2005-12-29 Paul Brook <paul@codesourcery.com>
[official-gcc.git] / gcc / cse.c
blob284f3fd5bea24c9899bbb855728fbee3bed35473
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
20 02110-1301, USA. */
22 #include "config.h"
23 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "rtl.h"
28 #include "tm_p.h"
29 #include "hard-reg-set.h"
30 #include "regs.h"
31 #include "basic-block.h"
32 #include "flags.h"
33 #include "real.h"
34 #include "insn-config.h"
35 #include "recog.h"
36 #include "function.h"
37 #include "expr.h"
38 #include "toplev.h"
39 #include "output.h"
40 #include "ggc.h"
41 #include "timevar.h"
42 #include "except.h"
43 #include "target.h"
44 #include "params.h"
45 #include "rtlhooks-def.h"
46 #include "tree-pass.h"
48 /* The basic idea of common subexpression elimination is to go
49 through the code, keeping a record of expressions that would
50 have the same value at the current scan point, and replacing
51 expressions encountered with the cheapest equivalent expression.
53 It is too complicated to keep track of the different possibilities
54 when control paths merge in this code; so, at each label, we forget all
55 that is known and start fresh. This can be described as processing each
56 extended basic block separately. We have a separate pass to perform
57 global CSE.
59 Note CSE can turn a conditional or computed jump into a nop or
60 an unconditional jump. When this occurs we arrange to run the jump
61 optimizer after CSE to delete the unreachable code.
63 We use two data structures to record the equivalent expressions:
64 a hash table for most expressions, and a vector of "quantity
65 numbers" to record equivalent (pseudo) registers.
67 The use of the special data structure for registers is desirable
68 because it is faster. It is possible because registers references
69 contain a fairly small number, the register number, taken from
70 a contiguously allocated series, and two register references are
71 identical if they have the same number. General expressions
72 do not have any such thing, so the only way to retrieve the
73 information recorded on an expression other than a register
74 is to keep it in a hash table.
76 Registers and "quantity numbers":
78 At the start of each basic block, all of the (hardware and pseudo)
79 registers used in the function are given distinct quantity
80 numbers to indicate their contents. During scan, when the code
81 copies one register into another, we copy the quantity number.
82 When a register is loaded in any other way, we allocate a new
83 quantity number to describe the value generated by this operation.
84 `REG_QTY (N)' records what quantity register N is currently thought
85 of as containing.
87 All real quantity numbers are greater than or equal to zero.
88 If register N has not been assigned a quantity, `REG_QTY (N)' will
89 equal -N - 1, which is always negative.
91 Quantity numbers below zero do not exist and none of the `qty_table'
92 entries should be referenced with a negative index.
94 We also maintain a bidirectional chain of registers for each
95 quantity number. The `qty_table` members `first_reg' and `last_reg',
96 and `reg_eqv_table' members `next' and `prev' hold these chains.
98 The first register in a chain is the one whose lifespan is least local.
99 Among equals, it is the one that was seen first.
100 We replace any equivalent register with that one.
102 If two registers have the same quantity number, it must be true that
103 REG expressions with qty_table `mode' must be in the hash table for both
104 registers and must be in the same class.
106 The converse is not true. Since hard registers may be referenced in
107 any mode, two REG expressions might be equivalent in the hash table
108 but not have the same quantity number if the quantity number of one
109 of the registers is not the same mode as those expressions.
111 Constants and quantity numbers
113 When a quantity has a known constant value, that value is stored
114 in the appropriate qty_table `const_rtx'. This is in addition to
115 putting the constant in the hash table as is usual for non-regs.
117 Whether a reg or a constant is preferred is determined by the configuration
118 macro CONST_COSTS and will often depend on the constant value. In any
119 event, expressions containing constants can be simplified, by fold_rtx.
121 When a quantity has a known nearly constant value (such as an address
122 of a stack slot), that value is stored in the appropriate qty_table
123 `const_rtx'.
125 Integer constants don't have a machine mode. However, cse
126 determines the intended machine mode from the destination
127 of the instruction that moves the constant. The machine mode
128 is recorded in the hash table along with the actual RTL
129 constant expression so that different modes are kept separate.
131 Other expressions:
133 To record known equivalences among expressions in general
134 we use a hash table called `table'. It has a fixed number of buckets
135 that contain chains of `struct table_elt' elements for expressions.
136 These chains connect the elements whose expressions have the same
137 hash codes.
139 Other chains through the same elements connect the elements which
140 currently have equivalent values.
142 Register references in an expression are canonicalized before hashing
143 the expression. This is done using `reg_qty' and qty_table `first_reg'.
144 The hash code of a register reference is computed using the quantity
145 number, not the register number.
147 When the value of an expression changes, it is necessary to remove from the
148 hash table not just that expression but all expressions whose values
149 could be different as a result.
151 1. If the value changing is in memory, except in special cases
152 ANYTHING referring to memory could be changed. That is because
153 nobody knows where a pointer does not point.
154 The function `invalidate_memory' removes what is necessary.
156 The special cases are when the address is constant or is
157 a constant plus a fixed register such as the frame pointer
158 or a static chain pointer. When such addresses are stored in,
159 we can tell exactly which other such addresses must be invalidated
160 due to overlap. `invalidate' does this.
161 All expressions that refer to non-constant
162 memory addresses are also invalidated. `invalidate_memory' does this.
164 2. If the value changing is a register, all expressions
165 containing references to that register, and only those,
166 must be removed.
168 Because searching the entire hash table for expressions that contain
169 a register is very slow, we try to figure out when it isn't necessary.
170 Precisely, this is necessary only when expressions have been
171 entered in the hash table using this register, and then the value has
172 changed, and then another expression wants to be added to refer to
173 the register's new value. This sequence of circumstances is rare
174 within any one basic block.
176 `REG_TICK' and `REG_IN_TABLE', accessors for members of
177 cse_reg_info, are used to detect this case. REG_TICK (i) is
178 incremented whenever a value is stored in register i.
179 REG_IN_TABLE (i) holds -1 if no references to register i have been
180 entered in the table; otherwise, it contains the value REG_TICK (i)
181 had when the references were entered. If we want to enter a
182 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
183 remove old references. Until we want to enter a new entry, the
184 mere fact that the two vectors don't match makes the entries be
185 ignored if anyone tries to match them.
187 Registers themselves are entered in the hash table as well as in
188 the equivalent-register chains. However, `REG_TICK' and
189 `REG_IN_TABLE' do not apply to expressions which are simple
190 register references. These expressions are removed from the table
191 immediately when they become invalid, and this can be done even if
192 we do not immediately search for all the expressions that refer to
193 the register.
195 A CLOBBER rtx in an instruction invalidates its operand for further
196 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
197 invalidates everything that resides in memory.
199 Related expressions:
201 Constant expressions that differ only by an additive integer
202 are called related. When a constant expression is put in
203 the table, the related expression with no constant term
204 is also entered. These are made to point at each other
205 so that it is possible to find out if there exists any
206 register equivalent to an expression related to a given expression. */
208 /* Length of qty_table vector. We know in advance we will not need
209 a quantity number this big. */
211 static int max_qty;
213 /* Next quantity number to be allocated.
214 This is 1 + the largest number needed so far. */
216 static int next_qty;
218 /* Per-qty information tracking.
220 `first_reg' and `last_reg' track the head and tail of the
221 chain of registers which currently contain this quantity.
223 `mode' contains the machine mode of this quantity.
225 `const_rtx' holds the rtx of the constant value of this
226 quantity, if known. A summations of the frame/arg pointer
227 and a constant can also be entered here. When this holds
228 a known value, `const_insn' is the insn which stored the
229 constant value.
231 `comparison_{code,const,qty}' are used to track when a
232 comparison between a quantity and some constant or register has
233 been passed. In such a case, we know the results of the comparison
234 in case we see it again. These members record a comparison that
235 is known to be true. `comparison_code' holds the rtx code of such
236 a comparison, else it is set to UNKNOWN and the other two
237 comparison members are undefined. `comparison_const' holds
238 the constant being compared against, or zero if the comparison
239 is not against a constant. `comparison_qty' holds the quantity
240 being compared against when the result is known. If the comparison
241 is not with a register, `comparison_qty' is -1. */
243 struct qty_table_elem
245 rtx const_rtx;
246 rtx const_insn;
247 rtx comparison_const;
248 int comparison_qty;
249 unsigned int first_reg, last_reg;
250 /* The sizes of these fields should match the sizes of the
251 code and mode fields of struct rtx_def (see rtl.h). */
252 ENUM_BITFIELD(rtx_code) comparison_code : 16;
253 ENUM_BITFIELD(machine_mode) mode : 8;
256 /* The table of all qtys, indexed by qty number. */
257 static struct qty_table_elem *qty_table;
259 /* Structure used to pass arguments via for_each_rtx to function
260 cse_change_cc_mode. */
261 struct change_cc_mode_args
263 rtx insn;
264 rtx newreg;
267 #ifdef HAVE_cc0
268 /* For machines that have a CC0, we do not record its value in the hash
269 table since its use is guaranteed to be the insn immediately following
270 its definition and any other insn is presumed to invalidate it.
272 Instead, we store below the value last assigned to CC0. If it should
273 happen to be a constant, it is stored in preference to the actual
274 assigned value. In case it is a constant, we store the mode in which
275 the constant should be interpreted. */
277 static rtx prev_insn_cc0;
278 static enum machine_mode prev_insn_cc0_mode;
280 /* Previous actual insn. 0 if at first insn of basic block. */
282 static rtx prev_insn;
283 #endif
285 /* Insn being scanned. */
287 static rtx this_insn;
289 /* Index by register number, gives the number of the next (or
290 previous) register in the chain of registers sharing the same
291 value.
293 Or -1 if this register is at the end of the chain.
295 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
297 /* Per-register equivalence chain. */
298 struct reg_eqv_elem
300 int next, prev;
303 /* The table of all register equivalence chains. */
304 static struct reg_eqv_elem *reg_eqv_table;
306 struct cse_reg_info
308 /* The timestamp at which this register is initialized. */
309 unsigned int timestamp;
311 /* The quantity number of the register's current contents. */
312 int reg_qty;
314 /* The number of times the register has been altered in the current
315 basic block. */
316 int reg_tick;
318 /* The REG_TICK value at which rtx's containing this register are
319 valid in the hash table. If this does not equal the current
320 reg_tick value, such expressions existing in the hash table are
321 invalid. */
322 int reg_in_table;
324 /* The SUBREG that was set when REG_TICK was last incremented. Set
325 to -1 if the last store was to the whole register, not a subreg. */
326 unsigned int subreg_ticked;
329 /* A table of cse_reg_info indexed by register numbers. */
330 static struct cse_reg_info *cse_reg_info_table;
332 /* The size of the above table. */
333 static unsigned int cse_reg_info_table_size;
335 /* The index of the first entry that has not been initialized. */
336 static unsigned int cse_reg_info_table_first_uninitialized;
338 /* The timestamp at the beginning of the current run of
339 cse_basic_block. We increment this variable at the beginning of
340 the current run of cse_basic_block. The timestamp field of a
341 cse_reg_info entry matches the value of this variable if and only
342 if the entry has been initialized during the current run of
343 cse_basic_block. */
344 static unsigned int cse_reg_info_timestamp;
346 /* A HARD_REG_SET containing all the hard registers for which there is
347 currently a REG expression in the hash table. Note the difference
348 from the above variables, which indicate if the REG is mentioned in some
349 expression in the table. */
351 static HARD_REG_SET hard_regs_in_table;
353 /* CUID of insn that starts the basic block currently being cse-processed. */
355 static int cse_basic_block_start;
357 /* CUID of insn that ends the basic block currently being cse-processed. */
359 static int cse_basic_block_end;
361 /* Vector mapping INSN_UIDs to cuids.
362 The cuids are like uids but increase monotonically always.
363 We use them to see whether a reg is used outside a given basic block. */
365 static int *uid_cuid;
367 /* Highest UID in UID_CUID. */
368 static int max_uid;
370 /* Get the cuid of an insn. */
372 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
374 /* Nonzero if this pass has made changes, and therefore it's
375 worthwhile to run the garbage collector. */
377 static int cse_altered;
379 /* Nonzero if cse has altered conditional jump insns
380 in such a way that jump optimization should be redone. */
382 static int cse_jumps_altered;
384 /* Nonzero if we put a LABEL_REF into the hash table for an INSN without a
385 REG_LABEL, we have to rerun jump after CSE to put in the note. */
386 static int recorded_label_ref;
388 /* canon_hash stores 1 in do_not_record
389 if it notices a reference to CC0, PC, or some other volatile
390 subexpression. */
392 static int do_not_record;
394 /* canon_hash stores 1 in hash_arg_in_memory
395 if it notices a reference to memory within the expression being hashed. */
397 static int hash_arg_in_memory;
399 /* The hash table contains buckets which are chains of `struct table_elt's,
400 each recording one expression's information.
401 That expression is in the `exp' field.
403 The canon_exp field contains a canonical (from the point of view of
404 alias analysis) version of the `exp' field.
406 Those elements with the same hash code are chained in both directions
407 through the `next_same_hash' and `prev_same_hash' fields.
409 Each set of expressions with equivalent values
410 are on a two-way chain through the `next_same_value'
411 and `prev_same_value' fields, and all point with
412 the `first_same_value' field at the first element in
413 that chain. The chain is in order of increasing cost.
414 Each element's cost value is in its `cost' field.
416 The `in_memory' field is nonzero for elements that
417 involve any reference to memory. These elements are removed
418 whenever a write is done to an unidentified location in memory.
419 To be safe, we assume that a memory address is unidentified unless
420 the address is either a symbol constant or a constant plus
421 the frame pointer or argument pointer.
423 The `related_value' field is used to connect related expressions
424 (that differ by adding an integer).
425 The related expressions are chained in a circular fashion.
426 `related_value' is zero for expressions for which this
427 chain is not useful.
429 The `cost' field stores the cost of this element's expression.
430 The `regcost' field stores the value returned by approx_reg_cost for
431 this element's expression.
433 The `is_const' flag is set if the element is a constant (including
434 a fixed address).
436 The `flag' field is used as a temporary during some search routines.
438 The `mode' field is usually the same as GET_MODE (`exp'), but
439 if `exp' is a CONST_INT and has no machine mode then the `mode'
440 field is the mode it was being used as. Each constant is
441 recorded separately for each mode it is used with. */
443 struct table_elt
445 rtx exp;
446 rtx canon_exp;
447 struct table_elt *next_same_hash;
448 struct table_elt *prev_same_hash;
449 struct table_elt *next_same_value;
450 struct table_elt *prev_same_value;
451 struct table_elt *first_same_value;
452 struct table_elt *related_value;
453 int cost;
454 int regcost;
455 /* The size of this field should match the size
456 of the mode field of struct rtx_def (see rtl.h). */
457 ENUM_BITFIELD(machine_mode) mode : 8;
458 char in_memory;
459 char is_const;
460 char flag;
463 /* We don't want a lot of buckets, because we rarely have very many
464 things stored in the hash table, and a lot of buckets slows
465 down a lot of loops that happen frequently. */
466 #define HASH_SHIFT 5
467 #define HASH_SIZE (1 << HASH_SHIFT)
468 #define HASH_MASK (HASH_SIZE - 1)
470 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
471 register (hard registers may require `do_not_record' to be set). */
473 #define HASH(X, M) \
474 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
475 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
476 : canon_hash (X, M)) & HASH_MASK)
478 /* Like HASH, but without side-effects. */
479 #define SAFE_HASH(X, M) \
480 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
481 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
482 : safe_hash (X, M)) & HASH_MASK)
484 /* Determine whether register number N is considered a fixed register for the
485 purpose of approximating register costs.
486 It is desirable to replace other regs with fixed regs, to reduce need for
487 non-fixed hard regs.
488 A reg wins if it is either the frame pointer or designated as fixed. */
489 #define FIXED_REGNO_P(N) \
490 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
491 || fixed_regs[N] || global_regs[N])
493 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
494 hard registers and pointers into the frame are the cheapest with a cost
495 of 0. Next come pseudos with a cost of one and other hard registers with
496 a cost of 2. Aside from these special cases, call `rtx_cost'. */
498 #define CHEAP_REGNO(N) \
499 (REGNO_PTR_FRAME_P(N) \
500 || (HARD_REGISTER_NUM_P (N) \
501 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
503 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
504 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
506 /* Get the number of times this register has been updated in this
507 basic block. */
509 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
511 /* Get the point at which REG was recorded in the table. */
513 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
515 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
516 SUBREG). */
518 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
520 /* Get the quantity number for REG. */
522 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
524 /* Determine if the quantity number for register X represents a valid index
525 into the qty_table. */
527 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
529 static struct table_elt *table[HASH_SIZE];
531 /* Chain of `struct table_elt's made so far for this function
532 but currently removed from the table. */
534 static struct table_elt *free_element_chain;
536 /* Set to the cost of a constant pool reference if one was found for a
537 symbolic constant. If this was found, it means we should try to
538 convert constants into constant pool entries if they don't fit in
539 the insn. */
541 static int constant_pool_entries_cost;
542 static int constant_pool_entries_regcost;
544 /* This data describes a block that will be processed by cse_basic_block. */
546 struct cse_basic_block_data
548 /* Lowest CUID value of insns in block. */
549 int low_cuid;
550 /* Highest CUID value of insns in block. */
551 int high_cuid;
552 /* Total number of SETs in block. */
553 int nsets;
554 /* Last insn in the block. */
555 rtx last;
556 /* Size of current branch path, if any. */
557 int path_size;
558 /* Current branch path, indicating which branches will be taken. */
559 struct branch_path
561 /* The branch insn. */
562 rtx branch;
563 /* Whether it should be taken or not. AROUND is the same as taken
564 except that it is used when the destination label is not preceded
565 by a BARRIER. */
566 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
567 } *path;
570 static bool fixed_base_plus_p (rtx x);
571 static int notreg_cost (rtx, enum rtx_code);
572 static int approx_reg_cost_1 (rtx *, void *);
573 static int approx_reg_cost (rtx);
574 static int preferable (int, int, int, int);
575 static void new_basic_block (void);
576 static void make_new_qty (unsigned int, enum machine_mode);
577 static void make_regs_eqv (unsigned int, unsigned int);
578 static void delete_reg_equiv (unsigned int);
579 static int mention_regs (rtx);
580 static int insert_regs (rtx, struct table_elt *, int);
581 static void remove_from_table (struct table_elt *, unsigned);
582 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
583 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
584 static rtx lookup_as_function (rtx, enum rtx_code);
585 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
586 enum machine_mode);
587 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
588 static void invalidate (rtx, enum machine_mode);
589 static int cse_rtx_varies_p (rtx, int);
590 static void remove_invalid_refs (unsigned int);
591 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
592 enum machine_mode);
593 static void rehash_using_reg (rtx);
594 static void invalidate_memory (void);
595 static void invalidate_for_call (void);
596 static rtx use_related_value (rtx, struct table_elt *);
598 static inline unsigned canon_hash (rtx, enum machine_mode);
599 static inline unsigned safe_hash (rtx, enum machine_mode);
600 static unsigned hash_rtx_string (const char *);
602 static rtx canon_reg (rtx, rtx);
603 static void find_best_addr (rtx, rtx *, enum machine_mode);
604 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
605 enum machine_mode *,
606 enum machine_mode *);
607 static rtx fold_rtx (rtx, rtx);
608 static rtx equiv_constant (rtx);
609 static void record_jump_equiv (rtx, int);
610 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
611 int);
612 static void cse_insn (rtx, rtx);
613 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
614 int, int);
615 static int addr_affects_sp_p (rtx);
616 static void invalidate_from_clobbers (rtx);
617 static rtx cse_process_notes (rtx, rtx);
618 static void invalidate_skipped_set (rtx, rtx, void *);
619 static void invalidate_skipped_block (rtx);
620 static rtx cse_basic_block (rtx, rtx, struct branch_path *);
621 static void count_reg_usage (rtx, int *, rtx, int);
622 static int check_for_label_ref (rtx *, void *);
623 extern void dump_class (struct table_elt*);
624 static void get_cse_reg_info_1 (unsigned int regno);
625 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
626 static int check_dependence (rtx *, void *);
628 static void flush_hash_table (void);
629 static bool insn_live_p (rtx, int *);
630 static bool set_live_p (rtx, rtx, int *);
631 static bool dead_libcall_p (rtx, int *);
632 static int cse_change_cc_mode (rtx *, void *);
633 static void cse_change_cc_mode_insn (rtx, rtx);
634 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
635 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
638 #undef RTL_HOOKS_GEN_LOWPART
639 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
641 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
643 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
644 virtual regs here because the simplify_*_operation routines are called
645 by integrate.c, which is called before virtual register instantiation. */
647 static bool
648 fixed_base_plus_p (rtx x)
650 switch (GET_CODE (x))
652 case REG:
653 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
654 return true;
655 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
656 return true;
657 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
658 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
659 return true;
660 return false;
662 case PLUS:
663 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
664 return false;
665 return fixed_base_plus_p (XEXP (x, 0));
667 default:
668 return false;
672 /* Dump the expressions in the equivalence class indicated by CLASSP.
673 This function is used only for debugging. */
674 void
675 dump_class (struct table_elt *classp)
677 struct table_elt *elt;
679 fprintf (stderr, "Equivalence chain for ");
680 print_rtl (stderr, classp->exp);
681 fprintf (stderr, ": \n");
683 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
685 print_rtl (stderr, elt->exp);
686 fprintf (stderr, "\n");
690 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
692 static int
693 approx_reg_cost_1 (rtx *xp, void *data)
695 rtx x = *xp;
696 int *cost_p = data;
698 if (x && REG_P (x))
700 unsigned int regno = REGNO (x);
702 if (! CHEAP_REGNO (regno))
704 if (regno < FIRST_PSEUDO_REGISTER)
706 if (SMALL_REGISTER_CLASSES)
707 return 1;
708 *cost_p += 2;
710 else
711 *cost_p += 1;
715 return 0;
718 /* Return an estimate of the cost of the registers used in an rtx.
719 This is mostly the number of different REG expressions in the rtx;
720 however for some exceptions like fixed registers we use a cost of
721 0. If any other hard register reference occurs, return MAX_COST. */
723 static int
724 approx_reg_cost (rtx x)
726 int cost = 0;
728 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
729 return MAX_COST;
731 return cost;
734 /* Returns a canonical version of X for the address, from the point of view,
735 that all multiplications are represented as MULT instead of the multiply
736 by a power of 2 being represented as ASHIFT. */
738 static rtx
739 canon_for_address (rtx x)
741 enum rtx_code code;
742 enum machine_mode mode;
743 rtx new = 0;
744 int i;
745 const char *fmt;
747 if (!x)
748 return x;
750 code = GET_CODE (x);
751 mode = GET_MODE (x);
753 switch (code)
755 case ASHIFT:
756 if (GET_CODE (XEXP (x, 1)) == CONST_INT
757 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)
758 && INTVAL (XEXP (x, 1)) >= 0)
760 new = canon_for_address (XEXP (x, 0));
761 new = gen_rtx_MULT (mode, new,
762 gen_int_mode ((HOST_WIDE_INT) 1
763 << INTVAL (XEXP (x, 1)),
764 mode));
766 break;
767 default:
768 break;
771 if (new)
772 return new;
774 /* Now recursively process each operand of this operation. */
775 fmt = GET_RTX_FORMAT (code);
776 for (i = 0; i < GET_RTX_LENGTH (code); i++)
777 if (fmt[i] == 'e')
779 new = canon_for_address (XEXP (x, i));
780 XEXP (x, i) = new;
782 return x;
785 /* Return a negative value if an rtx A, whose costs are given by COST_A
786 and REGCOST_A, is more desirable than an rtx B.
787 Return a positive value if A is less desirable, or 0 if the two are
788 equally good. */
789 static int
790 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
792 /* First, get rid of cases involving expressions that are entirely
793 unwanted. */
794 if (cost_a != cost_b)
796 if (cost_a == MAX_COST)
797 return 1;
798 if (cost_b == MAX_COST)
799 return -1;
802 /* Avoid extending lifetimes of hardregs. */
803 if (regcost_a != regcost_b)
805 if (regcost_a == MAX_COST)
806 return 1;
807 if (regcost_b == MAX_COST)
808 return -1;
811 /* Normal operation costs take precedence. */
812 if (cost_a != cost_b)
813 return cost_a - cost_b;
814 /* Only if these are identical consider effects on register pressure. */
815 if (regcost_a != regcost_b)
816 return regcost_a - regcost_b;
817 return 0;
820 /* Internal function, to compute cost when X is not a register; called
821 from COST macro to keep it simple. */
823 static int
824 notreg_cost (rtx x, enum rtx_code outer)
826 return ((GET_CODE (x) == SUBREG
827 && REG_P (SUBREG_REG (x))
828 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
829 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
830 && (GET_MODE_SIZE (GET_MODE (x))
831 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
832 && subreg_lowpart_p (x)
833 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
834 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
836 : rtx_cost (x, outer) * 2);
840 /* Initialize CSE_REG_INFO_TABLE. */
842 static void
843 init_cse_reg_info (unsigned int nregs)
845 /* Do we need to grow the table? */
846 if (nregs > cse_reg_info_table_size)
848 unsigned int new_size;
850 if (cse_reg_info_table_size < 2048)
852 /* Compute a new size that is a power of 2 and no smaller
853 than the large of NREGS and 64. */
854 new_size = (cse_reg_info_table_size
855 ? cse_reg_info_table_size : 64);
857 while (new_size < nregs)
858 new_size *= 2;
860 else
862 /* If we need a big table, allocate just enough to hold
863 NREGS registers. */
864 new_size = nregs;
867 /* Reallocate the table with NEW_SIZE entries. */
868 if (cse_reg_info_table)
869 free (cse_reg_info_table);
870 cse_reg_info_table = xmalloc (sizeof (struct cse_reg_info)
871 * new_size);
872 cse_reg_info_table_size = new_size;
873 cse_reg_info_table_first_uninitialized = 0;
876 /* Do we have all of the first NREGS entries initialized? */
877 if (cse_reg_info_table_first_uninitialized < nregs)
879 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
880 unsigned int i;
882 /* Put the old timestamp on newly allocated entries so that they
883 will all be considered out of date. We do not touch those
884 entries beyond the first NREGS entries to be nice to the
885 virtual memory. */
886 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
887 cse_reg_info_table[i].timestamp = old_timestamp;
889 cse_reg_info_table_first_uninitialized = nregs;
893 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
895 static void
896 get_cse_reg_info_1 (unsigned int regno)
898 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
899 entry will be considered to have been initialized. */
900 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
902 /* Initialize the rest of the entry. */
903 cse_reg_info_table[regno].reg_tick = 1;
904 cse_reg_info_table[regno].reg_in_table = -1;
905 cse_reg_info_table[regno].subreg_ticked = -1;
906 cse_reg_info_table[regno].reg_qty = -regno - 1;
909 /* Find a cse_reg_info entry for REGNO. */
911 static inline struct cse_reg_info *
912 get_cse_reg_info (unsigned int regno)
914 struct cse_reg_info *p = &cse_reg_info_table[regno];
916 /* If this entry has not been initialized, go ahead and initialize
917 it. */
918 if (p->timestamp != cse_reg_info_timestamp)
919 get_cse_reg_info_1 (regno);
921 return p;
924 /* Clear the hash table and initialize each register with its own quantity,
925 for a new basic block. */
927 static void
928 new_basic_block (void)
930 int i;
932 next_qty = 0;
934 /* Invalidate cse_reg_info_table. */
935 cse_reg_info_timestamp++;
937 /* Clear out hash table state for this pass. */
938 CLEAR_HARD_REG_SET (hard_regs_in_table);
940 /* The per-quantity values used to be initialized here, but it is
941 much faster to initialize each as it is made in `make_new_qty'. */
943 for (i = 0; i < HASH_SIZE; i++)
945 struct table_elt *first;
947 first = table[i];
948 if (first != NULL)
950 struct table_elt *last = first;
952 table[i] = NULL;
954 while (last->next_same_hash != NULL)
955 last = last->next_same_hash;
957 /* Now relink this hash entire chain into
958 the free element list. */
960 last->next_same_hash = free_element_chain;
961 free_element_chain = first;
965 #ifdef HAVE_cc0
966 prev_insn = 0;
967 prev_insn_cc0 = 0;
968 #endif
971 /* Say that register REG contains a quantity in mode MODE not in any
972 register before and initialize that quantity. */
974 static void
975 make_new_qty (unsigned int reg, enum machine_mode mode)
977 int q;
978 struct qty_table_elem *ent;
979 struct reg_eqv_elem *eqv;
981 gcc_assert (next_qty < max_qty);
983 q = REG_QTY (reg) = next_qty++;
984 ent = &qty_table[q];
985 ent->first_reg = reg;
986 ent->last_reg = reg;
987 ent->mode = mode;
988 ent->const_rtx = ent->const_insn = NULL_RTX;
989 ent->comparison_code = UNKNOWN;
991 eqv = &reg_eqv_table[reg];
992 eqv->next = eqv->prev = -1;
995 /* Make reg NEW equivalent to reg OLD.
