2005-08-16 Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
[official-gcc.git] / gcc / cse.c
blob1efcf308ae4ced72e6314dae719ca3749cd900ef
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 || exp_equiv_p (p->exp, p->exp, 1, false)))
3004 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3005 p->exp, op1);
3006 int new_cost;
3008 /* Get the canonical version of the address so we can accept
3009 more. */
3010 new = canon_for_address (new);
3012 new_cost = address_cost (new, mode);
3014 if (new_cost < best_addr_cost
3015 || (new_cost == best_addr_cost
3016 && (COST (new) + 1) >> 1 > best_rtx_cost))
3018 found_better = 1;
3019 best_addr_cost = new_cost;
3020 best_rtx_cost = (COST (new) + 1) >> 1;
3021 best_elt = p;
3022 best_rtx = new;
3026 if (found_better)
3028 if (validate_change (insn, loc,
3029 canon_reg (copy_rtx (best_rtx),
3030 NULL_RTX), 0))
3031 return;
3032 else
3033 best_elt->flag = 1;
3039 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3040 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3041 what values are being compared.
3043 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3044 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3045 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3046 compared to produce cc0.
3048 The return value is the comparison operator and is either the code of
3049 A or the code corresponding to the inverse of the comparison. */
3051 static enum rtx_code
3052 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3053 enum machine_mode *pmode1, enum machine_mode *pmode2)
3055 rtx arg1, arg2;
3057 arg1 = *parg1, arg2 = *parg2;
3059 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3061 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3063 /* Set nonzero when we find something of interest. */
3064 rtx x = 0;
3065 int reverse_code = 0;
3066 struct table_elt *p = 0;
3068 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3069 On machines with CC0, this is the only case that can occur, since
3070 fold_rtx will return the COMPARE or item being compared with zero
3071 when given CC0. */
3073 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3074 x = arg1;
3076 /* If ARG1 is a comparison operator and CODE is testing for
3077 STORE_FLAG_VALUE, get the inner arguments. */
3079 else if (COMPARISON_P (arg1))
3081 #ifdef FLOAT_STORE_FLAG_VALUE
3082 REAL_VALUE_TYPE fsfv;
3083 #endif
3085 if (code == NE
3086 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3087 && code == LT && STORE_FLAG_VALUE == -1)
3088 #ifdef FLOAT_STORE_FLAG_VALUE
3089 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3090 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3091 REAL_VALUE_NEGATIVE (fsfv)))
3092 #endif
3094 x = arg1;
3095 else if (code == EQ
3096 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3097 && code == GE && STORE_FLAG_VALUE == -1)
3098 #ifdef FLOAT_STORE_FLAG_VALUE
3099 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_FLOAT
3100 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3101 REAL_VALUE_NEGATIVE (fsfv)))
3102 #endif
3104 x = arg1, reverse_code = 1;
3107 /* ??? We could also check for
3109 (ne (and (eq (...) (const_int 1))) (const_int 0))
3111 and related forms, but let's wait until we see them occurring. */
3113 if (x == 0)
3114 /* Look up ARG1 in the hash table and see if it has an equivalence
3115 that lets us see what is being compared. */
3116 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3117 if (p)
3119 p = p->first_same_value;
3121 /* If what we compare is already known to be constant, that is as
3122 good as it gets.
3123 We need to break the loop in this case, because otherwise we
3124 can have an infinite loop when looking at a reg that is known
3125 to be a constant which is the same as a comparison of a reg
3126 against zero which appears later in the insn stream, which in
3127 turn is constant and the same as the comparison of the first reg
3128 against zero... */
3129 if (p->is_const)
3130 break;
3133 for (; p; p = p->next_same_value)
3135 enum machine_mode inner_mode = GET_MODE (p->exp);
3136 #ifdef FLOAT_STORE_FLAG_VALUE
3137 REAL_VALUE_TYPE fsfv;
3138 #endif
3140 /* If the entry isn't valid, skip it. */
3141 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3142 continue;
3144 if (GET_CODE (p->exp) == COMPARE
3145 /* Another possibility is that this machine has a compare insn
3146 that includes the comparison code. In that case, ARG1 would
3147 be equivalent to a comparison operation that would set ARG1 to
3148 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3149 ORIG_CODE is the actual comparison being done; if it is an EQ,
3150 we must reverse ORIG_CODE. On machine with a negative value
3151 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3152 || ((code == NE
3153 || (code == LT
3154 && GET_MODE_CLASS (inner_mode) == MODE_INT
3155 && (GET_MODE_BITSIZE (inner_mode)
3156 <= HOST_BITS_PER_WIDE_INT)
3157 && (STORE_FLAG_VALUE
3158 & ((HOST_WIDE_INT) 1
3159 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3160 #ifdef FLOAT_STORE_FLAG_VALUE
3161 || (code == LT
3162 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3163 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3164 REAL_VALUE_NEGATIVE (fsfv)))
3165 #endif
3167 && COMPARISON_P (p->exp)))
3169 x = p->exp;
3170 break;
3172 else if ((code == EQ
3173 || (code == GE
3174 && GET_MODE_CLASS (inner_mode) == MODE_INT
3175 && (GET_MODE_BITSIZE (inner_mode)
3176 <= HOST_BITS_PER_WIDE_INT)
3177 && (STORE_FLAG_VALUE
3178 & ((HOST_WIDE_INT) 1
3179 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3180 #ifdef FLOAT_STORE_FLAG_VALUE
3181 || (code == GE
3182 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
3183 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3184 REAL_VALUE_NEGATIVE (fsfv)))
3185 #endif
3187 && COMPARISON_P (p->exp))
3189 reverse_code = 1;
3190 x = p->exp;
3191 break;
3194 /* If this non-trapping address, e.g. fp + constant, the
3195 equivalent is a better operand since it may let us predict
3196 the value of the comparison. */
3197 else if (!rtx_addr_can_trap_p (p->exp))
3199 arg1 = p->exp;
3200 continue;
3204 /* If we didn't find a useful equivalence for ARG1, we are done.
3205 Otherwise, set up for the next iteration. */
3206 if (x == 0)
3207 break;
3209 /* If we need to reverse the comparison, make sure that that is
3210 possible -- we can't necessarily infer the value of GE from LT
3211 with floating-point operands. */
3212 if (reverse_code)
3214 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3215 if (reversed == UNKNOWN)
3216 break;
3217 else
3218 code = reversed;
3220 else if (COMPARISON_P (x))
3221 code = GET_CODE (x);
3222 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3225 /* Return our results. Return the modes from before fold_rtx
3226 because fold_rtx might produce const_int, and then it's too late. */
3227 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3228 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3230 return code;
3233 /* Fold SUBREG. */
3235 static rtx
3236 fold_rtx_subreg (rtx x, rtx insn)
3238 enum machine_mode mode = GET_MODE (x);
3239 rtx folded_arg0;
3240 rtx const_arg0;
3241 rtx new;
3243 /* See if we previously assigned a constant value to this SUBREG. */
3244 if ((new = lookup_as_function (x, CONST_INT)) != 0
3245 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3246 return new;
3248 /* If this is a paradoxical SUBREG, we have no idea what value the
3249 extra bits would have. However, if the operand is equivalent to
3250 a SUBREG whose operand is the same as our mode, and all the modes
3251 are within a word, we can just use the inner operand because
3252 these SUBREGs just say how to treat the register.
3254 Similarly if we find an integer constant. */
3256 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3258 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3259 struct table_elt *elt;
3261 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3262 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3263 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3264 imode)) != 0)
3265 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3267 if (CONSTANT_P (elt->exp)
3268 && GET_MODE (elt->exp) == VOIDmode)
3269 return elt->exp;
3271 if (GET_CODE (elt->exp) == SUBREG
3272 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3273 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3274 return copy_rtx (SUBREG_REG (elt->exp));
3277 return x;
3280 /* Fold SUBREG_REG. If it changed, see if we can simplify the
3281 SUBREG. We might be able to if the SUBREG is extracting a single
3282 word in an integral mode or extracting the low part. */
3284 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3285 const_arg0 = equiv_constant (folded_arg0);
3286 if (const_arg0)
3287 folded_arg0 = const_arg0;
3289 if (folded_arg0 != SUBREG_REG (x))
3291 new = simplify_subreg (mode, folded_arg0,
3292 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3293 if (new)
3294 return new;
3297 if (REG_P (folded_arg0)
3298 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3300 struct table_elt *elt;
3302 elt = lookup (folded_arg0,
3303 HASH (folded_arg0, GET_MODE (folded_arg0)),
3304 GET_MODE (folded_arg0));
3306 if (elt)
3307 elt = elt->first_same_value;
3309 if (subreg_lowpart_p (x))
3310 /* If this is a narrowing SUBREG and our operand is a REG, see
3311 if we can find an equivalence for REG that is an arithmetic
3312 operation in a wider mode where both operands are
3313 paradoxical SUBREGs from objects of our result mode. In
3314 that case, we couldn-t report an equivalent value for that
3315 operation, since we don't know what the extra bits will be.
3316 But we can find an equivalence for this SUBREG by folding
3317 that operation in the narrow mode. This allows us to fold
3318 arithmetic in narrow modes when the machine only supports
3319 word-sized arithmetic.
3321 Also look for a case where we have a SUBREG whose operand
3322 is the same as our result. If both modes are smaller than
3323 a word, we are simply interpreting a register in different
3324 modes and we can use the inner value. */
3326 for (; elt; elt = elt->next_same_value)
3328 enum rtx_code eltcode = GET_CODE (elt->exp);
3330 /* Just check for unary and binary operations. */
3331 if (UNARY_P (elt->exp)
3332 && eltcode != SIGN_EXTEND
3333 && eltcode != ZERO_EXTEND
3334 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3335 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3336 && (GET_MODE_CLASS (mode)
3337 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3339 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3341 if (!REG_P (op0) && ! CONSTANT_P (op0))
3342 op0 = fold_rtx (op0, NULL_RTX);
3344 op0 = equiv_constant (op0);
3345 if (op0)
3346 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3347 op0, mode);
3349 else if (ARITHMETIC_P (elt->exp)
3350 && eltcode != DIV && eltcode != MOD
3351 && eltcode != UDIV && eltcode != UMOD
3352 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3353 && eltcode != ROTATE && eltcode != ROTATERT
3354 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3355 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3356 == mode))
3357 || CONSTANT_P (XEXP (elt->exp, 0)))
3358 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3359 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3360 == mode))
3361 || CONSTANT_P (XEXP (elt->exp, 1))))
3363 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3364 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3366 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3367 op0 = fold_rtx (op0, NULL_RTX);
3369 if (op0)
3370 op0 = equiv_constant (op0);
3372 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3373 op1 = fold_rtx (op1, NULL_RTX);
3375 if (op1)
3376 op1 = equiv_constant (op1);
3378 /* If we are looking for the low SImode part of
3379 (ashift:DI c (const_int 32)), it doesn't work to
3380 compute that in SImode, because a 32-bit shift in
3381 SImode is unpredictable. We know the value is
3382 0. */
3383 if (op0 && op1
3384 && GET_CODE (elt->exp) == ASHIFT
3385 && GET_CODE (op1) == CONST_INT
3386 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3388 if (INTVAL (op1)
3389 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3390 /* If the count fits in the inner mode's width,
3391 but exceeds the outer mode's width, the value
3392 will get truncated to 0 by the subreg. */
3393 new = CONST0_RTX (mode);
3394 else
3395 /* If the count exceeds even the inner mode's width,
3396 don't fold this expression. */
3397 new = 0;
3399 else if (op0 && op1)
3400 new = simplify_binary_operation (GET_CODE (elt->exp),
3401 mode, op0, op1);
3404 else if (GET_CODE (elt->exp) == SUBREG
3405 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3406 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3407 <= UNITS_PER_WORD)
3408 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3409 new = copy_rtx (SUBREG_REG (elt->exp));
3411 if (new)
3412 return new;
3414 else
3415 /* A SUBREG resulting from a zero extension may fold to zero
3416 if it extracts higher bits than the ZERO_EXTEND's source
3417 bits. FIXME: if combine tried to, er, combine these
3418 instructions, this transformation may be moved to
3419 simplify_subreg. */
3420 for (; elt; elt = elt->next_same_value)
3422 if (GET_CODE (elt->exp) == ZERO_EXTEND
3423 && subreg_lsb (x)
3424 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3425 return CONST0_RTX (mode);
3429 return x;
3432 /* Fold MEM. */
3434 static rtx
3435 fold_rtx_mem (rtx x, rtx insn)
3437 enum machine_mode mode = GET_MODE (x);
3438 rtx new;
3440 /* If we are not actually processing an insn, don't try to find the
3441 best address. Not only don't we care, but we could modify the
3442 MEM in an invalid way since we have no insn to validate
3443 against. */
3444 if (insn != 0)
3445 find_best_addr (insn, &XEXP (x, 0), mode);
3448 /* Even if we don't fold in the insn itself, we can safely do so
3449 here, in hopes of getting a constant. */
3450 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3451 rtx base = 0;
3452 HOST_WIDE_INT offset = 0;
3454 if (REG_P (addr)
3455 && REGNO_QTY_VALID_P (REGNO (addr)))
3457 int addr_q = REG_QTY (REGNO (addr));
3458 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3460 if (GET_MODE (addr) == addr_ent->mode
3461 && addr_ent->const_rtx != NULL_RTX)
3462 addr = addr_ent->const_rtx;
3465 /* If address is constant, split it into a base and integer
3466 offset. */
3467 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3468 base = addr;
3469 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3470 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3472 base = XEXP (XEXP (addr, 0), 0);
3473 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3475 else if (GET_CODE (addr) == LO_SUM
3476 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3477 base = XEXP (addr, 1);
3479 /* If this is a constant pool reference, we can fold it into its
3480 constant to allow better value tracking. */
3481 if (base && GET_CODE (base) == SYMBOL_REF
3482 && CONSTANT_POOL_ADDRESS_P (base))
3484 rtx constant = get_pool_constant (base);
3485 enum machine_mode const_mode = get_pool_mode (base);
3486 rtx new;
3488 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3490 constant_pool_entries_cost = COST (constant);
3491 constant_pool_entries_regcost = approx_reg_cost (constant);
3494 /* If we are loading the full constant, we have an
3495 equivalence. */
3496 if (offset == 0 && mode == const_mode)
3497 return constant;
3499 /* If this actually isn't a constant (weird!), we can't do
3500 anything. Otherwise, handle the two most common cases:
3501 extracting a word from a multi-word constant, and
3502 extracting the low-order bits. Other cases don't seem
3503 common enough to worry about. */
3504 if (! CONSTANT_P (constant))
3505 return x;
3507 if (GET_MODE_CLASS (mode) == MODE_INT
3508 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3509 && offset % UNITS_PER_WORD == 0
3510 && (new = operand_subword (constant,
3511 offset / UNITS_PER_WORD,
3512 0, const_mode)) != 0)
3513 return new;
3515 if (((BYTES_BIG_ENDIAN
3516 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3517 || (! BYTES_BIG_ENDIAN && offset == 0))
3518 && (new = gen_lowpart (mode, constant)) != 0)
3519 return new;
3522 /* If this is a reference to a label at a known position in a jump
3523 table, we also know its value. */
3524 if (base && GET_CODE (base) == LABEL_REF)
3526 rtx label = XEXP (base, 0);
3527 rtx table_insn = NEXT_INSN (label);
3529 if (table_insn && JUMP_P (table_insn)
3530 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3532 rtx table = PATTERN (table_insn);
3534 if (offset >= 0
3535 && (offset / GET_MODE_SIZE (GET_MODE (table))
3536 < XVECLEN (table, 0)))
3538 rtx label = XVECEXP
3539 (table, 0, offset / GET_MODE_SIZE (GET_MODE (table)));
3540 rtx set;
3542 /* If we have an insn that loads the label from the
3543 jumptable into a reg, we don't want to set the reg
3544 to the label, because this may cause a reference to
3545 the label to remain after the label is removed in
3546 some very obscure cases (PR middle-end/18628). */
3547 if (!insn)
3548 return label;
3550 set = single_set (insn);
3552 if (! set || SET_SRC (set) != x)
3553 return x;
3555 /* If it's a jump, it's safe to reference the label. */
3556 if (SET_DEST (set) == pc_rtx)
3557 return label;
3559 return x;
3562 if (table_insn && JUMP_P (table_insn)
3563 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3565 rtx table = PATTERN (table_insn);
3567 if (offset >= 0
3568 && (offset / GET_MODE_SIZE (GET_MODE (table))
3569 < XVECLEN (table, 1)))
3571 offset /= GET_MODE_SIZE (GET_MODE (table));
3572 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3573 XEXP (table, 0));
3575 if (GET_MODE (table) != Pmode)
3576 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3578 /* Indicate this is a constant. This isn't a valid
3579 form of CONST, but it will only be used to fold the
3580 next insns and then discarded, so it should be
3581 safe.
3583 Note this expression must be explicitly discarded,
3584 by cse_insn, else it may end up in a REG_EQUAL note
3585 and "escape" to cause problems elsewhere. */
3586 return gen_rtx_CONST (GET_MODE (new), new);
3591 return x;
3595 /* If X is a nontrivial arithmetic operation on an argument
3596 for which a constant value can be determined, return
3597 the result of operating on that value, as a constant.
3598 Otherwise, return X, possibly with one or more operands
3599 modified by recursive calls to this function.
3601 If X is a register whose contents are known, we do NOT
3602 return those contents here. equiv_constant is called to
3603 perform that task.
