropriate qty_table
122 `const_rtx'.
124 Integer constants don't have a machine mode. However, cse
125 determines the intended machine mode from the destination
126 of the instruction that moves the constant. The machine mode
127 is recorded in the hash table along with the actual RTL
128 constant expression so that different modes are kept separate.
130 Other expressions:
132 To record known equivalences among expressions in general
133 we use a hash table called `table'. It has a fixed number of buckets
134 that contain chains of `struct table_elt' elements for expressions.
135 These chains connect the elements whose expressions have the same
136 hash codes.
138 Other chains through the same elements connect the elements which
139 currently have equivalent values.
141 Register references in an expression are canonicalized before hashing
142 the expression. This is done using `reg_qty' and qty_table `first_reg'.
143 The hash code of a register reference is computed using the quantity
144 number, not the register number.
146 When the value of an expression changes, it is necessary to remove from the
147 hash table not just that expression but all expressions whose values
148 could be different as a result.
150 1. If the value changing is in memory, except in special cases
151 ANYTHING referring to memory could be changed. That is because
152 nobody knows where a pointer does not point.
153 The function `invalidate_memory' removes what is necessary.
155 The special cases are when the address is constant or is
156 a constant plus a fixed register such as the frame pointer
157 or a static chain pointer. When such addresses are stored in,
158 we can tell exactly which other such addresses must be invalidated
159 due to overlap. `invalidate' does this.
160 All expressions that refer to non-constant
161 memory addresses are also invalidated. `invalidate_memory' does this.
163 2. If the value changing is a register, all expressions
164 containing references to that register, and only those,
165 must be removed.
167 Because searching the entire hash table for expressions that contain
168 a register is very slow, we try to figure out when it isn't necessary.
169 Precisely, this is necessary only when expressions have been
170 entered in the hash table using this register, and then the value has
171 changed, and then another expression wants to be added to refer to
172 the register's new value. This sequence of circumstances is rare
173 within any one basic block.
175 `REG_TICK' and `REG_IN_TABLE', accessors for members of
176 cse_reg_info, are used to detect this case. REG_TICK (i) is
177 incremented whenever a value is stored in register i.
178 REG_IN_TABLE (i) holds -1 if no references to register i have been
179 entered in the table; otherwise, it contains the value REG_TICK (i)
180 had when the references were entered. If we want to enter a
181 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
182 remove old references. Until we want to enter a new entry, the
183 mere fact that the two vectors don't match makes the entries be
184 ignored if anyone tries to match them.
186 Registers themselves are entered in the hash table as well as in
187 the equivalent-register chains. However, `REG_TICK' and
188 `REG_IN_TABLE' do not apply to expressions which are simple
189 register references. These expressions are removed from the table
190 immediately when they become invalid, and this can be done even if
191 we do not immediately search for all the expressions that refer to
192 the register.
194 A CLOBBER rtx in an instruction invalidates its operand for further
195 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
196 invalidates everything that resides in memory.
198 Related expressions:
200 Constant expressions that differ only by an additive integer
201 are called related. When a constant expression is put in
202 the table, the related expression with no constant term
203 is also entered. These are made to point at each other
204 so that it is possible to find out if there exists any
205 register equivalent to an expression related to a given expression. */
207 /* Length of qty_table vector. We know in advance we will not need
208 a quantity number this big. */
210 static int max_qty;
212 /* Next quantity number to be allocated.
213 This is 1 + the largest number needed so far. */
215 static int next_qty;
217 /* Per-qty information tracking.
219 `first_reg' and `last_reg' track the head and tail of the
220 chain of registers which currently contain this quantity.
222 `mode' contains the machine mode of this quantity.
224 `const_rtx' holds the rtx of the constant value of this
225 quantity, if known. A summations of the frame/arg pointer
226 and a constant can also be entered here. When this holds
227 a known value, `const_insn' is the insn which stored the
228 constant value.
230 `comparison_{code,const,qty}' are used to track when a
231 comparison between a quantity and some constant or register has
232 been passed. In such a case, we know the results of the comparison
233 in case we see it again. These members record a comparison that
234 is known to be true. `comparison_code' holds the rtx code of such
235 a comparison, else it is set to UNKNOWN and the other two
236 comparison members are undefined. `comparison_const' holds
237 the constant being compared against, or zero if the comparison
238 is not against a constant. `comparison_qty' holds the quantity
239 being compared against when the result is known. If the comparison
240 is not with a register, `comparison_qty' is -1. */
242 struct qty_table_elem
244 rtx const_rtx;
245 rtx const_insn;
246 rtx comparison_const;
247 int comparison_qty;
248 unsigned int first_reg, last_reg;
249 /* The sizes of these fields should match the sizes of the
250 code and mode fields of struct rtx_def (see rtl.h). */
251 ENUM_BITFIELD(rtx_code) comparison_code : 16;
252 ENUM_BITFIELD(machine_mode) mode : 8;
255 /* The table of all qtys, indexed by qty number. */
256 static struct qty_table_elem *qty_table;
258 /* Structure used to pass arguments via for_each_rtx to function
259 cse_change_cc_mode. */
260 struct change_cc_mode_args
262 rtx insn;
263 rtx newreg;
266 #ifdef HAVE_cc0
267 /* For machines that have a CC0, we do not record its value in the hash
268 table since its use is guaranteed to be the insn immediately following
269 its definition and any other insn is presumed to invalidate it.
271 Instead, we store below the value last assigned to CC0. If it should
272 happen to be a constant, it is stored in preference to the actual
273 assigned value. In case it is a constant, we store the mode in which
274 the constant should be interpreted. */
276 static rtx prev_insn_cc0;
277 static enum machine_mode prev_insn_cc0_mode;
279 /* Previous actual insn. 0 if at first insn of basic block. */
281 static rtx prev_insn;
282 #endif
284 /* Insn being scanned. */
286 static rtx this_insn;
288 /* Index by register number, gives the number of the next (or
289 previous) register in the chain of registers sharing the same
290 value.
292 Or -1 if this register is at the end of the chain.
294 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
296 /* Per-register equivalence chain. */
297 struct reg_eqv_elem
299 int next, prev;
302 /* The table of all register equivalence chains. */
303 static struct reg_eqv_elem *reg_eqv_table;
305 struct cse_reg_info
307 /* The timestamp at which this register is initialized. */
308 unsigned int timestamp;
310 /* The quantity number of the register's current contents. */
311 int reg_qty;
313 /* The number of times the register has been altered in the current
314 basic block. */
315 int reg_tick;
317 /* The REG_TICK value at which rtx's containing this register are
318 valid in the hash table. If this does not equal the current
319 reg_tick value, such expressions existing in the hash table are
320 invalid. */
321 int reg_in_table;
323 /* The SUBREG that was set when REG_TICK was last incremented. Set
324 to -1 if the last store was to the whole register, not a subreg. */
325 unsigned int subreg_ticked;
328 /* A table of cse_reg_info indexed by register numbers. */
329 static struct cse_reg_info *cse_reg_info_table;
331 /* The size of the above table. */
332 static unsigned int cse_reg_info_table_size;
334 /* The index of the first entry that has not been initialized. */
335 static unsigned int cse_reg_info_table_first_uninitialized;
337 /* The timestamp at the beginning of the current run of
338 cse_basic_block. We increment this variable at the beginning of
339 the current run of cse_basic_block. The timestamp field of a
340 cse_reg_info entry matches the value of this variable if and only
341 if the entry has been initialized during the current run of
342 cse_basic_block. */
343 static unsigned int cse_reg_info_timestamp;
345 /* A HARD_REG_SET containing all the hard registers for which there is
346 currently a REG expression in the hash table. Note the difference
347 from the above variables, which indicate if the REG is mentioned in some
348 expression in the table. */
350 static HARD_REG_SET hard_regs_in_table;
352 /* CUID of insn that starts the basic block currently being cse-processed. */
354 static int cse_basic_block_start;
356 /* CUID of insn that ends the basic block currently being cse-processed. */
358 static int cse_basic_block_end;
360 /* Vector mapping INSN_UIDs to cuids.
361 The cuids are like uids but increase monotonically always.
362 We use them to see whether a reg is used outside a given basic block. */
364 static int *uid_cuid;
366 /* Highest UID in UID_CUID. */
367 static int max_uid;
369 /* Get the cuid of an insn. */
371 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
373 /* Nonzero if this pass has made changes, and therefore it's
374 worthwhile to run the garbage collector. */
376 static int cse_altered;
378 /* Nonzero if cse has altered conditional jump insns
379 in such a way that jump optimization should be redone. */
381 static int cse_jumps_altered;
383 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
384 REG_LABEL, we have to rerun jump after CSE to put in the note. */
385 static int recorded_label_ref;
387 /* canon_hash stores 1 in do_not_record
388 if it notices a reference to CC0, PC, or some other volatile
389 subexpression. */
391 static int do_not_record;
393 /* canon_hash stores 1 in hash_arg_in_memory
394 if it notices a reference to memory within the expression being hashed. */
396 static int hash_arg_in_memory;
398 /* The hash table contains buckets which are chains of `struct table_elt's,
399 each recording one expression's information.
400 That expression is in the `exp' field.
402 The canon_exp field contains a canonical (from the point of view of
403 alias analysis) version of the `exp' field.
405 Those elements with the same hash code are chained in both directions
406 through the `next_same_hash' and `prev_same_hash' fields.
408 Each set of expressions with equivalent values
409 are on a two-way chain through the `next_same_value'
410 and `prev_same_value' fields, and all point with
411 the `first_same_value' field at the first element in
412 that chain. The chain is in order of increasing cost.
413 Each element's cost value is in its `cost' field.
415 The `in_memory' field is nonzero for elements that
416 involve any reference to memory. These elements are removed
417 whenever a write is done to an unidentified location in memory.
418 To be safe, we assume that a memory address is unidentified unless
419 the address is either a symbol constant or a constant plus
420 the frame pointer or argument pointer.
422 The `related_value' field is used to connect related expressions
423 (that differ by adding an integer).
424 The related expressions are chained in a circular fashion.
425 `related_value' is zero for expressions for which this
426 chain is not useful.
428 The `cost' field stores the cost of this element's expression.
429 The `regcost' field stores the value returned by approx_reg_cost for
430 this element's expression.
432 The `is_const' flag is set if the element is a constant (including
433 a fixed address).
435 The `flag' field is used as a temporary during some search routines.
437 The `mode' field is usually the same as GET_MODE (`exp'), but
438 if `exp' is a CONST_INT and has no machine mode then the `mode'
439 field is the mode it was being used as. Each constant is
440 recorded separately for each mode it is used with. */
442 struct table_elt
444 rtx exp;
445 rtx canon_exp;
446 struct table_elt *next_same_hash;
447 struct table_elt *prev_same_hash;
448 struct table_elt *next_same_value;
449 struct table_elt *prev_same_value;
450 struct table_elt *first_same_value;
451 struct table_elt *related_value;
452 int cost;
453 int regcost;
454 /* The size of this field should match the size
455 of the mode field of struct rtx_def (see rtl.h). */
456 ENUM_BITFIELD(machine_mode) mode : 8;
457 char in_memory;
458 char is_const;
459 char flag;
462 /* We don't want a lot of buckets, because we rarely have very many
463 things stored in the hash table, and a lot of buckets slows
464 down a lot of loops that happen frequently. */
465 #define HASH_SHIFT 5
466 #define HASH_SIZE (1 << HASH_SHIFT)
467 #define HASH_MASK (HASH_SIZE - 1)
469 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
470 register (hard registers may require `do_not_record' to be set). */
472 #define HASH(X, M) \
473 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
474 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
475 : canon_hash (X, M)) & HASH_MASK)
477 /* Like HASH, but without side-effects. */
478 #define SAFE_HASH(X, M) \
479 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
480 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
481 : safe_hash (X, M)) & HASH_MASK)
483 /* Determine whether register number N is considered a fixed register for the
484 purpose of approximating register costs.
485 It is desirable to replace other regs with fixed regs, to reduce need for
486 non-fixed hard regs.
487 A reg wins if it is either the frame pointer or designated as fixed. */
488 #define FIXED_REGNO_P(N) \
489 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
490 || fixed_regs[N] || global_regs[N])
492 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
493 hard registers and pointers into the frame are the cheapest with a cost
494 of 0. Next come pseudos with a cost of one and other hard registers with
495 a cost of 2. Aside from these special cases, call `rtx_cost'. */
497 #define CHEAP_REGNO(N) \
498 (REGNO_PTR_FRAME_P(N) \
499 || (HARD_REGISTER_NUM_P (N) \
500 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
502 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
503 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
505 /* Get the number of times this register has been updated in this
506 basic block. */
508 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
510 /* Get the point at which REG was recorded in the table. */
512 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
514 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
515 SUBREG). */
517 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
519 /* Get the quantity number for REG. */
521 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
523 /* Determine if the quantity number for register X represents a valid index
524 into the qty_table. */
526 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
528 static struct table_elt *table[HASH_SIZE];
530 /* Number of elements in the hash table. */
532 static unsigned int table_size;
534 /* Chain of `struct table_elt's made so far for this function
535 but currently removed from the table. */
537 static struct table_elt *free_element_chain;
539 /* Set to the cost of a constant pool reference if one was found for a
540 symbolic constant. If this was found, it means we should try to
541 convert constants into constant pool entries if they don't fit in
542 the insn. */
544 static int constant_pool_entries_cost;
545 static int constant_pool_entries_regcost;
547 /* This data describes a block that will be processed by cse_basic_block. */
549 struct cse_basic_block_data
551 /* Lowest CUID value of insns in block. */
552 int low_cuid;
553 /* Highest CUID value of insns in block. */
554 int high_cuid;
555 /* Total number of SETs in block. */
556 int nsets;
557 /* Last insn in the block. */
558 rtx last;
559 /* Size of current branch path, if any. */
560 int path_size;
561 /* Current branch path, indicating which branches will be taken. */
562 struct branch_path
564 /* The branch insn. */
565 rtx branch;
566 /* Whether it should be taken or not. AROUND is the same as taken
567 except that it is used when the destination label is not preceded
568 by a BARRIER. */
569 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
570 } *path;
573 static bool fixed_base_plus_p (rtx x);
574 static int notreg_cost (rtx, enum rtx_code);
575 static int approx_reg_cost_1 (rtx *, void *);
576 static int approx_reg_cost (rtx);
577 static int preferable (int, int, int, int);
578 static void new_basic_block (void);
579 static void make_new_qty (unsigned int, enum machine_mode);
580 static void make_regs_eqv (unsigned int, unsigned int);
581 static void delete_reg_equiv (unsigned int);
582 static int mention_regs (rtx);
583 static int insert_regs (rtx, struct table_elt *, int);
584 static void remove_from_table (struct table_elt *, unsigned);
585 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
586 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
587 static rtx lookup_as_function (rtx, enum rtx_code);
588 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
589 enum machine_mode);
590 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
591 static void invalidate (rtx, enum machine_mode);
592 static int cse_rtx_varies_p (rtx, int);
593 static void remove_invalid_refs (unsigned int);
594 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
595 enum machine_mode);
596 static void rehash_using_reg (rtx);
597 static void invalidate_memory (void);
598 static void invalidate_for_call (void);
599 static rtx use_related_value (rtx, struct table_elt *);
601 static inline unsigned canon_hash (rtx, enum machine_mode);
602 static inline unsigned safe_hash (rtx, enum machine_mode);
603 static unsigned hash_rtx_string (const char *);
605 static rtx canon_reg (rtx, rtx);
606 static void find_best_addr (rtx, rtx *, enum machine_mode);
607 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
608 enum machine_mode *,
609 enum machine_mode *);
610 static rtx fold_rtx (rtx, rtx);
611 static rtx equiv_constant (rtx);
612 static void record_jump_equiv (rtx, int);
613 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
614 int);
615 static void cse_insn (rtx, rtx);
616 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
617 int, int);
618 static int addr_affects_sp_p (rtx);
619 static void invalidate_from_clobbers (rtx);
620 static rtx cse_process_notes (rtx, rtx);
621 static void invalidate_skipped_set (rtx, rtx, void *);
622 static void invalidate_skipped_block (rtx);
623 static rtx cse_basic_block (rtx, rtx, struct branch_path *);
624 static void count_reg_usage (rtx, int *, rtx, int);
625 static int check_for_label_ref (rtx *, void *);
626 extern void dump_class (struct table_elt*);
627 static void get_cse_reg_info_1 (unsigned int regno);
628 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
629 static int check_dependence (rtx *, void *);
631 static void flush_hash_table (void);
632 static bool insn_live_p (rtx, int *);
633 static bool set_live_p (rtx, rtx, int *);
634 static bool dead_libcall_p (rtx, int *);
635 static int cse_change_cc_mode (rtx *, void *);
636 static void cse_change_cc_mode_insn (rtx, rtx);
637 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
638 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
641 #undef RTL_HOOKS_GEN_LOWPART
642 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
644 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
646 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
647 virtual regs here because the simplify_*_operation routines are called
648 by integrate.c, which is called before virtual register instantiation. */
650 static bool
651 fixed_base_plus_p (rtx x)
653 switch (GET_CODE (x))
655 case REG:
656 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
657 return true;
658 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
659 return true;
660 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
661 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
662 return true;
663 return false;
665 case PLUS:
666 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
667 return false;
668 return fixed_base_plus_p (XEXP (x, 0));
670 default:
671 return false;
675 /* Dump the expressions in the equivalence class indicated by CLASSP.
676 This function is used only for debugging. */
677 void
678 dump_class (struct table_elt *classp)
680 struct table_elt *elt;
682 fprintf (stderr, "Equivalence chain for ");
683 print_rtl (stderr, classp->exp);
684 fprintf (stderr, ": \n");
686 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
688 print_rtl (stderr, elt->exp);
689 fprintf (stderr, "\n");
693 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
695 static int
696 approx_reg_cost_1 (rtx *xp, void *data)
698 rtx x = *xp;
699 int *cost_p = data;
701 if (x && REG_P (x))
703 unsigned int regno = REGNO (x);
705 if (! CHEAP_REGNO (regno))
707 if (regno < FIRST_PSEUDO_REGISTER)
709 if (SMALL_REGISTER_CLASSES)
710 return 1;
711 *cost_p += 2;
713 else
714 *cost_p += 1;
718 return 0;
721 /* Return an estimate of the cost of the registers used in an rtx.
722 This is mostly the number of different REG expressions in the rtx;
723 however for some exceptions like fixed registers we use a cost of
724 0. If any other hard register reference occurs, return MAX_COST. */
726 static int
727 approx_reg_cost (rtx x)
729 int cost = 0;
731 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
732 return MAX_COST;
734 return cost;
737 /* Returns a canonical version of X for the address, from the point of view,
738 that all multiplications are represented as MULT instead of the multiply
739 by a power of 2 being represented as ASHIFT. */
741 static rtx
742 canon_for_address (rtx x)
744 enum rtx_code code;
745 enum machine_mode mode;
746 rtx new = 0;
747 int i;
748 const char *fmt;
750 if (!x)
751 return x;
753 code = GET_CODE (x);
754 mode = GET_MODE (x);
756 switch (code)
758 case ASHIFT:
759 if (GET_CODE (XEXP (x, 1)) == CONST_INT
760 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)
761 && INTVAL (XEXP (x, 1)) >= 0)
763 new = canon_for_address (XEXP (x, 0));
764 new = gen_rtx_MULT (mode, new,
765 gen_int_mode ((HOST_WIDE_INT) 1
766 << INTVAL (XEXP (x, 1)),
767 mode));
769 break;
770 default:
771 break;
774 if (new)
775 return new;
777 /* Now recursively process each operand of this operation. */
778 fmt = GET_RTX_FORMAT (code);
779 for (i = 0; i < GET_RTX_LENGTH (code); i++)
780 if (fmt[i] == 'e')
782 new = canon_for_address (XEXP (x, i));
783 XEXP (x, i) = new;
785 return x;
788 /* Return a negative value if an rtx A, whose costs are given by COST_A
789 and REGCOST_A, is more desirable than an rtx B.
790 Return a positive value if A is less desirable, or 0 if the two are
791 equally good. */
792 static int
793 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
795 /* First, get rid of cases involving expressions that are entirely
796 unwanted. */
797 if (cost_a != cost_b)
799 if (cost_a == MAX_COST)
800 return 1;
801 if (cost_b == MAX_COST)
802 return -1;
805 /* Avoid extending lifetimes of hardregs. */
806 if (regcost_a != regcost_b)
808 if (regcost_a == MAX_COST)
809 return 1;
810 if (regcost_b == MAX_COST)
811 return -1;
814 /* Normal operation costs take precedence. */
815 if (cost_a != cost_b)
816 return cost_a - cost_b;
817 /* Only if these are identical consider effects on register pressure. */
818 if (regcost_a != regcost_b)
819 return regcost_a - regcost_b;
820 return 0;
823 /* Internal function, to compute cost when X is not a register; called
824 from COST macro to keep it simple. */
826 static int
827 notreg_cost (rtx x, enum rtx_code outer)
829 return ((GET_CODE (x) == SUBREG
830 && REG_P (SUBREG_REG (x))
831 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
832 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
833 && (GET_MODE_SIZE (GET_MODE (x))
834 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
835 && subreg_lowpart_p (x)
836 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
837 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
839 : rtx_cost (x, outer) * 2);
843 /* Initialize CSE_REG_INFO_TABLE. */
845 static void
846 init_cse_reg_info (unsigned int nregs)
848 /* Do we need to grow the table? */
849 if (nregs > cse_reg_info_table_size)
851 unsigned int new_size;
853 if (cse_reg_info_table_size < 2048)
855 /* Compute a new size that is a power of 2 and no smaller
856 than the large of NREGS and 64. */
857 new_size = (cse_reg_info_table_size
858 ? cse_reg_info_table_size : 64);
860 while (new_size < nregs)
861 new_size *= 2;
863 else
865 /* If we need a big table, allocate just enough to hold
866 NREGS registers. */
867 new_size = nregs;
870 /* Reallocate the table with NEW_SIZE entries. */
871 if (cse_reg_info_table)
872 free (cse_reg_info_table);
873 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
874 cse_reg_info_table_size = new_size;
875 cse_reg_info_table_first_uninitialized = 0;
878 /* Do we have all of the first NREGS entries initialized? */
879 if (cse_reg_info_table_first_uninitialized < nregs)
881 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
882 unsigned int i;
884 /* Put the old timestamp on newly allocated entries so that they
885 will all be considered out of date. We do not touch those
886 entries beyond the first NREGS entries to be nice to the
887 virtual memory. */
888 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
889 cse_reg_info_table[i].timestamp = old_timestamp;
891 cse_reg_info_table_first_uninitialized = nregs;
895 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
897 static void
898 get_cse_reg_info_1 (unsigned int regno)
900 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
901 entry will be considered to have been initialized. */
902 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
904 /* Initialize the rest of the entry. */
905 cse_reg_info_table[regno].reg_tick = 1;
906 cse_reg_info_table[regno].reg_in_table = -1;
907 cse_reg_info_table[regno].subreg_ticked = -1;
908 cse_reg_info_table[regno].reg_qty = -regno - 1;
911 /* Find a cse_reg_info entry for REGNO. */
913 static inline struct cse_reg_info *
914 get_cse_reg_info (unsigned int regno)
916 struct cse_reg_info *p = &cse_reg_info_table[regno];
918 /* If this entry has not been initialized, go ahead and initialize
919 it. */
920 if (p->timestamp != cse_reg_info_timestamp)
921 get_cse_reg_info_1 (regno);
923 return p;
926 /* Clear the hash table and initialize each register with its own quantity,
927 for a new basic block. */
929 static void
930 new_basic_block (void)
932 int i;
934 next_qty = 0;
936 /* Invalidate cse_reg_info_table. */
937 cse_reg_info_timestamp++;
939 /* Clear out hash table state for this pass. */
940 CLEAR_HARD_REG_SET (hard_regs_in_table);
942 /* The per-quantity values used to be initialized here, but it is
943 much faster to initialize each as it is made in `make_new_qty'. */
945 for (i = 0; i < HASH_SIZE; i++)
947 struct table_elt *first;
949 first = table[i];
950 if (first != NULL)
952 struct table_elt *last = first;
954 table[i] = NULL;
956 while (last->next_same_hash != NULL)
957 last = last->next_same_hash;
959 /* Now relink this hash entire chain into
960 the free element list. */
962 last->next_same_hash = free_element_chain;
963 free_element_chain = first;
967 table_size = 0;
969 #ifdef HAVE_cc0
970 prev_insn = 0;
971 prev_insn_cc0 = 0;
972 #endif
975 /* Say that register REG contains a quantity in mode MODE not in any
976 register before and initialize that quantity. */
978 static void
979 make_new_qty (unsigned int reg, enum machine_mode mode)
981 int q;
982 struct qty_table_elem *ent;
983 struct reg_eqv_elem *eqv;
985 gcc_assert (next_qty < max_qty);
987 q = REG_QTY (reg) = next_qty++;
988 ent = &qty_table[q];
989 ent->first_reg = reg;
990 ent->last_reg = reg;
991 ent->mode = mode;
992 ent->const_rtx = ent->const_insn = NULL_RTX;
993 ent->comparison_code = UNKNOWN;
995 eqv = &reg_eqv_table[reg];
996 eqv->next = eqv->prev = -1;
999 /* Make reg NEW equivalent to reg OLD.
1000 OLD is not changing; NEW is. */
1002 static void
1003 make_regs_eqv (unsigned int new, unsigned int old)
1005 unsigned int lastr, firstr;
1006 int q = REG_QTY (old);
1007 struct qty_table_elem *ent;
1009 ent = &qty_table[q];
1011 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1012 gcc_assert (REGNO_QTY_VALID_P (old));
1014 REG_QTY (new) = q;
1015 firstr = ent->first_reg;
1016 lastr = ent->last_reg;
1018 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1019 hard regs. Among pseudos, if NEW will live longer than any other reg
1020 of the same qty, and that is beyond the current basic block,
1021 make it the new canonical replacement for this qty. */
1022 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1023 /* Certain fixed registers might be of the class NO_REGS. This means
1024 that not only can they not be allocated by the compiler, but
1025 they cannot be used in substitutions or canonicalizations
1026 either. */
1027 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1028 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1029 || (new >= FIRST_PSEUDO_REGISTER
1030 && (firstr < FIRST_PSEUDO_REGISTER
1031 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1032 || (uid_cuid[REGNO_FIRST_UID (new)]
1033 < cse_basic_block_start))
1034 && (uid_cuid[REGNO_LAST_UID (new)]
1035 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1037 reg_eqv_table[firstr].prev = new;
1038 reg_eqv_table[new].next = firstr;
1039 reg_eqv_table[new].prev = -1;
1040 ent->first_reg = new;
1042 else
1044 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1045 Otherwise, insert before any non-fixed hard regs that are at the
1046 end. Registers of class NO_REGS cannot be used as an
1047 equivalent for anything. */
1048 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1049 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1050 && new >= FIRST_PSEUDO_REGISTER)
1051 lastr = reg_eqv_table[lastr].prev;
1052 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1053 if (reg_eqv_table[lastr].next >= 0)
1054 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1055 else
1056 qty_table[q].last_reg = new;
1057 reg_eqv_table[lastr].next = new;
1058 reg_eqv_table[new].prev = lastr;
1062 /* Remove REG from its equivalence class. */
1064 static void
1065 delete_reg_equiv (unsigned int reg)
1067 struct qty_table_elem *ent;
1068 int q = REG_QTY (reg);
1069 int p, n;
1071 /* If invalid, do nothing. */
1072 if (! REGNO_QTY_VALID_P (reg))
1073 return;
1075 ent = &qty_table[q];
1077 p = reg_eqv_table[reg].prev;
1078 n = reg_eqv_table[reg].next;
1080 if (n != -1)
1081 reg_eqv_table[n].prev = p;
1082 else
1083 ent->last_reg = p;
1084 if (p != -1)
1085 reg_eqv_table[p].next = n;
1086 else
1087 ent->first_reg = n;
1089 REG_QTY (reg) = -reg - 1;
1092 /* Remove any invalid expressions from the hash table
1093 that refer to any of the registers contained in expression X.
1095 Make sure that newly inserted references to those registers
1096 as subexpressions will be considered valid.
1098 mention_regs is not called when a register itself
1099 is being stored in the table.