996 OLD is not changing; NEW is. */
998 static void
999 make_regs_eqv (unsigned int new, unsigned int old)
1001 unsigned int lastr, firstr;
1002 int q = REG_QTY (old);
1003 struct qty_table_elem *ent;
1005 ent = &qty_table[q];
1007 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1008 gcc_assert (REGNO_QTY_VALID_P (old));
1010 REG_QTY (new) = q;
1011 firstr = ent->first_reg;
1012 lastr = ent->last_reg;
1014 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1015 hard regs. Among pseudos, if NEW will live longer than any other reg
1016 of the same qty, and that is beyond the current basic block,
1017 make it the new canonical replacement for this qty. */
1018 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1019 /* Certain fixed registers might be of the class NO_REGS. This means
1020 that not only can they not be allocated by the compiler, but
1021 they cannot be used in substitutions or canonicalizations
1022 either. */
1023 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1024 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1025 || (new >= FIRST_PSEUDO_REGISTER
1026 && (firstr < FIRST_PSEUDO_REGISTER
1027 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1028 || (uid_cuid[REGNO_FIRST_UID (new)]
1029 < cse_basic_block_start))
1030 && (uid_cuid[REGNO_LAST_UID (new)]
1031 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1033 reg_eqv_table[firstr].prev = new;
1034 reg_eqv_table[new].next = firstr;
1035 reg_eqv_table[new].prev = -1;
1036 ent->first_reg = new;
1038 else
1040 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1041 Otherwise, insert before any non-fixed hard regs that are at the
1042 end. Registers of class NO_REGS cannot be used as an
1043 equivalent for anything. */
1044 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1045 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1046 && new >= FIRST_PSEUDO_REGISTER)
1047 lastr = reg_eqv_table[lastr].prev;
1048 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1049 if (reg_eqv_table[lastr].next >= 0)
1050 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1051 else
1052 qty_table[q].last_reg = new;
1053 reg_eqv_table[lastr].next = new;
1054 reg_eqv_table[new].prev = lastr;
1058 /* Remove REG from its equivalence class. */
1060 static void
1061 delete_reg_equiv (unsigned int reg)
1063 struct qty_table_elem *ent;
1064 int q = REG_QTY (reg);
1065 int p, n;
1067 /* If invalid, do nothing. */
1068 if (! REGNO_QTY_VALID_P (reg))
1069 return;
1071 ent = &qty_table[q];
1073 p = reg_eqv_table[reg].prev;
1074 n = reg_eqv_table[reg].next;
1076 if (n != -1)
1077 reg_eqv_table[n].prev = p;
1078 else
1079 ent->last_reg = p;
1080 if (p != -1)
1081 reg_eqv_table[p].next = n;
1082 else
1083 ent->first_reg = n;
1085 REG_QTY (reg) = -reg - 1;
1088 /* Remove any invalid expressions from the hash table
1089 that refer to any of the registers contained in expression X.
1091 Make sure that newly inserted references to those registers
1092 as subexpressions will be considered valid.
1094 mention_regs is not called when a register itself
1095 is being stored in the table.
1097 Return 1 if we have done something that may have changed the hash code
1098 of X. */
1100 static int
1101 mention_regs (rtx x)
1103 enum rtx_code code;
1104 int i, j;
1105 const char *fmt;
1106 int changed = 0;
1108 if (x == 0)
1109 return 0;
1111 code = GET_CODE (x);
1112 if (code == REG)
1114 unsigned int regno = REGNO (x);
1115 unsigned int endregno
1116 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1117 : hard_regno_nregs[regno][GET_MODE (x)]);
1118 unsigned int i;
1120 for (i = regno; i < endregno; i++)
1122 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1123 remove_invalid_refs (i);
1125 REG_IN_TABLE (i) = REG_TICK (i);
1126 SUBREG_TICKED (i) = -1;
1129 return 0;
1132 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1133 pseudo if they don't use overlapping words. We handle only pseudos
1134 here for simplicity. */
1135 if (code == SUBREG && REG_P (SUBREG_REG (x))
1136 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1138 unsigned int i = REGNO (SUBREG_REG (x));
1140 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1142 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1143 the last store to this register really stored into this
1144 subreg, then remove the memory of this subreg.
1145 Otherwise, remove any memory of the entire register and
1146 all its subregs from the table. */
1147 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1148 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1149 remove_invalid_refs (i);
1150 else
1151 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1154 REG_IN_TABLE (i) = REG_TICK (i);
1155 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1156 return 0;
1159 /* If X is a comparison or a COMPARE and either operand is a register
1160 that does not have a quantity, give it one. This is so that a later
1161 call to record_jump_equiv won't cause X to be assigned a different
1162 hash code and not found in the table after that call.
1164 It is not necessary to do this here, since rehash_using_reg can
1165 fix up the table later, but doing this here eliminates the need to
1166 call that expensive function in the most common case where the only
1167 use of the register is in the comparison. */
1169 if (code == COMPARE || COMPARISON_P (x))
1171 if (REG_P (XEXP (x, 0))
1172 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1173 if (insert_regs (XEXP (x, 0), NULL, 0))
1175 rehash_using_reg (XEXP (x, 0));
1176 changed = 1;
1179 if (REG_P (XEXP (x, 1))
1180 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1181 if (insert_regs (XEXP (x, 1), NULL, 0))
1183 rehash_using_reg (XEXP (x, 1));
1184 changed = 1;
1188 fmt = GET_RTX_FORMAT (code);
1189 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1190 if (fmt[i] == 'e')
1191 changed |= mention_regs (XEXP (x, i));
1192 else if (fmt[i] == 'E')
1193 for (j = 0; j < XVECLEN (x, i); j++)
1194 changed |= mention_regs (XVECEXP (x, i, j));
1196 return changed;
1199 /* Update the register quantities for inserting X into the hash table
1200 with a value equivalent to CLASSP.
1201 (If the class does not contain a REG, it is irrelevant.)
1202 If MODIFIED is nonzero, X is a destination; it is being modified.
1203 Note that delete_reg_equiv should be called on a register
1204 before insert_regs is done on that register with MODIFIED != 0.
1206 Nonzero value means that elements of reg_qty have changed
1207 so X's hash code may be different. */
1209 static int
1210 insert_regs (rtx x, struct table_elt *classp, int modified)
1212 if (REG_P (x))
1214 unsigned int regno = REGNO (x);
1215 int qty_valid;
1217 /* If REGNO is in the equivalence table already but is of the
1218 wrong mode for that equivalence, don't do anything here. */
1220 qty_valid = REGNO_QTY_VALID_P (regno);
1221 if (qty_valid)
1223 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1225 if (ent->mode != GET_MODE (x))
1226 return 0;
1229 if (modified || ! qty_valid)
1231 if (classp)
1232 for (classp = classp->first_same_value;
1233 classp != 0;
1234 classp = classp->next_same_value)
1235 if (REG_P (classp->exp)
1236 && GET_MODE (classp->exp) == GET_MODE (x))
1238 unsigned c_regno = REGNO (classp->exp);
1240 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1242 /* Suppose that 5 is hard reg and 100 and 101 are
1243 pseudos. Consider
1245 (set (reg:si 100) (reg:si 5))
1246 (set (reg:si 5) (reg:si 100))
1247 (set (reg:di 101) (reg:di 5))
1249 We would now set REG_QTY (101) = REG_QTY (5), but the
1250 entry for 5 is in SImode. When we use this later in
1251 copy propagation, we get the register in wrong mode. */
1252 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1253 continue;
1255 make_regs_eqv (regno, c_regno);
1256 return 1;
1259 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1260 than REG_IN_TABLE to find out if there was only a single preceding
1261 invalidation - for the SUBREG - or another one, which would be
1262 for the full register. However, if we find here that REG_TICK
1263 indicates that the register is invalid, it means that it has
1264 been invalidated in a separate operation. The SUBREG might be used
1265 now (then this is a recursive call), or we might use the full REG
1266 now and a SUBREG of it later. So bump up REG_TICK so that
1267 mention_regs will do the right thing. */
1268 if (! modified
1269 && REG_IN_TABLE (regno) >= 0
1270 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1271 REG_TICK (regno)++;
1272 make_new_qty (regno, GET_MODE (x));
1273 return 1;
1276 return 0;
1279 /* If X is a SUBREG, we will likely be inserting the inner register in the
1280 table. If that register doesn't have an assigned quantity number at
1281 this point but does later, the insertion that we will be doing now will
1282 not be accessible because its hash code will have changed. So assign
1283 a quantity number now. */
1285 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1286 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1288 insert_regs (SUBREG_REG (x), NULL, 0);
1289 mention_regs (x);
1290 return 1;
1292 else
1293 return mention_regs (x);
1296 /* Look in or update the hash table. */
1298 /* Remove table element ELT from use in the table.
1299 HASH is its hash code, made using the HASH macro.
1300 It's an argument because often that is known in advance
1301 and we save much time not recomputing it. */
1303 static void
1304 remove_from_table (struct table_elt *elt, unsigned int hash)
1306 if (elt == 0)
1307 return;
1309 /* Mark this element as removed. See cse_insn. */
1310 elt->first_same_value = 0;
1312 /* Remove the table element from its equivalence class. */
1315 struct table_elt *prev = elt->prev_same_value;
1316 struct table_elt *next = elt->next_same_value;
1318 if (next)
1319 next->prev_same_value = prev;
1321 if (prev)
1322 prev->next_same_value = next;
1323 else
1325 struct table_elt *newfirst = next;
1326 while (next)
1328 next->first_same_value = newfirst;
1329 next = next->next_same_value;
1334 /* Remove the table element from its hash bucket. */
1337 struct table_elt *prev = elt->prev_same_hash;
1338 struct table_elt *next = elt->next_same_hash;
1340 if (next)
1341 next->prev_same_hash = prev;
1343 if (prev)
1344 prev->next_same_hash = next;
1345 else if (table[hash] == elt)
1346 table[hash] = next;
1347 else
1349 /* This entry is not in the proper hash bucket. This can happen
1350 when two classes were merged by `merge_equiv_classes'. Search
1351 for the hash bucket that it heads. This happens only very
1352 rarely, so the cost is acceptable. */
1353 for (hash = 0; hash < HASH_SIZE; hash++)
1354 if (table[hash] == elt)
1355 table[hash] = next;
1359 /* Remove the table element from its related-value circular chain. */
1361 if (elt->related_value != 0 && elt->related_value != elt)
1363 struct table_elt *p = elt->related_value;
1365 while (p->related_value != elt)
1366 p = p->related_value;
1367 p->related_value = elt->related_value;
1368 if (p->related_value == p)
1369 p->related_value = 0;
1372 /* Now add it to the free element chain. */
1373 elt->next_same_hash = free_element_chain;
1374 free_element_chain = elt;
1377 /* Look up X in the hash table and return its table element,
1378 or 0 if X is not in the table.
1380 MODE is the machine-mode of X, or if X is an integer constant
1381 with VOIDmode then MODE is the mode with which X will be used.
1383 Here we are satisfied to find an expression whose tree structure
1384 looks like X. */
1386 static struct table_elt *
1387 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1389 struct table_elt *p;
1391 for (p = table[hash]; p; p = p->next_same_hash)
1392 if (mode == p->mode && ((x == p->exp && REG_P (x))
1393 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1394 return p;
1396 return 0;
1399 /* Like `lookup' but don't care whether the table element uses invalid regs.
1400 Also ignore discrepancies in the machine mode of a register. */
1402 static struct table_elt *
1403 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1405 struct table_elt *p;
1407 if (REG_P (x))
1409 unsigned int regno = REGNO (x);
1411 /* Don't check the machine mode when comparing registers;
1412 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1413 for (p = table[hash]; p; p = p->next_same_hash)
1414 if (REG_P (p->exp)
1415 && REGNO (p->exp) == regno)
1416 return p;
1418 else
1420 for (p = table[hash]; p; p = p->next_same_hash)
1421 if (mode == p->mode
1422 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1423 return p;
1426 return 0;
1429 /* Look for an expression equivalent to X and with code CODE.
1430 If one is found, return that expression. */
1432 static rtx
1433 lookup_as_function (rtx x, enum rtx_code code)
1435 struct table_elt *p
1436 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1438 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1439 long as we are narrowing. So if we looked in vain for a mode narrower
1440 than word_mode before, look for word_mode now. */
1441 if (p == 0 && code == CONST_INT
1442 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1444 x = copy_rtx (x);
1445 PUT_MODE (x, word_mode);
1446 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1449 if (p == 0)
1450 return 0;
1452 for (p = p->first_same_value; p; p = p->next_same_value)
1453 if (GET_CODE (p->exp) == code
1454 /* Make sure this is a valid entry in the table. */
1455 && exp_equiv_p (p->exp, p->exp, 1, false))
1456 return p->exp;
1458 return 0;
1461 /* Insert X in the hash table, assuming HASH is its hash code
1462 and CLASSP is an element of the class it should go in
1463 (or 0 if a new class should be made).
1464 It is inserted at the proper position to keep the class in
1465 the order cheapest first.
1467 MODE is the machine-mode of X, or if X is an integer constant
1468 with VOIDmode then MODE is the mode with which X will be used.
1470 For elements of equal cheapness, the most recent one
1471 goes in front, except that the first element in the list
1472 remains first unless a cheaper element is added. The order of
1473 pseudo-registers does not matter, as canon_reg will be called to
1474 find the cheapest when a register is retrieved from the table.
1476 The in_memory field in the hash table element is set to 0.
1477 The caller must set it nonzero if appropriate.
1479 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1480 and if insert_regs returns a nonzero value
1481 you must then recompute its hash code before calling here.
1483 If necessary, update table showing constant values of quantities. */
1485 #define CHEAPER(X, Y) \
1486 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1488 static struct table_elt *
1489 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1491 struct table_elt *elt;
1493 /* If X is a register and we haven't made a quantity for it,
1494 something is wrong. */
1495 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1497 /* If X is a hard register, show it is being put in the table. */
1498 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1500 unsigned int regno = REGNO (x);
1501 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1502 unsigned int i;
1504 for (i = regno; i < endregno; i++)
1505 SET_HARD_REG_BIT (hard_regs_in_table, i);
1508 /* Put an element for X into the right hash bucket. */
1510 elt = free_element_chain;
1511 if (elt)
1512 free_element_chain = elt->next_same_hash;
1513 else
1514 elt = xmalloc (sizeof (struct table_elt));
1516 elt->exp = x;
1517 elt->canon_exp = NULL_RTX;
1518 elt->cost = COST (x);
1519 elt->regcost = approx_reg_cost (x);
1520 elt->next_same_value = 0;
1521 elt->prev_same_value = 0;
1522 elt->next_same_hash = table[hash];
1523 elt->prev_same_hash = 0;
1524 elt->related_value = 0;
1525 elt->in_memory = 0;
1526 elt->mode = mode;
1527 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1529 if (table[hash])
1530 table[hash]->prev_same_hash = elt;
1531 table[hash] = elt;
1533 /* Put it into the proper value-class. */
1534 if (classp)
1536 classp = classp->first_same_value;
1537 if (CHEAPER (elt, classp))
1538 /* Insert at the head of the class. */
1540 struct table_elt *p;
1541 elt->next_same_value = classp;
1542 classp->prev_same_value = elt;
1543 elt->first_same_value = elt;
1545 for (p = classp; p; p = p->next_same_value)
1546 p->first_same_value = elt;
1548 else
1550 /* Insert not at head of the class. */
1551 /* Put it after the last element cheaper than X. */
1552 struct table_elt *p, *next;
1554 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1555 p = next);
1557 /* Put it after P and before NEXT. */
1558 elt->next_same_value = next;
1559 if (next)
1560 next->prev_same_value = elt;
1562 elt->prev_same_value = p;
1563 p->next_same_value = elt;
1564 elt->first_same_value = classp;
1567 else
1568 elt->first_same_value = elt;
1570 /* If this is a constant being set equivalent to a register or a register
1571 being set equivalent to a constant, note the constant equivalence.
1573 If this is a constant, it cannot be equivalent to a different constant,
1574 and a constant is the only thing that can be cheaper than a register. So
1575 we know the register is the head of the class (before the constant was
1576 inserted).
1578 If this is a register that is not already known equivalent to a
1579 constant, we must check the entire class.
1581 If this is a register that is already known equivalent to an insn,
1582 update the qtys `const_insn' to show that `this_insn' is the latest
1583 insn making that quantity equivalent to the constant. */
1585 if (elt->is_const && classp && REG_P (classp->exp)
1586 && !REG_P (x))
1588 int exp_q = REG_QTY (REGNO (classp->exp));
1589 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1591 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1592 exp_ent->const_insn = this_insn;
1595 else if (REG_P (x)
1596 && classp
1597 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1598 && ! elt->is_const)
1600 struct table_elt *p;
1602 for (p = classp; p != 0; p = p->next_same_value)
1604 if (p->is_const && !REG_P (p->exp))
1606 int x_q = REG_QTY (REGNO (x));
1607 struct qty_table_elem *x_ent = &qty_table[x_q];
1609 x_ent->const_rtx
1610 = gen_lowpart (GET_MODE (x), p->exp);
1611 x_ent->const_insn = this_insn;
1612 break;
1617 else if (REG_P (x)
1618 && qty_table[REG_QTY (REGNO (x))].const_rtx
1619 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1620 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1622 /* If this is a constant with symbolic value,
1623 and it has a term with an explicit integer value,
1624 link it up with related expressions. */
1625 if (GET_CODE (x) == CONST)
1627 rtx subexp = get_related_value (x);
1628 unsigned subhash;
1629 struct table_elt *subelt, *subelt_prev;
1631 if (subexp != 0)
1633 /* Get the integer-free subexpression in the hash table. */
1634 subhash = SAFE_HASH (subexp, mode);
1635 subelt = lookup (subexp, subhash, mode);
1636 if (subelt == 0)
1637 subelt = insert (subexp, NULL, subhash, mode);
1638 /* Initialize SUBELT's circular chain if it has none. */
1639 if (subelt->related_value == 0)
1640 subelt->related_value = subelt;
1641 /* Find the element in the circular chain that precedes SUBELT. */
1642 subelt_prev = subelt;
1643 while (subelt_prev->related_value != subelt)
1644 subelt_prev = subelt_prev->related_value;
1645 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1646 This way the element that follows SUBELT is the oldest one. */
1647 elt->related_value = subelt_prev->related_value;
1648 subelt_prev->related_value = elt;
1652 return elt;
1655 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1656 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1657 the two classes equivalent.
1659 CLASS1 will be the surviving class; CLASS2 should not be used after this
1660 call.
1662 Any invalid entries in CLASS2 will not be copied. */
1664 static void
1665 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1667 struct table_elt *elt, *next, *new;
1669 /* Ensure we start with the head of the classes. */
1670 class1 = class1->first_same_value;
1671 class2 = class2->first_same_value;
1673 /* If they were already equal, forget it. */
1674 if (class1 == class2)
1675 return;
1677 for (elt = class2; elt; elt = next)
1679 unsigned int hash;
1680 rtx exp = elt->exp;
1681 enum machine_mode mode = elt->mode;
1683 next = elt->next_same_value;
1685 /* Remove old entry, make a new one in CLASS1's class.
1686 Don't do this for invalid entries as we cannot find their
1687 hash code (it also isn't necessary). */
1688 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1690 bool need_rehash = false;
1692 hash_arg_in_memory = 0;
1693 hash = HASH (exp, mode);
1695 if (REG_P (exp))
1697 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1698 delete_reg_equiv (REGNO (exp));
1701 remove_from_table (elt, hash);
1703 if (insert_regs (exp, class1, 0) || need_rehash)
1705 rehash_using_reg (exp);
1706 hash = HASH (exp, mode);
1708 new = insert (exp, class1, hash, mode);
1709 new->in_memory = hash_arg_in_memory;
1714 /* Flush the entire hash table. */
1716 static void
1717 flush_hash_table (void)
1719 int i;
1720 struct table_elt *p;
1722 for (i = 0; i < HASH_SIZE; i++)
1723 for (p = table[i]; p; p = table[i])
1725 /* Note that invalidate can remove elements
1726 after P in the current hash chain. */
1727 if (REG_P (p->exp))
1728 invalidate (p->exp, p->mode);
1729 else
1730 remove_from_table (p, i);
1734 /* Function called for each rtx to check whether true dependence exist. */
1735 struct check_dependence_data
1737 enum machine_mode mode;
1738 rtx exp;
1739 rtx addr;
1742 static int
1743 check_dependence (rtx *x, void *data)
1745 struct check_dependence_data *d = (struct check_dependence_data *) data;
1746 if (*x && MEM_P (*x))
1747 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1748 cse_rtx_varies_p);
1749 else
1750 return 0;
1753 /* Remove from the hash table, or mark as invalid, all expressions whose
1754 values could be altered by storing in X. X is a register, a subreg, or
1755 a memory reference with nonvarying address (because, when a memory
1756 reference with a varying address is stored in, all memory references are
1757 removed by invalidate_memory so specific invalidation is superfluous).
1758 FULL_MODE, if not VOIDmode, indicates that this much should be
1759 invalidated instead of just the amount indicated by the mode of X. This
1760 is only used for bitfield stores into memory.
1762 A nonvarying address may be just a register or just a symbol reference,
1763 or it may be either of those plus a numeric offset. */
1765 static void
1766 invalidate (rtx x, enum machine_mode full_mode)
1768 int i;
1769 struct table_elt *p;
1770 rtx addr;
1772 switch (GET_CODE (x))
1774 case REG:
1776 /* If X is a register, dependencies on its contents are recorded
1777 through the qty number mechanism. Just change the qty number of
1778 the register, mark it as invalid for expressions that refer to it,
1779 and remove it itself. */
1780 unsigned int regno = REGNO (x);
1781 unsigned int hash = HASH (x, GET_MODE (x));
1783 /* Remove REGNO from any quantity list it might be on and indicate
1784 that its value might have changed. If it is a pseudo, remove its
1785 entry from the hash table.
1787 For a hard register, we do the first two actions above for any
1788 additional hard registers corresponding to X. Then, if any of these
1789 registers are in the table, we must remove any REG entries that
1790 overlap these registers. */
1792 delete_reg_equiv (regno);
1793 REG_TICK (regno)++;
1794 SUBREG_TICKED (regno) = -1;
1796 if (regno >= FIRST_PSEUDO_REGISTER)
1798 /* Because a register can be referenced in more than one mode,
1799 we might have to remove more than one table entry. */
1800 struct table_elt *elt;
1802 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1803 remove_from_table (elt, hash);
1805 else
1807 HOST_WIDE_INT in_table
1808 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1809 unsigned int endregno
1810 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1811 unsigned int tregno, tendregno, rn;
1812 struct table_elt *p, *next;
1814 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1816 for (rn = regno + 1; rn < endregno; rn++)
1818 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1819 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1820 delete_reg_equiv (rn);
1821 REG_TICK (rn)++;
1822 SUBREG_TICKED (rn) = -1;
1825 if (in_table)
1826 for (hash = 0; hash < HASH_SIZE; hash++)
1827 for (p = table[hash]; p; p = next)
1829 next = p->next_same_hash;
1831 if (!REG_P (p->exp)
1832 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1833 continue;
1835 tregno = REGNO (p->exp);
1836 tendregno
1837 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1838 if (tendregno > regno && tregno < endregno)
1839 remove_from_table (p, hash);
1843 return;
1845 case SUBREG:
1846 invalidate (SUBREG_REG (x), VOIDmode);
1847 return;
1849 case PARALLEL:
1850 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1851 invalidate (XVECEXP (x, 0, i), VOIDmode);
1852 return;
1854 case EXPR_LIST:
1855 /* This is part of a disjoint return value; extract the location in
1856 question ignoring the offset. */
1857 invalidate (XEXP (x, 0), VOIDmode);
1858 return;
1860 case MEM:
1861 addr = canon_rtx (get_addr (XEXP (x, 0)));
1862 /* Calculate the canonical version of X here so that
1863 true_dependence doesn't generate new RTL for X on each call. */
1864 x = canon_rtx (x);
1866 /* Remove all hash table elements that refer to overlapping pieces of
1867 memory. */
1868 if (full_mode == VOIDmode)
1869 full_mode = GET_MODE (x);
1871 for (i = 0; i < HASH_SIZE; i++)
1873 struct table_elt *next;
1875 for (p = table[i]; p; p = next)
1877 next = p->next_same_hash;
1878 if (p->in_memory)
1880 struct check_dependence_data d;
1882 /* Just canonicalize the expression once;
1883 otherwise each time we call invalidate
1884 true_dependence will canonicalize the
1885 expression again. */
1886 if (!p->canon_exp)
1887 p->canon_exp = canon_rtx (p->exp);
1888 d.exp = x;
1889 d.addr = addr;
1890 d.mode = full_mode;
1891 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1892 remove_from_table (p, i);
1896 return;
1898 default:
1899 gcc_unreachable ();
1903 /* Remove all expressions that refer to register REGNO,
1904 since they are already invalid, and we are about to
1905 mark that register valid again and don't want the old
1906 expressions to reappear as valid. */
1908 static void
1909 remove_invalid_refs (unsigned int regno)
1911 unsigned int i;
1912 struct table_elt *p, *next;
1914 for (i = 0; i < HASH_SIZE; i++)
1915 for (p = table[i]; p; p = next)
1917 next = p->next_same_hash;
1918 if (!REG_P (p->exp)
1919 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1920 remove_from_table (p, i);
1924 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1925 and mode MODE. */
1926 static void
1927 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1928 enum machine_mode mode)
1930 unsigned int i;
1931 struct table_elt *p, *next;
1932 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1934 for (i = 0; i < HASH_SIZE; i++)
1935 for (p = table[i]; p; p = next)
1937 rtx exp = p->exp;
1938 next = p->next_same_hash;
1940 if (!REG_P (exp)
1941 && (GET_CODE (exp) != SUBREG
1942 || !REG_P (SUBREG_REG (exp))
1943 || REGNO (SUBREG_REG (exp)) != regno
1944 || (((SUBREG_BYTE (exp)
1945 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1946 && SUBREG_BYTE (exp) <= end))
1947 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1948 remove_from_table (p, i);
1952 /* Recompute the hash codes of any valid entries in the hash table that
1953 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1955 This is called when we make a jump equivalence. */
1957 static void
1958 rehash_using_reg (rtx x)
1960 unsigned int i;
1961 struct table_elt *p, *next;
1962 unsigned hash;
1964 if (GET_CODE (x) == SUBREG)
1965 x = SUBREG_REG (x);
1967 /* If X is not a register or if the register is known not to be in any
1968 valid entries in the table, we have no work to do. */
1970 if (!REG_P (x)
1971 || REG_IN_TABLE (REGNO (x)) < 0
1972 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1973 return;
1975 /* Scan all hash chains looking for valid entries that mention X.