3605 INSN is the insn that we may be modifying. If it is 0, make a copy
3606 of X before modifying it. */
3608 static rtx
3609 fold_rtx (rtx x, rtx insn)
3611 enum rtx_code code;
3612 enum machine_mode mode;
3613 const char *fmt;
3614 int i;
3615 rtx new = 0;
3616 int copied = 0;
3617 int must_swap = 0;
3619 /* Folded equivalents of first two operands of X. */
3620 rtx folded_arg0;
3621 rtx folded_arg1;
3623 /* Constant equivalents of first three operands of X;
3624 0 when no such equivalent is known. */
3625 rtx const_arg0;
3626 rtx const_arg1;
3627 rtx const_arg2;
3629 /* The mode of the first operand of X. We need this for sign and zero
3630 extends. */
3631 enum machine_mode mode_arg0;
3633 if (x == 0)
3634 return x;
3636 mode = GET_MODE (x);
3637 code = GET_CODE (x);
3638 switch (code)
3640 case CONST:
3641 case CONST_INT:
3642 case CONST_DOUBLE:
3643 case CONST_VECTOR:
3644 case SYMBOL_REF:
3645 case LABEL_REF:
3646 case REG:
3647 case PC:
3648 /* No use simplifying an EXPR_LIST
3649 since they are used only for lists of args
3650 in a function call's REG_EQUAL note. */
3651 case EXPR_LIST:
3652 return x;
3654 #ifdef HAVE_cc0
3655 case CC0:
3656 return prev_insn_cc0;
3657 #endif
3659 case SUBREG:
3660 return fold_rtx_subreg (x, insn);
3662 case NOT:
3663 case NEG:
3664 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3665 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3666 new = lookup_as_function (XEXP (x, 0), code);
3667 if (new)
3668 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3669 break;
3671 case MEM:
3672 return fold_rtx_mem (x, insn);
3674 #ifdef NO_FUNCTION_CSE
3675 case CALL:
3676 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3677 return x;
3678 break;
3679 #endif
3681 case ASM_OPERANDS:
3682 if (insn)
3684 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3685 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3686 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3688 break;
3690 default:
3691 break;
3694 const_arg0 = 0;
3695 const_arg1 = 0;
3696 const_arg2 = 0;
3697 mode_arg0 = VOIDmode;
3699 /* Try folding our operands.
3700 Then see which ones have constant values known. */
3702 fmt = GET_RTX_FORMAT (code);
3703 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3704 if (fmt[i] == 'e')
3706 rtx arg = XEXP (x, i);
3707 rtx folded_arg = arg, const_arg = 0;
3708 enum machine_mode mode_arg = GET_MODE (arg);
3709 rtx cheap_arg, expensive_arg;
3710 rtx replacements[2];
3711 int j;
3712 int old_cost = COST_IN (XEXP (x, i), code);
3714 /* Most arguments are cheap, so handle them specially. */
3715 switch (GET_CODE (arg))
3717 case REG:
3718 /* This is the same as calling equiv_constant; it is duplicated
3719 here for speed. */
3720 if (REGNO_QTY_VALID_P (REGNO (arg)))
3722 int arg_q = REG_QTY (REGNO (arg));
3723 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3725 if (arg_ent->const_rtx != NULL_RTX
3726 && !REG_P (arg_ent->const_rtx)
3727 && GET_CODE (arg_ent->const_rtx) != PLUS)
3728 const_arg
3729 = gen_lowpart (GET_MODE (arg),
3730 arg_ent->const_rtx);
3732 break;
3734 case CONST:
3735 case CONST_INT:
3736 case SYMBOL_REF:
3737 case LABEL_REF:
3738 case CONST_DOUBLE:
3739 case CONST_VECTOR:
3740 const_arg = arg;
3741 break;
3743 #ifdef HAVE_cc0
3744 case CC0:
3745 folded_arg = prev_insn_cc0;
3746 mode_arg = prev_insn_cc0_mode;
3747 const_arg = equiv_constant (folded_arg);
3748 break;
3749 #endif
3751 default:
3752 folded_arg = fold_rtx (arg, insn);
3753 const_arg = equiv_constant (folded_arg);
3756 /* For the first three operands, see if the operand
3757 is constant or equivalent to a constant. */
3758 switch (i)
3760 case 0:
3761 folded_arg0 = folded_arg;
3762 const_arg0 = const_arg;
3763 mode_arg0 = mode_arg;
3764 break;
3765 case 1:
3766 folded_arg1 = folded_arg;
3767 const_arg1 = const_arg;
3768 break;
3769 case 2:
3770 const_arg2 = const_arg;
3771 break;
3774 /* Pick the least expensive of the folded argument and an
3775 equivalent constant argument. */
3776 if (const_arg == 0 || const_arg == folded_arg
3777 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3778 cheap_arg = folded_arg, expensive_arg = const_arg;
3779 else
3780 cheap_arg = const_arg, expensive_arg = folded_arg;
3782 /* Try to replace the operand with the cheapest of the two
3783 possibilities. If it doesn't work and this is either of the first
3784 two operands of a commutative operation, try swapping them.
3785 If THAT fails, try the more expensive, provided it is cheaper
3786 than what is already there. */
3788 if (cheap_arg == XEXP (x, i))
3789 continue;
3791 if (insn == 0 && ! copied)
3793 x = copy_rtx (x);
3794 copied = 1;
3797 /* Order the replacements from cheapest to most expensive. */
3798 replacements[0] = cheap_arg;
3799 replacements[1] = expensive_arg;
3801 for (j = 0; j < 2 && replacements[j]; j++)
3803 int new_cost = COST_IN (replacements[j], code);
3805 /* Stop if what existed before was cheaper. Prefer constants
3806 in the case of a tie. */
3807 if (new_cost > old_cost
3808 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3809 break;
3811 /* It's not safe to substitute the operand of a conversion
3812 operator with a constant, as the conversion's identity
3813 depends upon the mode of its operand. This optimization
3814 is handled by the call to simplify_unary_operation. */
3815 if (GET_RTX_CLASS (code) == RTX_UNARY
3816 && GET_MODE (replacements[j]) != mode_arg0
3817 && (code == ZERO_EXTEND
3818 || code == SIGN_EXTEND
3819 || code == TRUNCATE
3820 || code == FLOAT_TRUNCATE
3821 || code == FLOAT_EXTEND
3822 || code == FLOAT
3823 || code == FIX
3824 || code == UNSIGNED_FLOAT
3825 || code == UNSIGNED_FIX))
3826 continue;
3828 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3829 break;
3831 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3832 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3834 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3835 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3837 if (apply_change_group ())
3839 /* Swap them back to be invalid so that this loop can
3840 continue and flag them to be swapped back later. */
3841 rtx tem;
3843 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3844 XEXP (x, 1) = tem;
3845 must_swap = 1;
3846 break;
3852 else
3854 if (fmt[i] == 'E')
3855 /* Don't try to fold inside of a vector of expressions.
3856 Doing nothing is harmless. */
3860 /* If a commutative operation, place a constant integer as the second
3861 operand unless the first operand is also a constant integer. Otherwise,
3862 place any constant second unless the first operand is also a constant. */
3864 if (COMMUTATIVE_P (x))
3866 if (must_swap
3867 || swap_commutative_operands_p (const_arg0 ? const_arg0
3868 : XEXP (x, 0),
3869 const_arg1 ? const_arg1
3870 : XEXP (x, 1)))
3872 rtx tem = XEXP (x, 0);
3874 if (insn == 0 && ! copied)
3876 x = copy_rtx (x);
3877 copied = 1;
3880 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3881 validate_change (insn, &XEXP (x, 1), tem, 1);
3882 if (apply_change_group ())
3884 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3885 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3890 /* If X is an arithmetic operation, see if we can simplify it. */
3892 switch (GET_RTX_CLASS (code))
3894 case RTX_UNARY:
3896 int is_const = 0;
3898 /* We can't simplify extension ops unless we know the
3899 original mode. */
3900 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3901 && mode_arg0 == VOIDmode)
3902 break;
3904 /* If we had a CONST, strip it off and put it back later if we
3905 fold. */
3906 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3907 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3909 new = simplify_unary_operation (code, mode,
3910 const_arg0 ? const_arg0 : folded_arg0,
3911 mode_arg0);
3912 /* NEG of PLUS could be converted into MINUS, but that causes
3913 expressions of the form
3914 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3915 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3916 FIXME: those ports should be fixed. */
3917 if (new != 0 && is_const
3918 && GET_CODE (new) == PLUS
3919 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3920 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3921 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3922 new = gen_rtx_CONST (mode, new);
3924 break;
3926 case RTX_COMPARE:
3927 case RTX_COMM_COMPARE:
3928 /* See what items are actually being compared and set FOLDED_ARG[01]
3929 to those values and CODE to the actual comparison code. If any are
3930 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3931 do anything if both operands are already known to be constant. */
3933 /* ??? Vector mode comparisons are not supported yet. */
3934 if (VECTOR_MODE_P (mode))
3935 break;
3937 if (const_arg0 == 0 || const_arg1 == 0)
3939 struct table_elt *p0, *p1;
3940 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3941 enum machine_mode mode_arg1;
3943 #ifdef FLOAT_STORE_FLAG_VALUE
3944 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
3946 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3947 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3948 false_rtx = CONST0_RTX (mode);
3950 #endif
3952 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3953 &mode_arg0, &mode_arg1);
3955 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3956 what kinds of things are being compared, so we can't do
3957 anything with this comparison. */
3959 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3960 break;
3962 const_arg0 = equiv_constant (folded_arg0);
3963 const_arg1 = equiv_constant (folded_arg1);
3965 /* If we do not now have two constants being compared, see
3966 if we can nevertheless deduce some things about the
3967 comparison. */
3968 if (const_arg0 == 0 || const_arg1 == 0)
3970 /* Some addresses are known to be nonzero. We don't know
3971 their sign, but equality comparisons are known. */
3972 if (const_arg1 == const0_rtx
3973 && nonzero_address_p (folded_arg0))
3975 if (code == EQ)
3976 return false_rtx;
3977 else if (code == NE)
3978 return true_rtx;
3981 /* See if the two operands are the same. */
3983 if (folded_arg0 == folded_arg1
3984 || (REG_P (folded_arg0)
3985 && REG_P (folded_arg1)
3986 && (REG_QTY (REGNO (folded_arg0))
3987 == REG_QTY (REGNO (folded_arg1))))
3988 || ((p0 = lookup (folded_arg0,
3989 SAFE_HASH (folded_arg0, mode_arg0),
3990 mode_arg0))
3991 && (p1 = lookup (folded_arg1,
3992 SAFE_HASH (folded_arg1, mode_arg0),
3993 mode_arg0))
3994 && p0->first_same_value == p1->first_same_value))
3996 /* Sadly two equal NaNs are not equivalent. */
3997 if (!HONOR_NANS (mode_arg0))
3998 return ((code == EQ || code == LE || code == GE
3999 || code == LEU || code == GEU || code == UNEQ
4000 || code == UNLE || code == UNGE
4001 || code == ORDERED)
4002 ? true_rtx : false_rtx);
4003 /* Take care for the FP compares we can resolve. */
4004 if (code == UNEQ || code == UNLE || code == UNGE)
4005 return true_rtx;
4006 if (code == LTGT || code == LT || code == GT)
4007 return false_rtx;
4010 /* If FOLDED_ARG0 is a register, see if the comparison we are
4011 doing now is either the same as we did before or the reverse
4012 (we only check the reverse if not floating-point). */
4013 else if (REG_P (folded_arg0))
4015 int qty = REG_QTY (REGNO (folded_arg0));
4017 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4019 struct qty_table_elem *ent = &qty_table[qty];
4021 if ((comparison_dominates_p (ent->comparison_code, code)
4022 || (! FLOAT_MODE_P (mode_arg0)
4023 && comparison_dominates_p (ent->comparison_code,
4024 reverse_condition (code))))
4025 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4026 || (const_arg1
4027 && rtx_equal_p (ent->comparison_const,
4028 const_arg1))
4029 || (REG_P (folded_arg1)
4030 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4031 return (comparison_dominates_p (ent->comparison_code, code)
4032 ? true_rtx : false_rtx);
4038 /* If we are comparing against zero, see if the first operand is
4039 equivalent to an IOR with a constant. If so, we may be able to
4040 determine the result of this comparison. */
4042 if (const_arg1 == const0_rtx)
4044 rtx y = lookup_as_function (folded_arg0, IOR);
4045 rtx inner_const;
4047 if (y != 0
4048 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4049 && GET_CODE (inner_const) == CONST_INT
4050 && INTVAL (inner_const) != 0)
4052 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4053 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4054 && (INTVAL (inner_const)
4055 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4056 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4058 #ifdef FLOAT_STORE_FLAG_VALUE
4059 if (GET_MODE_CLASS (mode) == MODE_FLOAT)
4061 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4062 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4063 false_rtx = CONST0_RTX (mode);
4065 #endif
4067 switch (code)
4069 case EQ:
4070 return false_rtx;
4071 case NE:
4072 return true_rtx;
4073 case LT: case LE:
4074 if (has_sign)
4075 return true_rtx;
4076 break;
4077 case GT: case GE:
4078 if (has_sign)
4079 return false_rtx;
4080 break;
4081 default:
4082 break;
4088 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4089 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4090 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4092 break;
4094 case RTX_BIN_ARITH:
4095 case RTX_COMM_ARITH:
4096 switch (code)
4098 case PLUS:
4099 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4100 with that LABEL_REF as its second operand. If so, the result is
4101 the first operand of that MINUS. This handles switches with an
4102 ADDR_DIFF_VEC table. */
4103 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4105 rtx y
4106 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4107 : lookup_as_function (folded_arg0, MINUS);
4109 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4110 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4111 return XEXP (y, 0);
4113 /* Now try for a CONST of a MINUS like the above. */
4114 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4115 : lookup_as_function (folded_arg0, CONST))) != 0
4116 && GET_CODE (XEXP (y, 0)) == MINUS
4117 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4118 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4119 return XEXP (XEXP (y, 0), 0);
4122 /* Likewise if the operands are in the other order. */
4123 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4125 rtx y
4126 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4127 : lookup_as_function (folded_arg1, MINUS);
4129 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4130 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4131 return XEXP (y, 0);
4133 /* Now try for a CONST of a MINUS like the above. */
4134 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4135 : lookup_as_function (folded_arg1, CONST))) != 0
4136 && GET_CODE (XEXP (y, 0)) == MINUS
4137 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4138 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4139 return XEXP (XEXP (y, 0), 0);
4142 /* If second operand is a register equivalent to a negative
4143 CONST_INT, see if we can find a register equivalent to the
4144 positive constant. Make a MINUS if so. Don't do this for
4145 a non-negative constant since we might then alternate between
4146 choosing positive and negative constants. Having the positive
4147 constant previously-used is the more common case. Be sure
4148 the resulting constant is non-negative; if const_arg1 were
4149 the smallest negative number this would overflow: depending
4150 on the mode, this would either just be the same value (and
4151 hence not save anything) or be incorrect. */
4152 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4153 && INTVAL (const_arg1) < 0
4154 /* This used to test
4156 -INTVAL (const_arg1) >= 0
4158 But The Sun V5.0 compilers mis-compiled that test. So
4159 instead we test for the problematic value in a more direct
4160 manner and hope the Sun compilers get it correct. */
4161 && INTVAL (const_arg1) !=
4162 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4163 && REG_P (folded_arg1))
4165 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4166 struct table_elt *p
4167 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4169 if (p)
4170 for (p = p->first_same_value; p; p = p->next_same_value)
4171 if (REG_P (p->exp))
4172 return simplify_gen_binary (MINUS, mode, folded_arg0,
4173 canon_reg (p->exp, NULL_RTX));
4175 goto from_plus;
4177 case MINUS:
4178 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4179 If so, produce (PLUS Z C2-C). */
4180 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4182 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4183 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4184 return fold_rtx (plus_constant (copy_rtx (y),
4185 -INTVAL (const_arg1)),
4186 NULL_RTX);
4189 /* Fall through. */
4191 from_plus:
4192 case SMIN: case SMAX: case UMIN: case UMAX:
4193 case IOR: case AND: case XOR:
4194 case MULT:
4195 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4196 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4197 is known to be of similar form, we may be able to replace the
4198 operation with a combined operation. This may eliminate the
4199 intermediate operation if every use is simplified in this way.
4200 Note that the similar optimization done by combine.c only works
4201 if the intermediate operation's result has only one reference. */
4203 if (REG_P (folded_arg0)
4204 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4206 int is_shift
4207 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4208 rtx y = lookup_as_function (folded_arg0, code);
4209 rtx inner_const;
4210 enum rtx_code associate_code;
4211 rtx new_const;
4213 if (y == 0
4214 || 0 == (inner_const
4215 = equiv_constant (fold_rtx (XEXP (y, 1), 0)))
4216 || GET_CODE (inner_const) != CONST_INT
4217 /* If we have compiled a statement like
4218 "if (x == (x & mask1))", and now are looking at
4219 "x & mask2", we will have a case where the first operand
4220 of Y is the same as our first operand. Unless we detect
4221 this case, an infinite loop will result. */
4222 || XEXP (y, 0) == folded_arg0)
4223 break;
4225 /* Don't associate these operations if they are a PLUS with the
4226 same constant and it is a power of two. These might be doable
4227 with a pre- or post-increment. Similarly for two subtracts of
4228 identical powers of two with post decrement. */
4230 if (code == PLUS && const_arg1 == inner_const
4231 && ((HAVE_PRE_INCREMENT
4232 && exact_log2 (INTVAL (const_arg1)) >= 0)
4233 || (HAVE_POST_INCREMENT
4234 && exact_log2 (INTVAL (const_arg1)) >= 0)
4235 || (HAVE_PRE_DECREMENT
4236 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4237 || (HAVE_POST_DECREMENT
4238 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4239 break;
4241 /* Compute the code used to compose the constants. For example,
4242 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4244 associate_code = (is_shift || code == MINUS ? PLUS : code);
4246 new_const = simplify_binary_operation (associate_code, mode,
4247 const_arg1, inner_const);
4249 if (new_const == 0)
4250 break;
4252 /* If we are associating shift operations, don't let this
4253 produce a shift of the size of the object or larger.
