2006-08-06 Paolo Carlini <pcarlini@suse.de>
[official-gcc.git] / gcc / cse.c
blobfc15511dbc8a4ab0325e12161f6ec52095715e9f
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 = XNEWVEC (struct cse_reg_info, new_size);
871 cse_reg_info_table_size = new_size;
872 cse_reg_info_table_first_uninitialized = 0;
875 /* Do we have all of the first NREGS entries initialized? */
876 if (cse_reg_info_table_first_uninitialized < nregs)
878 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
879 unsigned int i;
881 /* Put the old timestamp on newly allocated entries so that they
882 will all be considered out of date. We do not touch those
883 entries beyond the first NREGS entries to be nice to the
884 virtual memory. */
885 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
886 cse_reg_info_table[i].timestamp = old_timestamp;
888 cse_reg_info_table_first_uninitialized = nregs;
892 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
894 static void
895 get_cse_reg_info_1 (unsigned int regno)
897 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
898 entry will be considered to have been initialized. */
899 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
901 /* Initialize the rest of the entry. */
902 cse_reg_info_table[regno].reg_tick = 1;
903 cse_reg_info_table[regno].reg_in_table = -1;
904 cse_reg_info_table[regno].subreg_ticked = -1;
905 cse_reg_info_table[regno].reg_qty = -regno - 1;
908 /* Find a cse_reg_info entry for REGNO. */
910 static inline struct cse_reg_info *
911 get_cse_reg_info (unsigned int regno)
913 struct cse_reg_info *p = &cse_reg_info_table[regno];
915 /* If this entry has not been initialized, go ahead and initialize
916 it. */
917 if (p->timestamp != cse_reg_info_timestamp)
918 get_cse_reg_info_1 (regno);
920 return p;
923 /* Clear the hash table and initialize each register with its own quantity,
924 for a new basic block. */
926 static void
927 new_basic_block (void)
929 int i;
931 next_qty = 0;
933 /* Invalidate cse_reg_info_table. */
934 cse_reg_info_timestamp++;
936 /* Clear out hash table state for this pass. */
937 CLEAR_HARD_REG_SET (hard_regs_in_table);
939 /* The per-quantity values used to be initialized here, but it is
940 much faster to initialize each as it is made in `make_new_qty'. */
942 for (i = 0; i < HASH_SIZE; i++)
944 struct table_elt *first;
946 first = table[i];
947 if (first != NULL)
949 struct table_elt *last = first;
951 table[i] = NULL;
953 while (last->next_same_hash != NULL)
954 last = last->next_same_hash;
956 /* Now relink this hash entire chain into
957 the free element list. */
959 last->next_same_hash = free_element_chain;
960 free_element_chain = first;
964 #ifdef HAVE_cc0
965 prev_insn = 0;
966 prev_insn_cc0 = 0;
967 #endif
970 /* Say that register REG contains a quantity in mode MODE not in any
971 register before and initialize that quantity. */
973 static void
974 make_new_qty (unsigned int reg, enum machine_mode mode)
976 int q;
977 struct qty_table_elem *ent;
978 struct reg_eqv_elem *eqv;
980 gcc_assert (next_qty < max_qty);
982 q = REG_QTY (reg) = next_qty++;
983 ent = &qty_table[q];
984 ent->first_reg = reg;
985 ent->last_reg = reg;
986 ent->mode = mode;
987 ent->const_rtx = ent->const_insn = NULL_RTX;
988 ent->comparison_code = UNKNOWN;
990 eqv = &reg_eqv_table[reg];
991 eqv->next = eqv->prev = -1;
994 /* Make reg NEW equivalent to reg OLD.
995 OLD is not changing; NEW is. */
997 static void
998 make_regs_eqv (unsigned int new, unsigned int old)
1000 unsigned int lastr, firstr;
1001 int q = REG_QTY (old);
1002 struct qty_table_elem *ent;
1004 ent = &qty_table[q];
1006 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1007 gcc_assert (REGNO_QTY_VALID_P (old));
1009 REG_QTY (new) = q;
1010 firstr = ent->first_reg;
1011 lastr = ent->last_reg;
1013 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1014 hard regs. Among pseudos, if NEW will live longer than any other reg
1015 of the same qty, and that is beyond the current basic block,
1016 make it the new canonical replacement for this qty. */
1017 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1018 /* Certain fixed registers might be of the class NO_REGS. This means
1019 that not only can they not be allocated by the compiler, but
1020 they cannot be used in substitutions or canonicalizations
1021 either. */
1022 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1023 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1024 || (new >= FIRST_PSEUDO_REGISTER
1025 && (firstr < FIRST_PSEUDO_REGISTER
1026 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1027 || (uid_cuid[REGNO_FIRST_UID (new)]
1028 < cse_basic_block_start))
1029 && (uid_cuid[REGNO_LAST_UID (new)]
1030 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1032 reg_eqv_table[firstr].prev = new;
1033 reg_eqv_table[new].next = firstr;
1034 reg_eqv_table[new].prev = -1;
1035 ent->first_reg = new;
1037 else
1039 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1040 Otherwise, insert before any non-fixed hard regs that are at the
1041 end. Registers of class NO_REGS cannot be used as an
1042 equivalent for anything. */
1043 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1044 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1045 && new >= FIRST_PSEUDO_REGISTER)
1046 lastr = reg_eqv_table[lastr].prev;
1047 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1048 if (reg_eqv_table[lastr].next >= 0)
1049 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1050 else
1051 qty_table[q].last_reg = new;
1052 reg_eqv_table[lastr].next = new;
1053 reg_eqv_table[new].prev = lastr;
1057 /* Remove REG from its equivalence class. */
1059 static void
1060 delete_reg_equiv (unsigned int reg)
1062 struct qty_table_elem *ent;
1063 int q = REG_QTY (reg);
1064 int p, n;
1066 /* If invalid, do nothing. */
1067 if (! REGNO_QTY_VALID_P (reg))
1068 return;
1070 ent = &qty_table[q];
1072 p = reg_eqv_table[reg].prev;
1073 n = reg_eqv_table[reg].next;
1075 if (n != -1)
1076 reg_eqv_table[n].prev = p;
1077 else
1078 ent->last_reg = p;
1079 if (p != -1)
1080 reg_eqv_table[p].next = n;
1081 else
1082 ent->first_reg = n;
1084 REG_QTY (reg) = -reg - 1;
1087 /* Remove any invalid expressions from the hash table
1088 that refer to any of the registers contained in expression X.
1090 Make sure that newly inserted references to those registers
1091 as subexpressions will be considered valid.
1093 mention_regs is not called when a register itself
1094 is being stored in the table.
1096 Return 1 if we have done something that may have changed the hash code
1097 of X. */
1099 static int
1100 mention_regs (rtx x)
1102 enum rtx_code code;
1103 int i, j;
1104 const char *fmt;
1105 int changed = 0;
1107 if (x == 0)
1108 return 0;
1110 code = GET_CODE (x);
1111 if (code == REG)
1113 unsigned int regno = REGNO (x);
1114 unsigned int endregno
1115 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1116 : hard_regno_nregs[regno][GET_MODE (x)]);
1117 unsigned int i;
1119 for (i = regno; i < endregno; i++)
1121 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1122 remove_invalid_refs (i);
1124 REG_IN_TABLE (i) = REG_TICK (i);
1125 SUBREG_TICKED (i) = -1;
1128 return 0;
1131 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1132 pseudo if they don't use overlapping words. We handle only pseudos
1133 here for simplicity. */
1134 if (code == SUBREG && REG_P (SUBREG_REG (x))
1135 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1137 unsigned int i = REGNO (SUBREG_REG (x));
1139 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1141 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1142 the last store to this register really stored into this
1143 subreg, then remove the memory of this subreg.
1144 Otherwise, remove any memory of the entire register and
1145 all its subregs from the table. */
1146 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1147 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1148 remove_invalid_refs (i);
1149 else
1150 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1153 REG_IN_TABLE (i) = REG_TICK (i);
1154 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1155 return 0;
1158 /* If X is a comparison or a COMPARE and either operand is a register
1159 that does not have a quantity, give it one. This is so that a later
1160 call to record_jump_equiv won't cause X to be assigned a different
1161 hash code and not found in the table after that call.
1163 It is not necessary to do this here, since rehash_using_reg can
1164 fix up the table later, but doing this here eliminates the need to
1165 call that expensive function in the most common case where the only
1166 use of the register is in the comparison. */
1168 if (code == COMPARE || COMPARISON_P (x))
1170 if (REG_P (XEXP (x, 0))
1171 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1172 if (insert_regs (XEXP (x, 0), NULL, 0))
1174 rehash_using_reg (XEXP (x, 0));
1175 changed = 1;
1178 if (REG_P (XEXP (x, 1))
1179 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1180 if (insert_regs (XEXP (x, 1), NULL, 0))
1182 rehash_using_reg (XEXP (x, 1));
1183 changed = 1;
1187 fmt = GET_RTX_FORMAT (code);
1188 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1189 if (fmt[i] == 'e')
1190 changed |= mention_regs (XEXP (x, i));
1191 else if (fmt[i] == 'E')
1192 for (j = 0; j < XVECLEN (x, i); j++)
1193 changed |= mention_regs (XVECEXP (x, i, j));
1195 return changed;
1198 /* Update the register quantities for inserting X into the hash table
1199 with a value equivalent to CLASSP.
1200 (If the class does not contain a REG, it is irrelevant.)
1201 If MODIFIED is nonzero, X is a destination; it is being modified.
1202 Note that delete_reg_equiv should be called on a register
1203 before insert_regs is done on that register with MODIFIED != 0.
1205 Nonzero value means that elements of reg_qty have changed
1206 so X's hash code may be different. */
1208 static int
1209 insert_regs (rtx x, struct table_elt *classp, int modified)
1211 if (REG_P (x))
1213 unsigned int regno = REGNO (x);
1214 int qty_valid;
1216 /* If REGNO is in the equivalence table already but is of the
1217 wrong mode for that equivalence, don't do anything here. */
1219 qty_valid = REGNO_QTY_VALID_P (regno);
1220 if (qty_valid)
1222 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1224 if (ent->mode != GET_MODE (x))
1225 return 0;
1228 if (modified || ! qty_valid)
1230 if (classp)
1231 for (classp = classp->first_same_value;
1232 classp != 0;
1233 classp = classp->next_same_value)
1234 if (REG_P (classp->exp)
1235 && GET_MODE (classp->exp) == GET_MODE (x))
1237 unsigned c_regno = REGNO (classp->exp);
1239 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1241 /* Suppose that 5 is hard reg and 100 and 101 are
1242 pseudos. Consider
1244 (set (reg:si 100) (reg:si 5))
1245 (set (reg:si 5) (reg:si 100))
1246 (set (reg:di 101) (reg:di 5))
1248 We would now set REG_QTY (101) = REG_QTY (5), but the
1249 entry for 5 is in SImode. When we use this later in
1250 copy propagation, we get the register in wrong mode. */
1251 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1252 continue;
1254 make_regs_eqv (regno, c_regno);
1255 return 1;
1258 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1259 than REG_IN_TABLE to find out if there was only a single preceding
1260 invalidation - for the SUBREG - or another one, which would be
1261 for the full register. However, if we find here that REG_TICK
1262 indicates that the register is invalid, it means that it has
1263 been invalidated in a separate operation. The SUBREG might be used
1264 now (then this is a recursive call), or we might use the full REG
1265 now and a SUBREG of it later. So bump up REG_TICK so that
1266 mention_regs will do the right thing. */
1267 if (! modified
1268 && REG_IN_TABLE (regno) >= 0
1269 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1270 REG_TICK (regno)++;
1271 make_new_qty (regno, GET_MODE (x));
1272 return 1;
1275 return 0;
1278 /* If X is a SUBREG, we will likely be inserting the inner register in the
1279 table. If that register doesn't have an assigned quantity number at
1280 this point but does later, the insertion that we will be doing now will
1281 not be accessible because its hash code will have changed. So assign
1282 a quantity number now. */
1284 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1285 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1287 insert_regs (SUBREG_REG (x), NULL, 0);
1288 mention_regs (x);
1289 return 1;
1291 else
1292 return mention_regs (x);
1295 /* Look in or update the hash table. */
1297 /* Remove table element ELT from use in the table.
1298 HASH is its hash code, made using the HASH macro.
1299 It's an argument because often that is known in advance
1300 and we save much time not recomputing it. */
1302 static void
1303 remove_from_table (struct table_elt *elt, unsigned int hash)
1305 if (elt == 0)
1306 return;
1308 /* Mark this element as removed. See cse_insn. */
1309 elt->first_same_value = 0;
1311 /* Remove the table element from its equivalence class. */
1314 struct table_elt *prev = elt->prev_same_value;
1315 struct table_elt *next = elt->next_same_value;
1317 if (next)
1318 next->prev_same_value = prev;
1320 if (prev)
1321 prev->next_same_value = next;
1322 else
1324 struct table_elt *newfirst = next;
1325 while (next)
1327 next->first_same_value = newfirst;
1328 next = next->next_same_value;
1333 /* Remove the table element from its hash bucket. */
1336 struct table_elt *prev = elt->prev_same_hash;
1337 struct table_elt *next = elt->next_same_hash;
1339 if (next)
1340 next->prev_same_hash = prev;
1342 if (prev)
1343 prev->next_same_hash = next;
1344 else if (table[hash] == elt)
1345 table[hash] = next;
1346 else
1348 /* This entry is not in the proper hash bucket. This can happen
1349 when two classes were merged by `merge_equiv_classes'. Search
1350 for the hash bucket that it heads. This happens only very
1351 rarely, so the cost is acceptable. */
1352 for (hash = 0; hash < HASH_SIZE; hash++)
1353 if (table[hash] == elt)
1354 table[hash] = next;
1358 /* Remove the table element from its related-value circular chain. */
1360 if (elt->related_value != 0 && elt->related_value != elt)
1362 struct table_elt *p = elt->related_value;
1364 while (p->related_value != elt)
1365 p = p->related_value;
1366 p->related_value = elt->related_value;
1367 if (p->related_value == p)
1368 p->related_value = 0;
1371 /* Now add it to the free element chain. */
1372 elt->next_same_hash = free_element_chain;
1373 free_element_chain = elt;
1376 /* Look up X in the hash table and return its table element,
1377 or 0 if X is not in the table.
1379 MODE is the machine-mode of X, or if X is an integer constant
1380 with VOIDmode then MODE is the mode with which X will be used.
1382 Here we are satisfied to find an expression whose tree structure
1383 looks like X. */
1385 static struct table_elt *
1386 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1388 struct table_elt *p;
1390 for (p = table[hash]; p; p = p->next_same_hash)
1391 if (mode == p->mode && ((x == p->exp && REG_P (x))
1392 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1393 return p;
1395 return 0;
1398 /* Like `lookup' but don't care whether the table element uses invalid regs.
1399 Also ignore discrepancies in the machine mode of a register. */
1401 static struct table_elt *
1402 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1404 struct table_elt *p;
1406 if (REG_P (x))
1408 unsigned int regno = REGNO (x);
1410 /* Don't check the machine mode when comparing registers;
1411 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1412 for (p = table[hash]; p; p = p->next_same_hash)
1413 if (REG_P (p->exp)
1414 && REGNO (p->exp) == regno)
1415 return p;
1417 else
1419 for (p = table[hash]; p; p = p->next_same_hash)
1420 if (mode == p->mode
1421 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1422 return p;
1425 return 0;
1428 /* Look for an expression equivalent to X and with code CODE.
1429 If one is found, return that expression. */
1431 static rtx
1432 lookup_as_function (rtx x, enum rtx_code code)
1434 struct table_elt *p
1435 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1437 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1438 long as we are narrowing. So if we looked in vain for a mode narrower
1439 than word_mode before, look for word_mode now. */
1440 if (p == 0 && code == CONST_INT
1441 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1443 x = copy_rtx (x);
1444 PUT_MODE (x, word_mode);
1445 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1448 if (p == 0)
1449 return 0;
1451 for (p = p->first_same_value; p; p = p->next_same_value)
1452 if (GET_CODE (p->exp) == code
1453 /* Make sure this is a valid entry in the table. */
1454 && exp_equiv_p (p->exp, p->exp, 1, false))
1455 return p->exp;
1457 return 0;
1460 /* Insert X in the hash table, assuming HASH is its hash code
1461 and CLASSP is an element of the class it should go in
1462 (or 0 if a new class should be made).
1463 It is inserted at the proper position to keep the class in
1464 the order cheapest first.
1466 MODE is the machine-mode of X, or if X is an integer constant
1467 with VOIDmode then MODE is the mode with which X will be used.
1469 For elements of equal cheapness, the most recent one
1470 goes in front, except that the first element in the list
1471 remains first unless a cheaper element is added. The order of
1472 pseudo-registers does not matter, as canon_reg will be called to
1473 find the cheapest when a register is retrieved from the table.
1475 The in_memory field in the hash table element is set to 0.
1476 The caller must set it nonzero if appropriate.
1478 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1479 and if insert_regs returns a nonzero value
1480 you must then recompute its hash code before calling here.
1482 If necessary, update table showing constant values of quantities. */
1484 #define CHEAPER(X, Y) \
1485 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1487 static struct table_elt *
1488 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1490 struct table_elt *elt;
1492 /* If X is a register and we haven't made a quantity for it,
1493 something is wrong. */
1494 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1496 /* If X is a hard register, show it is being put in the table. */
1497 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1499 unsigned int regno = REGNO (x);
1500 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1501 unsigned int i;
1503 for (i = regno; i < endregno; i++)
1504 SET_HARD_REG_BIT (hard_regs_in_table, i);
1507 /* Put an element for X into the right hash bucket. */
1509 elt = free_element_chain;
1510 if (elt)
1511 free_element_chain = elt->next_same_hash;
1512 else
1513 elt = XNEW (struct table_elt);
1515 elt->exp = x;
1516 elt->canon_exp = NULL_RTX;
1517 elt->cost = COST (x);
1518 elt->regcost = approx_reg_cost (x);
1519 elt->next_same_value = 0;
1520 elt->prev_same_value = 0;
1521 elt->next_same_hash = table[hash];
1522 elt->prev_same_hash = 0;
1523 elt->related_value = 0;
1524 elt->in_memory = 0;
1525 elt->mode = mode;
1526 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1528 if (table[hash])
1529 table[hash]->prev_same_hash = elt;
1530 table[hash] = elt;
1532 /* Put it into the proper value-class. */
1533 if (classp)
1535 classp = classp->first_same_value;
1536 if (CHEAPER (elt, classp))
1537 /* Insert at the head of the class. */
1539 struct table_elt *p;
1540 elt->next_same_value = classp;
1541 classp->prev_same_value = elt;
1542 elt->first_same_value = elt;
1544 for (p = classp; p; p = p->next_same_value)
1545 p->first_same_value = elt;
1547 else
1549 /* Insert not at head of the class. */
1550 /* Put it after the last element cheaper than X. */
1551 struct table_elt *p, *next;
1553 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1554 p = next);
1556 /* Put it after P and before NEXT. */
1557 elt->next_same_value = next;
1558 if (next)
1559 next->prev_same_value = elt;
1561 elt->prev_same_value = p;
1562 p->next_same_value = elt;
1563 elt->first_same_value = classp;
1566 else
1567 elt->first_same_value = elt;
1569 /* If this is a constant being set equivalent to a register or a register
1570 being set equivalent to a constant, note the constant equivalence.