1976 If we find one and it is in the wrong hash chain, move it. */
1978 for (i = 0; i < HASH_SIZE; i++)
1979 for (p = table[i]; p; p = next)
1981 next = p->next_same_hash;
1982 if (reg_mentioned_p (x, p->exp)
1983 && exp_equiv_p (p->exp, p->exp, 1, false)
1984 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1986 if (p->next_same_hash)
1987 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1989 if (p->prev_same_hash)
1990 p->prev_same_hash->next_same_hash = p->next_same_hash;
1991 else
1992 table[i] = p->next_same_hash;
1994 p->next_same_hash = table[hash];
1995 p->prev_same_hash = 0;
1996 if (table[hash])
1997 table[hash]->prev_same_hash = p;
1998 table[hash] = p;
2003 /* Remove from the hash table any expression that is a call-clobbered
2004 register. Also update their TICK values. */
2006 static void
2007 invalidate_for_call (void)
2009 unsigned int regno, endregno;
2010 unsigned int i;
2011 unsigned hash;
2012 struct table_elt *p, *next;
2013 int in_table = 0;
2015 /* Go through all the hard registers. For each that is clobbered in
2016 a CALL_INSN, remove the register from quantity chains and update
2017 reg_tick if defined. Also see if any of these registers is currently
2018 in the table. */
2020 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2021 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2023 delete_reg_equiv (regno);
2024 if (REG_TICK (regno) >= 0)
2026 REG_TICK (regno)++;
2027 SUBREG_TICKED (regno) = -1;
2030 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2033 /* In the case where we have no call-clobbered hard registers in the
2034 table, we are done. Otherwise, scan the table and remove any
2035 entry that overlaps a call-clobbered register. */
2037 if (in_table)
2038 for (hash = 0; hash < HASH_SIZE; hash++)
2039 for (p = table[hash]; p; p = next)
2041 next = p->next_same_hash;
2043 if (!REG_P (p->exp)
2044 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2045 continue;
2047 regno = REGNO (p->exp);
2048 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
2050 for (i = regno; i < endregno; i++)
2051 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2053 remove_from_table (p, hash);
2054 break;
2059 /* Given an expression X of type CONST,
2060 and ELT which is its table entry (or 0 if it
2061 is not in the hash table),
2062 return an alternate expression for X as a register plus integer.
2063 If none can be found, return 0. */
2065 static rtx
2066 use_related_value (rtx x, struct table_elt *elt)
2068 struct table_elt *relt = 0;
2069 struct table_elt *p, *q;
2070 HOST_WIDE_INT offset;
2072 /* First, is there anything related known?
2073 If we have a table element, we can tell from that.
2074 Otherwise, must look it up. */
2076 if (elt != 0 && elt->related_value != 0)
2077 relt = elt;
2078 else if (elt == 0 && GET_CODE (x) == CONST)
2080 rtx subexp = get_related_value (x);
2081 if (subexp != 0)
2082 relt = lookup (subexp,
2083 SAFE_HASH (subexp, GET_MODE (subexp)),
2084 GET_MODE (subexp));
2087 if (relt == 0)
2088 return 0;
2090 /* Search all related table entries for one that has an
2091 equivalent register. */
2093 p = relt;
2094 while (1)
2096 /* This loop is strange in that it is executed in two different cases.
2097 The first is when X is already in the table. Then it is searching
2098 the RELATED_VALUE list of X's class (RELT). The second case is when
2099 X is not in the table. Then RELT points to a class for the related
2100 value.
2102 Ensure that, whatever case we are in, that we ignore classes that have
2103 the same value as X. */
2105 if (rtx_equal_p (x, p->exp))
2106 q = 0;
2107 else
2108 for (q = p->first_same_value; q; q = q->next_same_value)
2109 if (REG_P (q->exp))
2110 break;
2112 if (q)
2113 break;
2115 p = p->related_value;
2117 /* We went all the way around, so there is nothing to be found.
2118 Alternatively, perhaps RELT was in the table for some other reason
2119 and it has no related values recorded. */
2120 if (p == relt || p == 0)
2121 break;
2124 if (q == 0)
2125 return 0;
2127 offset = (get_integer_term (x) - get_integer_term (p->exp));
2128 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2129 return plus_constant (q->exp, offset);
2132 /* Hash a string. Just add its bytes up. */
2133 static inline unsigned
2134 hash_rtx_string (const char *ps)
2136 unsigned hash = 0;
2137 const unsigned char *p = (const unsigned char *) ps;
2139 if (p)
2140 while (*p)
2141 hash += *p++;
2143 return hash;
2146 /* Hash an rtx. We are careful to make sure the value is never negative.
2147 Equivalent registers hash identically.
2148 MODE is used in hashing for CONST_INTs only;
2149 otherwise the mode of X is used.
2151 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2153 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2154 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2156 Note that cse_insn knows that the hash code of a MEM expression
2157 is just (int) MEM plus the hash code of the address. */
2159 unsigned
2160 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2161 int *hash_arg_in_memory_p, bool have_reg_qty)
2163 int i, j;
2164 unsigned hash = 0;
2165 enum rtx_code code;
2166 const char *fmt;
2168 /* Used to turn recursion into iteration. We can't rely on GCC's
2169 tail-recursion elimination since we need to keep accumulating values
2170 in HASH. */
2171 repeat:
2172 if (x == 0)
2173 return hash;
2175 code = GET_CODE (x);
2176 switch (code)
2178 case REG:
2180 unsigned int regno = REGNO (x);
2182 if (!reload_completed)
2184 /* On some machines, we can't record any non-fixed hard register,
2185 because extending its life will cause reload problems. We
2186 consider ap, fp, sp, gp to be fixed for this purpose.
2188 We also consider CCmode registers to be fixed for this purpose;
2189 failure to do so leads to failure to simplify 0<100 type of
2190 conditionals.
2192 On all machines, we can't record any global registers.
2193 Nor should we record any register that is in a small
2194 class, as defined by CLASS_LIKELY_SPILLED_P. */
2195 bool record;
2197 if (regno >= FIRST_PSEUDO_REGISTER)
2198 record = true;
2199 else if (x == frame_pointer_rtx
2200 || x == hard_frame_pointer_rtx
2201 || x == arg_pointer_rtx
2202 || x == stack_pointer_rtx
2203 || x == pic_offset_table_rtx)
2204 record = true;
2205 else if (global_regs[regno])
2206 record = false;
2207 else if (fixed_regs[regno])
2208 record = true;
2209 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2210 record = true;
2211 else if (SMALL_REGISTER_CLASSES)
2212 record = false;
2213 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2214 record = false;
2215 else
2216 record = true;
2218 if (!record)
2220 *do_not_record_p = 1;
2221 return 0;
2225 hash += ((unsigned int) REG << 7);
2226 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2227 return hash;
2230 /* We handle SUBREG of a REG specially because the underlying
2231 reg changes its hash value with every value change; we don't
2232 want to have to forget unrelated subregs when one subreg changes. */
2233 case SUBREG:
2235 if (REG_P (SUBREG_REG (x)))
2237 hash += (((unsigned int) SUBREG << 7)
2238 + REGNO (SUBREG_REG (x))
2239 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2240 return hash;
2242 break;
2245 case CONST_INT:
2246 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2247 + (unsigned int) INTVAL (x));
2248 return hash;
2250 case CONST_DOUBLE:
2251 /* This is like the general case, except that it only counts
2252 the integers representing the constant. */
2253 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2254 if (GET_MODE (x) != VOIDmode)
2255 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2256 else
2257 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2258 + (unsigned int) CONST_DOUBLE_HIGH (x));
2259 return hash;
2261 case CONST_VECTOR:
2263 int units;
2264 rtx elt;
2266 units = CONST_VECTOR_NUNITS (x);
2268 for (i = 0; i < units; ++i)
2270 elt = CONST_VECTOR_ELT (x, i);
2271 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2272 hash_arg_in_memory_p, have_reg_qty);
2275 return hash;
2278 /* Assume there is only one rtx object for any given label. */
2279 case LABEL_REF:
2280 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2281 differences and differences between each stage's debugging dumps. */
2282 hash += (((unsigned int) LABEL_REF << 7)
2283 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2284 return hash;
2286 case SYMBOL_REF:
2288 /* Don't hash on the symbol's address to avoid bootstrap differences.
2289 Different hash values may cause expressions to be recorded in
2290 different orders and thus different registers to be used in the
2291 final assembler. This also avoids differences in the dump files
2292 between various stages. */
2293 unsigned int h = 0;
2294 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2296 while (*p)
2297 h += (h << 7) + *p++; /* ??? revisit */
2299 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2300 return hash;
2303 case MEM:
2304 /* We don't record if marked volatile or if BLKmode since we don't
2305 know the size of the move. */
2306 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2308 *do_not_record_p = 1;
2309 return 0;
2311 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2312 *hash_arg_in_memory_p = 1;
2314 /* Now that we have already found this special case,
2315 might as well speed it up as much as possible. */
2316 hash += (unsigned) MEM;
2317 x = XEXP (x, 0);
2318 goto repeat;
2320 case USE:
2321 /* A USE that mentions non-volatile memory needs special
2322 handling since the MEM may be BLKmode which normally
2323 prevents an entry from being made. Pure calls are
2324 marked by a USE which mentions BLKmode memory.
2325 See calls.c:emit_call_1. */
2326 if (MEM_P (XEXP (x, 0))
2327 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2329 hash += (unsigned) USE;
2330 x = XEXP (x, 0);
2332 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2333 *hash_arg_in_memory_p = 1;
2335 /* Now that we have already found this special case,
2336 might as well speed it up as much as possible. */
2337 hash += (unsigned) MEM;
2338 x = XEXP (x, 0);
2339 goto repeat;
2341 break;
2343 case PRE_DEC:
2344 case PRE_INC:
2345 case POST_DEC:
2346 case POST_INC:
2347 case PRE_MODIFY:
2348 case POST_MODIFY:
2349 case PC:
2350 case CC0:
2351 case CALL:
2352 case UNSPEC_VOLATILE:
2353 *do_not_record_p = 1;
2354 return 0;
2356 case ASM_OPERANDS:
2357 if (MEM_VOLATILE_P (x))
2359 *do_not_record_p = 1;
2360 return 0;
2362 else
2364 /* We don't want to take the filename and line into account. */
2365 hash += (unsigned) code + (unsigned) GET_MODE (x)
2366 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2367 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2368 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2370 if (ASM_OPERANDS_INPUT_LENGTH (x))
2372 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2374 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2375 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2376 do_not_record_p, hash_arg_in_memory_p,
2377 have_reg_qty)
2378 + hash_rtx_string
2379 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2382 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2383 x = ASM_OPERANDS_INPUT (x, 0);
2384 mode = GET_MODE (x);
2385 goto repeat;
2388 return hash;
2390 break;
2392 default:
2393 break;
2396 i = GET_RTX_LENGTH (code) - 1;
2397 hash += (unsigned) code + (unsigned) GET_MODE (x);
2398 fmt = GET_RTX_FORMAT (code);
2399 for (; i >= 0; i--)
2401 switch (fmt[i])
2403 case 'e':
2404 /* If we are about to do the last recursive call
2405 needed at this level, change it into iteration.
2406 This function is called enough to be worth it. */
2407 if (i == 0)
2409 x = XEXP (x, i);
2410 goto repeat;
2413 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2414 hash_arg_in_memory_p, have_reg_qty);
2415 break;
2417 case 'E':
2418 for (j = 0; j < XVECLEN (x, i); j++)
2419 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2420 hash_arg_in_memory_p, have_reg_qty);
2421 break;
2423 case 's':
2424 hash += hash_rtx_string (XSTR (x, i));
2425 break;
2427 case 'i':
2428 hash += (unsigned int) XINT (x, i);
2429 break;
2431 case '0': case 't':
2432 /* Unused. */
2433 break;
2435 default:
2436 gcc_unreachable ();
2440 return hash;
2443 /* Hash an rtx X for cse via hash_rtx.
2444 Stores 1 in do_not_record if any subexpression is volatile.
2445 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2446 does not have the RTX_UNCHANGING_P bit set. */
2448 static inline unsigned
2449 canon_hash (rtx x, enum machine_mode mode)
2451 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2454 /* Like canon_hash but with no side effects, i.e. do_not_record
2455 and hash_arg_in_memory are not changed. */
2457 static inline unsigned
2458 safe_hash (rtx x, enum machine_mode mode)
2460 int dummy_do_not_record;
2461 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2464 /* Return 1 iff X and Y would canonicalize into the same thing,
2465 without actually constructing the canonicalization of either one.
2466 If VALIDATE is nonzero,
2467 we assume X is an expression being processed from the rtl
2468 and Y was found in the hash table. We check register refs
2469 in Y for being marked as valid.
2471 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2474 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2476 int i, j;
2477 enum rtx_code code;
2478 const char *fmt;
2480 /* Note: it is incorrect to assume an expression is equivalent to itself
2481 if VALIDATE is nonzero. */
2482 if (x == y && !validate)
2483 return 1;
2485 if (x == 0 || y == 0)
2486 return x == y;
2488 code = GET_CODE (x);
2489 if (code != GET_CODE (y))
2490 return 0;
2492 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2493 if (GET_MODE (x) != GET_MODE (y))
2494 return 0;
2496 switch (code)
2498 case PC:
2499 case CC0:
2500 case CONST_INT:
2501 case CONST_DOUBLE:
2502 return x == y;
2504 case LABEL_REF:
2505 return XEXP (x, 0) == XEXP (y, 0);
2507 case SYMBOL_REF:
2508 return XSTR (x, 0) == XSTR (y, 0);
2510 case REG:
2511 if (for_gcse)
2512 return REGNO (x) == REGNO (y);
2513 else
2515 unsigned int regno = REGNO (y);
2516 unsigned int i;
2517 unsigned int endregno
2518 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2519 : hard_regno_nregs[regno][GET_MODE (y)]);
2521 /* If the quantities are not the same, the expressions are not
2522 equivalent. If there are and we are not to validate, they
2523 are equivalent. Otherwise, ensure all regs are up-to-date. */
2525 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2526 return 0;
2528 if (! validate)
2529 return 1;
2531 for (i = regno; i < endregno; i++)
2532 if (REG_IN_TABLE (i) != REG_TICK (i))
2533 return 0;
2535 return 1;
2538 case MEM:
2539 if (for_gcse)
2541 /* Can't merge two expressions in different alias sets, since we
2542 can decide that the expression is transparent in a block when
2543 it isn't, due to it being set with the different alias set. */
2544 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
2545 return 0;
2547 /* A volatile mem should not be considered equivalent to any
2548 other. */
2549 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2550 return 0;
2552 break;
2554 /* For commutative operations, check both orders. */
2555 case PLUS:
2556 case MULT:
2557 case AND:
2558 case IOR:
2559 case XOR:
2560 case NE:
2561 case EQ:
2562 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2563 validate, for_gcse)
2564 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2565 validate, for_gcse))
2566 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2567 validate, for_gcse)
2568 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2569 validate, for_gcse)));
2571 case ASM_OPERANDS:
2572 /* We don't use the generic code below because we want to
2573 disregard filename and line numbers. */
2575 /* A volatile asm isn't equivalent to any other. */
2576 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2577 return 0;
2579 if (GET_MODE (x) != GET_MODE (y)
2580 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2581 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2582 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2583 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2584 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2585 return 0;
2587 if (ASM_OPERANDS_INPUT_LENGTH (x))
2589 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2590 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2591 ASM_OPERANDS_INPUT (y, i),
2592 validate, for_gcse)
2593 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2594 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2595 return 0;
2598 return 1;
2600 default:
2601 break;
2604 /* Compare the elements. If any pair of corresponding elements
2605 fail to match, return 0 for the whole thing. */
2607 fmt = GET_RTX_FORMAT (code);
2608 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2610 switch (fmt[i])
2612 case 'e':
2613 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2614 validate, for_gcse))
2615 return 0;
2616 break;
2618 case 'E':
2619 if (XVECLEN (x, i) != XVECLEN (y, i))
2620 return 0;
2621 for (j = 0; j < XVECLEN (x, i); j++)
2622 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2623 validate, for_gcse))
2624 return 0;
2625 break;
2627 case 's':
2628 if (strcmp (XSTR (x, i), XSTR (y, i)))
2629 return 0;
2630 break;
2632 case 'i':
2633 if (XINT (x, i) != XINT (y, i))
2634 return 0;
2635 break;
2637 case 'w':
2638 if (XWINT (x, i) != XWINT (y, i))
2639 return 0;
2640 break;
2642 case '0':
2643 case 't':
2644 break;
2646 default:
2647 gcc_unreachable ();
2651 return 1;
2654 /* Return 1 if X has a value that can vary even between two
2655 executions of the program. 0 means X can be compared reliably
2656 against certain constants or near-constants. */
2658 static int
2659 cse_rtx_varies_p (rtx x, int from_alias)
2661 /* We need not check for X and the equivalence class being of the same
2662 mode because if X is equivalent to a constant in some mode, it
2663 doesn't vary in any mode. */
2665 if (REG_P (x)
2666 && REGNO_QTY_VALID_P (REGNO (x)))
2668 int x_q = REG_QTY (REGNO (x));
2669 struct qty_table_elem *x_ent = &qty_table[x_q];
2671 if (GET_MODE (x) == x_ent->mode
2672 && x_ent->const_rtx != NULL_RTX)
2673 return 0;
2676 if (GET_CODE (x) == PLUS
2677 && GET_CODE (XEXP (x, 1)) == CONST_INT
2678 && REG_P (XEXP (x, 0))
2679 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2681 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2682 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2684 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2685 && x0_ent->const_rtx != NULL_RTX)
2686 return 0;
2689 /* This can happen as the result of virtual register instantiation, if
2690 the initial constant is too large to be a valid address. This gives
2691 us a three instruction sequence, load large offset into a register,
2692 load fp minus a constant into a register, then a MEM which is the
2693 sum of the two `constant' registers. */
2694 if (GET_CODE (x) == PLUS
2695 && REG_P (XEXP (x, 0))
2696 && REG_P (XEXP (x, 1))
2697 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2698 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2700 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2701 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2702 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2703 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2705 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2706 && x0_ent->const_rtx != NULL_RTX
2707 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2708 && x1_ent->const_rtx != NULL_RTX)
2709 return 0;
2712 return rtx_varies_p (x, from_alias);
2715 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2716 the result if necessary. INSN is as for canon_reg. */
2718 static void
2719 validate_canon_reg (rtx *xloc, rtx insn)
2721 rtx new = canon_reg (*xloc, insn);
2722 int insn_code;
2724 /* If replacing pseudo with hard reg or vice versa, ensure the
2725 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2726 if (insn != 0 && new != 0
2727 && REG_P (new) && REG_P (*xloc)
2728 && (((REGNO (new) < FIRST_PSEUDO_REGISTER)
2729 != (REGNO (*xloc) < FIRST_PSEUDO_REGISTER))
2730 || GET_MODE (new) != GET_MODE (*xloc)
2731 || (insn_code = recog_memoized (insn)) < 0
2732 || insn_data[insn_code].n_dups > 0))
2733 validate_change (insn, xloc, new, 1);
2734 else
2735 *xloc = new;
2738 /* Canonicalize an expression:
2739 replace each register reference inside it
2740 with the "oldest" equivalent register.
2742 If INSN is nonzero and we are replacing a pseudo with a hard register
2743 or vice versa, validate_change is used to ensure that INSN remains valid
2744 after we make our substitution. The calls are made with IN_GROUP nonzero
2745 so apply_change_group must be called upon the outermost return from this
2746 function (unless INSN is zero). The result of apply_change_group can
2747 generally be discarded since the changes we are making are optional. */
2749 static rtx
2750 canon_reg (rtx x, rtx insn)
2752 int i;
2753 enum rtx_code code;
2754 const char *fmt;
2756 if (x == 0)
2757 return x;
2759 code = GET_CODE (x);
2760 switch (code)
2762 case PC:
2763 case CC0:
2764 case CONST:
2765 case CONST_INT:
2766 case CONST_DOUBLE:
2767 case CONST_VECTOR:
2768 case SYMBOL_REF:
2769 case LABEL_REF:
2770 case ADDR_VEC:
2771 case ADDR_DIFF_VEC:
2772 return x;
2774 case REG:
2776 int first;
2777 int q;
2778 struct qty_table_elem *ent;
2780 /* Never replace a hard reg, because hard regs can appear
2781 in more than one machine mode, and we must preserve the mode
2782 of each occurrence. Also, some hard regs appear in
2783 MEMs that are shared and mustn't be altered. Don't try to
2784 replace any reg that maps to a reg of class NO_REGS. */
2785 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2786 || ! REGNO_QTY_VALID_P (REGNO (x)))
2787 return x;
2789 q = REG_QTY (REGNO (x));
2790 ent = &qty_table[q];
2791 first = ent->first_reg;
2792 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2793 : REGNO_REG_CLASS (first) == NO_REGS ? x
2794 : gen_rtx_REG (ent->mode, first));
2797 default:
2798 break;
2801 fmt = GET_RTX_FORMAT (code);
2802 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2804 int j;
2806 if (fmt[i] == 'e')
2807 validate_canon_reg (&XEXP (x, i), insn);
2808 else if (fmt[i] == 'E')
2809 for (j = 0; j < XVECLEN (x, i); j++)
2810 validate_canon_reg (&XVECEXP (x, i, j), insn);
2813 return x;
2816 /* LOC is a location within INSN that is an operand address (the contents of
2817 a MEM). Find the best equivalent address to use that is valid for this
2818 insn.
2820 On most CISC machines, complicated address modes are costly, and rtx_cost
2821 is a good approximation for that cost. However, most RISC machines have
2822 only a few (usually only one) memory reference formats. If an address is
2823 valid at all, it is often just as cheap as any other address. Hence, for
2824 RISC machines, we use `address_cost' to compare the costs of various
2825 addresses. For two addresses of equal cost, choose the one with the
2826 highest `rtx_cost' value as that has the potential of eliminating the
2827 most insns. For equal costs, we choose the first in the equivalence
2828 class. Note that we ignore the fact that pseudo registers are cheaper than
2829 hard registers here because we would also prefer the pseudo registers. */
2831 static void
2832 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2834 struct table_elt *elt;
2835 rtx addr = *loc;
2836 struct table_elt *p;
2837 int found_better = 1;
2838 int save_do_not_record = do_not_record;
2839 int save_hash_arg_in_memory = hash_arg_in_memory;
2840 int addr_volatile;
2841 int regno;
2842 unsigned hash;
2844 /* Do not try to replace constant addresses or addresses of local and
2845 argument slots. These MEM expressions are made only once and inserted
2846 in many instructions, as well as being used to control symbol table
2847 output. It is not safe to clobber them.
2849 There are some uncommon cases where the address is already in a register
2850 for some reason, but we cannot take advantage of that because we have
2851 no easy way to unshare the MEM. In addition, looking up all stack
2852 addresses is costly. */
2853 if ((GET_CODE (addr) == PLUS
2854 && REG_P (XEXP (addr, 0))
2855 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2856 && (regno = REGNO (XEXP (addr, 0)),
2857 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2858 || regno == ARG_POINTER_REGNUM))
2859 || (REG_P (addr)
2860 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2861 || regno == HARD_FRAME_POINTER_REGNUM
2862 || regno == ARG_POINTER_REGNUM))
2863 || CONSTANT_ADDRESS_P (addr))
2864 return;
2866 /* If this address is not simply a register, try to fold it. This will
2867 sometimes simplify the expression. Many simplifications
2868 will not be valid, but some, usually applying the associative rule, will
2869 be valid and produce better code. */
2870 if (!REG_P (addr))
2872 rtx folded = canon_for_address (fold_rtx (addr, NULL_RTX));
2874 if (folded != addr)
2876 int addr_folded_cost = address_cost (folded, mode);
2877 int addr_cost = address_cost (addr, mode);
2879 if ((addr_folded_cost < addr_cost
2880 || (addr_folded_cost == addr_cost
2881 /* ??? The rtx_cost comparison is left over from an older
2882 version of this code. It is probably no longer helpful.*/
2883 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2884 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2885 && validate_change (insn, loc, folded, 0))
2886 addr = folded;
2890 /* If this address is not in the hash table, we can't look for equivalences
2891 of the whole address. Also, ignore if volatile. */
2893 do_not_record = 0;
2894 hash = HASH (addr, Pmode);
2895 addr_volatile = do_not_record;
2896 do_not_record = save_do_not_record;
2897 hash_arg_in_memory = save_hash_arg_in_memory;
2899 if (addr_volatile)
2900 return;
2902 elt = lookup (addr, hash, Pmode);
2904 if (elt)
2906 /* We need to find the best (under the criteria documented above) entry
2907 in the class that is valid. We use the `flag' field to indicate
2908 choices that were invalid and iterate until we can't find a better
2909 one that hasn't already been tried. */
2911 for (p = elt->first_same_value; p; p = p->next_same_value)
2912 p->flag = 0;
2914 while (found_better)
2916 int best_addr_cost = address_cost (*loc, mode);
2917 int best_rtx_cost = (elt->cost + 1) >> 1;
2918 int exp_cost;
2919 struct table_elt *best_elt = elt;
2921 found_better = 0;
2922 for (p = elt->first_same_value; p; p = p->next_same_value)
2923 if (! p->flag)
2925 if ((REG_P (p->exp)
2926 || exp_equiv_p (p->exp, p->exp, 1, false))
2927 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2928 || (exp_cost == best_addr_cost
2929 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2931 found_better = 1;
2932 best_addr_cost = exp_cost;
2933 best_rtx_cost = (p->cost + 1) >> 1;
2934 best_elt = p;
2938 if (found_better)
2940 if (validate_change (insn, loc,
2941 canon_reg (copy_rtx (best_elt->exp),
2942 NULL_RTX), 0))
2943 return;
2944 else
2945 best_elt->flag = 1;
2950 /* If the address is a binary operation with the first operand a register
2951 and the second a constant, do the same as above, but looking for
2952 equivalences of the register. Then try to simplify before checking for
2953 the best address to use. This catches a few cases: First is when we
2954 have REG+const and the register is another REG+const. We can often merge
2955 the constants and eliminate one insn and one register. It may also be
2956 that a machine has a cheap REG+REG+const. Finally, this improves the
2957 code on the Alpha for unaligned byte stores. */
2959 if (flag_expensive_optimizations
2960 && ARITHMETIC_P (*loc)
2961 && REG_P (XEXP (*loc, 0)))
2963 rtx op1 = XEXP (*loc, 1);
2965 do_not_record = 0;
2966 hash = HASH (XEXP (*loc, 0), Pmode);
2967 do_not_record = save_do_not_record;
2968 hash_arg_in_memory = save_hash_arg_in_memory;
2970 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2971 if (elt == 0)
2972 return;
2974 /* We need to find the best (under the criteria documented above) entry
2975 in the class that is valid. We use the `flag' field to indicate
2976 choices that were invalid and iterate until we can't find a better
2977 one that hasn't already been tried. */
2979 for (p = elt->first_same_value; p; p = p->next_same_value)
2980 p->flag = 0;
2982 while (found_better)
2984 int best_addr_cost = address_cost (*loc, mode);
2985 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2986 struct table_elt *best_elt = elt;
2987 rtx best_rtx = *loc;
2988 int count;
2990 /* This is at worst case an O(n^2) algorithm, so limit our search
2991 to the first 32 elements on the list. This avoids trouble
2992 compiling code with very long basic blocks that can easily
2993 call simplify_gen_binary so many times that we run out of
2994 memory. */
2996 found_better = 0;
2997 for (p = elt->first_same_value, count = 0;
2998 p && count < 32;
2999 p = p->next_same_value, count++)
3000 if (! p->flag
3001 && (REG_P (p->exp)
3002 || (GET_CODE (p->exp) != EXPR_LIST
3003 && exp_equiv_p (p->exp, p->exp, 1, false))))
3006 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3007 p->exp, op1);
3008 int new_cost;
3010 /* Get the canonical version of the address so we can accept
3011 more. */
3012 new = canon_for_address (new);
3014 new_cost = address_cost (new, mode);
3016 if (new_cost < best_addr_cost
3017 || (new_cost == best_addr_cost
3018 && (COST (new) + 1) >> 1 > best_rtx_cost))
3020 found_better = 1;
3021 best_addr_cost = new_cost;
3022 best_rtx_cost = (COST (new) + 1) >> 1;
3023 best_elt = p;
3024 best_rtx = new;
3028 if (found_better)
3030 if (validate_change (insn, loc,
3031 canon_reg (copy_rtx (best_rtx),
3032 NULL_RTX), 0))
3033 return;
3034 else
3035 best_elt->flag = 1;
3041 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3042 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3043 what values are being compared.