4254 This could occur when we follow a sign-extend by a right
4255 shift on a machine that does a sign-extend as a pair
4256 of shifts. */
4258 if (is_shift && GET_CODE (new_const) == CONST_INT
4259 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4261 /* As an exception, we can turn an ASHIFTRT of this
4262 form into a shift of the number of bits - 1. */
4263 if (code == ASHIFTRT)
4264 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4265 else
4266 break;
4269 y = copy_rtx (XEXP (y, 0));
4271 /* If Y contains our first operand (the most common way this
4272 can happen is if Y is a MEM), we would do into an infinite
4273 loop if we tried to fold it. So don't in that case. */
4275 if (! reg_mentioned_p (folded_arg0, y))
4276 y = fold_rtx (y, insn);
4278 return simplify_gen_binary (code, mode, y, new_const);
4280 break;
4282 case DIV: case UDIV:
4283 /* ??? The associative optimization performed immediately above is
4284 also possible for DIV and UDIV using associate_code of MULT.
4285 However, we would need extra code to verify that the
4286 multiplication does not overflow, that is, there is no overflow
4287 in the calculation of new_const. */
4288 break;
4290 default:
4291 break;
4294 new = simplify_binary_operation (code, mode,
4295 const_arg0 ? const_arg0 : folded_arg0,
4296 const_arg1 ? const_arg1 : folded_arg1);
4297 break;
4299 case RTX_OBJ:
4300 /* (lo_sum (high X) X) is simply X. */
4301 if (code == LO_SUM && const_arg0 != 0
4302 && GET_CODE (const_arg0) == HIGH
4303 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4304 return const_arg1;
4305 break;
4307 case RTX_TERNARY:
4308 case RTX_BITFIELD_OPS:
4309 new = simplify_ternary_operation (code, mode, mode_arg0,
4310 const_arg0 ? const_arg0 : folded_arg0,
4311 const_arg1 ? const_arg1 : folded_arg1,
4312 const_arg2 ? const_arg2 : XEXP (x, 2));
4313 break;
4315 default:
4316 break;
4319 return new ? new : x;
4322 /* Return a constant value currently equivalent to X.
4323 Return 0 if we don't know one. */
4325 static rtx
4326 equiv_constant (rtx x)
4328 if (REG_P (x)
4329 && REGNO_QTY_VALID_P (REGNO (x)))
4331 int x_q = REG_QTY (REGNO (x));
4332 struct qty_table_elem *x_ent = &qty_table[x_q];
4334 if (x_ent->const_rtx)
4335 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4338 if (x == 0 || CONSTANT_P (x))
4339 return x;
4341 /* If X is a MEM, try to fold it outside the context of any insn to see if
4342 it might be equivalent to a constant. That handles the case where it
4343 is a constant-pool reference. Then try to look it up in the hash table
4344 in case it is something whose value we have seen before. */
4346 if (MEM_P (x))
4348 struct table_elt *elt;
4350 x = fold_rtx (x, NULL_RTX);
4351 if (CONSTANT_P (x))
4352 return x;
4354 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4355 if (elt == 0)
4356 return 0;
4358 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4359 if (elt->is_const && CONSTANT_P (elt->exp))
4360 return elt->exp;
4363 return 0;
4366 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4367 branch. It will be zero if not.
4369 In certain cases, this can cause us to add an equivalence. For example,
4370 if we are following the taken case of
4371 if (i == 2)
4372 we can add the fact that `i' and '2' are now equivalent.
4374 In any case, we can record that this comparison was passed. If the same
4375 comparison is seen later, we will know its value. */
4377 static void
4378 record_jump_equiv (rtx insn, int taken)
4380 int cond_known_true;
4381 rtx op0, op1;
4382 rtx set;
4383 enum machine_mode mode, mode0, mode1;
4384 int reversed_nonequality = 0;
4385 enum rtx_code code;
4387 /* Ensure this is the right kind of insn. */
4388 if (! any_condjump_p (insn))
4389 return;
4390 set = pc_set (insn);
4392 /* See if this jump condition is known true or false. */
4393 if (taken)
4394 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4395 else
4396 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4398 /* Get the type of comparison being done and the operands being compared.
4399 If we had to reverse a non-equality condition, record that fact so we
4400 know that it isn't valid for floating-point. */
4401 code = GET_CODE (XEXP (SET_SRC (set), 0));
4402 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4403 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4405 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4406 if (! cond_known_true)
4408 code = reversed_comparison_code_parts (code, op0, op1, insn);
4410 /* Don't remember if we can't find the inverse. */
4411 if (code == UNKNOWN)
4412 return;
4415 /* The mode is the mode of the non-constant. */
4416 mode = mode0;
4417 if (mode1 != VOIDmode)
4418 mode = mode1;
4420 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4423 /* Yet another form of subreg creation. In this case, we want something in
4424 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4426 static rtx
4427 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4429 enum machine_mode op_mode = GET_MODE (op);
4430 if (op_mode == mode || op_mode == VOIDmode)
4431 return op;
4432 return lowpart_subreg (mode, op, op_mode);
4435 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4436 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4437 Make any useful entries we can with that information. Called from
4438 above function and called recursively. */
4440 static void
4441 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4442 rtx op1, int reversed_nonequality)
4444 unsigned op0_hash, op1_hash;
4445 int op0_in_memory, op1_in_memory;
4446 struct table_elt *op0_elt, *op1_elt;
4448 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4449 we know that they are also equal in the smaller mode (this is also
4450 true for all smaller modes whether or not there is a SUBREG, but
4451 is not worth testing for with no SUBREG). */
4453 /* Note that GET_MODE (op0) may not equal MODE. */
4454 if (code == EQ && GET_CODE (op0) == SUBREG
4455 && (GET_MODE_SIZE (GET_MODE (op0))
4456 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4458 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4459 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4460 if (tem)
4461 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4462 reversed_nonequality);
4465 if (code == EQ && GET_CODE (op1) == SUBREG
4466 && (GET_MODE_SIZE (GET_MODE (op1))
4467 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4469 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4470 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4471 if (tem)
4472 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4473 reversed_nonequality);
4476 /* Similarly, if this is an NE comparison, and either is a SUBREG
4477 making a smaller mode, we know the whole thing is also NE. */
4479 /* Note that GET_MODE (op0) may not equal MODE;
4480 if we test MODE instead, we can get an infinite recursion
4481 alternating between two modes each wider than MODE. */
4483 if (code == NE && GET_CODE (op0) == SUBREG
4484 && subreg_lowpart_p (op0)
4485 && (GET_MODE_SIZE (GET_MODE (op0))
4486 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4488 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4489 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4490 if (tem)
4491 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4492 reversed_nonequality);
4495 if (code == NE && GET_CODE (op1) == SUBREG
4496 && subreg_lowpart_p (op1)
4497 && (GET_MODE_SIZE (GET_MODE (op1))
4498 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4500 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4501 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4502 if (tem)
4503 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4504 reversed_nonequality);
4507 /* Hash both operands. */
4509 do_not_record = 0;
4510 hash_arg_in_memory = 0;
4511 op0_hash = HASH (op0, mode);
4512 op0_in_memory = hash_arg_in_memory;
4514 if (do_not_record)
4515 return;
4517 do_not_record = 0;
4518 hash_arg_in_memory = 0;
4519 op1_hash = HASH (op1, mode);
4520 op1_in_memory = hash_arg_in_memory;
4522 if (do_not_record)
4523 return;
4525 /* Look up both operands. */
4526 op0_elt = lookup (op0, op0_hash, mode);
4527 op1_elt = lookup (op1, op1_hash, mode);
4529 /* If both operands are already equivalent or if they are not in the
4530 table but are identical, do nothing. */
4531 if ((op0_elt != 0 && op1_elt != 0
4532 && op0_elt->first_same_value == op1_elt->first_same_value)
4533 || op0 == op1 || rtx_equal_p (op0, op1))
4534 return;
4536 /* If we aren't setting two things equal all we can do is save this
4537 comparison. Similarly if this is floating-point. In the latter
4538 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4539 If we record the equality, we might inadvertently delete code
4540 whose intent was to change -0 to +0. */
4542 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4544 struct qty_table_elem *ent;
4545 int qty;
4547 /* If we reversed a floating-point comparison, if OP0 is not a
4548 register, or if OP1 is neither a register or constant, we can't
4549 do anything. */
4551 if (!REG_P (op1))
4552 op1 = equiv_constant (op1);
4554 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4555 || !REG_P (op0) || op1 == 0)
4556 return;
4558 /* Put OP0 in the hash table if it isn't already. This gives it a
4559 new quantity number. */
4560 if (op0_elt == 0)
4562 if (insert_regs (op0, NULL, 0))
4564 rehash_using_reg (op0);
4565 op0_hash = HASH (op0, mode);
4567 /* If OP0 is contained in OP1, this changes its hash code
4568 as well. Faster to rehash than to check, except
4569 for the simple case of a constant. */
4570 if (! CONSTANT_P (op1))
4571 op1_hash = HASH (op1,mode);
4574 op0_elt = insert (op0, NULL, op0_hash, mode);
4575 op0_elt->in_memory = op0_in_memory;
4578 qty = REG_QTY (REGNO (op0));
4579 ent = &qty_table[qty];
4581 ent->comparison_code = code;
4582 if (REG_P (op1))
4584 /* Look it up again--in case op0 and op1 are the same. */
4585 op1_elt = lookup (op1, op1_hash, mode);
4587 /* Put OP1 in the hash table so it gets a new quantity number. */
4588 if (op1_elt == 0)
4590 if (insert_regs (op1, NULL, 0))
4592 rehash_using_reg (op1);
4593 op1_hash = HASH (op1, mode);
4596 op1_elt = insert (op1, NULL, op1_hash, mode);
4597 op1_elt->in_memory = op1_in_memory;
4600 ent->comparison_const = NULL_RTX;
4601 ent->comparison_qty = REG_QTY (REGNO (op1));
4603 else
4605 ent->comparison_const = op1;
4606 ent->comparison_qty = -1;
4609 return;
4612 /* If either side is still missing an equivalence, make it now,
4613 then merge the equivalences. */
4615 if (op0_elt == 0)
4617 if (insert_regs (op0, NULL, 0))
4619 rehash_using_reg (op0);
4620 op0_hash = HASH (op0, mode);
4623 op0_elt = insert (op0, NULL, op0_hash, mode);
4624 op0_elt->in_memory = op0_in_memory;
4627 if (op1_elt == 0)
4629 if (insert_regs (op1, NULL, 0))
4631 rehash_using_reg (op1);
4632 op1_hash = HASH (op1, mode);
4635 op1_elt = insert (op1, NULL, op1_hash, mode);
4636 op1_elt->in_memory = op1_in_memory;
4639 merge_equiv_classes (op0_elt, op1_elt);
4642 /* CSE processing for one instruction.
4643 First simplify sources and addresses of all assignments
4644 in the instruction, using previously-computed equivalents values.
4645 Then install the new sources and destinations in the table
4646 of available values.
4648 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4649 the insn. It means that INSN is inside libcall block. In this
4650 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4652 /* Data on one SET contained in the instruction. */
4654 struct set
4656 /* The SET rtx itself. */
4657 rtx rtl;
4658 /* The SET_SRC of the rtx (the original value, if it is changing). */
4659 rtx src;
4660 /* The hash-table element for the SET_SRC of the SET. */
4661 struct table_elt *src_elt;
4662 /* Hash value for the SET_SRC. */
4663 unsigned src_hash;
4664 /* Hash value for the SET_DEST. */
4665 unsigned dest_hash;
4666 /* The SET_DEST, with SUBREG, etc., stripped. */
4667 rtx inner_dest;
4668 /* Nonzero if the SET_SRC is in memory. */
4669 char src_in_memory;
4670 /* Nonzero if the SET_SRC contains something
4671 whose value cannot be predicted and understood. */
4672 char src_volatile;
4673 /* Original machine mode, in case it becomes a CONST_INT.
4674 The size of this field should match the size of the mode
4675 field of struct rtx_def (see rtl.h). */
4676 ENUM_BITFIELD(machine_mode) mode : 8;
4677 /* A constant equivalent for SET_SRC, if any. */
4678 rtx src_const;
4679 /* Original SET_SRC value used for libcall notes. */
4680 rtx orig_src;
4681 /* Hash value of constant equivalent for SET_SRC. */
4682 unsigned src_const_hash;
4683 /* Table entry for constant equivalent for SET_SRC, if any. */
4684 struct table_elt *src_const_elt;
4687 static void
4688 cse_insn (rtx insn, rtx libcall_insn)
4690 rtx x = PATTERN (insn);
4691 int i;
4692 rtx tem;
4693 int n_sets = 0;
4695 #ifdef HAVE_cc0
4696 /* Records what this insn does to set CC0. */
4697 rtx this_insn_cc0 = 0;
4698 enum machine_mode this_insn_cc0_mode = VOIDmode;
4699 #endif
4701 rtx src_eqv = 0;
4702 struct table_elt *src_eqv_elt = 0;
4703 int src_eqv_volatile = 0;
4704 int src_eqv_in_memory = 0;
4705 unsigned src_eqv_hash = 0;
4707 struct set *sets = (struct set *) 0;
4709 this_insn = insn;
4711 /* Find all the SETs and CLOBBERs in this instruction.
4712 Record all the SETs in the array `set' and count them.
4713 Also determine whether there is a CLOBBER that invalidates
4714 all memory references, or all references at varying addresses. */
4716 if (CALL_P (insn))
4718 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4720 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4721 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4722 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4726 if (GET_CODE (x) == SET)
4728 sets = alloca (sizeof (struct set));
4729 sets[0].rtl = x;
4731 /* Ignore SETs that are unconditional jumps.
4732 They never need cse processing, so this does not hurt.
4733 The reason is not efficiency but rather
4734 so that we can test at the end for instructions
4735 that have been simplified to unconditional jumps
4736 and not be misled by unchanged instructions
4737 that were unconditional jumps to begin with. */
4738 if (SET_DEST (x) == pc_rtx
4739 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4742 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4743 The hard function value register is used only once, to copy to
4744 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4745 Ensure we invalidate the destination register. On the 80386 no
4746 other code would invalidate it since it is a fixed_reg.
4747 We need not check the return of apply_change_group; see canon_reg. */
4749 else if (GET_CODE (SET_SRC (x)) == CALL)
4751 canon_reg (SET_SRC (x), insn);
4752 apply_change_group ();
4753 fold_rtx (SET_SRC (x), insn);
4754 invalidate (SET_DEST (x), VOIDmode);
4756 else
4757 n_sets = 1;
4759 else if (GET_CODE (x) == PARALLEL)
4761 int lim = XVECLEN (x, 0);
4763 sets = alloca (lim * sizeof (struct set));
4765 /* Find all regs explicitly clobbered in this insn,
4766 and ensure they are not replaced with any other regs
4767 elsewhere in this insn.
4768 When a reg that is clobbered is also used for input,
4769 we should presume that that is for a reason,
4770 and we should not substitute some other register
4771 which is not supposed to be clobbered.
4772 Therefore, this loop cannot be merged into the one below
4773 because a CALL may precede a CLOBBER and refer to the
4774 value clobbered. We must not let a canonicalization do
4775 anything in that case. */
4776 for (i = 0; i < lim; i++)
4778 rtx y = XVECEXP (x, 0, i);
4779 if (GET_CODE (y) == CLOBBER)
4781 rtx clobbered = XEXP (y, 0);
4783 if (REG_P (clobbered)
4784 || GET_CODE (clobbered) == SUBREG)
4785 invalidate (clobbered, VOIDmode);
4786 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4787 || GET_CODE (clobbered) == ZERO_EXTRACT)
4788 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4792 for (i = 0; i < lim; i++)
4794 rtx y = XVECEXP (x, 0, i);
4795 if (GET_CODE (y) == SET)
4797 /* As above, we ignore unconditional jumps and call-insns and
4798 ignore the result of apply_change_group. */
4799 if (GET_CODE (SET_SRC (y)) == CALL)
4801 canon_reg (SET_SRC (y), insn);
4802 apply_change_group ();
4803 fold_rtx (SET_SRC (y), insn);
4804 invalidate (SET_DEST (y), VOIDmode);
4806 else if (SET_DEST (y) == pc_rtx
4807 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4809 else
4810 sets[n_sets++].rtl = y;
4812 else if (GET_CODE (y) == CLOBBER)
4814 /* If we clobber memory, canon the address.
4815 This does nothing when a register is clobbered
4816 because we have already invalidated the reg. */
4817 if (MEM_P (XEXP (y, 0)))
4818 canon_reg (XEXP (y, 0), NULL_RTX);
4820 else if (GET_CODE (y) == USE
4821 && ! (REG_P (XEXP (y, 0))
4822 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4823 canon_reg (y, NULL_RTX);
4824 else if (GET_CODE (y) == CALL)
4826 /* The result of apply_change_group can be ignored; see
4827 canon_reg. */
4828 canon_reg (y, insn);
4829 apply_change_group ();
4830 fold_rtx (y, insn);
4834 else if (GET_CODE (x) == CLOBBER)
4836 if (MEM_P (XEXP (x, 0)))
4837 canon_reg (XEXP (x, 0), NULL_RTX);
4840 /* Canonicalize a USE of a pseudo register or memory location. */
4841 else if (GET_CODE (x) == USE
4842 && ! (REG_P (XEXP (x, 0))
4843 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4844 canon_reg (XEXP (x, 0), NULL_RTX);
4845 else if (GET_CODE (x) == CALL)
4847 /* The result of apply_change_group can be ignored; see canon_reg. */
4848 canon_reg (x, insn);
4849 apply_change_group ();
4850 fold_rtx (x, insn);
4853 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4854 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4855 is handled specially for this case, and if it isn't set, then there will
4856 be no equivalence for the destination. */
4857 if (n_sets == 1 && REG_NOTES (insn) != 0
4858 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4859 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4860 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4862 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4863 XEXP (tem, 0) = src_eqv;
4866 /* Canonicalize sources and addresses of destinations.