1572 If this is a constant, it cannot be equivalent to a different constant,
1573 and a constant is the only thing that can be cheaper than a register. So
1574 we know the register is the head of the class (before the constant was
1575 inserted).
1577 If this is a register that is not already known equivalent to a
1578 constant, we must check the entire class.
1580 If this is a register that is already known equivalent to an insn,
1581 update the qtys `const_insn' to show that `this_insn' is the latest
1582 insn making that quantity equivalent to the constant. */
1584 if (elt->is_const && classp && REG_P (classp->exp)
1585 && !REG_P (x))
1587 int exp_q = REG_QTY (REGNO (classp->exp));
1588 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1590 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1591 exp_ent->const_insn = this_insn;
1594 else if (REG_P (x)
1595 && classp
1596 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1597 && ! elt->is_const)
1599 struct table_elt *p;
1601 for (p = classp; p != 0; p = p->next_same_value)
1603 if (p->is_const && !REG_P (p->exp))
1605 int x_q = REG_QTY (REGNO (x));
1606 struct qty_table_elem *x_ent = &qty_table[x_q];
1608 x_ent->const_rtx
1609 = gen_lowpart (GET_MODE (x), p->exp);
1610 x_ent->const_insn = this_insn;
1611 break;
1616 else if (REG_P (x)
1617 && qty_table[REG_QTY (REGNO (x))].const_rtx
1618 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1619 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1621 /* If this is a constant with symbolic value,
1622 and it has a term with an explicit integer value,
1623 link it up with related expressions. */
1624 if (GET_CODE (x) == CONST)
1626 rtx subexp = get_related_value (x);
1627 unsigned subhash;
1628 struct table_elt *subelt, *subelt_prev;
1630 if (subexp != 0)
1632 /* Get the integer-free subexpression in the hash table. */
1633 subhash = SAFE_HASH (subexp, mode);
1634 subelt = lookup (subexp, subhash, mode);
1635 if (subelt == 0)
1636 subelt = insert (subexp, NULL, subhash, mode);
1637 /* Initialize SUBELT's circular chain if it has none. */
1638 if (subelt->related_value == 0)
1639 subelt->related_value = subelt;
1640 /* Find the element in the circular chain that precedes SUBELT. */
1641 subelt_prev = subelt;
1642 while (subelt_prev->related_value != subelt)
1643 subelt_prev = subelt_prev->related_value;
1644 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1645 This way the element that follows SUBELT is the oldest one. */
1646 elt->related_value = subelt_prev->related_value;
1647 subelt_prev->related_value = elt;
1651 return elt;
1654 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1655 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1656 the two classes equivalent.
1658 CLASS1 will be the surviving class; CLASS2 should not be used after this
1659 call.
1661 Any invalid entries in CLASS2 will not be copied. */
1663 static void
1664 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1666 struct table_elt *elt, *next, *new;
1668 /* Ensure we start with the head of the classes. */
1669 class1 = class1->first_same_value;
1670 class2 = class2->first_same_value;
1672 /* If they were already equal, forget it. */
1673 if (class1 == class2)
1674 return;
1676 for (elt = class2; elt; elt = next)
1678 unsigned int hash;
1679 rtx exp = elt->exp;
1680 enum machine_mode mode = elt->mode;
1682 next = elt->next_same_value;
1684 /* Remove old entry, make a new one in CLASS1's class.
1685 Don't do this for invalid entries as we cannot find their
1686 hash code (it also isn't necessary). */
1687 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1689 bool need_rehash = false;
1691 hash_arg_in_memory = 0;
1692 hash = HASH (exp, mode);
1694 if (REG_P (exp))
1696 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1697 delete_reg_equiv (REGNO (exp));
1700 remove_from_table (elt, hash);
1702 if (insert_regs (exp, class1, 0) || need_rehash)
1704 rehash_using_reg (exp);
1705 hash = HASH (exp, mode);
1707 new = insert (exp, class1, hash, mode);
1708 new->in_memory = hash_arg_in_memory;
1713 /* Flush the entire hash table. */
1715 static void
1716 flush_hash_table (void)
1718 int i;
1719 struct table_elt *p;
1721 for (i = 0; i < HASH_SIZE; i++)
1722 for (p = table[i]; p; p = table[i])
1724 /* Note that invalidate can remove elements
1725 after P in the current hash chain. */
1726 if (REG_P (p->exp))
1727 invalidate (p->exp, VOIDmode);
1728 else
1729 remove_from_table (p, i);
1733 /* Function called for each rtx to check whether true dependence exist. */
1734 struct check_dependence_data
1736 enum machine_mode mode;
1737 rtx exp;
1738 rtx addr;
1741 static int
1742 check_dependence (rtx *x, void *data)
1744 struct check_dependence_data *d = (struct check_dependence_data *) data;
1745 if (*x && MEM_P (*x))
1746 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1747 cse_rtx_varies_p);
1748 else
1749 return 0;
1752 /* Remove from the hash table, or mark as invalid, all expressions whose
1753 values could be altered by storing in X. X is a register, a subreg, or
1754 a memory reference with nonvarying address (because, when a memory
1755 reference with a varying address is stored in, all memory references are
1756 removed by invalidate_memory so specific invalidation is superfluous).
1757 FULL_MODE, if not VOIDmode, indicates that this much should be
1758 invalidated instead of just the amount indicated by the mode of X. This
1759 is only used for bitfield stores into memory.
1761 A nonvarying address may be just a register or just a symbol reference,
1762 or it may be either of those plus a numeric offset. */
1764 static void
1765 invalidate (rtx x, enum machine_mode full_mode)
1767 int i;
1768 struct table_elt *p;
1769 rtx addr;
1771 switch (GET_CODE (x))
1773 case REG:
1775 /* If X is a register, dependencies on its contents are recorded
1776 through the qty number mechanism. Just change the qty number of
1777 the register, mark it as invalid for expressions that refer to it,
1778 and remove it itself. */
1779 unsigned int regno = REGNO (x);
1780 unsigned int hash = HASH (x, GET_MODE (x));
1782 /* Remove REGNO from any quantity list it might be on and indicate
1783 that its value might have changed. If it is a pseudo, remove its
1784 entry from the hash table.
1786 For a hard register, we do the first two actions above for any
1787 additional hard registers corresponding to X. Then, if any of these
1788 registers are in the table, we must remove any REG entries that
1789 overlap these registers. */
1791 delete_reg_equiv (regno);
1792 REG_TICK (regno)++;
1793 SUBREG_TICKED (regno) = -1;
1795 if (regno >= FIRST_PSEUDO_REGISTER)
1797 /* Because a register can be referenced in more than one mode,
1798 we might have to remove more than one table entry. */
1799 struct table_elt *elt;
1801 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1802 remove_from_table (elt, hash);
1804 else
1806 HOST_WIDE_INT in_table
1807 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1808 unsigned int endregno
1809 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1810 unsigned int tregno, tendregno, rn;
1811 struct table_elt *p, *next;
1813 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1815 for (rn = regno + 1; rn < endregno; rn++)
1817 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1818 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1819 delete_reg_equiv (rn);
1820 REG_TICK (rn)++;
1821 SUBREG_TICKED (rn) = -1;
1824 if (in_table)
1825 for (hash = 0; hash < HASH_SIZE; hash++)
1826 for (p = table[hash]; p; p = next)
1828 next = p->next_same_hash;
1830 if (!REG_P (p->exp)
1831 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1832 continue;
1834 tregno = REGNO (p->exp);
1835 tendregno
1836 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1837 if (tendregno > regno && tregno < endregno)
1838 remove_from_table (p, hash);
1842 return;
1844 case SUBREG:
1845 invalidate (SUBREG_REG (x), VOIDmode);
1846 return;
1848 case PARALLEL:
1849 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1850 invalidate (XVECEXP (x, 0, i), VOIDmode);
1851 return;
1853 case EXPR_LIST:
1854 /* This is part of a disjoint return value; extract the location in
1855 question ignoring the offset. */
1856 invalidate (XEXP (x, 0), VOIDmode);
1857 return;
1859 case MEM:
1860 addr = canon_rtx (get_addr (XEXP (x, 0)));
1861 /* Calculate the canonical version of X here so that
1862 true_dependence doesn't generate new RTL for X on each call. */
1863 x = canon_rtx (x);
1865 /* Remove all hash table elements that refer to overlapping pieces of
1866 memory. */
1867 if (full_mode == VOIDmode)
1868 full_mode = GET_MODE (x);
1870 for (i = 0; i < HASH_SIZE; i++)
1872 struct table_elt *next;
1874 for (p = table[i]; p; p = next)
1876 next = p->next_same_hash;
1877 if (p->in_memory)
1879 struct check_dependence_data d;
1881 /* Just canonicalize the expression once;
1882 otherwise each time we call invalidate
1883 true_dependence will canonicalize the
1884 expression again. */
1885 if (!p->canon_exp)
1886 p->canon_exp = canon_rtx (p->exp);
1887 d.exp = x;
1888 d.addr = addr;
1889 d.mode = full_mode;
1890 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1891 remove_from_table (p, i);
1895 return;
1897 default:
1898 gcc_unreachable ();
1902 /* Remove all expressions that refer to register REGNO,
1903 since they are already invalid, and we are about to
1904 mark that register valid again and don't want the old
1905 expressions to reappear as valid. */
1907 static void
1908 remove_invalid_refs (unsigned int regno)
1910 unsigned int i;
1911 struct table_elt *p, *next;
1913 for (i = 0; i < HASH_SIZE; i++)
1914 for (p = table[i]; p; p = next)
1916 next = p->next_same_hash;
1917 if (!REG_P (p->exp)
1918 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1919 remove_from_table (p, i);
1923 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1924 and mode MODE. */
1925 static void
1926 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1927 enum machine_mode mode)
1929 unsigned int i;
1930 struct table_elt *p, *next;
1931 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1933 for (i = 0; i < HASH_SIZE; i++)
1934 for (p = table[i]; p; p = next)
1936 rtx exp = p->exp;
1937 next = p->next_same_hash;
1939 if (!REG_P (exp)
1940 && (GET_CODE (exp) != SUBREG
1941 || !REG_P (SUBREG_REG (exp))
1942 || REGNO (SUBREG_REG (exp)) != regno
1943 || (((SUBREG_BYTE (exp)
1944 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1945 && SUBREG_BYTE (exp) <= end))
1946 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1947 remove_from_table (p, i);
1951 /* Recompute the hash codes of any valid entries in the hash table that
1952 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1954 This is called when we make a jump equivalence. */
1956 static void
1957 rehash_using_reg (rtx x)
1959 unsigned int i;
1960 struct table_elt *p, *next;
1961 unsigned hash;
1963 if (GET_CODE (x) == SUBREG)
1964 x = SUBREG_REG (x);
1966 /* If X is not a register or if the register is known not to be in any
1967 valid entries in the table, we have no work to do. */
1969 if (!REG_P (x)
1970 || REG_IN_TABLE (REGNO (x)) < 0
1971 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1972 return;
1974 /* Scan all hash chains looking for valid entries that mention X.
1975 If we find one and it is in the wrong hash chain, move it. */
1977 for (i = 0; i < HASH_SIZE; i++)
1978 for (p = table[i]; p; p = next)
1980 next = p->next_same_hash;
1981 if (reg_mentioned_p (x, p->exp)
1982 && exp_equiv_p (p->exp, p->exp, 1, false)
1983 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1985 if (p->next_same_hash)
1986 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1988 if (p->prev_same_hash)
1989 p->prev_same_hash->next_same_hash = p->next_same_hash;
1990 else
1991 table[i] = p->next_same_hash;
1993 p->next_same_hash = table[hash];
1994 p->prev_same_hash = 0;
1995 if (table[hash])
1996 table[hash]->prev_same_hash = p;
1997 table[hash] = p;
2002 /* Remove from the hash table any expression that is a call-clobbered
2003 register. Also update their TICK values. */
2005 static void
2006 invalidate_for_call (void)
2008 unsigned int regno, endregno;
2009 unsigned int i;
2010 unsigned hash;
2011 struct table_elt *p, *next;
2012 int in_table = 0;
2014 /* Go through all the hard registers. For each that is clobbered in
2015 a CALL_INSN, remove the register from quantity chains and update
2016 reg_tick if defined. Also see if any of these registers is currently
2017 in the table. */
2019 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2020 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2022 delete_reg_equiv (regno);
2023 if (REG_TICK (regno) >= 0)
2025 REG_TICK (regno)++;
2026 SUBREG_TICKED (regno) = -1;
2029 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2032 /* In the case where we have no call-clobbered hard registers in the
2033 table, we are done. Otherwise, scan the table and remove any
2034 entry that overlaps a call-clobbered register. */
2036 if (in_table)
2037 for (hash = 0; hash < HASH_SIZE; hash++)
2038 for (p = table[hash]; p; p = next)
2040 next = p->next_same_hash;
2042 if (!REG_P (p->exp)
2043 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2044 continue;
2046 regno = REGNO (p->exp);
2047 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
2049 for (i = regno; i < endregno; i++)
2050 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2052 remove_from_table (p, hash);
2053 break;
2058 /* Given an expression X of type CONST,
2059 and ELT which is its table entry (or 0 if it
2060 is not in the hash table),
2061 return an alternate expression for X as a register plus integer.
2062 If none can be found, return 0. */
2064 static rtx
2065 use_related_value (rtx x, struct table_elt *elt)
2067 struct table_elt *relt = 0;
2068 struct table_elt *p, *q;
2069 HOST_WIDE_INT offset;
2071 /* First, is there anything related known?
2072 If we have a table element, we can tell from that.
2073 Otherwise, must look it up. */
2075 if (elt != 0 && elt->related_value != 0)
2076 relt = elt;
2077 else if (elt == 0 && GET_CODE (x) == CONST)
2079 rtx subexp = get_related_value (x);
2080 if (subexp != 0)
2081 relt = lookup (subexp,
2082 SAFE_HASH (subexp, GET_MODE (subexp)),
2083 GET_MODE (subexp));
2086 if (relt == 0)
2087 return 0;
2089 /* Search all related table entries for one that has an
2090 equivalent register. */
2092 p = relt;
2093 while (1)
2095 /* This loop is strange in that it is executed in two different cases.
2096 The first is when X is already in the table. Then it is searching
2097 the RELATED_VALUE list of X's class (RELT). The second case is when
2098 X is not in the table. Then RELT points to a class for the related
2099 value.
2101 Ensure that, whatever case we are in, that we ignore classes that have
2102 the same value as X. */
2104 if (rtx_equal_p (x, p->exp))
2105 q = 0;
2106 else
2107 for (q = p->first_same_value; q; q = q->next_same_value)
2108 if (REG_P (q->exp))
2109 break;
2111 if (q)
2112 break;
2114 p = p->related_value;
2116 /* We went all the way around, so there is nothing to be found.
2117 Alternatively, perhaps RELT was in the table for some other reason
2118 and it has no related values recorded. */
2119 if (p == relt || p == 0)
2120 break;
2123 if (q == 0)
2124 return 0;
2126 offset = (get_integer_term (x) - get_integer_term (p->exp));
2127 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2128 return plus_constant (q->exp, offset);
2131 /* Hash a string. Just add its bytes up. */
2132 static inline unsigned
2133 hash_rtx_string (const char *ps)
2135 unsigned hash = 0;
2136 const unsigned char *p = (const unsigned char *) ps;
2138 if (p)
2139 while (*p)
2140 hash += *p++;
2142 return hash;
2145 /* Hash an rtx. We are careful to make sure the value is never negative.
2146 Equivalent registers hash identically.
2147 MODE is used in hashing for CONST_INTs only;
2148 otherwise the mode of X is used.
2150 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2152 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2153 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2155 Note that cse_insn knows that the hash code of a MEM expression
2156 is just (int) MEM plus the hash code of the address. */
2158 unsigned
2159 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2160 int *hash_arg_in_memory_p, bool have_reg_qty)
2162 int i, j;
2163 unsigned hash = 0;
2164 enum rtx_code code;
2165 const char *fmt;
2167 /* Used to turn recursion into iteration. We can't rely on GCC's
2168 tail-recursion elimination since we need to keep accumulating values
2169 in HASH. */
2170 repeat:
2171 if (x == 0)
2172 return hash;
2174 code = GET_CODE (x);
2175 switch (code)
2177 case REG:
2179 unsigned int regno = REGNO (x);
2181 if (!reload_completed)
2183 /* On some machines, we can't record any non-fixed hard register,
2184 because extending its life will cause reload problems. We
2185 consider ap, fp, sp, gp to be fixed for this purpose.
2187 We also consider CCmode registers to be fixed for this purpose;
2188 failure to do so leads to failure to simplify 0<100 type of
2189 conditionals.
2191 On all machines, we can't record any global registers.
2192 Nor should we record any register that is in a small
2193 class, as defined by CLASS_LIKELY_SPILLED_P. */
2194 bool record;
2196 if (regno >= FIRST_PSEUDO_REGISTER)
2197 record = true;
2198 else if (x == frame_pointer_rtx
2199 || x == hard_frame_pointer_rtx
2200 || x == arg_pointer_rtx
2201 || x == stack_pointer_rtx
2202 || x == pic_offset_table_rtx)
2203 record = true;
2204 else if (global_regs[regno])
2205 record = false;
2206 else if (fixed_regs[regno])
2207 record = true;
2208 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2209 record = true;
2210 else if (SMALL_REGISTER_CLASSES)
2211 record = false;
2212 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2213 record = false;
2214 else
2215 record = true;
2217 if (!record)
2219 *do_not_record_p = 1;
2220 return 0;
2224 hash += ((unsigned int) REG << 7);
2225 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2226 return hash;
2229 /* We handle SUBREG of a REG specially because the underlying
2230 reg changes its hash value with every value change; we don't
2231 want to have to forget unrelated subregs when one subreg changes. */
2232 case SUBREG:
2234 if (REG_P (SUBREG_REG (x)))
2236 hash += (((unsigned int) SUBREG << 7)
2237 + REGNO (SUBREG_REG (x))
2238 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2239 return hash;
2241 break;
2244 case CONST_INT:
2245 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2246 + (unsigned int) INTVAL (x));
2247 return hash;
2249 case CONST_DOUBLE:
2250 /* This is like the general case, except that it only counts
2251 the integers representing the constant. */
2252 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2253 if (GET_MODE (x) != VOIDmode)
2254 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2255 else
2256 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2257 + (unsigned int) CONST_DOUBLE_HIGH (x));
2258 return hash;
2260 case CONST_VECTOR:
2262 int units;
2263 rtx elt;
2265 units = CONST_VECTOR_NUNITS (x);
2267 for (i = 0; i < units; ++i)
2269 elt = CONST_VECTOR_ELT (x, i);
2270 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2271 hash_arg_in_memory_p, have_reg_qty);
2274 return hash;
2277 /* Assume there is only one rtx object for any given label. */
2278 case LABEL_REF:
2279 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2280 differences and differences between each stage's debugging dumps. */
2281 hash += (((unsigned int) LABEL_REF << 7)
2282 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2283 return hash;
2285 case SYMBOL_REF:
2287 /* Don't hash on the symbol's address to avoid bootstrap differences.