3045 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3046 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3047 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3048 compared to produce cc0.
3050 The return value is the comparison operator and is either the code of
3051 A or the code corresponding to the inverse of the comparison. */
3053 static enum rtx_code
3054 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3055 enum machine_mode *pmode1, enum machine_mode *pmode2)
3057 rtx arg1, arg2;
3059 arg1 = *parg1, arg2 = *parg2;
3061 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3063 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3065 /* Set nonzero when we find something of interest. */
3066 rtx x = 0;
3067 int reverse_code = 0;
3068 struct table_elt *p = 0;
3070 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3071 On machines with CC0, this is the only case that can occur, since
3072 fold_rtx will return the COMPARE or item being compared with zero
3073 when given CC0. */
3075 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3076 x = arg1;
3078 /* If ARG1 is a comparison operator and CODE is testing for
3079 STORE_FLAG_VALUE, get the inner arguments. */
3081 else if (COMPARISON_P (arg1))
3083 #ifdef FLOAT_STORE_FLAG_VALUE
3084 REAL_VALUE_TYPE fsfv;
3085 #endif
3087 if (code == NE
3088 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3089 && code == LT && STORE_FLAG_VALUE == -1)
3090 #ifdef FLOAT_STORE_FLAG_VALUE
3091 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3092 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3093 REAL_VALUE_NEGATIVE (fsfv)))
3094 #endif
3096 x = arg1;
3097 else if (code == EQ
3098 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3099 && code == GE && 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, reverse_code = 1;
3109 /* ??? We could also check for
3111 (ne (and (eq (...) (const_int 1))) (const_int 0))
3113 and related forms, but let's wait until we see them occurring. */
3115 if (x == 0)
3116 /* Look up ARG1 in the hash table and see if it has an equivalence
3117 that lets us see what is being compared. */
3118 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3119 if (p)
3121 p = p->first_same_value;
3123 /* If what we compare is already known to be constant, that is as
3124 good as it gets.
3125 We need to break the loop in this case, because otherwise we
3126 can have an infinite loop when looking at a reg that is known
3127 to be a constant which is the same as a comparison of a reg
3128 against zero which appears later in the insn stream, which in
3129 turn is constant and the same as the comparison of the first reg
3130 against zero... */
3131 if (p->is_const)
3132 break;
3135 for (; p; p = p->next_same_value)
3137 enum machine_mode inner_mode = GET_MODE (p->exp);
3138 #ifdef FLOAT_STORE_FLAG_VALUE
3139 REAL_VALUE_TYPE fsfv;
3140 #endif
3142 /* If the entry isn't valid, skip it. */
3143 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3144 continue;
3146 if (GET_CODE (p->exp) == COMPARE
3147 /* Another possibility is that this machine has a compare insn
3148 that includes the comparison code. In that case, ARG1 would
3149 be equivalent to a comparison operation that would set ARG1 to
3150 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3151 ORIG_CODE is the actual comparison being done; if it is an EQ,
3152 we must reverse ORIG_CODE. On machine with a negative value
3153 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3154 || ((code == NE
3155 || (code == LT
3156 && GET_MODE_CLASS (inner_mode) == MODE_INT
3157 && (GET_MODE_BITSIZE (inner_mode)
3158 <= HOST_BITS_PER_WIDE_INT)
3159 && (STORE_FLAG_VALUE
3160 & ((HOST_WIDE_INT) 1
3161 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3162 #ifdef FLOAT_STORE_FLAG_VALUE
3163 || (code == LT
3164 && SCALAR_FLOAT_MODE_P (inner_mode)
3165 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3166 REAL_VALUE_NEGATIVE (fsfv)))
3167 #endif
3169 && COMPARISON_P (p->exp)))
3171 x = p->exp;
3172 break;
3174 else if ((code == EQ
3175 || (code == GE
3176 && GET_MODE_CLASS (inner_mode) == MODE_INT
3177 && (GET_MODE_BITSIZE (inner_mode)
3178 <= HOST_BITS_PER_WIDE_INT)
3179 && (STORE_FLAG_VALUE
3180 & ((HOST_WIDE_INT) 1
3181 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3182 #ifdef FLOAT_STORE_FLAG_VALUE
3183 || (code == GE
3184 && SCALAR_FLOAT_MODE_P (inner_mode)
3185 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3186 REAL_VALUE_NEGATIVE (fsfv)))
3187 #endif
3189 && COMPARISON_P (p->exp))
3191 reverse_code = 1;
3192 x = p->exp;
3193 break;
3196 /* If this non-trapping address, e.g. fp + constant, the
3197 equivalent is a better operand since it may let us predict
3198 the value of the comparison. */
3199 else if (!rtx_addr_can_trap_p (p->exp))
3201 arg1 = p->exp;
3202 continue;
3206 /* If we didn't find a useful equivalence for ARG1, we are done.
3207 Otherwise, set up for the next iteration. */
3208 if (x == 0)
3209 break;
3211 /* If we need to reverse the comparison, make sure that that is
3212 possible -- we can't necessarily infer the value of GE from LT
3213 with floating-point operands. */
3214 if (reverse_code)
3216 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3217 if (reversed == UNKNOWN)
3218 break;
3219 else
3220 code = reversed;
3222 else if (COMPARISON_P (x))
3223 code = GET_CODE (x);
3224 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3227 /* Return our results. Return the modes from before fold_rtx
3228 because fold_rtx might produce const_int, and then it's too late. */
3229 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3230 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3232 return code;
3235 /* Fold SUBREG. */
3237 static rtx
3238 fold_rtx_subreg (rtx x, rtx insn)
3240 enum machine_mode mode = GET_MODE (x);
3241 rtx folded_arg0;
3242 rtx const_arg0;
3243 rtx new;
3245 /* See if we previously assigned a constant value to this SUBREG. */
3246 if ((new = lookup_as_function (x, CONST_INT)) != 0
3247 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3248 return new;
3250 /* If this is a paradoxical SUBREG, we have no idea what value the
3251 extra bits would have. However, if the operand is equivalent to
3252 a SUBREG whose operand is the same as our mode, and all the modes
3253 are within a word, we can just use the inner operand because
3254 these SUBREGs just say how to treat the register.
3256 Similarly if we find an integer constant. */
3258 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3260 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3261 struct table_elt *elt;
3263 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3264 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3265 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3266 imode)) != 0)
3267 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3269 if (CONSTANT_P (elt->exp)
3270 && GET_MODE (elt->exp) == VOIDmode)
3271 return elt->exp;
3273 if (GET_CODE (elt->exp) == SUBREG
3274 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3275 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3276 return copy_rtx (SUBREG_REG (elt->exp));
3279 return x;
3282 /* Fold SUBREG_REG. If it changed, see if we can simplify the
3283 SUBREG. We might be able to if the SUBREG is extracting a single
3284 word in an integral mode or extracting the low part. */
3286 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3287 const_arg0 = equiv_constant (folded_arg0);
3288 if (const_arg0)
3289 folded_arg0 = const_arg0;
3291 if (folded_arg0 != SUBREG_REG (x))
3293 new = simplify_subreg (mode, folded_arg0,
3294 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3295 if (new)
3296 return new;
3299 if (REG_P (folded_arg0)
3300 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3302 struct table_elt *elt;
3304 elt = lookup (folded_arg0,
3305 HASH (folded_arg0, GET_MODE (folded_arg0)),
3306 GET_MODE (folded_arg0));
3308 if (elt)
3309 elt = elt->first_same_value;
3311 if (subreg_lowpart_p (x))
3312 /* If this is a narrowing SUBREG and our operand is a REG, see
3313 if we can find an equivalence for REG that is an arithmetic
3314 operation in a wider mode where both operands are
3315 paradoxical SUBREGs from objects of our result mode. In
3316 that case, we couldn-t report an equivalent value for that
3317 operation, since we don't know what the extra bits will be.
3318 But we can find an equivalence for this SUBREG by folding
3319 that operation in the narrow mode. This allows us to fold
3320 arithmetic in narrow modes when the machine only supports
3321 word-sized arithmetic.
3323 Also look for a case where we have a SUBREG whose operand
3324 is the same as our result. If both modes are smaller than
3325 a word, we are simply interpreting a register in different
3326 modes and we can use the inner value. */
3328 for (; elt; elt = elt->next_same_value)
3330 enum rtx_code eltcode = GET_CODE (elt->exp);
3332 /* Just check for unary and binary operations. */
3333 if (UNARY_P (elt->exp)
3334 && eltcode != SIGN_EXTEND
3335 && eltcode != ZERO_EXTEND
3336 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3337 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3338 && (GET_MODE_CLASS (mode)
3339 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3341 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3343 if (!REG_P (op0) && ! CONSTANT_P (op0))
3344 op0 = fold_rtx (op0, NULL_RTX);
3346 op0 = equiv_constant (op0);
3347 if (op0)
3348 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3349 op0, mode);
3351 else if (ARITHMETIC_P (elt->exp)
3352 && eltcode != DIV && eltcode != MOD
3353 && eltcode != UDIV && eltcode != UMOD
3354 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3355 && eltcode != ROTATE && eltcode != ROTATERT
3356 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3357 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3358 == mode))
3359 || CONSTANT_P (XEXP (elt->exp, 0)))
3360 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3361 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3362 == mode))
3363 || CONSTANT_P (XEXP (elt->exp, 1))))
3365 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3366 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3368 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3369 op0 = fold_rtx (op0, NULL_RTX);
3371 if (op0)
3372 op0 = equiv_constant (op0);
3374 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3375 op1 = fold_rtx (op1, NULL_RTX);
3377 if (op1)
3378 op1 = equiv_constant (op1);
3380 /* If we are looking for the low SImode part of
3381 (ashift:DI c (const_int 32)), it doesn't work to
3382 compute that in SImode, because a 32-bit shift in
3383 SImode is unpredictable. We know the value is
3384 0. */
3385 if (op0 && op1
3386 && GET_CODE (elt->exp) == ASHIFT
3387 && GET_CODE (op1) == CONST_INT
3388 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3390 if (INTVAL (op1)
3391 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3392 /* If the count fits in the inner mode's width,
3393 but exceeds the outer mode's width, the value
3394 will get truncated to 0 by the subreg. */
3395 new = CONST0_RTX (mode);
3396 else
3397 /* If the count exceeds even the inner mode's width,
3398 don't fold this expression. */
3399 new = 0;
3401 else if (op0 && op1)
3402 new = simplify_binary_operation (GET_CODE (elt->exp),
3403 mode, op0, op1);
3406 else if (GET_CODE (elt->exp) == SUBREG
3407 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3408 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3409 <= UNITS_PER_WORD)
3410 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3411 new = copy_rtx (SUBREG_REG (elt->exp));
3413 if (new)
3414 return new;
3416 else
3417 /* A SUBREG resulting from a zero extension may fold to zero
3418 if it extracts higher bits than the ZERO_EXTEND's source
3419 bits. FIXME: if combine tried to, er, combine these
3420 instructions, this transformation may be moved to
3421 simplify_subreg. */
3422 for (; elt; elt = elt->next_same_value)
3424 if (GET_CODE (elt->exp) == ZERO_EXTEND
3425 && subreg_lsb (x)
3426 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3427 return CONST0_RTX (mode);
3431 return x;
3434 /* Fold MEM. */
3436 static rtx
3437 fold_rtx_mem (rtx x, rtx insn)
3439 enum machine_mode mode = GET_MODE (x);
3440 rtx new;
3442 /* If we are not actually processing an insn, don't try to find the
3443 best address. Not only don't we care, but we could modify the
3444 MEM in an invalid way since we have no insn to validate
3445 against. */
3446 if (insn != 0)
3447 find_best_addr (insn, &XEXP (x, 0), mode);
3450 /* Even if we don't fold in the insn itself, we can safely do so
3451 here, in hopes of getting a constant. */
3452 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3453 rtx base = 0;
3454 HOST_WIDE_INT offset = 0;
3456 if (REG_P (addr)
3457 && REGNO_QTY_VALID_P (REGNO (addr)))
3459 int addr_q = REG_QTY (REGNO (addr));
3460 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3462 if (GET_MODE (addr) == addr_ent->mode
3463 && addr_ent->const_rtx != NULL_RTX)
3464 addr = addr_ent->const_rtx;
3467 /* Call target hook to avoid the effects of -fpic etc.... */
3468 addr = targetm.delegitimize_address (addr);
3470 /* If address is constant, split it into a base and integer
3471 offset. */
3472 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3473 base = addr;
3474 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3475 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3477 base = XEXP (XEXP (addr, 0), 0);
3478 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3480 else if (GET_CODE (addr) == LO_SUM
3481 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3482 base = XEXP (addr, 1);
3484 /* If this is a constant pool reference, we can fold it into its
3485 constant to allow better value tracking. */
3486 if (base && GET_CODE (base) == SYMBOL_REF
3487 && CONSTANT_POOL_ADDRESS_P (base))
3489 rtx constant = get_pool_constant (base);
3490 enum machine_mode const_mode = get_pool_mode (base);
3491 rtx new;
3493 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3495 constant_pool_entries_cost = COST (constant);
3496 constant_pool_entries_regcost = approx_reg_cost (constant);
3499 /* If we are loading the full constant, we have an
3500 equivalence. */
3501 if (offset == 0 && mode == const_mode)
3502 return constant;
3504 /* If this actually isn't a constant (weird!), we can't do
3505 anything. Otherwise, handle the two most common cases:
3506 extracting a word from a multi-word constant, and
3507 extracting the low-order bits. Other cases don't seem
3508 common enough to worry about. */
3509 if (! CONSTANT_P (constant))
3510 return x;
3512 if (GET_MODE_CLASS (mode) == MODE_INT
3513 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3514 && offset % UNITS_PER_WORD == 0
3515 && (new = operand_subword (constant,
3516 offset / UNITS_PER_WORD,
3517 0, const_mode)) != 0)
3518 return new;
3520 if (((BYTES_BIG_ENDIAN
3521 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3522 || (! BYTES_BIG_ENDIAN && offset == 0))
3523 && (new = gen_lowpart (mode, constant)) != 0)
3524 return new;
3527 /* If this is a reference to a label at a known position in a jump
3528 table, we also know its value. */
3529 if (base && GET_CODE (base) == LABEL_REF)
3531 rtx label = XEXP (base, 0);
3532 rtx table_insn = NEXT_INSN (label);
3534 if (table_insn && JUMP_P (table_insn)
3535 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3537 rtx table = PATTERN (table_insn);
3539 if (offset >= 0
3540 && (offset / GET_MODE_SIZE (GET_MODE (table))
3541 < XVECLEN (table, 0)))
3543 rtx label = XVECEXP
3544 (table, 0, offset / GET_MODE_SIZE (GET_MODE (table)));
3545 rtx set;
3547 /* If we have an insn that loads the label from the
3548 jumptable into a reg, we don't want to set the reg
3549 to the label, because this may cause a reference to
3550 the label to remain after the label is removed in
3551 some very obscure cases (PR middle-end/18628). */
3552 if (!insn)
3553 return label;
3555 set = single_set (insn);
3557 if (! set || SET_SRC (set) != x)
3558 return x;
3560 /* If it's a jump, it's safe to reference the label. */
3561 if (SET_DEST (set) == pc_rtx)
3562 return label;
3564 return x;
3567 if (table_insn && JUMP_P (table_insn)
3568 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3570 rtx table = PATTERN (table_insn);
3572 if (offset >= 0
3573 && (offset / GET_MODE_SIZE (GET_MODE (table))
3574 < XVECLEN (table, 1)))
3576 offset /= GET_MODE_SIZE (GET_MODE (table));
3577 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3578 XEXP (table, 0));
3580 if (GET_MODE (table) != Pmode)
3581 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3583 /* Indicate this is a constant. This isn't a valid
3584 form of CONST, but it will only be used to fold the
3585 next insns and then discarded, so it should be
3586 safe.
3588 Note this expression must be explicitly discarded,
3589 by cse_insn, else it may end up in a REG_EQUAL note
3590 and "escape" to cause problems elsewhere. */
3591 return gen_rtx_CONST (GET_MODE (new), new);
3596 return x;
3600 /* If X is a nontrivial arithmetic operation on an argument
3601 for which a constant value can be determined, return
3602 the result of operating on that value, as a constant.
3603 Otherwise, return X, possibly with one or more operands
3604 modified by recursive calls to this function.
3606 If X is a register whose contents are known, we do NOT
3607 return those contents here. equiv_constant is called to
3608 perform that task.
3610 INSN is the insn that we may be modifying. If it is 0, make a copy
3611 of X before modifying it. */
3613 static rtx
3614 fold_rtx (rtx x, rtx insn)
3616 enum rtx_code code;
3617 enum machine_mode mode;
3618 const char *fmt;
3619 int i;
3620 rtx new = 0;
3621 int copied = 0;
3622 int must_swap = 0;
3624 /* Folded equivalents of first two operands of X. */
3625 rtx folded_arg0;
3626 rtx folded_arg1;
3628 /* Constant equivalents of first three operands of X;
3629 0 when no such equivalent is known. */
3630 rtx const_arg0;
3631 rtx const_arg1;
3632 rtx const_arg2;
3634 /* The mode of the first operand of X. We need this for sign and zero
3635 extends. */
3636 enum machine_mode mode_arg0;
3638 if (x == 0)
3639 return x;
3641 mode = GET_MODE (x);
3642 code = GET_CODE (x);
3643 switch (code)
3645 case CONST:
3646 case CONST_INT:
3647 case CONST_DOUBLE:
3648 case CONST_VECTOR:
3649 case SYMBOL_REF:
3650 case LABEL_REF:
3651 case REG:
3652 case PC:
3653 /* No use simplifying an EXPR_LIST
3654 since they are used only for lists of args
3655 in a function call's REG_EQUAL note. */
3656 case EXPR_LIST:
3657 return x;
3659 #ifdef HAVE_cc0
3660 case CC0:
3661 return prev_insn_cc0;
3662 #endif
3664 case SUBREG:
3665 return fold_rtx_subreg (x, insn);
3667 case NOT:
3668 case NEG:
3669 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3670 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3671 new = lookup_as_function (XEXP (x, 0), code);
3672 if (new)
3673 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3674 break;
3676 case MEM:
3677 return fold_rtx_mem (x, insn);
3679 #ifdef NO_FUNCTION_CSE
3680 case CALL:
3681 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3682 return x;
3683 break;
3684 #endif
3686 case ASM_OPERANDS:
3687 if (insn)
3689 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3690 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3691 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3693 break;
3695 default:
3696 break;
3699 const_arg0 = 0;
3700 const_arg1 = 0;
3701 const_arg2 = 0;
3702 mode_arg0 = VOIDmode;
3704 /* Try folding our operands.
3705 Then see which ones have constant values known. */
3707 fmt = GET_RTX_FORMAT (code);
3708 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3709 if (fmt[i] == 'e')
3711 rtx arg = XEXP (x, i);
3712 rtx folded_arg = arg, const_arg = 0;
3713 enum machine_mode mode_arg = GET_MODE (arg);
3714 rtx cheap_arg, expensive_arg;
3715 rtx replacements[2];
3716 int j;
3717 int old_cost = COST_IN (XEXP (x, i), code);
3719 /* Most arguments are cheap, so handle them specially. */
3720 switch (GET_CODE (arg))
3722 case REG:
3723 /* This is the same as calling equiv_constant; it is duplicated
3724 here for speed. */
3725 if (REGNO_QTY_VALID_P (REGNO (arg)))
3727 int arg_q = REG_QTY (REGNO (arg));
3728 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3730 if (arg_ent->const_rtx != NULL_RTX
3731 && !REG_P (arg_ent->const_rtx)
3732 && GET_CODE (arg_ent->const_rtx) != PLUS)
3733 const_arg
3734 = gen_lowpart (GET_MODE (arg),
3735 arg_ent->const_rtx);
3737 break;
3739 case CONST:
3740 case CONST_INT:
3741 case SYMBOL_REF:
3742 case LABEL_REF:
3743 case CONST_DOUBLE:
3744 case CONST_VECTOR:
3745 const_arg = arg;
3746 break;
3748 #ifdef HAVE_cc0
3749 case CC0:
3750 folded_arg = prev_insn_cc0;
3751 mode_arg = prev_insn_cc0_mode;
3752 const_arg = equiv_constant (folded_arg);
3753 break;
3754 #endif
3756 default:
3757 folded_arg = fold_rtx (arg, insn);
3758 const_arg = equiv_constant (folded_arg);
3761 /* For the first three operands, see if the operand
3762 is constant or equivalent to a constant. */
3763 switch (i)
3765 case 0:
3766 folded_arg0 = folded_arg;
3767 const_arg0 = const_arg;
3768 mode_arg0 = mode_arg;
3769 break;
3770 case 1:
3771 folded_arg1 = folded_arg;
3772 const_arg1 = const_arg;
3773 break;
3774 case 2:
3775 const_arg2 = const_arg;
3776 break;
3779 /* Pick the least expensive of the folded argument and an
3780 equivalent constant argument. */
3781 if (const_arg == 0 || const_arg == folded_arg
3782 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3783 cheap_arg = folded_arg, expensive_arg = const_arg;
3784 else
3785 cheap_arg = const_arg, expensive_arg = folded_arg;
3787 /* Try to replace the operand with the cheapest of the two
3788 possibilities. If it doesn't work and this is either of the first
3789 two operands of a commutative operation, try swapping them.
3790 If THAT fails, try the more expensive, provided it is cheaper
3791 than what is already there. */
3793 if (cheap_arg == XEXP (x, i))
3794 continue;
3796 if (insn == 0 && ! copied)
3798 x = copy_rtx (x);
3799 copied = 1;
3802 /* Order the replacements from cheapest to most expensive. */
3803 replacements[0] = cheap_arg;
3804 replacements[1] = expensive_arg;
3806 for (j = 0; j < 2 && replacements[j]; j++)
3808 int new_cost = COST_IN (replacements[j], code);
3810 /* Stop if what existed before was cheaper. Prefer constants
3811 in the case of a tie. */
3812 if (new_cost > old_cost
3813 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3814 break;
3816 /* It's not safe to substitute the operand of a conversion
3817 operator with a constant, as the conversion's identity
3818 depends upon the mode of its operand. This optimization
3819 is handled by the call to simplify_unary_operation. */
3820 if (GET_RTX_CLASS (code) == RTX_UNARY
3821 && GET_MODE (replacements[j]) != mode_arg0
3822 && (code == ZERO_EXTEND
3823 || code == SIGN_EXTEND
3824 || code == TRUNCATE
3825 || code == FLOAT_TRUNCATE
3826 || code == FLOAT_EXTEND
3827 || code == FLOAT
3828 || code == FIX
3829 || code == UNSIGNED_FLOAT
3830 || code == UNSIGNED_FIX))
3831 continue;
3833 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3834 break;
3836 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3837 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3839 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3840 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3842 if (apply_change_group ())
3844 /* Swap them back to be invalid so that this loop can
3845 continue and flag them to be swapped back later. */
3846 rtx tem;
3848 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3849 XEXP (x, 1) = tem;
3850 must_swap = 1;
3851 break;
3857 else
3859 if (fmt[i] == 'E')
3860 /* Don't try to fold inside of a vector of expressions.
3861 Doing nothing is harmless. */
3865 /* If a commutative operation, place a constant integer as the second
3866 operand unless the first operand is also a constant integer. Otherwise,
3867 place any constant second unless the first operand is also a constant. */
3869 if (COMMUTATIVE_P (x))
3871 if (must_swap
3872 || swap_commutative_operands_p (const_arg0 ? const_arg0
3873 : XEXP (x, 0),
3874 const_arg1 ? const_arg1
3875 : XEXP (x, 1)))
3877 rtx tem = XEXP (x, 0);
3879 if (insn == 0 && ! copied)
3881 x = copy_rtx (x);
3882 copied = 1;
3885 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3886 validate_change (insn, &XEXP (x, 1), tem, 1);
3887 if (apply_change_group ())
3889 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3890 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3895 /* If X is an arithmetic operation, see if we can simplify it. */
3897 switch (GET_RTX_CLASS (code))
3899 case RTX_UNARY:
3901 int is_const = 0;
3903 /* We can't simplify extension ops unless we know the
3904 original mode. */
3905 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3906 && mode_arg0 == VOIDmode)
3907 break;
3909 /* If we had a CONST, strip it off and put it back later if we
3910 fold. */
3911 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3912 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3914 new = simplify_unary_operation (code, mode,
3915 const_arg0 ? const_arg0 : folded_arg0,
3916 mode_arg0);
3917 /* NEG of PLUS could be converted into MINUS, but that causes
3918 expressions of the form
3919 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3920 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3921 FIXME: those ports should be fixed. */
3922 if (new != 0 && is_const
3923 && GET_CODE (new) == PLUS
3924 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3925 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3926 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3927 new = gen_rtx_CONST (mode, new);
3929 break;
3931 case RTX_COMPARE:
3932 case RTX_COMM_COMPARE:
3933 /* See what items are actually being compared and set FOLDED_ARG[01]
3934 to those values and CODE to the actual comparison code. If any are
3935 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3936 do anything if both operands are already known to be constant. */
3938 /* ??? Vector mode comparisons are not supported yet. */
3939 if (VECTOR_MODE_P (mode))
3940 break;
3942 if (const_arg0 == 0 || const_arg1 == 0)
3944 struct table_elt *p0, *p1;
3945 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3946 enum machine_mode mode_arg1;
3948 #ifdef FLOAT_STORE_FLAG_VALUE
3949 if (SCALAR_FLOAT_MODE_P (mode))
3951 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3952 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3953 false_rtx = CONST0_RTX (mode);
3955 #endif
3957 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3958 &mode_arg0, &mode_arg1);
3960 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3961 what kinds of things are being compared, so we can't do
3962 anything with this comparison. */
3964 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3965 break;
3967 const_arg0 = equiv_constant (folded_arg0);
3968 const_arg1 = equiv_constant (folded_arg1);
3970 /* If we do not now have two constants being compared, see
3971 if we can nevertheless deduce some things about the
3972 comparison. */
3973 if (const_arg0 == 0 || const_arg1 == 0)
3975 /* Some addresses are known to be nonzero. We don't know
3976 their sign, but equality comparisons are known. */
3977 if (const_arg1 == const0_rtx
3978 && nonzero_address_p (folded_arg0))
3980 if (code == EQ)
3981 return false_rtx;
3982 else if (code == NE)
3983 return true_rtx;
3986 /* See if the two operands are the same. */
3988 if (folded_arg0 == folded_arg1
3989 || (REG_P (folded_arg0)
3990 && REG_P (folded_arg1)
3991 && (REG_QTY (REGNO (folded_arg0))
3992 == REG_QTY (REGNO (folded_arg1))))
3993 || ((p0 = lookup (folded_arg0,
3994 SAFE_HASH (folded_arg0, mode_arg0),
3995 mode_arg0))
3996 && (p1 = lookup (folded_arg1,
3997 SAFE_HASH (folded_arg1, mode_arg0),
3998 mode_arg0))
3999 && p0->first_same_value == p1->first_same_value))
4001 /* Sadly two equal NaNs are not equivalent. */
4002 if (!HONOR_NANS (mode_arg0))
4003 return ((code == EQ || code == LE || code == GE
4004 || code == LEU || code == GEU || code == UNEQ
4005 || code == UNLE || code == UNGE
4006 || code == ORDERED)
4007 ? true_rtx : false_rtx);
4008 /* Take care for the FP compares we can resolve. */
4009 if (code == UNEQ || code == UNLE || code == UNGE)
4010 return true_rtx;
4011 if (code == LTGT || code == LT || code == GT)
4012 return false_rtx;
4015 /* If FOLDED_ARG0 is a register, see if the comparison we are
4016 doing now is either the same as we did before or the reverse
4017 (we only check the reverse if not floating-point). */
4018 else if (REG_P (folded_arg0))
4020 int qty = REG_QTY (REGNO (folded_arg0));
4022 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4024 struct qty_table_elem *ent = &qty_table[qty];
4026 if ((comparison_dominates_p (ent->comparison_code, code)
4027 || (! FLOAT_MODE_P (mode_arg0)
4028 && comparison_dominates_p (ent->comparison_code,
4029 reverse_condition (code))))
4030 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4031 || (const_arg1
4032 && rtx_equal_p (ent->comparison_const,
4033 const_arg1))
4034 || (REG_P (folded_arg1)
4035 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4036 return (comparison_dominates_p (ent->comparison_code, code)
4037 ? true_rtx : false_rtx);
4043 /* If we are comparing against zero, see if the first operand is
4044 equivalent to an IOR with a constant. If so, we may be able to
4045 determine the result of this comparison. */
4047 if (const_arg1 == const0_rtx)
4049 rtx y = lookup_as_function (folded_arg0, IOR);
4050 rtx inner_const;
4052 if (y != 0
4053 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4054 && GET_CODE (inner_const) == CONST_INT
4055 && INTVAL (inner_const) != 0)
4057 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4058 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4059 && (INTVAL (inner_const)
4060 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4061 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4063 #ifdef FLOAT_STORE_FLAG_VALUE
4064 if (SCALAR_FLOAT_MODE_P (mode))
4066 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4067 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4068 false_rtx = CONST0_RTX (mode);
4070 #endif
4072 switch (code)
4074 case EQ:
4075 return false_rtx;
4076 case NE:
4077 return true_rtx;
4078 case LT: case LE:
4079 if (has_sign)
4080 return true_rtx;
4081 break;
4082 case GT: case GE:
4083 if (has_sign)
4084 return false_rtx;
4085 break;
4086 default:
4087 break;
4093 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4094 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4095 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4097 break;
4099 case RTX_BIN_ARITH:
4100 case RTX_COMM_ARITH:
4101 switch (code)
4103 case PLUS:
4104 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4105 with that LABEL_REF as its second operand. If so, the result is
4106 the first operand of that MINUS. This handles switches with an
4107 ADDR_DIFF_VEC table. */
4108 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4110 rtx y
4111 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4112 : lookup_as_function (folded_arg0, MINUS);
4114 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4115 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4116 return XEXP (y, 0);
4118 /* Now try for a CONST of a MINUS like the above. */
4119 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4120 : lookup_as_function (folded_arg0, CONST))) != 0
4121 && GET_CODE (XEXP (y, 0)) == MINUS
4122 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4123 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4124 return XEXP (XEXP (y, 0), 0);
4127 /* Likewise if the operands are in the other order. */
4128 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4130 rtx y
4131 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4132 : lookup_as_function (folded_arg1, MINUS);
4134 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4135 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4136 return XEXP (y, 0);
4138 /* Now try for a CONST of a MINUS like the above. */
4139 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4140 : lookup_as_function (folded_arg1, CONST))) != 0
4141 && GET_CODE (XEXP (y, 0)) == MINUS
4142 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4143 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4144 return XEXP (XEXP (y, 0), 0);
4147 /* If second operand is a register equivalent to a negative
4148 CONST_INT, see if we can find a register equivalent to the
4149 positive constant. Make a MINUS if so. Don't do this for
4150 a non-negative constant since we might then alternate between
4151 choosing positive and negative constants. Having the positive
4152 constant previously-used is the more common case. Be sure
4153 the resulting constant is non-negative; if const_arg1 were
4154 the smallest negative number this would overflow: depending
4155 on the mode, this would either just be the same value (and
4156 hence not save anything) or be incorrect. */
4157 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4158 && INTVAL (const_arg1) < 0
4159 /* This used to test
4161 -INTVAL (const_arg1) >= 0
4163 But The Sun V5.0 compilers mis-compiled that test. So
4164 instead we test for the problematic value in a more direct
4165 manner and hope the Sun compilers get it correct. */
4166 && INTVAL (const_arg1) !=
4167 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4168 && REG_P (folded_arg1))
4170 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4171 struct table_elt *p
4172 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4174 if (p)
4175 for (p = p->first_same_value; p; p = p->next_same_value)
4176 if (REG_P (p->exp))
4177 return simplify_gen_binary (MINUS, mode, folded_arg0,
4178 canon_reg (p->exp, NULL_RTX));
4180 goto from_plus;
4182 case MINUS:
4183 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4184 If so, produce (PLUS Z C2-C). */
4185 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4187 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4188 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4189 return fold_rtx (plus_constant (copy_rtx (y),
4190 -INTVAL (const_arg1)),
4191 NULL_RTX);
4194 /* Fall through. */
4196 from_plus:
4197 case SMIN: case SMAX: case UMIN: case UMAX:
4198 case IOR: case AND: case XOR:
4199 case MULT:
4200 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4201 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4202 is known to be of similar form, we may be able to replace the
4203 operation with a combined operation. This may eliminate the
4204 intermediate operation if every use is simplified in this way.