4867 We do this in a separate pass to avoid problems when a MATCH_DUP is
4868 present in the insn pattern. In that case, we want to ensure that
4869 we don't break the duplicate nature of the pattern. So we will replace
4870 both operands at the same time. Otherwise, we would fail to find an
4871 equivalent substitution in the loop calling validate_change below.
4873 We used to suppress canonicalization of DEST if it appears in SRC,
4874 but we don't do this any more. */
4876 for (i = 0; i < n_sets; i++)
4878 rtx dest = SET_DEST (sets[i].rtl);
4879 rtx src = SET_SRC (sets[i].rtl);
4880 rtx new = canon_reg (src, insn);
4881 int insn_code;
4883 sets[i].orig_src = src;
4884 if ((REG_P (new) && REG_P (src)
4885 && ((REGNO (new) < FIRST_PSEUDO_REGISTER)
4886 != (REGNO (src) < FIRST_PSEUDO_REGISTER)))
4887 || (insn_code = recog_memoized (insn)) < 0
4888 || insn_data[insn_code].n_dups > 0)
4889 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4890 else
4891 SET_SRC (sets[i].rtl) = new;
4893 if (GET_CODE (dest) == ZERO_EXTRACT)
4895 validate_change (insn, &XEXP (dest, 1),
4896 canon_reg (XEXP (dest, 1), insn), 1);
4897 validate_change (insn, &XEXP (dest, 2),
4898 canon_reg (XEXP (dest, 2), insn), 1);
4901 while (GET_CODE (dest) == SUBREG
4902 || GET_CODE (dest) == ZERO_EXTRACT
4903 || GET_CODE (dest) == STRICT_LOW_PART)
4904 dest = XEXP (dest, 0);
4906 if (MEM_P (dest))
4907 canon_reg (dest, insn);
4910 /* Now that we have done all the replacements, we can apply the change
4911 group and see if they all work. Note that this will cause some
4912 canonicalizations that would have worked individually not to be applied
4913 because some other canonicalization didn't work, but this should not
4914 occur often.
4916 The result of apply_change_group can be ignored; see canon_reg. */
4918 apply_change_group ();
4920 /* Set sets[i].src_elt to the class each source belongs to.
4921 Detect assignments from or to volatile things
4922 and set set[i] to zero so they will be ignored
4923 in the rest of this function.
4925 Nothing in this loop changes the hash table or the register chains. */
4927 for (i = 0; i < n_sets; i++)
4929 rtx src, dest;
4930 rtx src_folded;
4931 struct table_elt *elt = 0, *p;
4932 enum machine_mode mode;
4933 rtx src_eqv_here;
4934 rtx src_const = 0;
4935 rtx src_related = 0;
4936 struct table_elt *src_const_elt = 0;
4937 int src_cost = MAX_COST;
4938 int src_eqv_cost = MAX_COST;
4939 int src_folded_cost = MAX_COST;
4940 int src_related_cost = MAX_COST;
4941 int src_elt_cost = MAX_COST;
4942 int src_regcost = MAX_COST;
4943 int src_eqv_regcost = MAX_COST;
4944 int src_folded_regcost = MAX_COST;
4945 int src_related_regcost = MAX_COST;
4946 int src_elt_regcost = MAX_COST;
4947 /* Set nonzero if we need to call force_const_mem on with the
4948 contents of src_folded before using it. */
4949 int src_folded_force_flag = 0;
4951 dest = SET_DEST (sets[i].rtl);
4952 src = SET_SRC (sets[i].rtl);
4954 /* If SRC is a constant that has no machine mode,
4955 hash it with the destination's machine mode.
4956 This way we can keep different modes separate. */
4958 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4959 sets[i].mode = mode;
4961 if (src_eqv)
4963 enum machine_mode eqvmode = mode;
4964 if (GET_CODE (dest) == STRICT_LOW_PART)
4965 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4966 do_not_record = 0;
4967 hash_arg_in_memory = 0;
4968 src_eqv_hash = HASH (src_eqv, eqvmode);
4970 /* Find the equivalence class for the equivalent expression. */
4972 if (!do_not_record)
4973 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4975 src_eqv_volatile = do_not_record;
4976 src_eqv_in_memory = hash_arg_in_memory;
4979 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4980 value of the INNER register, not the destination. So it is not
4981 a valid substitution for the source. But save it for later. */
4982 if (GET_CODE (dest) == STRICT_LOW_PART)
4983 src_eqv_here = 0;
4984 else
4985 src_eqv_here = src_eqv;
4987 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4988 simplified result, which may not necessarily be valid. */
4989 src_folded = fold_rtx (src, insn);
4991 #if 0
4992 /* ??? This caused bad code to be generated for the m68k port with -O2.
4993 Suppose src is (CONST_INT -1), and that after truncation src_folded
4994 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4995 At the end we will add src and src_const to the same equivalence
4996 class. We now have 3 and -1 on the same equivalence class. This
4997 causes later instructions to be mis-optimized. */
4998 /* If storing a constant in a bitfield, pre-truncate the constant
4999 so we will be able to record it later. */
5000 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5002 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5004 if (GET_CODE (src) == CONST_INT
5005 && GET_CODE (width) == CONST_INT
5006 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5007 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5008 src_folded
5009 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5010 << INTVAL (width)) - 1));
5012 #endif
5014 /* Compute SRC's hash code, and also notice if it
5015 should not be recorded at all. In that case,
5016 prevent any further processing of this assignment. */
5017 do_not_record = 0;
5018 hash_arg_in_memory = 0;
5020 sets[i].src = src;
5021 sets[i].src_hash = HASH (src, mode);
5022 sets[i].src_volatile = do_not_record;
5023 sets[i].src_in_memory = hash_arg_in_memory;
5025 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5026 a pseudo, do not record SRC. Using SRC as a replacement for
5027 anything else will be incorrect in that situation. Note that
5028 this usually occurs only for stack slots, in which case all the
5029 RTL would be referring to SRC, so we don't lose any optimization
5030 opportunities by not having SRC in the hash table. */
5032 if (MEM_P (src)
5033 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5034 && REG_P (dest)
5035 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5036 sets[i].src_volatile = 1;
5038 #if 0
5039 /* It is no longer clear why we used to do this, but it doesn't
5040 appear to still be needed. So let's try without it since this
5041 code hurts cse'ing widened ops. */
5042 /* If source is a paradoxical subreg (such as QI treated as an SI),
5043 treat it as volatile. It may do the work of an SI in one context
5044 where the extra bits are not being used, but cannot replace an SI
5045 in general. */
5046 if (GET_CODE (src) == SUBREG
5047 && (GET_MODE_SIZE (GET_MODE (src))
5048 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5049 sets[i].src_volatile = 1;
5050 #endif
5052 /* Locate all possible equivalent forms for SRC. Try to replace
5053 SRC in the insn with each cheaper equivalent.
5055 We have the following types of equivalents: SRC itself, a folded
5056 version, a value given in a REG_EQUAL note, or a value related
5057 to a constant.
5059 Each of these equivalents may be part of an additional class
5060 of equivalents (if more than one is in the table, they must be in
5061 the same class; we check for this).
5063 If the source is volatile, we don't do any table lookups.
5065 We note any constant equivalent for possible later use in a
5066 REG_NOTE. */
5068 if (!sets[i].src_volatile)
5069 elt = lookup (src, sets[i].src_hash, mode);
5071 sets[i].src_elt = elt;
5073 if (elt && src_eqv_here && src_eqv_elt)
5075 if (elt->first_same_value != src_eqv_elt->first_same_value)
5077 /* The REG_EQUAL is indicating that two formerly distinct
5078 classes are now equivalent. So merge them. */
5079 merge_equiv_classes (elt, src_eqv_elt);
5080 src_eqv_hash = HASH (src_eqv, elt->mode);
5081 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5084 src_eqv_here = 0;
5087 else if (src_eqv_elt)
5088 elt = src_eqv_elt;
5090 /* Try to find a constant somewhere and record it in `src_const'.
5091 Record its table element, if any, in `src_const_elt'. Look in
5092 any known equivalences first. (If the constant is not in the
5093 table, also set `sets[i].src_const_hash'). */
5094 if (elt)
5095 for (p = elt->first_same_value; p; p = p->next_same_value)
5096 if (p->is_const)
5098 src_const = p->exp;
5099 src_const_elt = elt;
5100 break;
5103 if (src_const == 0
5104 && (CONSTANT_P (src_folded)
5105 /* Consider (minus (label_ref L1) (label_ref L2)) as
5106 "constant" here so we will record it. This allows us
5107 to fold switch statements when an ADDR_DIFF_VEC is used. */
5108 || (GET_CODE (src_folded) == MINUS
5109 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5110 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5111 src_const = src_folded, src_const_elt = elt;
5112 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5113 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5115 /* If we don't know if the constant is in the table, get its
5116 hash code and look it up. */
5117 if (src_const && src_const_elt == 0)
5119 sets[i].src_const_hash = HASH (src_const, mode);
5120 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5123 sets[i].src_const = src_const;
5124 sets[i].src_const_elt = src_const_elt;
5126 /* If the constant and our source are both in the table, mark them as
5127 equivalent. Otherwise, if a constant is in the table but the source
5128 isn't, set ELT to it. */
5129 if (src_const_elt && elt
5130 && src_const_elt->first_same_value != elt->first_same_value)
5131 merge_equiv_classes (elt, src_const_elt);
5132 else if (src_const_elt && elt == 0)
5133 elt = src_const_elt;
5135 /* See if there is a register linearly related to a constant
5136 equivalent of SRC. */
5137 if (src_const
5138 && (GET_CODE (src_const) == CONST
5139 || (src_const_elt && src_const_elt->related_value != 0)))
5141 src_related = use_related_value (src_const, src_const_elt);
5142 if (src_related)
5144 struct table_elt *src_related_elt
5145 = lookup (src_related, HASH (src_related, mode), mode);
5146 if (src_related_elt && elt)
5148 if (elt->first_same_value
5149 != src_related_elt->first_same_value)
5150 /* This can occur when we previously saw a CONST
5151 involving a SYMBOL_REF and then see the SYMBOL_REF
5152 twice. Merge the involved classes. */
5153 merge_equiv_classes (elt, src_related_elt);
5155 src_related = 0;
5156 src_related_elt = 0;
5158 else if (src_related_elt && elt == 0)
5159 elt = src_related_elt;
5163 /* See if we have a CONST_INT that is already in a register in a
5164 wider mode. */
5166 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5167 && GET_MODE_CLASS (mode) == MODE_INT
5168 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5170 enum machine_mode wider_mode;
5172 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5173 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5174 && src_related == 0;
5175 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5177 struct table_elt *const_elt
5178 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5180 if (const_elt == 0)
5181 continue;
5183 for (const_elt = const_elt->first_same_value;
5184 const_elt; const_elt = const_elt->next_same_value)
5185 if (REG_P (const_elt->exp))
5187 src_related = gen_lowpart (mode,
5188 const_elt->exp);
5189 break;
5194 /* Another possibility is that we have an AND with a constant in
5195 a mode narrower than a word. If so, it might have been generated
5196 as part of an "if" which would narrow the AND. If we already
5197 have done the AND in a wider mode, we can use a SUBREG of that
5198 value. */
5200 if (flag_expensive_optimizations && ! src_related
5201 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5202 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5204 enum machine_mode tmode;
5205 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5207 for (tmode = GET_MODE_WIDER_MODE (mode);
5208 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5209 tmode = GET_MODE_WIDER_MODE (tmode))
5211 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5212 struct table_elt *larger_elt;
5214 if (inner)
5216 PUT_MODE (new_and, tmode);
5217 XEXP (new_and, 0) = inner;
5218 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5219 if (larger_elt == 0)
5220 continue;
5222 for (larger_elt = larger_elt->first_same_value;
5223 larger_elt; larger_elt = larger_elt->next_same_value)
5224 if (REG_P (larger_elt->exp))
5226 src_related
5227 = gen_lowpart (mode, larger_elt->exp);
5228 break;
5231 if (src_related)
5232 break;
5237 #ifdef LOAD_EXTEND_OP
5238 /* See if a MEM has already been loaded with a widening operation;
5239 if it has, we can use a subreg of that. Many CISC machines
5240 also have such operations, but this is only likely to be
5241 beneficial on these machines. */
5243 if (flag_expensive_optimizations && src_related == 0
5244 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5245 && GET_MODE_CLASS (mode) == MODE_INT
5246 && MEM_P (src) && ! do_not_record
5247 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5249 struct rtx_def memory_extend_buf;
5250 rtx memory_extend_rtx = &memory_extend_buf;
5251 enum machine_mode tmode;
5253 /* Set what we are trying to extend and the operation it might
5254 have been extended with. */
5255 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5256 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5257 XEXP (memory_extend_rtx, 0) = src;
5259 for (tmode = GET_MODE_WIDER_MODE (mode);
5260 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5261 tmode = GET_MODE_WIDER_MODE (tmode))
5263 struct table_elt *larger_elt;
5265 PUT_MODE (memory_extend_rtx, tmode);
5266 larger_elt = lookup (memory_extend_rtx,
5267 HASH (memory_extend_rtx, tmode), tmode);
5268 if (larger_elt == 0)
5269 continue;
5271 for (larger_elt = larger_elt->first_same_value;
5272 larger_elt; larger_elt = larger_elt->next_same_value)
5273 if (REG_P (larger_elt->exp))
5275 src_related = gen_lowpart (mode,
5276 larger_elt->exp);
5277 break;
5280 if (src_related)
5281 break;
5284 #endif /* LOAD_EXTEND_OP */
5286 if (src == src_folded)
5287 src_folded = 0;
5289 /* At this point, ELT, if nonzero, points to a class of expressions
5290 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5291 and SRC_RELATED, if nonzero, each contain additional equivalent
5292 expressions. Prune these latter expressions by deleting expressions
5293 already in the equivalence class.
5295 Check for an equivalent identical to the destination. If found,
5296 this is the preferred equivalent since it will likely lead to
5297 elimination of the insn. Indicate this by placing it in
5298 `src_related'. */
5300 if (elt)
5301 elt = elt->first_same_value;
5302 for (p = elt; p; p = p->next_same_value)
5304 enum rtx_code code = GET_CODE (p->exp);
5306 /* If the expression is not valid, ignore it. Then we do not
5307 have to check for validity below. In most cases, we can use
5308 `rtx_equal_p', since canonicalization has already been done. */
5309 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5310 continue;
5312 /* Also skip paradoxical subregs, unless that's what we're
5313 looking for. */
5314 if (code == SUBREG
5315 && (GET_MODE_SIZE (GET_MODE (p->exp))
5316 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5317 && ! (src != 0
5318 && GET_CODE (src) == SUBREG
5319 && GET_MODE (src) == GET_MODE (p->exp)
5320 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5321 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5322 continue;
5324 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5325 src = 0;
5326 else if (src_folded && GET_CODE (src_folded) == code
5327 && rtx_equal_p (src_folded, p->exp))
5328 src_folded = 0;
5329 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5330 && rtx_equal_p (src_eqv_here, p->exp))
5331 src_eqv_here = 0;
5332 else if (src_related && GET_CODE (src_related) == code
5333 && rtx_equal_p (src_related, p->exp))
5334 src_related = 0;
5336 /* This is the same as the destination of the insns, we want
5337 to prefer it. Copy it to src_related. The code below will
5338 then give it a negative cost. */
5339 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5340 src_related = dest;
5343 /* Find the cheapest valid equivalent, trying all the available
5344 possibilities. Prefer items not in the hash table to ones
5345 that are when they are equal cost. Note that we can never
5346 worsen an insn as the current contents will also succeed.