1101 Return 1 if we have done something that may have changed the hash code
1102 of X. */
1104 static int
1105 mention_regs (rtx x)
1107 enum rtx_code code;
1108 int i, j;
1109 const char *fmt;
1110 int changed = 0;
1112 if (x == 0)
1113 return 0;
1115 code = GET_CODE (x);
1116 if (code == REG)
1118 unsigned int regno = REGNO (x);
1119 unsigned int endregno
1120 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1121 : hard_regno_nregs[regno][GET_MODE (x)]);
1122 unsigned int i;
1124 for (i = regno; i < endregno; i++)
1126 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1127 remove_invalid_refs (i);
1129 REG_IN_TABLE (i) = REG_TICK (i);
1130 SUBREG_TICKED (i) = -1;
1133 return 0;
1136 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1137 pseudo if they don't use overlapping words. We handle only pseudos
1138 here for simplicity. */
1139 if (code == SUBREG && REG_P (SUBREG_REG (x))
1140 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1142 unsigned int i = REGNO (SUBREG_REG (x));
1144 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1146 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1147 the last store to this register really stored into this
1148 subreg, then remove the memory of this subreg.
1149 Otherwise, remove any memory of the entire register and
1150 all its subregs from the table. */
1151 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1152 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1153 remove_invalid_refs (i);
1154 else
1155 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1158 REG_IN_TABLE (i) = REG_TICK (i);
1159 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1160 return 0;
1163 /* If X is a comparison or a COMPARE and either operand is a register
1164 that does not have a quantity, give it one. This is so that a later
1165 call to record_jump_equiv won't cause X to be assigned a different
1166 hash code and not found in the table after that call.
1168 It is not necessary to do this here, since rehash_using_reg can
1169 fix up the table later, but doing this here eliminates the need to
1170 call that expensive function in the most common case where the only
1171 use of the register is in the comparison. */
1173 if (code == COMPARE || COMPARISON_P (x))
1175 if (REG_P (XEXP (x, 0))
1176 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1177 if (insert_regs (XEXP (x, 0), NULL, 0))
1179 rehash_using_reg (XEXP (x, 0));
1180 changed = 1;
1183 if (REG_P (XEXP (x, 1))
1184 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1185 if (insert_regs (XEXP (x, 1), NULL, 0))
1187 rehash_using_reg (XEXP (x, 1));
1188 changed = 1;
1192 fmt = GET_RTX_FORMAT (code);
1193 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1194 if (fmt[i] == 'e')
1195 changed |= mention_regs (XEXP (x, i));
1196 else if (fmt[i] == 'E')
1197 for (j = 0; j < XVECLEN (x, i); j++)
1198 changed |= mention_regs (XVECEXP (x, i, j));
1200 return changed;
1203 /* Update the register quantities for inserting X into the hash table
1204 with a value equivalent to CLASSP.
1205 (If the class does not contain a REG, it is irrelevant.)
1206 If MODIFIED is nonzero, X is a destination; it is being modified.
1207 Note that delete_reg_equiv should be called on a register
1208 before insert_regs is done on that register with MODIFIED != 0.
1210 Nonzero value means that elements of reg_qty have changed
1211 so X's hash code may be different. */
1213 static int
1214 insert_regs (rtx x, struct table_elt *classp, int modified)
1216 if (REG_P (x))
1218 unsigned int regno = REGNO (x);
1219 int qty_valid;
1221 /* If REGNO is in the equivalence table already but is of the
1222 wrong mode for that equivalence, don't do anything here. */
1224 qty_valid = REGNO_QTY_VALID_P (regno);
1225 if (qty_valid)
1227 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1229 if (ent->mode != GET_MODE (x))
1230 return 0;
1233 if (modified || ! qty_valid)
1235 if (classp)
1236 for (classp = classp->first_same_value;
1237 classp != 0;
1238 classp = classp->next_same_value)
1239 if (REG_P (classp->exp)
1240 && GET_MODE (classp->exp) == GET_MODE (x))
1242 unsigned c_regno = REGNO (classp->exp);
1244 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1246 /* Suppose that 5 is hard reg and 100 and 101 are
1247 pseudos. Consider
1249 (set (reg:si 100) (reg:si 5))
1250 (set (reg:si 5) (reg:si 100))
1251 (set (reg:di 101) (reg:di 5))
1253 We would now set REG_QTY (101) = REG_QTY (5), but the
1254 entry for 5 is in SImode. When we use this later in
1255 copy propagation, we get the register in wrong mode. */
1256 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1257 continue;
1259 make_regs_eqv (regno, c_regno);
1260 return 1;
1263 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1264 than REG_IN_TABLE to find out if there was only a single preceding
1265 invalidation - for the SUBREG - or another one, which would be
1266 for the full register. However, if we find here that REG_TICK
1267 indicates that the register is invalid, it means that it has
1268 been invalidated in a separate operation. The SUBREG might be used
1269 now (then this is a recursive call), or we might use the full REG
1270 now and a SUBREG of it later. So bump up REG_TICK so that
1271 mention_regs will do the right thing. */
1272 if (! modified
1273 && REG_IN_TABLE (regno) >= 0
1274 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1275 REG_TICK (regno)++;
1276 make_new_qty (regno, GET_MODE (x));
1277 return 1;
1280 return 0;
1283 /* If X is a SUBREG, we will likely be inserting the inner register in the
1284 table. If that register doesn't have an assigned quantity number at
1285 this point but does later, the insertion that we will be doing now will
1286 not be accessible because its hash code will have changed. So assign
1287 a quantity number now. */
1289 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1290 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1292 insert_regs (SUBREG_REG (x), NULL, 0);
1293 mention_regs (x);
1294 return 1;
1296 else
1297 return mention_regs (x);
1300 /* Look in or update the hash table. */
1302 /* Remove table element ELT from use in the table.
1303 HASH is its hash code, made using the HASH macro.
1304 It's an argument because often that is known in advance
1305 and we save much time not recomputing it. */
1307 static void
1308 remove_from_table (struct table_elt *elt, unsigned int hash)
1310 if (elt == 0)
1311 return;
1313 /* Mark this element as removed. See cse_insn. */
1314 elt->first_same_value = 0;
1316 /* Remove the table element from its equivalence class. */
1319 struct table_elt *prev = elt->prev_same_value;
1320 struct table_elt *next = elt->next_same_value;
1322 if (next)
1323 next->prev_same_value = prev;
1325 if (prev)
1326 prev->next_same_value = next;
1327 else
1329 struct table_elt *newfirst = next;
1330 while (next)
1332 next->first_same_value = newfirst;
1333 next = next->next_same_value;
1338 /* Remove the table element from its hash bucket. */
1341 struct table_elt *prev = elt->prev_same_hash;
1342 struct table_elt *next = elt->next_same_hash;
1344 if (next)
1345 next->prev_same_hash = prev;
1347 if (prev)
1348 prev->next_same_hash = next;
1349 else if (table[hash] == elt)
1350 table[hash] = next;
1351 else
1353 /* This entry is not in the proper hash bucket. This can happen
1354 when two classes were merged by `merge_equiv_classes'. Search
1355 for the hash bucket that it heads. This happens only very
1356 rarely, so the cost is acceptable. */
1357 for (hash = 0; hash < HASH_SIZE; hash++)
1358 if (table[hash] == elt)
1359 table[hash] = next;
1363 /* Remove the table element from its related-value circular chain. */
1365 if (elt->related_value != 0 && elt->related_value != elt)
1367 struct table_elt *p = elt->related_value;
1369 while (p->related_value != elt)
1370 p = p->related_value;
1371 p->related_value = elt->related_value;
1372 if (p->related_value == p)
1373 p->related_value = 0;
1376 /* Now add it to the free element chain. */
1377 elt->next_same_hash = free_element_chain;
1378 free_element_chain = elt;
1380 table_size--;
1383 /* Look up X in the hash table and return its table element,
1384 or 0 if X is not in the table.
1386 MODE is the machine-mode of X, or if X is an integer constant
1387 with VOIDmode then MODE is the mode with which X will be used.
1389 Here we are satisfied to find an expression whose tree structure
1390 looks like X. */
1392 static struct table_elt *
1393 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1395 struct table_elt *p;
1397 for (p = table[hash]; p; p = p->next_same_hash)
1398 if (mode == p->mode && ((x == p->exp && REG_P (x))
1399 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1400 return p;
1402 return 0;
1405 /* Like `lookup' but don't care whether the table element uses invalid regs.
1406 Also ignore discrepancies in the machine mode of a register. */
1408 static struct table_elt *
1409 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1411 struct table_elt *p;
1413 if (REG_P (x))
1415 unsigned int regno = REGNO (x);
1417 /* Don't check the machine mode when comparing registers;
1418 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1419 for (p = table[hash]; p; p = p->next_same_hash)
1420 if (REG_P (p->exp)
1421 && REGNO (p->exp) == regno)
1422 return p;
1424 else
1426 for (p = table[hash]; p; p = p->next_same_hash)
1427 if (mode == p->mode
1428 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1429 return p;
1432 return 0;
1435 /* Look for an expression equivalent to X and with code CODE.
1436 If one is found, return that expression. */
1438 static rtx
1439 lookup_as_function (rtx x, enum rtx_code code)
1441 struct table_elt *p
1442 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1444 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1445 long as we are narrowing. So if we looked in vain for a mode narrower
1446 than word_mode before, look for word_mode now. */
1447 if (p == 0 && code == CONST_INT
1448 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1450 x = copy_rtx (x);
1451 PUT_MODE (x, word_mode);
1452 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1455 if (p == 0)
1456 return 0;
1458 for (p = p->first_same_value; p; p = p->next_same_value)
1459 if (GET_CODE (p->exp) == code
1460 /* Make sure this is a valid entry in the table. */
1461 && exp_equiv_p (p->exp, p->exp, 1, false))
1462 return p->exp;
1464 return 0;
1467 /* Insert X in the hash table, assuming HASH is its hash code
1468 and CLASSP is an element of the class it should go in
1469 (or 0 if a new class should be made).
1470 It is inserted at the proper position to keep the class in
1471 the order cheapest first.
1473 MODE is the machine-mode of X, or if X is an integer constant
1474 with VOIDmode then MODE is the mode with which X will be used.
1476 For elements of equal cheapness, the most recent one
1477 goes in front, except that the first element in the list
1478 remains first unless a cheaper element is added. The order of
1479 pseudo-registers does not matter, as canon_reg will be called to
1480 find the cheapest when a register is retrieved from the table.
1482 The in_memory field in the hash table element is set to 0.
1483 The caller must set it nonzero if appropriate.
1485 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1486 and if insert_regs returns a nonzero value
1487 you must then recompute its hash code before calling here.
1489 If necessary, update table showing constant values of quantities. */
1491 #define CHEAPER(X, Y) \
1492 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1494 static struct table_elt *
1495 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1497 struct table_elt *elt;
1499 /* If X is a register and we haven't made a quantity for it,
1500 something is wrong. */
1501 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1503 /* If X is a hard register, show it is being put in the table. */
1504 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1506 unsigned int regno = REGNO (x);
1507 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1508 unsigned int i;
1510 for (i = regno; i < endregno; i++)
1511 SET_HARD_REG_BIT (hard_regs_in_table, i);
1514 /* Put an element for X into the right hash bucket. */
1516 elt = free_element_chain;
1517 if (elt)
1518 free_element_chain = elt->next_same_hash;
1519 else
1520 elt = XNEW (struct table_elt);
1522 elt->exp = x;
1523 elt->canon_exp = NULL_RTX;
1524 elt->cost = COST (x);
1525 elt->regcost = approx_reg_cost (x);
1526 elt->next_same_value = 0;
1527 elt->prev_same_value = 0;
1528 elt->next_same_hash = table[hash];
1529 elt->prev_same_hash = 0;
1530 elt->related_value = 0;
1531 elt->in_memory = 0;
1532 elt->mode = mode;
1533 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1535 if (table[hash])
1536 table[hash]->prev_same_hash = elt;
1537 table[hash] = elt;
1539 /* Put it into the proper value-class. */
1540 if (classp)
1542 classp = classp->first_same_value;
1543 if (CHEAPER (elt, classp))
1544 /* Insert at the head of the class. */
1546 struct table_elt *p;
1547 elt->next_same_value = classp;
1548 classp->prev_same_value = elt;
1549 elt->first_same_value = elt;
1551 for (p = classp; p; p = p->next_same_value)
1552 p->first_same_value = elt;
1554 else
1556 /* Insert not at head of the class. */
1557 /* Put it after the last element cheaper than X. */
1558 struct table_elt *p, *next;
1560 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1561 p = next);
1563 /* Put it after P and before NEXT. */
1564 elt->next_same_value = next;
1565 if (next)
1566 next->prev_same_value = elt;
1568 elt->prev_same_value = p;
1569 p->next_same_value = elt;
1570 elt->first_same_value = classp;
1573 else
1574 elt->first_same_value = elt;
1576 /* If this is a constant being set equivalent to a register or a register
1577 being set equivalent to a constant, note the constant equivalence.
1579 If this is a constant, it cannot be equivalent to a different constant,
1580 and a constant is the only thing that can be cheaper than a register. So
1581 we know the register is the head of the class (before the constant was
1582 inserted).
1584 If this is a register that is not already known equivalent to a
1585 constant, we must check the entire class.
1587 If this is a register that is already known equivalent to an insn,
1588 update the qtys `const_insn' to show that `this_insn' is the latest
1589 insn making that quantity equivalent to the constant. */
1591 if (elt->is_const && classp && REG_P (classp->exp)
1592 && !REG_P (x))
1594 int exp_q = REG_QTY (REGNO (classp->exp));
1595 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1597 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1598 exp_ent->const_insn = this_insn;
1601 else if (REG_P (x)
1602 && classp
1603 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1604 && ! elt->is_const)
1606 struct table_elt *p;
1608 for (p = classp; p != 0; p = p->next_same_value)
1610 if (p->is_const && !REG_P (p->exp))
1612 int x_q = REG_QTY (REGNO (x));
1613 struct qty_table_elem *x_ent = &qty_table[x_q];
1615 x_ent->const_rtx
1616 = gen_lowpart (GET_MODE (x), p->exp);
1617 x_ent->const_insn = this_insn;
1618 break;
1623 else if (REG_P (x)
1624 && qty_table[REG_QTY (REGNO (x))].const_rtx
1625 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1626 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1628 /* If this is a constant with symbolic value,
1629 and it has a term with an explicit integer value,
1630 link it up with related expressions. */
1631 if (GET_CODE (x) == CONST)
1633 rtx subexp = get_related_value (x);
1634 unsigned subhash;
1635 struct table_elt *subelt, *subelt_prev;
1637 if (subexp != 0)
1639 /* Get the integer-free subexpression in the hash table. */
1640 subhash = SAFE_HASH (subexp, mode);
1641 subelt = lookup (subexp, subhash, mode);
1642 if (subelt == 0)
1643 subelt = insert (subexp, NULL, subhash, mode);
1644 /* Initialize SUBELT's circular chain if it has none. */
1645 if (subelt->related_value == 0)
1646 subelt->related_value = subelt;
1647 /* Find the element in the circular chain that precedes SUBELT. */
1648 subelt_prev = subelt;
1649 while (subelt_prev->related_value != subelt)
1650 subelt_prev = subelt_prev->related_value;
1651 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1652 This way the element that follows SUBELT is the oldest one. */
1653 elt->related_value = subelt_prev->related_value;
1654 subelt_prev->related_value = elt;
1658 table_size++;
1660 return elt;
1663 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1664 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1665 the two classes equivalent.
1667 CLASS1 will be the surviving class; CLASS2 should not be used after this
1668 call.
1670 Any invalid entries in CLASS2 will not be copied. */
1672 static void
1673 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1675 struct table_elt *elt, *next, *new;
1677 /* Ensure we start with the head of the classes. */
1678 class1 = class1->first_same_value;
1679 class2 = class2->first_same_value;
1681 /* If they were already equal, forget it. */
1682 if (class1 == class2)
1683 return;
1685 for (elt = class2; elt; elt = next)
1687 unsigned int hash;
1688 rtx exp = elt->exp;
1689 enum machine_mode mode = elt->mode;
1691 next = elt->next_same_value;
1693 /* Remove old entry, make a new one in CLASS1's class.
1694 Don't do this for invalid entries as we cannot find their
1695 hash code (it also isn't necessary). */
1696 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1698 bool need_rehash = false;
1700 hash_arg_in_memory = 0;
1701 hash = HASH (exp, mode);
1703 if (REG_P (exp))
1705 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1706 delete_reg_equiv (REGNO (exp));
1709 remove_from_table (elt, hash);
1711 if (insert_regs (exp, class1, 0) || need_rehash)
1713 rehash_using_reg (exp);
1714 hash = HASH (exp, mode);
1716 new = insert (exp, class1, hash, mode);
1717 new->in_memory = hash_arg_in_memory;
1722 /* Flush the entire hash table. */
1724 static void
1725 flush_hash_table (void)
1727 int i;
1728 struct table_elt *p;
1730 for (i = 0; i < HASH_SIZE; i++)
1731 for (p = table[i]; p; p = table[i])
1733 /* Note that invalidate can remove elements
1734 after P in the current hash chain. */
1735 if (REG_P (p->exp))
1736 invalidate (p->exp, VOIDmode);
1737 else
1738 remove_from_table (p, i);
1742 /* Function called for each rtx to check whether true dependence exist. */
1743 struct check_dependence_data
1745 enum machine_mode mode;
1746 rtx exp;
1747 rtx addr;
1750 static int
1751 check_dependence (rtx *x, void *data)
1753 struct check_dependence_data *d = (struct check_dependence_data *) data;
1754 if (*x && MEM_P (*x))
1755 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1756 cse_rtx_varies_p);
1757 else
1758 return 0;
1761 /* Remove from the hash table, or mark as invalid, all expressions whose
1762 values could be altered by storing in X. X is a register, a subreg, or
1763 a memory reference with nonvarying address (because, when a memory
1764 reference with a varying address is stored in, all memory references are
1765 removed by invalidate_memory so specific invalidation is superfluous).
1766 FULL_MODE, if not VOIDmode, indicates that this much should be
1767 invalidated instead of just the amount indicated by the mode of X. This
1768 is only used for bitfield stores into memory.
1770 A nonvarying address may be just a register or just a symbol reference,
1771 or it may be either of those plus a numeric offset. */
1773 static void
1774 invalidate (rtx x, enum machine_mode full_mode)
1776 int i;
1777 struct table_elt *p;
1778 rtx addr;
1780 switch (GET_CODE (x))
1782 case REG:
1784 /* If X is a register, dependencies on its contents are recorded
1785 through the qty number mechanism. Just change the qty number of
1786 the register, mark it as invalid for expressions that refer to it,
1787 and remove it itself. */
1788 unsigned int regno = REGNO (x);
1789 unsigned int hash = HASH (x, GET_MODE (x));
1791 /* Remove REGNO from any quantity list it might be on and indicate
1792 that its value might have changed. If it is a pseudo, remove its
1793 entry from the hash table.
1795 For a hard register, we do the first two actions above for any
1796 additional hard registers corresponding to X. Then, if any of these
1797 registers are in the table, we must remove any REG entries that
1798 overlap these registers. */
1800 delete_reg_equiv (regno);
1801 REG_TICK (regno)++;
1802 SUBREG_TICKED (regno) = -1;
1804 if (regno >= FIRST_PSEUDO_REGISTER)
1806 /* Because a register can be referenced in more than one mode,
1807 we might have to remove more than one table entry. */
1808 struct table_elt *elt;
1810 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1811 remove_from_table (elt, hash);
1813 else
1815 HOST_WIDE_INT in_table
1816 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1817 unsigned int endregno
1818 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1819 unsigned int tregno, tendregno, rn;
1820 struct table_elt *p, *next;
1822 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1824 for (rn = regno + 1; rn < endregno; rn++)
1826 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1827 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1828 delete_reg_equiv (rn);
1829 REG_TICK (rn)++;
1830 SUBREG_TICKED (rn) = -1;
1833 if (in_table)
1834 for (hash = 0; hash < HASH_SIZE; hash++)
1835 for (p = table[hash]; p; p = next)
1837 next = p->next_same_hash;
1839 if (!REG_P (p->exp)
1840 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1841 continue;
1843 tregno = REGNO (p->exp);
1844 tendregno
1845 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1846 if (tendregno > regno && tregno < endregno)
1847 remove_from_table (p, hash);
1851 return;
1853 case SUBREG:
1854 invalidate (SUBREG_REG (x), VOIDmode);
1855 return;
1857 case PARALLEL:
1858 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1859 invalidate (XVECEXP (x, 0, i), VOIDmode);
1860 return;
1862 case EXPR_LIST:
1863 /* This is part of a disjoint return value; extract the location in
1864 question ignoring the offset. */
1865 invalidate (XEXP (x, 0), VOIDmode);
1866 return;
1868 case MEM:
1869 addr = canon_rtx (get_addr (XEXP (x, 0)));
1870 /* Calculate the canonical version of X here so that
1871 true_dependence doesn't generate new RTL for X on each call. */
1872 x = canon_rtx (x);
1874 /* Remove all hash table elements that refer to overlapping pieces of
1875 memory. */
1876 if (full_mode == VOIDmode)
1877 full_mode = GET_MODE (x);
1879 for (i = 0; i < HASH_SIZE; i++)
1881 struct table_elt *next;
1883 for (p = table[i]; p; p = next)
1885 next = p->next_same_hash;
1886 if (p->in_memory)
1888 struct check_dependence_data d;
1890 /* Just canonicalize the expression once;
1891 otherwise each time we call invalidate
1892 true_dependence will canonicalize the
1893 expression again. */
1894 if (!p->canon_exp)
1895 p->canon_exp = canon_rtx (p->exp);
1896 d.exp = x;
1897 d.addr = addr;
1898 d.mode = full_mode;
1899 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1900 remove_from_table (p, i);
1904 return;
1906 default:
1907 gcc_unreachable ();
1911 /* Remove all expressions that refer to register REGNO,
1912 since they are already invalid, and we are about to
1913 mark that register valid again and don't want the old
1914 expressions to reappear as valid. */
1916 static void
1917 remove_invalid_refs (unsigned int regno)
1919 unsigned int i;
1920 struct table_elt *p, *next;
1922 for (i = 0; i < HASH_SIZE; i++)
1923 for (p = table[i]; p; p = next)
1925 next = p->next_same_hash;
1926 if (!REG_P (p->exp)
1927 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1928 remove_from_table (p, i);
1932 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1933 and mode MODE. */
1934 static void
1935 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1936 enum machine_mode mode)
1938 unsigned int i;
1939 struct table_elt *p, *next;
1940 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1942 for (i = 0; i < HASH_SIZE; i++)
1943 for (p = table[i]; p; p = next)
1945 rtx exp = p->exp;
1946 next = p->next_same_hash;
1948 if (!REG_P (exp)
1949 && (GET_CODE (exp) != SUBREG
1950 || !REG_P (SUBREG_REG (exp))
1951 || REGNO (SUBREG_REG (exp)) != regno
1952 || (((SUBREG_BYTE (exp)
1953 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1954 && SUBREG_BYTE (exp) <= end))
1955 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1956 remove_from_table (p, i);
1960 /* Recompute the hash codes of any valid entries in the hash table that
1961 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1963 This is called when we make a jump equivalence. */
1965 static void
1966 rehash_using_reg (rtx x)
1968 unsigned int i;
1969 struct table_elt *p, *next;
1970 unsigned hash;
1972 if (GET_CODE (x) == SUBREG)
1973 x = SUBREG_REG (x);
1975 /* If X is not a register or if the register is known not to be in any
1976 valid entries in the table, we have no work to do. */
1978 if (!REG_P (x)
1979 || REG_IN_TABLE (REGNO (x)) < 0
1980 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1981 return;
1983 /* Scan all hash chains looking for valid entries that mention X.
1984 If we find one and it is in the wrong hash chain, move it. */
1986 for (i = 0; i < HASH_SIZE; i++)
1987 for (p = table[i]; p; p = next)
1989 next = p->next_same_hash;
1990 if (reg_mentioned_p (x, p->exp)
1991 && exp_equiv_p (p->exp, p->exp, 1, false)
1992 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1994 if (p->next_same_hash)
1995 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1997 if (p->prev_same_hash)
1998 p->prev_same_hash->next_same_hash = p->next_same_hash;
1999 else
2000 table[i] = p->next_same_hash;
2002 p->next_same_hash = table[hash];
2003 p->prev_same_hash = 0;
2004 if (table[hash])
2005 table[hash]->prev_same_hash = p;
2006 table[hash] = p;
2011 /* Remove from the hash table any expression that is a call-clobbered
2012 register. Also update their TICK values. */
2014 static void
2015 invalidate_for_call (void)
2017 unsigned int regno, endregno;
2018 unsigned int i;
2019 unsigned hash;
2020 struct table_elt *p, *next;
2021 int in_table = 0;
2023 /* Go through all the hard registers. For each that is clobbered in
2024 a CALL_INSN, remove the register from quantity chains and update
2025 reg_tick if defined. Also see if any of these registers is currently
2026 in the table. */
2028 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2029 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2031 delete_reg_equiv (regno);
2032 if (REG_TICK (regno) >= 0)
2034 REG_TICK (regno)++;
2035 SUBREG_TICKED (regno) = -1;
2038 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2041 /* In the case where we have no call-clobbered hard registers in the
2042 table, we are done. Otherwise, scan the table and remove any
2043 entry that overlaps a call-clobbered register. */
2045 if (in_table)
2046 for (hash = 0; hash < HASH_SIZE; hash++)
2047 for (p = table[hash]; p; p = next)
2049 next = p->next_same_hash;
2051 if (!REG_P (p->exp)
2052 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2053 continue;
2055 regno = REGNO (p->exp);
2056 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
2058 for (i = regno; i < endregno; i++)
2059 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2061 remove_from_table (p, hash);
2062 break;
2067 /* Given an expression X of type CONST,
2068 and ELT which is its table entry (or 0 if it
2069 is not in the hash table),
2070 return an alternate expression for X as a register plus integer.
2071 If none can be found, return 0. */
2073 static rtx
2074 use_related_value (rtx x, struct table_elt *elt)
2076 struct table_elt *relt = 0;
2077 struct table_elt *p, *q;
2078 HOST_WIDE_INT offset;
2080 /* First, is there anything related known?
2081 If we have a table element, we can tell from that.
2082 Otherwise, must look it up. */
2084 if (elt != 0 && elt->related_value != 0)
2085 relt = elt;
2086 else if (elt == 0 && GET_CODE (x) == CONST)
2088 rtx subexp = get_related_value (x);
2089 if (subexp != 0)
2090 relt = lookup (subexp,
2091 SAFE_HASH (subexp, GET_MODE (subexp)),
2092 GET_MODE (subexp));
2095 if (relt == 0)
2096 return 0;
2098 /* Search all related table entries for one that has an
2099 equivalent register. */
2101 p = relt;
2102 while (1)
2104 /* This loop is strange in that it is executed in two different cases.
2105 The first is when X is already in the table. Then it is searching
2106 the RELATED_VALUE list of X's class (RELT). The second case is when
2107 X is not in the table. Then RELT points to a class for the related
2108 value.
2110 Ensure that, whatever case we are in, that we ignore classes that have
2111 the same value as X. */
2113 if (rtx_equal_p (x, p->exp))
2114 q = 0;
2115 else
2116 for (q = p->first_same_value; q; q = q->next_same_value)
2117 if (REG_P (q->exp))
2118 break;
2120 if (q)
2121 break;
2123 p = p->related_value;
2125 /* We went all the way around, so there is nothing to be found.
2126 Alternatively, perhaps RELT was in the table for some other reason
2127 and it has no related values recorded. */
2128 if (p == relt || p == 0)
2129 break;
2132 if (q == 0)
2133 return 0;
2135 offset = (get_integer_term (x) - get_integer_term (p->exp));
2136 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2137 return plus_constant (q->exp, offset);
2140 /* Hash a string. Just add its bytes up. */
2141 static inline unsigned
2142 hash_rtx_string (const char *ps)
2144 unsigned hash = 0;
2145 const unsigned char *p = (const unsigned char *) ps;
2147 if (p)
2148 while (*p)
2149 hash += *p++;
2151 return hash;
2154 /* Hash an rtx. We are careful to make sure the value is never negative.