4205 Note that the similar optimization done by combine.c only works
4206 if the intermediate operation's result has only one reference. */
4208 if (REG_P (folded_arg0)
4209 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4211 int is_shift
4212 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4213 rtx y = lookup_as_function (folded_arg0, code);
4214 rtx inner_const;
4215 enum rtx_code associate_code;
4216 rtx new_const;
4218 if (y == 0
4219 || 0 == (inner_const
4220 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4221 || GET_CODE (inner_const) != CONST_INT
4222 /* If we have compiled a statement like
4223 "if (x == (x & mask1))", and now are looking at
4224 "x & mask2", we will have a case where the first operand
4225 of Y is the same as our first operand. Unless we detect
4226 this case, an infinite loop will result. */
4227 || XEXP (y, 0) == folded_arg0)
4228 break;
4230 /* Don't associate these operations if they are a PLUS with the
4231 same constant and it is a power of two. These might be doable
4232 with a pre- or post-increment. Similarly for two subtracts of
4233 identical powers of two with post decrement. */
4235 if (code == PLUS && const_arg1 == inner_const
4236 && ((HAVE_PRE_INCREMENT
4237 && exact_log2 (INTVAL (const_arg1)) >= 0)
4238 || (HAVE_POST_INCREMENT
4239 && exact_log2 (INTVAL (const_arg1)) >= 0)
4240 || (HAVE_PRE_DECREMENT
4241 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4242 || (HAVE_POST_DECREMENT
4243 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4244 break;
4246 /* Compute the code used to compose the constants. For example,
4247 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4249 associate_code = (is_shift || code == MINUS ? PLUS : code);
4251 new_const = simplify_binary_operation (associate_code, mode,
4252 const_arg1, inner_const);
4254 if (new_const == 0)
4255 break;
4257 /* If we are associating shift operations, don't let this
4258 produce a shift of the size of the object or larger.
4259 This could occur when we follow a sign-extend by a right
4260 shift on a machine that does a sign-extend as a pair
4261 of shifts. */
4263 if (is_shift && GET_CODE (new_const) == CONST_INT
4264 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4266 /* As an exception, we can turn an ASHIFTRT of this
4267 form into a shift of the number of bits - 1. */
4268 if (code == ASHIFTRT)
4269 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4270 else
4271 break;
4274 y = copy_rtx (XEXP (y, 0));
4276 /* If Y contains our first operand (the most common way this
4277 can happen is if Y is a MEM), we would do into an infinite
4278 loop if we tried to fold it. So don't in that case. */
4280 if (! reg_mentioned_p (folded_arg0, y))
4281 y = fold_rtx (y, insn);
4283 return simplify_gen_binary (code, mode, y, new_const);
4285 break;
4287 case DIV: case UDIV:
4288 /* ??? The associative optimization performed immediately above is
4289 also possible for DIV and UDIV using associate_code of MULT.
4290 However, we would need extra code to verify that the
4291 multiplication does not overflow, that is, there is no overflow
4292 in the calculation of new_const. */
4293 break;
4295 default:
4296 break;
4299 new = simplify_binary_operation (code, mode,
4300 const_arg0 ? const_arg0 : folded_arg0,
4301 const_arg1 ? const_arg1 : folded_arg1);
4302 break;
4304 case RTX_OBJ:
4305 /* (lo_sum (high X) X) is simply X. */
4306 if (code == LO_SUM && const_arg0 != 0
4307 && GET_CODE (const_arg0) == HIGH
4308 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4309 return const_arg1;
4310 break;
4312 case RTX_TERNARY:
4313 case RTX_BITFIELD_OPS:
4314 new = simplify_ternary_operation (code, mode, mode_arg0,
4315 const_arg0 ? const_arg0 : folded_arg0,
4316 const_arg1 ? const_arg1 : folded_arg1,
4317 const_arg2 ? const_arg2 : XEXP (x, 2));
4318 break;
4320 default:
4321 break;
4324 return new ? new : x;
4327 /* Return a constant value currently equivalent to X.
4328 Return 0 if we don't know one. */
4330 static rtx
4331 equiv_constant (rtx x)
4333 if (REG_P (x)
4334 && REGNO_QTY_VALID_P (REGNO (x)))
4336 int x_q = REG_QTY (REGNO (x));
4337 struct qty_table_elem *x_ent = &qty_table[x_q];
4339 if (x_ent->const_rtx)
4340 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4343 if (x == 0 || CONSTANT_P (x))
4344 return x;
4346 /* If X is a MEM, try to fold it outside the context of any insn to see if
4347 it might be equivalent to a constant. That handles the case where it
4348 is a constant-pool reference. Then try to look it up in the hash table
4349 in case it is something whose value we have seen before. */
4351 if (MEM_P (x))
4353 struct table_elt *elt;
4355 x = fold_rtx (x, NULL_RTX);
4356 if (CONSTANT_P (x))
4357 return x;
4359 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4360 if (elt == 0)
4361 return 0;
4363 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4364 if (elt->is_const && CONSTANT_P (elt->exp))
4365 return elt->exp;
4368 return 0;
4371 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4372 branch. It will be zero if not.
4374 In certain cases, this can cause us to add an equivalence. For example,
4375 if we are following the taken case of
4376 if (i == 2)
4377 we can add the fact that `i' and '2' are now equivalent.
4379 In any case, we can record that this comparison was passed. If the same
4380 comparison is seen later, we will know its value. */
4382 static void
4383 record_jump_equiv (rtx insn, int taken)
4385 int cond_known_true;
4386 rtx op0, op1;
4387 rtx set;
4388 enum machine_mode mode, mode0, mode1;
4389 int reversed_nonequality = 0;
4390 enum rtx_code code;
4392 /* Ensure this is the right kind of insn. */
4393 if (! any_condjump_p (insn))
4394 return;
4395 set = pc_set (insn);
4397 /* See if this jump condition is known true or false. */
4398 if (taken)
4399 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4400 else
4401 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4403 /* Get the type of comparison being done and the operands being compared.
4404 If we had to reverse a non-equality condition, record that fact so we
4405 know that it isn't valid for floating-point. */
4406 code = GET_CODE (XEXP (SET_SRC (set), 0));
4407 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4408 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4410 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4411 if (! cond_known_true)
4413 code = reversed_comparison_code_parts (code, op0, op1, insn);
4415 /* Don't remember if we can't find the inverse. */
4416 if (code == UNKNOWN)
4417 return;
4420 /* The mode is the mode of the non-constant. */
4421 mode = mode0;
4422 if (mode1 != VOIDmode)
4423 mode = mode1;
4425 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4428 /* Yet another form of subreg creation. In this case, we want something in
4429 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4431 static rtx
4432 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4434 enum machine_mode op_mode = GET_MODE (op);
4435 if (op_mode == mode || op_mode == VOIDmode)
4436 return op;
4437 return lowpart_subreg (mode, op, op_mode);
4440 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4441 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4442 Make any useful entries we can with that information. Called from
4443 above function and called recursively. */
4445 static void
4446 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4447 rtx op1, int reversed_nonequality)
4449 unsigned op0_hash, op1_hash;
4450 int op0_in_memory, op1_in_memory;
4451 struct table_elt *op0_elt, *op1_elt;
4453 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4454 we know that they are also equal in the smaller mode (this is also
4455 true for all smaller modes whether or not there is a SUBREG, but
4456 is not worth testing for with no SUBREG). */
4458 /* Note that GET_MODE (op0) may not equal MODE. */
4459 if (code == EQ && GET_CODE (op0) == SUBREG
4460 && (GET_MODE_SIZE (GET_MODE (op0))
4461 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4463 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4464 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4465 if (tem)
4466 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4467 reversed_nonequality);
4470 if (code == EQ && GET_CODE (op1) == SUBREG
4471 && (GET_MODE_SIZE (GET_MODE (op1))
4472 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4474 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4475 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4476 if (tem)
4477 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4478 reversed_nonequality);
4481 /* Similarly, if this is an NE comparison, and either is a SUBREG
4482 making a smaller mode, we know the whole thing is also NE. */
4484 /* Note that GET_MODE (op0) may not equal MODE;
4485 if we test MODE instead, we can get an infinite recursion
4486 alternating between two modes each wider than MODE. */
4488 if (code == NE && GET_CODE (op0) == SUBREG
4489 && subreg_lowpart_p (op0)
4490 && (GET_MODE_SIZE (GET_MODE (op0))
4491 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4493 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4494 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4495 if (tem)
4496 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4497 reversed_nonequality);
4500 if (code == NE && GET_CODE (op1) == SUBREG
4501 && subreg_lowpart_p (op1)
4502 && (GET_MODE_SIZE (GET_MODE (op1))
4503 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4505 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4506 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4507 if (tem)
4508 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4509 reversed_nonequality);
4512 /* Hash both operands. */
4514 do_not_record = 0;
4515 hash_arg_in_memory = 0;
4516 op0_hash = HASH (op0, mode);
4517 op0_in_memory = hash_arg_in_memory;
4519 if (do_not_record)
4520 return;
4522 do_not_record = 0;
4523 hash_arg_in_memory = 0;
4524 op1_hash = HASH (op1, mode);
4525 op1_in_memory = hash_arg_in_memory;
4527 if (do_not_record)
4528 return;
4530 /* Look up both operands. */
4531 op0_elt = lookup (op0, op0_hash, mode);
4532 op1_elt = lookup (op1, op1_hash, mode);
4534 /* If both operands are already equivalent or if they are not in the
4535 table but are identical, do nothing. */
4536 if ((op0_elt != 0 && op1_elt != 0
4537 && op0_elt->first_same_value == op1_elt->first_same_value)
4538 || op0 == op1 || rtx_equal_p (op0, op1))
4539 return;
4541 /* If we aren't setting two things equal all we can do is save this
4542 comparison. Similarly if this is floating-point. In the latter
4543 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4544 If we record the equality, we might inadvertently delete code
4545 whose intent was to change -0 to +0. */
4547 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4549 struct qty_table_elem *ent;
4550 int qty;
4552 /* If we reversed a floating-point comparison, if OP0 is not a
4553 register, or if OP1 is neither a register or constant, we can't
4554 do anything. */
4556 if (!REG_P (op1))
4557 op1 = equiv_constant (op1);
4559 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4560 || !REG_P (op0) || op1 == 0)
4561 return;
4563 /* Put OP0 in the hash table if it isn't already. This gives it a
4564 new quantity number. */
4565 if (op0_elt == 0)
4567 if (insert_regs (op0, NULL, 0))
4569 rehash_using_reg (op0);
4570 op0_hash = HASH (op0, mode);
4572 /* If OP0 is contained in OP1, this changes its hash code
4573 as well. Faster to rehash than to check, except
4574 for the simple case of a constant. */
4575 if (! CONSTANT_P (op1))
4576 op1_hash = HASH (op1,mode);
4579 op0_elt = insert (op0, NULL, op0_hash, mode);
4580 op0_elt->in_memory = op0_in_memory;
4583 qty = REG_QTY (REGNO (op0));
4584 ent = &qty_table[qty];
4586 ent->comparison_code = code;
4587 if (REG_P (op1))
4589 /* Look it up again--in case op0 and op1 are the same. */
4590 op1_elt = lookup (op1, op1_hash, mode);
4592 /* Put OP1 in the hash table so it gets a new quantity number. */
4593 if (op1_elt == 0)
4595 if (insert_regs (op1, NULL, 0))
4597 rehash_using_reg (op1);
4598 op1_hash = HASH (op1, mode);
4601 op1_elt = insert (op1, NULL, op1_hash, mode);
4602 op1_elt->in_memory = op1_in_memory;
4605 ent->comparison_const = NULL_RTX;
4606 ent->comparison_qty = REG_QTY (REGNO (op1));
4608 else
4610 ent->comparison_const = op1;
4611 ent->comparison_qty = -1;
4614 return;
4617 /* If either side is still missing an equivalence, make it now,
4618 then merge the equivalences. */
4620 if (op0_elt == 0)
4622 if (insert_regs (op0, NULL, 0))
4624 rehash_using_reg (op0);
4625 op0_hash = HASH (op0, mode);
4628 op0_elt = insert (op0, NULL, op0_hash, mode);
4629 op0_elt->in_memory = op0_in_memory;
4632 if (op1_elt == 0)
4634 if (insert_regs (op1, NULL, 0))
4636 rehash_using_reg (op1);
4637 op1_hash = HASH (op1, mode);
4640 op1_elt = insert (op1, NULL, op1_hash, mode);
4641 op1_elt->in_memory = op1_in_memory;
4644 merge_equiv_classes (op0_elt, op1_elt);
4647 /* CSE processing for one instruction.
4648 First simplify sources and addresses of all assignments
4649 in the instruction, using previously-computed equivalents values.
4650 Then install the new sources and destinations in the table
4651 of available values.
4653 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4654 the insn. It means that INSN is inside libcall block. In this
4655 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4657 /* Data on one SET contained in the instruction. */
4659 struct set
4661 /* The SET rtx itself. */
4662 rtx rtl;
4663 /* The SET_SRC of the rtx (the original value, if it is changing). */
4664 rtx src;
4665 /* The hash-table element for the SET_SRC of the SET. */
4666 struct table_elt *src_elt;
4667 /* Hash value for the SET_SRC. */
4668 unsigned src_hash;
4669 /* Hash value for the SET_DEST. */
4670 unsigned dest_hash;
4671 /* The SET_DEST, with SUBREG, etc., stripped. */
4672 rtx inner_dest;
4673 /* Nonzero if the SET_SRC is in memory. */
4674 char src_in_memory;
4675 /* Nonzero if the SET_SRC contains something
4676 whose value cannot be predicted and understood. */
4677 char src_volatile;
4678 /* Original machine mode, in case it becomes a CONST_INT.
4679 The size of this field should match the size of the mode
4680 field of struct rtx_def (see rtl.h). */
4681 ENUM_BITFIELD(machine_mode) mode : 8;
4682 /* A constant equivalent for SET_SRC, if any. */
4683 rtx src_const;
4684 /* Original SET_SRC value used for libcall notes. */
4685 rtx orig_src;
4686 /* Hash value of constant equivalent for SET_SRC. */
4687 unsigned src_const_hash;
4688 /* Table entry for constant equivalent for SET_SRC, if any. */
4689 struct table_elt *src_const_elt;
4692 static void
4693 cse_insn (rtx insn, rtx libcall_insn)
4695 rtx x = PATTERN (insn);
4696 int i;
4697 rtx tem;
4698 int n_sets = 0;
4700 #ifdef HAVE_cc0
4701 /* Records what this insn does to set CC0. */
4702 rtx this_insn_cc0 = 0;
4703 enum machine_mode this_insn_cc0_mode = VOIDmode;
4704 #endif
4706 rtx src_eqv = 0;
4707 struct table_elt *src_eqv_elt = 0;
4708 int src_eqv_volatile = 0;
4709 int src_eqv_in_memory = 0;
4710 unsigned src_eqv_hash = 0;
4712 struct set *sets = (struct set *) 0;
4714 this_insn = insn;
4716 /* Find all the SETs and CLOBBERs in this instruction.
4717 Record all the SETs in the array `set' and count them.
4718 Also determine whether there is a CLOBBER that invalidates
4719 all memory references, or all references at varying addresses. */
4721 if (CALL_P (insn))
4723 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4725 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4726 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4727 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4731 if (GET_CODE (x) == SET)
4733 sets = alloca (sizeof (struct set));
4734 sets[0].rtl = x;
4736 /* Ignore SETs that are unconditional jumps.
4737 They never need cse processing, so this does not hurt.
4738 The reason is not efficiency but rather
4739 so that we can test at the end for instructions
4740 that have been simplified to unconditional jumps
4741 and not be misled by unchanged instructions
4742 that were unconditional jumps to begin with. */
4743 if (SET_DEST (x) == pc_rtx
4744 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4747 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4748 The hard function value register is used only once, to copy to
4749 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4750 Ensure we invalidate the destination register. On the 80386 no
4751 other code would invalidate it since it is a fixed_reg.
4752 We need not check the return of apply_change_group; see canon_reg. */
4754 else if (GET_CODE (SET_SRC (x)) == CALL)
4756 canon_reg (SET_SRC (x), insn);
4757 apply_change_group ();
4758 fold_rtx (SET_SRC (x), insn);
4759 invalidate (SET_DEST (x), VOIDmode);
4761 else
4762 n_sets = 1;
4764 else if (GET_CODE (x) == PARALLEL)
4766 int lim = XVECLEN (x, 0);
4768 sets = alloca (lim * sizeof (struct set));
4770 /* Find all regs explicitly clobbered in this insn,
4771 and ensure they are not replaced with any other regs
4772 elsewhere in this insn.
4773 When a reg that is clobbered is also used for input,
4774 we should presume that that is for a reason,
4775 and we should not substitute some other register
4776 which is not supposed to be clobbered.
4777 Therefore, this loop cannot be merged into the one below
4778 because a CALL may precede a CLOBBER and refer to the
4779 value clobbered. We must not let a canonicalization do
4780 anything in that case. */
4781 for (i = 0; i < lim; i++)
4783 rtx y = XVECEXP (x, 0, i);
4784 if (GET_CODE (y) == CLOBBER)
4786 rtx clobbered = XEXP (y, 0);
4788 if (REG_P (clobbered)
4789 || GET_CODE (clobbered) == SUBREG)
4790 invalidate (clobbered, VOIDmode);
4791 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4792 || GET_CODE (clobbered) == ZERO_EXTRACT)
4793 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4797 for (i = 0; i < lim; i++)
4799 rtx y = XVECEXP (x, 0, i);
4800 if (GET_CODE (y) == SET)
4802 /* As above, we ignore unconditional jumps and call-insns and
4803 ignore the result of apply_change_group. */
4804 if (GET_CODE (SET_SRC (y)) == CALL)
4806 canon_reg (SET_SRC (y), insn);
4807 apply_change_group ();
4808 fold_rtx (SET_SRC (y), insn);
4809 invalidate (SET_DEST (y), VOIDmode);
4811 else if (SET_DEST (y) == pc_rtx
4812 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4814 else
4815 sets[n_sets++].rtl = y;
4817 else if (GET_CODE (y) == CLOBBER)
4819 /* If we clobber memory, canon the address.
4820 This does nothing when a register is clobbered
4821 because we have already invalidated the reg. */
4822 if (MEM_P (XEXP (y, 0)))
4823 canon_reg (XEXP (y, 0), NULL_RTX);
4825 else if (GET_CODE (y) == USE
4826 && ! (REG_P (XEXP (y, 0))
4827 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4828 canon_reg (y, NULL_RTX);
4829 else if (GET_CODE (y) == CALL)
4831 /* The result of apply_change_group can be ignored; see
4832 canon_reg. */
4833 canon_reg (y, insn);
4834 apply_change_group ();
4835 fold_rtx (y, insn);
4839 else if (GET_CODE (x) == CLOBBER)
4841 if (MEM_P (XEXP (x, 0)))
4842 canon_reg (XEXP (x, 0), NULL_RTX);
4845 /* Canonicalize a USE of a pseudo register or memory location. */
4846 else if (GET_CODE (x) == USE
4847 && ! (REG_P (XEXP (x, 0))
4848 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4849 canon_reg (XEXP (x, 0), NULL_RTX);
4850 else if (GET_CODE (x) == CALL)
4852 /* The result of apply_change_group can be ignored; see canon_reg. */
4853 canon_reg (x, insn);
4854 apply_change_group ();
4855 fold_rtx (x, insn);
4858 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4859 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4860 is handled specially for this case, and if it isn't set, then there will
4861 be no equivalence for the destination. */
4862 if (n_sets == 1 && REG_NOTES (insn) != 0
4863 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4864 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4865 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4867 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4868 XEXP (tem, 0) = src_eqv;
4871 /* Canonicalize sources and addresses of destinations.
4872 We do this in a separate pass to avoid problems when a MATCH_DUP is
4873 present in the insn pattern. In that case, we want to ensure that
4874 we don't break the duplicate nature of the pattern. So we will replace
4875 both operands at the same time. Otherwise, we would fail to find an
4876 equivalent substitution in the loop calling validate_change below.
4878 We used to suppress canonicalization of DEST if it appears in SRC,
4879 but we don't do this any more. */
4881 for (i = 0; i < n_sets; i++)
4883 rtx dest = SET_DEST (sets[i].rtl);
4884 rtx src = SET_SRC (sets[i].rtl);
4885 rtx new = canon_reg (src, insn);
4886 int insn_code;
4888 sets[i].orig_src = src;
4889 if ((REG_P (new) && REG_P (src)
4890 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4891 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4892 || (insn_code = recog_memoized (insn)) < 0
4893 || insn_data[insn_code].n_dups > 0)
4894 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4895 else
4896 SET_SRC (sets[i].rtl) = new;
4898 if (GET_CODE (dest) == ZERO_EXTRACT)
4900 validate_change (insn, &XEXP (dest, 1),
4901 canon_reg (XEXP (dest, 1), insn), 1);
4902 validate_change (insn, &XEXP (dest, 2),
4903 canon_reg (XEXP (dest, 2), insn), 1);
4906 while (GET_CODE (dest) == SUBREG
4907 || GET_CODE (dest) == ZERO_EXTRACT
4908 || GET_CODE (dest) == STRICT_LOW_PART)
4909 dest = XEXP (dest, 0);
4911 if (MEM_P (dest))
4912 canon_reg (dest, insn);
4915 /* Now that we have done all the replacements, we can apply the change
4916 group and see if they all work. Note that this will cause some
4917 canonicalizations that would have worked individually not to be applied
4918 because some other canonicalization didn't work, but this should not
4919 occur often.
4921 The result of apply_change_group can be ignored; see canon_reg. */
4923 apply_change_group ();
4925 /* Set sets[i].src_elt to the class each source belongs to.
4926 Detect assignments from or to volatile things
4927 and set set[i] to zero so they will be ignored
4928 in the rest of this function.
4930 Nothing in this loop changes the hash table or the register chains. */
4932 for (i = 0; i < n_sets; i++)
4934 rtx src, dest;
4935 rtx src_folded;
4936 struct table_elt *elt = 0, *p;
4937 enum machine_mode mode;
4938 rtx src_eqv_here;
4939 rtx src_const = 0;
4940 rtx src_related = 0;
4941 struct table_elt *src_const_elt = 0;
4942 int src_cost = MAX_COST;
4943 int src_eqv_cost = MAX_COST;
4944 int src_folded_cost = MAX_COST;
4945 int src_related_cost = MAX_COST;
4946 int src_elt_cost = MAX_COST;
4947 int src_regcost = MAX_COST;
4948 int src_eqv_regcost = MAX_COST;
4949 int src_folded_regcost = MAX_COST;
4950 int src_related_regcost = MAX_COST;
4951 int src_elt_regcost = MAX_COST;
4952 /* Set nonzero if we need to call force_const_mem on with the
4953 contents of src_folded before using it. */
4954 int src_folded_force_flag = 0;
4956 dest = SET_DEST (sets[i].rtl);
4957 src = SET_SRC (sets[i].rtl);
4959 /* If SRC is a constant that has no machine mode,
4960 hash it with the destination's machine mode.