2288 Different hash values may cause expressions to be recorded in
2289 different orders and thus different registers to be used in the
2290 final assembler. This also avoids differences in the dump files
2291 between various stages. */
2292 unsigned int h = 0;
2293 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2295 while (*p)
2296 h += (h << 7) + *p++; /* ??? revisit */
2298 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2299 return hash;
2302 case MEM:
2303 /* We don't record if marked volatile or if BLKmode since we don't
2304 know the size of the move. */
2305 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2307 *do_not_record_p = 1;
2308 return 0;
2310 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2311 *hash_arg_in_memory_p = 1;
2313 /* Now that we have already found this special case,
2314 might as well speed it up as much as possible. */
2315 hash += (unsigned) MEM;
2316 x = XEXP (x, 0);
2317 goto repeat;
2319 case USE:
2320 /* A USE that mentions non-volatile memory needs special
2321 handling since the MEM may be BLKmode which normally
2322 prevents an entry from being made. Pure calls are
2323 marked by a USE which mentions BLKmode memory.
2324 See calls.c:emit_call_1. */
2325 if (MEM_P (XEXP (x, 0))
2326 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2328 hash += (unsigned) USE;
2329 x = XEXP (x, 0);
2331 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2332 *hash_arg_in_memory_p = 1;
2334 /* Now that we have already found this special case,
2335 might as well speed it up as much as possible. */
2336 hash += (unsigned) MEM;
2337 x = XEXP (x, 0);
2338 goto repeat;
2340 break;
2342 case PRE_DEC:
2343 case PRE_INC:
2344 case POST_DEC:
2345 case POST_INC:
2346 case PRE_MODIFY:
2347 case POST_MODIFY:
2348 case PC:
2349 case CC0:
2350 case CALL:
2351 case UNSPEC_VOLATILE:
2352 *do_not_record_p = 1;
2353 return 0;
2355 case ASM_OPERANDS:
2356 if (MEM_VOLATILE_P (x))
2358 *do_not_record_p = 1;
2359 return 0;
2361 else
2363 /* We don't want to take the filename and line into account. */
2364 hash += (unsigned) code + (unsigned) GET_MODE (x)
2365 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2366 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2367 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2369 if (ASM_OPERANDS_INPUT_LENGTH (x))
2371 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2373 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2374 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2375 do_not_record_p, hash_arg_in_memory_p,
2376 have_reg_qty)
2377 + hash_rtx_string
2378 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2381 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2382 x = ASM_OPERANDS_INPUT (x, 0);
2383 mode = GET_MODE (x);
2384 goto repeat;
2387 return hash;
2389 break;
2391 default:
2392 break;
2395 i = GET_RTX_LENGTH (code) - 1;
2396 hash += (unsigned) code + (unsigned) GET_MODE (x);
2397 fmt = GET_RTX_FORMAT (code);
2398 for (; i >= 0; i--)
2400 switch (fmt[i])
2402 case 'e':
2403 /* If we are about to do the last recursive call
2404 needed at this level, change it into iteration.
2405 This function is called enough to be worth it. */
2406 if (i == 0)
2408 x = XEXP (x, i);
2409 goto repeat;
2412 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2413 hash_arg_in_memory_p, have_reg_qty);
2414 break;
2416 case 'E':
2417 for (j = 0; j < XVECLEN (x, i); j++)
2418 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2419 hash_arg_in_memory_p, have_reg_qty);
2420 break;
2422 case 's':
2423 hash += hash_rtx_string (XSTR (x, i));
2424 break;
2426 case 'i':
2427 hash += (unsigned int) XINT (x, i);
2428 break;
2430 case '0': case 't':
2431 /* Unused. */
2432 break;
2434 default:
2435 gcc_unreachable ();
2439 return hash;
2442 /* Hash an rtx X for cse via hash_rtx.
2443 Stores 1 in do_not_record if any subexpression is volatile.
2444 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2445 does not have the RTX_UNCHANGING_P bit set. */
2447 static inline unsigned
2448 canon_hash (rtx x, enum machine_mode mode)
2450 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2453 /* Like canon_hash but with no side effects, i.e. do_not_record
2454 and hash_arg_in_memory are not changed. */
2456 static inline unsigned
2457 safe_hash (rtx x, enum machine_mode mode)
2459 int dummy_do_not_record;
2460 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2463 /* Return 1 iff X and Y would canonicalize into the same thing,
2464 without actually constructing the canonicalization of either one.
2465 If VALIDATE is nonzero,
2466 we assume X is an expression being processed from the rtl
2467 and Y was found in the hash table. We check register refs
2468 in Y for being marked as valid.
2470 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2473 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2475 int i, j;
2476 enum rtx_code code;
2477 const char *fmt;
2479 /* Note: it is incorrect to assume an expression is equivalent to itself
2480 if VALIDATE is nonzero. */
2481 if (x == y && !validate)
2482 return 1;
2484 if (x == 0 || y == 0)
2485 return x == y;
2487 code = GET_CODE (x);
2488 if (code != GET_CODE (y))
2489 return 0;
2491 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2492 if (GET_MODE (x) != GET_MODE (y))
2493 return 0;
2495 switch (code)
2497 case PC:
2498 case CC0:
2499 case CONST_INT:
2500 case CONST_DOUBLE:
2501 return x == y;
2503 case LABEL_REF:
2504 return XEXP (x, 0) == XEXP (y, 0);
2506 case SYMBOL_REF:
2507 return XSTR (x, 0) == XSTR (y, 0);
2509 case REG:
2510 if (for_gcse)
2511 return REGNO (x) == REGNO (y);
2512 else
2514 unsigned int regno = REGNO (y);
2515 unsigned int i;
2516 unsigned int endregno
2517 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2518 : hard_regno_nregs[regno][GET_MODE (y)]);
2520 /* If the quantities are not the same, the expressions are not
2521 equivalent. If there are and we are not to validate, they
2522 are equivalent. Otherwise, ensure all regs are up-to-date. */
2524 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2525 return 0;
2527 if (! validate)
2528 return 1;
2530 for (i = regno; i < endregno; i++)
2531 if (REG_IN_TABLE (i) != REG_TICK (i))
2532 return 0;
2534 return 1;
2537 case MEM:
2538 if (for_gcse)
2540 /* A volatile mem should not be considered equivalent to any
2541 other. */
2542 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2543 return 0;
2545 /* Can't merge two expressions in different alias sets, since we
2546 can decide that the expression is transparent in a block when
2547 it isn't, due to it being set with the different alias set.
2549 Also, can't merge two expressions with different MEM_ATTRS.
2550 They could e.g. be two different entities allocated into the
2551 same space on the stack (see e.g. PR25130). In that case, the
2552 MEM addresses can be the same, even though the two MEMs are
2553 absolutely not equivalent.
2555 But because really all MEM attributes should be the same for
2556 equivalent MEMs, we just use the invariant that MEMs that have
2557 the same attributes share the same mem_attrs data structure. */
2558 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2559 return 0;
2561 break;
2563 /* For commutative operations, check both orders. */
2564 case PLUS:
2565 case MULT:
2566 case AND:
2567 case IOR:
2568 case XOR:
2569 case NE:
2570 case EQ:
2571 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2572 validate, for_gcse)
2573 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2574 validate, for_gcse))
2575 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2576 validate, for_gcse)
2577 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2578 validate, for_gcse)));
2580 case ASM_OPERANDS:
2581 /* We don't use the generic code below because we want to
2582 disregard filename and line numbers. */
2584 /* A volatile asm isn't equivalent to any other. */
2585 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2586 return 0;
2588 if (GET_MODE (x) != GET_MODE (y)
2589 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2590 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2591 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2592 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2593 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2594 return 0;
2596 if (ASM_OPERANDS_INPUT_LENGTH (x))
2598 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2599 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2600 ASM_OPERANDS_INPUT (y, i),
2601 validate, for_gcse)
2602 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2603 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2604 return 0;
2607 return 1;
2609 default:
2610 break;
2613 /* Compare the elements. If any pair of corresponding elements
2614 fail to match, return 0 for the whole thing. */
2616 fmt = GET_RTX_FORMAT (code);
2617 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2619 switch (fmt[i])
2621 case 'e':
2622 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2623 validate, for_gcse))
2624 return 0;
2625 break;
2627 case 'E':
2628 if (XVECLEN (x, i) != XVECLEN (y, i))
2629 return 0;
2630 for (j = 0; j < XVECLEN (x, i); j++)
2631 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2632 validate, for_gcse))
2633 return 0;
2634 break;
2636 case 's':
2637 if (strcmp (XSTR (x, i), XSTR (y, i)))
2638 return 0;
2639 break;
2641 case 'i':
2642 if (XINT (x, i) != XINT (y, i))
2643 return 0;
2644 break;
2646 case 'w':
2647 if (XWINT (x, i) != XWINT (y, i))
2648 return 0;
2649 break;
2651 case '0':
2652 case 't':
2653 break;
2655 default:
2656 gcc_unreachable ();
2660 return 1;
2663 /* Return 1 if X has a value that can vary even between two
2664 executions of the program. 0 means X can be compared reliably
2665 against certain constants or near-constants. */
2667 static int
2668 cse_rtx_varies_p (rtx x, int from_alias)
2670 /* We need not check for X and the equivalence class being of the same
2671 mode because if X is equivalent to a constant in some mode, it
2672 doesn't vary in any mode. */
2674 if (REG_P (x)
2675 && REGNO_QTY_VALID_P (REGNO (x)))
2677 int x_q = REG_QTY (REGNO (x));
2678 struct qty_table_elem *x_ent = &qty_table[x_q];
2680 if (GET_MODE (x) == x_ent->mode
2681 && x_ent->const_rtx != NULL_RTX)
2682 return 0;
2685 if (GET_CODE (x) == PLUS
2686 && GET_CODE (XEXP (x, 1)) == CONST_INT
2687 && REG_P (XEXP (x, 0))
2688 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2690 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2691 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2693 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2694 && x0_ent->const_rtx != NULL_RTX)
2695 return 0;
2698 /* This can happen as the result of virtual register instantiation, if
2699 the initial constant is too large to be a valid address. This gives
2700 us a three instruction sequence, load large offset into a register,
2701 load fp minus a constant into a register, then a MEM which is the
2702 sum of the two `constant' registers. */
2703 if (GET_CODE (x) == PLUS
2704 && REG_P (XEXP (x, 0))
2705 && REG_P (XEXP (x, 1))
2706 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2707 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2709 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2710 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2711 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2712 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2714 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2715 && x0_ent->const_rtx != NULL_RTX
2716 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2717 && x1_ent->const_rtx != NULL_RTX)
2718 return 0;
2721 return rtx_varies_p (x, from_alias);
2724 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2725 the result if necessary. INSN is as for canon_reg. */
2727 static void
2728 validate_canon_reg (rtx *xloc, rtx insn)
2730 rtx new = canon_reg (*xloc, insn);
2732 /* If replacing pseudo with hard reg or vice versa, ensure the
2733 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2734 if (insn != 0 && new != 0)
2735 validate_change (insn, xloc, new, 1);
2736 else
2737 *xloc = new;
2740 /* Canonicalize an expression:
2741 replace each register reference inside it
2742 with the "oldest" equivalent register.
2744 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2745 after we make our substitution. The calls are made with IN_GROUP nonzero
2746 so apply_change_group must be called upon the outermost return from this
2747 function (unless INSN is zero). The result of apply_change_group can
2748 generally be discarded since the changes we are making are optional. */
2750 static rtx
2751 canon_reg (rtx x, rtx insn)
2753 int i;
2754 enum rtx_code code;
2755 const char *fmt;
2757 if (x == 0)
2758 return x;
2760 code = GET_CODE (x);
2761 switch (code)
2763 case PC:
2764 case CC0:
2765 case CONST:
2766 case CONST_INT:
2767 case CONST_DOUBLE:
2768 case CONST_VECTOR:
2769 case SYMBOL_REF:
2770 case LABEL_REF:
2771 case ADDR_VEC:
2772 case ADDR_DIFF_VEC:
2773 return x;
2775 case REG:
2777 int first;
2778 int q;
2779 struct qty_table_elem *ent;
2781 /* Never replace a hard reg, because hard regs can appear
2782 in more than one machine mode, and we must preserve the mode
2783 of each occurrence. Also, some hard regs appear in
2784 MEMs that are shared and mustn't be altered. Don't try to
2785 replace any reg that maps to a reg of class NO_REGS. */
2786 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2787 || ! REGNO_QTY_VALID_P (REGNO (x)))
2788 return x;
2790 q = REG_QTY (REGNO (x));
2791 ent = &qty_table[q];
2792 first = ent->first_reg;
2793 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2794 : REGNO_REG_CLASS (first) == NO_REGS ? x
2795 : gen_rtx_REG (ent->mode, first));
2798 default:
2799 break;
2802 fmt = GET_RTX_FORMAT (code);
2803 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2805 int j;
2807 if (fmt[i] == 'e')
2808 validate_canon_reg (&XEXP (x, i), insn);
2809 else if (fmt[i] == 'E')
2810 for (j = 0; j < XVECLEN (x, i); j++)
2811 validate_canon_reg (&XVECEXP (x, i, j), insn);
2814 return x;
2817 /* LOC is a location within INSN that is an operand address (the contents of
2818 a MEM). Find the best equivalent address to use that is valid for this
2819 insn.
2821 On most CISC machines, complicated address modes are costly, and rtx_cost
2822 is a good approximation for that cost. However, most RISC machines have
2823 only a few (usually only one) memory reference formats. If an address is
2824 valid at all, it is often just as cheap as any other address. Hence, for
2825 RISC machines, we use `address_cost' to compare the costs of various
2826 addresses. For two addresses of equal cost, choose the one with the
2827 highest `rtx_cost' value as that has the potential of eliminating the
2828 most insns. For equal costs, we choose the first in the equivalence
2829 class. Note that we ignore the fact that pseudo registers are cheaper than
2830 hard registers here because we would also prefer the pseudo registers. */
2832 static void
2833 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2835 struct table_elt *elt;
2836 rtx addr = *loc;
2837 struct table_elt *p;
2838 int found_better = 1;
2839 int save_do_not_record = do_not_record;
2840 int save_hash_arg_in_memory = hash_arg_in_memory;
2841 int addr_volatile;
2842 int regno;
2843 unsigned hash;
2845 /* Do not try to replace constant addresses or addresses of local and
2846 argument slots. These MEM expressions are made only once and inserted
2847 in many instructions, as well as being used to control symbol table
2848 output. It is not safe to clobber them.
2850 There are some uncommon cases where the address is already in a register
2851 for some reason, but we cannot take advantage of that because we have
2852 no easy way to unshare the MEM. In addition, looking up all stack
2853 addresses is costly. */
2854 if ((GET_CODE (addr) == PLUS
2855 && REG_P (XEXP (addr, 0))
2856 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2857 && (regno = REGNO (XEXP (addr, 0)),
2858 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2859 || regno == ARG_POINTER_REGNUM))
2860 || (REG_P (addr)
2861 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2862 || regno == HARD_FRAME_POINTER_REGNUM
2863 || regno == ARG_POINTER_REGNUM))
2864 || CONSTANT_ADDRESS_P (addr))
2865 return;
2867 /* If this address is not simply a register, try to fold it. This will
2868 sometimes simplify the expression. Many simplifications
2869 will not be valid, but some, usually applying the associative rule, will
2870 be valid and produce better code. */
2871 if (!REG_P (addr))
2873 rtx folded = canon_for_address (fold_rtx (addr, NULL_RTX));
2875 if (folded != addr)
2877 int addr_folded_cost = address_cost (folded, mode);
2878 int addr_cost = address_cost (addr, mode);
2880 if ((addr_folded_cost < addr_cost
2881 || (addr_folded_cost == addr_cost
2882 /* ??? The rtx_cost comparison is left over from an older
2883 version of this code. It is probably no longer helpful.*/
2884 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2885 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2886 && validate_change (insn, loc, folded, 0))
2887 addr = folded;
2891 /* If this address is not in the hash table, we can't look for equivalences
2892 of the whole address. Also, ignore if volatile. */
2894 do_not_record = 0;
2895 hash = HASH (addr, Pmode);
2896 addr_volatile = do_not_record;
2897 do_not_record = save_do_not_record;
2898 hash_arg_in_memory = save_hash_arg_in_memory;
2900 if (addr_volatile)
2901 return;
2903 elt = lookup (addr, hash, Pmode);
2905 if (elt)
2907 /* We need to find the best (under the criteria documented above) entry
2908 in the class that is valid. We use the `flag' field to indicate
2909 choices that were invalid and iterate until we can't find a better
2910 one that hasn't already been tried. */
2912 for (p = elt->first_same_value; p; p = p->next_same_value)
2913 p->flag = 0;
2915 while (found_better)
2917 int best_addr_cost = address_cost (*loc, mode);
2918 int best_rtx_cost = (elt->cost + 1) >> 1;
2919 int exp_cost;
2920 struct table_elt *best_elt = elt;
2922 found_better = 0;
2923 for (p = elt->first_same_value; p; p = p->next_same_value)
2924 if (! p->flag)
2926 if ((REG_P (p->exp)
2927 || exp_equiv_p (p->exp, p->exp, 1, false))
2928 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2929 || (exp_cost == best_addr_cost
2930 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2932 found_better = 1;
2933 best_addr_cost = exp_cost;
2934 best_rtx_cost = (p->cost + 1) >> 1;
2935 best_elt = p;
2939 if (found_better)
2941 if (validate_change (insn, loc,
2942 canon_reg (copy_rtx (best_elt->exp),
2943 NULL_RTX), 0))
2944 return;
2945 else
2946 best_elt->flag = 1;
2951 /* If the address is a binary operation with the first operand a register
2952 and the second a constant, do the same as above, but looking for
2953 equivalences of the register. Then try to simplify before checking for
2954 the best address to use. This catches a few cases: First is when we
2955 have REG+const and the register is another REG+const. We can often merge
2956 the constants and eliminate one insn and one register. It may also be
2957 that a machine has a cheap REG+REG+const. Finally, this improves the
2958 code on the Alpha for unaligned byte stores. */
2960 if (flag_expensive_optimizations
2961 && ARITHMETIC_P (*loc)
2962 && REG_P (XEXP (*loc, 0)))
2964 rtx op1 = XEXP (*loc, 1);
2966 do_not_record = 0;
2967 hash = HASH (XEXP (*loc, 0), Pmode);
2968 do_not_record = save_do_not_record;
2969 hash_arg_in_memory = save_hash_arg_in_memory;
2971 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2972 if (elt == 0)
2973 return;
2975 /* We need to find the best (under the criteria documented above) entry
2976 in the class that is valid. We use the `flag' field to indicate
2977 choices that were invalid and iterate until we can't find a better
2978 one that hasn't already been tried. */
2980 for (p = elt->first_same_value; p; p = p->next_same_value)
2981 p->flag = 0;
2983 while (found_better)
2985 int best_addr_cost = address_cost (*loc, mode);
2986 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2987 struct table_elt *best_elt = elt;
2988 rtx best_rtx = *loc;
2989 int count;
2991 /* This is at worst case an O(n^2) algorithm, so limit our search
2992 to the first 32 elements on the list. This avoids trouble
2993 compiling code with very long basic blocks that can easily
2994 call simplify_gen_binary so many times that we run out of
2995 memory. */
2997 found_better = 0;
2998 for (p = elt->first_same_value, count = 0;
2999 p && count < 32;
3000 p = p->next_same_value, count++)
3001 if (! p->flag
3002 && (REG_P (p->exp)
3003 || (GET_CODE (p->exp) != EXPR_LIST
3004 && exp_equiv_p (p->exp, p->exp, 1, false))))
3007 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3008 p->exp, op1);
3009 int new_cost;
3011 /* Get the canonical version of the address so we can accept
3012 more. */
3013 new = canon_for_address (new);
3015 new_cost = address_cost (new, mode);
3017 if (new_cost < best_addr_cost
3018 || (new_cost == best_addr_cost
3019 && (COST (new) + 1) >> 1 > best_rtx_cost))
3021 found_better = 1;
3022 best_addr_cost = new_cost;
3023 best_rtx_cost = (COST (new) + 1) >> 1;
3024 best_elt = p;
3025 best_rtx = new;
3029 if (found_better)
3031 if (validate_change (insn, loc,
3032 canon_reg (copy_rtx (best_rtx),
3033 NULL_RTX), 0))
3034 return;
3035 else
3036 best_elt->flag = 1;
3042 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3043 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3044 what values are being compared.