2155 Equivalent registers hash identically.
2156 MODE is used in hashing for CONST_INTs only;
2157 otherwise the mode of X is used.
2159 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2161 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2162 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2164 Note that cse_insn knows that the hash code of a MEM expression
2165 is just (int) MEM plus the hash code of the address. */
2167 unsigned
2168 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2169 int *hash_arg_in_memory_p, bool have_reg_qty)
2171 int i, j;
2172 unsigned hash = 0;
2173 enum rtx_code code;
2174 const char *fmt;
2176 /* Used to turn recursion into iteration. We can't rely on GCC's
2177 tail-recursion elimination since we need to keep accumulating values
2178 in HASH. */
2179 repeat:
2180 if (x == 0)
2181 return hash;
2183 code = GET_CODE (x);
2184 switch (code)
2186 case REG:
2188 unsigned int regno = REGNO (x);
2190 if (!reload_completed)
2192 /* On some machines, we can't record any non-fixed hard register,
2193 because extending its life will cause reload problems. We
2194 consider ap, fp, sp, gp to be fixed for this purpose.
2196 We also consider CCmode registers to be fixed for this purpose;
2197 failure to do so leads to failure to simplify 0<100 type of
2198 conditionals.
2200 On all machines, we can't record any global registers.
2201 Nor should we record any register that is in a small
2202 class, as defined by CLASS_LIKELY_SPILLED_P. */
2203 bool record;
2205 if (regno >= FIRST_PSEUDO_REGISTER)
2206 record = true;
2207 else if (x == frame_pointer_rtx
2208 || x == hard_frame_pointer_rtx
2209 || x == arg_pointer_rtx
2210 || x == stack_pointer_rtx
2211 || x == pic_offset_table_rtx)
2212 record = true;
2213 else if (global_regs[regno])
2214 record = false;
2215 else if (fixed_regs[regno])
2216 record = true;
2217 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2218 record = true;
2219 else if (SMALL_REGISTER_CLASSES)
2220 record = false;
2221 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2222 record = false;
2223 else
2224 record = true;
2226 if (!record)
2228 *do_not_record_p = 1;
2229 return 0;
2233 hash += ((unsigned int) REG << 7);
2234 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2235 return hash;
2238 /* We handle SUBREG of a REG specially because the underlying
2239 reg changes its hash value with every value change; we don't
2240 want to have to forget unrelated subregs when one subreg changes. */
2241 case SUBREG:
2243 if (REG_P (SUBREG_REG (x)))
2245 hash += (((unsigned int) SUBREG << 7)
2246 + REGNO (SUBREG_REG (x))
2247 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2248 return hash;
2250 break;
2253 case CONST_INT:
2254 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2255 + (unsigned int) INTVAL (x));
2256 return hash;
2258 case CONST_DOUBLE:
2259 /* This is like the general case, except that it only counts
2260 the integers representing the constant. */
2261 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2262 if (GET_MODE (x) != VOIDmode)
2263 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2264 else
2265 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2266 + (unsigned int) CONST_DOUBLE_HIGH (x));
2267 return hash;
2269 case CONST_VECTOR:
2271 int units;
2272 rtx elt;
2274 units = CONST_VECTOR_NUNITS (x);
2276 for (i = 0; i < units; ++i)
2278 elt = CONST_VECTOR_ELT (x, i);
2279 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2280 hash_arg_in_memory_p, have_reg_qty);
2283 return hash;
2286 /* Assume there is only one rtx object for any given label. */
2287 case LABEL_REF:
2288 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2289 differences and differences between each stage's debugging dumps. */
2290 hash += (((unsigned int) LABEL_REF << 7)
2291 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2292 return hash;
2294 case SYMBOL_REF:
2296 /* Don't hash on the symbol's address to avoid bootstrap differences.
2297 Different hash values may cause expressions to be recorded in
2298 different orders and thus different registers to be used in the
2299 final assembler. This also avoids differences in the dump files
2300 between various stages. */
2301 unsigned int h = 0;
2302 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2304 while (*p)
2305 h += (h << 7) + *p++; /* ??? revisit */
2307 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2308 return hash;
2311 case MEM:
2312 /* We don't record if marked volatile or if BLKmode since we don't
2313 know the size of the move. */
2314 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2316 *do_not_record_p = 1;
2317 return 0;
2319 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2320 *hash_arg_in_memory_p = 1;
2322 /* Now that we have already found this special case,
2323 might as well speed it up as much as possible. */
2324 hash += (unsigned) MEM;
2325 x = XEXP (x, 0);
2326 goto repeat;
2328 case USE:
2329 /* A USE that mentions non-volatile memory needs special
2330 handling since the MEM may be BLKmode which normally
2331 prevents an entry from being made. Pure calls are
2332 marked by a USE which mentions BLKmode memory.
2333 See calls.c:emit_call_1. */
2334 if (MEM_P (XEXP (x, 0))
2335 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2337 hash += (unsigned) USE;
2338 x = XEXP (x, 0);
2340 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2341 *hash_arg_in_memory_p = 1;
2343 /* Now that we have already found this special case,
2344 might as well speed it up as much as possible. */
2345 hash += (unsigned) MEM;
2346 x = XEXP (x, 0);
2347 goto repeat;
2349 break;
2351 case PRE_DEC:
2352 case PRE_INC:
2353 case POST_DEC:
2354 case POST_INC:
2355 case PRE_MODIFY:
2356 case POST_MODIFY:
2357 case PC:
2358 case CC0:
2359 case CALL:
2360 case UNSPEC_VOLATILE:
2361 *do_not_record_p = 1;
2362 return 0;
2364 case ASM_OPERANDS:
2365 if (MEM_VOLATILE_P (x))
2367 *do_not_record_p = 1;
2368 return 0;
2370 else
2372 /* We don't want to take the filename and line into account. */
2373 hash += (unsigned) code + (unsigned) GET_MODE (x)
2374 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2375 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2376 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2378 if (ASM_OPERANDS_INPUT_LENGTH (x))
2380 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2382 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2383 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2384 do_not_record_p, hash_arg_in_memory_p,
2385 have_reg_qty)
2386 + hash_rtx_string
2387 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2390 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2391 x = ASM_OPERANDS_INPUT (x, 0);
2392 mode = GET_MODE (x);
2393 goto repeat;
2396 return hash;
2398 break;
2400 default:
2401 break;
2404 i = GET_RTX_LENGTH (code) - 1;
2405 hash += (unsigned) code + (unsigned) GET_MODE (x);
2406 fmt = GET_RTX_FORMAT (code);
2407 for (; i >= 0; i--)
2409 switch (fmt[i])
2411 case 'e':
2412 /* If we are about to do the last recursive call
2413 needed at this level, change it into iteration.
2414 This function is called enough to be worth it. */
2415 if (i == 0)
2417 x = XEXP (x, i);
2418 goto repeat;
2421 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2422 hash_arg_in_memory_p, have_reg_qty);
2423 break;
2425 case 'E':
2426 for (j = 0; j < XVECLEN (x, i); j++)
2427 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2428 hash_arg_in_memory_p, have_reg_qty);
2429 break;
2431 case 's':
2432 hash += hash_rtx_string (XSTR (x, i));
2433 break;
2435 case 'i':
2436 hash += (unsigned int) XINT (x, i);
2437 break;
2439 case '0': case 't':
2440 /* Unused. */
2441 break;
2443 default:
2444 gcc_unreachable ();
2448 return hash;
2451 /* Hash an rtx X for cse via hash_rtx.
2452 Stores 1 in do_not_record if any subexpression is volatile.
2453 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2454 does not have the RTX_UNCHANGING_P bit set. */
2456 static inline unsigned
2457 canon_hash (rtx x, enum machine_mode mode)
2459 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2462 /* Like canon_hash but with no side effects, i.e. do_not_record
2463 and hash_arg_in_memory are not changed. */
2465 static inline unsigned
2466 safe_hash (rtx x, enum machine_mode mode)
2468 int dummy_do_not_record;
2469 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2472 /* Return 1 iff X and Y would canonicalize into the same thing,
2473 without actually constructing the canonicalization of either one.
2474 If VALIDATE is nonzero,
2475 we assume X is an expression being processed from the rtl
2476 and Y was found in the hash table. We check register refs
2477 in Y for being marked as valid.
2479 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2482 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2484 int i, j;
2485 enum rtx_code code;
2486 const char *fmt;
2488 /* Note: it is incorrect to assume an expression is equivalent to itself
2489 if VALIDATE is nonzero. */
2490 if (x == y && !validate)
2491 return 1;
2493 if (x == 0 || y == 0)
2494 return x == y;
2496 code = GET_CODE (x);
2497 if (code != GET_CODE (y))
2498 return 0;
2500 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2501 if (GET_MODE (x) != GET_MODE (y))
2502 return 0;
2504 switch (code)
2506 case PC:
2507 case CC0:
2508 case CONST_INT:
2509 case CONST_DOUBLE:
2510 return x == y;
2512 case LABEL_REF:
2513 return XEXP (x, 0) == XEXP (y, 0);
2515 case SYMBOL_REF:
2516 return XSTR (x, 0) == XSTR (y, 0);
2518 case REG:
2519 if (for_gcse)
2520 return REGNO (x) == REGNO (y);
2521 else
2523 unsigned int regno = REGNO (y);
2524 unsigned int i;
2525 unsigned int endregno
2526 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2527 : hard_regno_nregs[regno][GET_MODE (y)]);
2529 /* If the quantities are not the same, the expressions are not
2530 equivalent. If there are and we are not to validate, they
2531 are equivalent. Otherwise, ensure all regs are up-to-date. */
2533 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2534 return 0;
2536 if (! validate)
2537 return 1;
2539 for (i = regno; i < endregno; i++)
2540 if (REG_IN_TABLE (i) != REG_TICK (i))
2541 return 0;
2543 return 1;
2546 case MEM:
2547 if (for_gcse)
2549 /* A volatile mem should not be considered equivalent to any
2550 other. */
2551 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2552 return 0;
2554 /* Can't merge two expressions in different alias sets, since we
2555 can decide that the expression is transparent in a block when
2556 it isn't, due to it being set with the different alias set.
2558 Also, can't merge two expressions with different MEM_ATTRS.
2559 They could e.g. be two different entities allocated into the
2560 same space on the stack (see e.g. PR25130). In that case, the
2561 MEM addresses can be the same, even though the two MEMs are
2562 absolutely not equivalent.
2564 But because really all MEM attributes should be the same for
2565 equivalent MEMs, we just use the invariant that MEMs that have
2566 the same attributes share the same mem_attrs data structure. */
2567 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2568 return 0;
2570 break;
2572 /* For commutative operations, check both orders. */
2573 case PLUS:
2574 case MULT:
2575 case AND:
2576 case IOR:
2577 case XOR:
2578 case NE:
2579 case EQ:
2580 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2581 validate, for_gcse)
2582 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2583 validate, for_gcse))
2584 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2585 validate, for_gcse)
2586 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2587 validate, for_gcse)));
2589 case ASM_OPERANDS:
2590 /* We don't use the generic code below because we want to
2591 disregard filename and line numbers. */
2593 /* A volatile asm isn't equivalent to any other. */
2594 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2595 return 0;
2597 if (GET_MODE (x) != GET_MODE (y)
2598 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2599 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2600 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2601 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2602 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2603 return 0;
2605 if (ASM_OPERANDS_INPUT_LENGTH (x))
2607 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2608 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2609 ASM_OPERANDS_INPUT (y, i),
2610 validate, for_gcse)
2611 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2612 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2613 return 0;
2616 return 1;
2618 default:
2619 break;
2622 /* Compare the elements. If any pair of corresponding elements
2623 fail to match, return 0 for the whole thing. */
2625 fmt = GET_RTX_FORMAT (code);
2626 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2628 switch (fmt[i])
2630 case 'e':
2631 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2632 validate, for_gcse))
2633 return 0;
2634 break;
2636 case 'E':
2637 if (XVECLEN (x, i) != XVECLEN (y, i))
2638 return 0;
2639 for (j = 0; j < XVECLEN (x, i); j++)
2640 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2641 validate, for_gcse))
2642 return 0;
2643 break;
2645 case 's':
2646 if (strcmp (XSTR (x, i), XSTR (y, i)))
2647 return 0;
2648 break;
2650 case 'i':
2651 if (XINT (x, i) != XINT (y, i))
2652 return 0;
2653 break;
2655 case 'w':
2656 if (XWINT (x, i) != XWINT (y, i))
2657 return 0;
2658 break;
2660 case '0':
2661 case 't':
2662 break;
2664 default:
2665 gcc_unreachable ();
2669 return 1;
2672 /* Return 1 if X has a value that can vary even between two
2673 executions of the program. 0 means X can be compared reliably
2674 against certain constants or near-constants. */
2676 static int
2677 cse_rtx_varies_p (rtx x, int from_alias)
2679 /* We need not check for X and the equivalence class being of the same
2680 mode because if X is equivalent to a constant in some mode, it
2681 doesn't vary in any mode. */
2683 if (REG_P (x)
2684 && REGNO_QTY_VALID_P (REGNO (x)))
2686 int x_q = REG_QTY (REGNO (x));
2687 struct qty_table_elem *x_ent = &qty_table[x_q];
2689 if (GET_MODE (x) == x_ent->mode
2690 && x_ent->const_rtx != NULL_RTX)
2691 return 0;
2694 if (GET_CODE (x) == PLUS
2695 && GET_CODE (XEXP (x, 1)) == CONST_INT
2696 && REG_P (XEXP (x, 0))
2697 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2699 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2700 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2702 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2703 && x0_ent->const_rtx != NULL_RTX)
2704 return 0;
2707 /* This can happen as the result of virtual register instantiation, if
2708 the initial constant is too large to be a valid address. This gives
2709 us a three instruction sequence, load large offset into a register,
2710 load fp minus a constant into a register, then a MEM which is the
2711 sum of the two `constant' registers. */
2712 if (GET_CODE (x) == PLUS
2713 && REG_P (XEXP (x, 0))
2714 && REG_P (XEXP (x, 1))
2715 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2716 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2718 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2719 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2720 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2721 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2723 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2724 && x0_ent->const_rtx != NULL_RTX
2725 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2726 && x1_ent->const_rtx != NULL_RTX)
2727 return 0;
2730 return rtx_varies_p (x, from_alias);
2733 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2734 the result if necessary. INSN is as for canon_reg. */
2736 static void
2737 validate_canon_reg (rtx *xloc, rtx insn)
2739 rtx new = canon_reg (*xloc, insn);
2741 /* If replacing pseudo with hard reg or vice versa, ensure the
2742 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2743 if (insn != 0 && new != 0)
2744 validate_change (insn, xloc, new, 1);
2745 else
2746 *xloc = new;
2749 /* Canonicalize an expression:
2750 replace each register reference inside it
2751 with the "oldest" equivalent register.
2753 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2754 after we make our substitution. The calls are made with IN_GROUP nonzero
2755 so apply_change_group must be called upon the outermost return from this
2756 function (unless INSN is zero). The result of apply_change_group can
2757 generally be discarded since the changes we are making are optional. */
2759 static rtx
2760 canon_reg (rtx x, rtx insn)
2762 int i;
2763 enum rtx_code code;
2764 const char *fmt;
2766 if (x == 0)
2767 return x;
2769 code = GET_CODE (x);
2770 switch (code)
2772 case PC:
2773 case CC0:
2774 case CONST:
2775 case CONST_INT:
2776 case CONST_DOUBLE:
2777 case CONST_VECTOR:
2778 case SYMBOL_REF:
2779 case LABEL_REF:
2780 case ADDR_VEC:
2781 case ADDR_DIFF_VEC:
2782 return x;
2784 case REG:
2786 int first;
2787 int q;
2788 struct qty_table_elem *ent;
2790 /* Never replace a hard reg, because hard regs can appear
2791 in more than one machine mode, and we must preserve the mode
2792 of each occurrence. Also, some hard regs appear in
2793 MEMs that are shared and mustn't be altered. Don't try to
2794 replace any reg that maps to a reg of class NO_REGS. */
2795 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2796 || ! REGNO_QTY_VALID_P (REGNO (x)))
2797 return x;
2799 q = REG_QTY (REGNO (x));
2800 ent = &qty_table[q];
2801 first = ent->first_reg;
2802 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2803 : REGNO_REG_CLASS (first) == NO_REGS ? x
2804 : gen_rtx_REG (ent->mode, first));
2807 default:
2808 break;
2811 fmt = GET_RTX_FORMAT (code);
2812 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2814 int j;
2816 if (fmt[i] == 'e')
2817 validate_canon_reg (&XEXP (x, i), insn);
2818 else if (fmt[i] == 'E')
2819 for (j = 0; j < XVECLEN (x, i); j++)
2820 validate_canon_reg (&XVECEXP (x, i, j), insn);
2823 return x;
2826 /* LOC is a location within INSN that is an operand address (the contents of
2827 a MEM). Find the best equivalent address to use that is valid for this
2828 insn.
2830 On most CISC machines, complicated address modes are costly, and rtx_cost
2831 is a good approximation for that cost. However, most RISC machines have
2832 only a few (usually only one) memory reference formats. If an address is
2833 valid at all, it is often just as cheap as any other address. Hence, for
2834 RISC machines, we use `address_cost' to compare the costs of various
2835 addresses. For two addresses of equal cost, choose the one with the
2836 highest `rtx_cost' value as that has the potential of eliminating the
2837 most insns. For equal costs, we choose the first in the equivalence
2838 class. Note that we ignore the fact that pseudo registers are cheaper than
2839 hard registers here because we would also prefer the pseudo registers. */
2841 static void
2842 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2844 struct table_elt *elt;
2845 rtx addr = *loc;
2846 struct table_elt *p;
2847 int found_better = 1;
2848 int save_do_not_record = do_not_record;
2849 int save_hash_arg_in_memory = hash_arg_in_memory;
2850 int addr_volatile;
2851 int regno;
2852 unsigned hash;
2854 /* Do not try to replace constant addresses or addresses of local and
2855 argument slots. These MEM expressions are made only once and inserted
2856 in many instructions, as well as being used to control symbol table
2857 output. It is not safe to clobber them.
2859 There are some uncommon cases where the address is already in a register
2860 for some reason, but we cannot take advantage of that because we have
2861 no easy way to unshare the MEM. In addition, looking up all stack
2862 addresses is costly. */
2863 if ((GET_CODE (addr) == PLUS
2864 && REG_P (XEXP (addr, 0))
2865 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2866 && (regno = REGNO (XEXP (addr, 0)),
2867 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2868 || regno == ARG_POINTER_REGNUM))
2869 || (REG_P (addr)
2870 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2871 || regno == HARD_FRAME_POINTER_REGNUM
2872 || regno == ARG_POINTER_REGNUM))
2873 || CONSTANT_ADDRESS_P (addr))
2874 return;
2876 /* If this address is not simply a register, try to fold it. This will
2877 sometimes simplify the expression. Many simplifications
2878 will not be valid, but some, usually applying the associative rule, will
2879 be valid and produce better code. */
2880 if (!REG_P (addr))
2882 rtx folded = canon_for_address (fold_rtx (addr, NULL_RTX));
2884 if (folded != addr)
2886 int addr_folded_cost = address_cost (folded, mode);
2887 int addr_cost = address_cost (addr, mode);
2889 if ((addr_folded_cost < addr_cost
2890 || (addr_folded_cost == addr_cost
2891 /* ??? The rtx_cost comparison is left over from an older
2892 version of this code. It is probably no longer helpful.*/
2893 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2894 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2895 && validate_change (insn, loc, folded, 0))
2896 addr = folded;
2900 /* If this address is not in the hash table, we can't look for equivalences
2901 of the whole address. Also, ignore if volatile. */
2903 do_not_record = 0;
2904 hash = HASH (addr, Pmode);
2905 addr_volatile = do_not_record;
2906 do_not_record = save_do_not_record;
2907 hash_arg_in_memory = save_hash_arg_in_memory;
2909 if (addr_volatile)
2910 return;
2912 elt = lookup (addr, hash, Pmode);
2914 if (elt)
2916 /* We need to find the best (under the criteria documented above) entry
2917 in the class that is valid. We use the `flag' field to indicate
2918 choices that were invalid and iterate until we can't find a better
2919 one that hasn't already been tried. */
2921 for (p = elt->first_same_value; p; p = p->next_same_value)
2922 p->flag = 0;
2924 while (found_better)
2926 int best_addr_cost = address_cost (*loc, mode);
2927 int best_rtx_cost = (elt->cost + 1) >> 1;
2928 int exp_cost;
2929 struct table_elt *best_elt = elt;
2931 found_better = 0;
2932 for (p = elt->first_same_value; p; p = p->next_same_value)
2933 if (! p->flag)
2935 if ((REG_P (p->exp)
2936 || exp_equiv_p (p->exp, p->exp, 1, false))
2937 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2938 || (exp_cost == best_addr_cost
2939 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2941 found_better = 1;
2942 best_addr_cost = exp_cost;
2943 best_rtx_cost = (p->cost + 1) >> 1;
2944 best_elt = p;
2948 if (found_better)
2950 if (validate_change (insn, loc,
2951 canon_reg (copy_rtx (best_elt->exp),
2952 NULL_RTX), 0))
2953 return;
2954 else
2955 best_elt->flag = 1;
2960 /* If the address is a binary operation with the first operand a register
2961 and the second a constant, do the same as above, but looking for
2962 equivalences of the register. Then try to simplify before checking for
2963 the best address to use. This catches a few cases: First is when we
2964 have REG+const and the register is another REG+const. We can often merge
2965 the constants and eliminate one insn and one register. It may also be
2966 that a machine has a cheap REG+REG+const. Finally, this improves the
2967 code on the Alpha for unaligned byte stores. */
2969 if (flag_expensive_optimizations
2970 && ARITHMETIC_P (*loc)
2971 && REG_P (XEXP (*loc, 0)))
2973 rtx op1 = XEXP (*loc, 1);
2975 do_not_record = 0;
2976 hash = HASH (XEXP (*loc, 0), Pmode);
2977 do_not_record = save_do_not_record;
2978 hash_arg_in_memory = save_hash_arg_in_memory;
2980 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2981 if (elt == 0)
2982 return;
2984 /* We need to find the best (under the criteria documented above) entry
2985 in the class that is valid. We use the `flag' field to indicate
2986 choices that were invalid and iterate until we can't find a better
2987 one that hasn't already been tried. */
2989 for (p = elt->first_same_value; p; p = p->next_same_value)
2990 p->flag = 0;
2992 while (found_better)
2994 int best_addr_cost = address_cost (*loc, mode);
2995 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2996 struct table_elt *best_elt = elt;
2997 rtx best_rtx = *loc;
2998 int count;
3000 /* This is at worst case an O(n^2) algorithm, so limit our search
3001 to the first 32 elements on the list. This avoids trouble
3002 compiling code with very long basic blocks that can easily
3003 call simplify_gen_binary so many times that we run out of
3004 memory. */
3006 found_better = 0;
3007 for (p = elt->first_same_value, count = 0;
3008 p && count < 32;
3009 p = p->next_same_value, count++)
3010 if (! p->flag
3011 && (REG_P (p->exp)
3012 || (GET_CODE (p->exp) != EXPR_LIST
3013 && exp_equiv_p (p->exp, p->exp, 1, false))))
3016 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3017 p->exp, op1);
3018 int new_cost;
3020 /* Get the canonical version of the address so we can accept
3021 more. */
3022 new = canon_for_address (new);
3024 new_cost = address_cost (new, mode);
3026 if (new_cost < best_addr_cost
3027 || (new_cost == best_addr_cost
3028 && (COST (new) + 1) >> 1 > best_rtx_cost))
3030 found_better = 1;
3031 best_addr_cost = new_cost;
3032 best_rtx_cost = (COST (new) + 1) >> 1;
3033 best_elt = p;
3034 best_rtx = new;
3038 if (found_better)
3040 if (validate_change (insn, loc,
3041 canon_reg (copy_rtx (best_rtx),
3042 NULL_RTX), 0))
3043 return;
3044 else
3045 best_elt->flag = 1;
3051 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3052 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3053 what values are being compared.
3055 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3056 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3057 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3058 compared to produce cc0.
3060 The return value is the comparison operator and is either the code of
3061 A or the code corresponding to the inverse of the comparison. */
3063 static enum rtx_code
3064 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3065 enum machine_mode *pmode1, enum machine_mode *pmode2)
3067 rtx arg1, arg2;
3069 arg1 = *parg1, arg2 = *parg2;
3071 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3073 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3075 /* Set nonzero when we find something of interest. */
3076 rtx x = 0;
3077 int reverse_code = 0;
3078 struct table_elt *p = 0;
3080 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3081 On machines with CC0, this is the only case that can occur, since
3082 fold_rtx will return the COMPARE or item being compared with zero
3083 when given CC0. */
3085 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3086 x = arg1;
3088 /* If ARG1 is a comparison operator and CODE is testing for
3089 STORE_FLAG_VALUE, get the inner arguments. */
3091 else if (COMPARISON_P (arg1))
3093 #ifdef FLOAT_STORE_FLAG_VALUE
3094 REAL_VALUE_TYPE fsfv;
3095 #endif
3097 if (code == NE
3098 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3099 && code == LT && STORE_FLAG_VALUE == -1)
3100 #ifdef FLOAT_STORE_FLAG_VALUE
3101 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3102 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3103 REAL_VALUE_NEGATIVE (fsfv)))
3104 #endif
3106 x = arg1;
3107 else if (code == EQ
3108 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3109 && code == GE && STORE_FLAG_VALUE == -1)
3110 #ifdef FLOAT_STORE_FLAG_VALUE
3111 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3112 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3113 REAL_VALUE_NEGATIVE (fsfv)))
3114 #endif
3116 x = arg1, reverse_code = 1;
3119 /* ??? We could also check for
3121 (ne (and (eq (...) (const_int 1))) (const_int 0))
3123 and related forms, but let's wait until we see them occurring. */
3125 if (x == 0)
3126 /* Look up ARG1 in the hash table and see if it has an equivalence
3127 that lets us see what is being compared. */
3128 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3129 if (p)
3131 p = p->first_same_value;
3133 /* If what we compare is already known to be constant, that is as
3134 good as it gets.
3135 We need to break the loop in this case, because otherwise we
3136 can have an infinite loop when looking at a reg that is known
3137 to be a constant which is the same as a comparison of a reg
3138 against zero which appears later in the insn stream, which in
3139 turn is constant and the same as the comparison of the first reg
3140 against zero... */
3141 if (p->is_const)
3142 break;
3145 for (; p; p = p->next_same_value)
3147 enum machine_mode inner_mode = GET_MODE (p->exp);
3148 #ifdef FLOAT_STORE_FLAG_VALUE
3149 REAL_VALUE_TYPE fsfv;
3150 #endif
3152 /* If the entry isn't valid, skip it. */
3153 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3154 continue;
3156 if (GET_CODE (p->exp) == COMPARE
3157 /* Another possibility is that this machine has a compare insn
3158 that includes the comparison code. In that case, ARG1 would
3159 be equivalent to a comparison operation that would set ARG1 to
3160 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3161 ORIG_CODE is the actual comparison being done; if it is an EQ,
3162 we must reverse ORIG_CODE. On machine with a negative value
3163 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3164 || ((code == NE
3165 || (code == LT
3166 && GET_MODE_CLASS (inner_mode) == MODE_INT
3167 && (GET_MODE_BITSIZE (inner_mode)
3168 <= HOST_BITS_PER_WIDE_INT)
3169 && (STORE_FLAG_VALUE
3170 & ((HOST_WIDE_INT) 1
3171 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3172 #ifdef FLOAT_STORE_FLAG_VALUE
3173 || (code == LT
3174 && SCALAR_FLOAT_MODE_P (inner_mode)
3175 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3176 REAL_VALUE_NEGATIVE (fsfv)))
3177 #endif
3179 && COMPARISON_P (p->exp)))
3181 x = p->exp;
3182 break;
3184 else if ((code == EQ
3185 || (code == GE
3186 && GET_MODE_CLASS (inner_mode) == MODE_INT
3187 && (GET_MODE_BITSIZE (inner_mode)
3188 <= HOST_BITS_PER_WIDE_INT)
3189 && (STORE_FLAG_VALUE
3190 & ((HOST_WIDE_INT) 1
3191 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3192 #ifdef FLOAT_STORE_FLAG_VALUE
3193 || (code == GE
3194 && SCALAR_FLOAT_MODE_P (inner_mode)
3195 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3196 REAL_VALUE_NEGATIVE (fsfv)))
3197 #endif
3199 && COMPARISON_P (p->exp))
3201 reverse_code = 1;
3202 x = p->exp;
3203 break;
3206 /* If this non-trapping address, e.g. fp + constant, the
3207 equivalent is a better operand since it may let us predict
3208 the value of the comparison. */
3209 else if (!rtx_addr_can_trap_p (p->exp))
3211 arg1 = p->exp;
3212 continue;
3216 /* If we didn't find a useful equivalence for ARG1, we are done.