4961 This way we can keep different modes separate. */
4963 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4964 sets[i].mode = mode;
4966 if (src_eqv)
4968 enum machine_mode eqvmode = mode;
4969 if (GET_CODE (dest) == STRICT_LOW_PART)
4970 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4971 do_not_record = 0;
4972 hash_arg_in_memory = 0;
4973 src_eqv_hash = HASH (src_eqv, eqvmode);
4975 /* Find the equivalence class for the equivalent expression. */
4977 if (!do_not_record)
4978 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4980 src_eqv_volatile = do_not_record;
4981 src_eqv_in_memory = hash_arg_in_memory;
4984 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4985 value of the INNER register, not the destination. So it is not
4986 a valid substitution for the source. But save it for later. */
4987 if (GET_CODE (dest) == STRICT_LOW_PART)
4988 src_eqv_here = 0;
4989 else
4990 src_eqv_here = src_eqv;
4992 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4993 simplified result, which may not necessarily be valid. */
4994 src_folded = fold_rtx (src, insn);
4996 #if 0
4997 /* ??? This caused bad code to be generated for the m68k port with -O2.
4998 Suppose src is (CONST_INT -1), and that after truncation src_folded
4999 is (CONST_INT 3). Suppose src_folded is then used for src_const.
5000 At the end we will add src and src_const to the same equivalence
5001 class. We now have 3 and -1 on the same equivalence class. This
5002 causes later instructions to be mis-optimized. */
5003 /* If storing a constant in a bitfield, pre-truncate the constant
5004 so we will be able to record it later. */
5005 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5007 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5009 if (GET_CODE (src) == CONST_INT
5010 && GET_CODE (width) == CONST_INT
5011 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5012 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5013 src_folded
5014 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5015 << INTVAL (width)) - 1));
5017 #endif
5019 /* Compute SRC's hash code, and also notice if it
5020 should not be recorded at all. In that case,
5021 prevent any further processing of this assignment. */
5022 do_not_record = 0;
5023 hash_arg_in_memory = 0;
5025 sets[i].src = src;
5026 sets[i].src_hash = HASH (src, mode);
5027 sets[i].src_volatile = do_not_record;
5028 sets[i].src_in_memory = hash_arg_in_memory;
5030 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5031 a pseudo, do not record SRC. Using SRC as a replacement for
5032 anything else will be incorrect in that situation. Note that
5033 this usually occurs only for stack slots, in which case all the
5034 RTL would be referring to SRC, so we don't lose any optimization
5035 opportunities by not having SRC in the hash table. */
5037 if (MEM_P (src)
5038 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5039 && REG_P (dest)
5040 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5041 sets[i].src_volatile = 1;
5043 #if 0
5044 /* It is no longer clear why we used to do this, but it doesn't
5045 appear to still be needed. So let's try without it since this
5046 code hurts cse'ing widened ops. */
5047 /* If source is a paradoxical subreg (such as QI treated as an SI),
5048 treat it as volatile. It may do the work of an SI in one context
5049 where the extra bits are not being used, but cannot replace an SI
5050 in general. */
5051 if (GET_CODE (src) == SUBREG
5052 && (GET_MODE_SIZE (GET_MODE (src))
5053 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5054 sets[i].src_volatile = 1;
5055 #endif
5057 /* Locate all possible equivalent forms for SRC. Try to replace
5058 SRC in the insn with each cheaper equivalent.
5060 We have the following types of equivalents: SRC itself, a folded
5061 version, a value given in a REG_EQUAL note, or a value related
5062 to a constant.
5064 Each of these equivalents may be part of an additional class
5065 of equivalents (if more than one is in the table, they must be in
5066 the same class; we check for this).
5068 If the source is volatile, we don't do any table lookups.
5070 We note any constant equivalent for possible later use in a
5071 REG_NOTE. */
5073 if (!sets[i].src_volatile)
5074 elt = lookup (src, sets[i].src_hash, mode);
5076 sets[i].src_elt = elt;
5078 if (elt && src_eqv_here && src_eqv_elt)
5080 if (elt->first_same_value != src_eqv_elt->first_same_value)
5082 /* The REG_EQUAL is indicating that two formerly distinct
5083 classes are now equivalent. So merge them. */
5084 merge_equiv_classes (elt, src_eqv_elt);
5085 src_eqv_hash = HASH (src_eqv, elt->mode);
5086 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5089 src_eqv_here = 0;
5092 else if (src_eqv_elt)
5093 elt = src_eqv_elt;
5095 /* Try to find a constant somewhere and record it in `src_const'.
5096 Record its table element, if any, in `src_const_elt'. Look in
5097 any known equivalences first. (If the constant is not in the
5098 table, also set `sets[i].src_const_hash'). */
5099 if (elt)
5100 for (p = elt->first_same_value; p; p = p->next_same_value)
5101 if (p->is_const)
5103 src_const = p->exp;
5104 src_const_elt = elt;
5105 break;
5108 if (src_const == 0
5109 && (CONSTANT_P (src_folded)
5110 /* Consider (minus (label_ref L1) (label_ref L2)) as
5111 "constant" here so we will record it. This allows us
5112 to fold switch statements when an ADDR_DIFF_VEC is used. */
5113 || (GET_CODE (src_folded) == MINUS
5114 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5115 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5116 src_const = src_folded, src_const_elt = elt;
5117 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5118 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5120 /* If we don't know if the constant is in the table, get its
5121 hash code and look it up. */
5122 if (src_const && src_const_elt == 0)
5124 sets[i].src_const_hash = HASH (src_const, mode);
5125 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5128 sets[i].src_const = src_const;
5129 sets[i].src_const_elt = src_const_elt;
5131 /* If the constant and our source are both in the table, mark them as
5132 equivalent. Otherwise, if a constant is in the table but the source
5133 isn't, set ELT to it. */
5134 if (src_const_elt && elt
5135 && src_const_elt->first_same_value != elt->first_same_value)
5136 merge_equiv_classes (elt, src_const_elt);
5137 else if (src_const_elt && elt == 0)
5138 elt = src_const_elt;
5140 /* See if there is a register linearly related to a constant
5141 equivalent of SRC. */
5142 if (src_const
5143 && (GET_CODE (src_const) == CONST
5144 || (src_const_elt && src_const_elt->related_value != 0)))
5146 src_related = use_related_value (src_const, src_const_elt);
5147 if (src_related)
5149 struct table_elt *src_related_elt
5150 = lookup (src_related, HASH (src_related, mode), mode);
5151 if (src_related_elt && elt)
5153 if (elt->first_same_value
5154 != src_related_elt->first_same_value)
5155 /* This can occur when we previously saw a CONST
5156 involving a SYMBOL_REF and then see the SYMBOL_REF
5157 twice. Merge the involved classes. */
5158 merge_equiv_classes (elt, src_related_elt);
5160 src_related = 0;
5161 src_related_elt = 0;
5163 else if (src_related_elt && elt == 0)
5164 elt = src_related_elt;
5168 /* See if we have a CONST_INT that is already in a register in a
5169 wider mode. */
5171 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5172 && GET_MODE_CLASS (mode) == MODE_INT
5173 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5175 enum machine_mode wider_mode;
5177 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5178 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5179 && src_related == 0;
5180 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5182 struct table_elt *const_elt
5183 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5185 if (const_elt == 0)
5186 continue;
5188 for (const_elt = const_elt->first_same_value;
5189 const_elt; const_elt = const_elt->next_same_value)
5190 if (REG_P (const_elt->exp))
5192 src_related = gen_lowpart (mode,
5193 const_elt->exp);
5194 break;
5199 /* Another possibility is that we have an AND with a constant in
5200 a mode narrower than a word. If so, it might have been generated
5201 as part of an "if" which would narrow the AND. If we already
5202 have done the AND in a wider mode, we can use a SUBREG of that
5203 value. */
5205 if (flag_expensive_optimizations && ! src_related
5206 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5207 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5209 enum machine_mode tmode;
5210 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5212 for (tmode = GET_MODE_WIDER_MODE (mode);
5213 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5214 tmode = GET_MODE_WIDER_MODE (tmode))
5216 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5217 struct table_elt *larger_elt;
5219 if (inner)
5221 PUT_MODE (new_and, tmode);
5222 XEXP (new_and, 0) = inner;
5223 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5224 if (larger_elt == 0)
5225 continue;
5227 for (larger_elt = larger_elt->first_same_value;
5228 larger_elt; larger_elt = larger_elt->next_same_value)
5229 if (REG_P (larger_elt->exp))
5231 src_related
5232 = gen_lowpart (mode, larger_elt->exp);
5233 break;
5236 if (src_related)
5237 break;
5242 #ifdef LOAD_EXTEND_OP
5243 /* See if a MEM has already been loaded with a widening operation;
5244 if it has, we can use a subreg of that. Many CISC machines
5245 also have such operations, but this is only likely to be
5246 beneficial on these machines. */
5248 if (flag_expensive_optimizations && src_related == 0
5249 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5250 && GET_MODE_CLASS (mode) == MODE_INT
5251 && MEM_P (src) && ! do_not_record
5252 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5254 struct rtx_def memory_extend_buf;
5255 rtx memory_extend_rtx = &memory_extend_buf;
5256 enum machine_mode tmode;
5258 /* Set what we are trying to extend and the operation it might
5259 have been extended with. */
5260 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5261 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5262 XEXP (memory_extend_rtx, 0) = src;
5264 for (tmode = GET_MODE_WIDER_MODE (mode);
5265 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5266 tmode = GET_MODE_WIDER_MODE (tmode))
5268 struct table_elt *larger_elt;
5270 PUT_MODE (memory_extend_rtx, tmode);
5271 larger_elt = lookup (memory_extend_rtx,
5272 HASH (memory_extend_rtx, tmode), tmode);
5273 if (larger_elt == 0)
5274 continue;
5276 for (larger_elt = larger_elt->first_same_value;
5277 larger_elt; larger_elt = larger_elt->next_same_value)
5278 if (REG_P (larger_elt->exp))
5280 src_related = gen_lowpart (mode,
5281 larger_elt->exp);
5282 break;
5285 if (src_related)
5286 break;
5289 #endif /* LOAD_EXTEND_OP */
5291 if (src == src_folded)
5292 src_folded = 0;
5294 /* At this point, ELT, if nonzero, points to a class of expressions
5295 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5296 and SRC_RELATED, if nonzero, each contain additional equivalent
5297 expressions. Prune these latter expressions by deleting expressions
5298 already in the equivalence class.
5300 Check for an equivalent identical to the destination. If found,
5301 this is the preferred equivalent since it will likely lead to
5302 elimination of the insn. Indicate this by placing it in
5303 `src_related'. */
5305 if (elt)
5306 elt = elt->first_same_value;
5307 for (p = elt; p; p = p->next_same_value)
5309 enum rtx_code code = GET_CODE (p->exp);
5311 /* If the expression is not valid, ignore it. Then we do not
5312 have to check for validity below. In most cases, we can use
5313 `rtx_equal_p', since canonicalization has already been done. */
5314 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5315 continue;
5317 /* Also skip paradoxical subregs, unless that's what we're
5318 looking for. */
5319 if (code == SUBREG
5320 && (GET_MODE_SIZE (GET_MODE (p->exp))
5321 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5322 && ! (src != 0
5323 && GET_CODE (src) == SUBREG
5324 && GET_MODE (src) == GET_MODE (p->exp)
5325 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5326 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5327 continue;
5329 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5330 src = 0;
5331 else if (src_folded && GET_CODE (src_folded) == code
5332 && rtx_equal_p (src_folded, p->exp))
5333 src_folded = 0;
5334 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5335 && rtx_equal_p (src_eqv_here, p->exp))
5336 src_eqv_here = 0;
5337 else if (src_related && GET_CODE (src_related) == code
5338 && rtx_equal_p (src_related, p->exp))
5339 src_related = 0;
5341 /* This is the same as the destination of the insns, we want
5342 to prefer it. Copy it to src_related. The code below will
5343 then give it a negative cost. */
5344 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5345 src_related = dest;
5348 /* Find the cheapest valid equivalent, trying all the available
5349 possibilities. Prefer items not in the hash table to ones
5350 that are when they are equal cost. Note that we can never
5351 worsen an insn as the current contents will also succeed.
5352 If we find an equivalent identical to the destination, use it as best,
5353 since this insn will probably be eliminated in that case. */
5354 if (src)
5356 if (rtx_equal_p (src, dest))
5357 src_cost = src_regcost = -1;
5358 else
5360 src_cost = COST (src);
5361 src_regcost = approx_reg_cost (src);
5365 if (src_eqv_here)
5367 if (rtx_equal_p (src_eqv_here, dest))
5368 src_eqv_cost = src_eqv_regcost = -1;
5369 else
5371 src_eqv_cost = COST (src_eqv_here);
5372 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5376 if (src_folded)
5378 if (rtx_equal_p (src_folded, dest))
5379 src_folded_cost = src_folded_regcost = -1;
5380 else
5382 src_folded_cost = COST (src_folded);
5383 src_folded_regcost = approx_reg_cost (src_folded);
5387 if (src_related)
5389 if (rtx_equal_p (src_related, dest))
5390 src_related_cost = src_related_regcost = -1;
5391 else
5393 src_related_cost = COST (src_related);
5394 src_related_regcost = approx_reg_cost (src_related);
5398 /* If this was an indirect jump insn, a known label will really be
5399 cheaper even though it looks more expensive. */
5400 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5401 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5403 /* Terminate loop when replacement made. This must terminate since
5404 the current contents will be tested and will always be valid. */
5405 while (1)
5407 rtx trial;
5409 /* Skip invalid entries. */
5410 while (elt && !REG_P (elt->exp)
5411 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5412 elt = elt->next_same_value;
5414 /* A paradoxical subreg would be bad here: it'll be the right
5415 size, but later may be adjusted so that the upper bits aren't
5416 what we want. So reject it. */
5417 if (elt != 0
5418 && GET_CODE (elt->exp) == SUBREG
5419 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5420 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5421 /* It is okay, though, if the rtx we're trying to match
5422 will ignore any of the bits we can't predict. */
5423 && ! (src != 0
5424 && GET_CODE (src) == SUBREG
5425 && GET_MODE (src) == GET_MODE (elt->exp)
5426 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5427 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5429 elt = elt->next_same_value;
5430 continue;
5433 if (elt)
5435 src_elt_cost = elt->cost;
5436 src_elt_regcost = elt->regcost;
5439 /* Find cheapest and skip it for the next time. For items
5440 of equal cost, use this order:
5441 src_folded, src, src_eqv, src_related and hash table entry. */
5442 if (src_folded
5443 && preferable (src_folded_cost, src_folded_regcost,
5444 src_cost, src_regcost) <= 0
5445 && preferable (src_folded_cost, src_folded_regcost,
5446 src_eqv_cost, src_eqv_regcost) <= 0
5447 && preferable (src_folded_cost, src_folded_regcost,
5448 src_related_cost, src_related_regcost) <= 0
5449 && preferable (src_folded_cost, src_folded_regcost,
5450 src_elt_cost, src_elt_regcost) <= 0)
5452 trial = src_folded, src_folded_cost = MAX_COST;
5453 if (src_folded_force_flag)
5455 rtx forced = force_const_mem (mode, trial);
5456 if (forced)
5457 trial = forced;
5460 else if (src
5461 && preferable (src_cost, src_regcost,
5462 src_eqv_cost, src_eqv_regcost) <= 0
5463 && preferable (src_cost, src_regcost,
5464 src_related_cost, src_related_regcost) <= 0
5465 && preferable (src_cost, src_regcost,
5466 src_elt_cost, src_elt_regcost) <= 0)
5467 trial = src, src_cost = MAX_COST;
5468 else if (src_eqv_here
5469 && preferable (src_eqv_cost, src_eqv_regcost,
5470 src_related_cost, src_related_regcost) <= 0
5471 && preferable (src_eqv_cost, src_eqv_regcost,
5472 src_elt_cost, src_elt_regcost) <= 0)
5473 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5474 else if (src_related
5475 && preferable (src_related_cost, src_related_regcost,
5476 src_elt_cost, src_elt_regcost) <= 0)
5477 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5478 else
5480 trial = copy_rtx (elt->exp);
5481 elt = elt->next_same_value;
5482 src_elt_cost = MAX_COST;
5485 /* We don't normally have an insn matching (set (pc) (pc)), so
5486 check for this separately here. We will delete such an
5487 insn below.
5489 For other cases such as a table jump or conditional jump
5490 where we know the ultimate target, go ahead and replace the
5491 operand. While that may not make a valid insn, we will
5492 reemit the jump below (and also insert any necessary
5493 barriers). */
5494 if (n_sets == 1 && dest == pc_rtx
5495 && (trial == pc_rtx
5496 || (GET_CODE (trial) == LABEL_REF
5497 && ! condjump_p (insn))))
5499 /* Don't substitute non-local labels, this confuses CFG. */
5500 if (GET_CODE (trial) == LABEL_REF
5501 && LABEL_REF_NONLOCAL_P (trial))
5502 continue;
5504 SET_SRC (sets[i].rtl) = trial;
5505 cse_jumps_altered = 1;
5506 break;
5509 /* Reject certain invalid forms of CONST that we create. */
5510 else if (CONSTANT_P (trial)
5511 && GET_CODE (trial) == CONST
5512 /* Reject cases that will cause decode_rtx_const to
5513 die. On the alpha when simplifying a switch, we
5514 get (const (truncate (minus (label_ref)
5515 (label_ref)))). */
5516 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5517 /* Likewise on IA-64, except without the
5518 truncate. */
5519 || (GET_CODE (XEXP (trial, 0)) == MINUS
5520 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5521 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5522 /* Do nothing for this case. */
5525 /* Look for a substitution that makes a valid insn. */
5526 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5528 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5530 /* If we just made a substitution inside a libcall, then we
5531 need to make the same substitution in any notes attached
5532 to the RETVAL insn. */
5533 if (libcall_insn
5534 && (REG_P (sets[i].orig_src)
5535 || GET_CODE (sets[i].orig_src) == SUBREG
5536 || MEM_P (sets[i].orig_src)))
5538 rtx note = find_reg_equal_equiv_note (libcall_insn);
5539 if (note != 0)
5540 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5541 sets[i].orig_src,
5542 copy_rtx (new));
5545 /* The result of apply_change_group can be ignored; see
5546 canon_reg. */
5548 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5549 apply_change_group ();
5550 break;
5553 /* If we previously found constant pool entries for
5554 constants and this is a constant, try making a
5555 pool entry. Put it in src_folded unless we already have done
5556 this since that is where it likely came from. */
5558 else if (constant_pool_entries_cost
5559 && CONSTANT_P (trial)
5560 && (src_folded == 0
5561 || (!MEM_P (src_folded)
5562 && ! src_folded_force_flag))
5563 && GET_MODE_CLASS (mode) != MODE_CC
5564 && mode != VOIDmode)
5566 src_folded_force_flag = 1;
5567 src_folded = trial;
5568 src_folded_cost = constant_pool_entries_cost;
5569 src_folded_regcost = constant_pool_entries_regcost;
5573 src = SET_SRC (sets[i].rtl);
5575 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5576 However, there is an important exception: If both are registers
5577 that are not the head of their equivalence class, replace SET_SRC
5578 with the head of the class. If we do not do this, we will have
5579 both registers live over a portion of the basic block. This way,
5580 their lifetimes will likely abut instead of overlapping. */
5581 if (REG_P (dest)
5582 && REGNO_QTY_VALID_P (REGNO (dest)))
5584 int dest_q = REG_QTY (REGNO (dest));
5585 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5587 if (dest_ent->mode == GET_MODE (dest)
5588 && dest_ent->first_reg != REGNO (dest)
5589 && REG_P (src) && REGNO (src) == REGNO (dest)
5590 /* Don't do this if the original insn had a hard reg as
5591 SET_SRC or SET_DEST. */
5592 && (!REG_P (sets[i].src)
5593 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5594 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5595 /* We can't call canon_reg here because it won't do anything if
5596 SRC is a hard register. */
5598 int src_q = REG_QTY (REGNO (src));
5599 struct qty_table_elem *src_ent = &qty_table[src_q];
5600 int first = src_ent->first_reg;
5601 rtx new_src
5602 = (first >= FIRST_PSEUDO_REGISTER
5603 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5605 /* We must use validate-change even for this, because this
5606 might be a special no-op instruction, suitable only to
5607 tag notes onto. */
5608 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5610 src = new_src;
5611 /* If we had a constant that is cheaper than what we are now
5612 setting SRC to, use that constant. We ignored it when we
5613 thought we could make this into a no-op. */
5614 if (src_const && COST (src_const) < COST (src)
5615 && validate_change (insn, &SET_SRC (sets[i].rtl),
5616 src_const, 0))
5617 src = src_const;
5622 /* If we made a change, recompute SRC values. */
5623 if (src != sets[i].src)
5625 cse_altered = 1;
5626 do_not_record = 0;
5627 hash_arg_in_memory = 0;
5628 sets[i].src = src;
5629 sets[i].src_hash = HASH (src, mode);
5630 sets[i].src_volatile = do_not_record;
5631 sets[i].src_in_memory = hash_arg_in_memory;
5632 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5635 /* If this is a single SET, we are setting a register, and we have an
5636 equivalent constant, we want to add a REG_NOTE. We don't want
5637 to write a REG_EQUAL note for a constant pseudo since verifying that
5638 that pseudo hasn't been eliminated is a pain. Such a note also
5639 won't help anything.
5641 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5642 which can be created for a reference to a compile time computable
5643 entry in a jump table. */
5645 if (n_sets == 1 && src_const && REG_P (dest)
5646 && !REG_P (src_const)
5647 && ! (GET_CODE (src_const) == CONST
5648 && GET_CODE (XEXP (src_const, 0)) == MINUS
5649 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5650 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5652 /* We only want a REG_EQUAL note if src_const != src. */
5653 if (! rtx_equal_p (src, src_const))
5655 /* Make sure that the rtx is not shared. */
5656 src_const = copy_rtx (src_const);
5658 /* Record the actual constant value in a REG_EQUAL note,
5659 making a new one if one does not already exist. */
5660 set_unique_reg_note (insn, REG_EQUAL, src_const);
5664 /* Now deal with the destination. */
5665 do_not_record = 0;
5667 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5668 while (GET_CODE (dest) == SUBREG
5669 || GET_CODE (dest) == ZERO_EXTRACT
5670 || GET_CODE (dest) == STRICT_LOW_PART)
5671 dest = XEXP (dest, 0);
5673 sets[i].inner_dest = dest;
5675 if (MEM_P (dest))
5677 #ifdef PUSH_ROUNDING
5678 /* Stack pushes invalidate the stack pointer. */
5679 rtx addr = XEXP (dest, 0);
5680 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5681 && XEXP (addr, 0) == stack_pointer_rtx)
5682 invalidate (stack_pointer_rtx, Pmode);
5683 #endif
5684 dest = fold_rtx (dest, insn);
5687 /* Compute the hash code of the destination now,
5688 before the effects of this instruction are recorded,
5689 since the register values used in the address computation
5690 are those before this instruction. */
5691 sets[i].dest_hash = HASH (dest, mode);
5693 /* Don't enter a bit-field in the hash table
5694 because the value in it after the store
5695 may not equal what was stored, due to truncation. */
5697 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5699 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5701 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5702 && GET_CODE (width) == CONST_INT
5703 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5704 && ! (INTVAL (src_const)
5705 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5706 /* Exception: if the value is constant,
5707 and it won't be truncated, record it. */
5709 else
5711 /* This is chosen so that the destination will be invalidated
5712 but no new value will be recorded.
5713 We must invalidate because sometimes constant
5714 values can be recorded for bitfields. */
5715 sets[i].src_elt = 0;
5716 sets[i].src_volatile = 1;
5717 src_eqv = 0;
5718 src_eqv_elt = 0;
5722 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5723 the insn. */
5724 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5726 /* One less use of the label this insn used to jump to. */
5727 delete_insn (insn);
5728 cse_jumps_altered = 1;
5729 /* No more processing for this set. */
5730 sets[i].rtl = 0;
5733 /* If this SET is now setting PC to a label, we know it used to
5734 be a conditional or computed branch. */
5735 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5736 && !LABEL_REF_NONLOCAL_P (src))
5738 /* Now emit a BARRIER after the unconditional jump. */
5739 if (NEXT_INSN (insn) == 0
5740 || !BARRIER_P (NEXT_INSN (insn)))
5741 emit_barrier_after (insn);
5743 /* We reemit the jump in as many cases as possible just in
5744 case the form of an unconditional jump is significantly
5745 different than a computed jump or conditional jump.
5747 If this insn has multiple sets, then reemitting the
5748 jump is nontrivial. So instead we just force rerecognition
5749 and hope for the best. */
5750 if (n_sets == 1)
5752 rtx new, note;
5754 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5755 JUMP_LABEL (new) = XEXP (src, 0);
5756 LABEL_NUSES (XEXP (src, 0))++;
5758 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5759 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5760 if (note)
5762 XEXP (note, 1) = NULL_RTX;
5763 REG_NOTES (new) = note;
5766 delete_insn (insn);
5767 insn = new;
5769 /* Now emit a BARRIER after the unconditional jump. */
5770 if (NEXT_INSN (insn) == 0
5771 || !BARRIER_P (NEXT_INSN (insn)))
5772 emit_barrier_after (insn);
5774 else
5775 INSN_CODE (insn) = -1;
5777 /* Do not bother deleting any unreachable code,
5778 let jump/flow do that. */
5780 cse_jumps_altered = 1;
5781 sets[i].rtl = 0;
5784 /* If destination is volatile, invalidate it and then do no further
5785 processing for this assignment. */
5787 else if (do_not_record)
5789 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5790 invalidate (dest, VOIDmode);
5791 else if (MEM_P (dest))
5792 invalidate (dest, VOIDmode);
5793 else if (GET_CODE (dest) == STRICT_LOW_PART
5794 || GET_CODE (dest) == ZERO_EXTRACT)
5795 invalidate (XEXP (dest, 0), GET_MODE (dest));
5796 sets[i].rtl = 0;
5799 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5800 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5802 #ifdef HAVE_cc0
5803 /* If setting CC0, record what it was set to, or a constant, if it
5804 is equivalent to a constant. If it is being set to a floating-point
5805 value, make a COMPARE with the appropriate constant of 0. If we
5806 don't do this, later code can interpret this as a test against
5807 const0_rtx, which can cause problems if we try to put it into an
5808 insn as a floating-point operand. */
5809 if (dest == cc0_rtx)
5811 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5812 this_insn_cc0_mode = mode;
5813 if (FLOAT_MODE_P (mode))
5814 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5815 CONST0_RTX (mode));
5817 #endif
5820 /* Now enter all non-volatile source expressions in the hash table
5821 if they are not already present.
5822 Record their equivalence classes in src_elt.
5823 This way we can insert the corresponding destinations into
5824 the same classes even if the actual sources are no longer in them
5825 (having been invalidated). */
5827 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5828 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5830 struct table_elt *elt;
5831 struct table_elt *classp = sets[0].src_elt;
5832 rtx dest = SET_DEST (sets[0].rtl);
5833 enum machine_mode eqvmode = GET_MODE (dest);
5835 if (GET_CODE (dest) == STRICT_LOW_PART)
5837 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5838 classp = 0;
5840 if (insert_regs (src_eqv, classp, 0))
5842 rehash_using_reg (src_eqv);
5843 src_eqv_hash = HASH (src_eqv, eqvmode);
5845 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5846 elt->in_memory = src_eqv_in_memory;
5847 src_eqv_elt = elt;
5849 /* Check to see if src_eqv_elt is the same as a set source which
5850 does not yet have an elt, and if so set the elt of the set source
5851 to src_eqv_elt. */
5852 for (i = 0; i < n_sets; i++)
5853 if (sets[i].rtl && sets[i].src_elt == 0
5854 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5855 sets[i].src_elt = src_eqv_elt;
5858 for (i = 0; i < n_sets; i++)
5859 if (sets[i].rtl && ! sets[i].src_volatile
5860 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5862 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5864 /* REG_EQUAL in setting a STRICT_LOW_PART
5865 gives an equivalent for the entire destination register,
5866 not just for the subreg being stored in now.
5867 This is a more interesting equivalence, so we arrange later
5868 to treat the entire reg as the destination. */
5869 sets[i].src_elt = src_eqv_elt;
5870 sets[i].src_hash = src_eqv_hash;
5872 else
5874 /* Insert source and constant equivalent into hash table, if not
5875 already present. */
5876 struct table_elt *classp = src_eqv_elt;
5877 rtx src = sets[i].src;
5878 rtx dest = SET_DEST (sets[i].rtl);
5879 enum machine_mode mode
5880 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5882 /* It's possible that we have a source value known to be
5883 constant but don't have a REG_EQUAL note on the insn.