3046 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3047 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3048 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3049 compared to produce cc0.
3051 The return value is the comparison operator and is either the code of
3052 A or the code corresponding to the inverse of the comparison. */
3054 static enum rtx_code
3055 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3056 enum machine_mode *pmode1, enum machine_mode *pmode2)
3058 rtx arg1, arg2;
3060 arg1 = *parg1, arg2 = *parg2;
3062 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3064 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3066 /* Set nonzero when we find something of interest. */
3067 rtx x = 0;
3068 int reverse_code = 0;
3069 struct table_elt *p = 0;
3071 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3072 On machines with CC0, this is the only case that can occur, since
3073 fold_rtx will return the COMPARE or item being compared with zero
3074 when given CC0. */
3076 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3077 x = arg1;
3079 /* If ARG1 is a comparison operator and CODE is testing for
3080 STORE_FLAG_VALUE, get the inner arguments. */
3082 else if (COMPARISON_P (arg1))
3084 #ifdef FLOAT_STORE_FLAG_VALUE
3085 REAL_VALUE_TYPE fsfv;
3086 #endif
3088 if (code == NE
3089 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3090 && code == LT && STORE_FLAG_VALUE == -1)
3091 #ifdef FLOAT_STORE_FLAG_VALUE
3092 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3093 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3094 REAL_VALUE_NEGATIVE (fsfv)))
3095 #endif
3097 x = arg1;
3098 else if (code == EQ
3099 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3100 && code == GE && STORE_FLAG_VALUE == -1)
3101 #ifdef FLOAT_STORE_FLAG_VALUE
3102 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3103 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3104 REAL_VALUE_NEGATIVE (fsfv)))
3105 #endif
3107 x = arg1, reverse_code = 1;
3110 /* ??? We could also check for
3112 (ne (and (eq (...) (const_int 1))) (const_int 0))
3114 and related forms, but let's wait until we see them occurring. */
3116 if (x == 0)
3117 /* Look up ARG1 in the hash table and see if it has an equivalence
3118 that lets us see what is being compared. */
3119 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3120 if (p)
3122 p = p->first_same_value;
3124 /* If what we compare is already known to be constant, that is as
3125 good as it gets.
3126 We need to break the loop in this case, because otherwise we
3127 can have an infinite loop when looking at a reg that is known
3128 to be a constant which is the same as a comparison of a reg
3129 against zero which appears later in the insn stream, which in
3130 turn is constant and the same as the comparison of the first reg
3131 against zero... */
3132 if (p->is_const)
3133 break;
3136 for (; p; p = p->next_same_value)
3138 enum machine_mode inner_mode = GET_MODE (p->exp);
3139 #ifdef FLOAT_STORE_FLAG_VALUE
3140 REAL_VALUE_TYPE fsfv;
3141 #endif
3143 /* If the entry isn't valid, skip it. */
3144 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3145 continue;
3147 if (GET_CODE (p->exp) == COMPARE
3148 /* Another possibility is that this machine has a compare insn
3149 that includes the comparison code. In that case, ARG1 would
3150 be equivalent to a comparison operation that would set ARG1 to
3151 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3152 ORIG_CODE is the actual comparison being done; if it is an EQ,
3153 we must reverse ORIG_CODE. On machine with a negative value
3154 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3155 || ((code == NE
3156 || (code == LT
3157 && GET_MODE_CLASS (inner_mode) == MODE_INT
3158 && (GET_MODE_BITSIZE (inner_mode)
3159 <= HOST_BITS_PER_WIDE_INT)
3160 && (STORE_FLAG_VALUE
3161 & ((HOST_WIDE_INT) 1
3162 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3163 #ifdef FLOAT_STORE_FLAG_VALUE
3164 || (code == LT
3165 && SCALAR_FLOAT_MODE_P (inner_mode)
3166 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3167 REAL_VALUE_NEGATIVE (fsfv)))
3168 #endif
3170 && COMPARISON_P (p->exp)))
3172 x = p->exp;
3173 break;
3175 else if ((code == EQ
3176 || (code == GE
3177 && GET_MODE_CLASS (inner_mode) == MODE_INT
3178 && (GET_MODE_BITSIZE (inner_mode)
3179 <= HOST_BITS_PER_WIDE_INT)
3180 && (STORE_FLAG_VALUE
3181 & ((HOST_WIDE_INT) 1
3182 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3183 #ifdef FLOAT_STORE_FLAG_VALUE
3184 || (code == GE
3185 && SCALAR_FLOAT_MODE_P (inner_mode)
3186 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3187 REAL_VALUE_NEGATIVE (fsfv)))
3188 #endif
3190 && COMPARISON_P (p->exp))
3192 reverse_code = 1;
3193 x = p->exp;
3194 break;
3197 /* If this non-trapping address, e.g. fp + constant, the
3198 equivalent is a better operand since it may let us predict
3199 the value of the comparison. */
3200 else if (!rtx_addr_can_trap_p (p->exp))
3202 arg1 = p->exp;
3203 continue;
3207 /* If we didn't find a useful equivalence for ARG1, we are done.
3208 Otherwise, set up for the next iteration. */
3209 if (x == 0)
3210 break;
3212 /* If we need to reverse the comparison, make sure that that is
3213 possible -- we can't necessarily infer the value of GE from LT
3214 with floating-point operands. */
3215 if (reverse_code)
3217 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3218 if (reversed == UNKNOWN)
3219 break;
3220 else
3221 code = reversed;
3223 else if (COMPARISON_P (x))
3224 code = GET_CODE (x);
3225 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3228 /* Return our results. Return the modes from before fold_rtx
3229 because fold_rtx might produce const_int, and then it's too late. */
3230 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3231 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3233 return code;
3236 /* Fold SUBREG. */
3238 static rtx
3239 fold_rtx_subreg (rtx x, rtx insn)
3241 enum machine_mode mode = GET_MODE (x);
3242 rtx folded_arg0;
3243 rtx const_arg0;
3244 rtx new;
3246 /* See if we previously assigned a constant value to this SUBREG. */
3247 if ((new = lookup_as_function (x, CONST_INT)) != 0
3248 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3249 return new;
3251 /* If this is a paradoxical SUBREG, we have no idea what value the
3252 extra bits would have. However, if the operand is equivalent to
3253 a SUBREG whose operand is the same as our mode, and all the modes
3254 are within a word, we can just use the inner operand because
3255 these SUBREGs just say how to treat the register.
3257 Similarly if we find an integer constant. */
3259 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3261 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3262 struct table_elt *elt;
3264 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3265 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3266 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3267 imode)) != 0)
3268 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3270 if (CONSTANT_P (elt->exp)
3271 && GET_MODE (elt->exp) == VOIDmode)
3272 return elt->exp;
3274 if (GET_CODE (elt->exp) == SUBREG
3275 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3276 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3277 return copy_rtx (SUBREG_REG (elt->exp));
3280 return x;
3283 /* Fold SUBREG_REG. If it changed, see if we can simplify the
3284 SUBREG. We might be able to if the SUBREG is extracting a single
3285 word in an integral mode or extracting the low part. */
3287 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3288 const_arg0 = equiv_constant (folded_arg0);
3289 if (const_arg0)
3290 folded_arg0 = const_arg0;
3292 if (folded_arg0 != SUBREG_REG (x))
3294 new = simplify_subreg (mode, folded_arg0,
3295 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3296 if (new)
3297 return new;
3300 if (REG_P (folded_arg0)
3301 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3303 struct table_elt *elt;
3305 elt = lookup (folded_arg0,
3306 HASH (folded_arg0, GET_MODE (folded_arg0)),
3307 GET_MODE (folded_arg0));
3309 if (elt)
3310 elt = elt->first_same_value;
3312 if (subreg_lowpart_p (x))
3313 /* If this is a narrowing SUBREG and our operand is a REG, see
3314 if we can find an equivalence for REG that is an arithmetic
3315 operation in a wider mode where both operands are
3316 paradoxical SUBREGs from objects of our result mode. In
3317 that case, we couldn-t report an equivalent value for that
3318 operation, since we don't know what the extra bits will be.
3319 But we can find an equivalence for this SUBREG by folding
3320 that operation in the narrow mode. This allows us to fold
3321 arithmetic in narrow modes when the machine only supports
3322 word-sized arithmetic.
3324 Also look for a case where we have a SUBREG whose operand
3325 is the same as our result. If both modes are smaller than
3326 a word, we are simply interpreting a register in different
3327 modes and we can use the inner value. */
3329 for (; elt; elt = elt->next_same_value)
3331 enum rtx_code eltcode = GET_CODE (elt->exp);
3333 /* Just check for unary and binary operations. */
3334 if (UNARY_P (elt->exp)
3335 && eltcode != SIGN_EXTEND
3336 && eltcode != ZERO_EXTEND
3337 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3338 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3339 && (GET_MODE_CLASS (mode)
3340 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3342 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3344 if (!REG_P (op0) && ! CONSTANT_P (op0))
3345 op0 = fold_rtx (op0, NULL_RTX);
3347 op0 = equiv_constant (op0);
3348 if (op0)
3349 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3350 op0, mode);
3352 else if (ARITHMETIC_P (elt->exp)
3353 && eltcode != DIV && eltcode != MOD
3354 && eltcode != UDIV && eltcode != UMOD
3355 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3356 && eltcode != ROTATE && eltcode != ROTATERT
3357 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3358 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3359 == mode))
3360 || CONSTANT_P (XEXP (elt->exp, 0)))
3361 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3362 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3363 == mode))
3364 || CONSTANT_P (XEXP (elt->exp, 1))))
3366 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3367 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3369 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3370 op0 = fold_rtx (op0, NULL_RTX);
3372 if (op0)
3373 op0 = equiv_constant (op0);
3375 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3376 op1 = fold_rtx (op1, NULL_RTX);
3378 if (op1)
3379 op1 = equiv_constant (op1);
3381 /* If we are looking for the low SImode part of
3382 (ashift:DI c (const_int 32)), it doesn't work to
3383 compute that in SImode, because a 32-bit shift in
3384 SImode is unpredictable. We know the value is
3385 0. */
3386 if (op0 && op1
3387 && GET_CODE (elt->exp) == ASHIFT
3388 && GET_CODE (op1) == CONST_INT
3389 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3391 if (INTVAL (op1)
3392 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3393 /* If the count fits in the inner mode's width,
3394 but exceeds the outer mode's width, the value
3395 will get truncated to 0 by the subreg. */
3396 new = CONST0_RTX (mode);
3397 else
3398 /* If the count exceeds even the inner mode's width,
3399 don't fold this expression. */
3400 new = 0;
3402 else if (op0 && op1)
3403 new = simplify_binary_operation (GET_CODE (elt->exp),
3404 mode, op0, op1);
3407 else if (GET_CODE (elt->exp) == SUBREG
3408 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3409 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3410 <= UNITS_PER_WORD)
3411 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3412 new = copy_rtx (SUBREG_REG (elt->exp));
3414 if (new)
3415 return new;
3417 else
3418 /* A SUBREG resulting from a zero extension may fold to zero
3419 if it extracts higher bits than the ZERO_EXTEND's source
3420 bits. FIXME: if combine tried to, er, combine these
3421 instructions, this transformation may be moved to
3422 simplify_subreg. */
3423 for (; elt; elt = elt->next_same_value)
3425 if (GET_CODE (elt->exp) == ZERO_EXTEND
3426 && subreg_lsb (x)
3427 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3428 return CONST0_RTX (mode);
3432 return x;
3435 /* Fold MEM. */
3437 static rtx
3438 fold_rtx_mem (rtx x, rtx insn)
3440 enum machine_mode mode = GET_MODE (x);
3441 rtx new;
3443 /* If we are not actually processing an insn, don't try to find the
3444 best address. Not only don't we care, but we could modify the
3445 MEM in an invalid way since we have no insn to validate
3446 against. */
3447 if (insn != 0)
3448 find_best_addr (insn, &XEXP (x, 0), mode);
3451 /* Even if we don't fold in the insn itself, we can safely do so
3452 here, in hopes of getting a constant. */
3453 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3454 rtx base = 0;
3455 HOST_WIDE_INT offset = 0;
3457 if (REG_P (addr)
3458 && REGNO_QTY_VALID_P (REGNO (addr)))
3460 int addr_q = REG_QTY (REGNO (addr));
3461 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3463 if (GET_MODE (addr) == addr_ent->mode
3464 && addr_ent->const_rtx != NULL_RTX)
3465 addr = addr_ent->const_rtx;
3468 /* Call target hook to avoid the effects of -fpic etc.... */
3469 addr = targetm.delegitimize_address (addr);
3471 /* If address is constant, split it into a base and integer
3472 offset. */
3473 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3474 base = addr;
3475 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3476 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3478 base = XEXP (XEXP (addr, 0), 0);
3479 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3481 else if (GET_CODE (addr) == LO_SUM
3482 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3483 base = XEXP (addr, 1);
3485 /* If this is a constant pool reference, we can fold it into its
3486 constant to allow better value tracking. */
3487 if (base && GET_CODE (base) == SYMBOL_REF
3488 && CONSTANT_POOL_ADDRESS_P (base))
3490 rtx constant = get_pool_constant (base);
3491 enum machine_mode const_mode = get_pool_mode (base);
3492 rtx new;
3494 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3496 constant_pool_entries_cost = COST (constant);
3497 constant_pool_entries_regcost = approx_reg_cost (constant);
3500 /* If we are loading the full constant, we have an
3501 equivalence. */
3502 if (offset == 0 && mode == const_mode)
3503 return constant;
3505 /* If this actually isn't a constant (weird!), we can't do
3506 anything. Otherwise, handle the two most common cases:
3507 extracting a word from a multi-word constant, and
3508 extracting the low-order bits. Other cases don't seem
3509 common enough to worry about. */
3510 if (! CONSTANT_P (constant))
3511 return x;
3513 if (GET_MODE_CLASS (mode) == MODE_INT
3514 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3515 && offset % UNITS_PER_WORD == 0
3516 && (new = operand_subword (constant,
3517 offset / UNITS_PER_WORD,
3518 0, const_mode)) != 0)
3519 return new;
3521 if (((BYTES_BIG_ENDIAN
3522 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3523 || (! BYTES_BIG_ENDIAN && offset == 0))
3524 && (new = gen_lowpart (mode, constant)) != 0)
3525 return new;
3528 /* If this is a reference to a label at a known position in a jump
3529 table, we also know its value. */
3530 if (base && GET_CODE (base) == LABEL_REF)
3532 rtx label = XEXP (base, 0);
3533 rtx table_insn = NEXT_INSN (label);
3535 if (table_insn && JUMP_P (table_insn)
3536 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3538 rtx table = PATTERN (table_insn);
3540 if (offset >= 0
3541 && (offset / GET_MODE_SIZE (GET_MODE (table))
3542 < XVECLEN (table, 0)))
3544 rtx label = XVECEXP
3545 (table, 0, offset / GET_MODE_SIZE (GET_MODE (table)));
3546 rtx set;
3548 /* If we have an insn that loads the label from the
3549 jumptable into a reg, we don't want to set the reg
3550 to the label, because this may cause a reference to
3551 the label to remain after the label is removed in
3552 some very obscure cases (PR middle-end/18628). */
3553 if (!insn)
3554 return label;
3556 set = single_set (insn);
3558 if (! set || SET_SRC (set) != x)
3559 return x;
3561 /* If it's a jump, it's safe to reference the label. */
3562 if (SET_DEST (set) == pc_rtx)
3563 return label;
3565 return x;
3568 if (table_insn && JUMP_P (table_insn)
3569 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3571 rtx table = PATTERN (table_insn);
3573 if (offset >= 0
3574 && (offset / GET_MODE_SIZE (GET_MODE (table))
3575 < XVECLEN (table, 1)))
3577 offset /= GET_MODE_SIZE (GET_MODE (table));
3578 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3579 XEXP (table, 0));
3581 if (GET_MODE (table) != Pmode)
3582 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3584 /* Indicate this is a constant. This isn't a valid
3585 form of CONST, but it will only be used to fold the
3586 next insns and then discarded, so it should be
3587 safe.
3589 Note this expression must be explicitly discarded,
3590 by cse_insn, else it may end up in a REG_EQUAL note
3591 and "escape" to cause problems elsewhere. */
3592 return gen_rtx_CONST (GET_MODE (new), new);
3597 return x;
3601 /* If X is a nontrivial arithmetic operation on an argument
3602 for which a constant value can be determined, return
3603 the result of operating on that value, as a constant.
3604 Otherwise, return X, possibly with one or more operands
3605 modified by recursive calls to this function.
3607 If X is a register whose contents are known, we do NOT
3608 return those contents here. equiv_constant is called to
3609 perform that task.
3611 INSN is the insn that we may be modifying. If it is 0, make a copy
3612 of X before modifying it. */
3614 static rtx
3615 fold_rtx (rtx x, rtx insn)
3617 enum rtx_code code;
3618 enum machine_mode mode;
3619 const char *fmt;
3620 int i;
3621 rtx new = 0;
3622 int copied = 0;
3623 int must_swap = 0;
3625 /* Folded equivalents of first two operands of X. */
3626 rtx folded_arg0;
3627 rtx folded_arg1;
3629 /* Constant equivalents of first three operands of X;
3630 0 when no such equivalent is known. */
3631 rtx const_arg0;
3632 rtx const_arg1;
3633 rtx const_arg2;
3635 /* The mode of the first operand of X. We need this for sign and zero
3636 extends. */
3637 enum machine_mode mode_arg0;
3639 if (x == 0)
3640 return x;
3642 mode = GET_MODE (x);
3643 code = GET_CODE (x);
3644 switch (code)
3646 case CONST:
3647 case CONST_INT:
3648 case CONST_DOUBLE:
3649 case CONST_VECTOR:
3650 case SYMBOL_REF:
3651 case LABEL_REF:
3652 case REG:
3653 case PC:
3654 /* No use simplifying an EXPR_LIST
3655 since they are used only for lists of args
3656 in a function call's REG_EQUAL note. */
3657 case EXPR_LIST:
3658 return x;
3660 #ifdef HAVE_cc0
3661 case CC0:
3662 return prev_insn_cc0;
3663 #endif
3665 case SUBREG:
3666 return fold_rtx_subreg (x, insn);
3668 case NOT:
3669 case NEG:
3670 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3671 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3672 new = lookup_as_function (XEXP (x, 0), code);
3673 if (new)
3674 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3675 break;
3677 case MEM:
3678 return fold_rtx_mem (x, insn);
3680 #ifdef NO_FUNCTION_CSE
3681 case CALL:
3682 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3683 return x;
3684 break;
3685 #endif
3687 case ASM_OPERANDS:
3688 if (insn)
3690 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3691 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3692 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3694 break;
3696 default:
3697 break;
3700 const_arg0 = 0;
3701 const_arg1 = 0;
3702 const_arg2 = 0;
3703 mode_arg0 = VOIDmode;
3705 /* Try folding our operands.