3217 Otherwise, set up for the next iteration. */
3218 if (x == 0)
3219 break;
3221 /* If we need to reverse the comparison, make sure that that is
3222 possible -- we can't necessarily infer the value of GE from LT
3223 with floating-point operands. */
3224 if (reverse_code)
3226 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3227 if (reversed == UNKNOWN)
3228 break;
3229 else
3230 code = reversed;
3232 else if (COMPARISON_P (x))
3233 code = GET_CODE (x);
3234 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3237 /* Return our results. Return the modes from before fold_rtx
3238 because fold_rtx might produce const_int, and then it's too late. */
3239 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3240 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3242 return code;
3245 /* Fold SUBREG. */
3247 static rtx
3248 fold_rtx_subreg (rtx x, rtx insn)
3250 enum machine_mode mode = GET_MODE (x);
3251 rtx folded_arg0;
3252 rtx const_arg0;
3253 rtx new;
3255 /* See if we previously assigned a constant value to this SUBREG. */
3256 if ((new = lookup_as_function (x, CONST_INT)) != 0
3257 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3258 return new;
3260 /* If this is a paradoxical SUBREG, we have no idea what value the
3261 extra bits would have. However, if the operand is equivalent to
3262 a SUBREG whose operand is the same as our mode, and all the modes
3263 are within a word, we can just use the inner operand because
3264 these SUBREGs just say how to treat the register.
3266 Similarly if we find an integer constant. */
3268 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3270 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3271 struct table_elt *elt;
3273 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3274 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3275 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3276 imode)) != 0)
3277 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3279 if (CONSTANT_P (elt->exp)
3280 && GET_MODE (elt->exp) == VOIDmode)
3281 return elt->exp;
3283 if (GET_CODE (elt->exp) == SUBREG
3284 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3285 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3286 return copy_rtx (SUBREG_REG (elt->exp));
3289 return x;
3292 /* Fold SUBREG_REG. If it changed, see if we can simplify the
3293 SUBREG. We might be able to if the SUBREG is extracting a single
3294 word in an integral mode or extracting the low part. */
3296 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3297 const_arg0 = equiv_constant (folded_arg0);
3298 if (const_arg0)
3299 folded_arg0 = const_arg0;
3301 if (folded_arg0 != SUBREG_REG (x))
3303 new = simplify_subreg (mode, folded_arg0,
3304 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3305 if (new)
3306 return new;
3309 if (REG_P (folded_arg0)
3310 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3312 struct table_elt *elt;
3314 elt = lookup (folded_arg0,
3315 HASH (folded_arg0, GET_MODE (folded_arg0)),
3316 GET_MODE (folded_arg0));
3318 if (elt)
3319 elt = elt->first_same_value;
3321 if (subreg_lowpart_p (x))
3322 /* If this is a narrowing SUBREG and our operand is a REG, see
3323 if we can find an equivalence for REG that is an arithmetic
3324 operation in a wider mode where both operands are
3325 paradoxical SUBREGs from objects of our result mode. In
3326 that case, we couldn-t report an equivalent value for that
3327 operation, since we don't know what the extra bits will be.
3328 But we can find an equivalence for this SUBREG by folding
3329 that operation in the narrow mode. This allows us to fold
3330 arithmetic in narrow modes when the machine only supports
3331 word-sized arithmetic.
3333 Also look for a case where we have a SUBREG whose operand
3334 is the same as our result. If both modes are smaller than
3335 a word, we are simply interpreting a register in different
3336 modes and we can use the inner value. */
3338 for (; elt; elt = elt->next_same_value)
3340 enum rtx_code eltcode = GET_CODE (elt->exp);
3342 /* Just check for unary and binary operations. */
3343 if (UNARY_P (elt->exp)
3344 && eltcode != SIGN_EXTEND
3345 && eltcode != ZERO_EXTEND
3346 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3347 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3348 && (GET_MODE_CLASS (mode)
3349 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3351 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3353 if (!REG_P (op0) && ! CONSTANT_P (op0))
3354 op0 = fold_rtx (op0, NULL_RTX);
3356 op0 = equiv_constant (op0);
3357 if (op0)
3358 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3359 op0, mode);
3361 else if (ARITHMETIC_P (elt->exp)
3362 && eltcode != DIV && eltcode != MOD
3363 && eltcode != UDIV && eltcode != UMOD
3364 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3365 && eltcode != ROTATE && eltcode != ROTATERT
3366 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3367 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3368 == mode))
3369 || CONSTANT_P (XEXP (elt->exp, 0)))
3370 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3371 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3372 == mode))
3373 || CONSTANT_P (XEXP (elt->exp, 1))))
3375 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3376 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3378 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3379 op0 = fold_rtx (op0, NULL_RTX);
3381 if (op0)
3382 op0 = equiv_constant (op0);
3384 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3385 op1 = fold_rtx (op1, NULL_RTX);
3387 if (op1)
3388 op1 = equiv_constant (op1);
3390 /* If we are looking for the low SImode part of
3391 (ashift:DI c (const_int 32)), it doesn't work to
3392 compute that in SImode, because a 32-bit shift in
3393 SImode is unpredictable. We know the value is
3394 0. */
3395 if (op0 && op1
3396 && GET_CODE (elt->exp) == ASHIFT
3397 && GET_CODE (op1) == CONST_INT
3398 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3400 if (INTVAL (op1)
3401 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3402 /* If the count fits in the inner mode's width,
3403 but exceeds the outer mode's width, the value
3404 will get truncated to 0 by the subreg. */
3405 new = CONST0_RTX (mode);
3406 else
3407 /* If the count exceeds even the inner mode's width,
3408 don't fold this expression. */
3409 new = 0;
3411 else if (op0 && op1)
3412 new = simplify_binary_operation (GET_CODE (elt->exp),
3413 mode, op0, op1);
3416 else if (GET_CODE (elt->exp) == SUBREG
3417 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3418 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3419 <= UNITS_PER_WORD)
3420 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3421 new = copy_rtx (SUBREG_REG (elt->exp));
3423 if (new)
3424 return new;
3426 else
3427 /* A SUBREG resulting from a zero extension may fold to zero
3428 if it extracts higher bits than the ZERO_EXTEND's source
3429 bits. FIXME: if combine tried to, er, combine these
3430 instructions, this transformation may be moved to
3431 simplify_subreg. */
3432 for (; elt; elt = elt->next_same_value)
3434 if (GET_CODE (elt->exp) == ZERO_EXTEND
3435 && subreg_lsb (x)
3436 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3437 return CONST0_RTX (mode);
3441 return x;
3444 /* Fold MEM. Not to be called directly, see fold_rtx_mem instead. */
3446 static rtx
3447 fold_rtx_mem_1 (rtx x, rtx insn)
3449 enum machine_mode mode = GET_MODE (x);
3450 rtx new;
3452 /* If we are not actually processing an insn, don't try to find the
3453 best address. Not only don't we care, but we could modify the
3454 MEM in an invalid way since we have no insn to validate
3455 against. */
3456 if (insn != 0)
3457 find_best_addr (insn, &XEXP (x, 0), mode);
3460 /* Even if we don't fold in the insn itself, we can safely do so
3461 here, in hopes of getting a constant. */
3462 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3463 rtx base = 0;
3464 HOST_WIDE_INT offset = 0;
3466 if (REG_P (addr)
3467 && REGNO_QTY_VALID_P (REGNO (addr)))
3469 int addr_q = REG_QTY (REGNO (addr));
3470 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3472 if (GET_MODE (addr) == addr_ent->mode
3473 && addr_ent->const_rtx != NULL_RTX)
3474 addr = addr_ent->const_rtx;
3477 /* Call target hook to avoid the effects of -fpic etc.... */
3478 addr = targetm.delegitimize_address (addr);
3480 /* If address is constant, split it into a base and integer
3481 offset. */
3482 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3483 base = addr;
3484 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3485 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3487 base = XEXP (XEXP (addr, 0), 0);
3488 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3490 else if (GET_CODE (addr) == LO_SUM
3491 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3492 base = XEXP (addr, 1);
3494 /* If this is a constant pool reference, we can fold it into its
3495 constant to allow better value tracking. */
3496 if (base && GET_CODE (base) == SYMBOL_REF
3497 && CONSTANT_POOL_ADDRESS_P (base))
3499 rtx constant = get_pool_constant (base);
3500 enum machine_mode const_mode = get_pool_mode (base);
3501 rtx new;
3503 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3505 constant_pool_entries_cost = COST (constant);
3506 constant_pool_entries_regcost = approx_reg_cost (constant);
3509 /* If we are loading the full constant, we have an
3510 equivalence. */
3511 if (offset == 0 && mode == const_mode)
3512 return constant;
3514 /* If this actually isn't a constant (weird!), we can't do
3515 anything. Otherwise, handle the two most common cases:
3516 extracting a word from a multi-word constant, and
3517 extracting the low-order bits. Other cases don't seem
3518 common enough to worry about. */
3519 if (! CONSTANT_P (constant))
3520 return x;
3522 if (GET_MODE_CLASS (mode) == MODE_INT
3523 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3524 && offset % UNITS_PER_WORD == 0
3525 && (new = operand_subword (constant,
3526 offset / UNITS_PER_WORD,
3527 0, const_mode)) != 0)
3528 return new;
3530 if (((BYTES_BIG_ENDIAN
3531 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3532 || (! BYTES_BIG_ENDIAN && offset == 0))
3533 && (new = gen_lowpart (mode, constant)) != 0)
3534 return new;
3537 /* If this is a reference to a label at a known position in a jump
3538 table, we also know its value. */
3539 if (base && GET_CODE (base) == LABEL_REF)
3541 rtx label = XEXP (base, 0);
3542 rtx table_insn = NEXT_INSN (label);
3544 if (table_insn && JUMP_P (table_insn)
3545 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3547 rtx table = PATTERN (table_insn);
3549 if (offset >= 0
3550 && (offset / GET_MODE_SIZE (GET_MODE (table))
3551 < XVECLEN (table, 0)))
3553 rtx label = XVECEXP
3554 (table, 0, offset / GET_MODE_SIZE (GET_MODE (table)));
3555 rtx set;
3557 /* If we have an insn that loads the label from the
3558 jumptable into a reg, we don't want to set the reg
3559 to the label, because this may cause a reference to
3560 the label to remain after the label is removed in
3561 some very obscure cases (PR middle-end/18628). */
3562 if (!insn)
3563 return label;
3565 set = single_set (insn);
3567 if (! set || SET_SRC (set) != x)
3568 return x;
3570 /* If it's a jump, it's safe to reference the label. */
3571 if (SET_DEST (set) == pc_rtx)
3572 return label;
3574 return x;
3577 if (table_insn && JUMP_P (table_insn)
3578 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3580 rtx table = PATTERN (table_insn);
3582 if (offset >= 0
3583 && (offset / GET_MODE_SIZE (GET_MODE (table))
3584 < XVECLEN (table, 1)))
3586 offset /= GET_MODE_SIZE (GET_MODE (table));
3587 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3588 XEXP (table, 0));
3590 if (GET_MODE (table) != Pmode)
3591 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3593 /* Indicate this is a constant. This isn't a valid
3594 form of CONST, but it will only be used to fold the
3595 next insns and then discarded, so it should be
3596 safe.
3598 Note this expression must be explicitly discarded,
3599 by cse_insn, else it may end up in a REG_EQUAL note
3600 and "escape" to cause problems elsewhere. */
3601 return gen_rtx_CONST (GET_MODE (new), new);
3606 return x;
3610 /* Fold MEM. */
3612 static rtx
3613 fold_rtx_mem (rtx x, rtx insn)
3615 /* To avoid infinite oscillations between fold_rtx and fold_rtx_mem,
3616 refuse to allow recursion of the latter past n levels. This can
3617 happen because fold_rtx_mem will try to fold the address of the
3618 memory reference it is passed, i.e. conceptually throwing away
3619 the MEM and reinjecting the bare address into fold_rtx. As a
3620 result, patterns like
3622 set (reg1)
3623 (plus (reg)
3624 (mem (plus (reg2) (const_int))))
3626 set (reg2)
3627 (plus (reg)
3628 (mem (plus (reg1) (const_int))))
3630 will defeat any "first-order" short-circuit put in either
3631 function to prevent these infinite oscillations.
3633 The heuristics for determining n is as follows: since each time
3634 it is invoked fold_rtx_mem throws away a MEM, and since MEMs
3635 are generically not nested, we assume that each invocation of
3636 fold_rtx_mem corresponds to a new "top-level" operand, i.e.
3637 the source or the destination of a SET. So fold_rtx_mem is
3638 bound to stop or cycle before n recursions, n being the number
3639 of expressions recorded in the hash table. We also leave some
3640 play to account for the initial steps. */
3642 static unsigned int depth;
3643 rtx ret;
3645 if (depth > 3 + table_size)
3646 return x;
3648 depth++;
3649 ret = fold_rtx_mem_1 (x, insn);
3650 depth--;
3652 return ret;
3655 /* If X is a nontrivial arithmetic operation on an argument
3656 for which a constant value can be determined, return
3657 the result of operating on that value, as a constant.
3658 Otherwise, return X, possibly with one or more operands
3659 modified by recursive calls to this function.
3661 If X is a register whose contents are known, we do NOT
3662 return those contents here. equiv_constant is called to
3663 perform that task.
3665 INSN is the insn that we may be modifying. If it is 0, make a copy
3666 of X before modifying it. */
3668 static rtx
3669 fold_rtx (rtx x, rtx insn)
3671 enum rtx_code code;
3672 enum machine_mode mode;
3673 const char *fmt;
3674 int i;
3675 rtx new = 0;
3676 int copied = 0;
3677 int must_swap = 0;
3679 /* Folded equivalents of first two operands of X. */
3680 rtx folded_arg0;
3681 rtx folded_arg1;
3683 /* Constant equivalents of first three operands of X;
3684 0 when no such equivalent is known. */
3685 rtx const_arg0;
3686 rtx const_arg1;
3687 rtx const_arg2;
3689 /* The mode of the first operand of X. We need this for sign and zero
3690 extends. */
3691 enum machine_mode mode_arg0;
3693 if (x == 0)
3694 return x;
3696 mode = GET_MODE (x);
3697 code = GET_CODE (x);
3698 switch (code)
3700 case CONST:
3701 case CONST_INT:
3702 case CONST_DOUBLE:
3703 case CONST_VECTOR:
3704 case SYMBOL_REF:
3705 case LABEL_REF:
3706 case REG:
3707 case PC:
3708 /* No use simplifying an EXPR_LIST
3709 since they are used only for lists of args
3710 in a function call's REG_EQUAL note. */
3711 case EXPR_LIST:
3712 return x;
3714 #ifdef HAVE_cc0
3715 case CC0:
3716 return prev_insn_cc0;
3717 #endif
3719 case SUBREG:
3720 return fold_rtx_subreg (x, insn);
3722 case NOT:
3723 case NEG:
3724 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3725 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3726 new = lookup_as_function (XEXP (x, 0), code);
3727 if (new)
3728 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3729 break;
3731 case MEM:
3732 return fold_rtx_mem (x, insn);
3734 #ifdef NO_FUNCTION_CSE
3735 case CALL:
3736 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3737 return x;
3738 break;
3739 #endif
3741 case ASM_OPERANDS:
3742 if (insn)
3744 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3745 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3746 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3748 break;
3750 default:
3751 break;
3754 const_arg0 = 0;
3755 const_arg1 = 0;
3756 const_arg2 = 0;
3757 mode_arg0 = VOIDmode;
3759 /* Try folding our operands.
3760 Then see which ones have constant values known. */
3762 fmt = GET_RTX_FORMAT (code);
3763 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3764 if (fmt[i] == 'e')
3766 rtx arg = XEXP (x, i);
3767 rtx folded_arg = arg, const_arg = 0;
3768 enum machine_mode mode_arg = GET_MODE (arg);
3769 rtx cheap_arg, expensive_arg;
3770 rtx replacements[2];
3771 int j;
3772 int old_cost = COST_IN (XEXP (x, i), code);
3774 /* Most arguments are cheap, so handle them specially. */
3775 switch (GET_CODE (arg))
3777 case REG:
3778 /* This is the same as calling equiv_constant; it is duplicated
3779 here for speed. */
3780 if (REGNO_QTY_VALID_P (REGNO (arg)))
3782 int arg_q = REG_QTY (REGNO (arg));
3783 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3785 if (arg_ent->const_rtx != NULL_RTX
3786 && !REG_P (arg_ent->const_rtx)
3787 && GET_CODE (arg_ent->const_rtx) != PLUS)
3788 const_arg
3789 = gen_lowpart (GET_MODE (arg),
3790 arg_ent->const_rtx);
3792 break;
3794 case CONST:
3795 case CONST_INT:
3796 case SYMBOL_REF:
3797 case LABEL_REF:
3798 case CONST_DOUBLE:
3799 case CONST_VECTOR:
3800 const_arg = arg;
3801 break;
3803 #ifdef HAVE_cc0
3804 case CC0:
3805 folded_arg = prev_insn_cc0;
3806 mode_arg = prev_insn_cc0_mode;
3807 const_arg = equiv_constant (folded_arg);
3808 break;
3809 #endif
3811 default:
3812 folded_arg = fold_rtx (arg, insn);
3813 const_arg = equiv_constant (folded_arg);
3816 /* For the first three operands, see if the operand
3817 is constant or equivalent to a constant. */
3818 switch (i)
3820 case 0:
3821 folded_arg0 = folded_arg;
3822 const_arg0 = const_arg;
3823 mode_arg0 = mode_arg;
3824 break;
3825 case 1:
3826 folded_arg1 = folded_arg;
3827 const_arg1 = const_arg;
3828 break;
3829 case 2:
3830 const_arg2 = const_arg;
3831 break;
3834 /* Pick the least expensive of the folded argument and an
3835 equivalent constant argument. */
3836 if (const_arg == 0 || const_arg == folded_arg
3837 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3838 cheap_arg = folded_arg, expensive_arg = const_arg;
3839 else
3840 cheap_arg = const_arg, expensive_arg = folded_arg;
3842 /* Try to replace the operand with the cheapest of the two
3843 possibilities. If it doesn't work and this is either of the first
3844 two operands of a commutative operation, try swapping them.
3845 If THAT fails, try the more expensive, provided it is cheaper
3846 than what is already there. */
3848 if (cheap_arg == XEXP (x, i))
3849 continue;
3851 if (insn == 0 && ! copied)
3853 x = copy_rtx (x);
3854 copied = 1;
3857 /* Order the replacements from cheapest to most expensive. */
3858 replacements[0] = cheap_arg;
3859 replacements[1] = expensive_arg;
3861 for (j = 0; j < 2 && replacements[j]; j++)
3863 int new_cost = COST_IN (replacements[j], code);
3865 /* Stop if what existed before was cheaper. Prefer constants
3866 in the case of a tie. */
3867 if (new_cost > old_cost
3868 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3869 break;
3871 /* It's not safe to substitute the operand of a conversion
3872 operator with a constant, as the conversion's identity
3873 depends upon the mode of its operand. This optimization
3874 is handled by the call to simplify_unary_operation. */
3875 if (GET_RTX_CLASS (code) == RTX_UNARY
3876 && GET_MODE (replacements[j]) != mode_arg0
3877 && (code == ZERO_EXTEND
3878 || code == SIGN_EXTEND
3879 || code == TRUNCATE
3880 || code == FLOAT_TRUNCATE
3881 || code == FLOAT_EXTEND
3882 || code == FLOAT
3883 || code == FIX
3884 || code == UNSIGNED_FLOAT
3885 || code == UNSIGNED_FIX))
3886 continue;
3888 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3889 break;
3891 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3892 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3894 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3895 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3897 if (apply_change_group ())
3899 /* Swap them back to be invalid so that this loop can
3900 continue and flag them to be swapped back later. */
3901 rtx tem;
3903 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3904 XEXP (x, 1) = tem;
3905 must_swap = 1;
3906 break;
3912 else
3914 if (fmt[i] == 'E')
3915 /* Don't try to fold inside of a vector of expressions.
3916 Doing nothing is harmless. */
3920 /* If a commutative operation, place a constant integer as the second
3921 operand unless the first operand is also a constant integer. Otherwise,
3922 place any constant second unless the first operand is also a constant. */
3924 if (COMMUTATIVE_P (x))
3926 if (must_swap
3927 || swap_commutative_operands_p (const_arg0 ? const_arg0
3928 : XEXP (x, 0),
3929 const_arg1 ? const_arg1
3930 : XEXP (x, 1)))
3932 rtx tem = XEXP (x, 0);
3934 if (insn == 0 && ! copied)
3936 x = copy_rtx (x);
3937 copied = 1;
3940 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3941 validate_change (insn, &XEXP (x, 1), tem, 1);
3942 if (apply_change_group ())
3944 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3945 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3950 /* If X is an arithmetic operation, see if we can simplify it. */
3952 switch (GET_RTX_CLASS (code))
3954 case RTX_UNARY:
3956 int is_const = 0;
3958 /* We can't simplify extension ops unless we know the
3959 original mode. */
3960 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3961 && mode_arg0 == VOIDmode)
3962 break;
3964 /* If we had a CONST, strip it off and put it back later if we
3965 fold. */
3966 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3967 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3969 new = simplify_unary_operation (code, mode,
3970 const_arg0 ? const_arg0 : folded_arg0,
3971 mode_arg0);
3972 /* NEG of PLUS could be converted into MINUS, but that causes
3973 expressions of the form
3974 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3975 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3976 FIXME: those ports should be fixed. */
3977 if (new != 0 && is_const
3978 && GET_CODE (new) == PLUS
3979 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3980 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3981 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3982 new = gen_rtx_CONST (mode, new);
3984 break;
3986 case RTX_COMPARE:
3987 case RTX_COMM_COMPARE:
3988 /* See what items are actually being compared and set FOLDED_ARG[01]
3989 to those values and CODE to the actual comparison code. If any are
3990 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3991 do anything if both operands are already known to be constant. */
3993 /* ??? Vector mode comparisons are not supported yet. */
3994 if (VECTOR_MODE_P (mode))
3995 break;
3997 if (const_arg0 == 0 || const_arg1 == 0)
3999 struct table_elt *p0, *p1;
4000 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4001 enum machine_mode mode_arg1;
4003 #ifdef FLOAT_STORE_FLAG_VALUE
4004 if (SCALAR_FLOAT_MODE_P (mode))
4006 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4007 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4008 false_rtx = CONST0_RTX (mode);
4010 #endif
4012 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
4013 &mode_arg0, &mode_arg1);
4015 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
4016 what kinds of things are being compared, so we can't do
4017 anything with this comparison. */
4019 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
4020 break;
4022 const_arg0 = equiv_constant (folded_arg0);
4023 const_arg1 = equiv_constant (folded_arg1);
4025 /* If we do not now have two constants being compared, see
4026 if we can nevertheless deduce some things about the
4027 comparison. */
4028 if (const_arg0 == 0 || const_arg1 == 0)
4030 if (const_arg1 != NULL)
4032 rtx cheapest_simplification;
4033 int cheapest_cost;
4034 rtx simp_result;
4035 struct table_elt *p;
4037 /* See if we can find an equivalent of folded_arg0
4038 that gets us a cheaper expression, possibly a
4039 constant through simplifications. */
4040 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
4041 mode_arg0);
4043 if (p != NULL)
4045 cheapest_simplification = x;
4046 cheapest_cost = COST (x);
4048 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
4050 int cost;
4052 /* If the entry isn't valid, skip it. */
4053 if (! exp_equiv_p (p->exp, p->exp, 1, false))
4054 continue;
4056 /* Try to simplify using this equivalence. */
4057 simp_result
4058 = simplify_relational_operation (code, mode,
4059 mode_arg0,
4060 p->exp,
4061 const_arg1);
4063 if (simp_result == NULL)
4064 continue;
4066 cost = COST (simp_result);
4067 if (cost < cheapest_cost)
4069 cheapest_cost = cost;
4070 cheapest_simplification = simp_result;
4074 /* If we have a cheaper expression now, use that
4075 and try folding it further, from the top. */
4076 if (cheapest_simplification != x)
4077 return fold_rtx (cheapest_simplification, insn);
4081 /* Some addresses are known to be nonzero. We don't know
4082 their sign, but equality comparisons are known. */
4083 if (const_arg1 == const0_rtx
4084 && nonzero_address_p (folded_arg0))
4086 if (code == EQ)
4087 return false_rtx;
4088 else if (code == NE)
4089 return true_rtx;
4092 /* See if the two operands are the same. */
4094 if (folded_arg0 == folded_arg1
4095 || (REG_P (folded_arg0)
4096 && REG_P (folded_arg1)
4097 && (REG_QTY (REGNO (folded_arg0))
4098 == REG_QTY (REGNO (folded_arg1))))
4099 || ((p0 = lookup (folded_arg0,
4100 SAFE_HASH (folded_arg0, mode_arg0),
4101 mode_arg0))
4102 && (p1 = lookup (folded_arg1,
4103 SAFE_HASH (folded_arg1, mode_arg0),
4104 mode_arg0))
4105 && p0->first_same_value == p1->first_same_value))
4107 /* Sadly two equal NaNs are not equivalent. */
4108 if (!HONOR_NANS (mode_arg0))
4109 return ((code == EQ || code == LE || code == GE
4110 || code == LEU || code == GEU || code == UNEQ
4111 || code == UNLE || code == UNGE
4112 || code == ORDERED)
4113 ? true_rtx : false_rtx);
4114 /* Take care for the FP compares we can resolve. */
4115 if (code == UNEQ || code == UNLE || code == UNGE)
4116 return true_rtx;
4117 if (code == LTGT || code == LT || code == GT)
4118 return false_rtx;
4121 /* If FOLDED_ARG0 is a register, see if the comparison we are
4122 doing now is either the same as we did before or the reverse
4123 (we only check the reverse if not floating-point). */
4124 else if (REG_P (folded_arg0))
4126 int qty = REG_QTY (REGNO (folded_arg0));
4128 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4130 struct qty_table_elem *ent = &qty_table[qty];
4132 if ((comparison_dominates_p (ent->comparison_code, code)
4133 || (! FLOAT_MODE_P (mode_arg0)
4134 && comparison_dominates_p (ent->comparison_code,
4135 reverse_condition (code))))
4136 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4137 || (const_arg1
4138 && rtx_equal_p (ent->comparison_const,
4139 const_arg1))
4140 || (REG_P (folded_arg1)
4141 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4142 return (comparison_dominates_p (ent->comparison_code, code)
4143 ? true_rtx : false_rtx);
4149 /* If we are comparing against zero, see if the first operand is
4150 equivalent to an IOR with a constant. If so, we may be able to
4151 determine the result of this comparison. */
4153 if (const_arg1 == const0_rtx)
4155 rtx y = lookup_as_function (folded_arg0, IOR);
4156 rtx inner_const;
4158 if (y != 0
4159 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4160 && GET_CODE (inner_const) == CONST_INT
4161 && INTVAL (inner_const) != 0)
4163 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4164 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4165 && (INTVAL (inner_const)
4166 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4167 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4169 #ifdef FLOAT_STORE_FLAG_VALUE
4170 if (SCALAR_FLOAT_MODE_P (mode))
4172 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4173 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4174 false_rtx = CONST0_RTX (mode);
4176 #endif
4178 switch (code)
4180 case EQ:
4181 return false_rtx;
4182 case NE:
4183 return true_rtx;
4184 case LT: case LE:
4185 if (has_sign)
4186 return true_rtx;
4187 break;
4188 case GT: case GE:
4189 if (has_sign)
4190 return false_rtx;
4191 break;
4192 default:
4193 break;
4199 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4200 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4201 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4203 break;
4205 case RTX_BIN_ARITH:
4206 case RTX_COMM_ARITH:
4207 switch (code)
4209 case PLUS:
4210 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4211 with that LABEL_REF as its second operand. If so, the result is
4212 the first operand of that MINUS. This handles switches with an
4213 ADDR_DIFF_VEC table. */
4214 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4216 rtx y
4217 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4218 : lookup_as_function (folded_arg0, MINUS);
4220 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4221 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4222 return XEXP (y, 0);
4224 /* Now try for a CONST of a MINUS like the above. */
4225 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4226 : lookup_as_function (folded_arg0, CONST))) != 0
4227 && GET_CODE (XEXP (y, 0)) == MINUS
4228 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4229 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4230 return XEXP (XEXP (y, 0), 0);
4233 /* Likewise if the operands are in the other order. */
4234 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4236 rtx y
4237 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4238 : lookup_as_function (folded_arg1, MINUS);
4240 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4241 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4242 return XEXP (y, 0);
4244 /* Now try for a CONST of a MINUS like the above. */
4245 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4246 : lookup_as_function (folded_arg1, CONST))) != 0
4247 && GET_CODE (XEXP (y, 0)) == MINUS
4248 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4249 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4250 return XEXP (XEXP (y, 0), 0);
4253 /* If second operand is a register equivalent to a negative
4254 CONST_INT, see if we can find a register equivalent to the
4255 positive constant. Make a MINUS if so. Don't do this for
4256 a non-negative constant since we might then alternate between
4257 choosing positive and negative constants. Having the positive
4258 constant previously-used is the more common case. Be sure
4259 the resulting constant is non-negative; if const_arg1 were
4260 the smallest negative number this would overflow: depending
4261 on the mode, this would either just be the same value (and
4262 hence not save anything) or be incorrect. */
4263 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4264 && INTVAL (const_arg1) < 0
4265 /* This used to test
4267 -INTVAL (const_arg1) >= 0
4269 But The Sun V5.0 compilers mis-compiled that test. So
4270 instead we test for the problematic value in a more direct
4271 manner and hope the Sun compilers get it correct. */
4272 && INTVAL (const_arg1) !=
4273 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4274 && REG_P (folded_arg1))
4276 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4277 struct table_elt *p
4278 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4280 if (p)
4281 for (p = p->first_same_value; p; p = p->next_same_value)
4282 if (REG_P (p->exp))
4283 return simplify_gen_binary (MINUS, mode, folded_arg0,
4284 canon_reg (p->exp, NULL_RTX));
4286 goto from_plus;
4288 case MINUS:
4289 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4290 If so, produce (PLUS Z C2-C). */
4291 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4293 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4294 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4295 return fold_rtx (plus_constant (copy_rtx (y),
4296 -INTVAL (const_arg1)),
4297 NULL_RTX);
4300 /* Fall through. */
4302 from_plus:
4303 case SMIN: case SMAX: case UMIN: case UMAX:
4304 case IOR: case AND: case XOR:
4305 case MULT:
4306 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4307 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4308 is known to be of similar form, we may be able to replace the
4309 operation with a combined operation. This may eliminate the
4310 intermediate operation if every use is simplified in this way.