5347 If we find an equivalent identical to the destination, use it as best,
5348 since this insn will probably be eliminated in that case. */
5349 if (src)
5351 if (rtx_equal_p (src, dest))
5352 src_cost = src_regcost = -1;
5353 else
5355 src_cost = COST (src);
5356 src_regcost = approx_reg_cost (src);
5360 if (src_eqv_here)
5362 if (rtx_equal_p (src_eqv_here, dest))
5363 src_eqv_cost = src_eqv_regcost = -1;
5364 else
5366 src_eqv_cost = COST (src_eqv_here);
5367 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5371 if (src_folded)
5373 if (rtx_equal_p (src_folded, dest))
5374 src_folded_cost = src_folded_regcost = -1;
5375 else
5377 src_folded_cost = COST (src_folded);
5378 src_folded_regcost = approx_reg_cost (src_folded);
5382 if (src_related)
5384 if (rtx_equal_p (src_related, dest))
5385 src_related_cost = src_related_regcost = -1;
5386 else
5388 src_related_cost = COST (src_related);
5389 src_related_regcost = approx_reg_cost (src_related);
5393 /* If this was an indirect jump insn, a known label will really be
5394 cheaper even though it looks more expensive. */
5395 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5396 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5398 /* Terminate loop when replacement made. This must terminate since
5399 the current contents will be tested and will always be valid. */
5400 while (1)
5402 rtx trial;
5404 /* Skip invalid entries. */
5405 while (elt && !REG_P (elt->exp)
5406 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5407 elt = elt->next_same_value;
5409 /* A paradoxical subreg would be bad here: it'll be the right
5410 size, but later may be adjusted so that the upper bits aren't
5411 what we want. So reject it. */
5412 if (elt != 0
5413 && GET_CODE (elt->exp) == SUBREG
5414 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5415 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5416 /* It is okay, though, if the rtx we're trying to match
5417 will ignore any of the bits we can't predict. */
5418 && ! (src != 0
5419 && GET_CODE (src) == SUBREG
5420 && GET_MODE (src) == GET_MODE (elt->exp)
5421 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5422 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5424 elt = elt->next_same_value;
5425 continue;
5428 if (elt)
5430 src_elt_cost = elt->cost;
5431 src_elt_regcost = elt->regcost;
5434 /* Find cheapest and skip it for the next time. For items
5435 of equal cost, use this order:
5436 src_folded, src, src_eqv, src_related and hash table entry. */
5437 if (src_folded
5438 && preferable (src_folded_cost, src_folded_regcost,
5439 src_cost, src_regcost) <= 0
5440 && preferable (src_folded_cost, src_folded_regcost,
5441 src_eqv_cost, src_eqv_regcost) <= 0
5442 && preferable (src_folded_cost, src_folded_regcost,
5443 src_related_cost, src_related_regcost) <= 0
5444 && preferable (src_folded_cost, src_folded_regcost,
5445 src_elt_cost, src_elt_regcost) <= 0)
5447 trial = src_folded, src_folded_cost = MAX_COST;
5448 if (src_folded_force_flag)
5450 rtx forced = force_const_mem (mode, trial);
5451 if (forced)
5452 trial = forced;
5455 else if (src
5456 && preferable (src_cost, src_regcost,
5457 src_eqv_cost, src_eqv_regcost) <= 0
5458 && preferable (src_cost, src_regcost,
5459 src_related_cost, src_related_regcost) <= 0
5460 && preferable (src_cost, src_regcost,
5461 src_elt_cost, src_elt_regcost) <= 0)
5462 trial = src, src_cost = MAX_COST;
5463 else if (src_eqv_here
5464 && preferable (src_eqv_cost, src_eqv_regcost,
5465 src_related_cost, src_related_regcost) <= 0
5466 && preferable (src_eqv_cost, src_eqv_regcost,
5467 src_elt_cost, src_elt_regcost) <= 0)
5468 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5469 else if (src_related
5470 && preferable (src_related_cost, src_related_regcost,
5471 src_elt_cost, src_elt_regcost) <= 0)
5472 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5473 else
5475 trial = copy_rtx (elt->exp);
5476 elt = elt->next_same_value;
5477 src_elt_cost = MAX_COST;
5480 /* We don't normally have an insn matching (set (pc) (pc)), so
5481 check for this separately here. We will delete such an
5482 insn below.
5484 For other cases such as a table jump or conditional jump
5485 where we know the ultimate target, go ahead and replace the
5486 operand. While that may not make a valid insn, we will
5487 reemit the jump below (and also insert any necessary
5488 barriers). */
5489 if (n_sets == 1 && dest == pc_rtx
5490 && (trial == pc_rtx
5491 || (GET_CODE (trial) == LABEL_REF
5492 && ! condjump_p (insn))))
5494 /* Don't substitute non-local labels, this confuses CFG. */
5495 if (GET_CODE (trial) == LABEL_REF
5496 && LABEL_REF_NONLOCAL_P (trial))
5497 continue;
5499 SET_SRC (sets[i].rtl) = trial;
5500 cse_jumps_altered = 1;
5501 break;
5504 /* Look for a substitution that makes a valid insn. */
5505 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5507 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5509 /* If we just made a substitution inside a libcall, then we
5510 need to make the same substitution in any notes attached
5511 to the RETVAL insn. */
5512 if (libcall_insn
5513 && (REG_P (sets[i].orig_src)
5514 || GET_CODE (sets[i].orig_src) == SUBREG
5515 || MEM_P (sets[i].orig_src)))
5517 rtx note = find_reg_equal_equiv_note (libcall_insn);
5518 if (note != 0)
5519 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5520 sets[i].orig_src,
5521 copy_rtx (new));
5524 /* The result of apply_change_group can be ignored; see
5525 canon_reg. */
5527 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5528 apply_change_group ();
5529 break;
5532 /* If we previously found constant pool entries for
5533 constants and this is a constant, try making a
5534 pool entry. Put it in src_folded unless we already have done
5535 this since that is where it likely came from. */
5537 else if (constant_pool_entries_cost
5538 && CONSTANT_P (trial)
5539 /* Reject cases that will cause decode_rtx_const to
5540 die. On the alpha when simplifying a switch, we
5541 get (const (truncate (minus (label_ref)
5542 (label_ref)))). */
5543 && ! (GET_CODE (trial) == CONST
5544 && GET_CODE (XEXP (trial, 0)) == TRUNCATE)
5545 /* Likewise on IA-64, except without the truncate. */
5546 && ! (GET_CODE (trial) == CONST
5547 && GET_CODE (XEXP (trial, 0)) == MINUS
5548 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5549 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)
5550 && (src_folded == 0
5551 || (!MEM_P (src_folded)
5552 && ! src_folded_force_flag))
5553 && GET_MODE_CLASS (mode) != MODE_CC
5554 && mode != VOIDmode)
5556 src_folded_force_flag = 1;
5557 src_folded = trial;
5558 src_folded_cost = constant_pool_entries_cost;
5559 src_folded_regcost = constant_pool_entries_regcost;
5563 src = SET_SRC (sets[i].rtl);
5565 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5566 However, there is an important exception: If both are registers
5567 that are not the head of their equivalence class, replace SET_SRC
5568 with the head of the class. If we do not do this, we will have
5569 both registers live over a portion of the basic block. This way,
5570 their lifetimes will likely abut instead of overlapping. */
5571 if (REG_P (dest)
5572 && REGNO_QTY_VALID_P (REGNO (dest)))
5574 int dest_q = REG_QTY (REGNO (dest));
5575 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5577 if (dest_ent->mode == GET_MODE (dest)
5578 && dest_ent->first_reg != REGNO (dest)
5579 && REG_P (src) && REGNO (src) == REGNO (dest)
5580 /* Don't do this if the original insn had a hard reg as
5581 SET_SRC or SET_DEST. */
5582 && (!REG_P (sets[i].src)
5583 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5584 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5585 /* We can't call canon_reg here because it won't do anything if
5586 SRC is a hard register. */
5588 int src_q = REG_QTY (REGNO (src));
5589 struct qty_table_elem *src_ent = &qty_table[src_q];
5590 int first = src_ent->first_reg;
5591 rtx new_src
5592 = (first >= FIRST_PSEUDO_REGISTER
5593 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5595 /* We must use validate-change even for this, because this
5596 might be a special no-op instruction, suitable only to
5597 tag notes onto. */
5598 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5600 src = new_src;
5601 /* If we had a constant that is cheaper than what we are now
5602 setting SRC to, use that constant. We ignored it when we
5603 thought we could make this into a no-op. */
5604 if (src_const && COST (src_const) < COST (src)
5605 && validate_change (insn, &SET_SRC (sets[i].rtl),
5606 src_const, 0))
5607 src = src_const;
5612 /* If we made a change, recompute SRC values. */
5613 if (src != sets[i].src)
5615 cse_altered = 1;
5616 do_not_record = 0;
5617 hash_arg_in_memory = 0;
5618 sets[i].src = src;
5619 sets[i].src_hash = HASH (src, mode);
5620 sets[i].src_volatile = do_not_record;
5621 sets[i].src_in_memory = hash_arg_in_memory;
5622 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5625 /* If this is a single SET, we are setting a register, and we have an
5626 equivalent constant, we want to add a REG_NOTE. We don't want
5627 to write a REG_EQUAL note for a constant pseudo since verifying that
5628 that pseudo hasn't been eliminated is a pain. Such a note also
5629 won't help anything.
5631 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5632 which can be created for a reference to a compile time computable
5633 entry in a jump table. */
5635 if (n_sets == 1 && src_const && REG_P (dest)
5636 && !REG_P (src_const)
5637 && ! (GET_CODE (src_const) == CONST
5638 && GET_CODE (XEXP (src_const, 0)) == MINUS
5639 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5640 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5642 /* We only want a REG_EQUAL note if src_const != src. */
5643 if (! rtx_equal_p (src, src_const))
5645 /* Make sure that the rtx is not shared. */
5646 src_const = copy_rtx (src_const);
5648 /* Record the actual constant value in a REG_EQUAL note,
5649 making a new one if one does not already exist. */
5650 set_unique_reg_note (insn, REG_EQUAL, src_const);
5654 /* Now deal with the destination. */
5655 do_not_record = 0;
5657 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5658 while (GET_CODE (dest) == SUBREG
5659 || GET_CODE (dest) == ZERO_EXTRACT
5660 || GET_CODE (dest) == STRICT_LOW_PART)
5661 dest = XEXP (dest, 0);
5663 sets[i].inner_dest = dest;
5665 if (MEM_P (dest))
5667 #ifdef PUSH_ROUNDING
5668 /* Stack pushes invalidate the stack pointer. */
5669 rtx addr = XEXP (dest, 0);
5670 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5671 && XEXP (addr, 0) == stack_pointer_rtx)
5672 invalidate (stack_pointer_rtx, Pmode);
5673 #endif
5674 dest = fold_rtx (dest, insn);
5677 /* Compute the hash code of the destination now,
5678 before the effects of this instruction are recorded,
5679 since the register values used in the address computation
5680 are those before this instruction. */
5681 sets[i].dest_hash = HASH (dest, mode);
5683 /* Don't enter a bit-field in the hash table
5684 because the value in it after the store
5685 may not equal what was stored, due to truncation. */
5687 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5689 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5691 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5692 && GET_CODE (width) == CONST_INT
5693 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5694 && ! (INTVAL (src_const)
5695 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5696 /* Exception: if the value is constant,
5697 and it won't be truncated, record it. */
5699 else
5701 /* This is chosen so that the destination will be invalidated
5702 but no new value will be recorded.
5703 We must invalidate because sometimes constant
5704 values can be recorded for bitfields. */
5705 sets[i].src_elt = 0;
5706 sets[i].src_volatile = 1;
5707 src_eqv = 0;
5708 src_eqv_elt = 0;
5712 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5713 the insn. */
5714 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5716 /* One less use of the label this insn used to jump to. */
5717 delete_insn (insn);
5718 cse_jumps_altered = 1;
5719 /* No more processing for this set. */
5720 sets[i].rtl = 0;
5723 /* If this SET is now setting PC to a label, we know it used to
5724 be a conditional or computed branch. */
5725 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5726 && !LABEL_REF_NONLOCAL_P (src))
5728 /* Now emit a BARRIER after the unconditional jump. */
5729 if (NEXT_INSN (insn) == 0
5730 || !BARRIER_P (NEXT_INSN (insn)))
5731 emit_barrier_after (insn);
5733 /* We reemit the jump in as many cases as possible just in
5734 case the form of an unconditional jump is significantly
5735 different than a computed jump or conditional jump.
5737 If this insn has multiple sets, then reemitting the
5738 jump is nontrivial. So instead we just force rerecognition
5739 and hope for the best. */
5740 if (n_sets == 1)
5742 rtx new, note;
5744 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5745 JUMP_LABEL (new) = XEXP (src, 0);
5746 LABEL_NUSES (XEXP (src, 0))++;
5748 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5749 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5750 if (note)
5752 XEXP (note, 1) = NULL_RTX;
5753 REG_NOTES (new) = note;
5756 delete_insn (insn);
5757 insn = new;
5759 /* Now emit a BARRIER after the unconditional jump. */
5760 if (NEXT_INSN (insn) == 0
5761 || !BARRIER_P (NEXT_INSN (insn)))
5762 emit_barrier_after (insn);
5764 else
5765 INSN_CODE (insn) = -1;
5767 /* Do not bother deleting any unreachable code,
5768 let jump/flow do that. */
5770 cse_jumps_altered = 1;
5771 sets[i].rtl = 0;
5774 /* If destination is volatile, invalidate it and then do no further
5775 processing for this assignment. */
5777 else if (do_not_record)
5779 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5780 invalidate (dest, VOIDmode);
5781 else if (MEM_P (dest))
5782 invalidate (dest, VOIDmode);
5783 else if (GET_CODE (dest) == STRICT_LOW_PART
5784 || GET_CODE (dest) == ZERO_EXTRACT)
5785 invalidate (XEXP (dest, 0), GET_MODE (dest));
5786 sets[i].rtl = 0;
5789 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5790 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5792 #ifdef HAVE_cc0
5793 /* If setting CC0, record what it was set to, or a constant, if it
5794 is equivalent to a constant. If it is being set to a floating-point
5795 value, make a COMPARE with the appropriate constant of 0. If we
5796 don't do this, later code can interpret this as a test against
5797 const0_rtx, which can cause problems if we try to put it into an
5798 insn as a floating-point operand. */
5799 if (dest == cc0_rtx)
5801 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5802 this_insn_cc0_mode = mode;
5803 if (FLOAT_MODE_P (mode))
5804 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5805 CONST0_RTX (mode));
5807 #endif
5810 /* Now enter all non-volatile source expressions in the hash table
5811 if they are not already present.
5812 Record their equivalence classes in src_elt.
5813 This way we can insert the corresponding destinations into
5814 the same classes even if the actual sources are no longer in them
5815 (having been invalidated). */
5817 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5818 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5820 struct table_elt *elt;
5821 struct table_elt *classp = sets[0].src_elt;
5822 rtx dest = SET_DEST (sets[0].rtl);
5823 enum machine_mode eqvmode = GET_MODE (dest);
5825 if (GET_CODE (dest) == STRICT_LOW_PART)
5827 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5828 classp = 0;
5830 if (insert_regs (src_eqv, classp, 0))
5832 rehash_using_reg (src_eqv);
5833 src_eqv_hash = HASH (src_eqv, eqvmode);
5835 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5836 elt->in_memory = src_eqv_in_memory;
5837 src_eqv_elt = elt;
5839 /* Check to see if src_eqv_elt is the same as a set source which
5840 does not yet have an elt, and if so set the elt of the set source
5841 to src_eqv_elt. */
5842 for (i = 0; i < n_sets; i++)
5843 if (sets[i].rtl && sets[i].src_elt == 0
5844 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5845 sets[i].src_elt = src_eqv_elt;
5848 for (i = 0; i < n_sets; i++)
5849 if (sets[i].rtl && ! sets[i].src_volatile
5850 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5852 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5854 /* REG_EQUAL in setting a STRICT_LOW_PART
5855 gives an equivalent for the entire destination register,
5856 not just for the subreg being stored in now.
5857 This is a more interesting equivalence, so we arrange later
5858 to treat the entire reg as the destination. */
5859 sets[i].src_elt = src_eqv_elt;
5860 sets[i].src_hash = src_eqv_hash;
5862 else
5864 /* Insert source and constant equivalent into hash table, if not
5865 already present. */
5866 struct table_elt *classp = src_eqv_elt;
5867 rtx src = sets[i].src;
5868 rtx dest = SET_DEST (sets[i].rtl);
5869 enum machine_mode mode
5870 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5872 /* It's possible that we have a source value known to be
5873 constant but don't have a REG_EQUAL note on the insn.
5874 Lack of a note will mean src_eqv_elt will be NULL. This
5875 can happen where we've generated a SUBREG to access a
5876 CONST_INT that is already in a register in a wider mode.
5877 Ensure that the source expression is put in the proper
5878 constant class. */
5879 if (!classp)
5880 classp = sets[i].src_const_elt;
5882 if (sets[i].src_elt == 0)
5884 /* Don't put a hard register source into the table if this is
5885 the last insn of a libcall. In this case, we only need
5886 to put src_eqv_elt in src_elt. */
5887 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5889 struct table_elt *elt;
5891 /* Note that these insert_regs calls cannot remove
5892 any of the src_elt's, because they would have failed to
5893 match if not still valid. */
5894 if (insert_regs (src, classp, 0))
5896 rehash_using_reg (src);
5897 sets[i].src_hash = HASH (src, mode);
5899 elt = insert (src, classp, sets[i].src_hash, mode);
5900 elt->in_memory = sets[i].src_in_memory;
5901 sets[i].src_elt = classp = elt;
5903 else
5904 sets[i].src_elt = classp;
5906 if (sets[i].src_const && sets[i].src_const_elt == 0
5907 && src != sets[i].src_const
5908 && ! rtx_equal_p (sets[i].src_const, src))
5909 sets[i].src_elt = insert (sets[i].src_const, classp,
5910 sets[i].src_const_hash, mode);
5913 else if (sets[i].src_elt == 0)
5914 /* If we did not insert the source into the hash table (e.g., it was
5915 volatile), note the equivalence class for the REG_EQUAL value, if any,
5916 so that the destination goes into that class. */
5917 sets[i].src_elt = src_eqv_elt;
5919 invalidate_from_clobbers (x);
5921 /* Some registers are invalidated by subroutine calls. Memory is
5922 invalidated by non-constant calls. */
5924 if (CALL_P (insn))
5926 if (! CONST_OR_PURE_CALL_P (insn))
5927 invalidate_memory ();
5928 invalidate_for_call ();
5931 /* Now invalidate everything set by this instruction.
5932 If a SUBREG or other funny destination is being set,
5933 sets[i].rtl is still nonzero, so here we invalidate the reg
5934 a part of which is being set. */
5936 for (i = 0; i < n_sets; i++)
5937 if (sets[i].rtl)
5939 /* We can't use the inner dest, because the mode associated with
5940 a ZERO_EXTRACT is significant. */
5941 rtx dest = SET_DEST (sets[i].rtl);
5943 /* Needed for registers to remove the register from its
5944 previous quantity's chain.
5945 Needed for memory if this is a nonvarying address, unless
5946 we have just done an invalidate_memory that covers even those. */
5947 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5948 invalidate (dest, VOIDmode);
5949 else if (MEM_P (dest))
5950 invalidate (dest, VOIDmode);
5951 else if (GET_CODE (dest) == STRICT_LOW_PART
5952 || GET_CODE (dest) == ZERO_EXTRACT)
5953 invalidate (XEXP (dest, 0), GET_MODE (dest));
5956 /* A volatile ASM invalidates everything. */
5957 if (NONJUMP_INSN_P (insn)
5958 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5959 && MEM_VOLATILE_P (PATTERN (insn)))
5960 flush_hash_table ();
5962 /* Make sure registers mentioned in destinations
5963 are safe for use in an expression to be inserted.