3706 Then see which ones have constant values known. */
3708 fmt = GET_RTX_FORMAT (code);
3709 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3710 if (fmt[i] == 'e')
3712 rtx arg = XEXP (x, i);
3713 rtx folded_arg = arg, const_arg = 0;
3714 enum machine_mode mode_arg = GET_MODE (arg);
3715 rtx cheap_arg, expensive_arg;
3716 rtx replacements[2];
3717 int j;
3718 int old_cost = COST_IN (XEXP (x, i), code);
3720 /* Most arguments are cheap, so handle them specially. */
3721 switch (GET_CODE (arg))
3723 case REG:
3724 /* This is the same as calling equiv_constant; it is duplicated
3725 here for speed. */
3726 if (REGNO_QTY_VALID_P (REGNO (arg)))
3728 int arg_q = REG_QTY (REGNO (arg));
3729 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3731 if (arg_ent->const_rtx != NULL_RTX
3732 && !REG_P (arg_ent->const_rtx)
3733 && GET_CODE (arg_ent->const_rtx) != PLUS)
3734 const_arg
3735 = gen_lowpart (GET_MODE (arg),
3736 arg_ent->const_rtx);
3738 break;
3740 case CONST:
3741 case CONST_INT:
3742 case SYMBOL_REF:
3743 case LABEL_REF:
3744 case CONST_DOUBLE:
3745 case CONST_VECTOR:
3746 const_arg = arg;
3747 break;
3749 #ifdef HAVE_cc0
3750 case CC0:
3751 folded_arg = prev_insn_cc0;
3752 mode_arg = prev_insn_cc0_mode;
3753 const_arg = equiv_constant (folded_arg);
3754 break;
3755 #endif
3757 default:
3758 folded_arg = fold_rtx (arg, insn);
3759 const_arg = equiv_constant (folded_arg);
3762 /* For the first three operands, see if the operand
3763 is constant or equivalent to a constant. */
3764 switch (i)
3766 case 0:
3767 folded_arg0 = folded_arg;
3768 const_arg0 = const_arg;
3769 mode_arg0 = mode_arg;
3770 break;
3771 case 1:
3772 folded_arg1 = folded_arg;
3773 const_arg1 = const_arg;
3774 break;
3775 case 2:
3776 const_arg2 = const_arg;
3777 break;
3780 /* Pick the least expensive of the folded argument and an
3781 equivalent constant argument. */
3782 if (const_arg == 0 || const_arg == folded_arg
3783 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3784 cheap_arg = folded_arg, expensive_arg = const_arg;
3785 else
3786 cheap_arg = const_arg, expensive_arg = folded_arg;
3788 /* Try to replace the operand with the cheapest of the two
3789 possibilities. If it doesn't work and this is either of the first
3790 two operands of a commutative operation, try swapping them.
3791 If THAT fails, try the more expensive, provided it is cheaper
3792 than what is already there. */
3794 if (cheap_arg == XEXP (x, i))
3795 continue;
3797 if (insn == 0 && ! copied)
3799 x = copy_rtx (x);
3800 copied = 1;
3803 /* Order the replacements from cheapest to most expensive. */
3804 replacements[0] = cheap_arg;
3805 replacements[1] = expensive_arg;
3807 for (j = 0; j < 2 && replacements[j]; j++)
3809 int new_cost = COST_IN (replacements[j], code);
3811 /* Stop if what existed before was cheaper. Prefer constants
3812 in the case of a tie. */
3813 if (new_cost > old_cost
3814 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3815 break;
3817 /* It's not safe to substitute the operand of a conversion
3818 operator with a constant, as the conversion's identity
3819 depends upon the mode of its operand. This optimization
3820 is handled by the call to simplify_unary_operation. */
3821 if (GET_RTX_CLASS (code) == RTX_UNARY
3822 && GET_MODE (replacements[j]) != mode_arg0
3823 && (code == ZERO_EXTEND
3824 || code == SIGN_EXTEND
3825 || code == TRUNCATE
3826 || code == FLOAT_TRUNCATE
3827 || code == FLOAT_EXTEND
3828 || code == FLOAT
3829 || code == FIX
3830 || code == UNSIGNED_FLOAT
3831 || code == UNSIGNED_FIX))
3832 continue;
3834 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3835 break;
3837 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3838 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3840 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3841 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3843 if (apply_change_group ())
3845 /* Swap them back to be invalid so that this loop can
3846 continue and flag them to be swapped back later. */
3847 rtx tem;
3849 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3850 XEXP (x, 1) = tem;
3851 must_swap = 1;
3852 break;
3858 else
3860 if (fmt[i] == 'E')
3861 /* Don't try to fold inside of a vector of expressions.
3862 Doing nothing is harmless. */
3866 /* If a commutative operation, place a constant integer as the second
3867 operand unless the first operand is also a constant integer. Otherwise,
3868 place any constant second unless the first operand is also a constant. */
3870 if (COMMUTATIVE_P (x))
3872 if (must_swap
3873 || swap_commutative_operands_p (const_arg0 ? const_arg0
3874 : XEXP (x, 0),
3875 const_arg1 ? const_arg1
3876 : XEXP (x, 1)))
3878 rtx tem = XEXP (x, 0);
3880 if (insn == 0 && ! copied)
3882 x = copy_rtx (x);
3883 copied = 1;
3886 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3887 validate_change (insn, &XEXP (x, 1), tem, 1);
3888 if (apply_change_group ())
3890 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3891 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3896 /* If X is an arithmetic operation, see if we can simplify it. */
3898 switch (GET_RTX_CLASS (code))
3900 case RTX_UNARY:
3902 int is_const = 0;
3904 /* We can't simplify extension ops unless we know the
3905 original mode. */
3906 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3907 && mode_arg0 == VOIDmode)
3908 break;
3910 /* If we had a CONST, strip it off and put it back later if we
3911 fold. */
3912 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3913 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3915 new = simplify_unary_operation (code, mode,
3916 const_arg0 ? const_arg0 : folded_arg0,
3917 mode_arg0);
3918 /* NEG of PLUS could be converted into MINUS, but that causes
3919 expressions of the form
3920 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3921 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3922 FIXME: those ports should be fixed. */
3923 if (new != 0 && is_const
3924 && GET_CODE (new) == PLUS
3925 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3926 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3927 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3928 new = gen_rtx_CONST (mode, new);
3930 break;
3932 case RTX_COMPARE:
3933 case RTX_COMM_COMPARE:
3934 /* See what items are actually being compared and set FOLDED_ARG[01]
3935 to those values and CODE to the actual comparison code. If any are
3936 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3937 do anything if both operands are already known to be constant. */
3939 /* ??? Vector mode comparisons are not supported yet. */
3940 if (VECTOR_MODE_P (mode))
3941 break;
3943 if (const_arg0 == 0 || const_arg1 == 0)
3945 struct table_elt *p0, *p1;
3946 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
3947 enum machine_mode mode_arg1;
3949 #ifdef FLOAT_STORE_FLAG_VALUE
3950 if (SCALAR_FLOAT_MODE_P (mode))
3952 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3953 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3954 false_rtx = CONST0_RTX (mode);
3956 #endif
3958 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3959 &mode_arg0, &mode_arg1);
3961 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3962 what kinds of things are being compared, so we can't do
3963 anything with this comparison. */
3965 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3966 break;
3968 const_arg0 = equiv_constant (folded_arg0);
3969 const_arg1 = equiv_constant (folded_arg1);
3971 /* If we do not now have two constants being compared, see
3972 if we can nevertheless deduce some things about the
3973 comparison. */
3974 if (const_arg0 == 0 || const_arg1 == 0)
3976 if (const_arg1 != NULL)
3978 rtx cheapest_simplification;
3979 int cheapest_cost;
3980 rtx simp_result;
3981 struct table_elt *p;
3983 /* See if we can find an equivalent of folded_arg0
3984 that gets us a cheaper expression, possibly a
3985 constant through simplifications. */
3986 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3987 mode_arg0);
3989 if (p != NULL)
3991 cheapest_simplification = x;
3992 cheapest_cost = COST (x);
3994 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3996 int cost;
3998 /* If the entry isn't valid, skip it. */
3999 if (! exp_equiv_p (p->exp, p->exp, 1, false))
4000 continue;
4002 /* Try to simplify using this equivalence. */
4003 simp_result
4004 = simplify_relational_operation (code, mode,
4005 mode_arg0,
4006 p->exp,
4007 const_arg1);
4009 if (simp_result == NULL)
4010 continue;
4012 cost = COST (simp_result);
4013 if (cost < cheapest_cost)
4015 cheapest_cost = cost;
4016 cheapest_simplification = simp_result;
4020 /* If we have a cheaper expression now, use that
4021 and try folding it further, from the top. */
4022 if (cheapest_simplification != x)
4023 return fold_rtx (cheapest_simplification, insn);
4027 /* Some addresses are known to be nonzero. We don't know
4028 their sign, but equality comparisons are known. */
4029 if (const_arg1 == const0_rtx
4030 && nonzero_address_p (folded_arg0))
4032 if (code == EQ)
4033 return false_rtx;
4034 else if (code == NE)
4035 return true_rtx;
4038 /* See if the two operands are the same. */
4040 if (folded_arg0 == folded_arg1
4041 || (REG_P (folded_arg0)
4042 && REG_P (folded_arg1)
4043 && (REG_QTY (REGNO (folded_arg0))
4044 == REG_QTY (REGNO (folded_arg1))))
4045 || ((p0 = lookup (folded_arg0,
4046 SAFE_HASH (folded_arg0, mode_arg0),
4047 mode_arg0))
4048 && (p1 = lookup (folded_arg1,
4049 SAFE_HASH (folded_arg1, mode_arg0),
4050 mode_arg0))
4051 && p0->first_same_value == p1->first_same_value))
4053 /* Sadly two equal NaNs are not equivalent. */
4054 if (!HONOR_NANS (mode_arg0))
4055 return ((code == EQ || code == LE || code == GE
4056 || code == LEU || code == GEU || code == UNEQ
4057 || code == UNLE || code == UNGE
4058 || code == ORDERED)
4059 ? true_rtx : false_rtx);
4060 /* Take care for the FP compares we can resolve. */
4061 if (code == UNEQ || code == UNLE || code == UNGE)
4062 return true_rtx;
4063 if (code == LTGT || code == LT || code == GT)
4064 return false_rtx;
4067 /* If FOLDED_ARG0 is a register, see if the comparison we are
4068 doing now is either the same as we did before or the reverse
4069 (we only check the reverse if not floating-point). */
4070 else if (REG_P (folded_arg0))
4072 int qty = REG_QTY (REGNO (folded_arg0));
4074 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4076 struct qty_table_elem *ent = &qty_table[qty];
4078 if ((comparison_dominates_p (ent->comparison_code, code)
4079 || (! FLOAT_MODE_P (mode_arg0)
4080 && comparison_dominates_p (ent->comparison_code,
4081 reverse_condition (code))))
4082 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4083 || (const_arg1
4084 && rtx_equal_p (ent->comparison_const,
4085 const_arg1))
4086 || (REG_P (folded_arg1)
4087 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4088 return (comparison_dominates_p (ent->comparison_code, code)
4089 ? true_rtx : false_rtx);
4095 /* If we are comparing against zero, see if the first operand is
4096 equivalent to an IOR with a constant. If so, we may be able to
4097 determine the result of this comparison. */
4099 if (const_arg1 == const0_rtx)
4101 rtx y = lookup_as_function (folded_arg0, IOR);
4102 rtx inner_const;
4104 if (y != 0
4105 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4106 && GET_CODE (inner_const) == CONST_INT
4107 && INTVAL (inner_const) != 0)
4109 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4110 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4111 && (INTVAL (inner_const)
4112 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4113 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4115 #ifdef FLOAT_STORE_FLAG_VALUE
4116 if (SCALAR_FLOAT_MODE_P (mode))
4118 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4119 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4120 false_rtx = CONST0_RTX (mode);
4122 #endif
4124 switch (code)
4126 case EQ:
4127 return false_rtx;
4128 case NE:
4129 return true_rtx;
4130 case LT: case LE:
4131 if (has_sign)
4132 return true_rtx;
4133 break;
4134 case GT: case GE:
4135 if (has_sign)
4136 return false_rtx;
4137 break;
4138 default:
4139 break;
4145 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4146 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4147 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4149 break;
4151 case RTX_BIN_ARITH:
4152 case RTX_COMM_ARITH:
4153 switch (code)
4155 case PLUS:
4156 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4157 with that LABEL_REF as its second operand. If so, the result is
4158 the first operand of that MINUS. This handles switches with an
4159 ADDR_DIFF_VEC table. */
4160 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4162 rtx y
4163 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4164 : lookup_as_function (folded_arg0, MINUS);
4166 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4167 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4168 return XEXP (y, 0);
4170 /* Now try for a CONST of a MINUS like the above. */
4171 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4172 : lookup_as_function (folded_arg0, CONST))) != 0
4173 && GET_CODE (XEXP (y, 0)) == MINUS
4174 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4175 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4176 return XEXP (XEXP (y, 0), 0);
4179 /* Likewise if the operands are in the other order. */
4180 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4182 rtx y
4183 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4184 : lookup_as_function (folded_arg1, MINUS);
4186 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4187 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4188 return XEXP (y, 0);
4190 /* Now try for a CONST of a MINUS like the above. */
4191 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4192 : lookup_as_function (folded_arg1, CONST))) != 0
4193 && GET_CODE (XEXP (y, 0)) == MINUS
4194 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4195 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4196 return XEXP (XEXP (y, 0), 0);
4199 /* If second operand is a register equivalent to a negative
4200 CONST_INT, see if we can find a register equivalent to the
4201 positive constant. Make a MINUS if so. Don't do this for
4202 a non-negative constant since we might then alternate between
4203 choosing positive and negative constants. Having the positive
4204 constant previously-used is the more common case. Be sure
4205 the resulting constant is non-negative; if const_arg1 were
4206 the smallest negative number this would overflow: depending
4207 on the mode, this would either just be the same value (and
4208 hence not save anything) or be incorrect. */
4209 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4210 && INTVAL (const_arg1) < 0
4211 /* This used to test
4213 -INTVAL (const_arg1) >= 0
4215 But The Sun V5.0 compilers mis-compiled that test. So
4216 instead we test for the problematic value in a more direct
4217 manner and hope the Sun compilers get it correct. */
4218 && INTVAL (const_arg1) !=
4219 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4220 && REG_P (folded_arg1))
4222 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4223 struct table_elt *p
4224 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4226 if (p)
4227 for (p = p->first_same_value; p; p = p->next_same_value)
4228 if (REG_P (p->exp))
4229 return simplify_gen_binary (MINUS, mode, folded_arg0,
4230 canon_reg (p->exp, NULL_RTX));
4232 goto from_plus;
4234 case MINUS:
4235 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4236 If so, produce (PLUS Z C2-C). */
4237 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4239 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4240 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4241 return fold_rtx (plus_constant (copy_rtx (y),
4242 -INTVAL (const_arg1)),
4243 NULL_RTX);
4246 /* Fall through. */
4248 from_plus:
4249 case SMIN: case SMAX: case UMIN: case UMAX:
4250 case IOR: case AND: case XOR:
4251 case MULT:
4252 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4253 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4254 is known to be of similar form, we may be able to replace the
4255 operation with a combined operation. This may eliminate the
4256 intermediate operation if every use is simplified in this way.
4257 Note that the similar optimization done by combine.c only works
4258 if the intermediate operation's result has only one reference. */
4260 if (REG_P (folded_arg0)
4261 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4263 int is_shift
4264 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4265 rtx y = lookup_as_function (folded_arg0, code);
4266 rtx inner_const;
4267 enum rtx_code associate_code;
4268 rtx new_const;
4270 if (is_shift
4271 && (INTVAL (const_arg1) >= GET_MODE_BITSIZE (mode)
4272 || INTVAL (const_arg1) < 0))
4274 if (SHIFT_COUNT_TRUNCATED)
4275 const_arg1 = GEN_INT (INTVAL (const_arg1)
4276 & (GET_MODE_BITSIZE (mode) - 1));
4277 else
4278 break;
4281 if (y == 0)
4282 break;
4283 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
4284 if (!inner_const || GET_CODE (inner_const) != CONST_INT)
4285 break;
4287 /* If we have compiled a statement like
4288 "if (x == (x & mask1))", and now are looking at
4289 "x & mask2", we will have a case where the first operand
4290 of Y is the same as our first operand. Unless we detect
4291 this case, an infinite loop will result. */
4292 if (XEXP (y, 0) == folded_arg0)
4293 break;
4295 /* Don't associate these operations if they are a PLUS with the
4296 same constant and it is a power of two. These might be doable
4297 with a pre- or post-increment. Similarly for two subtracts of
4298 identical powers of two with post decrement. */
4300 if (code == PLUS && const_arg1 == inner_const
4301 && ((HAVE_PRE_INCREMENT
4302 && exact_log2 (INTVAL (const_arg1)) >= 0)
4303 || (HAVE_POST_INCREMENT
4304 && exact_log2 (INTVAL (const_arg1)) >= 0)
4305 || (HAVE_PRE_DECREMENT
4306 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4307 || (HAVE_POST_DECREMENT
4308 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4309 break;
4311 if (is_shift
4312 && (INTVAL (inner_const) >= GET_MODE_BITSIZE (mode)
4313 || INTVAL (inner_const) < 0))
4315 if (SHIFT_COUNT_TRUNCATED)
4316 inner_const = GEN_INT (INTVAL (inner_const)
4317 & (GET_MODE_BITSIZE (mode) - 1));
4318 else
4319 break;
4322 /* Compute the code used to compose the constants. For example,
4323 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4325 associate_code = (is_shift || code == MINUS ? PLUS : code);
4327 new_const = simplify_binary_operation (associate_code, mode,
4328 const_arg1, inner_const);
4330 if (new_const == 0)
4331 break;
4333 /* If we are associating shift operations, don't let this
4334 produce a shift of the size of the object or larger.
4335 This could occur when we follow a sign-extend by a right
4336 shift on a machine that does a sign-extend as a pair
4337 of shifts. */
4339 if (is_shift
4340 && GET_CODE (new_const) == CONST_INT
4341 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4343 /* As an exception, we can turn an ASHIFTRT of this
4344 form into a shift of the number of bits - 1. */
4345 if (code == ASHIFTRT)
4346 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4347 else if (!side_effects_p (XEXP (y, 0)))
4348 return CONST0_RTX (mode);
4349 else
4350 break;
4353 y = copy_rtx (XEXP (y, 0));
4355 /* If Y contains our first operand (the most common way this
4356 can happen is if Y is a MEM), we would do into an infinite
4357 loop if we tried to fold it. So don't in that case. */
4359 if (! reg_mentioned_p (folded_arg0, y))
4360 y = fold_rtx (y, insn);
4362 return simplify_gen_binary (code, mode, y, new_const);
4364 break;
4366 case DIV: case UDIV:
4367 /* ??? The associative optimization performed immediately above is
4368 also possible for DIV and UDIV using associate_code of MULT.