4311 Note that the similar optimization done by combine.c only works
4312 if the intermediate operation's result has only one reference. */
4314 if (REG_P (folded_arg0)
4315 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4317 int is_shift
4318 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4319 rtx y, inner_const, new_const;
4320 enum rtx_code associate_code;
4322 if (is_shift
4323 && (INTVAL (const_arg1) >= GET_MODE_BITSIZE (mode)
4324 || INTVAL (const_arg1) < 0))
4326 if (SHIFT_COUNT_TRUNCATED)
4327 const_arg1 = GEN_INT (INTVAL (const_arg1)
4328 & (GET_MODE_BITSIZE (mode) - 1));
4329 else
4330 break;
4333 y = lookup_as_function (folded_arg0, code);
4334 if (y == 0)
4335 break;
4337 /* If we have compiled a statement like
4338 "if (x == (x & mask1))", and now are looking at
4339 "x & mask2", we will have a case where the first operand
4340 of Y is the same as our first operand. Unless we detect
4341 this case, an infinite loop will result. */
4342 if (XEXP (y, 0) == folded_arg0)
4343 break;
4345 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
4346 if (!inner_const || GET_CODE (inner_const) != CONST_INT)
4347 break;
4349 /* Don't associate these operations if they are a PLUS with the
4350 same constant and it is a power of two. These might be doable
4351 with a pre- or post-increment. Similarly for two subtracts of
4352 identical powers of two with post decrement. */
4354 if (code == PLUS && const_arg1 == inner_const
4355 && ((HAVE_PRE_INCREMENT
4356 && exact_log2 (INTVAL (const_arg1)) >= 0)
4357 || (HAVE_POST_INCREMENT
4358 && exact_log2 (INTVAL (const_arg1)) >= 0)
4359 || (HAVE_PRE_DECREMENT
4360 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4361 || (HAVE_POST_DECREMENT
4362 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4363 break;
4365 if (is_shift
4366 && (INTVAL (inner_const) >= GET_MODE_BITSIZE (mode)
4367 || INTVAL (inner_const) < 0))
4369 if (SHIFT_COUNT_TRUNCATED)
4370 inner_const = GEN_INT (INTVAL (inner_const)
4371 & (GET_MODE_BITSIZE (mode) - 1));
4372 else
4373 break;
4376 /* Compute the code used to compose the constants. For example,
4377 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4379 associate_code = (is_shift || code == MINUS ? PLUS : code);
4381 new_const = simplify_binary_operation (associate_code, mode,
4382 const_arg1, inner_const);
4384 if (new_const == 0)
4385 break;
4387 /* If we are associating shift operations, don't let this
4388 produce a shift of the size of the object or larger.
4389 This could occur when we follow a sign-extend by a right
4390 shift on a machine that does a sign-extend as a pair
4391 of shifts. */
4393 if (is_shift
4394 && GET_CODE (new_const) == CONST_INT
4395 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4397 /* As an exception, we can turn an ASHIFTRT of this
4398 form into a shift of the number of bits - 1. */
4399 if (code == ASHIFTRT)
4400 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4401 else if (!side_effects_p (XEXP (y, 0)))
4402 return CONST0_RTX (mode);
4403 else
4404 break;
4407 y = copy_rtx (XEXP (y, 0));
4409 /* If Y contains our first operand (the most common way this
4410 can happen is if Y is a MEM), we would do into an infinite
4411 loop if we tried to fold it. So don't in that case. */
4413 if (! reg_mentioned_p (folded_arg0, y))
4414 y = fold_rtx (y, insn);
4416 return simplify_gen_binary (code, mode, y, new_const);
4418 break;
4420 case DIV: case UDIV:
4421 /* ??? The associative optimization performed immediately above is
4422 also possible for DIV and UDIV using associate_code of MULT.
4423 However, we would need extra code to verify that the
4424 multiplication does not overflow, that is, there is no overflow
4425 in the calculation of new_const. */
4426 break;
4428 default:
4429 break;
4432 new = simplify_binary_operation (code, mode,
4433 const_arg0 ? const_arg0 : folded_arg0,
4434 const_arg1 ? const_arg1 : folded_arg1);
4435 break;
4437 case RTX_OBJ:
4438 /* (lo_sum (high X) X) is simply X. */
4439 if (code == LO_SUM && const_arg0 != 0
4440 && GET_CODE (const_arg0) == HIGH
4441 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4442 return const_arg1;
4443 break;
4445 case RTX_TERNARY:
4446 case RTX_BITFIELD_OPS:
4447 new = simplify_ternary_operation (code, mode, mode_arg0,
4448 const_arg0 ? const_arg0 : folded_arg0,
4449 const_arg1 ? const_arg1 : folded_arg1,
4450 const_arg2 ? const_arg2 : XEXP (x, 2));
4451 break;
4453 default:
4454 break;
4457 return new ? new : x;
4460 /* Return a constant value currently equivalent to X.
4461 Return 0 if we don't know one. */
4463 static rtx
4464 equiv_constant (rtx x)
4466 if (REG_P (x)
4467 && REGNO_QTY_VALID_P (REGNO (x)))
4469 int x_q = REG_QTY (REGNO (x));
4470 struct qty_table_elem *x_ent = &qty_table[x_q];
4472 if (x_ent->const_rtx)
4473 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4476 if (x == 0 || CONSTANT_P (x))
4477 return x;
4479 /* If X is a MEM, try to fold it outside the context of any insn to see if
4480 it might be equivalent to a constant. That handles the case where it
4481 is a constant-pool reference. Then try to look it up in the hash table
4482 in case it is something whose value we have seen before. */
4484 if (MEM_P (x))
4486 struct table_elt *elt;
4488 x = fold_rtx (x, NULL_RTX);
4489 if (CONSTANT_P (x))
4490 return x;
4492 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4493 if (elt == 0)
4494 return 0;
4496 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4497 if (elt->is_const && CONSTANT_P (elt->exp))
4498 return elt->exp;
4501 return 0;
4504 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4505 branch. It will be zero if not.
4507 In certain cases, this can cause us to add an equivalence. For example,
4508 if we are following the taken case of
4509 if (i == 2)
4510 we can add the fact that `i' and '2' are now equivalent.
4512 In any case, we can record that this comparison was passed. If the same
4513 comparison is seen later, we will know its value. */
4515 static void
4516 record_jump_equiv (rtx insn, int taken)
4518 int cond_known_true;
4519 rtx op0, op1;
4520 rtx set;
4521 enum machine_mode mode, mode0, mode1;
4522 int reversed_nonequality = 0;
4523 enum rtx_code code;
4525 /* Ensure this is the right kind of insn. */
4526 if (! any_condjump_p (insn))
4527 return;
4528 set = pc_set (insn);
4530 /* See if this jump condition is known true or false. */
4531 if (taken)
4532 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4533 else
4534 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4536 /* Get the type of comparison being done and the operands being compared.
4537 If we had to reverse a non-equality condition, record that fact so we
4538 know that it isn't valid for floating-point. */
4539 code = GET_CODE (XEXP (SET_SRC (set), 0));
4540 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4541 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4543 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4545 /* If the mode is a MODE_CC mode, we don't know what kinds of things
4546 are being compared, so we can't do anything with this
4547 comparison. */
4549 if (GET_MODE_CLASS (mode0) == MODE_CC)
4550 return;
4552 if (! cond_known_true)
4554 code = reversed_comparison_code_parts (code, op0, op1, insn);
4556 /* Don't remember if we can't find the inverse. */
4557 if (code == UNKNOWN)
4558 return;
4561 /* The mode is the mode of the non-constant. */
4562 mode = mode0;
4563 if (mode1 != VOIDmode)
4564 mode = mode1;
4566 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4569 /* Yet another form of subreg creation. In this case, we want something in
4570 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4572 static rtx
4573 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4575 enum machine_mode op_mode = GET_MODE (op);
4576 if (op_mode == mode || op_mode == VOIDmode)
4577 return op;
4578 return lowpart_subreg (mode, op, op_mode);
4581 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4582 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4583 Make any useful entries we can with that information. Called from
4584 above function and called recursively. */
4586 static void
4587 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4588 rtx op1, int reversed_nonequality)
4590 unsigned op0_hash, op1_hash;
4591 int op0_in_memory, op1_in_memory;
4592 struct table_elt *op0_elt, *op1_elt;
4594 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4595 we know that they are also equal in the smaller mode (this is also
4596 true for all smaller modes whether or not there is a SUBREG, but
4597 is not worth testing for with no SUBREG). */
4599 /* Note that GET_MODE (op0) may not equal MODE. */
4600 if (code == EQ && GET_CODE (op0) == SUBREG
4601 && (GET_MODE_SIZE (GET_MODE (op0))
4602 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4604 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4605 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4606 if (tem)
4607 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4608 reversed_nonequality);
4611 if (code == EQ && GET_CODE (op1) == SUBREG
4612 && (GET_MODE_SIZE (GET_MODE (op1))
4613 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4615 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4616 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4617 if (tem)
4618 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4619 reversed_nonequality);
4622 /* Similarly, if this is an NE comparison, and either is a SUBREG
4623 making a smaller mode, we know the whole thing is also NE. */
4625 /* Note that GET_MODE (op0) may not equal MODE;
4626 if we test MODE instead, we can get an infinite recursion
4627 alternating between two modes each wider than MODE. */
4629 if (code == NE && GET_CODE (op0) == SUBREG
4630 && subreg_lowpart_p (op0)
4631 && (GET_MODE_SIZE (GET_MODE (op0))
4632 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4634 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4635 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4636 if (tem)
4637 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4638 reversed_nonequality);
4641 if (code == NE && GET_CODE (op1) == SUBREG
4642 && subreg_lowpart_p (op1)
4643 && (GET_MODE_SIZE (GET_MODE (op1))
4644 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4646 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4647 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4648 if (tem)
4649 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4650 reversed_nonequality);
4653 /* Hash both operands. */
4655 do_not_record = 0;
4656 hash_arg_in_memory = 0;
4657 op0_hash = HASH (op0, mode);
4658 op0_in_memory = hash_arg_in_memory;
4660 if (do_not_record)
4661 return;
4663 do_not_record = 0;
4664 hash_arg_in_memory = 0;
4665 op1_hash = HASH (op1, mode);
4666 op1_in_memory = hash_arg_in_memory;
4668 if (do_not_record)
4669 return;
4671 /* Look up both operands. */
4672 op0_elt = lookup (op0, op0_hash, mode);
4673 op1_elt = lookup (op1, op1_hash, mode);
4675 /* If both operands are already equivalent or if they are not in the
4676 table but are identical, do nothing. */
4677 if ((op0_elt != 0 && op1_elt != 0
4678 && op0_elt->first_same_value == op1_elt->first_same_value)
4679 || op0 == op1 || rtx_equal_p (op0, op1))
4680 return;
4682 /* If we aren't setting two things equal all we can do is save this
4683 comparison. Similarly if this is floating-point. In the latter
4684 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4685 If we record the equality, we might inadvertently delete code
4686 whose intent was to change -0 to +0. */
4688 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4690 struct qty_table_elem *ent;
4691 int qty;
4693 /* If we reversed a floating-point comparison, if OP0 is not a
4694 register, or if OP1 is neither a register or constant, we can't
4695 do anything. */
4697 if (!REG_P (op1))
4698 op1 = equiv_constant (op1);
4700 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4701 || !REG_P (op0) || op1 == 0)
4702 return;
4704 /* Put OP0 in the hash table if it isn't already. This gives it a
4705 new quantity number. */
4706 if (op0_elt == 0)
4708 if (insert_regs (op0, NULL, 0))
4710 rehash_using_reg (op0);
4711 op0_hash = HASH (op0, mode);
4713 /* If OP0 is contained in OP1, this changes its hash code
4714 as well. Faster to rehash than to check, except
4715 for the simple case of a constant. */
4716 if (! CONSTANT_P (op1))
4717 op1_hash = HASH (op1,mode);
4720 op0_elt = insert (op0, NULL, op0_hash, mode);
4721 op0_elt->in_memory = op0_in_memory;
4724 qty = REG_QTY (REGNO (op0));
4725 ent = &qty_table[qty];
4727 ent->comparison_code = code;
4728 if (REG_P (op1))
4730 /* Look it up again--in case op0 and op1 are the same. */
4731 op1_elt = lookup (op1, op1_hash, mode);
4733 /* Put OP1 in the hash table so it gets a new quantity number. */
4734 if (op1_elt == 0)
4736 if (insert_regs (op1, NULL, 0))
4738 rehash_using_reg (op1);
4739 op1_hash = HASH (op1, mode);
4742 op1_elt = insert (op1, NULL, op1_hash, mode);
4743 op1_elt->in_memory = op1_in_memory;
4746 ent->comparison_const = NULL_RTX;
4747 ent->comparison_qty = REG_QTY (REGNO (op1));
4749 else
4751 ent->comparison_const = op1;
4752 ent->comparison_qty = -1;
4755 return;
4758 /* If either side is still missing an equivalence, make it now,
4759 then merge the equivalences. */
4761 if (op0_elt == 0)
4763 if (insert_regs (op0, NULL, 0))
4765 rehash_using_reg (op0);
4766 op0_hash = HASH (op0, mode);
4769 op0_elt = insert (op0, NULL, op0_hash, mode);
4770 op0_elt->in_memory = op0_in_memory;
4773 if (op1_elt == 0)
4775 if (insert_regs (op1, NULL, 0))
4777 rehash_using_reg (op1);
4778 op1_hash = HASH (op1, mode);
4781 op1_elt = insert (op1, NULL, op1_hash, mode);
4782 op1_elt->in_memory = op1_in_memory;
4785 merge_equiv_classes (op0_elt, op1_elt);
4788 /* CSE processing for one instruction.
4789 First simplify sources and addresses of all assignments
4790 in the instruction, using previously-computed equivalents values.
4791 Then install the new sources and destinations in the table
4792 of available values.
4794 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4795 the insn. It means that INSN is inside libcall block. In this
4796 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4798 /* Data on one SET contained in the instruction. */
4800 struct set
4802 /* The SET rtx itself. */
4803 rtx rtl;
4804 /* The SET_SRC of the rtx (the original value, if it is changing). */
4805 rtx src;
4806 /* The hash-table element for the SET_SRC of the SET. */
4807 struct table_elt *src_elt;
4808 /* Hash value for the SET_SRC. */
4809 unsigned src_hash;
4810 /* Hash value for the SET_DEST. */
4811 unsigned dest_hash;
4812 /* The SET_DEST, with SUBREG, etc., stripped. */
4813 rtx inner_dest;
4814 /* Nonzero if the SET_SRC is in memory. */
4815 char src_in_memory;
4816 /* Nonzero if the SET_SRC contains something
4817 whose value cannot be predicted and understood. */
4818 char src_volatile;
4819 /* Original machine mode, in case it becomes a CONST_INT.
4820 The size of this field should match the size of the mode
4821 field of struct rtx_def (see rtl.h). */
4822 ENUM_BITFIELD(machine_mode) mode : 8;
4823 /* A constant equivalent for SET_SRC, if any. */
4824 rtx src_const;
4825 /* Original SET_SRC value used for libcall notes. */
4826 rtx orig_src;
4827 /* Hash value of constant equivalent for SET_SRC. */
4828 unsigned src_const_hash;
4829 /* Table entry for constant equivalent for SET_SRC, if any. */
4830 struct table_elt *src_const_elt;
4831 /* Table entry for the destination address. */
4832 struct table_elt *dest_addr_elt;
4835 static void
4836 cse_insn (rtx insn, rtx libcall_insn)
4838 rtx x = PATTERN (insn);
4839 int i;
4840 rtx tem;
4841 int n_sets = 0;
4843 #ifdef HAVE_cc0
4844 /* Records what this insn does to set CC0. */
4845 rtx this_insn_cc0 = 0;
4846 enum machine_mode this_insn_cc0_mode = VOIDmode;
4847 #endif
4849 rtx src_eqv = 0;
4850 struct table_elt *src_eqv_elt = 0;
4851 int src_eqv_volatile = 0;
4852 int src_eqv_in_memory = 0;
4853 unsigned src_eqv_hash = 0;
4855 struct set *sets = (struct set *) 0;
4857 this_insn = insn;
4859 /* Find all the SETs and CLOBBERs in this instruction.
4860 Record all the SETs in the array `set' and count them.
4861 Also determine whether there is a CLOBBER that invalidates
4862 all memory references, or all references at varying addresses. */
4864 if (CALL_P (insn))
4866 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4868 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4869 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4870 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4874 if (GET_CODE (x) == SET)
4876 sets = alloca (sizeof (struct set));
4877 sets[0].rtl = x;
4879 /* Ignore SETs that are unconditional jumps.
4880 They never need cse processing, so this does not hurt.
4881 The reason is not efficiency but rather
4882 so that we can test at the end for instructions
4883 that have been simplified to unconditional jumps
4884 and not be misled by unchanged instructions
4885 that were unconditional jumps to begin with. */
4886 if (SET_DEST (x) == pc_rtx
4887 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4890 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4891 The hard function value register is used only once, to copy to
4892 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4893 Ensure we invalidate the destination register. On the 80386 no
4894 other code would invalidate it since it is a fixed_reg.
4895 We need not check the return of apply_change_group; see canon_reg. */
4897 else if (GET_CODE (SET_SRC (x)) == CALL)
4899 canon_reg (SET_SRC (x), insn);
4900 apply_change_group ();
4901 fold_rtx (SET_SRC (x), insn);
4902 invalidate (SET_DEST (x), VOIDmode);
4904 else
4905 n_sets = 1;
4907 else if (GET_CODE (x) == PARALLEL)
4909 int lim = XVECLEN (x, 0);
4911 sets = alloca (lim * sizeof (struct set));
4913 /* Find all regs explicitly clobbered in this insn,
4914 and ensure they are not replaced with any other regs
4915 elsewhere in this insn.
4916 When a reg that is clobbered is also used for input,
4917 we should presume that that is for a reason,
4918 and we should not substitute some other register
4919 which is not supposed to be clobbered.
4920 Therefore, this loop cannot be merged into the one below
4921 because a CALL may precede a CLOBBER and refer to the
4922 value clobbered. We must not let a canonicalization do
4923 anything in that case. */
4924 for (i = 0; i < lim; i++)
4926 rtx y = XVECEXP (x, 0, i);
4927 if (GET_CODE (y) == CLOBBER)
4929 rtx clobbered = XEXP (y, 0);
4931 if (REG_P (clobbered)
4932 || GET_CODE (clobbered) == SUBREG)
4933 invalidate (clobbered, VOIDmode);
4934 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4935 || GET_CODE (clobbered) == ZERO_EXTRACT)
4936 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4940 for (i = 0; i < lim; i++)
4942 rtx y = XVECEXP (x, 0, i);
4943 if (GET_CODE (y) == SET)
4945 /* As above, we ignore unconditional jumps and call-insns and
4946 ignore the result of apply_change_group. */
4947 if (GET_CODE (SET_SRC (y)) == CALL)
4949 canon_reg (SET_SRC (y), insn);
4950 apply_change_group ();
4951 fold_rtx (SET_SRC (y), insn);
4952 invalidate (SET_DEST (y), VOIDmode);
4954 else if (SET_DEST (y) == pc_rtx
4955 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4957 else
4958 sets[n_sets++].rtl = y;
4960 else if (GET_CODE (y) == CLOBBER)
4962 /* If we clobber memory, canon the address.
4963 This does nothing when a register is clobbered
4964 because we have already invalidated the reg. */
4965 if (MEM_P (XEXP (y, 0)))
4966 canon_reg (XEXP (y, 0), NULL_RTX);
4968 else if (GET_CODE (y) == USE
4969 && ! (REG_P (XEXP (y, 0))
4970 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4971 canon_reg (y, NULL_RTX);
4972 else if (GET_CODE (y) == CALL)
4974 /* The result of apply_change_group can be ignored; see
4975 canon_reg. */
4976 canon_reg (y, insn);
4977 apply_change_group ();
4978 fold_rtx (y, insn);
4982 else if (GET_CODE (x) == CLOBBER)
4984 if (MEM_P (XEXP (x, 0)))
4985 canon_reg (XEXP (x, 0), NULL_RTX);
4988 /* Canonicalize a USE of a pseudo register or memory location. */
4989 else if (GET_CODE (x) == USE
4990 && ! (REG_P (XEXP (x, 0))
4991 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4992 canon_reg (XEXP (x, 0), NULL_RTX);
4993 else if (GET_CODE (x) == CALL)
4995 /* The result of apply_change_group can be ignored; see canon_reg. */
4996 canon_reg (x, insn);
4997 apply_change_group ();
4998 fold_rtx (x, insn);
5001 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
5002 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
5003 is handled specially for this case, and if it isn't set, then there will
5004 be no equivalence for the destination. */
5005 if (n_sets == 1 && REG_NOTES (insn) != 0
5006 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
5007 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
5008 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
5010 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
5011 XEXP (tem, 0) = src_eqv;
5014 /* Canonicalize sources and addresses of destinations.
5015 We do this in a separate pass to avoid problems when a MATCH_DUP is
5016 present in the insn pattern. In that case, we want to ensure that
5017 we don't break the duplicate nature of the pattern. So we will replace
5018 both operands at the same time. Otherwise, we would fail to find an
5019 equivalent substitution in the loop calling validate_change below.
5021 We used to suppress canonicalization of DEST if it appears in SRC,
5022 but we don't do this any more. */
5024 for (i = 0; i < n_sets; i++)
5026 rtx dest = SET_DEST (sets[i].rtl);
5027 rtx src = SET_SRC (sets[i].rtl);
5028 rtx new = canon_reg (src, insn);
5030 sets[i].orig_src = src;
5031 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5033 if (GET_CODE (dest) == ZERO_EXTRACT)
5035 validate_change (insn, &XEXP (dest, 1),
5036 canon_reg (XEXP (dest, 1), insn), 1);
5037 validate_change (insn, &XEXP (dest, 2),
5038 canon_reg (XEXP (dest, 2), insn), 1);
5041 while (GET_CODE (dest) == SUBREG
5042 || GET_CODE (dest) == ZERO_EXTRACT
5043 || GET_CODE (dest) == STRICT_LOW_PART)
5044 dest = XEXP (dest, 0);
5046 if (MEM_P (dest))
5047 canon_reg (dest, insn);
5050 /* Now that we have done all the replacements, we can apply the change
5051 group and see if they all work. Note that this will cause some
5052 canonicalizations that would have worked individually not to be applied
5053 because some other canonicalization didn't work, but this should not
5054 occur often.
5056 The result of apply_change_group can be ignored; see canon_reg. */
5058 apply_change_group ();
5060 /* Set sets[i].src_elt to the class each source belongs to.
5061 Detect assignments from or to volatile things
5062 and set set[i] to zero so they will be ignored
5063 in the rest of this function.
5065 Nothing in this loop changes the hash table or the register chains. */
5067 for (i = 0; i < n_sets; i++)
5069 rtx src, dest;
5070 rtx src_folded;
5071 struct table_elt *elt = 0, *p;
5072 enum machine_mode mode;
5073 rtx src_eqv_here;
5074 rtx src_const = 0;
5075 rtx src_related = 0;
5076 struct table_elt *src_const_elt = 0;
5077 int src_cost = MAX_COST;
5078 int src_eqv_cost = MAX_COST;
5079 int src_folded_cost = MAX_COST;
5080 int src_related_cost = MAX_COST;
5081 int src_elt_cost = MAX_COST;
5082 int src_regcost = MAX_COST;
5083 int src_eqv_regcost = MAX_COST;
5084 int src_folded_regcost = MAX_COST;
5085 int src_related_regcost = MAX_COST;
5086 int src_elt_regcost = MAX_COST;
5087 /* Set nonzero if we need to call force_const_mem on with the
5088 contents of src_folded before using it. */
5089 int src_folded_force_flag = 0;
5091 dest = SET_DEST (sets[i].rtl);
5092 src = SET_SRC (sets[i].rtl);
5094 /* If SRC is a constant that has no machine mode,
5095 hash it with the destination's machine mode.
5096 This way we can keep different modes separate. */
5098 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5099 sets[i].mode = mode;
5101 if (src_eqv)
5103 enum machine_mode eqvmode = mode;
5104 if (GET_CODE (dest) == STRICT_LOW_PART)
5105 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5106 do_not_record = 0;
5107 hash_arg_in_memory = 0;
5108 src_eqv_hash = HASH (src_eqv, eqvmode);
5110 /* Find the equivalence class for the equivalent expression. */
5112 if (!do_not_record)
5113 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
5115 src_eqv_volatile = do_not_record;
5116 src_eqv_in_memory = hash_arg_in_memory;
5119 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
5120 value of the INNER register, not the destination. So it is not
5121 a valid substitution for the source. But save it for later. */
5122 if (GET_CODE (dest) == STRICT_LOW_PART)
5123 src_eqv_here = 0;
5124 else
5125 src_eqv_here = src_eqv;
5127 /* Simplify and foldable subexpressions in SRC. Then get the fully-
5128 simplified result, which may not necessarily be valid. */
5129 src_folded = fold_rtx (src, insn);
5131 #if 0
5132 /* ??? This caused bad code to be generated for the m68k port with -O2.
5133 Suppose src is (CONST_INT -1), and that after truncation src_folded
5134 is (CONST_INT 3). Suppose src_folded is then used for src_const.