5964 This removes from the hash table
5965 any invalid entry that refers to one of these registers.
5967 We don't care about the return value from mention_regs because
5968 we are going to hash the SET_DEST values unconditionally. */
5970 for (i = 0; i < n_sets; i++)
5972 if (sets[i].rtl)
5974 rtx x = SET_DEST (sets[i].rtl);
5976 if (!REG_P (x))
5977 mention_regs (x);
5978 else
5980 /* We used to rely on all references to a register becoming
5981 inaccessible when a register changes to a new quantity,
5982 since that changes the hash code. However, that is not
5983 safe, since after HASH_SIZE new quantities we get a
5984 hash 'collision' of a register with its own invalid
5985 entries. And since SUBREGs have been changed not to
5986 change their hash code with the hash code of the register,
5987 it wouldn't work any longer at all. So we have to check
5988 for any invalid references lying around now.
5989 This code is similar to the REG case in mention_regs,
5990 but it knows that reg_tick has been incremented, and
5991 it leaves reg_in_table as -1 . */
5992 unsigned int regno = REGNO (x);
5993 unsigned int endregno
5994 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
5995 : hard_regno_nregs[regno][GET_MODE (x)]);
5996 unsigned int i;
5998 for (i = regno; i < endregno; i++)
6000 if (REG_IN_TABLE (i) >= 0)
6002 remove_invalid_refs (i);
6003 REG_IN_TABLE (i) = -1;
6010 /* We may have just removed some of the src_elt's from the hash table.
6011 So replace each one with the current head of the same class. */
6013 for (i = 0; i < n_sets; i++)
6014 if (sets[i].rtl)
6016 if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6017 /* If elt was removed, find current head of same class,
6018 or 0 if nothing remains of that class. */
6020 struct table_elt *elt = sets[i].src_elt;
6022 while (elt && elt->prev_same_value)
6023 elt = elt->prev_same_value;
6025 while (elt && elt->first_same_value == 0)
6026 elt = elt->next_same_value;
6027 sets[i].src_elt = elt ? elt->first_same_value : 0;
6031 /* Now insert the destinations into their equivalence classes. */
6033 for (i = 0; i < n_sets; i++)
6034 if (sets[i].rtl)
6036 rtx dest = SET_DEST (sets[i].rtl);
6037 struct table_elt *elt;
6039 /* Don't record value if we are not supposed to risk allocating
6040 floating-point values in registers that might be wider than
6041 memory. */
6042 if ((flag_float_store
6043 && MEM_P (dest)
6044 && FLOAT_MODE_P (GET_MODE (dest)))
6045 /* Don't record BLKmode values, because we don't know the
6046 size of it, and can't be sure that other BLKmode values
6047 have the same or smaller size. */
6048 || GET_MODE (dest) == BLKmode
6049 /* Don't record values of destinations set inside a libcall block
6050 since we might delete the libcall. Things should have been set
6051 up so we won't want to reuse such a value, but we play it safe
6052 here. */
6053 || libcall_insn
6054 /* If we didn't put a REG_EQUAL value or a source into the hash
6055 table, there is no point is recording DEST. */
6056 || sets[i].src_elt == 0
6057 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6058 or SIGN_EXTEND, don't record DEST since it can cause
6059 some tracking to be wrong.
6061 ??? Think about this more later. */
6062 || (GET_CODE (dest) == SUBREG
6063 && (GET_MODE_SIZE (GET_MODE (dest))
6064 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6065 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6066 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6067 continue;
6069 /* STRICT_LOW_PART isn't part of the value BEING set,
6070 and neither is the SUBREG inside it.
6071 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6072 if (GET_CODE (dest) == STRICT_LOW_PART)
6073 dest = SUBREG_REG (XEXP (dest, 0));
6075 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6076 /* Registers must also be inserted into chains for quantities. */
6077 if (insert_regs (dest, sets[i].src_elt, 1))
6079 /* If `insert_regs' changes something, the hash code must be
6080 recalculated. */
6081 rehash_using_reg (dest);
6082 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6085 elt = insert (dest, sets[i].src_elt,
6086 sets[i].dest_hash, GET_MODE (dest));
6088 elt->in_memory = (MEM_P (sets[i].inner_dest)
6089 && !MEM_READONLY_P (sets[i].inner_dest));
6091 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6092 narrower than M2, and both M1 and M2 are the same number of words,
6093 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6094 make that equivalence as well.
6096 However, BAR may have equivalences for which gen_lowpart
6097 will produce a simpler value than gen_lowpart applied to
6098 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6099 BAR's equivalences. If we don't get a simplified form, make
6100 the SUBREG. It will not be used in an equivalence, but will
6101 cause two similar assignments to be detected.
6103 Note the loop below will find SUBREG_REG (DEST) since we have
6104 already entered SRC and DEST of the SET in the table. */
6106 if (GET_CODE (dest) == SUBREG
6107 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6108 / UNITS_PER_WORD)
6109 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6110 && (GET_MODE_SIZE (GET_MODE (dest))
6111 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6112 && sets[i].src_elt != 0)
6114 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6115 struct table_elt *elt, *classp = 0;
6117 for (elt = sets[i].src_elt->first_same_value; elt;
6118 elt = elt->next_same_value)
6120 rtx new_src = 0;
6121 unsigned src_hash;
6122 struct table_elt *src_elt;
6123 int byte = 0;
6125 /* Ignore invalid entries. */
6126 if (!REG_P (elt->exp)
6127 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6128 continue;
6130 /* We may have already been playing subreg games. If the
6131 mode is already correct for the destination, use it. */
6132 if (GET_MODE (elt->exp) == new_mode)
6133 new_src = elt->exp;
6134 else
6136 /* Calculate big endian correction for the SUBREG_BYTE.
6137 We have already checked that M1 (GET_MODE (dest))
6138 is not narrower than M2 (new_mode). */
6139 if (BYTES_BIG_ENDIAN)
6140 byte = (GET_MODE_SIZE (GET_MODE (dest))
6141 - GET_MODE_SIZE (new_mode));
6143 new_src = simplify_gen_subreg (new_mode, elt->exp,
6144 GET_MODE (dest), byte);
6147 /* The call to simplify_gen_subreg fails if the value
6148 is VOIDmode, yet we can't do any simplification, e.g.
6149 for EXPR_LISTs denoting function call results.
6150 It is invalid to construct a SUBREG with a VOIDmode
6151 SUBREG_REG, hence a zero new_src means we can't do
6152 this substitution. */
6153 if (! new_src)
6154 continue;
6156 src_hash = HASH (new_src, new_mode);
6157 src_elt = lookup (new_src, src_hash, new_mode);
6159 /* Put the new source in the hash table is if isn't
6160 already. */
6161 if (src_elt == 0)
6163 if (insert_regs (new_src, classp, 0))
6165 rehash_using_reg (new_src);
6166 src_hash = HASH (new_src, new_mode);
6168 src_elt = insert (new_src, classp, src_hash, new_mode);
6169 src_elt->in_memory = elt->in_memory;
6171 else if (classp && classp != src_elt->first_same_value)
6172 /* Show that two things that we've seen before are
6173 actually the same. */
6174 merge_equiv_classes (src_elt, classp);
6176 classp = src_elt->first_same_value;
6177 /* Ignore invalid entries. */
6178 while (classp
6179 && !REG_P (classp->exp)
6180 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6181 classp = classp->next_same_value;
6186 /* Special handling for (set REG0 REG1) where REG0 is the
6187 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6188 be used in the sequel, so (if easily done) change this insn to
6189 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6190 that computed their value. Then REG1 will become a dead store
6191 and won't cloud the situation for later optimizations.
6193 Do not make this change if REG1 is a hard register, because it will
6194 then be used in the sequel and we may be changing a two-operand insn
6195 into a three-operand insn.
6197 Also do not do this if we are operating on a copy of INSN.
6199 Also don't do this if INSN ends a libcall; this would cause an unrelated
6200 register to be set in the middle of a libcall, and we then get bad code
6201 if the libcall is deleted. */
6203 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6204 && NEXT_INSN (PREV_INSN (insn)) == insn
6205 && REG_P (SET_SRC (sets[0].rtl))
6206 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6207 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6209 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6210 struct qty_table_elem *src_ent = &qty_table[src_q];
6212 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6213 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6215 rtx prev = insn;
6216 /* Scan for the previous nonnote insn, but stop at a basic
6217 block boundary. */
6220 prev = PREV_INSN (prev);
6222 while (prev && NOTE_P (prev)
6223 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6225 /* Do not swap the registers around if the previous instruction
6226 attaches a REG_EQUIV note to REG1.
6228 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6229 from the pseudo that originally shadowed an incoming argument
6230 to another register. Some uses of REG_EQUIV might rely on it
6231 being attached to REG1 rather than REG2.
6233 This section previously turned the REG_EQUIV into a REG_EQUAL
6234 note. We cannot do that because REG_EQUIV may provide an
6235 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6237 if (prev != 0 && NONJUMP_INSN_P (prev)
6238 && GET_CODE (PATTERN (prev)) == SET
6239 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6240 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6242 rtx dest = SET_DEST (sets[0].rtl);
6243 rtx src = SET_SRC (sets[0].rtl);
6244 rtx note;
6246 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6247 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6248 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6249 apply_change_group ();
6251 /* If INSN has a REG_EQUAL note, and this note mentions
6252 REG0, then we must delete it, because the value in
6253 REG0 has changed. If the note's value is REG1, we must
6254 also delete it because that is now this insn's dest. */
6255 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6256 if (note != 0
6257 && (reg_mentioned_p (dest, XEXP (note, 0))
6258 || rtx_equal_p (src, XEXP (note, 0))))
6259 remove_note (insn, note);
6264 /* If this is a conditional jump insn, record any known equivalences due to
6265 the condition being tested. */
6267 if (JUMP_P (insn)
6268 && n_sets == 1 && GET_CODE (x) == SET
6269 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6270 record_jump_equiv (insn, 0);
6272 #ifdef HAVE_cc0
6273 /* If the previous insn set CC0 and this insn no longer references CC0,
6274 delete the previous insn. Here we use the fact that nothing expects CC0
6275 to be valid over an insn, which is true until the final pass. */
6276 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6277 && (tem = single_set (prev_insn)) != 0
6278 && SET_DEST (tem) == cc0_rtx
6279 && ! reg_mentioned_p (cc0_rtx, x))
6280 delete_insn (prev_insn);
6282 prev_insn_cc0 = this_insn_cc0;
6283 prev_insn_cc0_mode = this_insn_cc0_mode;
6284 prev_insn = insn;
6285 #endif
6288 /* Remove from the hash table all expressions that reference memory. */
6290 static void
6291 invalidate_memory (void)
6293 int i;
6294 struct table_elt *p, *next;
6296 for (i = 0; i < HASH_SIZE; i++)
6297 for (p = table[i]; p; p = next)
6299 next = p->next_same_hash;
6300 if (p->in_memory)
6301 remove_from_table (p, i);
6305 /* If ADDR is an address that implicitly affects the stack pointer, return
6306 1 and update the register tables to show the effect. Else, return 0. */
6308 static int
6309 addr_affects_sp_p (rtx addr)
6311 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6312 && REG_P (XEXP (addr, 0))
6313 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6315 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6317 REG_TICK (STACK_POINTER_REGNUM)++;
6318 /* Is it possible to use a subreg of SP? */
6319 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6322 /* This should be *very* rare. */
6323 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6324 invalidate (stack_pointer_rtx, VOIDmode);
6326 return 1;
6329 return 0;
6332 /* Perform invalidation on the basis of everything about an insn
6333 except for invalidating the actual places that are SET in it.
6334 This includes the places CLOBBERed, and anything that might
6335 alias with something that is SET or CLOBBERed.
6337 X is the pattern of the insn. */
6339 static void
6340 invalidate_from_clobbers (rtx x)
6342 if (GET_CODE (x) == CLOBBER)
6344 rtx ref = XEXP (x, 0);
6345 if (ref)
6347 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6348 || MEM_P (ref))
6349 invalidate (ref, VOIDmode);
6350 else if (GET_CODE (ref) == STRICT_LOW_PART
6351 || GET_CODE (ref) == ZERO_EXTRACT)
6352 invalidate (XEXP (ref, 0), GET_MODE (ref));
6355 else if (GET_CODE (x) == PARALLEL)
6357 int i;
6358 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6360 rtx y = XVECEXP (x, 0, i);
6361 if (GET_CODE (y) == CLOBBER)
6363 rtx ref = XEXP (y, 0);
6364 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6365 || MEM_P (ref))
6366 invalidate (ref, VOIDmode);
6367 else if (GET_CODE (ref) == STRICT_LOW_PART
6368 || GET_CODE (ref) == ZERO_EXTRACT)
6369 invalidate (XEXP (ref, 0), GET_MODE (ref));
6375 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6376 and replace any registers in them with either an equivalent constant
6377 or the canonical form of the register. If we are inside an address,
6378 only do this if the address remains valid.
6380 OBJECT is 0 except when within a MEM in which case it is the MEM.
6382 Return the replacement for X. */
6384 static rtx
6385 cse_process_notes (rtx x, rtx object)
6387 enum rtx_code code = GET_CODE (x);
6388 const char *fmt = GET_RTX_FORMAT (code);
6389 int i;
6391 switch (code)
6393 case CONST_INT:
6394 case CONST:
6395 case SYMBOL_REF:
6396 case LABEL_REF:
6397 case CONST_DOUBLE:
6398 case CONST_VECTOR:
6399 case PC:
6400 case CC0:
6401 case LO_SUM:
6402 return x;
6404 case MEM:
6405 validate_change (x, &XEXP (x, 0),
6406 cse_process_notes (XEXP (x, 0), x), 0);
6407 return x;
6409 case EXPR_LIST:
6410 case INSN_LIST:
6411 if (REG_NOTE_KIND (x) == REG_EQUAL)
6412 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6413 if (XEXP (x, 1))
6414 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6415 return x;
6417 case SIGN_EXTEND:
6418 case ZERO_EXTEND:
6419 case SUBREG:
6421 rtx new = cse_process_notes (XEXP (x, 0), object);
6422 /* We don't substitute VOIDmode constants into these rtx,
6423 since they would impede folding. */
6424 if (GET_MODE (new) != VOIDmode)
6425 validate_change (object, &XEXP (x, 0), new, 0);
6426 return x;
6429 case REG:
6430 i = REG_QTY (REGNO (x));
6432 /* Return a constant or a constant register. */
6433 if (REGNO_QTY_VALID_P (REGNO (x)))
6435 struct qty_table_elem *ent = &qty_table[i];
6437 if (ent->const_rtx != NULL_RTX
6438 && (CONSTANT_P (ent->const_rtx)
6439 || REG_P (ent->const_rtx)))
6441 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6442 if (new)
6443 return new;
6447 /* Otherwise, canonicalize this register. */
6448 return canon_reg (x, NULL_RTX);
6450 default:
6451 break;
6454 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6455 if (fmt[i] == 'e')
6456 validate_change (object, &XEXP (x, i),
6457 cse_process_notes (XEXP (x, i), object), 0);
6459 return x;
6462 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6463 since they are done elsewhere. This function is called via note_stores. */
6465 static void
6466 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6468 enum rtx_code code = GET_CODE (dest);
6470 if (code == MEM
6471 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6472 /* There are times when an address can appear varying and be a PLUS
6473 during this scan when it would be a fixed address were we to know
6474 the proper equivalences. So invalidate all memory if there is
6475 a BLKmode or nonscalar memory reference or a reference to a
6476 variable address. */
6477 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6478 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6480 invalidate_memory ();
6481 return;
6484 if (GET_CODE (set) == CLOBBER
6485 || CC0_P (dest)
6486 || dest == pc_rtx)
6487 return;
6489 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6490 invalidate (XEXP (dest, 0), GET_MODE (dest));
6491 else if (code == REG || code == SUBREG || code == MEM)
6492 invalidate (dest, VOIDmode);
6495 /* Invalidate all insns from START up to the end of the function or the
6496 next label. This called when we wish to CSE around a block that is
6497 conditionally executed. */
6499 static void
6500 invalidate_skipped_block (rtx start)
6502 rtx insn;
6504 for (insn = start; insn && !LABEL_P (insn);
6505 insn = NEXT_INSN (insn))
6507 if (! INSN_P (insn))
6508 continue;
6510 if (CALL_P (insn))
6512 if (! CONST_OR_PURE_CALL_P (insn))
6513 invalidate_memory ();
6514 invalidate_for_call ();
6517 invalidate_from_clobbers (PATTERN (insn));
6518 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6522 /* Find the end of INSN's basic block and return its range,
6523 the total number of SETs in all the insns of the block, the last insn of the
6524 block, and the branch path.