5884 Lack of a note will mean src_eqv_elt will be NULL. This
5885 can happen where we've generated a SUBREG to access a
5886 CONST_INT that is already in a register in a wider mode.
5887 Ensure that the source expression is put in the proper
5888 constant class. */
5889 if (!classp)
5890 classp = sets[i].src_const_elt;
5892 if (sets[i].src_elt == 0)
5894 /* Don't put a hard register source into the table if this is
5895 the last insn of a libcall. In this case, we only need
5896 to put src_eqv_elt in src_elt. */
5897 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5899 struct table_elt *elt;
5901 /* Note that these insert_regs calls cannot remove
5902 any of the src_elt's, because they would have failed to
5903 match if not still valid. */
5904 if (insert_regs (src, classp, 0))
5906 rehash_using_reg (src);
5907 sets[i].src_hash = HASH (src, mode);
5909 elt = insert (src, classp, sets[i].src_hash, mode);
5910 elt->in_memory = sets[i].src_in_memory;
5911 sets[i].src_elt = classp = elt;
5913 else
5914 sets[i].src_elt = classp;
5916 if (sets[i].src_const && sets[i].src_const_elt == 0
5917 && src != sets[i].src_const
5918 && ! rtx_equal_p (sets[i].src_const, src))
5919 sets[i].src_elt = insert (sets[i].src_const, classp,
5920 sets[i].src_const_hash, mode);
5923 else if (sets[i].src_elt == 0)
5924 /* If we did not insert the source into the hash table (e.g., it was
5925 volatile), note the equivalence class for the REG_EQUAL value, if any,
5926 so that the destination goes into that class. */
5927 sets[i].src_elt = src_eqv_elt;
5929 invalidate_from_clobbers (x);
5931 /* Some registers are invalidated by subroutine calls. Memory is
5932 invalidated by non-constant calls. */
5934 if (CALL_P (insn))
5936 if (! CONST_OR_PURE_CALL_P (insn))
5937 invalidate_memory ();
5938 invalidate_for_call ();
5941 /* Now invalidate everything set by this instruction.
5942 If a SUBREG or other funny destination is being set,
5943 sets[i].rtl is still nonzero, so here we invalidate the reg
5944 a part of which is being set. */
5946 for (i = 0; i < n_sets; i++)
5947 if (sets[i].rtl)
5949 /* We can't use the inner dest, because the mode associated with
5950 a ZERO_EXTRACT is significant. */
5951 rtx dest = SET_DEST (sets[i].rtl);
5953 /* Needed for registers to remove the register from its
5954 previous quantity's chain.
5955 Needed for memory if this is a nonvarying address, unless
5956 we have just done an invalidate_memory that covers even those. */
5957 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5958 invalidate (dest, VOIDmode);
5959 else if (MEM_P (dest))
5960 invalidate (dest, VOIDmode);
5961 else if (GET_CODE (dest) == STRICT_LOW_PART
5962 || GET_CODE (dest) == ZERO_EXTRACT)
5963 invalidate (XEXP (dest, 0), GET_MODE (dest));
5966 /* A volatile ASM invalidates everything. */
5967 if (NONJUMP_INSN_P (insn)
5968 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5969 && MEM_VOLATILE_P (PATTERN (insn)))
5970 flush_hash_table ();
5972 /* Make sure registers mentioned in destinations
5973 are safe for use in an expression to be inserted.
5974 This removes from the hash table
5975 any invalid entry that refers to one of these registers.
5977 We don't care about the return value from mention_regs because
5978 we are going to hash the SET_DEST values unconditionally. */
5980 for (i = 0; i < n_sets; i++)
5982 if (sets[i].rtl)
5984 rtx x = SET_DEST (sets[i].rtl);
5986 if (!REG_P (x))
5987 mention_regs (x);
5988 else
5990 /* We used to rely on all references to a register becoming
5991 inaccessible when a register changes to a new quantity,
5992 since that changes the hash code. However, that is not
5993 safe, since after HASH_SIZE new quantities we get a
5994 hash 'collision' of a register with its own invalid
5995 entries. And since SUBREGs have been changed not to
5996 change their hash code with the hash code of the register,
5997 it wouldn't work any longer at all. So we have to check
5998 for any invalid references lying around now.
5999 This code is similar to the REG case in mention_regs,
6000 but it knows that reg_tick has been incremented, and
6001 it leaves reg_in_table as -1 . */
6002 unsigned int regno = REGNO (x);
6003 unsigned int endregno
6004 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
6005 : hard_regno_nregs[regno][GET_MODE (x)]);
6006 unsigned int i;
6008 for (i = regno; i < endregno; i++)
6010 if (REG_IN_TABLE (i) >= 0)
6012 remove_invalid_refs (i);
6013 REG_IN_TABLE (i) = -1;
6020 /* We may have just removed some of the src_elt's from the hash table.
6021 So replace each one with the current head of the same class. */
6023 for (i = 0; i < n_sets; i++)
6024 if (sets[i].rtl)
6026 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6027 /* If elt was removed, find current head of same class,
6028 or 0 if nothing remains of that class. */
6030 struct table_elt *elt = sets[i].src_elt;
6032 while (elt && elt->prev_same_value)
6033 elt = elt->prev_same_value;
6035 while (elt && elt->first_same_value == 0)
6036 elt = elt->next_same_value;
6037 sets[i].src_elt = elt ? elt->first_same_value : 0;
6041 /* Now insert the destinations into their equivalence classes. */
6043 for (i = 0; i < n_sets; i++)
6044 if (sets[i].rtl)
6046 rtx dest = SET_DEST (sets[i].rtl);
6047 struct table_elt *elt;
6049 /* Don't record value if we are not supposed to risk allocating
6050 floating-point values in registers that might be wider than
6051 memory. */
6052 if ((flag_float_store
6053 && MEM_P (dest)
6054 && FLOAT_MODE_P (GET_MODE (dest)))
6055 /* Don't record BLKmode values, because we don't know the
6056 size of it, and can't be sure that other BLKmode values
6057 have the same or smaller size. */
6058 || GET_MODE (dest) == BLKmode
6059 /* Don't record values of destinations set inside a libcall block
6060 since we might delete the libcall. Things should have been set
6061 up so we won't want to reuse such a value, but we play it safe
6062 here. */
6063 || libcall_insn
6064 /* If we didn't put a REG_EQUAL value or a source into the hash
6065 table, there is no point is recording DEST. */
6066 || sets[i].src_elt == 0
6067 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6068 or SIGN_EXTEND, don't record DEST since it can cause
6069 some tracking to be wrong.
6071 ??? Think about this more later. */
6072 || (GET_CODE (dest) == SUBREG
6073 && (GET_MODE_SIZE (GET_MODE (dest))
6074 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6075 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6076 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6077 continue;
6079 /* STRICT_LOW_PART isn't part of the value BEING set,
6080 and neither is the SUBREG inside it.
6081 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6082 if (GET_CODE (dest) == STRICT_LOW_PART)
6083 dest = SUBREG_REG (XEXP (dest, 0));
6085 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6086 /* Registers must also be inserted into chains for quantities. */
6087 if (insert_regs (dest, sets[i].src_elt, 1))
6089 /* If `insert_regs' changes something, the hash code must be
6090 recalculated. */
6091 rehash_using_reg (dest);
6092 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6095 elt = insert (dest, sets[i].src_elt,
6096 sets[i].dest_hash, GET_MODE (dest));
6098 elt->in_memory = (MEM_P (sets[i].inner_dest)
6099 && !MEM_READONLY_P (sets[i].inner_dest));
6101 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6102 narrower than M2, and both M1 and M2 are the same number of words,
6103 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6104 make that equivalence as well.
6106 However, BAR may have equivalences for which gen_lowpart
6107 will produce a simpler value than gen_lowpart applied to
6108 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6109 BAR's equivalences. If we don't get a simplified form, make
6110 the SUBREG. It will not be used in an equivalence, but will
6111 cause two similar assignments to be detected.
6113 Note the loop below will find SUBREG_REG (DEST) since we have
6114 already entered SRC and DEST of the SET in the table. */
6116 if (GET_CODE (dest) == SUBREG
6117 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6118 / UNITS_PER_WORD)
6119 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6120 && (GET_MODE_SIZE (GET_MODE (dest))
6121 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6122 && sets[i].src_elt != 0)
6124 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6125 struct table_elt *elt, *classp = 0;
6127 for (elt = sets[i].src_elt->first_same_value; elt;
6128 elt = elt->next_same_value)
6130 rtx new_src = 0;
6131 unsigned src_hash;
6132 struct table_elt *src_elt;
6133 int byte = 0;
6135 /* Ignore invalid entries. */
6136 if (!REG_P (elt->exp)
6137 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6138 continue;
6140 /* We may have already been playing subreg games. If the
6141 mode is already correct for the destination, use it. */
6142 if (GET_MODE (elt->exp) == new_mode)
6143 new_src = elt->exp;
6144 else
6146 /* Calculate big endian correction for the SUBREG_BYTE.
6147 We have already checked that M1 (GET_MODE (dest))
6148 is not narrower than M2 (new_mode). */
6149 if (BYTES_BIG_ENDIAN)
6150 byte = (GET_MODE_SIZE (GET_MODE (dest))
6151 - GET_MODE_SIZE (new_mode));
6153 new_src = simplify_gen_subreg (new_mode, elt->exp,
6154 GET_MODE (dest), byte);
6157 /* The call to simplify_gen_subreg fails if the value
6158 is VOIDmode, yet we can't do any simplification, e.g.
6159 for EXPR_LISTs denoting function call results.
6160 It is invalid to construct a SUBREG with a VOIDmode
6161 SUBREG_REG, hence a zero new_src means we can't do
6162 this substitution. */
6163 if (! new_src)
6164 continue;
6166 src_hash = HASH (new_src, new_mode);
6167 src_elt = lookup (new_src, src_hash, new_mode);
6169 /* Put the new source in the hash table is if isn't
6170 already. */
6171 if (src_elt == 0)
6173 if (insert_regs (new_src, classp, 0))
6175 rehash_using_reg (new_src);
6176 src_hash = HASH (new_src, new_mode);
6178 src_elt = insert (new_src, classp, src_hash, new_mode);
6179 src_elt->in_memory = elt->in_memory;
6181 else if (classp && classp != src_elt->first_same_value)
6182 /* Show that two things that we've seen before are
6183 actually the same. */
6184 merge_equiv_classes (src_elt, classp);
6186 classp = src_elt->first_same_value;
6187 /* Ignore invalid entries. */
6188 while (classp
6189 && !REG_P (classp->exp)
6190 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6191 classp = classp->next_same_value;
6196 /* Special handling for (set REG0 REG1) where REG0 is the
6197 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6198 be used in the sequel, so (if easily done) change this insn to
6199 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6200 that computed their value. Then REG1 will become a dead store
6201 and won't cloud the situation for later optimizations.
6203 Do not make this change if REG1 is a hard register, because it will
6204 then be used in the sequel and we may be changing a two-operand insn
6205 into a three-operand insn.
6207 Also do not do this if we are operating on a copy of INSN.
6209 Also don't do this if INSN ends a libcall; this would cause an unrelated
6210 register to be set in the middle of a libcall, and we then get bad code
6211 if the libcall is deleted. */
6213 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6214 && NEXT_INSN (PREV_INSN (insn)) == insn
6215 && REG_P (SET_SRC (sets[0].rtl))
6216 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6217 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6219 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6220 struct qty_table_elem *src_ent = &qty_table[src_q];
6222 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6223 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6225 rtx prev = insn;
6226 /* Scan for the previous nonnote insn, but stop at a basic
6227 block boundary. */
6230 prev = PREV_INSN (prev);
6232 while (prev && NOTE_P (prev)
6233 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6235 /* Do not swap the registers around if the previous instruction
6236 attaches a REG_EQUIV note to REG1.
6238 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6239 from the pseudo that originally shadowed an incoming argument
6240 to another register. Some uses of REG_EQUIV might rely on it
6241 being attached to REG1 rather than REG2.
6243 This section previously turned the REG_EQUIV into a REG_EQUAL
6244 note. We cannot do that because REG_EQUIV may provide an
6245 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6247 if (prev != 0 && NONJUMP_INSN_P (prev)
6248 && GET_CODE (PATTERN (prev)) == SET
6249 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6250 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6252 rtx dest = SET_DEST (sets[0].rtl);
6253 rtx src = SET_SRC (sets[0].rtl);
6254 rtx note;
6256 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6257 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6258 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6259 apply_change_group ();
6261 /* If INSN has a REG_EQUAL note, and this note mentions
6262 REG0, then we must delete it, because the value in
6263 REG0 has changed. If the note's value is REG1, we must
6264 also delete it because that is now this insn's dest. */
6265 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6266 if (note != 0
6267 && (reg_mentioned_p (dest, XEXP (note, 0))
6268 || rtx_equal_p (src, XEXP (note, 0))))
6269 remove_note (insn, note);
6274 /* If this is a conditional jump insn, record any known equivalences due to
6275 the condition being tested. */
6277 if (JUMP_P (insn)
6278 && n_sets == 1 && GET_CODE (x) == SET
6279 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6280 record_jump_equiv (insn, 0);
6282 #ifdef HAVE_cc0
6283 /* If the previous insn set CC0 and this insn no longer references CC0,
6284 delete the previous insn. Here we use the fact that nothing expects CC0
6285 to be valid over an insn, which is true until the final pass. */
6286 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6287 && (tem = single_set (prev_insn)) != 0
6288 && SET_DEST (tem) == cc0_rtx
6289 && ! reg_mentioned_p (cc0_rtx, x))
6290 delete_insn (prev_insn);
6292 prev_insn_cc0 = this_insn_cc0;
6293 prev_insn_cc0_mode = this_insn_cc0_mode;
6294 prev_insn = insn;
6295 #endif
6298 /* Remove from the hash table all expressions that reference memory. */
6300 static void
6301 invalidate_memory (void)
6303 int i;
6304 struct table_elt *p, *next;
6306 for (i = 0; i < HASH_SIZE; i++)
6307 for (p = table[i]; p; p = next)
6309 next = p->next_same_hash;
6310 if (p->in_memory)
6311 remove_from_table (p, i);
6315 /* If ADDR is an address that implicitly affects the stack pointer, return
6316 1 and update the register tables to show the effect. Else, return 0. */
6318 static int
6319 addr_affects_sp_p (rtx addr)
6321 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6322 && REG_P (XEXP (addr, 0))
6323 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6325 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6327 REG_TICK (STACK_POINTER_REGNUM)++;
6328 /* Is it possible to use a subreg of SP? */
6329 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6332 /* This should be *very* rare. */
6333 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6334 invalidate (stack_pointer_rtx, VOIDmode);
6336 return 1;
6339 return 0;
6342 /* Perform invalidation on the basis of everything about an insn
6343 except for invalidating the actual places that are SET in it.
6344 This includes the places CLOBBERed, and anything that might
6345 alias with something that is SET or CLOBBERed.
6347 X is the pattern of the insn. */
6349 static void
6350 invalidate_from_clobbers (rtx x)
6352 if (GET_CODE (x) == CLOBBER)
6354 rtx ref = XEXP (x, 0);
6355 if (ref)
6357 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6358 || MEM_P (ref))
6359 invalidate (ref, VOIDmode);
6360 else if (GET_CODE (ref) == STRICT_LOW_PART
6361 || GET_CODE (ref) == ZERO_EXTRACT)
6362 invalidate (XEXP (ref, 0), GET_MODE (ref));
6365 else if (GET_CODE (x) == PARALLEL)
6367 int i;
6368 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6370 rtx y = XVECEXP (x, 0, i);
6371 if (GET_CODE (y) == CLOBBER)
6373 rtx ref = XEXP (y, 0);
6374 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6375 || MEM_P (ref))
6376 invalidate (ref, VOIDmode);
6377 else if (GET_CODE (ref) == STRICT_LOW_PART
6378 || GET_CODE (ref) == ZERO_EXTRACT)
6379 invalidate (XEXP (ref, 0), GET_MODE (ref));
6385 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6386 and replace any registers in them with either an equivalent constant
6387 or the canonical form of the register. If we are inside an address,
6388 only do this if the address remains valid.
6390 OBJECT is 0 except when within a MEM in which case it is the MEM.
6392 Return the replacement for X. */
6394 static rtx
6395 cse_process_notes (rtx x, rtx object)
6397 enum rtx_code code = GET_CODE (x);
6398 const char *fmt = GET_RTX_FORMAT (code);
6399 int i;
6401 switch (code)
6403 case CONST_INT:
6404 case CONST:
6405 case SYMBOL_REF:
6406 case LABEL_REF:
6407 case CONST_DOUBLE:
6408 case CONST_VECTOR:
6409 case PC:
6410 case CC0:
6411 case LO_SUM:
6412 return x;
6414 case MEM:
6415 validate_change (x, &XEXP (x, 0),
6416 cse_process_notes (XEXP (x, 0), x), 0);
6417 return x;
6419 case EXPR_LIST:
6420 case INSN_LIST:
6421 if (REG_NOTE_KIND (x) == REG_EQUAL)
6422 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6423 if (XEXP (x, 1))
6424 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6425 return x;
6427 case SIGN_EXTEND:
6428 case ZERO_EXTEND:
6429 case SUBREG:
6431 rtx new = cse_process_notes (XEXP (x, 0), object);
6432 /* We don't substitute VOIDmode constants into these rtx,
6433 since they would impede folding. */
6434 if (GET_MODE (new) != VOIDmode)
6435 validate_change (object, &XEXP (x, 0), new, 0);
6436 return x;
6439 case REG:
6440 i = REG_QTY (REGNO (x));
6442 /* Return a constant or a constant register. */
6443 if (REGNO_QTY_VALID_P (REGNO (x)))
6445 struct qty_table_elem *ent = &qty_table[i];
6447 if (ent->const_rtx != NULL_RTX
6448 && (CONSTANT_P (ent->const_rtx)
6449 || REG_P (ent->const_rtx)))
6451 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6452 if (new)
6453 return new;
6457 /* Otherwise, canonicalize this register. */
6458 return canon_reg (x, NULL_RTX);
6460 default:
6461 break;
6464 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6465 if (fmt[i] == 'e')
6466 validate_change (object, &XEXP (x, i),
6467 cse_process_notes (XEXP (x, i), object), 0);
6469 return x;
6472 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6473 since they are done elsewhere. This function is called via note_stores. */
6475 static void
6476 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6478 enum rtx_code code = GET_CODE (dest);
6480 if (code == MEM
6481 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6482 /* There are times when an address can appear varying and be a PLUS
6483 during this scan when it would be a fixed address were we to know
6484 the proper equivalences. So invalidate all memory if there is
6485 a BLKmode or nonscalar memory reference or a reference to a
6486 variable address. */
6487 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6488 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6490 invalidate_memory ();
6491 return;
6494 if (GET_CODE (set) == CLOBBER
6495 || CC0_P (dest)
6496 || dest == pc_rtx)
6497 return;
6499 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6500 invalidate (XEXP (dest, 0), GET_MODE (dest));
6501 else if (code == REG || code == SUBREG || code == MEM)
6502 invalidate (dest, VOIDmode);
6505 /* Invalidate all insns from START up to the end of the function or the
6506 next label. This called when we wish to CSE around a block that is
6507 conditionally executed. */
6509 static void
6510 invalidate_skipped_block (rtx start)
6512 rtx insn;
6514 for (insn = start; insn && !LABEL_P (insn);
6515 insn = NEXT_INSN (insn))
6517 if (! INSN_P (insn))
6518 continue;
6520 if (CALL_P (insn))
6522 if (! CONST_OR_PURE_CALL_P (insn))
6523 invalidate_memory ();
6524 invalidate_for_call ();
6527 invalidate_from_clobbers (PATTERN (insn));
6528 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6532 /* Find the end of INSN's basic block and return its range,
6533 the total number of SETs in all the insns of the block, the last insn of the
6534 block, and the branch path.
6536 The branch path indicates which branches should be followed. If a nonzero
6537 path size is specified, the block should be rescanned and a different set
6538 of branches will be taken. The branch path is only used if
6539 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6541 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6542 used to describe the block. It is filled in with the information about
6543 the current block. The incoming structure's branch path, if any, is used
6544 to construct the output branch path. */
6546 static void
6547 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6548 int follow_jumps, int skip_blocks)
6550 rtx p = insn, q;
6551 int nsets = 0;
6552 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6553 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6554 int path_size = data->path_size;
6555 int path_entry = 0;
6556 int i;
6558 /* Update the previous branch path, if any. If the last branch was
6559 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6560 If it was previously PATH_NOT_TAKEN,
6561 shorten the path by one and look at the previous branch. We know that
6562 at least one branch must have been taken if PATH_SIZE is nonzero. */
6563 while (path_size > 0)
6565 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6567 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6568 break;
6570 else
6571 path_size--;
6574 /* If the first instruction is marked with QImode, that means we've
6575 already processed this block. Our caller will look at DATA->LAST
6576 to figure out where to go next. We want to return the next block
6577 in the instruction stream, not some branched-to block somewhere
6578 else. We accomplish this by pretending our called forbid us to
6579 follow jumps, or skip blocks. */
6580 if (GET_MODE (insn) == QImode)
6581 follow_jumps = skip_blocks = 0;
6583 /* Scan to end of this basic block. */
6584 while (p && !LABEL_P (p))
6586 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6587 the regs restored by the longjmp come from
6588 a later time than the setjmp. */
6589 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6590 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6591 break;
6593 /* A PARALLEL can have lots of SETs in it,
6594 especially if it is really an ASM_OPERANDS. */
6595 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6596 nsets += XVECLEN (PATTERN (p), 0);
6597 else if (!NOTE_P (p))
6598 nsets += 1;
6600 /* Ignore insns made by CSE; they cannot affect the boundaries of
6601 the basic block. */
6603 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6604 high_cuid = INSN_CUID (p);
6605 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6606 low_cuid = INSN_CUID (p);
6608 /* See if this insn is in our branch path. If it is and we are to
6609 take it, do so. */
6610 if (path_entry < path_size && data->path[path_entry].branch == p)
6612 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6613 p = JUMP_LABEL (p);
6615 /* Point to next entry in path, if any. */
6616 path_entry++;
6619 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6620 was specified, we haven't reached our maximum path length, there are
6621 insns following the target of the jump, this is the only use of the
6622 jump label, and the target label is preceded by a BARRIER.
6624 Alternatively, we can follow the jump if it branches around a
6625 block of code and there are no other branches into the block.
6626 In this case invalidate_skipped_block will be called to invalidate any
6627 registers set in the block when following the jump. */
6629 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6630 && JUMP_P (p)
6631 && GET_CODE (PATTERN (p)) == SET
6632 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6633 && JUMP_LABEL (p) != 0
6634 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6635 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6637 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6638 if ((!NOTE_P (q)
6639 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6640 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6641 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6642 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6643 break;
6645 /* If we ran into a BARRIER, this code is an extension of the
6646 basic block when the branch is taken. */
6647 if (follow_jumps && q != 0 && BARRIER_P (q))
6649 /* Don't allow ourself to keep walking around an
6650 always-executed loop. */
6651 if (next_real_insn (q) == next)
6653 p = NEXT_INSN (p);
6654 continue;
6657 /* Similarly, don't put a branch in our path more than once. */
6658 for (i = 0; i < path_entry; i++)
6659 if (data->path[i].branch == p)
6660 break;
6662 if (i != path_entry)
6663 break;
6665 data->path[path_entry].branch = p;
6666 data->path[path_entry++].status = PATH_TAKEN;
6668 /* This branch now ends our path. It was possible that we
6669 didn't see this branch the last time around (when the
6670 insn in front of the target was a JUMP_INSN that was
6671 turned into a no-op). */
6672 path_size = path_entry;
6674 p = JUMP_LABEL (p);
6675 /* Mark block so we won't scan it again later. */
6676 PUT_MODE (NEXT_INSN (p), QImode);
6678 /* Detect a branch around a block of code. */
6679 else if (skip_blocks && q != 0 && !LABEL_P (q))
6681 rtx tmp;
6683 if (next_real_insn (q) == next)
6685 p = NEXT_INSN (p);
6686 continue;
6689 for (i = 0; i < path_entry; i++)
6690 if (data->path[i].branch == p)
6691 break;
6693 if (i != path_entry)
6694 break;
6696 /* This is no_labels_between_p (p, q) with an added check for
6697 reaching the end of a function (in case Q precedes P). */
6698 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6699 if (LABEL_P (tmp))
6700 break;
6702 if (tmp == q)
6704 data->path[path_entry].branch = p;
6705 data->path[path_entry++].status = PATH_AROUND;
6707 path_size = path_entry;
6709 p = JUMP_LABEL (p);
6710 /* Mark block so we won't scan it again later. */
6711 PUT_MODE (NEXT_INSN (p), QImode);
6715 p = NEXT_INSN (p);
6718 data->low_cuid = low_cuid;
6719 data->high_cuid = high_cuid;
6720 data->nsets = nsets;
6721 data->last = p;
6723 /* If all jumps in the path are not taken, set our path length to zero
6724 so a rescan won't be done. */
6725 for (i = path_size - 1; i >= 0; i--)
6726 if (data->path[i].status != PATH_NOT_TAKEN)
6727 break;
6729 if (i == -1)
6730 data->path_size = 0;
6731 else
6732 data->path_size = path_size;
6734 /* End the current branch path. */
6735 data->path[path_size].branch = 0;
6738 /* Perform cse on the instructions of a function.
6739 F is the first instruction.
6740 NREGS is one plus the highest pseudo-reg number used in the instruction.
6742 Returns 1 if jump_optimize should be redone due to simplifications
6743 in conditional jump instructions. */
6746 cse_main (rtx f, int nregs, FILE *file)
6748 struct cse_basic_block_data val;
6749 rtx insn = f;
6750 int i;
6752 init_cse_reg_info (nregs);
6754 val.path = xmalloc (sizeof (struct branch_path)
6755 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6757 cse_jumps_altered = 0;
6758 recorded_label_ref = 0;
6759 constant_pool_entries_cost = 0;
6760 constant_pool_entries_regcost = 0;
6761 val.path_size = 0;
6762 rtl_hooks = cse_rtl_hooks;
6764 init_recog ();
6765 init_alias_analysis ();
6767 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6769 /* Find the largest uid. */
6771 max_uid = get_max_uid ();
6772 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6774 /* Compute the mapping from uids to cuids.
6775 CUIDs are numbers assigned to insns, like uids,
6776 except that cuids increase monotonically through the code.
6777 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6778 between two insns is not affected by -g. */
6780 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6782 if (!NOTE_P (insn)
6783 || NOTE_LINE_NUMBER (insn) < 0)
6784 INSN_CUID (insn) = ++i;
6785 else
6786 /* Give a line number note the same cuid as preceding insn. */
6787 INSN_CUID (insn) = i;
6790 /* Loop over basic blocks.