5135 At the end we will add src and src_const to the same equivalence
5136 class. We now have 3 and -1 on the same equivalence class. This
5137 causes later instructions to be mis-optimized. */
5138 /* If storing a constant in a bitfield, pre-truncate the constant
5139 so we will be able to record it later. */
5140 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5142 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5144 if (GET_CODE (src) == CONST_INT
5145 && GET_CODE (width) == CONST_INT
5146 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5147 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5148 src_folded
5149 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5150 << INTVAL (width)) - 1));
5152 #endif
5154 /* Compute SRC's hash code, and also notice if it
5155 should not be recorded at all. In that case,
5156 prevent any further processing of this assignment. */
5157 do_not_record = 0;
5158 hash_arg_in_memory = 0;
5160 sets[i].src = src;
5161 sets[i].src_hash = HASH (src, mode);
5162 sets[i].src_volatile = do_not_record;
5163 sets[i].src_in_memory = hash_arg_in_memory;
5165 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5166 a pseudo, do not record SRC. Using SRC as a replacement for
5167 anything else will be incorrect in that situation. Note that
5168 this usually occurs only for stack slots, in which case all the
5169 RTL would be referring to SRC, so we don't lose any optimization
5170 opportunities by not having SRC in the hash table. */
5172 if (MEM_P (src)
5173 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5174 && REG_P (dest)
5175 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5176 sets[i].src_volatile = 1;
5178 #if 0
5179 /* It is no longer clear why we used to do this, but it doesn't
5180 appear to still be needed. So let's try without it since this
5181 code hurts cse'ing widened ops. */
5182 /* If source is a paradoxical subreg (such as QI treated as an SI),
5183 treat it as volatile. It may do the work of an SI in one context
5184 where the extra bits are not being used, but cannot replace an SI
5185 in general. */
5186 if (GET_CODE (src) == SUBREG
5187 && (GET_MODE_SIZE (GET_MODE (src))
5188 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5189 sets[i].src_volatile = 1;
5190 #endif
5192 /* Locate all possible equivalent forms for SRC. Try to replace
5193 SRC in the insn with each cheaper equivalent.
5195 We have the following types of equivalents: SRC itself, a folded
5196 version, a value given in a REG_EQUAL note, or a value related
5197 to a constant.
5199 Each of these equivalents may be part of an additional class
5200 of equivalents (if more than one is in the table, they must be in
5201 the same class; we check for this).
5203 If the source is volatile, we don't do any table lookups.
5205 We note any constant equivalent for possible later use in a
5206 REG_NOTE. */
5208 if (!sets[i].src_volatile)
5209 elt = lookup (src, sets[i].src_hash, mode);
5211 sets[i].src_elt = elt;
5213 if (elt && src_eqv_here && src_eqv_elt)
5215 if (elt->first_same_value != src_eqv_elt->first_same_value)
5217 /* The REG_EQUAL is indicating that two formerly distinct
5218 classes are now equivalent. So merge them. */
5219 merge_equiv_classes (elt, src_eqv_elt);
5220 src_eqv_hash = HASH (src_eqv, elt->mode);
5221 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5224 src_eqv_here = 0;
5227 else if (src_eqv_elt)
5228 elt = src_eqv_elt;
5230 /* Try to find a constant somewhere and record it in `src_const'.
5231 Record its table element, if any, in `src_const_elt'. Look in
5232 any known equivalences first. (If the constant is not in the
5233 table, also set `sets[i].src_const_hash'). */
5234 if (elt)
5235 for (p = elt->first_same_value; p; p = p->next_same_value)
5236 if (p->is_const)
5238 src_const = p->exp;
5239 src_const_elt = elt;
5240 break;
5243 if (src_const == 0
5244 && (CONSTANT_P (src_folded)
5245 /* Consider (minus (label_ref L1) (label_ref L2)) as
5246 "constant" here so we will record it. This allows us
5247 to fold switch statements when an ADDR_DIFF_VEC is used. */
5248 || (GET_CODE (src_folded) == MINUS
5249 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5250 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5251 src_const = src_folded, src_const_elt = elt;
5252 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5253 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5255 /* If we don't know if the constant is in the table, get its
5256 hash code and look it up. */
5257 if (src_const && src_const_elt == 0)
5259 sets[i].src_const_hash = HASH (src_const, mode);
5260 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5263 sets[i].src_const = src_const;
5264 sets[i].src_const_elt = src_const_elt;
5266 /* If the constant and our source are both in the table, mark them as
5267 equivalent. Otherwise, if a constant is in the table but the source
5268 isn't, set ELT to it. */
5269 if (src_const_elt && elt
5270 && src_const_elt->first_same_value != elt->first_same_value)
5271 merge_equiv_classes (elt, src_const_elt);
5272 else if (src_const_elt && elt == 0)
5273 elt = src_const_elt;
5275 /* See if there is a register linearly related to a constant
5276 equivalent of SRC. */
5277 if (src_const
5278 && (GET_CODE (src_const) == CONST
5279 || (src_const_elt && src_const_elt->related_value != 0)))
5281 src_related = use_related_value (src_const, src_const_elt);
5282 if (src_related)
5284 struct table_elt *src_related_elt
5285 = lookup (src_related, HASH (src_related, mode), mode);
5286 if (src_related_elt && elt)
5288 if (elt->first_same_value
5289 != src_related_elt->first_same_value)
5290 /* This can occur when we previously saw a CONST
5291 involving a SYMBOL_REF and then see the SYMBOL_REF
5292 twice. Merge the involved classes. */
5293 merge_equiv_classes (elt, src_related_elt);
5295 src_related = 0;
5296 src_related_elt = 0;
5298 else if (src_related_elt && elt == 0)
5299 elt = src_related_elt;
5303 /* See if we have a CONST_INT that is already in a register in a
5304 wider mode. */
5306 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5307 && GET_MODE_CLASS (mode) == MODE_INT
5308 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5310 enum machine_mode wider_mode;
5312 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5313 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5314 && src_related == 0;
5315 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5317 struct table_elt *const_elt
5318 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5320 if (const_elt == 0)
5321 continue;
5323 for (const_elt = const_elt->first_same_value;
5324 const_elt; const_elt = const_elt->next_same_value)
5325 if (REG_P (const_elt->exp))
5327 src_related = gen_lowpart (mode,
5328 const_elt->exp);
5329 break;
5334 /* Another possibility is that we have an AND with a constant in
5335 a mode narrower than a word. If so, it might have been generated
5336 as part of an "if" which would narrow the AND. If we already
5337 have done the AND in a wider mode, we can use a SUBREG of that
5338 value. */
5340 if (flag_expensive_optimizations && ! src_related
5341 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5342 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5344 enum machine_mode tmode;
5345 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5347 for (tmode = GET_MODE_WIDER_MODE (mode);
5348 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5349 tmode = GET_MODE_WIDER_MODE (tmode))
5351 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5352 struct table_elt *larger_elt;
5354 if (inner)
5356 PUT_MODE (new_and, tmode);
5357 XEXP (new_and, 0) = inner;
5358 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5359 if (larger_elt == 0)
5360 continue;
5362 for (larger_elt = larger_elt->first_same_value;
5363 larger_elt; larger_elt = larger_elt->next_same_value)
5364 if (REG_P (larger_elt->exp))
5366 src_related
5367 = gen_lowpart (mode, larger_elt->exp);
5368 break;
5371 if (src_related)
5372 break;
5377 #ifdef LOAD_EXTEND_OP
5378 /* See if a MEM has already been loaded with a widening operation;
5379 if it has, we can use a subreg of that. Many CISC machines
5380 also have such operations, but this is only likely to be
5381 beneficial on these machines. */
5383 if (flag_expensive_optimizations && src_related == 0
5384 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5385 && GET_MODE_CLASS (mode) == MODE_INT
5386 && MEM_P (src) && ! do_not_record
5387 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5389 struct rtx_def memory_extend_buf;
5390 rtx memory_extend_rtx = &memory_extend_buf;
5391 enum machine_mode tmode;
5393 /* Set what we are trying to extend and the operation it might
5394 have been extended with. */
5395 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5396 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5397 XEXP (memory_extend_rtx, 0) = src;
5399 for (tmode = GET_MODE_WIDER_MODE (mode);
5400 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5401 tmode = GET_MODE_WIDER_MODE (tmode))
5403 struct table_elt *larger_elt;
5405 PUT_MODE (memory_extend_rtx, tmode);
5406 larger_elt = lookup (memory_extend_rtx,
5407 HASH (memory_extend_rtx, tmode), tmode);
5408 if (larger_elt == 0)
5409 continue;
5411 for (larger_elt = larger_elt->first_same_value;
5412 larger_elt; larger_elt = larger_elt->next_same_value)
5413 if (REG_P (larger_elt->exp))
5415 src_related = gen_lowpart (mode,
5416 larger_elt->exp);
5417 break;
5420 if (src_related)
5421 break;
5424 #endif /* LOAD_EXTEND_OP */
5426 if (src == src_folded)
5427 src_folded = 0;
5429 /* At this point, ELT, if nonzero, points to a class of expressions
5430 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5431 and SRC_RELATED, if nonzero, each contain additional equivalent
5432 expressions. Prune these latter expressions by deleting expressions
5433 already in the equivalence class.
5435 Check for an equivalent identical to the destination. If found,
5436 this is the preferred equivalent since it will likely lead to
5437 elimination of the insn. Indicate this by placing it in
5438 `src_related'. */
5440 if (elt)
5441 elt = elt->first_same_value;
5442 for (p = elt; p; p = p->next_same_value)
5444 enum rtx_code code = GET_CODE (p->exp);
5446 /* If the expression is not valid, ignore it. Then we do not
5447 have to check for validity below. In most cases, we can use
5448 `rtx_equal_p', since canonicalization has already been done. */
5449 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5450 continue;
5452 /* Also skip paradoxical subregs, unless that's what we're
5453 looking for. */
5454 if (code == SUBREG
5455 && (GET_MODE_SIZE (GET_MODE (p->exp))
5456 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5457 && ! (src != 0
5458 && GET_CODE (src) == SUBREG
5459 && GET_MODE (src) == GET_MODE (p->exp)
5460 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5461 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5462 continue;
5464 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5465 src = 0;
5466 else if (src_folded && GET_CODE (src_folded) == code
5467 && rtx_equal_p (src_folded, p->exp))
5468 src_folded = 0;
5469 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5470 && rtx_equal_p (src_eqv_here, p->exp))
5471 src_eqv_here = 0;
5472 else if (src_related && GET_CODE (src_related) == code
5473 && rtx_equal_p (src_related, p->exp))
5474 src_related = 0;
5476 /* This is the same as the destination of the insns, we want
5477 to prefer it. Copy it to src_related. The code below will
5478 then give it a negative cost. */
5479 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5480 src_related = dest;
5483 /* Find the cheapest valid equivalent, trying all the available
5484 possibilities. Prefer items not in the hash table to ones
5485 that are when they are equal cost. Note that we can never
5486 worsen an insn as the current contents will also succeed.
5487 If we find an equivalent identical to the destination, use it as best,
5488 since this insn will probably be eliminated in that case. */
5489 if (src)
5491 if (rtx_equal_p (src, dest))
5492 src_cost = src_regcost = -1;
5493 else
5495 src_cost = COST (src);
5496 src_regcost = approx_reg_cost (src);
5500 if (src_eqv_here)
5502 if (rtx_equal_p (src_eqv_here, dest))
5503 src_eqv_cost = src_eqv_regcost = -1;
5504 else
5506 src_eqv_cost = COST (src_eqv_here);
5507 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5511 if (src_folded)
5513 if (rtx_equal_p (src_folded, dest))
5514 src_folded_cost = src_folded_regcost = -1;
5515 else
5517 src_folded_cost = COST (src_folded);
5518 src_folded_regcost = approx_reg_cost (src_folded);
5522 if (src_related)
5524 if (rtx_equal_p (src_related, dest))
5525 src_related_cost = src_related_regcost = -1;
5526 else
5528 src_related_cost = COST (src_related);
5529 src_related_regcost = approx_reg_cost (src_related);
5533 /* If this was an indirect jump insn, a known label will really be
5534 cheaper even though it looks more expensive. */
5535 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5536 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5538 /* Terminate loop when replacement made. This must terminate since
5539 the current contents will be tested and will always be valid. */
5540 while (1)
5542 rtx trial;
5544 /* Skip invalid entries. */
5545 while (elt && !REG_P (elt->exp)
5546 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5547 elt = elt->next_same_value;
5549 /* A paradoxical subreg would be bad here: it'll be the right
5550 size, but later may be adjusted so that the upper bits aren't
5551 what we want. So reject it. */
5552 if (elt != 0
5553 && GET_CODE (elt->exp) == SUBREG
5554 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5555 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5556 /* It is okay, though, if the rtx we're trying to match
5557 will ignore any of the bits we can't predict. */
5558 && ! (src != 0
5559 && GET_CODE (src) == SUBREG
5560 && GET_MODE (src) == GET_MODE (elt->exp)
5561 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5562 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5564 elt = elt->next_same_value;
5565 continue;
5568 if (elt)
5570 src_elt_cost = elt->cost;
5571 src_elt_regcost = elt->regcost;
5574 /* Find cheapest and skip it for the next time. For items
5575 of equal cost, use this order:
5576 src_folded, src, src_eqv, src_related and hash table entry. */
5577 if (src_folded
5578 && preferable (src_folded_cost, src_folded_regcost,
5579 src_cost, src_regcost) <= 0
5580 && preferable (src_folded_cost, src_folded_regcost,
5581 src_eqv_cost, src_eqv_regcost) <= 0
5582 && preferable (src_folded_cost, src_folded_regcost,
5583 src_related_cost, src_related_regcost) <= 0
5584 && preferable (src_folded_cost, src_folded_regcost,
5585 src_elt_cost, src_elt_regcost) <= 0)
5587 trial = src_folded, src_folded_cost = MAX_COST;
5588 if (src_folded_force_flag)
5590 rtx forced = force_const_mem (mode, trial);
5591 if (forced)
5592 trial = forced;
5595 else if (src
5596 && preferable (src_cost, src_regcost,
5597 src_eqv_cost, src_eqv_regcost) <= 0
5598 && preferable (src_cost, src_regcost,
5599 src_related_cost, src_related_regcost) <= 0
5600 && preferable (src_cost, src_regcost,
5601 src_elt_cost, src_elt_regcost) <= 0)
5602 trial = src, src_cost = MAX_COST;
5603 else if (src_eqv_here
5604 && preferable (src_eqv_cost, src_eqv_regcost,
5605 src_related_cost, src_related_regcost) <= 0
5606 && preferable (src_eqv_cost, src_eqv_regcost,
5607 src_elt_cost, src_elt_regcost) <= 0)
5608 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5609 else if (src_related
5610 && preferable (src_related_cost, src_related_regcost,
5611 src_elt_cost, src_elt_regcost) <= 0)
5612 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5613 else
5615 trial = copy_rtx (elt->exp);
5616 elt = elt->next_same_value;
5617 src_elt_cost = MAX_COST;
5620 /* We don't normally have an insn matching (set (pc) (pc)), so
5621 check for this separately here. We will delete such an
5622 insn below.
5624 For other cases such as a table jump or conditional jump
5625 where we know the ultimate target, go ahead and replace the
5626 operand. While that may not make a valid insn, we will
5627 reemit the jump below (and also insert any necessary
5628 barriers). */
5629 if (n_sets == 1 && dest == pc_rtx
5630 && (trial == pc_rtx
5631 || (GET_CODE (trial) == LABEL_REF
5632 && ! condjump_p (insn))))
5634 /* Don't substitute non-local labels, this confuses CFG. */
5635 if (GET_CODE (trial) == LABEL_REF
5636 && LABEL_REF_NONLOCAL_P (trial))
5637 continue;
5639 SET_SRC (sets[i].rtl) = trial;
5640 cse_jumps_altered = 1;
5641 break;
5644 /* Reject certain invalid forms of CONST that we create. */
5645 else if (CONSTANT_P (trial)
5646 && GET_CODE (trial) == CONST
5647 /* Reject cases that will cause decode_rtx_const to
5648 die. On the alpha when simplifying a switch, we
5649 get (const (truncate (minus (label_ref)
5650 (label_ref)))). */
5651 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5652 /* Likewise on IA-64, except without the
5653 truncate. */
5654 || (GET_CODE (XEXP (trial, 0)) == MINUS
5655 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5656 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5657 /* Do nothing for this case. */
5660 /* Look for a substitution that makes a valid insn. */
5661 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5663 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5665 /* If we just made a substitution inside a libcall, then we
5666 need to make the same substitution in any notes attached
5667 to the RETVAL insn. */
5668 if (libcall_insn
5669 && (REG_P (sets[i].orig_src)
5670 || GET_CODE (sets[i].orig_src) == SUBREG
5671 || MEM_P (sets[i].orig_src)))
5673 rtx note = find_reg_equal_equiv_note (libcall_insn);
5674 if (note != 0)
5675 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5676 sets[i].orig_src,
5677 copy_rtx (new));
5680 /* The result of apply_change_group can be ignored; see
5681 canon_reg. */
5683 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5684 apply_change_group ();
5685 break;
5688 /* If we previously found constant pool entries for
5689 constants and this is a constant, try making a
5690 pool entry. Put it in src_folded unless we already have done
5691 this since that is where it likely came from. */
5693 else if (constant_pool_entries_cost
5694 && CONSTANT_P (trial)
5695 && (src_folded == 0
5696 || (!MEM_P (src_folded)
5697 && ! src_folded_force_flag))
5698 && GET_MODE_CLASS (mode) != MODE_CC
5699 && mode != VOIDmode)
5701 src_folded_force_flag = 1;
5702 src_folded = trial;
5703 src_folded_cost = constant_pool_entries_cost;
5704 src_folded_regcost = constant_pool_entries_regcost;
5708 src = SET_SRC (sets[i].rtl);
5710 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5711 However, there is an important exception: If both are registers
5712 that are not the head of their equivalence class, replace SET_SRC
5713 with the head of the class. If we do not do this, we will have
5714 both registers live over a portion of the basic block. This way,
5715 their lifetimes will likely abut instead of overlapping. */
5716 if (REG_P (dest)
5717 && REGNO_QTY_VALID_P (REGNO (dest)))
5719 int dest_q = REG_QTY (REGNO (dest));
5720 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5722 if (dest_ent->mode == GET_MODE (dest)
5723 && dest_ent->first_reg != REGNO (dest)
5724 && REG_P (src) && REGNO (src) == REGNO (dest)
5725 /* Don't do this if the original insn had a hard reg as
5726 SET_SRC or SET_DEST. */
5727 && (!REG_P (sets[i].src)
5728 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5729 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5730 /* We can't call canon_reg here because it won't do anything if
5731 SRC is a hard register. */
5733 int src_q = REG_QTY (REGNO (src));
5734 struct qty_table_elem *src_ent = &qty_table[src_q];
5735 int first = src_ent->first_reg;
5736 rtx new_src
5737 = (first >= FIRST_PSEUDO_REGISTER
5738 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5740 /* We must use validate-change even for this, because this
5741 might be a special no-op instruction, suitable only to
5742 tag notes onto. */
5743 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5745 src = new_src;
5746 /* If we had a constant that is cheaper than what we are now
5747 setting SRC to, use that constant. We ignored it when we
5748 thought we could make this into a no-op. */
5749 if (src_const && COST (src_const) < COST (src)
5750 && validate_change (insn, &SET_SRC (sets[i].rtl),
5751 src_const, 0))
5752 src = src_const;
5757 /* If we made a change, recompute SRC values. */
5758 if (src != sets[i].src)
5760 cse_altered = 1;
5761 do_not_record = 0;
5762 hash_arg_in_memory = 0;
5763 sets[i].src = src;
5764 sets[i].src_hash = HASH (src, mode);
5765 sets[i].src_volatile = do_not_record;
5766 sets[i].src_in_memory = hash_arg_in_memory;
5767 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5770 /* If this is a single SET, we are setting a register, and we have an
5771 equivalent constant, we want to add a REG_NOTE. We don't want
5772 to write a REG_EQUAL note for a constant pseudo since verifying that
5773 that pseudo hasn't been eliminated is a pain. Such a note also
5774 won't help anything.
5776 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5777 which can be created for a reference to a compile time computable
5778 entry in a jump table. */
5780 if (n_sets == 1 && src_const && REG_P (dest)
5781 && !REG_P (src_const)
5782 && ! (GET_CODE (src_const) == CONST
5783 && GET_CODE (XEXP (src_const, 0)) == MINUS
5784 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5785 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5787 /* We only want a REG_EQUAL note if src_const != src. */
5788 if (! rtx_equal_p (src, src_const))
5790 /* Make sure that the rtx is not shared. */
5791 src_const = copy_rtx (src_const);
5793 /* Record the actual constant value in a REG_EQUAL note,
5794 making a new one if one does not already exist. */
5795 set_unique_reg_note (insn, REG_EQUAL, src_const);
5799 /* Now deal with the destination. */
5800 do_not_record = 0;
5802 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5803 while (GET_CODE (dest) == SUBREG
5804 || GET_CODE (dest) == ZERO_EXTRACT
5805 || GET_CODE (dest) == STRICT_LOW_PART)
5806 dest = XEXP (dest, 0);
5808 sets[i].inner_dest = dest;
5810 if (MEM_P (dest))
5812 #ifdef PUSH_ROUNDING
5813 /* Stack pushes invalidate the stack pointer. */
5814 rtx addr = XEXP (dest, 0);
5815 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5816 && XEXP (addr, 0) == stack_pointer_rtx)
5817 invalidate (stack_pointer_rtx, VOIDmode);
5818 #endif
5819 dest = fold_rtx (dest, insn);
5822 /* Compute the hash code of the destination now,
5823 before the effects of this instruction are recorded,
5824 since the register values used in the address computation
5825 are those before this instruction. */
5826 sets[i].dest_hash = HASH (dest, mode);
5828 /* Don't enter a bit-field in the hash table
5829 because the value in it after the store
5830 may not equal what was stored, due to truncation. */
5832 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5834 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5836 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5837 && GET_CODE (width) == CONST_INT
5838 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5839 && ! (INTVAL (src_const)
5840 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5841 /* Exception: if the value is constant,
5842 and it won't be truncated, record it. */
5844 else
5846 /* This is chosen so that the destination will be invalidated
5847 but no new value will be recorded.
5848 We must invalidate because sometimes constant
5849 values can be recorded for bitfields. */
5850 sets[i].src_elt = 0;
5851 sets[i].src_volatile = 1;
5852 src_eqv = 0;
5853 src_eqv_elt = 0;
5857 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5858 the insn. */
5859 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5861 /* One less use of the label this insn used to jump to. */
5862 delete_insn (insn);
5863 cse_jumps_altered = 1;
5864 /* No more processing for this set. */
5865 sets[i].rtl = 0;
5868 /* If this SET is now setting PC to a label, we know it used to
5869 be a conditional or computed branch. */
5870 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5871 && !LABEL_REF_NONLOCAL_P (src))
5873 /* Now emit a BARRIER after the unconditional jump. */
5874 if (NEXT_INSN (insn) == 0
5875 || !BARRIER_P (NEXT_INSN (insn)))
5876 emit_barrier_after (insn);
5878 /* We reemit the jump in as many cases as possible just in
5879 case the form of an unconditional jump is significantly
5880 different than a computed jump or conditional jump.
5882 If this insn has multiple sets, then reemitting the
5883 jump is nontrivial. So instead we just force rerecognition
5884 and hope for the best. */
5885 if (n_sets == 1)
5887 rtx new, note;
5889 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5890 JUMP_LABEL (new) = XEXP (src, 0);
5891 LABEL_NUSES (XEXP (src, 0))++;
5893 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5894 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5895 if (note)
5897 XEXP (note, 1) = NULL_RTX;
5898 REG_NOTES (new) = note;
5901 delete_insn (insn);
5902 insn = new;
5904 /* Now emit a BARRIER after the unconditional jump. */
5905 if (NEXT_INSN (insn) == 0
5906 || !BARRIER_P (NEXT_INSN (insn)))
5907 emit_barrier_after (insn);
5909 else
5910 INSN_CODE (insn) = -1;
5912 /* Do not bother deleting any unreachable code,
5913 let jump/flow do that. */
5915 cse_jumps_altered = 1;
5916 sets[i].rtl = 0;
5919 /* If destination is volatile, invalidate it and then do no further
5920 processing for this assignment. */
5922 else if (do_not_record)
5924 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5925 invalidate (dest, VOIDmode);
5926 else if (MEM_P (dest))
5927 invalidate (dest, VOIDmode);
5928 else if (GET_CODE (dest) == STRICT_LOW_PART
5929 || GET_CODE (dest) == ZERO_EXTRACT)
5930 invalidate (XEXP (dest, 0), GET_MODE (dest));
5931 sets[i].rtl = 0;
5934 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5935 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5937 #ifdef HAVE_cc0
5938 /* If setting CC0, record what it was set to, or a constant, if it
5939 is equivalent to a constant. If it is being set to a floating-point
5940 value, make a COMPARE with the appropriate constant of 0. If we
5941 don't do this, later code can interpret this as a test against
5942 const0_rtx, which can cause problems if we try to put it into an
5943 insn as a floating-point operand. */
5944 if (dest == cc0_rtx)
5946 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5947 this_insn_cc0_mode = mode;
5948 if (FLOAT_MODE_P (mode))
5949 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5950 CONST0_RTX (mode));
5952 #endif
5955 /* Now enter all non-volatile source expressions in the hash table
5956 if they are not already present.
5957 Record their equivalence classes in src_elt.
5958 This way we can insert the corresponding destinations into
5959 the same classes even if the actual sources are no longer in them
5960 (having been invalidated). */
5962 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5963 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5965 struct table_elt *elt;
5966 struct table_elt *classp = sets[0].src_elt;
5967 rtx dest = SET_DEST (sets[0].rtl);
5968 enum machine_mode eqvmode = GET_MODE (dest);
5970 if (GET_CODE (dest) == STRICT_LOW_PART)
5972 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5973 classp = 0;
5975 if (insert_regs (src_eqv, classp, 0))
5977 rehash_using_reg (src_eqv);
5978 src_eqv_hash = HASH (src_eqv, eqvmode);
5980 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5981 elt->in_memory = src_eqv_in_memory;
5982 src_eqv_elt = elt;
5984 /* Check to see if src_eqv_elt is the same as a set source which
5985 does not yet have an elt, and if so set the elt of the set source
5986 to src_eqv_elt. */
5987 for (i = 0; i < n_sets; i++)
5988 if (sets[i].rtl && sets[i].src_elt == 0
5989 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5990 sets[i].src_elt = src_eqv_elt;
5993 for (i = 0; i < n_sets; i++)
5994 if (sets[i].rtl && ! sets[i].src_volatile
5995 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5997 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5999 /* REG_EQUAL in setting a STRICT_LOW_PART
6000 gives an equivalent for the entire destination register,
6001 not just for the subreg being stored in now.
6002 This is a more interesting equivalence, so we arrange later
6003 to treat the entire reg as the destination. */
6004 sets[i].src_elt = src_eqv_elt;
6005 sets[i].src_hash = src_eqv_hash;
6007 else
6009 /* Insert source and constant equivalent into hash table, if not
6010 already present. */
6011 struct table_elt *classp = src_eqv_elt;
6012 rtx src = sets[i].src;
6013 rtx dest = SET_DEST (sets[i].rtl);
6014 enum machine_mode mode
6015 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
6017 /* It's possible that we have a source value known to be
6018 constant but don't have a REG_EQUAL note on the insn.