4369 However, we would need extra code to verify that the
4370 multiplication does not overflow, that is, there is no overflow
4371 in the calculation of new_const. */
4372 break;
4374 default:
4375 break;
4378 new = simplify_binary_operation (code, mode,
4379 const_arg0 ? const_arg0 : folded_arg0,
4380 const_arg1 ? const_arg1 : folded_arg1);
4381 break;
4383 case RTX_OBJ:
4384 /* (lo_sum (high X) X) is simply X. */
4385 if (code == LO_SUM && const_arg0 != 0
4386 && GET_CODE (const_arg0) == HIGH
4387 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4388 return const_arg1;
4389 break;
4391 case RTX_TERNARY:
4392 case RTX_BITFIELD_OPS:
4393 new = simplify_ternary_operation (code, mode, mode_arg0,
4394 const_arg0 ? const_arg0 : folded_arg0,
4395 const_arg1 ? const_arg1 : folded_arg1,
4396 const_arg2 ? const_arg2 : XEXP (x, 2));
4397 break;
4399 default:
4400 break;
4403 return new ? new : x;
4406 /* Return a constant value currently equivalent to X.
4407 Return 0 if we don't know one. */
4409 static rtx
4410 equiv_constant (rtx x)
4412 if (REG_P (x)
4413 && REGNO_QTY_VALID_P (REGNO (x)))
4415 int x_q = REG_QTY (REGNO (x));
4416 struct qty_table_elem *x_ent = &qty_table[x_q];
4418 if (x_ent->const_rtx)
4419 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4422 if (x == 0 || CONSTANT_P (x))
4423 return x;
4425 /* If X is a MEM, try to fold it outside the context of any insn to see if
4426 it might be equivalent to a constant. That handles the case where it
4427 is a constant-pool reference. Then try to look it up in the hash table
4428 in case it is something whose value we have seen before. */
4430 if (MEM_P (x))
4432 struct table_elt *elt;
4434 x = fold_rtx (x, NULL_RTX);
4435 if (CONSTANT_P (x))
4436 return x;
4438 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4439 if (elt == 0)
4440 return 0;
4442 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4443 if (elt->is_const && CONSTANT_P (elt->exp))
4444 return elt->exp;
4447 return 0;
4450 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4451 branch. It will be zero if not.
4453 In certain cases, this can cause us to add an equivalence. For example,
4454 if we are following the taken case of
4455 if (i == 2)
4456 we can add the fact that `i' and '2' are now equivalent.
4458 In any case, we can record that this comparison was passed. If the same
4459 comparison is seen later, we will know its value. */
4461 static void
4462 record_jump_equiv (rtx insn, int taken)
4464 int cond_known_true;
4465 rtx op0, op1;
4466 rtx set;
4467 enum machine_mode mode, mode0, mode1;
4468 int reversed_nonequality = 0;
4469 enum rtx_code code;
4471 /* Ensure this is the right kind of insn. */
4472 if (! any_condjump_p (insn))
4473 return;
4474 set = pc_set (insn);
4476 /* See if this jump condition is known true or false. */
4477 if (taken)
4478 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4479 else
4480 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4482 /* Get the type of comparison being done and the operands being compared.
4483 If we had to reverse a non-equality condition, record that fact so we
4484 know that it isn't valid for floating-point. */
4485 code = GET_CODE (XEXP (SET_SRC (set), 0));
4486 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4487 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4489 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4490 if (! cond_known_true)
4492 code = reversed_comparison_code_parts (code, op0, op1, insn);
4494 /* Don't remember if we can't find the inverse. */
4495 if (code == UNKNOWN)
4496 return;
4499 /* The mode is the mode of the non-constant. */
4500 mode = mode0;
4501 if (mode1 != VOIDmode)
4502 mode = mode1;
4504 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4507 /* Yet another form of subreg creation. In this case, we want something in
4508 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4510 static rtx
4511 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4513 enum machine_mode op_mode = GET_MODE (op);
4514 if (op_mode == mode || op_mode == VOIDmode)
4515 return op;
4516 return lowpart_subreg (mode, op, op_mode);
4519 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4520 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4521 Make any useful entries we can with that information. Called from
4522 above function and called recursively. */
4524 static void
4525 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4526 rtx op1, int reversed_nonequality)
4528 unsigned op0_hash, op1_hash;
4529 int op0_in_memory, op1_in_memory;
4530 struct table_elt *op0_elt, *op1_elt;
4532 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4533 we know that they are also equal in the smaller mode (this is also
4534 true for all smaller modes whether or not there is a SUBREG, but
4535 is not worth testing for with no SUBREG). */
4537 /* Note that GET_MODE (op0) may not equal MODE. */
4538 if (code == EQ && GET_CODE (op0) == SUBREG
4539 && (GET_MODE_SIZE (GET_MODE (op0))
4540 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4542 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4543 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4544 if (tem)
4545 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4546 reversed_nonequality);
4549 if (code == EQ && GET_CODE (op1) == SUBREG
4550 && (GET_MODE_SIZE (GET_MODE (op1))
4551 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4553 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4554 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4555 if (tem)
4556 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4557 reversed_nonequality);
4560 /* Similarly, if this is an NE comparison, and either is a SUBREG
4561 making a smaller mode, we know the whole thing is also NE. */
4563 /* Note that GET_MODE (op0) may not equal MODE;
4564 if we test MODE instead, we can get an infinite recursion
4565 alternating between two modes each wider than MODE. */
4567 if (code == NE && GET_CODE (op0) == SUBREG
4568 && subreg_lowpart_p (op0)
4569 && (GET_MODE_SIZE (GET_MODE (op0))
4570 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4572 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4573 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4574 if (tem)
4575 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4576 reversed_nonequality);
4579 if (code == NE && GET_CODE (op1) == SUBREG
4580 && subreg_lowpart_p (op1)
4581 && (GET_MODE_SIZE (GET_MODE (op1))
4582 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4584 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4585 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4586 if (tem)
4587 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4588 reversed_nonequality);
4591 /* Hash both operands. */
4593 do_not_record = 0;
4594 hash_arg_in_memory = 0;
4595 op0_hash = HASH (op0, mode);
4596 op0_in_memory = hash_arg_in_memory;
4598 if (do_not_record)
4599 return;
4601 do_not_record = 0;
4602 hash_arg_in_memory = 0;
4603 op1_hash = HASH (op1, mode);
4604 op1_in_memory = hash_arg_in_memory;
4606 if (do_not_record)
4607 return;
4609 /* Look up both operands. */
4610 op0_elt = lookup (op0, op0_hash, mode);
4611 op1_elt = lookup (op1, op1_hash, mode);
4613 /* If both operands are already equivalent or if they are not in the
4614 table but are identical, do nothing. */
4615 if ((op0_elt != 0 && op1_elt != 0
4616 && op0_elt->first_same_value == op1_elt->first_same_value)
4617 || op0 == op1 || rtx_equal_p (op0, op1))
4618 return;
4620 /* If we aren't setting two things equal all we can do is save this
4621 comparison. Similarly if this is floating-point. In the latter
4622 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4623 If we record the equality, we might inadvertently delete code
4624 whose intent was to change -0 to +0. */
4626 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4628 struct qty_table_elem *ent;
4629 int qty;
4631 /* If we reversed a floating-point comparison, if OP0 is not a
4632 register, or if OP1 is neither a register or constant, we can't
4633 do anything. */
4635 if (!REG_P (op1))
4636 op1 = equiv_constant (op1);
4638 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4639 || !REG_P (op0) || op1 == 0)
4640 return;
4642 /* Put OP0 in the hash table if it isn't already. This gives it a
4643 new quantity number. */
4644 if (op0_elt == 0)
4646 if (insert_regs (op0, NULL, 0))
4648 rehash_using_reg (op0);
4649 op0_hash = HASH (op0, mode);
4651 /* If OP0 is contained in OP1, this changes its hash code
4652 as well. Faster to rehash than to check, except
4653 for the simple case of a constant. */
4654 if (! CONSTANT_P (op1))
4655 op1_hash = HASH (op1,mode);
4658 op0_elt = insert (op0, NULL, op0_hash, mode);
4659 op0_elt->in_memory = op0_in_memory;
4662 qty = REG_QTY (REGNO (op0));
4663 ent = &qty_table[qty];
4665 ent->comparison_code = code;
4666 if (REG_P (op1))
4668 /* Look it up again--in case op0 and op1 are the same. */
4669 op1_elt = lookup (op1, op1_hash, mode);
4671 /* Put OP1 in the hash table so it gets a new quantity number. */
4672 if (op1_elt == 0)
4674 if (insert_regs (op1, NULL, 0))
4676 rehash_using_reg (op1);
4677 op1_hash = HASH (op1, mode);
4680 op1_elt = insert (op1, NULL, op1_hash, mode);
4681 op1_elt->in_memory = op1_in_memory;
4684 ent->comparison_const = NULL_RTX;
4685 ent->comparison_qty = REG_QTY (REGNO (op1));
4687 else
4689 ent->comparison_const = op1;
4690 ent->comparison_qty = -1;
4693 return;
4696 /* If either side is still missing an equivalence, make it now,
4697 then merge the equivalences. */
4699 if (op0_elt == 0)
4701 if (insert_regs (op0, NULL, 0))
4703 rehash_using_reg (op0);
4704 op0_hash = HASH (op0, mode);
4707 op0_elt = insert (op0, NULL, op0_hash, mode);
4708 op0_elt->in_memory = op0_in_memory;
4711 if (op1_elt == 0)
4713 if (insert_regs (op1, NULL, 0))
4715 rehash_using_reg (op1);
4716 op1_hash = HASH (op1, mode);
4719 op1_elt = insert (op1, NULL, op1_hash, mode);
4720 op1_elt->in_memory = op1_in_memory;
4723 merge_equiv_classes (op0_elt, op1_elt);
4726 /* CSE processing for one instruction.
4727 First simplify sources and addresses of all assignments
4728 in the instruction, using previously-computed equivalents values.
4729 Then install the new sources and destinations in the table
4730 of available values.
4732 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4733 the insn. It means that INSN is inside libcall block. In this
4734 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4736 /* Data on one SET contained in the instruction. */
4738 struct set
4740 /* The SET rtx itself. */
4741 rtx rtl;
4742 /* The SET_SRC of the rtx (the original value, if it is changing). */
4743 rtx src;
4744 /* The hash-table element for the SET_SRC of the SET. */
4745 struct table_elt *src_elt;
4746 /* Hash value for the SET_SRC. */
4747 unsigned src_hash;
4748 /* Hash value for the SET_DEST. */
4749 unsigned dest_hash;
4750 /* The SET_DEST, with SUBREG, etc., stripped. */
4751 rtx inner_dest;
4752 /* Nonzero if the SET_SRC is in memory. */
4753 char src_in_memory;
4754 /* Nonzero if the SET_SRC contains something
4755 whose value cannot be predicted and understood. */
4756 char src_volatile;
4757 /* Original machine mode, in case it becomes a CONST_INT.
4758 The size of this field should match the size of the mode
4759 field of struct rtx_def (see rtl.h). */
4760 ENUM_BITFIELD(machine_mode) mode : 8;
4761 /* A constant equivalent for SET_SRC, if any. */
4762 rtx src_const;
4763 /* Original SET_SRC value used for libcall notes. */
4764 rtx orig_src;
4765 /* Hash value of constant equivalent for SET_SRC. */
4766 unsigned src_const_hash;
4767 /* Table entry for constant equivalent for SET_SRC, if any. */
4768 struct table_elt *src_const_elt;
4769 /* Table entry for the destination address. */
4770 struct table_elt *dest_addr_elt;
4773 static void
4774 cse_insn (rtx insn, rtx libcall_insn)
4776 rtx x = PATTERN (insn);
4777 int i;
4778 rtx tem;
4779 int n_sets = 0;
4781 #ifdef HAVE_cc0
4782 /* Records what this insn does to set CC0. */
4783 rtx this_insn_cc0 = 0;
4784 enum machine_mode this_insn_cc0_mode = VOIDmode;
4785 #endif
4787 rtx src_eqv = 0;
4788 struct table_elt *src_eqv_elt = 0;
4789 int src_eqv_volatile = 0;
4790 int src_eqv_in_memory = 0;
4791 unsigned src_eqv_hash = 0;
4793 struct set *sets = (struct set *) 0;
4795 this_insn = insn;
4797 /* Find all the SETs and CLOBBERs in this instruction.
4798 Record all the SETs in the array `set' and count them.
4799 Also determine whether there is a CLOBBER that invalidates
4800 all memory references, or all references at varying addresses. */
4802 if (CALL_P (insn))
4804 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4806 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4807 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4808 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4812 if (GET_CODE (x) == SET)
4814 sets = alloca (sizeof (struct set));
4815 sets[0].rtl = x;
4817 /* Ignore SETs that are unconditional jumps.
4818 They never need cse processing, so this does not hurt.
4819 The reason is not efficiency but rather
4820 so that we can test at the end for instructions
4821 that have been simplified to unconditional jumps
4822 and not be misled by unchanged instructions
4823 that were unconditional jumps to begin with. */
4824 if (SET_DEST (x) == pc_rtx
4825 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4828 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4829 The hard function value register is used only once, to copy to
4830 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4831 Ensure we invalidate the destination register. On the 80386 no
4832 other code would invalidate it since it is a fixed_reg.
4833 We need not check the return of apply_change_group; see canon_reg. */
4835 else if (GET_CODE (SET_SRC (x)) == CALL)
4837 canon_reg (SET_SRC (x), insn);
4838 apply_change_group ();
4839 fold_rtx (SET_SRC (x), insn);
4840 invalidate (SET_DEST (x), VOIDmode);
4842 else
4843 n_sets = 1;
4845 else if (GET_CODE (x) == PARALLEL)
4847 int lim = XVECLEN (x, 0);
4849 sets = alloca (lim * sizeof (struct set));
4851 /* Find all regs explicitly clobbered in this insn,
4852 and ensure they are not replaced with any other regs
4853 elsewhere in this insn.
4854 When a reg that is clobbered is also used for input,
4855 we should presume that that is for a reason,
4856 and we should not substitute some other register
4857 which is not supposed to be clobbered.
4858 Therefore, this loop cannot be merged into the one below
4859 because a CALL may precede a CLOBBER and refer to the
4860 value clobbered. We must not let a canonicalization do
4861 anything in that case. */
4862 for (i = 0; i < lim; i++)
4864 rtx y = XVECEXP (x, 0, i);
4865 if (GET_CODE (y) == CLOBBER)
4867 rtx clobbered = XEXP (y, 0);
4869 if (REG_P (clobbered)
4870 || GET_CODE (clobbered) == SUBREG)
4871 invalidate (clobbered, VOIDmode);
4872 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4873 || GET_CODE (clobbered) == ZERO_EXTRACT)
4874 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4878 for (i = 0; i < lim; i++)
4880 rtx y = XVECEXP (x, 0, i);
4881 if (GET_CODE (y) == SET)
4883 /* As above, we ignore unconditional jumps and call-insns and
4884 ignore the result of apply_change_group. */
4885 if (GET_CODE (SET_SRC (y)) == CALL)
4887 canon_reg (SET_SRC (y), insn);
4888 apply_change_group ();
4889 fold_rtx (SET_SRC (y), insn);
4890 invalidate (SET_DEST (y), VOIDmode);
4892 else if (SET_DEST (y) == pc_rtx
4893 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4895 else
4896 sets[n_sets++].rtl = y;
4898 else if (GET_CODE (y) == CLOBBER)
4900 /* If we clobber memory, canon the address.
4901 This does nothing when a register is clobbered
4902 because we have already invalidated the reg. */
4903 if (MEM_P (XEXP (y, 0)))
4904 canon_reg (XEXP (y, 0), NULL_RTX);
4906 else if (GET_CODE (y) == USE
4907 && ! (REG_P (XEXP (y, 0))
4908 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4909 canon_reg (y, NULL_RTX);
4910 else if (GET_CODE (y) == CALL)
4912 /* The result of apply_change_group can be ignored; see
4913 canon_reg. */
4914 canon_reg (y, insn);
4915 apply_change_group ();
4916 fold_rtx (y, insn);
4920 else if (GET_CODE (x) == CLOBBER)
4922 if (MEM_P (XEXP (x, 0)))
4923 canon_reg (XEXP (x, 0), NULL_RTX);
4926 /* Canonicalize a USE of a pseudo register or memory location. */
4927 else if (GET_CODE (x) == USE
4928 && ! (REG_P (XEXP (x, 0))
4929 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4930 canon_reg (XEXP (x, 0), NULL_RTX);
4931 else if (GET_CODE (x) == CALL)
4933 /* The result of apply_change_group can be ignored; see canon_reg. */
4934 canon_reg (x, insn);
4935 apply_change_group ();
4936 fold_rtx (x, insn);
4939 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4940 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4941 is handled specially for this case, and if it isn't set, then there will
4942 be no equivalence for the destination. */
4943 if (n_sets == 1 && REG_NOTES (insn) != 0
4944 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4945 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4946 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4948 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
4949 XEXP (tem, 0) = src_eqv;
4952 /* Canonicalize sources and addresses of destinations.
4953 We do this in a separate pass to avoid problems when a MATCH_DUP is
4954 present in the insn pattern. In that case, we want to ensure that
4955 we don't break the duplicate nature of the pattern. So we will replace
4956 both operands at the same time. Otherwise, we would fail to find an
4957 equivalent substitution in the loop calling validate_change below.
4959 We used to suppress canonicalization of DEST if it appears in SRC,
4960 but we don't do this any more. */
4962 for (i = 0; i < n_sets; i++)
4964 rtx dest = SET_DEST (sets[i].rtl);
4965 rtx src = SET_SRC (sets[i].rtl);
4966 rtx new = canon_reg (src, insn);
4968 sets[i].orig_src = src;
4969 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
4971 if (GET_CODE (dest) == ZERO_EXTRACT)
4973 validate_change (insn, &XEXP (dest, 1),
4974 canon_reg (XEXP (dest, 1), insn), 1);
4975 validate_change (insn, &XEXP (dest, 2),
4976 canon_reg (XEXP (dest, 2), insn), 1);
4979 while (GET_CODE (dest) == SUBREG
4980 || GET_CODE (dest) == ZERO_EXTRACT
4981 || GET_CODE (dest) == STRICT_LOW_PART)
4982 dest = XEXP (dest, 0);
4984 if (MEM_P (dest))
4985 canon_reg (dest, insn);
4988 /* Now that we have done all the replacements, we can apply the change
4989 group and see if they all work. Note that this will cause some
4990 canonicalizations that would have worked individually not to be applied
4991 because some other canonicalization didn't work, but this should not
4992 occur often.
4994 The result of apply_change_group can be ignored; see canon_reg. */
4996 apply_change_group ();
4998 /* Set sets[i].src_elt to the class each source belongs to.
4999 Detect assignments from or to volatile things
5000 and set set[i] to zero so they will be ignored
5001 in the rest of this function.