6526 The branch path indicates which branches should be followed. If a nonzero
6527 path size is specified, the block should be rescanned and a different set
6528 of branches will be taken. The branch path is only used if
6529 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6531 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6532 used to describe the block. It is filled in with the information about
6533 the current block. The incoming structure's branch path, if any, is used
6534 to construct the output branch path. */
6536 static void
6537 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6538 int follow_jumps, int skip_blocks)
6540 rtx p = insn, q;
6541 int nsets = 0;
6542 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6543 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6544 int path_size = data->path_size;
6545 int path_entry = 0;
6546 int i;
6548 /* Update the previous branch path, if any. If the last branch was
6549 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6550 If it was previously PATH_NOT_TAKEN,
6551 shorten the path by one and look at the previous branch. We know that
6552 at least one branch must have been taken if PATH_SIZE is nonzero. */
6553 while (path_size > 0)
6555 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6557 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6558 break;
6560 else
6561 path_size--;
6564 /* If the first instruction is marked with QImode, that means we've
6565 already processed this block. Our caller will look at DATA->LAST
6566 to figure out where to go next. We want to return the next block
6567 in the instruction stream, not some branched-to block somewhere
6568 else. We accomplish this by pretending our called forbid us to
6569 follow jumps, or skip blocks. */
6570 if (GET_MODE (insn) == QImode)
6571 follow_jumps = skip_blocks = 0;
6573 /* Scan to end of this basic block. */
6574 while (p && !LABEL_P (p))
6576 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6577 the regs restored by the longjmp come from
6578 a later time than the setjmp. */
6579 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6580 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6581 break;
6583 /* A PARALLEL can have lots of SETs in it,
6584 especially if it is really an ASM_OPERANDS. */
6585 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6586 nsets += XVECLEN (PATTERN (p), 0);
6587 else if (!NOTE_P (p))
6588 nsets += 1;
6590 /* Ignore insns made by CSE; they cannot affect the boundaries of
6591 the basic block. */
6593 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6594 high_cuid = INSN_CUID (p);
6595 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6596 low_cuid = INSN_CUID (p);
6598 /* See if this insn is in our branch path. If it is and we are to
6599 take it, do so. */
6600 if (path_entry < path_size && data->path[path_entry].branch == p)
6602 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6603 p = JUMP_LABEL (p);
6605 /* Point to next entry in path, if any. */
6606 path_entry++;
6609 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6610 was specified, we haven't reached our maximum path length, there are
6611 insns following the target of the jump, this is the only use of the
6612 jump label, and the target label is preceded by a BARRIER.
6614 Alternatively, we can follow the jump if it branches around a
6615 block of code and there are no other branches into the block.
6616 In this case invalidate_skipped_block will be called to invalidate any
6617 registers set in the block when following the jump. */
6619 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6620 && JUMP_P (p)
6621 && GET_CODE (PATTERN (p)) == SET
6622 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6623 && JUMP_LABEL (p) != 0
6624 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6625 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6627 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6628 if ((!NOTE_P (q)
6629 || NOTE_LINE_NUMBER (q) == NOTE_INSN_LOOP_END
6630 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6631 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6632 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6633 break;
6635 /* If we ran into a BARRIER, this code is an extension of the
6636 basic block when the branch is taken. */
6637 if (follow_jumps && q != 0 && BARRIER_P (q))
6639 /* Don't allow ourself to keep walking around an
6640 always-executed loop. */
6641 if (next_real_insn (q) == next)
6643 p = NEXT_INSN (p);
6644 continue;
6647 /* Similarly, don't put a branch in our path more than once. */
6648 for (i = 0; i < path_entry; i++)
6649 if (data->path[i].branch == p)
6650 break;
6652 if (i != path_entry)
6653 break;
6655 data->path[path_entry].branch = p;
6656 data->path[path_entry++].status = PATH_TAKEN;
6658 /* This branch now ends our path. It was possible that we
6659 didn't see this branch the last time around (when the
6660 insn in front of the target was a JUMP_INSN that was
6661 turned into a no-op). */
6662 path_size = path_entry;
6664 p = JUMP_LABEL (p);
6665 /* Mark block so we won't scan it again later. */
6666 PUT_MODE (NEXT_INSN (p), QImode);
6668 /* Detect a branch around a block of code. */
6669 else if (skip_blocks && q != 0 && !LABEL_P (q))
6671 rtx tmp;
6673 if (next_real_insn (q) == next)
6675 p = NEXT_INSN (p);
6676 continue;
6679 for (i = 0; i < path_entry; i++)
6680 if (data->path[i].branch == p)
6681 break;
6683 if (i != path_entry)
6684 break;
6686 /* This is no_labels_between_p (p, q) with an added check for
6687 reaching the end of a function (in case Q precedes P). */
6688 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6689 if (LABEL_P (tmp))
6690 break;
6692 if (tmp == q)
6694 data->path[path_entry].branch = p;
6695 data->path[path_entry++].status = PATH_AROUND;
6697 path_size = path_entry;
6699 p = JUMP_LABEL (p);
6700 /* Mark block so we won't scan it again later. */
6701 PUT_MODE (NEXT_INSN (p), QImode);
6705 p = NEXT_INSN (p);
6708 data->low_cuid = low_cuid;
6709 data->high_cuid = high_cuid;
6710 data->nsets = nsets;
6711 data->last = p;
6713 /* If all jumps in the path are not taken, set our path length to zero
6714 so a rescan won't be done. */
6715 for (i = path_size - 1; i >= 0; i--)
6716 if (data->path[i].status != PATH_NOT_TAKEN)
6717 break;
6719 if (i == -1)
6720 data->path_size = 0;
6721 else
6722 data->path_size = path_size;
6724 /* End the current branch path. */
6725 data->path[path_size].branch = 0;
6728 /* Perform cse on the instructions of a function.
6729 F is the first instruction.
6730 NREGS is one plus the highest pseudo-reg number used in the instruction.
6732 Returns 1 if jump_optimize should be redone due to simplifications
6733 in conditional jump instructions. */
6736 cse_main (rtx f, int nregs, FILE *file)
6738 struct cse_basic_block_data val;
6739 rtx insn = f;
6740 int i;
6742 init_cse_reg_info (nregs);
6744 val.path = xmalloc (sizeof (struct branch_path)
6745 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6747 cse_jumps_altered = 0;
6748 recorded_label_ref = 0;
6749 constant_pool_entries_cost = 0;
6750 constant_pool_entries_regcost = 0;
6751 val.path_size = 0;
6752 rtl_hooks = cse_rtl_hooks;
6754 init_recog ();
6755 init_alias_analysis ();
6757 reg_eqv_table = xmalloc (nregs * sizeof (struct reg_eqv_elem));
6759 /* Find the largest uid. */
6761 max_uid = get_max_uid ();
6762 uid_cuid = xcalloc (max_uid + 1, sizeof (int));
6764 /* Compute the mapping from uids to cuids.
6765 CUIDs are numbers assigned to insns, like uids,
6766 except that cuids increase monotonically through the code.
6767 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6768 between two insns is not affected by -g. */
6770 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6772 if (!NOTE_P (insn)
6773 || NOTE_LINE_NUMBER (insn) < 0)
6774 INSN_CUID (insn) = ++i;
6775 else
6776 /* Give a line number note the same cuid as preceding insn. */
6777 INSN_CUID (insn) = i;
6780 /* Loop over basic blocks.
6781 Compute the maximum number of qty's needed for each basic block
6782 (which is 2 for each SET). */
6783 insn = f;
6784 while (insn)
6786 cse_altered = 0;
6787 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6788 flag_cse_skip_blocks);
6790 /* If this basic block was already processed or has no sets, skip it. */
6791 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6793 PUT_MODE (insn, VOIDmode);
6794 insn = (val.last ? NEXT_INSN (val.last) : 0);
6795 val.path_size = 0;
6796 continue;
6799 cse_basic_block_start = val.low_cuid;
6800 cse_basic_block_end = val.high_cuid;
6801 max_qty = val.nsets * 2;
6803 if (file)
6804 fprintf (file, ";; Processing block from %d to %d, %d sets.\n",
6805 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6806 val.nsets);
6808 /* Make MAX_QTY bigger to give us room to optimize
6809 past the end of this basic block, if that should prove useful. */
6810 if (max_qty < 500)
6811 max_qty = 500;
6813 /* If this basic block is being extended by following certain jumps,
6814 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6815 Otherwise, we start after this basic block. */
6816 if (val.path_size > 0)
6817 cse_basic_block (insn, val.last, val.path);
6818 else
6820 int old_cse_jumps_altered = cse_jumps_altered;
6821 rtx temp;
6823 /* When cse changes a conditional jump to an unconditional
6824 jump, we want to reprocess the block, since it will give
6825 us a new branch path to investigate. */
6826 cse_jumps_altered = 0;
6827 temp = cse_basic_block (insn, val.last, val.path);
6828 if (cse_jumps_altered == 0
6829 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6830 insn = temp;
6832 cse_jumps_altered |= old_cse_jumps_altered;
6835 if (cse_altered)
6836 ggc_collect ();
6838 #ifdef USE_C_ALLOCA
6839 alloca (0);
6840 #endif
6843 /* Clean up. */
6844 end_alias_analysis ();
6845 free (uid_cuid);
6846 free (reg_eqv_table);
6847 free (val.path);
6848 rtl_hooks = general_rtl_hooks;
6850 return cse_jumps_altered || recorded_label_ref;
6853 /* Process a single basic block. FROM and TO and the limits of the basic
6854 block. NEXT_BRANCH points to the branch path when following jumps or
6855 a null path when not following jumps. */
6857 static rtx
6858 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
6860 rtx insn;
6861 int to_usage = 0;
6862 rtx libcall_insn = NULL_RTX;
6863 int num_insns = 0;
6864 int no_conflict = 0;
6866 /* Allocate the space needed by qty_table. */
6867 qty_table = xmalloc (max_qty * sizeof (struct qty_table_elem));
6869 new_basic_block ();
6871 /* TO might be a label. If so, protect it from being deleted. */
6872 if (to != 0 && LABEL_P (to))
6873 ++LABEL_NUSES (to);
6875 for (insn = from; insn != to; insn = NEXT_INSN (insn))
6877 enum rtx_code code = GET_CODE (insn);
6879 /* If we have processed 1,000 insns, flush the hash table to
6880 avoid extreme quadratic behavior. We must not include NOTEs
6881 in the count since there may be more of them when generating
6882 debugging information. If we clear the table at different
6883 times, code generated with -g -O might be different than code
6884 generated with -O but not -g.
6886 ??? This is a real kludge and needs to be done some other way.
6887 Perhaps for 2.9. */
6888 if (code != NOTE && num_insns++ > 1000)
6890 flush_hash_table ();
6891 num_insns = 0;
6894 /* See if this is a branch that is part of the path. If so, and it is
6895 to be taken, do so. */
6896 if (next_branch->branch == insn)
6898 enum taken status = next_branch++->status;
6899 if (status != PATH_NOT_TAKEN)
6901 if (status == PATH_TAKEN)
6902 record_jump_equiv (insn, 1);
6903 else
6904 invalidate_skipped_block (NEXT_INSN (insn));
6906 /* Set the last insn as the jump insn; it doesn't affect cc0.
6907 Then follow this branch. */
6908 #ifdef HAVE_cc0
6909 prev_insn_cc0 = 0;
6910 prev_insn = insn;
6911 #endif
6912 insn = JUMP_LABEL (insn);
6913 continue;
6917 if (GET_MODE (insn) == QImode)
6918 PUT_MODE (insn, VOIDmode);
6920 if (GET_RTX_CLASS (code) == RTX_INSN)
6922 rtx p;
6924 /* Process notes first so we have all notes in canonical forms when
6925 looking for duplicate operations. */
6927 if (REG_NOTES (insn))
6928 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
6930 /* Track when we are inside in LIBCALL block. Inside such a block,
6931 we do not want to record destinations. The last insn of a
6932 LIBCALL block is not considered to be part of the block, since
6933 its destination is the result of the block and hence should be
6934 recorded. */
6936 if (REG_NOTES (insn) != 0)
6938 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
6939 libcall_insn = XEXP (p, 0);
6940 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
6942 /* Keep libcall_insn for the last SET insn of a no-conflict
6943 block to prevent changing the destination. */
6944 if (! no_conflict)
6945 libcall_insn = 0;
6946 else
6947 no_conflict = -1;
6949 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
6950 no_conflict = 1;
6953 cse_insn (insn, libcall_insn);
6955 if (no_conflict == -1)
6957 libcall_insn = 0;
6958 no_conflict = 0;
6961 /* If we haven't already found an insn where we added a LABEL_REF,
6962 check this one. */
6963 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
6964 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6965 (void *) insn))
6966 recorded_label_ref = 1;
6969 /* If INSN is now an unconditional jump, skip to the end of our
6970 basic block by pretending that we just did the last insn in the
6971 basic block. If we are jumping to the end of our block, show
6972 that we can have one usage of TO. */
6974 if (any_uncondjump_p (insn))
6976 if (to == 0)
6978 free (qty_table);
6979 return 0;
6982 if (JUMP_LABEL (insn) == to)
6983 to_usage = 1;
6985 /* Maybe TO was deleted because the jump is unconditional.
6986 If so, there is nothing left in this basic block. */
6987 /* ??? Perhaps it would be smarter to set TO
6988 to whatever follows this insn,
6989 and pretend the basic block had always ended here. */
6990 if (INSN_DELETED_P (to))
6991 break;
6993 insn = PREV_INSN (to);
6996 /* See if it is ok to keep on going past the label
6997 which used to end our basic block. Remember that we incremented
6998 the count of that label, so we decrement it here. If we made
6999 a jump unconditional, TO_USAGE will be one; in that case, we don't
7000 want to count the use in that jump. */
7002 if (to != 0 && NEXT_INSN (insn) == to
7003 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7005 struct cse_basic_block_data val;
7006 rtx prev;
7008 insn = NEXT_INSN (to);
7010 /* If TO was the last insn in the function, we are done. */
7011 if (insn == 0)
7013 free (qty_table);
7014 return 0;
7017 /* If TO was preceded by a BARRIER we are done with this block
7018 because it has no continuation. */
7019 prev = prev_nonnote_insn (to);
7020 if (prev && BARRIER_P (prev))
7022 free (qty_table);
7023 return insn;
7026 /* Find the end of the following block. Note that we won't be
7027 following branches in this case. */
7028 to_usage = 0;
7029 val.path_size = 0;
7030 val.path = xmalloc (sizeof (struct branch_path)
7031 * PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7032 cse_end_of_basic_block (insn, &val, 0, 0);
7033 free (val.path);
7035 /* If the tables we allocated have enough space left
7036 to handle all the SETs in the next basic block,
7037 continue through it. Otherwise, return,
7038 and that block will be scanned individually. */
7039 if (val.nsets * 2 + next_qty > max_qty)
7040 break;
7042 cse_basic_block_start = val.low_cuid;
7043 cse_basic_block_end = val.high_cuid;
7044 to = val.last;
7046 /* Prevent TO from being deleted if it is a label. */
7047 if (to != 0 && LABEL_P (to))
7048 ++LABEL_NUSES (to);
7050 /* Back up so we process the first insn in the extension. */
7051 insn = PREV_INSN (insn);
7055 gcc_assert (next_qty <= max_qty);
7057 free (qty_table);
7059 return to ? NEXT_INSN (to) : 0;
7062 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7063 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7065 static int
7066 check_for_label_ref (rtx *rtl, void *data)
7068 rtx insn = (rtx) data;
7070 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7071 we must rerun jump since it needs to place the note. If this is a
7072 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7073 since no REG_LABEL will be added. */
7074 return (GET_CODE (*rtl) == LABEL_REF
7075 && ! LABEL_REF_NONLOCAL_P (*rtl)
7076 && LABEL_P (XEXP (*rtl, 0))
7077 && INSN_UID (XEXP (*rtl, 0)) != 0
7078 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7081 /* Count the number of times registers are used (not set) in X.
7082 COUNTS is an array in which we accumulate the count, INCR is how much
7083 we count each register usage.
7085 Don't count a usage of DEST, which is the SET_DEST of a SET which
7086 contains X in its SET_SRC. This is because such a SET does not
7087 modify the liveness of DEST.
7088 DEST is set to pc_rtx for a trapping insn, which means that we must count
7089 uses of a SET_DEST regardless because the insn can't be deleted here. */
7091 static void
7092 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
7094 enum rtx_code code;
7095 rtx note;
7096 const char *fmt;
7097 int i, j;
7099 if (x == 0)
7100 return;
7102 switch (code = GET_CODE (x))
7104 case REG:
7105 if (x != dest)
7106 counts[REGNO (x)] += incr;
7107 return;
7109 case PC:
7110 case CC0:
7111 case CONST:
7112 case CONST_INT:
7113 case CONST_DOUBLE:
7114 case CONST_VECTOR:
7115 case SYMBOL_REF:
7116 case LABEL_REF:
7117 return;
7119 case CLOBBER:
7120 /* If we are clobbering a MEM, mark any registers inside the address
7121 as being used. */
7122 if (MEM_P (XEXP (x, 0)))
7123 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7124 return;
7126 case SET:
7127 /* Unless we are setting a REG, count everything in SET_DEST. */
7128 if (!REG_P (SET_DEST (x)))
7129 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7130 count_reg_usage (SET_SRC (x), counts,
7131 dest ? dest : SET_DEST (x),
7132 incr);
7133 return;
7135 case CALL_INSN:
7136 case INSN:
7137 case JUMP_INSN:
7138 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
7139 this fact by setting DEST to pc_rtx. */
7140 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
7141 dest = pc_rtx;
7142 if (code == CALL_INSN)
7143 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
7144 count_reg_usage (PATTERN (x), counts, dest, incr);
7146 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7147 use them. */
7149 note = find_reg_equal_equiv_note (x);
7150 if (note)
7152 rtx eqv = XEXP (note, 0);
7154 if (GET_CODE (eqv) == EXPR_LIST)
7155 /* This REG_EQUAL note describes the result of a function call.