6019 Lack of a note will mean src_eqv_elt will be NULL. This
6020 can happen where we've generated a SUBREG to access a
6021 CONST_INT that is already in a register in a wider mode.
6022 Ensure that the source expression is put in the proper
6023 constant class. */
6024 if (!classp)
6025 classp = sets[i].src_const_elt;
6027 if (sets[i].src_elt == 0)
6029 /* Don't put a hard register source into the table if this is
6030 the last insn of a libcall. In this case, we only need
6031 to put src_eqv_elt in src_elt. */
6032 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6034 struct table_elt *elt;
6036 /* Note that these insert_regs calls cannot remove
6037 any of the src_elt's, because they would have failed to
6038 match if not still valid. */
6039 if (insert_regs (src, classp, 0))
6041 rehash_using_reg (src);
6042 sets[i].src_hash = HASH (src, mode);
6044 elt = insert (src, classp, sets[i].src_hash, mode);
6045 elt->in_memory = sets[i].src_in_memory;
6046 sets[i].src_elt = classp = elt;
6048 else
6049 sets[i].src_elt = classp;
6051 if (sets[i].src_const && sets[i].src_const_elt == 0
6052 && src != sets[i].src_const
6053 && ! rtx_equal_p (sets[i].src_const, src))
6054 sets[i].src_elt = insert (sets[i].src_const, classp,
6055 sets[i].src_const_hash, mode);
6058 else if (sets[i].src_elt == 0)
6059 /* If we did not insert the source into the hash table (e.g., it was
6060 volatile), note the equivalence class for the REG_EQUAL value, if any,
6061 so that the destination goes into that class. */
6062 sets[i].src_elt = src_eqv_elt;
6064 /* Record destination addresses in the hash table. This allows us to
6065 check if they are invalidated by other sets. */
6066 for (i = 0; i < n_sets; i++)
6068 if (sets[i].rtl)
6070 rtx x = sets[i].inner_dest;
6071 struct table_elt *elt;
6072 enum machine_mode mode;
6073 unsigned hash;
6075 if (MEM_P (x))
6077 x = XEXP (x, 0);
6078 mode = GET_MODE (x);
6079 hash = HASH (x, mode);
6080 elt = lookup (x, hash, mode);
6081 if (!elt)
6083 if (insert_regs (x, NULL, 0))
6085 rtx dest = SET_DEST (sets[i].rtl);
6087 rehash_using_reg (x);
6088 hash = HASH (x, mode);
6089 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6091 elt = insert (x, NULL, hash, mode);
6094 sets[i].dest_addr_elt = elt;
6096 else
6097 sets[i].dest_addr_elt = NULL;
6101 invalidate_from_clobbers (x);
6103 /* Some registers are invalidated by subroutine calls. Memory is
6104 invalidated by non-constant calls. */
6106 if (CALL_P (insn))
6108 if (! CONST_OR_PURE_CALL_P (insn))
6109 invalidate_memory ();
6110 invalidate_for_call ();
6113 /* Now invalidate everything set by this instruction.
6114 If a SUBREG or other funny destination is being set,
6115 sets[i].rtl is still nonzero, so here we invalidate the reg
6116 a part of which is being set. */
6118 for (i = 0; i < n_sets; i++)
6119 if (sets[i].rtl)
6121 /* We can't use the inner dest, because the mode associated with
6122 a ZERO_EXTRACT is significant. */
6123 rtx dest = SET_DEST (sets[i].rtl);
6125 /* Needed for registers to remove the register from its
6126 previous quantity's chain.
6127 Needed for memory if this is a nonvarying address, unless
6128 we have just done an invalidate_memory that covers even those. */
6129 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6130 invalidate (dest, VOIDmode);
6131 else if (MEM_P (dest))
6132 invalidate (dest, VOIDmode);
6133 else if (GET_CODE (dest) == STRICT_LOW_PART
6134 || GET_CODE (dest) == ZERO_EXTRACT)
6135 invalidate (XEXP (dest, 0), GET_MODE (dest));
6138 /* A volatile ASM invalidates everything. */
6139 if (NONJUMP_INSN_P (insn)
6140 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
6141 && MEM_VOLATILE_P (PATTERN (insn)))
6142 flush_hash_table ();
6144 /* Make sure registers mentioned in destinations
6145 are safe for use in an expression to be inserted.
6146 This removes from the hash table
6147 any invalid entry that refers to one of these registers.
6149 We don't care about the return value from mention_regs because
6150 we are going to hash the SET_DEST values unconditionally. */
6152 for (i = 0; i < n_sets; i++)
6154 if (sets[i].rtl)
6156 rtx x = SET_DEST (sets[i].rtl);
6158 if (!REG_P (x))
6159 mention_regs (x);
6160 else
6162 /* We used to rely on all references to a register becoming
6163 inaccessible when a register changes to a new quantity,
6164 since that changes the hash code. However, that is not
6165 safe, since after HASH_SIZE new quantities we get a
6166 hash 'collision' of a register with its own invalid
6167 entries. And since SUBREGs have been changed not to
6168 change their hash code with the hash code of the register,
6169 it wouldn't work any longer at all. So we have to check
6170 for any invalid references lying around now.
6171 This code is similar to the REG case in mention_regs,
6172 but it knows that reg_tick has been incremented, and
6173 it leaves reg_in_table as -1 . */
6174 unsigned int regno = REGNO (x);
6175 unsigned int endregno
6176 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
6177 : hard_regno_nregs[regno][GET_MODE (x)]);
6178 unsigned int i;
6180 for (i = regno; i < endregno; i++)
6182 if (REG_IN_TABLE (i) >= 0)
6184 remove_invalid_refs (i);
6185 REG_IN_TABLE (i) = -1;
6192 /* We may have just removed some of the src_elt's from the hash table.
6193 So replace each one with the current head of the same class.
6194 Also check if destination addresses have been removed. */
6196 for (i = 0; i < n_sets; i++)
6197 if (sets[i].rtl)
6199 if (sets[i].dest_addr_elt
6200 && sets[i].dest_addr_elt->first_same_value == 0)
6202 /* The elt was removed, which means this destination is not
6203 valid after this instruction. */
6204 sets[i].rtl = NULL_RTX;
6206 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6207 /* If elt was removed, find current head of same class,
6208 or 0 if nothing remains of that class. */
6210 struct table_elt *elt = sets[i].src_elt;
6212 while (elt && elt->prev_same_value)
6213 elt = elt->prev_same_value;
6215 while (elt && elt->first_same_value == 0)
6216 elt = elt->next_same_value;
6217 sets[i].src_elt = elt ? elt->first_same_value : 0;
6221 /* Now insert the destinations into their equivalence classes. */
6223 for (i = 0; i < n_sets; i++)
6224 if (sets[i].rtl)
6226 rtx dest = SET_DEST (sets[i].rtl);
6227 struct table_elt *elt;
6229 /* Don't record value if we are not supposed to risk allocating
6230 floating-point values in registers that might be wider than
6231 memory. */
6232 if ((flag_float_store
6233 && MEM_P (dest)
6234 && FLOAT_MODE_P (GET_MODE (dest)))
6235 /* Don't record BLKmode values, because we don't know the
6236 size of it, and can't be sure that other BLKmode values
6237 have the same or smaller size. */
6238 || GET_MODE (dest) == BLKmode
6239 /* Don't record values of destinations set inside a libcall block
6240 since we might delete the libcall. Things should have been set
6241 up so we won't want to reuse such a value, but we play it safe
6242 here. */
6243 || libcall_insn
6244 /* If we didn't put a REG_EQUAL value or a source into the hash
6245 table, there is no point is recording DEST. */
6246 || sets[i].src_elt == 0
6247 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6248 or SIGN_EXTEND, don't record DEST since it can cause
6249 some tracking to be wrong.
6251 ??? Think about this more later. */
6252 || (GET_CODE (dest) == SUBREG
6253 && (GET_MODE_SIZE (GET_MODE (dest))
6254 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6255 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6256 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6257 continue;
6259 /* STRICT_LOW_PART isn't part of the value BEING set,
6260 and neither is the SUBREG inside it.
6261 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6262 if (GET_CODE (dest) == STRICT_LOW_PART)
6263 dest = SUBREG_REG (XEXP (dest, 0));
6265 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6266 /* Registers must also be inserted into chains for quantities. */
6267 if (insert_regs (dest, sets[i].src_elt, 1))
6269 /* If `insert_regs' changes something, the hash code must be
6270 recalculated. */
6271 rehash_using_reg (dest);
6272 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6275 elt = insert (dest, sets[i].src_elt,
6276 sets[i].dest_hash, GET_MODE (dest));
6278 elt->in_memory = (MEM_P (sets[i].inner_dest)
6279 && !MEM_READONLY_P (sets[i].inner_dest));
6281 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6282 narrower than M2, and both M1 and M2 are the same number of words,
6283 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6284 make that equivalence as well.
6286 However, BAR may have equivalences for which gen_lowpart
6287 will produce a simpler value than gen_lowpart applied to
6288 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6289 BAR's equivalences. If we don't get a simplified form, make
6290 the SUBREG. It will not be used in an equivalence, but will
6291 cause two similar assignments to be detected.
6293 Note the loop below will find SUBREG_REG (DEST) since we have
6294 already entered SRC and DEST of the SET in the table. */
6296 if (GET_CODE (dest) == SUBREG
6297 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6298 / UNITS_PER_WORD)
6299 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6300 && (GET_MODE_SIZE (GET_MODE (dest))
6301 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6302 && sets[i].src_elt != 0)
6304 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6305 struct table_elt *elt, *classp = 0;
6307 for (elt = sets[i].src_elt->first_same_value; elt;
6308 elt = elt->next_same_value)
6310 rtx new_src = 0;
6311 unsigned src_hash;
6312 struct table_elt *src_elt;
6313 int byte = 0;
6315 /* Ignore invalid entries. */
6316 if (!REG_P (elt->exp)
6317 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6318 continue;
6320 /* We may have already been playing subreg games. If the
6321 mode is already correct for the destination, use it. */
6322 if (GET_MODE (elt->exp) == new_mode)
6323 new_src = elt->exp;
6324 else
6326 /* Calculate big endian correction for the SUBREG_BYTE.
6327 We have already checked that M1 (GET_MODE (dest))
6328 is not narrower than M2 (new_mode). */
6329 if (BYTES_BIG_ENDIAN)
6330 byte = (GET_MODE_SIZE (GET_MODE (dest))
6331 - GET_MODE_SIZE (new_mode));
6333 new_src = simplify_gen_subreg (new_mode, elt->exp,
6334 GET_MODE (dest), byte);
6337 /* The call to simplify_gen_subreg fails if the value
6338 is VOIDmode, yet we can't do any simplification, e.g.
6339 for EXPR_LISTs denoting function call results.
6340 It is invalid to construct a SUBREG with a VOIDmode
6341 SUBREG_REG, hence a zero new_src means we can't do
6342 this substitution. */
6343 if (! new_src)
6344 continue;
6346 src_hash = HASH (new_src, new_mode);
6347 src_elt = lookup (new_src, src_hash, new_mode);
6349 /* Put the new source in the hash table is if isn't
6350 already. */
6351 if (src_elt == 0)
6353 if (insert_regs (new_src, classp, 0))
6355 rehash_using_reg (new_src);
6356 src_hash = HASH (new_src, new_mode);
6358 src_elt = insert (new_src, classp, src_hash, new_mode);
6359 src_elt->in_memory = elt->in_memory;
6361 else if (classp && classp != src_elt->first_same_value)
6362 /* Show that two things that we've seen before are
6363 actually the same. */
6364 merge_equiv_classes (src_elt, classp);
6366 classp = src_elt->first_same_value;
6367 /* Ignore invalid entries. */
6368 while (classp
6369 && !REG_P (classp->exp)
6370 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6371 classp = classp->next_same_value;
6376 /* Special handling for (set REG0 REG1) where REG0 is the
6377 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6378 be used in the sequel, so (if easily done) change this insn to
6379 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6380 that computed their value. Then REG1 will become a dead store
6381 and won't cloud the situation for later optimizations.
6383 Do not make this change if REG1 is a hard register, because it will
6384 then be used in the sequel and we may be changing a two-operand insn
6385 into a three-operand insn.
6387 Also do not do this if we are operating on a copy of INSN.
6389 Also don't do this if INSN ends a libcall; this would cause an unrelated
6390 register to be set in the middle of a libcall, and we then get bad code
6391 if the libcall is deleted. */
6393 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6394 && NEXT_INSN (PREV_INSN (insn)) == insn
6395 && REG_P (SET_SRC (sets[0].rtl))
6396 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6397 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6399 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6400 struct qty_table_elem *src_ent = &qty_table[src_q];
6402 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6403 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6405 rtx prev = insn;
6406 /* Scan for the previous nonnote insn, but stop at a basic
6407 block boundary. */
6410 prev = PREV_INSN (prev);
6412 while (prev && NOTE_P (prev)
6413 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6415 /* Do not swap the registers around if the previous instruction
6416 attaches a REG_EQUIV note to REG1.
6418 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6419 from the pseudo that originally shadowed an incoming argument
6420 to another register. Some uses of REG_EQUIV might rely on it
6421 being attached to REG1 rather than REG2.
6423 This section previously turned the REG_EQUIV into a REG_EQUAL
6424 note. We cannot do that because REG_EQUIV may provide an
6425 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6427 if (prev != 0 && NONJUMP_INSN_P (prev)
6428 && GET_CODE (PATTERN (prev)) == SET
6429 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6430 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6432 rtx dest = SET_DEST (sets[0].rtl);
6433 rtx src = SET_SRC (sets[0].rtl);
6434 rtx note;
6436 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6437 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6438 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6439 apply_change_group ();
6441 /* If INSN has a REG_EQUAL note, and this note mentions
6442 REG0, then we must delete it, because the value in
6443 REG0 has changed. If the note's value is REG1, we must
6444 also delete it because that is now this insn's dest. */
6445 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6446 if (note != 0
6447 && (reg_mentioned_p (dest, XEXP (note, 0))
6448 || rtx_equal_p (src, XEXP (note, 0))))
6449 remove_note (insn, note);
6454 /* If this is a conditional jump insn, record any known equivalences due to
6455 the condition being tested. */
6457 if (JUMP_P (insn)
6458 && n_sets == 1 && GET_CODE (x) == SET
6459 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6460 record_jump_equiv (insn, 0);
6462 #ifdef HAVE_cc0
6463 /* If the previous insn set CC0 and this insn no longer references CC0,
6464 delete the previous insn. Here we use the fact that nothing expects CC0
6465 to be valid over an insn, which is true until the final pass. */
6466 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6467 && (tem = single_set (prev_insn)) != 0
6468 && SET_DEST (tem) == cc0_rtx
6469 && ! reg_mentioned_p (cc0_rtx, x))
6470 delete_insn (prev_insn);
6472 prev_insn_cc0 = this_insn_cc0;
6473 prev_insn_cc0_mode = this_insn_cc0_mode;
6474 prev_insn = insn;
6475 #endif
6478 /* Remove from the hash table all expressions that reference memory. */
6480 static void
6481 invalidate_memory (void)
6483 int i;
6484 struct table_elt *p, *next;
6486 for (i = 0; i < HASH_SIZE; i++)
6487 for (p = table[i]; p; p = next)
6489 next = p->next_same_hash;
6490 if (p->in_memory)
6491 remove_from_table (p, i);
6495 /* If ADDR is an address that implicitly affects the stack pointer, return
6496 1 and update the register tables to show the effect. Else, return 0. */
6498 static int
6499 addr_affects_sp_p (rtx addr)
6501 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6502 && REG_P (XEXP (addr, 0))
6503 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6505 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6507 REG_TICK (STACK_POINTER_REGNUM)++;
6508 /* Is it possible to use a subreg of SP? */
6509 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6512 /* This should be *very* rare. */
6513 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6514 invalidate (stack_pointer_rtx, VOIDmode);
6516 return 1;
6519 return 0;
6522 /* Perform invalidation on the basis of everything about an insn
6523 except for invalidating the actual places that are SET in it.
6524 This includes the places CLOBBERed, and anything that might
6525 alias with something that is SET or CLOBBERed.
6527 X is the pattern of the insn. */
6529 static void
6530 invalidate_from_clobbers (rtx x)
6532 if (GET_CODE (x) == CLOBBER)
6534 rtx ref = XEXP (x, 0);
6535 if (ref)
6537 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6538 || MEM_P (ref))
6539 invalidate (ref, VOIDmode);
6540 else if (GET_CODE (ref) == STRICT_LOW_PART
6541 || GET_CODE (ref) == ZERO_EXTRACT)
6542 invalidate (XEXP (ref, 0), GET_MODE (ref));
6545 else if (GET_CODE (x) == PARALLEL)
6547 int i;
6548 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6550 rtx y = XVECEXP (x, 0, i);
6551 if (GET_CODE (y) == CLOBBER)
6553 rtx ref = XEXP (y, 0);
6554 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6555 || MEM_P (ref))
6556 invalidate (ref, VOIDmode);
6557 else if (GET_CODE (ref) == STRICT_LOW_PART
6558 || GET_CODE (ref) == ZERO_EXTRACT)
6559 invalidate (XEXP (ref, 0), GET_MODE (ref));
6565 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6566 and replace any registers in them with either an equivalent constant
6567 or the canonical form of the register. If we are inside an address,
6568 only do this if the address remains valid.
6570 OBJECT is 0 except when within a MEM in which case it is the MEM.
6572 Return the replacement for X. */
6574 static rtx
6575 cse_process_notes (rtx x, rtx object)
6577 enum rtx_code code = GET_CODE (x);
6578 const char *fmt = GET_RTX_FORMAT (code);
6579 int i;
6581 switch (code)
6583 case CONST_INT:
6584 case CONST:
6585 case SYMBOL_REF:
6586 case LABEL_REF:
6587 case CONST_DOUBLE:
6588 case CONST_VECTOR:
6589 case PC:
6590 case CC0:
6591 case LO_SUM:
6592 return x;
6594 case MEM:
6595 validate_change (x, &XEXP (x, 0),
6596 cse_process_notes (XEXP (x, 0), x), 0);
6597 return x;
6599 case EXPR_LIST:
6600 case INSN_LIST:
6601 if (REG_NOTE_KIND (x) == REG_EQUAL)
6602 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6603 if (XEXP (x, 1))
6604 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6605 return x;
6607 case SIGN_EXTEND:
6608 case ZERO_EXTEND:
6609 case SUBREG:
6611 rtx new = cse_process_notes (XEXP (x, 0), object);
6612 /* We don't substitute VOIDmode constants into these rtx,
6613 since they would impede folding. */
6614 if (GET_MODE (new) != VOIDmode)
6615 validate_change (object, &XEXP (x, 0), new, 0);
6616 return x;
6619 case REG:
6620 i = REG_QTY (REGNO (x));
6622 /* Return a constant or a constant register. */
6623 if (REGNO_QTY_VALID_P (REGNO (x)))
6625 struct qty_table_elem *ent = &qty_table[i];
6627 if (ent->const_rtx != NULL_RTX
6628 && (CONSTANT_P (ent->const_rtx)
6629 || REG_P (ent->const_rtx)))
6631 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6632 if (new)
6633 return new;
6637 /* Otherwise, canonicalize this register. */
6638 return canon_reg (x, NULL_RTX);
6640 default:
6641 break;
6644 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6645 if (fmt[i] == 'e')
6646 validate_change (object, &XEXP (x, i),
6647 cse_process_notes (XEXP (x, i), object), 0);
6649 return x;
6652 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6653 since they are done elsewhere. This function is called via note_stores. */
6655 static void
6656 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6658 enum rtx_code code = GET_CODE (dest);
6660 if (code == MEM
6661 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6662 /* There are times when an address can appear varying and be a PLUS
6663 during this scan when it would be a fixed address were we to know
6664 the proper equivalences. So invalidate all memory if there is
6665 a BLKmode or nonscalar memory reference or a reference to a
6666 variable address. */
6667 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6668 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6670 invalidate_memory ();
6671 return;
6674 if (GET_CODE (set) == CLOBBER
6675 || CC0_P (dest)
6676 || dest == pc_rtx)
6677 return;
6679 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6680 invalidate (XEXP (dest, 0), GET_MODE (dest));
6681 else if (code == REG || code == SUBREG || code == MEM)
6682 invalidate (dest, VOIDmode);
6685 /* Invalidate all insns from START up to the end of the function or the
6686 next label. This called when we wish to CSE around a block that is
6687 conditionally executed. */
6689 static void
6690 invalidate_skipped_block (rtx start)
6692 rtx insn;
6694 for (insn = start; insn && !LABEL_P (insn);
6695 insn = NEXT_INSN (insn))
6697 if (! INSN_P (insn))
6698 continue;
6700 if (CALL_P (insn))
6702 if (! CONST_OR_PURE_CALL_P (insn))
6703 invalidate_memory ();
6704 invalidate_for_call ();
6707 invalidate_from_clobbers (PATTERN (insn));
6708 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6712 /* Find the end of INSN's basic block and return its range,
6713 the total number of SETs in all the insns of the block, the last insn of the
6714 block, and the branch path.
6716 The branch path indicates which branches should be followed. If a nonzero
6717 path size is specified, the block should be rescanned and a different set
6718 of branches will be taken. The branch path is only used if
6719 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6721 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6722 used to describe the block. It is filled in with the information about
6723 the current block. The incoming structure's branch path, if any, is used
6724 to construct the output branch path. */
6726 static void
6727 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6728 int follow_jumps, int skip_blocks)
6730 rtx p = insn, q;
6731 int nsets = 0;
6732 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6733 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6734 int path_size = data->path_size;
6735 int path_entry = 0;
6736 int i;
6738 /* Update the previous branch path, if any. If the last branch was
6739 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6740 If it was previously PATH_NOT_TAKEN,
6741 shorten the path by one and look at the previous branch. We know that
6742 at least one branch must have been taken if PATH_SIZE is nonzero. */
6743 while (path_size > 0)
6745 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6747 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6748 break;
6750 else
6751 path_size--;
6754 /* If the first instruction is marked with QImode, that means we've
6755 already processed this block. Our caller will look at DATA->LAST
6756 to figure out where to go next. We want to return the next block
6757 in the instruction stream, not some branched-to block somewhere
6758 else. We accomplish this by pretending our called forbid us to
6759 follow jumps, or skip blocks. */
6760 if (GET_MODE (insn) == QImode)
6761 follow_jumps = skip_blocks = 0;
6763 /* Scan to end of this basic block. */
6764 while (p && !LABEL_P (p))
6766 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6767 the regs restored by the longjmp come from
6768 a later time than the setjmp. */
6769 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6770 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6771 break;
6773 /* A PARALLEL can have lots of SETs in it,
6774 especially if it is really an ASM_OPERANDS. */
6775 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6776 nsets += XVECLEN (PATTERN (p), 0);
6777 else if (!NOTE_P (p))
6778 nsets += 1;
6780 /* Ignore insns made by CSE; they cannot affect the boundaries of
6781 the basic block. */
6783 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6784 high_cuid = INSN_CUID (p);
6785 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6786 low_cuid = INSN_CUID (p);
6788 /* See if this insn is in our branch path. If it is and we are to
6789 take it, do so. */
6790 if (path_entry < path_size && data->path[path_entry].branch == p)
6792 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6793 p = JUMP_LABEL (p);
6795 /* Point to next entry in path, if any. */
6796 path_entry++;
6799 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6800 was specified, we haven't reached our maximum path length, there are
6801 insns following the target of the jump, this is the only use of the
6802 jump label, and the target label is preceded by a BARRIER.
6804 Alternatively, we can follow the jump if it branches around a
6805 block of code and there are no other branches into the block.
6806 In this case invalidate_skipped_block will be called to invalidate any
6807 registers set in the block when following the jump. */
6809 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6810 && JUMP_P (p)
6811 && GET_CODE (PATTERN (p)) == SET
6812 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6813 && JUMP_LABEL (p) != 0
6814 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6815 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6817 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6818 if ((!NOTE_P (q)
6819 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6820 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6821 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6822 break;
6824 /* If we ran into a BARRIER, this code is an extension of the
6825 basic block when the branch is taken. */
6826 if (follow_jumps && q != 0 && BARRIER_P (q))
6828 /* Don't allow ourself to keep walking around an
6829 always-executed loop. */
6830 if (next_real_insn (q) == next)
6832 p = NEXT_INSN (p);
6833 continue;
6836 /* Similarly, don't put a branch in our path more than once. */
6837 for (i = 0; i < path_entry; i++)
6838 if (data->path[i].branch == p)
6839 break;
6841 if (i != path_entry)
6842 break;
6844 data->path[path_entry].branch = p;
6845 data->path[path_entry++].status = PATH_TAKEN;
6847 /* This branch now ends our path. It was possible that we
6848 didn't see this branch the last time around (when the
6849 insn in front of the target was a JUMP_INSN that was
6850 turned into a no-op). */
6851 path_size = path_entry;
6853 p = JUMP_LABEL (p);
6854 /* Mark block so we won't scan it again later. */
6855 PUT_MODE (NEXT_INSN (p), QImode);
6857 /* Detect a branch around a block of code. */
6858 else if (skip_blocks && q != 0 && !LABEL_P (q))
6860 rtx tmp;
6862 if (next_real_insn (q) == next)
6864 p = NEXT_INSN (p);
6865 continue;
6868 for (i = 0; i < path_entry; i++)
6869 if (data->path[i].branch == p)
6870 break;
6872 if (i != path_entry)
6873 break;
6875 /* This is no_labels_between_p (p, q) with an added check for
6876 reaching the end of a function (in case Q precedes P). */
6877 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6878 if (LABEL_P (tmp))
6879 break;
6881 if (tmp == q)
6883 data->path[path_entry].branch = p;
6884 data->path[path_entry++].status = PATH_AROUND;
6886 path_size = path_entry;
6888 p = JUMP_LABEL (p);
6889 /* Mark block so we won't scan it again later. */
6890 PUT_MODE (NEXT_INSN (p), QImode);
6894 p = NEXT_INSN (p);
6897 data->low_cuid = low_cuid;
6898 data->high_cuid = high_cuid;
6899 data->nsets = nsets;
6900 data->last = p;
6902 /* If all jumps in the path are not taken, set our path length to zero
6903 so a rescan won't be done. */
6904 for (i = path_size - 1; i >= 0; i--)
6905 if (data->path[i].status != PATH_NOT_TAKEN)
6906 break;
6908 if (i == -1)
6909 data->path_size = 0;
6910 else
6911 data->path_size = path_size;
6913 /* End the current branch path. */
6914 data->path[path_size].branch = 0;
6917 /* Perform cse on the instructions of a function.
6918 F is the first instruction.
6919 NREGS is one plus the highest pseudo-reg number used in the instruction.
6921 Returns 1 if jump_optimize should be redone due to simplifications
6922 in conditional jump instructions. */
6925 cse_main (rtx f, int nregs)
6927 struct cse_basic_block_data val;
6928 rtx insn = f;
6929 int i;
6931 init_cse_reg_info (nregs);
6933 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6935 cse_jumps_altered = 0;
6936 recorded_label_ref = 0;
6937 constant_pool_entries_cost = 0;
6938 constant_pool_entries_regcost = 0;
6939 val.path_size = 0;
6940 rtl_hooks = cse_rtl_hooks;
6942 init_recog ();
6943 init_alias_analysis ();
6945 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6947 /* Find the largest uid. */
6949 max_uid = get_max_uid ();
6950 uid_cuid = XCNEWVEC (int, max_uid + 1);
6952 /* Compute the mapping from uids to cuids.
6953 CUIDs are numbers assigned to insns, like uids,
6954 except that cuids increase monotonically through the code.
6955 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6956 between two insns is not affected by -g. */
6958 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6960 if (!NOTE_P (insn)
6961 || NOTE_LINE_NUMBER (insn) < 0)
6962 INSN_CUID (insn) = ++i;
6963 else
6964 /* Give a line number note the same cuid as preceding insn. */
6965 INSN_CUID (insn) = i;
6968 /* Loop over basic blocks.
6969 Compute the maximum number of qty's needed for each basic block
6970 (which is 2 for each SET). */
6971 insn = f;
6972 while (insn)
6974 cse_altered = 0;
6975 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6976 flag_cse_skip_blocks);
6978 /* If this basic block was already processed or has no sets, skip it. */
6979 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6981 PUT_MODE (insn, VOIDmode);
6982 insn = (val.last ? NEXT_INSN (val.last) : 0);
6983 val.path_size = 0;
6984 continue;
6987 cse_basic_block_start = val.low_cuid;
6988 cse_basic_block_end = val.high_cuid;
6989 max_qty = val.nsets * 2;
6991 if (dump_file)
6992 fprintf (dump_file, ";; Processing block from %d to %d, %d sets.\n",
6993 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6994 val.nsets);
6996 /* Make MAX_QTY bigger to give us room to optimize
6997 past the end of this basic block, if that should prove useful. */
6998 if (max_qty < 500)
6999 max_qty = 500;
7001 /* If this basic block is being extended by following certain jumps,
7002 (see `cse_end_of_basic_block'), we reprocess the code from the start.