6791 Compute the maximum number of qty's needed for each basic block
6792 (which is 2 for each SET). */
6793 insn = f;
6794 while (insn)
6796 cse_altered = 0;
6797 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6798 flag_cse_skip_blocks);
6800 /* If this basic block was already processed or has no sets, skip it. */
6801 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6803 PUT_MODE (insn, VOIDmode);
6804 insn = (val.last ? NEXT_INSN (val.last) : 0);
6805 val.path_size = 0;
6806 continue;
6809 cse_basic_block_start = val.low_cuid;
6810 cse_basic_block_end = val.high_cuid;
6811 max_qty = val.nsets * 2;
6813 if (file)
6814 fprintf (file, ";; Processing block from %d to %d, %d sets.\n",
6815 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6816 val.nsets);
6818 /* Make MAX_QTY bigger to give us room to optimize
6819 past the end of this basic block, if that should prove useful. */
6820 if (max_qty < 500)
6821 max_qty = 500;
6823 /* If this basic block is being extended by following certain jumps,
6824 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6825 Otherwise, we start after this basic block. */
6826 if (val.path_size > 0)
6827 cse_basic_block (insn, val.last, val.path);
6828 else
6830 int old_cse_jumps_altered = cse_jumps_altered;
6831 rtx temp;
6833 /* When cse changes a conditional jump to an unconditional
6834 jump, we want to reprocess the block, since it will give
6835 us a new branch path to investigate. */
6836 cse_jumps_altered = 0;
6837 temp = cse_basic_block (insn, val.last, val.path);
6838 if (cse_jumps_altered == 0
6839 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6840 insn = temp;
6842 cse_jumps_altered |= old_cse_jumps_altered;
6845 if (cse_altered)
6846 ggc_collect ();
6848 #ifdef USE_C_ALLOCA
6849 alloca (0);
6850 #endif
6853 /* Clean up. */
6854 end_alias_analysis ();
6855 free (uid_cuid);
6856 free (reg_eqv_table);
6857 free (val.path);
6858 rtl_hooks = general_rtl_hooks;
6860 return cse_jumps_altered || recorded_label_ref;
6863 /* Process a single basic block. FROM and TO and the limits of the basic
6864 block. NEXT_BRANCH points to the branch path when following jumps or
6865 a null path when not following jumps. */
6867 static rtx
6868 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
6870 rtx insn;
6871 int to_usage = 0;
6872 rtx libcall_insn = NULL_RTX;
6873 int num_insns = 0;
6874 int no_conflict = 0;
6876 /* Allocate the space needed by qty_table. */
6877 qty_table = xmalloc (max_qty * sizeof (struct qty_table_elem));
6879 new_basic_block ();
6881 /* TO might be a label. If so, protect it from being deleted. */
6882 if (to != 0 && LABEL_P (to))
6883 ++LABEL_NUSES (to);
6885 for (insn = from; insn != to; insn = NEXT_INSN (insn))
6887 enum rtx_code code = GET_CODE (insn);
6889 /* If we have processed 1,000 insns, flush the hash table to
6890 avoid extreme quadratic behavior. We must not include NOTEs
6891 in the count since there may be more of them when generating
6892 debugging information. If we clear the table at different
6893 times, code generated with -g -O might be different than code
6894 generated with -O but not -g.
6896 ??? This is a real kludge and needs to be done some other way.
6897 Perhaps for 2.9. */
6898 if (code != NOTE && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6900 flush_hash_table ();
6901 num_insns = 0;
6904 /* See if this is a branch that is part of the path. If so, and it is
6905 to be taken, do so. */
6906 if (next_branch->branch == insn)
6908 enum taken status = next_branch++->status;
6909 if (status != PATH_NOT_TAKEN)
6911 if (status == PATH_TAKEN)
6912 record_jump_equiv (insn, 1);
6913 else
6914 invalidate_skipped_block (NEXT_INSN (insn));
6916 /* Set the last insn as the jump insn; it doesn't affect cc0.
6917 Then follow this branch. */
6918 #ifdef HAVE_cc0
6919 prev_insn_cc0 = 0;
6920 prev_insn = insn;
6921 #endif
6922 insn = JUMP_LABEL (insn);
6923 continue;
6927 if (GET_MODE (insn) == QImode)
6928 PUT_MODE (insn, VOIDmode);
6930 if (GET_RTX_CLASS (code) == RTX_INSN)
6932 rtx p;
6934 /* Process notes first so we have all notes in canonical forms when
6935 looking for duplicate operations. */
6937 if (REG_NOTES (insn))
6938 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
6940 /* Track when we are inside in LIBCALL block. Inside such a block,
6941 we do not want to record destinations. The last insn of a
6942 LIBCALL block is not considered to be part of the block, since
6943 its destination is the result of the block and hence should be
6944 recorded. */
6946 if (REG_NOTES (insn) != 0)
6948 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
6949 libcall_insn = XEXP (p, 0);
6950 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
6952 /* Keep libcall_insn for the last SET insn of a no-conflict
6953 block to prevent changing the destination. */
6954 if (! no_conflict)
6955 libcall_insn = 0;
6956 else
6957 no_conflict = -1;
6959 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
6960 no_conflict = 1;
6963 cse_insn (insn, libcall_insn);
6965 if (no_conflict == -1)
6967 libcall_insn = 0;
6968 no_conflict = 0;
6971 /* If we haven't already found an insn where we added a LABEL_REF,
6972 check this one. */
6973 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
6974 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6975 (void *) insn))
6976 recorded_label_ref = 1;
6979 /* If INSN is now an unconditional jump, skip to the end of our
6980 basic block by pretending that we just did the last insn in the
6981 basic block. If we are jumping to the end of our block, show
6982 that we can have one usage of TO. */
6984 if (any_uncondjump_p (insn))
6986 if (to == 0)
6988 free (qty_table);
6989 return 0;
6992 if (JUMP_LABEL (insn) == to)
6993 to_usage = 1;
6995 /* Maybe TO was deleted because the jump is unconditional.
6996 If so, there is nothing left in this basic block. */
6997 /* ??? Perhaps it would be smarter to set TO
6998 to whatever follows this insn,
6999 and pretend the basic block had always ended here. */
7000 if (INSN_DELETED_P (to))
7001 break;
7003 insn = PREV_INSN (to);
7006 /* See if it is ok to keep on going past the label
7007 which used to end our basic block. Remember that we incremented
7008 the count of that label, so we decrement it here. If we made
7009 a jump unconditional, TO_USAGE will be one; in that case, we don't
7010 want to count the use in that jump. */
7012 if (to != 0 && NEXT_INSN (insn) == to
7013 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7015 struct cse_basic_block_data val;
7016 rtx prev;
7018 insn = NEXT_INSN (to);
7020 /* If TO was the last insn in the function, we are done. */
7021 if (insn == 0)
7023 free (qty_table);
7024 return 0;
7027 /* If TO was preceded by a BARRIER we are done with this block
7028 because it has no continuation. */
7029 prev = prev_nonnote_insn (to);
7030 if (prev && BARRIER_P (prev))
7032 free (qty_table);
7033 return insn;
7036 /* Find the end of the following block. Note that we won't be
7037 following branches in this case. */
7038 to_usage = 0;
7039 val.path_size = 0;
7040 val.path = xmalloc (sizeof (struct branch_path)
7041 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7042 cse_end_of_basic_block (insn, &val, 0, 0);
7043 free (val.path);
7045 /* If the tables we allocated have enough space left
7046 to handle all the SETs in the next basic block,
7047 continue through it. Otherwise, return,
7048 and that block will be scanned individually. */
7049 if (val.nsets * 2 + next_qty > max_qty)
7050 break;
7052 cse_basic_block_start = val.low_cuid;
7053 cse_basic_block_end = val.high_cuid;
7054 to = val.last;
7056 /* Prevent TO from being deleted if it is a label. */
7057 if (to != 0 && LABEL_P (to))
7058 ++LABEL_NUSES (to);
7060 /* Back up so we process the first insn in the extension. */
7061 insn = PREV_INSN (insn);
7065 gcc_assert (next_qty <= max_qty);
7067 free (qty_table);
7069 return to ? NEXT_INSN (to) : 0;
7072 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7073 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7075 static int
7076 check_for_label_ref (rtx *rtl, void *data)
7078 rtx insn = (rtx) data;
7080 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7081 we must rerun jump since it needs to place the note. If this is a
7082 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7083 since no REG_LABEL will be added. */
7084 return (GET_CODE (*rtl) == LABEL_REF
7085 && ! LABEL_REF_NONLOCAL_P (*rtl)
7086 && LABEL_P (XEXP (*rtl, 0))
7087 && INSN_UID (XEXP (*rtl, 0)) != 0
7088 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7091 /* Count the number of times registers are used (not set) in X.
7092 COUNTS is an array in which we accumulate the count, INCR is how much
7093 we count each register usage.
7095 Don't count a usage of DEST, which is the SET_DEST of a SET which
7096 contains X in its SET_SRC. This is because such a SET does not
7097 modify the liveness of DEST.
7098 DEST is set to pc_rtx for a trapping insn, which means that we must count
7099 uses of a SET_DEST regardless because the insn can't be deleted here. */
7101 static void
7102 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
7104 enum rtx_code code;
7105 rtx note;
7106 const char *fmt;
7107 int i, j;
7109 if (x == 0)
7110 return;
7112 switch (code = GET_CODE (x))
7114 case REG:
7115 if (x != dest)
7116 counts[REGNO (x)] += incr;
7117 return;
7119 case PC:
7120 case CC0:
7121 case CONST:
7122 case CONST_INT:
7123 case CONST_DOUBLE:
7124 case CONST_VECTOR:
7125 case SYMBOL_REF:
7126 case LABEL_REF:
7127 return;
7129 case CLOBBER:
7130 /* If we are clobbering a MEM, mark any registers inside the address
7131 as being used. */
7132 if (MEM_P (XEXP (x, 0)))
7133 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7134 return;
7136 case SET:
7137 /* Unless we are setting a REG, count everything in SET_DEST. */
7138 if (!REG_P (SET_DEST (x)))
7139 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7140 count_reg_usage (SET_SRC (x), counts,
7141 dest ? dest : SET_DEST (x),
7142 incr);
7143 return;
7145 case CALL_INSN:
7146 case INSN:
7147 case JUMP_INSN:
7148 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
7149 this fact by setting DEST to pc_rtx. */
7150 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
7151 dest = pc_rtx;
7152 if (code == CALL_INSN)
7153 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
7154 count_reg_usage (PATTERN (x), counts, dest, incr);
7156 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7157 use them. */
7159 note = find_reg_equal_equiv_note (x);
7160 if (note)
7162 rtx eqv = XEXP (note, 0);
7164 if (GET_CODE (eqv) == EXPR_LIST)
7165 /* This REG_EQUAL note describes the result of a function call.
7166 Process all the arguments. */
7169 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
7170 eqv = XEXP (eqv, 1);
7172 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7173 else
7174 count_reg_usage (eqv, counts, dest, incr);
7176 return;
7178 case EXPR_LIST:
7179 if (REG_NOTE_KIND (x) == REG_EQUAL
7180 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7181 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7182 involving registers in the address. */
7183 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7184 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7186 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7187 return;
7189 case ASM_OPERANDS:
7190 /* If the asm is volatile, then this insn cannot be deleted,
7191 and so the inputs *must* be live. */
7192 if (MEM_VOLATILE_P (x))
7193 dest = NULL_RTX;
7194 /* Iterate over just the inputs, not the constraints as well. */
7195 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7196 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
7197 return;
7199 case INSN_LIST:
7200 gcc_unreachable ();
7202 default:
7203 break;
7206 fmt = GET_RTX_FORMAT (code);
7207 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7209 if (fmt[i] == 'e')
7210 count_reg_usage (XEXP (x, i), counts, dest, incr);
7211 else if (fmt[i] == 'E')
7212 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7213 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7217 /* Return true if set is live. */
7218 static bool
7219 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7220 int *counts)
7222 #ifdef HAVE_cc0
7223 rtx tem;
7224 #endif
7226 if (set_noop_p (set))
7229 #ifdef HAVE_cc0
7230 else if (GET_CODE (SET_DEST (set)) == CC0
7231 && !side_effects_p (SET_SRC (set))
7232 && ((tem = next_nonnote_insn (insn)) == 0
7233 || !INSN_P (tem)
7234 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7235 return false;
7236 #endif
7237 else if (!REG_P (SET_DEST (set))
7238 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7239 || counts[REGNO (SET_DEST (set))] != 0
7240 || side_effects_p (SET_SRC (set)))
7241 return true;
7242 return false;
7245 /* Return true if insn is live. */
7247 static bool
7248 insn_live_p (rtx insn, int *counts)
7250 int i;
7251 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7252 return true;
7253 else if (GET_CODE (PATTERN (insn)) == SET)
7254 return set_live_p (PATTERN (insn), insn, counts);
7255 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7257 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7259 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7261 if (GET_CODE (elt) == SET)
7263 if (set_live_p (elt, insn, counts))
7264 return true;
7266 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7267 return true;
7269 return false;
7271 else
7272 return true;
7275 /* Return true if libcall is dead as a whole. */
7277 static bool
7278 dead_libcall_p (rtx insn, int *counts)
7280 rtx note, set, new;
7282 /* See if there's a REG_EQUAL note on this insn and try to
7283 replace the source with the REG_EQUAL expression.
7285 We assume that insns with REG_RETVALs can only be reg->reg
7286 copies at this point. */
7287 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7288 if (!note)
7289 return false;
7291 set = single_set (insn);
7292 if (!set)
7293 return false;
7295 new = simplify_rtx (XEXP (note, 0));
7296 if (!new)
7297 new = XEXP (note, 0);
7299 /* While changing insn, we must update the counts accordingly. */
7300 count_reg_usage (insn, counts, NULL_RTX, -1);
7302 if (validate_change (insn, &SET_SRC (set), new, 0))
7304 count_reg_usage (insn, counts, NULL_RTX, 1);
7305 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7306 remove_note (insn, note);
7307 return true;
7310 if (CONSTANT_P (new))
7312 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7313 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7315 count_reg_usage (insn, counts, NULL_RTX, 1);
7316 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7317 remove_note (insn, note);
7318 return true;
7322 count_reg_usage (insn, counts, NULL_RTX, 1);
7323 return false;
7326 /* Scan all the insns and delete any that are dead; i.e., they store a register
7327 that is never used or they copy a register to itself.
7329 This is used to remove insns made obviously dead by cse, loop or other
7330 optimizations. It improves the heuristics in loop since it won't try to
7331 move dead invariants out of loops or make givs for dead quantities. The
7332 remaining passes of the compilation are also sped up. */
7335 delete_trivially_dead_insns (rtx insns, int nreg)
7337 int *counts;
7338 rtx insn, prev;
7339 int in_libcall = 0, dead_libcall = 0;
7340 int ndead = 0;
7342 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7343 /* First count the number of times each register is used. */
7344 counts = xcalloc (nreg, sizeof (int));
7345 for (insn = insns; insn; insn = NEXT_INSN (insn))
7346 if (INSN_P (insn))
7347 count_reg_usage (insn, counts, NULL_RTX, 1);
7349 /* Go from the last insn to the first and delete insns that only set unused
7350 registers or copy a register to itself. As we delete an insn, remove
7351 usage counts for registers it uses.
7353 The first jump optimization pass may leave a real insn as the last
7354 insn in the function. We must not skip that insn or we may end
7355 up deleting code that is not really dead. */
7356 for (insn = get_last_insn (); insn; insn = prev)
7358 int live_insn = 0;
7360 prev = PREV_INSN (insn);
7361 if (!INSN_P (insn))
7362 continue;
7364 /* Don't delete any insns that are part of a libcall block unless
7365 we can delete the whole libcall block.
7367 Flow or loop might get confused if we did that. Remember
7368 that we are scanning backwards. */
7369 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7371 in_libcall = 1;
7372 live_insn = 1;
7373 dead_libcall = dead_libcall_p (insn, counts);
7375 else if (in_libcall)
7376 live_insn = ! dead_libcall;
7377 else
7378 live_insn = insn_live_p (insn, counts);
7380 /* If this is a dead insn, delete it and show registers in it aren't
7381 being used. */
7383 if (! live_insn)
7385 count_reg_usage (insn, counts, NULL_RTX, -1);
7386 delete_insn_and_edges (insn);
7387 ndead++;
7390 if (in_libcall && find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7392 in_libcall = 0;
7393 dead_libcall = 0;
7397 if (dump_file && ndead)
7398 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7399 ndead);
7400 /* Clean up. */
7401 free (counts);
7402 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7403 return ndead;
7406 /* This function is called via for_each_rtx. The argument, NEWREG, is
7407 a condition code register with the desired mode. If we are looking
7408 at the same register in a different mode, replace it with
7409 NEWREG. */
7411 static int
7412 cse_change_cc_mode (rtx *loc, void *data)
7414 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7416 if (*loc
7417 && REG_P (*loc)
7418 && REGNO (*loc) == REGNO (args->newreg)
7419 && GET_MODE (*loc) != GET_MODE (args->newreg))
7421 validate_change (args->insn, loc, args->newreg, 1);
7423 return -1;
7425 return 0;
7428 /* Change the mode of any reference to the register REGNO (NEWREG) to
7429 GET_MODE (NEWREG) in INSN. */
7431 static void
7432 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7434 struct change_cc_mode_args args;
7435 int success;
7437 if (!INSN_P (insn))
7438 return;
7440 args.insn = insn;
7441 args.newreg = newreg;
7443 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7444 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7446 /* If the following assertion was triggered, there is most probably
7447 something wrong with the cc_modes_compatible back end function.
7448 CC modes only can be considered compatible if the insn - with the mode
7449 replaced by any of the compatible modes - can still be recognized. */
7450 success = apply_change_group ();
7451 gcc_assert (success);
7454 /* Change the mode of any reference to the register REGNO (NEWREG) to
7455 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7456 any instruction which modifies NEWREG. */
7458 static void
7459 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7461 rtx insn;
7463 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7465 if (! INSN_P (insn))
7466 continue;
7468 if (reg_set_p (newreg, insn))
7469 return;
7471 cse_change_cc_mode_insn (insn, newreg);
7475 /* BB is a basic block which finishes with CC_REG as a condition code
7476 register which is set to CC_SRC. Look through the successors of BB
7477 to find blocks which have a single predecessor (i.e., this one),
7478 and look through those blocks for an assignment to CC_REG which is
7479 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7480 permitted to change the mode of CC_SRC to a compatible mode. This
7481 returns VOIDmode if no equivalent assignments were found.
7482 Otherwise it returns the mode which CC_SRC should wind up with.
7484 The main complexity in this function is handling the mode issues.
7485 We may have more than one duplicate which we can eliminate, and we
7486 try to find a mode which will work for multiple duplicates. */
7488 static enum machine_mode
7489 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7491 bool found_equiv;
7492 enum machine_mode mode;
7493 unsigned int insn_count;
7494 edge e;
7495 rtx insns[2];
7496 enum machine_mode modes[2];
7497 rtx last_insns[2];
7498 unsigned int i;
7499 rtx newreg;
7500 edge_iterator ei;
7502 /* We expect to have two successors. Look at both before picking
7503 the final mode for the comparison. If we have more successors
7504 (i.e., some sort of table jump, although that seems unlikely),
7505 then we require all beyond the first two to use the same
7506 mode. */
7508 found_equiv = false;
7509 mode = GET_MODE (cc_src);
7510 insn_count = 0;
7511 FOR_EACH_EDGE (e, ei, bb->succs)
7513 rtx insn;
7514 rtx end;
7516 if (e->flags & EDGE_COMPLEX)
7517 continue;
7519 if (EDGE_COUNT (e->dest->preds) != 1
7520 || e->dest == EXIT_BLOCK_PTR)
7521 continue;
7523 end = NEXT_INSN (BB_END (e->dest));
7524 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7526 rtx set;
7528 if (! INSN_P (insn))
7529 continue;
7531 /* If CC_SRC is modified, we have to stop looking for
7532 something which uses it. */
7533 if (modified_in_p (cc_src, insn))
7534 break;
7536 /* Check whether INSN sets CC_REG to CC_SRC. */
7537 set = single_set (insn);
7538 if (set
7539 && REG_P (SET_DEST (set))
7540 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7542 bool found;
7543 enum machine_mode set_mode;
7544 enum machine_mode comp_mode;
7546 found = false;
7547 set_mode = GET_MODE (SET_SRC (set));
7548 comp_mode = set_mode;
7549 if (rtx_equal_p (cc_src, SET_SRC (set)))
7550 found = true;
7551 else if (GET_CODE (cc_src) == COMPARE
7552 && GET_CODE (SET_SRC (set)) == COMPARE
7553 && mode != set_mode
7554 && rtx_equal_p (XEXP (cc_src, 0),
7555 XEXP (SET_SRC (set), 0))
7556 && rtx_equal_p (XEXP (cc_src, 1),
7557 XEXP (SET_SRC (set), 1)))
7560 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7561 if (comp_mode != VOIDmode
7562 && (can_change_mode || comp_mode == mode))
7563 found = true;
7566 if (found)
7568 found_equiv = true;
7569 if (insn_count < ARRAY_SIZE (insns))
7571 insns[insn_count] = insn;
7572 modes[insn_count] = set_mode;
7573 last_insns[insn_count] = end;
7574 ++insn_count;
7576 if (mode != comp_mode)
7578 gcc_assert (can_change_mode);
7579 mode = comp_mode;
7581 /* The modified insn will be re-recognized later. */
7582 PUT_MODE (cc_src, mode);
7585 else
7587 if (set_mode != mode)
7589 /* We found a matching expression in the
7590 wrong mode, but we don't have room to
7591 store it in the array. Punt. This case
7592 should be rare. */
7593 break;
7595 /* INSN sets CC_REG to a value equal to CC_SRC
7596 with the right mode. We can simply delete
7597 it. */
7598 delete_insn (insn);
7601 /* We found an instruction to delete. Keep looking,
7602 in the hopes of finding a three-way jump. */
7603 continue;
7606 /* We found an instruction which sets the condition
7607 code, so don't look any farther. */
7608 break;
7611 /* If INSN sets CC_REG in some other way, don't look any
7612 farther. */
7613 if (reg_set_p (cc_reg, insn))
7614 break;
7617 /* If we fell off the bottom of the block, we can keep looking
7618 through successors. We pass CAN_CHANGE_MODE as false because
7619 we aren't prepared to handle compatibility between the
7620 further blocks and this block. */
7621 if (insn == end)
7623 enum machine_mode submode;
7625 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7626 if (submode != VOIDmode)
7628 gcc_assert (submode == mode);
7629 found_equiv = true;
7630 can_change_mode = false;
7635 if (! found_equiv)
7636 return VOIDmode;
7638 /* Now INSN_COUNT is the number of instructions we found which set
7639 CC_REG to a value equivalent to CC_SRC. The instructions are in
7640 INSNS. The modes used by those instructions are in MODES. */
7642 newreg = NULL_RTX;
7643 for (i = 0; i < insn_count; ++i)
7645 if (modes[i] != mode)
7647 /* We need to change the mode of CC_REG in INSNS[i] and
7648 subsequent instructions. */
7649 if (! newreg)
7651 if (GET_MODE (cc_reg) == mode)
7652 newreg = cc_reg;
7653 else
7654 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7656 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7657 newreg);
7660 delete_insn (insns[i]);
7663 return mode;
7666 /* If we have a fixed condition code register (or two), walk through
7667 the instructions and try to eliminate duplicate assignments. */
7669 void
7670 cse_condition_code_reg (void)
7672 unsigned int cc_regno_1;
7673 unsigned int cc_regno_2;
7674 rtx cc_reg_1;
7675 rtx cc_reg_2;
7676 basic_block bb;
7678 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7679 return;
7681 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7682 if (cc_regno_2 != INVALID_REGNUM)
7683 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7684 else
7685 cc_reg_2 = NULL_RTX;
7687 FOR_EACH_BB (bb)
7689 rtx last_insn;
7690 rtx cc_reg;
7691 rtx insn;
7692 rtx cc_src_insn;
7693 rtx cc_src;
7694 enum machine_mode mode;
7695 enum machine_mode orig_mode;
7697 /* Look for blocks which end with a conditional jump based on a
7698 condition code register. Then look for the instruction which
7699 sets the condition code register. Then look through the
7700 successor blocks for instructions which set the condition
7701 code register to the same value. There are other possible
7702 uses of the condition code register, but these are by far the
7703 most common and the ones which we are most likely to be able
7704 to optimize. */
7706 last_insn = BB_END (bb);
7707 if (!JUMP_P (last_insn))
7708 continue;
7710 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7711 cc_reg = cc_reg_1;
7712 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7713 cc_reg = cc_reg_2;
7714 else
7715 continue;
7717 cc_src_insn = NULL_RTX;
7718 cc_src = NULL_RTX;
7719 for (insn = PREV_INSN (last_insn);
7720 insn && insn != PREV_INSN (BB_HEAD (bb));
7721 insn = PREV_INSN (insn))
7723 rtx set;
7725 if (! INSN_P (insn))
7726 continue;
7727 set = single_set (insn);
7728 if (set
7729 && REG_P (SET_DEST (set))
7730 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7732 cc_src_insn = insn;
7733 cc_src = SET_SRC (set);
7734 break;
7736 else if (reg_set_p (cc_reg, insn))
7737 break;
7740 if (! cc_src_insn)
7741 continue;
7743 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7744 continue;
7746 /* Now CC_REG is a condition code register used for a
7747 conditional jump at the end of the block, and CC_SRC, in
7748 CC_SRC_INSN, is the value to which that condition code
7749 register is set, and CC_SRC is still meaningful at the end of
7750 the basic block. */
7752 orig_mode = GET_MODE (cc_src);
7753 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7754 if (mode != VOIDmode)
7756 gcc_assert (mode == GET_MODE (cc_src));
7757 if (mode != orig_mode)
7759 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7761 cse_change_cc_mode_insn (cc_src_insn, newreg);
7763 /* Do the same in the following insns that use the
7764 current value of CC_REG within BB. */
7765 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7766 NEXT_INSN (last_insn),
7767 newreg);
7774 /* Perform common subexpression elimination. Nonzero value from
7775 `cse_main' means that jumps were simplified and some code may now
7776 be unreachable, so do jump optimization again. */
7777 static bool
7778 gate_handle_cse (void)
7780 return optimize > 0;
7783 static void
7784 rest_of_handle_cse (void)
7786 int tem;
7788 if (dump_file)
7789 dump_flow_info (dump_file);
7791 reg_scan (get_insns (), max_reg_num ());
7793 tem = cse_main (get_insns (), max_reg_num (), dump_file);
7794 if (tem)
7795 rebuild_jump_labels (get_insns ());
7796 if (purge_all_dead_edges ())
7797 delete_unreachable_blocks ();
7799 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7801 /* If we are not running more CSE passes, then we are no longer
7802 expecting CSE to be run. But always rerun it in a cheap mode. */
7803 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7805 if (tem)
7806 delete_dead_jumptables ();
7808 if (tem || optimize > 1)
7809 cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_PRE_LOOP);
7812 struct tree_opt_pass pass_cse =
7814 "cse1", /* name */
7815 gate_handle_cse, /* gate */
7816 rest_of_handle_cse, /* execute */
7817 NULL, /* sub */
7818 NULL, /* next */
7819 0, /* static_pass_number */
7820 TV_CSE, /* tv_id */
7821 0, /* properties_required */
7822 0, /* properties_provided */
7823 0, /* properties_destroyed */
7824 0, /* todo_flags_start */
7825 TODO_dump_func |
7826 TODO_ggc_collect, /* todo_flags_finish */
7827 's' /* letter */
7831 static bool
7832 gate_handle_cse2 (void)
7834 return optimize > 0 && flag_rerun_cse_after_loop;
7837 /* Run second CSE pass after loop optimizations. */
7838 static void
7839 rest_of_handle_cse2 (void)
7841 int tem;
7843 if (dump_file)
7844 dump_flow_info (dump_file);
7846 tem = cse_main (get_insns (), max_reg_num (), dump_file);
7848 /* Run a pass to eliminate duplicated assignments to condition code
7849 registers. We have to run this after bypass_jumps, because it
7850 makes it harder for that pass to determine whether a jump can be
7851 bypassed safely. */
7852 cse_condition_code_reg ();
7854 purge_all_dead_edges ();
7855 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7857 if (tem)
7859 timevar_push (TV_JUMP);
7860 rebuild_jump_labels (get_insns ());
7861 delete_dead_jumptables ();
7862 cleanup_cfg (CLEANUP_EXPENSIVE);
7863 timevar_pop (TV_JUMP);
7865 reg_scan (get_insns (), max_reg_num ());
7866 cse_not_expected = 1;
7870 struct tree_opt_pass pass_cse2 =
7872 "cse2", /* name */
7873 gate_handle_cse2, /* gate */
7874 rest_of_handle_cse2, /* execute */
7875 NULL, /* sub */
7876 NULL, /* next */
7877 0, /* static_pass_number */
7878 TV_CSE2, /* tv_id */
7879 0, /* properties_required */
7880 0, /* properties_provided */
7881 0, /* properties_destroyed */
7882 0, /* todo_flags_start */
7883 TODO_dump_func |
7884 TODO_ggc_collect, /* todo_flags_finish */
7885 't' /* letter */