5003 Nothing in this loop changes the hash table or the register chains. */
5005 for (i = 0; i < n_sets; i++)
5007 rtx src, dest;
5008 rtx src_folded;
5009 struct table_elt *elt = 0, *p;
5010 enum machine_mode mode;
5011 rtx src_eqv_here;
5012 rtx src_const = 0;
5013 rtx src_related = 0;
5014 struct table_elt *src_const_elt = 0;
5015 int src_cost = MAX_COST;
5016 int src_eqv_cost = MAX_COST;
5017 int src_folded_cost = MAX_COST;
5018 int src_related_cost = MAX_COST;
5019 int src_elt_cost = MAX_COST;
5020 int src_regcost = MAX_COST;
5021 int src_eqv_regcost = MAX_COST;
5022 int src_folded_regcost = MAX_COST;
5023 int src_related_regcost = MAX_COST;
5024 int src_elt_regcost = MAX_COST;
5025 /* Set nonzero if we need to call force_const_mem on with the
5026 contents of src_folded before using it. */
5027 int src_folded_force_flag = 0;
5029 dest = SET_DEST (sets[i].rtl);
5030 src = SET_SRC (sets[i].rtl);
5032 /* If SRC is a constant that has no machine mode,
5033 hash it with the destination's machine mode.
5034 This way we can keep different modes separate. */
5036 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5037 sets[i].mode = mode;
5039 if (src_eqv)
5041 enum machine_mode eqvmode = mode;
5042 if (GET_CODE (dest) == STRICT_LOW_PART)
5043 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5044 do_not_record = 0;
5045 hash_arg_in_memory = 0;
5046 src_eqv_hash = HASH (src_eqv, eqvmode);
5048 /* Find the equivalence class for the equivalent expression. */
5050 if (!do_not_record)
5051 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
5053 src_eqv_volatile = do_not_record;
5054 src_eqv_in_memory = hash_arg_in_memory;
5057 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
5058 value of the INNER register, not the destination. So it is not
5059 a valid substitution for the source. But save it for later. */
5060 if (GET_CODE (dest) == STRICT_LOW_PART)
5061 src_eqv_here = 0;
5062 else
5063 src_eqv_here = src_eqv;
5065 /* Simplify and foldable subexpressions in SRC. Then get the fully-
5066 simplified result, which may not necessarily be valid. */
5067 src_folded = fold_rtx (src, insn);
5069 #if 0
5070 /* ??? This caused bad code to be generated for the m68k port with -O2.
5071 Suppose src is (CONST_INT -1), and that after truncation src_folded
5072 is (CONST_INT 3). Suppose src_folded is then used for src_const.
5073 At the end we will add src and src_const to the same equivalence
5074 class. We now have 3 and -1 on the same equivalence class. This
5075 causes later instructions to be mis-optimized. */
5076 /* If storing a constant in a bitfield, pre-truncate the constant
5077 so we will be able to record it later. */
5078 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5080 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5082 if (GET_CODE (src) == CONST_INT
5083 && GET_CODE (width) == CONST_INT
5084 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5085 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5086 src_folded
5087 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5088 << INTVAL (width)) - 1));
5090 #endif
5092 /* Compute SRC's hash code, and also notice if it
5093 should not be recorded at all. In that case,
5094 prevent any further processing of this assignment. */
5095 do_not_record = 0;
5096 hash_arg_in_memory = 0;
5098 sets[i].src = src;
5099 sets[i].src_hash = HASH (src, mode);
5100 sets[i].src_volatile = do_not_record;
5101 sets[i].src_in_memory = hash_arg_in_memory;
5103 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5104 a pseudo, do not record SRC. Using SRC as a replacement for
5105 anything else will be incorrect in that situation. Note that
5106 this usually occurs only for stack slots, in which case all the
5107 RTL would be referring to SRC, so we don't lose any optimization
5108 opportunities by not having SRC in the hash table. */
5110 if (MEM_P (src)
5111 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5112 && REG_P (dest)
5113 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5114 sets[i].src_volatile = 1;
5116 #if 0
5117 /* It is no longer clear why we used to do this, but it doesn't
5118 appear to still be needed. So let's try without it since this
5119 code hurts cse'ing widened ops. */
5120 /* If source is a paradoxical subreg (such as QI treated as an SI),
5121 treat it as volatile. It may do the work of an SI in one context
5122 where the extra bits are not being used, but cannot replace an SI
5123 in general. */
5124 if (GET_CODE (src) == SUBREG
5125 && (GET_MODE_SIZE (GET_MODE (src))
5126 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5127 sets[i].src_volatile = 1;
5128 #endif
5130 /* Locate all possible equivalent forms for SRC. Try to replace
5131 SRC in the insn with each cheaper equivalent.
5133 We have the following types of equivalents: SRC itself, a folded
5134 version, a value given in a REG_EQUAL note, or a value related
5135 to a constant.
5137 Each of these equivalents may be part of an additional class
5138 of equivalents (if more than one is in the table, they must be in
5139 the same class; we check for this).
5141 If the source is volatile, we don't do any table lookups.
5143 We note any constant equivalent for possible later use in a
5144 REG_NOTE. */
5146 if (!sets[i].src_volatile)
5147 elt = lookup (src, sets[i].src_hash, mode);
5149 sets[i].src_elt = elt;
5151 if (elt && src_eqv_here && src_eqv_elt)
5153 if (elt->first_same_value != src_eqv_elt->first_same_value)
5155 /* The REG_EQUAL is indicating that two formerly distinct
5156 classes are now equivalent. So merge them. */
5157 merge_equiv_classes (elt, src_eqv_elt);
5158 src_eqv_hash = HASH (src_eqv, elt->mode);
5159 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5162 src_eqv_here = 0;
5165 else if (src_eqv_elt)
5166 elt = src_eqv_elt;
5168 /* Try to find a constant somewhere and record it in `src_const'.
5169 Record its table element, if any, in `src_const_elt'. Look in
5170 any known equivalences first. (If the constant is not in the
5171 table, also set `sets[i].src_const_hash'). */
5172 if (elt)
5173 for (p = elt->first_same_value; p; p = p->next_same_value)
5174 if (p->is_const)
5176 src_const = p->exp;
5177 src_const_elt = elt;
5178 break;
5181 if (src_const == 0
5182 && (CONSTANT_P (src_folded)
5183 /* Consider (minus (label_ref L1) (label_ref L2)) as
5184 "constant" here so we will record it. This allows us
5185 to fold switch statements when an ADDR_DIFF_VEC is used. */
5186 || (GET_CODE (src_folded) == MINUS
5187 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5188 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5189 src_const = src_folded, src_const_elt = elt;
5190 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5191 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5193 /* If we don't know if the constant is in the table, get its
5194 hash code and look it up. */
5195 if (src_const && src_const_elt == 0)
5197 sets[i].src_const_hash = HASH (src_const, mode);
5198 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5201 sets[i].src_const = src_const;
5202 sets[i].src_const_elt = src_const_elt;
5204 /* If the constant and our source are both in the table, mark them as
5205 equivalent. Otherwise, if a constant is in the table but the source
5206 isn't, set ELT to it. */
5207 if (src_const_elt && elt
5208 && src_const_elt->first_same_value != elt->first_same_value)
5209 merge_equiv_classes (elt, src_const_elt);
5210 else if (src_const_elt && elt == 0)
5211 elt = src_const_elt;
5213 /* See if there is a register linearly related to a constant
5214 equivalent of SRC. */
5215 if (src_const
5216 && (GET_CODE (src_const) == CONST
5217 || (src_const_elt && src_const_elt->related_value != 0)))
5219 src_related = use_related_value (src_const, src_const_elt);
5220 if (src_related)
5222 struct table_elt *src_related_elt
5223 = lookup (src_related, HASH (src_related, mode), mode);
5224 if (src_related_elt && elt)
5226 if (elt->first_same_value
5227 != src_related_elt->first_same_value)
5228 /* This can occur when we previously saw a CONST
5229 involving a SYMBOL_REF and then see the SYMBOL_REF
5230 twice. Merge the involved classes. */
5231 merge_equiv_classes (elt, src_related_elt);
5233 src_related = 0;
5234 src_related_elt = 0;
5236 else if (src_related_elt && elt == 0)
5237 elt = src_related_elt;
5241 /* See if we have a CONST_INT that is already in a register in a
5242 wider mode. */
5244 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5245 && GET_MODE_CLASS (mode) == MODE_INT
5246 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5248 enum machine_mode wider_mode;
5250 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5251 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5252 && src_related == 0;
5253 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5255 struct table_elt *const_elt
5256 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5258 if (const_elt == 0)
5259 continue;
5261 for (const_elt = const_elt->first_same_value;
5262 const_elt; const_elt = const_elt->next_same_value)
5263 if (REG_P (const_elt->exp))
5265 src_related = gen_lowpart (mode,
5266 const_elt->exp);
5267 break;
5272 /* Another possibility is that we have an AND with a constant in
5273 a mode narrower than a word. If so, it might have been generated
5274 as part of an "if" which would narrow the AND. If we already
5275 have done the AND in a wider mode, we can use a SUBREG of that
5276 value. */
5278 if (flag_expensive_optimizations && ! src_related
5279 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5280 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5282 enum machine_mode tmode;
5283 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5285 for (tmode = GET_MODE_WIDER_MODE (mode);
5286 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5287 tmode = GET_MODE_WIDER_MODE (tmode))
5289 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5290 struct table_elt *larger_elt;
5292 if (inner)
5294 PUT_MODE (new_and, tmode);
5295 XEXP (new_and, 0) = inner;
5296 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5297 if (larger_elt == 0)
5298 continue;
5300 for (larger_elt = larger_elt->first_same_value;
5301 larger_elt; larger_elt = larger_elt->next_same_value)
5302 if (REG_P (larger_elt->exp))
5304 src_related
5305 = gen_lowpart (mode, larger_elt->exp);
5306 break;
5309 if (src_related)
5310 break;
5315 #ifdef LOAD_EXTEND_OP
5316 /* See if a MEM has already been loaded with a widening operation;
5317 if it has, we can use a subreg of that. Many CISC machines
5318 also have such operations, but this is only likely to be
5319 beneficial on these machines. */
5321 if (flag_expensive_optimizations && src_related == 0
5322 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5323 && GET_MODE_CLASS (mode) == MODE_INT
5324 && MEM_P (src) && ! do_not_record
5325 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5327 struct rtx_def memory_extend_buf;
5328 rtx memory_extend_rtx = &memory_extend_buf;
5329 enum machine_mode tmode;
5331 /* Set what we are trying to extend and the operation it might
5332 have been extended with. */
5333 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5334 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5335 XEXP (memory_extend_rtx, 0) = src;
5337 for (tmode = GET_MODE_WIDER_MODE (mode);
5338 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5339 tmode = GET_MODE_WIDER_MODE (tmode))
5341 struct table_elt *larger_elt;
5343 PUT_MODE (memory_extend_rtx, tmode);
5344 larger_elt = lookup (memory_extend_rtx,
5345 HASH (memory_extend_rtx, tmode), tmode);
5346 if (larger_elt == 0)
5347 continue;
5349 for (larger_elt = larger_elt->first_same_value;
5350 larger_elt; larger_elt = larger_elt->next_same_value)
5351 if (REG_P (larger_elt->exp))
5353 src_related = gen_lowpart (mode,
5354 larger_elt->exp);
5355 break;
5358 if (src_related)
5359 break;
5362 #endif /* LOAD_EXTEND_OP */
5364 if (src == src_folded)
5365 src_folded = 0;
5367 /* At this point, ELT, if nonzero, points to a class of expressions
5368 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5369 and SRC_RELATED, if nonzero, each contain additional equivalent
5370 expressions. Prune these latter expressions by deleting expressions
5371 already in the equivalence class.
5373 Check for an equivalent identical to the destination. If found,
5374 this is the preferred equivalent since it will likely lead to
5375 elimination of the insn. Indicate this by placing it in
5376 `src_related'. */
5378 if (elt)
5379 elt = elt->first_same_value;
5380 for (p = elt; p; p = p->next_same_value)
5382 enum rtx_code code = GET_CODE (p->exp);
5384 /* If the expression is not valid, ignore it. Then we do not
5385 have to check for validity below. In most cases, we can use
5386 `rtx_equal_p', since canonicalization has already been done. */
5387 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5388 continue;
5390 /* Also skip paradoxical subregs, unless that's what we're
5391 looking for. */
5392 if (code == SUBREG
5393 && (GET_MODE_SIZE (GET_MODE (p->exp))
5394 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5395 && ! (src != 0
5396 && GET_CODE (src) == SUBREG
5397 && GET_MODE (src) == GET_MODE (p->exp)
5398 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5399 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5400 continue;
5402 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5403 src = 0;
5404 else if (src_folded && GET_CODE (src_folded) == code
5405 && rtx_equal_p (src_folded, p->exp))
5406 src_folded = 0;
5407 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5408 && rtx_equal_p (src_eqv_here, p->exp))
5409 src_eqv_here = 0;
5410 else if (src_related && GET_CODE (src_related) == code
5411 && rtx_equal_p (src_related, p->exp))
5412 src_related = 0;
5414 /* This is the same as the destination of the insns, we want
5415 to prefer it. Copy it to src_related. The code below will
5416 then give it a negative cost. */
5417 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5418 src_related = dest;
5421 /* Find the cheapest valid equivalent, trying all the available
5422 possibilities. Prefer items not in the hash table to ones
5423 that are when they are equal cost. Note that we can never
5424 worsen an insn as the current contents will also succeed.
5425 If we find an equivalent identical to the destination, use it as best,
5426 since this insn will probably be eliminated in that case. */
5427 if (src)
5429 if (rtx_equal_p (src, dest))
5430 src_cost = src_regcost = -1;
5431 else
5433 src_cost = COST (src);
5434 src_regcost = approx_reg_cost (src);
5438 if (src_eqv_here)
5440 if (rtx_equal_p (src_eqv_here, dest))
5441 src_eqv_cost = src_eqv_regcost = -1;
5442 else
5444 src_eqv_cost = COST (src_eqv_here);
5445 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5449 if (src_folded)
5451 if (rtx_equal_p (src_folded, dest))
5452 src_folded_cost = src_folded_regcost = -1;
5453 else
5455 src_folded_cost = COST (src_folded);
5456 src_folded_regcost = approx_reg_cost (src_folded);
5460 if (src_related)
5462 if (rtx_equal_p (src_related, dest))
5463 src_related_cost = src_related_regcost = -1;
5464 else
5466 src_related_cost = COST (src_related);
5467 src_related_regcost = approx_reg_cost (src_related);
5471 /* If this was an indirect jump insn, a known label will really be
5472 cheaper even though it looks more expensive. */
5473 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5474 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5476 /* Terminate loop when replacement made. This must terminate since
5477 the current contents will be tested and will always be valid. */
5478 while (1)
5480 rtx trial;
5482 /* Skip invalid entries. */
5483 while (elt && !REG_P (elt->exp)
5484 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5485 elt = elt->next_same_value;
5487 /* A paradoxical subreg would be bad here: it'll be the right
5488 size, but later may be adjusted so that the upper bits aren't
5489 what we want. So reject it. */
5490 if (elt != 0
5491 && GET_CODE (elt->exp) == SUBREG
5492 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5493 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5494 /* It is okay, though, if the rtx we're trying to match
5495 will ignore any of the bits we can't predict. */
5496 && ! (src != 0
5497 && GET_CODE (src) == SUBREG
5498 && GET_MODE (src) == GET_MODE (elt->exp)
5499 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5500 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5502 elt = elt->next_same_value;
5503 continue;
5506 if (elt)
5508 src_elt_cost = elt->cost;
5509 src_elt_regcost = elt->regcost;
5512 /* Find cheapest and skip it for the next time. For items
5513 of equal cost, use this order:
5514 src_folded, src, src_eqv, src_related and hash table entry. */
5515 if (src_folded
5516 && preferable (src_folded_cost, src_folded_regcost,
5517 src_cost, src_regcost) <= 0
5518 && preferable (src_folded_cost, src_folded_regcost,
5519 src_eqv_cost, src_eqv_regcost) <= 0
5520 && preferable (src_folded_cost, src_folded_regcost,
5521 src_related_cost, src_related_regcost) <= 0
5522 && preferable (src_folded_cost, src_folded_regcost,
5523 src_elt_cost, src_elt_regcost) <= 0)
5525 trial = src_folded, src_folded_cost = MAX_COST;
5526 if (src_folded_force_flag)
5528 rtx forced = force_const_mem (mode, trial);
5529 if (forced)
5530 trial = forced;
5533 else if (src
5534 && preferable (src_cost, src_regcost,
5535 src_eqv_cost, src_eqv_regcost) <= 0
5536 && preferable (src_cost, src_regcost,
5537 src_related_cost, src_related_regcost) <= 0
5538 && preferable (src_cost, src_regcost,
5539 src_elt_cost, src_elt_regcost) <= 0)
5540 trial = src, src_cost = MAX_COST;
5541 else if (src_eqv_here
5542 && preferable (src_eqv_cost, src_eqv_regcost,
5543 src_related_cost, src_related_regcost) <= 0
5544 && preferable (src_eqv_cost, src_eqv_regcost,
5545 src_elt_cost, src_elt_regcost) <= 0)
5546 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5547 else if (src_related
5548 && preferable (src_related_cost, src_related_regcost,
5549 src_elt_cost, src_elt_regcost) <= 0)
5550 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5551 else
5553 trial = copy_rtx (elt->exp);
5554 elt = elt->next_same_value;
5555 src_elt_cost = MAX_COST;
5558 /* We don't normally have an insn matching (set (pc) (pc)), so
5559 check for this separately here. We will delete such an
5560 insn below.