7003 Otherwise, we start after this basic block. */
7004 if (val.path_size > 0)
7005 cse_basic_block (insn, val.last, val.path);
7006 else
7008 int old_cse_jumps_altered = cse_jumps_altered;
7009 rtx temp;
7011 /* When cse changes a conditional jump to an unconditional
7012 jump, we want to reprocess the block, since it will give
7013 us a new branch path to investigate. */
7014 cse_jumps_altered = 0;
7015 temp = cse_basic_block (insn, val.last, val.path);
7016 if (cse_jumps_altered == 0
7017 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7018 insn = temp;
7020 cse_jumps_altered |= old_cse_jumps_altered;
7023 if (cse_altered)
7024 ggc_collect ();
7026 #ifdef USE_C_ALLOCA
7027 alloca (0);
7028 #endif
7031 /* Clean up. */
7032 end_alias_analysis ();
7033 free (uid_cuid);
7034 free (reg_eqv_table);
7035 free (val.path);
7036 rtl_hooks = general_rtl_hooks;
7038 return cse_jumps_altered || recorded_label_ref;
7041 /* Process a single basic block. FROM and TO and the limits of the basic
7042 block. NEXT_BRANCH points to the branch path when following jumps or
7043 a null path when not following jumps. */
7045 static rtx
7046 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
7048 rtx insn;
7049 int to_usage = 0;
7050 rtx libcall_insn = NULL_RTX;
7051 int num_insns = 0;
7052 int no_conflict = 0;
7054 /* Allocate the space needed by qty_table. */
7055 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
7057 new_basic_block ();
7059 /* TO might be a label. If so, protect it from being deleted. */
7060 if (to != 0 && LABEL_P (to))
7061 ++LABEL_NUSES (to);
7063 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7065 enum rtx_code code = GET_CODE (insn);
7067 /* If we have processed 1,000 insns, flush the hash table to
7068 avoid extreme quadratic behavior. We must not include NOTEs
7069 in the count since there may be more of them when generating
7070 debugging information. If we clear the table at different
7071 times, code generated with -g -O might be different than code
7072 generated with -O but not -g.
7074 ??? This is a real kludge and needs to be done some other way.
7075 Perhaps for 2.9. */
7076 if (code != NOTE && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
7078 flush_hash_table ();
7079 num_insns = 0;
7082 /* See if this is a branch that is part of the path. If so, and it is
7083 to be taken, do so. */
7084 if (next_branch->branch == insn)
7086 enum taken status = next_branch++->status;
7087 if (status != PATH_NOT_TAKEN)
7089 if (status == PATH_TAKEN)
7090 record_jump_equiv (insn, 1);
7091 else
7092 invalidate_skipped_block (NEXT_INSN (insn));
7094 /* Set the last insn as the jump insn; it doesn't affect cc0.
7095 Then follow this branch. */
7096 #ifdef HAVE_cc0
7097 prev_insn_cc0 = 0;
7098 prev_insn = insn;
7099 #endif
7100 insn = JUMP_LABEL (insn);
7101 continue;
7105 if (GET_MODE (insn) == QImode)
7106 PUT_MODE (insn, VOIDmode);
7108 if (GET_RTX_CLASS (code) == RTX_INSN)
7110 rtx p;
7112 /* Process notes first so we have all notes in canonical forms when
7113 looking for duplicate operations. */
7115 if (REG_NOTES (insn))
7116 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7118 /* Track when we are inside in LIBCALL block. Inside such a block,
7119 we do not want to record destinations. The last insn of a
7120 LIBCALL block is not considered to be part of the block, since
7121 its destination is the result of the block and hence should be
7122 recorded. */
7124 if (REG_NOTES (insn) != 0)
7126 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7127 libcall_insn = XEXP (p, 0);
7128 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7130 /* Keep libcall_insn for the last SET insn of a no-conflict
7131 block to prevent changing the destination. */
7132 if (! no_conflict)
7133 libcall_insn = 0;
7134 else
7135 no_conflict = -1;
7137 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
7138 no_conflict = 1;
7141 cse_insn (insn, libcall_insn);
7143 if (no_conflict == -1)
7145 libcall_insn = 0;
7146 no_conflict = 0;
7149 /* If we haven't already found an insn where we added a LABEL_REF,
7150 check this one. */
7151 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
7152 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
7153 (void *) insn))
7154 recorded_label_ref = 1;
7157 /* If INSN is now an unconditional jump, skip to the end of our
7158 basic block by pretending that we just did the last insn in the
7159 basic block. If we are jumping to the end of our block, show
7160 that we can have one usage of TO. */
7162 if (any_uncondjump_p (insn))
7164 if (to == 0)
7166 free (qty_table);
7167 return 0;
7170 if (JUMP_LABEL (insn) == to)
7171 to_usage = 1;
7173 /* Maybe TO was deleted because the jump is unconditional.
7174 If so, there is nothing left in this basic block. */
7175 /* ??? Perhaps it would be smarter to set TO
7176 to whatever follows this insn,
7177 and pretend the basic block had always ended here. */
7178 if (INSN_DELETED_P (to))
7179 break;
7181 insn = PREV_INSN (to);
7184 /* See if it is ok to keep on going past the label
7185 which used to end our basic block. Remember that we incremented
7186 the count of that label, so we decrement it here. If we made
7187 a jump unconditional, TO_USAGE will be one; in that case, we don't
7188 want to count the use in that jump. */
7190 if (to != 0 && NEXT_INSN (insn) == to
7191 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7193 struct cse_basic_block_data val;
7194 rtx prev;
7196 insn = NEXT_INSN (to);
7198 /* If TO was the last insn in the function, we are done. */
7199 if (insn == 0)
7201 free (qty_table);
7202 return 0;
7205 /* If TO was preceded by a BARRIER we are done with this block
7206 because it has no continuation. */
7207 prev = prev_nonnote_insn (to);
7208 if (prev && BARRIER_P (prev))
7210 free (qty_table);
7211 return insn;
7214 /* Find the end of the following block. Note that we won't be
7215 following branches in this case. */
7216 to_usage = 0;
7217 val.path_size = 0;
7218 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7219 cse_end_of_basic_block (insn, &val, 0, 0);
7220 free (val.path);
7222 /* If the tables we allocated have enough space left
7223 to handle all the SETs in the next basic block,
7224 continue through it. Otherwise, return,
7225 and that block will be scanned individually. */
7226 if (val.nsets * 2 + next_qty > max_qty)
7227 break;
7229 cse_basic_block_start = val.low_cuid;
7230 cse_basic_block_end = val.high_cuid;
7231 to = val.last;
7233 /* Prevent TO from being deleted if it is a label. */
7234 if (to != 0 && LABEL_P (to))
7235 ++LABEL_NUSES (to);
7237 /* Back up so we process the first insn in the extension. */
7238 insn = PREV_INSN (insn);
7242 gcc_assert (next_qty <= max_qty);
7244 free (qty_table);
7246 return to ? NEXT_INSN (to) : 0;
7249 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7250 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7252 static int
7253 check_for_label_ref (rtx *rtl, void *data)
7255 rtx insn = (rtx) data;
7257 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7258 we must rerun jump since it needs to place the note. If this is a
7259 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7260 since no REG_LABEL will be added. */
7261 return (GET_CODE (*rtl) == LABEL_REF
7262 && ! LABEL_REF_NONLOCAL_P (*rtl)
7263 && LABEL_P (XEXP (*rtl, 0))
7264 && INSN_UID (XEXP (*rtl, 0)) != 0
7265 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7268 /* Count the number of times registers are used (not set) in X.
7269 COUNTS is an array in which we accumulate the count, INCR is how much
7270 we count each register usage.
7272 Don't count a usage of DEST, which is the SET_DEST of a SET which
7273 contains X in its SET_SRC. This is because such a SET does not
7274 modify the liveness of DEST.
7275 DEST is set to pc_rtx for a trapping insn, which means that we must count
7276 uses of a SET_DEST regardless because the insn can't be deleted here. */
7278 static void
7279 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
7281 enum rtx_code code;
7282 rtx note;
7283 const char *fmt;
7284 int i, j;
7286 if (x == 0)
7287 return;
7289 switch (code = GET_CODE (x))
7291 case REG:
7292 if (x != dest)
7293 counts[REGNO (x)] += incr;
7294 return;
7296 case PC:
7297 case CC0:
7298 case CONST:
7299 case CONST_INT:
7300 case CONST_DOUBLE:
7301 case CONST_VECTOR:
7302 case SYMBOL_REF:
7303 case LABEL_REF:
7304 return;
7306 case CLOBBER:
7307 /* If we are clobbering a MEM, mark any registers inside the address
7308 as being used. */
7309 if (MEM_P (XEXP (x, 0)))
7310 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7311 return;
7313 case SET:
7314 /* Unless we are setting a REG, count everything in SET_DEST. */
7315 if (!REG_P (SET_DEST (x)))
7316 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7317 count_reg_usage (SET_SRC (x), counts,
7318 dest ? dest : SET_DEST (x),
7319 incr);
7320 return;
7322 case CALL_INSN:
7323 case INSN:
7324 case JUMP_INSN:
7325 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
7326 this fact by setting DEST to pc_rtx. */
7327 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
7328 dest = pc_rtx;
7329 if (code == CALL_INSN)
7330 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
7331 count_reg_usage (PATTERN (x), counts, dest, incr);
7333 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7334 use them. */
7336 note = find_reg_equal_equiv_note (x);
7337 if (note)
7339 rtx eqv = XEXP (note, 0);
7341 if (GET_CODE (eqv) == EXPR_LIST)
7342 /* This REG_EQUAL note describes the result of a function call.
7343 Process all the arguments. */
7346 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
7347 eqv = XEXP (eqv, 1);
7349 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7350 else
7351 count_reg_usage (eqv, counts, dest, incr);
7353 return;
7355 case EXPR_LIST:
7356 if (REG_NOTE_KIND (x) == REG_EQUAL
7357 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7358 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7359 involving registers in the address. */
7360 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7361 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7363 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7364 return;
7366 case ASM_OPERANDS:
7367 /* If the asm is volatile, then this insn cannot be deleted,
7368 and so the inputs *must* be live. */
7369 if (MEM_VOLATILE_P (x))
7370 dest = NULL_RTX;
7371 /* Iterate over just the inputs, not the constraints as well. */
7372 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7373 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
7374 return;
7376 case INSN_LIST:
7377 gcc_unreachable ();
7379 default:
7380 break;
7383 fmt = GET_RTX_FORMAT (code);
7384 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7386 if (fmt[i] == 'e')
7387 count_reg_usage (XEXP (x, i), counts, dest, incr);
7388 else if (fmt[i] == 'E')
7389 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7390 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7394 /* Return true if set is live. */
7395 static bool
7396 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7397 int *counts)
7399 #ifdef HAVE_cc0
7400 rtx tem;
7401 #endif
7403 if (set_noop_p (set))
7406 #ifdef HAVE_cc0
7407 else if (GET_CODE (SET_DEST (set)) == CC0
7408 && !side_effects_p (SET_SRC (set))
7409 && ((tem = next_nonnote_insn (insn)) == 0
7410 || !INSN_P (tem)
7411 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7412 return false;
7413 #endif
7414 else if (!REG_P (SET_DEST (set))
7415 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7416 || counts[REGNO (SET_DEST (set))] != 0
7417 || side_effects_p (SET_SRC (set)))
7418 return true;
7419 return false;
7422 /* Return true if insn is live. */
7424 static bool
7425 insn_live_p (rtx insn, int *counts)
7427 int i;
7428 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7429 return true;
7430 else if (GET_CODE (PATTERN (insn)) == SET)
7431 return set_live_p (PATTERN (insn), insn, counts);
7432 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7434 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7436 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7438 if (GET_CODE (elt) == SET)
7440 if (set_live_p (elt, insn, counts))
7441 return true;
7443 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7444 return true;
7446 return false;
7448 else
7449 return true;
7452 /* Return true if libcall is dead as a whole. */
7454 static bool
7455 dead_libcall_p (rtx insn, int *counts)
7457 rtx note, set, new;
7459 /* See if there's a REG_EQUAL note on this insn and try to
7460 replace the source with the REG_EQUAL expression.
7462 We assume that insns with REG_RETVALs can only be reg->reg
7463 copies at this point. */
7464 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7465 if (!note)
7466 return false;
7468 set = single_set (insn);
7469 if (!set)
7470 return false;
7472 new = simplify_rtx (XEXP (note, 0));
7473 if (!new)
7474 new = XEXP (note, 0);
7476 /* While changing insn, we must update the counts accordingly. */
7477 count_reg_usage (insn, counts, NULL_RTX, -1);
7479 if (validate_change (insn, &SET_SRC (set), new, 0))
7481 count_reg_usage (insn, counts, NULL_RTX, 1);
7482 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7483 remove_note (insn, note);
7484 return true;
7487 if (CONSTANT_P (new))
7489 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7490 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7492 count_reg_usage (insn, counts, NULL_RTX, 1);
7493 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7494 remove_note (insn, note);
7495 return true;
7499 count_reg_usage (insn, counts, NULL_RTX, 1);
7500 return false;
7503 /* Scan all the insns and delete any that are dead; i.e., they store a register
7504 that is never used or they copy a register to itself.
7506 This is used to remove insns made obviously dead by cse, loop or other
7507 optimizations. It improves the heuristics in loop since it won't try to
7508 move dead invariants out of loops or make givs for dead quantities. The
7509 remaining passes of the compilation are also sped up. */
7512 delete_trivially_dead_insns (rtx insns, int nreg)
7514 int *counts;
7515 rtx insn, prev;
7516 int in_libcall = 0, dead_libcall = 0;
7517 int ndead = 0;
7519 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7520 /* First count the number of times each register is used. */
7521 counts = XCNEWVEC (int, nreg);
7522 for (insn = insns; insn; insn = NEXT_INSN (insn))
7523 if (INSN_P (insn))
7524 count_reg_usage (insn, counts, NULL_RTX, 1);
7526 /* Go from the last insn to the first and delete insns that only set unused
7527 registers or copy a register to itself. As we delete an insn, remove
7528 usage counts for registers it uses.
7530 The first jump optimization pass may leave a real insn as the last
7531 insn in the function. We must not skip that insn or we may end
7532 up deleting code that is not really dead. */
7533 for (insn = get_last_insn (); insn; insn = prev)
7535 int live_insn = 0;
7537 prev = PREV_INSN (insn);
7538 if (!INSN_P (insn))
7539 continue;
7541 /* Don't delete any insns that are part of a libcall block unless
7542 we can delete the whole libcall block.
7544 Flow or loop might get confused if we did that. Remember
7545 that we are scanning backwards. */
7546 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7548 in_libcall = 1;
7549 live_insn = 1;
7550 dead_libcall = dead_libcall_p (insn, counts);
7552 else if (in_libcall)
7553 live_insn = ! dead_libcall;
7554 else
7555 live_insn = insn_live_p (insn, counts);
7557 /* If this is a dead insn, delete it and show registers in it aren't
7558 being used. */
7560 if (! live_insn)
7562 count_reg_usage (insn, counts, NULL_RTX, -1);
7563 delete_insn_and_edges (insn);
7564 ndead++;
7567 if (in_libcall && find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7569 in_libcall = 0;
7570 dead_libcall = 0;
7574 if (dump_file && ndead)
7575 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7576 ndead);
7577 /* Clean up. */
7578 free (counts);
7579 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7580 return ndead;
7583 /* This function is called via for_each_rtx. The argument, NEWREG, is
7584 a condition code register with the desired mode. If we are looking
7585 at the same register in a different mode, replace it with
7586 NEWREG. */
7588 static int
7589 cse_change_cc_mode (rtx *loc, void *data)
7591 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7593 if (*loc
7594 && REG_P (*loc)
7595 && REGNO (*loc) == REGNO (args->newreg)
7596 && GET_MODE (*loc) != GET_MODE (args->newreg))
7598 validate_change (args->insn, loc, args->newreg, 1);
7600 return -1;
7602 return 0;
7605 /* Change the mode of any reference to the register REGNO (NEWREG) to
7606 GET_MODE (NEWREG) in INSN. */
7608 static void
7609 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7611 struct change_cc_mode_args args;
7612 int success;
7614 if (!INSN_P (insn))
7615 return;
7617 args.insn = insn;
7618 args.newreg = newreg;
7620 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7621 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7623 /* If the following assertion was triggered, there is most probably
7624 something wrong with the cc_modes_compatible back end function.
7625 CC modes only can be considered compatible if the insn - with the mode
7626 replaced by any of the compatible modes - can still be recognized. */
7627 success = apply_change_group ();
7628 gcc_assert (success);
7631 /* Change the mode of any reference to the register REGNO (NEWREG) to
7632 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7633 any instruction which modifies NEWREG. */
7635 static void
7636 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7638 rtx insn;
7640 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7642 if (! INSN_P (insn))
7643 continue;
7645 if (reg_set_p (newreg, insn))
7646 return;
7648 cse_change_cc_mode_insn (insn, newreg);
7652 /* BB is a basic block which finishes with CC_REG as a condition code
7653 register which is set to CC_SRC. Look through the successors of BB
7654 to find blocks which have a single predecessor (i.e., this one),
7655 and look through those blocks for an assignment to CC_REG which is
7656 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7657 permitted to change the mode of CC_SRC to a compatible mode. This
7658 returns VOIDmode if no equivalent assignments were found.
7659 Otherwise it returns the mode which CC_SRC should wind up with.
7661 The main complexity in this function is handling the mode issues.
7662 We may have more than one duplicate which we can eliminate, and we
7663 try to find a mode which will work for multiple duplicates. */
7665 static enum machine_mode
7666 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7668 bool found_equiv;
7669 enum machine_mode mode;
7670 unsigned int insn_count;
7671 edge e;
7672 rtx insns[2];
7673 enum machine_mode modes[2];
7674 rtx last_insns[2];
7675 unsigned int i;
7676 rtx newreg;
7677 edge_iterator ei;
7679 /* We expect to have two successors. Look at both before picking
7680 the final mode for the comparison. If we have more successors
7681 (i.e., some sort of table jump, although that seems unlikely),
7682 then we require all beyond the first two to use the same
7683 mode. */
7685 found_equiv = false;
7686 mode = GET_MODE (cc_src);
7687 insn_count = 0;
7688 FOR_EACH_EDGE (e, ei, bb->succs)
7690 rtx insn;
7691 rtx end;
7693 if (e->flags & EDGE_COMPLEX)
7694 continue;
7696 if (EDGE_COUNT (e->dest->preds) != 1
7697 || e->dest == EXIT_BLOCK_PTR)
7698 continue;
7700 end = NEXT_INSN (BB_END (e->dest));
7701 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7703 rtx set;
7705 if (! INSN_P (insn))
7706 continue;
7708 /* If CC_SRC is modified, we have to stop looking for
7709 something which uses it. */
7710 if (modified_in_p (cc_src, insn))
7711 break;
7713 /* Check whether INSN sets CC_REG to CC_SRC. */
7714 set = single_set (insn);
7715 if (set
7716 && REG_P (SET_DEST (set))
7717 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7719 bool found;
7720 enum machine_mode set_mode;
7721 enum machine_mode comp_mode;
7723 found = false;
7724 set_mode = GET_MODE (SET_SRC (set));
7725 comp_mode = set_mode;
7726 if (rtx_equal_p (cc_src, SET_SRC (set)))
7727 found = true;
7728 else if (GET_CODE (cc_src) == COMPARE
7729 && GET_CODE (SET_SRC (set)) == COMPARE
7730 && mode != set_mode
7731 && rtx_equal_p (XEXP (cc_src, 0),
7732 XEXP (SET_SRC (set), 0))
7733 && rtx_equal_p (XEXP (cc_src, 1),
7734 XEXP (SET_SRC (set), 1)))
7737 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7738 if (comp_mode != VOIDmode
7739 && (can_change_mode || comp_mode == mode))
7740 found = true;
7743 if (found)
7745 found_equiv = true;
7746 if (insn_count < ARRAY_SIZE (insns))
7748 insns[insn_count] = insn;
7749 modes[insn_count] = set_mode;
7750 last_insns[insn_count] = end;
7751 ++insn_count;
7753 if (mode != comp_mode)
7755 gcc_assert (can_change_mode);
7756 mode = comp_mode;
7758 /* The modified insn will be re-recognized later. */
7759 PUT_MODE (cc_src, mode);
7762 else
7764 if (set_mode != mode)
7766 /* We found a matching expression in the
7767 wrong mode, but we don't have room to
7768 store it in the array. Punt. This case
7769 should be rare. */
7770 break;
7772 /* INSN sets CC_REG to a value equal to CC_SRC
7773 with the right mode. We can simply delete
7774 it. */
7775 delete_insn (insn);
7778 /* We found an instruction to delete. Keep looking,
7779 in the hopes of finding a three-way jump. */
7780 continue;
7783 /* We found an instruction which sets the condition
7784 code, so don't look any farther. */
7785 break;
7788 /* If INSN sets CC_REG in some other way, don't look any
7789 farther. */
7790 if (reg_set_p (cc_reg, insn))
7791 break;
7794 /* If we fell off the bottom of the block, we can keep looking
7795 through successors. We pass CAN_CHANGE_MODE as false because
7796 we aren't prepared to handle compatibility between the
7797 further blocks and this block. */
7798 if (insn == end)
7800 enum machine_mode submode;
7802 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7803 if (submode != VOIDmode)
7805 gcc_assert (submode == mode);
7806 found_equiv = true;
7807 can_change_mode = false;
7812 if (! found_equiv)
7813 return VOIDmode;
7815 /* Now INSN_COUNT is the number of instructions we found which set
7816 CC_REG to a value equivalent to CC_SRC. The instructions are in
7817 INSNS. The modes used by those instructions are in MODES. */
7819 newreg = NULL_RTX;
7820 for (i = 0; i < insn_count; ++i)
7822 if (modes[i] != mode)
7824 /* We need to change the mode of CC_REG in INSNS[i] and
7825 subsequent instructions. */
7826 if (! newreg)
7828 if (GET_MODE (cc_reg) == mode)
7829 newreg = cc_reg;
7830 else
7831 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7833 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7834 newreg);
7837 delete_insn (insns[i]);
7840 return mode;
7843 /* If we have a fixed condition code register (or two), walk through
7844 the instructions and try to eliminate duplicate assignments. */
7846 static void
7847 cse_condition_code_reg (void)
7849 unsigned int cc_regno_1;
7850 unsigned int cc_regno_2;
7851 rtx cc_reg_1;
7852 rtx cc_reg_2;
7853 basic_block bb;
7855 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7856 return;
7858 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7859 if (cc_regno_2 != INVALID_REGNUM)
7860 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7861 else
7862 cc_reg_2 = NULL_RTX;
7864 FOR_EACH_BB (bb)
7866 rtx last_insn;
7867 rtx cc_reg;
7868 rtx insn;
7869 rtx cc_src_insn;
7870 rtx cc_src;
7871 enum machine_mode mode;
7872 enum machine_mode orig_mode;
7874 /* Look for blocks which end with a conditional jump based on a
7875 condition code register. Then look for the instruction which
7876 sets the condition code register. Then look through the
7877 successor blocks for instructions which set the condition
7878 code register to the same value. There are other possible
7879 uses of the condition code register, but these are by far the
7880 most common and the ones which we are most likely to be able
7881 to optimize. */
7883 last_insn = BB_END (bb);
7884 if (!JUMP_P (last_insn))
7885 continue;
7887 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7888 cc_reg = cc_reg_1;
7889 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7890 cc_reg = cc_reg_2;
7891 else
7892 continue;
7894 cc_src_insn = NULL_RTX;
7895 cc_src = NULL_RTX;
7896 for (insn = PREV_INSN (last_insn);
7897 insn && insn != PREV_INSN (BB_HEAD (bb));
7898 insn = PREV_INSN (insn))
7900 rtx set;
7902 if (! INSN_P (insn))
7903 continue;
7904 set = single_set (insn);
7905 if (set
7906 && REG_P (SET_DEST (set))
7907 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7909 cc_src_insn = insn;
7910 cc_src = SET_SRC (set);
7911 break;
7913 else if (reg_set_p (cc_reg, insn))
7914 break;
7917 if (! cc_src_insn)
7918 continue;
7920 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7921 continue;
7923 /* Now CC_REG is a condition code register used for a
7924 conditional jump at the end of the block, and CC_SRC, in
7925 CC_SRC_INSN, is the value to which that condition code
7926 register is set, and CC_SRC is still meaningful at the end of
7927 the basic block. */
7929 orig_mode = GET_MODE (cc_src);
7930 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7931 if (mode != VOIDmode)
7933 gcc_assert (mode == GET_MODE (cc_src));
7934 if (mode != orig_mode)
7936 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7938 cse_change_cc_mode_insn (cc_src_insn, newreg);
7940 /* Do the same in the following insns that use the
7941 current value of CC_REG within BB. */
7942 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7943 NEXT_INSN (last_insn),
7944 newreg);
7951 /* Perform common subexpression elimination. Nonzero value from
7952 `cse_main' means that jumps were simplified and some code may now
7953 be unreachable, so do jump optimization again. */
7954 static bool
7955 gate_handle_cse (void)
7957 return optimize > 0;
7960 static unsigned int
7961 rest_of_handle_cse (void)
7963 int tem;
7965 if (dump_file)
7966 dump_flow_info (dump_file, dump_flags);
7968 reg_scan (get_insns (), max_reg_num ());
7970 tem = cse_main (get_insns (), max_reg_num ());
7971 if (tem)
7972 rebuild_jump_labels (get_insns ());
7973 if (purge_all_dead_edges ())
7974 delete_unreachable_blocks ();
7976 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7978 /* If we are not running more CSE passes, then we are no longer
7979 expecting CSE to be run. But always rerun it in a cheap mode. */
7980 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7982 if (tem)
7983 delete_dead_jumptables ();
7985 if (tem || optimize > 1)
7986 cleanup_cfg (CLEANUP_EXPENSIVE);
7987 return 0;
7990 struct tree_opt_pass pass_cse =
7992 "cse1", /* name */
7993 gate_handle_cse, /* gate */
7994 rest_of_handle_cse, /* execute */
7995 NULL, /* sub */
7996 NULL, /* next */
7997 0, /* static_pass_number */
7998 TV_CSE, /* tv_id */
7999 0, /* properties_required */
8000 0, /* properties_provided */
8001 0, /* properties_destroyed */
8002 0, /* todo_flags_start */
8003 TODO_dump_func |
8004 TODO_ggc_collect, /* todo_flags_finish */
8005 's' /* letter */
8009 static bool
8010 gate_handle_cse2 (void)
8012 return optimize > 0 && flag_rerun_cse_after_loop;
8015 /* Run second CSE pass after loop optimizations. */
8016 static unsigned int
8017 rest_of_handle_cse2 (void)
8019 int tem;
8021 if (dump_file)
8022 dump_flow_info (dump_file, dump_flags);
8024 tem = cse_main (get_insns (), max_reg_num ());
8026 /* Run a pass to eliminate duplicated assignments to condition code
8027 registers. We have to run this after bypass_jumps, because it
8028 makes it harder for that pass to determine whether a jump can be
8029 bypassed safely. */
8030 cse_condition_code_reg ();
8032 purge_all_dead_edges ();
8033 delete_trivially_dead_insns (get_insns (), max_reg_num ());
8035 if (tem)
8037 timevar_push (TV_JUMP);
8038 rebuild_jump_labels (get_insns ());
8039 delete_dead_jumptables ();
8040 cleanup_cfg (CLEANUP_EXPENSIVE);
8041 timevar_pop (TV_JUMP);
8043 reg_scan (get_insns (), max_reg_num ());
8044 cse_not_expected = 1;
8045 return 0;
8049 struct tree_opt_pass pass_cse2 =
8051 "cse2", /* name */
8052 gate_handle_cse2, /* gate */
8053 rest_of_handle_cse2, /* execute */
8054 NULL, /* sub */
8055 NULL, /* next */
8056 0, /* static_pass_number */
8057 TV_CSE2, /* tv_id */
8058 0, /* properties_required */
8059 0, /* properties_provided */
8060 0, /* properties_destroyed */
8061 0, /* todo_flags_start */
8062 TODO_dump_func |
8063 TODO_ggc_collect, /* todo_flags_finish */
8064 't' /* letter */