7156 Process all the arguments. */
7159 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
7160 eqv = XEXP (eqv, 1);
7162 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7163 else
7164 count_reg_usage (eqv, counts, dest, incr);
7166 return;
7168 case EXPR_LIST:
7169 if (REG_NOTE_KIND (x) == REG_EQUAL
7170 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7171 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7172 involving registers in the address. */
7173 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7174 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7176 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7177 return;
7179 case ASM_OPERANDS:
7180 /* If the asm is volatile, then this insn cannot be deleted,
7181 and so the inputs *must* be live. */
7182 if (MEM_VOLATILE_P (x))
7183 dest = NULL_RTX;
7184 /* Iterate over just the inputs, not the constraints as well. */
7185 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7186 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
7187 return;
7189 case INSN_LIST:
7190 gcc_unreachable ();
7192 default:
7193 break;
7196 fmt = GET_RTX_FORMAT (code);
7197 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7199 if (fmt[i] == 'e')
7200 count_reg_usage (XEXP (x, i), counts, dest, incr);
7201 else if (fmt[i] == 'E')
7202 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7203 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7207 /* Return true if set is live. */
7208 static bool
7209 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7210 int *counts)
7212 #ifdef HAVE_cc0
7213 rtx tem;
7214 #endif
7216 if (set_noop_p (set))
7219 #ifdef HAVE_cc0
7220 else if (GET_CODE (SET_DEST (set)) == CC0
7221 && !side_effects_p (SET_SRC (set))
7222 && ((tem = next_nonnote_insn (insn)) == 0
7223 || !INSN_P (tem)
7224 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7225 return false;
7226 #endif
7227 else if (!REG_P (SET_DEST (set))
7228 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7229 || counts[REGNO (SET_DEST (set))] != 0
7230 || side_effects_p (SET_SRC (set)))
7231 return true;
7232 return false;
7235 /* Return true if insn is live. */
7237 static bool
7238 insn_live_p (rtx insn, int *counts)
7240 int i;
7241 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7242 return true;
7243 else if (GET_CODE (PATTERN (insn)) == SET)
7244 return set_live_p (PATTERN (insn), insn, counts);
7245 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7247 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7249 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7251 if (GET_CODE (elt) == SET)
7253 if (set_live_p (elt, insn, counts))
7254 return true;
7256 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7257 return true;
7259 return false;
7261 else
7262 return true;
7265 /* Return true if libcall is dead as a whole. */
7267 static bool
7268 dead_libcall_p (rtx insn, int *counts)
7270 rtx note, set, new;
7272 /* See if there's a REG_EQUAL note on this insn and try to
7273 replace the source with the REG_EQUAL expression.
7275 We assume that insns with REG_RETVALs can only be reg->reg
7276 copies at this point. */
7277 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7278 if (!note)
7279 return false;
7281 set = single_set (insn);
7282 if (!set)
7283 return false;
7285 new = simplify_rtx (XEXP (note, 0));
7286 if (!new)
7287 new = XEXP (note, 0);
7289 /* While changing insn, we must update the counts accordingly. */
7290 count_reg_usage (insn, counts, NULL_RTX, -1);
7292 if (validate_change (insn, &SET_SRC (set), new, 0))
7294 count_reg_usage (insn, counts, NULL_RTX, 1);
7295 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7296 remove_note (insn, note);
7297 return true;
7300 if (CONSTANT_P (new))
7302 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7303 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7305 count_reg_usage (insn, counts, NULL_RTX, 1);
7306 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7307 remove_note (insn, note);
7308 return true;
7312 count_reg_usage (insn, counts, NULL_RTX, 1);
7313 return false;
7316 /* Scan all the insns and delete any that are dead; i.e., they store a register
7317 that is never used or they copy a register to itself.
7319 This is used to remove insns made obviously dead by cse, loop or other
7320 optimizations. It improves the heuristics in loop since it won't try to
7321 move dead invariants out of loops or make givs for dead quantities. The
7322 remaining passes of the compilation are also sped up. */
7325 delete_trivially_dead_insns (rtx insns, int nreg)
7327 int *counts;
7328 rtx insn, prev;
7329 int in_libcall = 0, dead_libcall = 0;
7330 int ndead = 0;
7332 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7333 /* First count the number of times each register is used. */
7334 counts = xcalloc (nreg, sizeof (int));
7335 for (insn = insns; insn; insn = NEXT_INSN (insn))
7336 if (INSN_P (insn))
7337 count_reg_usage (insn, counts, NULL_RTX, 1);
7339 /* Go from the last insn to the first and delete insns that only set unused
7340 registers or copy a register to itself. As we delete an insn, remove
7341 usage counts for registers it uses.
7343 The first jump optimization pass may leave a real insn as the last
7344 insn in the function. We must not skip that insn or we may end
7345 up deleting code that is not really dead. */
7346 for (insn = get_last_insn (); insn; insn = prev)
7348 int live_insn = 0;
7350 prev = PREV_INSN (insn);
7351 if (!INSN_P (insn))
7352 continue;
7354 /* Don't delete any insns that are part of a libcall block unless
7355 we can delete the whole libcall block.
7357 Flow or loop might get confused if we did that. Remember
7358 that we are scanning backwards. */
7359 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7361 in_libcall = 1;
7362 live_insn = 1;
7363 dead_libcall = dead_libcall_p (insn, counts);
7365 else if (in_libcall)
7366 live_insn = ! dead_libcall;
7367 else
7368 live_insn = insn_live_p (insn, counts);
7370 /* If this is a dead insn, delete it and show registers in it aren't
7371 being used. */
7373 if (! live_insn)
7375 count_reg_usage (insn, counts, NULL_RTX, -1);
7376 delete_insn_and_edges (insn);
7377 ndead++;
7380 if (in_libcall && find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7382 in_libcall = 0;
7383 dead_libcall = 0;
7387 if (dump_file && ndead)
7388 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7389 ndead);
7390 /* Clean up. */
7391 free (counts);
7392 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7393 return ndead;
7396 /* This function is called via for_each_rtx. The argument, NEWREG, is
7397 a condition code register with the desired mode. If we are looking
7398 at the same register in a different mode, replace it with
7399 NEWREG. */
7401 static int
7402 cse_change_cc_mode (rtx *loc, void *data)
7404 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7406 if (*loc
7407 && REG_P (*loc)
7408 && REGNO (*loc) == REGNO (args->newreg)
7409 && GET_MODE (*loc) != GET_MODE (args->newreg))
7411 validate_change (args->insn, loc, args->newreg, 1);
7413 return -1;
7415 return 0;
7418 /* Change the mode of any reference to the register REGNO (NEWREG) to
7419 GET_MODE (NEWREG) in INSN. */
7421 static void
7422 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7424 struct change_cc_mode_args args;
7425 int success;
7427 if (!INSN_P (insn))
7428 return;
7430 args.insn = insn;
7431 args.newreg = newreg;
7433 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7434 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7436 /* If the following assertion was triggered, there is most probably
7437 something wrong with the cc_modes_compatible back end function.
7438 CC modes only can be considered compatible if the insn - with the mode
7439 replaced by any of the compatible modes - can still be recognized. */
7440 success = apply_change_group ();
7441 gcc_assert (success);
7444 /* Change the mode of any reference to the register REGNO (NEWREG) to
7445 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7446 any instruction which modifies NEWREG. */
7448 static void
7449 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7451 rtx insn;
7453 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7455 if (! INSN_P (insn))
7456 continue;
7458 if (reg_set_p (newreg, insn))
7459 return;
7461 cse_change_cc_mode_insn (insn, newreg);
7465 /* BB is a basic block which finishes with CC_REG as a condition code
7466 register which is set to CC_SRC. Look through the successors of BB
7467 to find blocks which have a single predecessor (i.e., this one),
7468 and look through those blocks for an assignment to CC_REG which is
7469 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7470 permitted to change the mode of CC_SRC to a compatible mode. This
7471 returns VOIDmode if no equivalent assignments were found.
7472 Otherwise it returns the mode which CC_SRC should wind up with.
7474 The main complexity in this function is handling the mode issues.
7475 We may have more than one duplicate which we can eliminate, and we
7476 try to find a mode which will work for multiple duplicates. */
7478 static enum machine_mode
7479 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7481 bool found_equiv;
7482 enum machine_mode mode;
7483 unsigned int insn_count;
7484 edge e;
7485 rtx insns[2];
7486 enum machine_mode modes[2];
7487 rtx last_insns[2];
7488 unsigned int i;
7489 rtx newreg;
7490 edge_iterator ei;
7492 /* We expect to have two successors. Look at both before picking
7493 the final mode for the comparison. If we have more successors
7494 (i.e., some sort of table jump, although that seems unlikely),
7495 then we require all beyond the first two to use the same
7496 mode. */
7498 found_equiv = false;
7499 mode = GET_MODE (cc_src);
7500 insn_count = 0;
7501 FOR_EACH_EDGE (e, ei, bb->succs)
7503 rtx insn;
7504 rtx end;
7506 if (e->flags & EDGE_COMPLEX)
7507 continue;
7509 if (EDGE_COUNT (e->dest->preds) != 1
7510 || e->dest == EXIT_BLOCK_PTR)
7511 continue;
7513 end = NEXT_INSN (BB_END (e->dest));
7514 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7516 rtx set;
7518 if (! INSN_P (insn))
7519 continue;
7521 /* If CC_SRC is modified, we have to stop looking for
7522 something which uses it. */
7523 if (modified_in_p (cc_src, insn))
7524 break;
7526 /* Check whether INSN sets CC_REG to CC_SRC. */
7527 set = single_set (insn);
7528 if (set
7529 && REG_P (SET_DEST (set))
7530 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7532 bool found;
7533 enum machine_mode set_mode;
7534 enum machine_mode comp_mode;
7536 found = false;
7537 set_mode = GET_MODE (SET_SRC (set));
7538 comp_mode = set_mode;
7539 if (rtx_equal_p (cc_src, SET_SRC (set)))
7540 found = true;
7541 else if (GET_CODE (cc_src) == COMPARE
7542 && GET_CODE (SET_SRC (set)) == COMPARE
7543 && mode != set_mode
7544 && rtx_equal_p (XEXP (cc_src, 0),
7545 XEXP (SET_SRC (set), 0))
7546 && rtx_equal_p (XEXP (cc_src, 1),
7547 XEXP (SET_SRC (set), 1)))
7550 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7551 if (comp_mode != VOIDmode
7552 && (can_change_mode || comp_mode == mode))
7553 found = true;
7556 if (found)
7558 found_equiv = true;
7559 if (insn_count < ARRAY_SIZE (insns))
7561 insns[insn_count] = insn;
7562 modes[insn_count] = set_mode;
7563 last_insns[insn_count] = end;
7564 ++insn_count;
7566 if (mode != comp_mode)
7568 gcc_assert (can_change_mode);
7569 mode = comp_mode;
7571 /* The modified insn will be re-recognized later. */
7572 PUT_MODE (cc_src, mode);
7575 else
7577 if (set_mode != mode)
7579 /* We found a matching expression in the
7580 wrong mode, but we don't have room to
7581 store it in the array. Punt. This case
7582 should be rare. */
7583 break;
7585 /* INSN sets CC_REG to a value equal to CC_SRC
7586 with the right mode. We can simply delete
7587 it. */
7588 delete_insn (insn);
7591 /* We found an instruction to delete. Keep looking,
7592 in the hopes of finding a three-way jump. */
7593 continue;
7596 /* We found an instruction which sets the condition
7597 code, so don't look any farther. */
7598 break;
7601 /* If INSN sets CC_REG in some other way, don't look any
7602 farther. */
7603 if (reg_set_p (cc_reg, insn))
7604 break;
7607 /* If we fell off the bottom of the block, we can keep looking
7608 through successors. We pass CAN_CHANGE_MODE as false because
7609 we aren't prepared to handle compatibility between the
7610 further blocks and this block. */
7611 if (insn == end)
7613 enum machine_mode submode;
7615 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7616 if (submode != VOIDmode)
7618 gcc_assert (submode == mode);
7619 found_equiv = true;
7620 can_change_mode = false;
7625 if (! found_equiv)
7626 return VOIDmode;
7628 /* Now INSN_COUNT is the number of instructions we found which set
7629 CC_REG to a value equivalent to CC_SRC. The instructions are in
7630 INSNS. The modes used by those instructions are in MODES. */
7632 newreg = NULL_RTX;
7633 for (i = 0; i < insn_count; ++i)
7635 if (modes[i] != mode)
7637 /* We need to change the mode of CC_REG in INSNS[i] and
7638 subsequent instructions. */
7639 if (! newreg)
7641 if (GET_MODE (cc_reg) == mode)
7642 newreg = cc_reg;
7643 else
7644 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7646 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7647 newreg);
7650 delete_insn (insns[i]);
7653 return mode;
7656 /* If we have a fixed condition code register (or two), walk through
7657 the instructions and try to eliminate duplicate assignments. */
7659 void
7660 cse_condition_code_reg (void)
7662 unsigned int cc_regno_1;
7663 unsigned int cc_regno_2;
7664 rtx cc_reg_1;
7665 rtx cc_reg_2;
7666 basic_block bb;
7668 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7669 return;
7671 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7672 if (cc_regno_2 != INVALID_REGNUM)
7673 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7674 else
7675 cc_reg_2 = NULL_RTX;
7677 FOR_EACH_BB (bb)
7679 rtx last_insn;
7680 rtx cc_reg;
7681 rtx insn;
7682 rtx cc_src_insn;
7683 rtx cc_src;
7684 enum machine_mode mode;
7685 enum machine_mode orig_mode;
7687 /* Look for blocks which end with a conditional jump based on a
7688 condition code register. Then look for the instruction which
7689 sets the condition code register. Then look through the
7690 successor blocks for instructions which set the condition
7691 code register to the same value. There are other possible
7692 uses of the condition code register, but these are by far the
7693 most common and the ones which we are most likely to be able
7694 to optimize. */
7696 last_insn = BB_END (bb);
7697 if (!JUMP_P (last_insn))
7698 continue;
7700 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7701 cc_reg = cc_reg_1;
7702 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7703 cc_reg = cc_reg_2;
7704 else
7705 continue;
7707 cc_src_insn = NULL_RTX;
7708 cc_src = NULL_RTX;
7709 for (insn = PREV_INSN (last_insn);
7710 insn && insn != PREV_INSN (BB_HEAD (bb));
7711 insn = PREV_INSN (insn))
7713 rtx set;
7715 if (! INSN_P (insn))
7716 continue;
7717 set = single_set (insn);
7718 if (set
7719 && REG_P (SET_DEST (set))
7720 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7722 cc_src_insn = insn;
7723 cc_src = SET_SRC (set);
7724 break;
7726 else if (reg_set_p (cc_reg, insn))
7727 break;
7730 if (! cc_src_insn)
7731 continue;
7733 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7734 continue;
7736 /* Now CC_REG is a condition code register used for a
7737 conditional jump at the end of the block, and CC_SRC, in
7738 CC_SRC_INSN, is the value to which that condition code
7739 register is set, and CC_SRC is still meaningful at the end of
7740 the basic block. */
7742 orig_mode = GET_MODE (cc_src);
7743 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7744 if (mode != VOIDmode)
7746 gcc_assert (mode == GET_MODE (cc_src));
7747 if (mode != orig_mode)
7749 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7751 cse_change_cc_mode_insn (cc_src_insn, newreg);
7753 /* Do the same in the following insns that use the
7754 current value of CC_REG within BB. */
7755 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7756 NEXT_INSN (last_insn),
7757 newreg);
7764 /* Perform common subexpression elimination. Nonzero value from
7765 `cse_main' means that jumps were simplified and some code may now
7766 be unreachable, so do jump optimization again. */
7767 static bool
7768 gate_handle_cse (void)
7770 return optimize > 0;
7773 static void
7774 rest_of_handle_cse (void)
7776 int tem;
7778 if (dump_file)
7779 dump_flow_info (dump_file);
7781 reg_scan (get_insns (), max_reg_num ());
7783 tem = cse_main (get_insns (), max_reg_num (), dump_file);
7784 if (tem)
7785 rebuild_jump_labels (get_insns ());
7786 if (purge_all_dead_edges ())
7787 delete_unreachable_blocks ();
7789 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7791 /* If we are not running more CSE passes, then we are no longer
7792 expecting CSE to be run. But always rerun it in a cheap mode. */
7793 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7795 if (tem)
7796 delete_dead_jumptables ();
7798 if (tem || optimize > 1)
7799 cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_PRE_LOOP);
7802 struct tree_opt_pass pass_cse =
7804 "cse1", /* name */
7805 gate_handle_cse, /* gate */
7806 rest_of_handle_cse, /* execute */
7807 NULL, /* sub */
7808 NULL, /* next */
7809 0, /* static_pass_number */
7810 TV_CSE, /* tv_id */
7811 0, /* properties_required */
7812 0, /* properties_provided */
7813 0, /* properties_destroyed */
7814 0, /* todo_flags_start */
7815 TODO_dump_func |
7816 TODO_ggc_collect, /* todo_flags_finish */
7817 's' /* letter */
7821 static bool
7822 gate_handle_cse2 (void)
7824 return optimize > 0 && flag_rerun_cse_after_loop;
7827 /* Run second CSE pass after loop optimizations. */
7828 static void
7829 rest_of_handle_cse2 (void)
7831 int tem;
7833 if (dump_file)
7834 dump_flow_info (dump_file);
7836 tem = cse_main (get_insns (), max_reg_num (), dump_file);
7838 /* Run a pass to eliminate duplicated assignments to condition code
7839 registers. We have to run this after bypass_jumps, because it
7840 makes it harder for that pass to determine whether a jump can be
7841 bypassed safely. */
7842 cse_condition_code_reg ();
7844 purge_all_dead_edges ();
7845 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7847 if (tem)
7849 timevar_push (TV_JUMP);
7850 rebuild_jump_labels (get_insns ());
7851 delete_dead_jumptables ();
7852 cleanup_cfg (CLEANUP_EXPENSIVE);
7853 timevar_pop (TV_JUMP);
7855 reg_scan (get_insns (), max_reg_num ());
7856 cse_not_expected = 1;
7860 struct tree_opt_pass pass_cse2 =
7862 "cse2", /* name */
7863 gate_handle_cse2, /* gate */
7864 rest_of_handle_cse2, /* execute */
7865 NULL, /* sub */
7866 NULL, /* next */
7867 0, /* static_pass_number */
7868 TV_CSE2, /* tv_id */
7869 0, /* properties_required */
7870 0, /* properties_provided */
7871 0, /* properties_destroyed */
7872 0, /* todo_flags_start */
7873 TODO_dump_func |
7874 TODO_ggc_collect, /* todo_flags_finish */
7875 't' /* letter */