5562 For other cases such as a table jump or conditional jump
5563 where we know the ultimate target, go ahead and replace the
5564 operand. While that may not make a valid insn, we will
5565 reemit the jump below (and also insert any necessary
5566 barriers). */
5567 if (n_sets == 1 && dest == pc_rtx
5568 && (trial == pc_rtx
5569 || (GET_CODE (trial) == LABEL_REF
5570 && ! condjump_p (insn))))
5572 /* Don't substitute non-local labels, this confuses CFG. */
5573 if (GET_CODE (trial) == LABEL_REF
5574 && LABEL_REF_NONLOCAL_P (trial))
5575 continue;
5577 SET_SRC (sets[i].rtl) = trial;
5578 cse_jumps_altered = 1;
5579 break;
5582 /* Reject certain invalid forms of CONST that we create. */
5583 else if (CONSTANT_P (trial)
5584 && GET_CODE (trial) == CONST
5585 /* Reject cases that will cause decode_rtx_const to
5586 die. On the alpha when simplifying a switch, we
5587 get (const (truncate (minus (label_ref)
5588 (label_ref)))). */
5589 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5590 /* Likewise on IA-64, except without the
5591 truncate. */
5592 || (GET_CODE (XEXP (trial, 0)) == MINUS
5593 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5594 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5595 /* Do nothing for this case. */
5598 /* Look for a substitution that makes a valid insn. */
5599 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5601 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5603 /* If we just made a substitution inside a libcall, then we
5604 need to make the same substitution in any notes attached
5605 to the RETVAL insn. */
5606 if (libcall_insn
5607 && (REG_P (sets[i].orig_src)
5608 || GET_CODE (sets[i].orig_src) == SUBREG
5609 || MEM_P (sets[i].orig_src)))
5611 rtx note = find_reg_equal_equiv_note (libcall_insn);
5612 if (note != 0)
5613 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5614 sets[i].orig_src,
5615 copy_rtx (new));
5618 /* The result of apply_change_group can be ignored; see
5619 canon_reg. */
5621 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5622 apply_change_group ();
5623 break;
5626 /* If we previously found constant pool entries for
5627 constants and this is a constant, try making a
5628 pool entry. Put it in src_folded unless we already have done
5629 this since that is where it likely came from. */
5631 else if (constant_pool_entries_cost
5632 && CONSTANT_P (trial)
5633 && (src_folded == 0
5634 || (!MEM_P (src_folded)
5635 && ! src_folded_force_flag))
5636 && GET_MODE_CLASS (mode) != MODE_CC
5637 && mode != VOIDmode)
5639 src_folded_force_flag = 1;
5640 src_folded = trial;
5641 src_folded_cost = constant_pool_entries_cost;
5642 src_folded_regcost = constant_pool_entries_regcost;
5646 src = SET_SRC (sets[i].rtl);
5648 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5649 However, there is an important exception: If both are registers
5650 that are not the head of their equivalence class, replace SET_SRC
5651 with the head of the class. If we do not do this, we will have
5652 both registers live over a portion of the basic block. This way,
5653 their lifetimes will likely abut instead of overlapping. */
5654 if (REG_P (dest)
5655 && REGNO_QTY_VALID_P (REGNO (dest)))
5657 int dest_q = REG_QTY (REGNO (dest));
5658 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5660 if (dest_ent->mode == GET_MODE (dest)
5661 && dest_ent->first_reg != REGNO (dest)
5662 && REG_P (src) && REGNO (src) == REGNO (dest)
5663 /* Don't do this if the original insn had a hard reg as
5664 SET_SRC or SET_DEST. */
5665 && (!REG_P (sets[i].src)
5666 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5667 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5668 /* We can't call canon_reg here because it won't do anything if
5669 SRC is a hard register. */
5671 int src_q = REG_QTY (REGNO (src));
5672 struct qty_table_elem *src_ent = &qty_table[src_q];
5673 int first = src_ent->first_reg;
5674 rtx new_src
5675 = (first >= FIRST_PSEUDO_REGISTER
5676 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5678 /* We must use validate-change even for this, because this
5679 might be a special no-op instruction, suitable only to
5680 tag notes onto. */
5681 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5683 src = new_src;
5684 /* If we had a constant that is cheaper than what we are now
5685 setting SRC to, use that constant. We ignored it when we
5686 thought we could make this into a no-op. */
5687 if (src_const && COST (src_const) < COST (src)
5688 && validate_change (insn, &SET_SRC (sets[i].rtl),
5689 src_const, 0))
5690 src = src_const;
5695 /* If we made a change, recompute SRC values. */
5696 if (src != sets[i].src)
5698 cse_altered = 1;
5699 do_not_record = 0;
5700 hash_arg_in_memory = 0;
5701 sets[i].src = src;
5702 sets[i].src_hash = HASH (src, mode);
5703 sets[i].src_volatile = do_not_record;
5704 sets[i].src_in_memory = hash_arg_in_memory;
5705 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5708 /* If this is a single SET, we are setting a register, and we have an
5709 equivalent constant, we want to add a REG_NOTE. We don't want
5710 to write a REG_EQUAL note for a constant pseudo since verifying that
5711 that pseudo hasn't been eliminated is a pain. Such a note also
5712 won't help anything.
5714 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5715 which can be created for a reference to a compile time computable
5716 entry in a jump table. */
5718 if (n_sets == 1 && src_const && REG_P (dest)
5719 && !REG_P (src_const)
5720 && ! (GET_CODE (src_const) == CONST
5721 && GET_CODE (XEXP (src_const, 0)) == MINUS
5722 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5723 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5725 /* We only want a REG_EQUAL note if src_const != src. */
5726 if (! rtx_equal_p (src, src_const))
5728 /* Make sure that the rtx is not shared. */
5729 src_const = copy_rtx (src_const);
5731 /* Record the actual constant value in a REG_EQUAL note,
5732 making a new one if one does not already exist. */
5733 set_unique_reg_note (insn, REG_EQUAL, src_const);
5737 /* Now deal with the destination. */
5738 do_not_record = 0;
5740 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5741 while (GET_CODE (dest) == SUBREG
5742 || GET_CODE (dest) == ZERO_EXTRACT
5743 || GET_CODE (dest) == STRICT_LOW_PART)
5744 dest = XEXP (dest, 0);
5746 sets[i].inner_dest = dest;
5748 if (MEM_P (dest))
5750 #ifdef PUSH_ROUNDING
5751 /* Stack pushes invalidate the stack pointer. */
5752 rtx addr = XEXP (dest, 0);
5753 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5754 && XEXP (addr, 0) == stack_pointer_rtx)
5755 invalidate (stack_pointer_rtx, VOIDmode);
5756 #endif
5757 dest = fold_rtx (dest, insn);
5760 /* Compute the hash code of the destination now,
5761 before the effects of this instruction are recorded,
5762 since the register values used in the address computation
5763 are those before this instruction. */
5764 sets[i].dest_hash = HASH (dest, mode);
5766 /* Don't enter a bit-field in the hash table
5767 because the value in it after the store
5768 may not equal what was stored, due to truncation. */
5770 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5772 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5774 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5775 && GET_CODE (width) == CONST_INT
5776 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5777 && ! (INTVAL (src_const)
5778 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5779 /* Exception: if the value is constant,
5780 and it won't be truncated, record it. */
5782 else
5784 /* This is chosen so that the destination will be invalidated
5785 but no new value will be recorded.
5786 We must invalidate because sometimes constant
5787 values can be recorded for bitfields. */
5788 sets[i].src_elt = 0;
5789 sets[i].src_volatile = 1;
5790 src_eqv = 0;
5791 src_eqv_elt = 0;
5795 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5796 the insn. */
5797 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5799 /* One less use of the label this insn used to jump to. */
5800 delete_insn (insn);
5801 cse_jumps_altered = 1;
5802 /* No more processing for this set. */
5803 sets[i].rtl = 0;
5806 /* If this SET is now setting PC to a label, we know it used to
5807 be a conditional or computed branch. */
5808 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5809 && !LABEL_REF_NONLOCAL_P (src))
5811 /* Now emit a BARRIER after the unconditional jump. */
5812 if (NEXT_INSN (insn) == 0
5813 || !BARRIER_P (NEXT_INSN (insn)))
5814 emit_barrier_after (insn);
5816 /* We reemit the jump in as many cases as possible just in
5817 case the form of an unconditional jump is significantly
5818 different than a computed jump or conditional jump.
5820 If this insn has multiple sets, then reemitting the
5821 jump is nontrivial. So instead we just force rerecognition
5822 and hope for the best. */
5823 if (n_sets == 1)
5825 rtx new, note;
5827 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5828 JUMP_LABEL (new) = XEXP (src, 0);
5829 LABEL_NUSES (XEXP (src, 0))++;
5831 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5832 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5833 if (note)
5835 XEXP (note, 1) = NULL_RTX;
5836 REG_NOTES (new) = note;
5839 delete_insn (insn);
5840 insn = new;
5842 /* Now emit a BARRIER after the unconditional jump. */
5843 if (NEXT_INSN (insn) == 0
5844 || !BARRIER_P (NEXT_INSN (insn)))
5845 emit_barrier_after (insn);
5847 else
5848 INSN_CODE (insn) = -1;
5850 /* Do not bother deleting any unreachable code,
5851 let jump/flow do that. */
5853 cse_jumps_altered = 1;
5854 sets[i].rtl = 0;
5857 /* If destination is volatile, invalidate it and then do no further
5858 processing for this assignment. */
5860 else if (do_not_record)
5862 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5863 invalidate (dest, VOIDmode);
5864 else if (MEM_P (dest))
5865 invalidate (dest, VOIDmode);
5866 else if (GET_CODE (dest) == STRICT_LOW_PART
5867 || GET_CODE (dest) == ZERO_EXTRACT)
5868 invalidate (XEXP (dest, 0), GET_MODE (dest));
5869 sets[i].rtl = 0;
5872 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5873 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5875 #ifdef HAVE_cc0
5876 /* If setting CC0, record what it was set to, or a constant, if it
5877 is equivalent to a constant. If it is being set to a floating-point
5878 value, make a COMPARE with the appropriate constant of 0. If we
5879 don't do this, later code can interpret this as a test against
5880 const0_rtx, which can cause problems if we try to put it into an
5881 insn as a floating-point operand. */
5882 if (dest == cc0_rtx)
5884 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5885 this_insn_cc0_mode = mode;
5886 if (FLOAT_MODE_P (mode))
5887 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5888 CONST0_RTX (mode));
5890 #endif
5893 /* Now enter all non-volatile source expressions in the hash table
5894 if they are not already present.
5895 Record their equivalence classes in src_elt.
5896 This way we can insert the corresponding destinations into
5897 the same classes even if the actual sources are no longer in them
5898 (having been invalidated). */
5900 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5901 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5903 struct table_elt *elt;
5904 struct table_elt *classp = sets[0].src_elt;
5905 rtx dest = SET_DEST (sets[0].rtl);
5906 enum machine_mode eqvmode = GET_MODE (dest);
5908 if (GET_CODE (dest) == STRICT_LOW_PART)
5910 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5911 classp = 0;
5913 if (insert_regs (src_eqv, classp, 0))
5915 rehash_using_reg (src_eqv);
5916 src_eqv_hash = HASH (src_eqv, eqvmode);
5918 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5919 elt->in_memory = src_eqv_in_memory;
5920 src_eqv_elt = elt;
5922 /* Check to see if src_eqv_elt is the same as a set source which
5923 does not yet have an elt, and if so set the elt of the set source
5924 to src_eqv_elt. */
5925 for (i = 0; i < n_sets; i++)
5926 if (sets[i].rtl && sets[i].src_elt == 0
5927 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5928 sets[i].src_elt = src_eqv_elt;
5931 for (i = 0; i < n_sets; i++)
5932 if (sets[i].rtl && ! sets[i].src_volatile
5933 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5935 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5937 /* REG_EQUAL in setting a STRICT_LOW_PART
5938 gives an equivalent for the entire destination register,
5939 not just for the subreg being stored in now.
5940 This is a more interesting equivalence, so we arrange later
5941 to treat the entire reg as the destination. */
5942 sets[i].src_elt = src_eqv_elt;
5943 sets[i].src_hash = src_eqv_hash;
5945 else
5947 /* Insert source and constant equivalent into hash table, if not
5948 already present. */
5949 struct table_elt *classp = src_eqv_elt;
5950 rtx src = sets[i].src;
5951 rtx dest = SET_DEST (sets[i].rtl);
5952 enum machine_mode mode
5953 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5955 /* It's possible that we have a source value known to be
5956 constant but don't have a REG_EQUAL note on the insn.
5957 Lack of a note will mean src_eqv_elt will be NULL. This
5958 can happen where we've generated a SUBREG to access a
5959 CONST_INT that is already in a register in a wider mode.
5960 Ensure that the source expression is put in the proper
5961 constant class. */
5962 if (!classp)
5963 classp = sets[i].src_const_elt;
5965 if (sets[i].src_elt == 0)
5967 /* Don't put a hard register source into the table if this is
5968 the last insn of a libcall. In this case, we only need
5969 to put src_eqv_elt in src_elt. */
5970 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
5972 struct table_elt *elt;
5974 /* Note that these insert_regs calls cannot remove
5975 any of the src_elt's, because they would have failed to
5976 match if not still valid. */
5977 if (insert_regs (src, classp, 0))
5979 rehash_using_reg (src);
5980 sets[i].src_hash = HASH (src, mode);
5982 elt = insert (src, classp, sets[i].src_hash, mode);
5983 elt->in_memory = sets[i].src_in_memory;
5984 sets[i].src_elt = classp = elt;
5986 else
5987 sets[i].src_elt = classp;
5989 if (sets[i].src_const && sets[i].src_const_elt == 0
5990 && src != sets[i].src_const
5991 && ! rtx_equal_p (sets[i].src_const, src))
5992 sets[i].src_elt = insert (sets[i].src_const, classp,
5993 sets[i].src_const_hash, mode);
5996 else if (sets[i].src_elt == 0)
5997 /* If we did not insert the source into the hash table (e.g., it was
5998 volatile), note the equivalence class for the REG_EQUAL value, if any,
5999 so that the destination goes into that class. */
6000 sets[i].src_elt = src_eqv_elt;
6002 /* Record destination addresses in the hash table. This allows us to
6003 check if they are invalidated by other sets. */
6004 for (i = 0; i < n_sets; i++)
6006 if (sets[i].rtl)
6008 rtx x = sets[i].inner_dest;
6009 struct table_elt *elt;
6010 enum machine_mode mode;
6011 unsigned hash;
6013 if (MEM_P (x))
6015 x = XEXP (x, 0);
6016 mode = GET_MODE (x);
6017 hash = HASH (x, mode);
6018 elt = lookup (x, hash, mode);
6019 if (!elt)
6021 if (insert_regs (x, NULL, 0))
6023 rehash_using_reg (x);
6024 hash = HASH (x, mode);
6026 elt = insert (x, NULL, hash, mode);
6029 sets[i].dest_addr_elt = elt;
6031 else
6032 sets[i].dest_addr_elt = NULL;
6036 invalidate_from_clobbers (x);
6038 /* Some registers are invalidated by subroutine calls. Memory is
6039 invalidated by non-constant calls. */
6041 if (CALL_P (insn))
6043 if (! CONST_OR_PURE_CALL_P (insn))
6044 invalidate_memory ();
6045 invalidate_for_call ();
6048 /* Now invalidate everything set by this instruction.
6049 If a SUBREG or other funny destination is being set,
6050 sets[i].rtl is still nonzero, so here we invalidate the reg
6051 a part of which is being set. */
6053 for (i = 0; i < n_sets; i++)
6054 if (sets[i].rtl)
6056 /* We can't use the inner dest, because the mode associated with
6057 a ZERO_EXTRACT is significant. */
6058 rtx dest = SET_DEST (sets[i].rtl);
6060 /* Needed for registers to remove the register from its
6061 previous quantity's chain.
6062 Needed for memory if this is a nonvarying address, unless
6063 we have just done an invalidate_memory that covers even those. */
6064 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6065 invalidate (dest, VOIDmode);
6066 else if (MEM_P (dest))
6067 invalidate (dest, VOIDmode);
6068 else if (GET_CODE (dest) == STRICT_LOW_PART
6069 || GET_CODE (dest) == ZERO_EXTRACT)
6070 invalidate (XEXP (dest, 0), GET_MODE (dest));
6073 /* A volatile ASM invalidates everything. */
6074 if (NONJUMP_INSN_P (insn)
6075 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
6076 && MEM_VOLATILE_P (PATTERN (insn)))
6077 flush_hash_table ();
6079 /* Make sure registers mentioned in destinations
6080 are safe for use in an expression to be inserted.
6081 This removes from the hash table
6082 any invalid entry that refers to one of these registers.
6084 We don't care about the return value from mention_regs because
6085 we are going to hash the SET_DEST values unconditionally. */
6087 for (i = 0; i < n_sets; i++)
6089 if (sets[i].rtl)
6091 rtx x = SET_DEST (sets[i].rtl);
6093 if (!REG_P (x))
6094 mention_regs (x);
6095 else
6097 /* We used to rely on all references to a register becoming
6098 inaccessible when a register changes to a new quantity,
6099 since that changes the hash code. However, that is not
6100 safe, since after HASH_SIZE new quantities we get a
6101 hash 'collision' of a register with its own invalid
6102 entries. And since SUBREGs have been changed not to
6103 change their hash code with the hash code of the register,
6104 it wouldn't work any longer at all. So we have to check
6105 for any invalid references lying around now.
6106 This code is similar to the REG case in mention_regs,
6107 but it knows that reg_tick has been incremented, and
6108 it leaves reg_in_table as -1 . */
6109 unsigned int regno = REGNO (x);
6110 unsigned int endregno
6111 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
6112 : hard_regno_nregs[regno][GET_MODE (x)]);
6113 unsigned int i;
6115 for (i = regno; i < endregno; i++)
6117 if (REG_IN_TABLE (i) >= 0)
6119 remove_invalid_refs (i);
6120 REG_IN_TABLE (i) = -1;
6127 /* We may have just removed some of the src_elt's from the hash table.
6128 So replace each one with the current head of the same class.
6129 Also check if destination addresses have been removed. */
6131 for (i = 0; i < n_sets; i++)
6132 if (sets[i].rtl)
6134 if (sets[i].dest_addr_elt
6135 && sets[i].dest_addr_elt->first_same_value == 0)
6137 /* The elt was removed, which means this destination is not
6138 valid after this instruction. */
6139 sets[i].rtl = NULL_RTX;
6141 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6142 /* If elt was removed, find current head of same class,
6143 or 0 if nothing remains of that class. */
6145 struct table_elt *elt = sets[i].src_elt;
6147 while (elt && elt->prev_same_value)
6148 elt = elt->prev_same_value;
6150 while (elt && elt->first_same_value == 0)
6151 elt = elt->next_same_value;
6152 sets[i].src_elt = elt ? elt->first_same_value : 0;
6156 /* Now insert the destinations into their equivalence classes. */
6158 for (i = 0; i < n_sets; i++)
6159 if (sets[i].rtl)
6161 rtx dest = SET_DEST (sets[i].rtl);
6162 struct table_elt *elt;
6164 /* Don't record value if we are not supposed to risk allocating
6165 floating-point values in registers that might be wider than
6166 memory. */
6167 if ((flag_float_store
6168 && MEM_P (dest)
6169 && FLOAT_MODE_P (GET_MODE (dest)))
6170 /* Don't record BLKmode values, because we don't know the
6171 size of it, and can't be sure that other BLKmode values
6172 have the same or smaller size. */
6173 || GET_MODE (dest) == BLKmode
6174 /* Don't record values of destinations set inside a libcall block
6175 since we might delete the libcall. Things should have been set
6176 up so we won't want to reuse such a value, but we play it safe
6177 here. */
6178 || libcall_insn
6179 /* If we didn't put a REG_EQUAL value or a source into the hash
6180 table, there is no point is recording DEST. */
6181 || sets[i].src_elt == 0
6182 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6183 or SIGN_EXTEND, don't record DEST since it can cause
6184 some tracking to be wrong.
6186 ??? Think about this more later. */
6187 || (GET_CODE (dest) == SUBREG
6188 && (GET_MODE_SIZE (GET_MODE (dest))
6189 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6190 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6191 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6192 continue;
6194 /* STRICT_LOW_PART isn't part of the value BEING set,
6195 and neither is the SUBREG inside it.
6196 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6197 if (GET_CODE (dest) == STRICT_LOW_PART)
6198 dest = SUBREG_REG (XEXP (dest, 0));
6200 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6201 /* Registers must also be inserted into chains for quantities. */
6202 if (insert_regs (dest, sets[i].src_elt, 1))
6204 /* If `insert_regs' changes something, the hash code must be
6205 recalculated. */
6206 rehash_using_reg (dest);
6207 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6210 elt = insert (dest, sets[i].src_elt,
6211 sets[i].dest_hash, GET_MODE (dest));
6213 elt->in_memory = (MEM_P (sets[i].inner_dest)
6214 && !MEM_READONLY_P (sets[i].inner_dest));
6216 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6217 narrower than M2, and both M1 and M2 are the same number of words,
6218 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6219 make that equivalence as well.
6221 However, BAR may have equivalences for which gen_lowpart
6222 will produce a simpler value than gen_lowpart applied to
6223 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6224 BAR's equivalences. If we don't get a simplified form, make
6225 the SUBREG. It will not be used in an equivalence, but will
6226 cause two similar assignments to be detected.
6228 Note the loop below will find SUBREG_REG (DEST) since we have
6229 already entered SRC and DEST of the SET in the table. */
6231 if (GET_CODE (dest) == SUBREG
6232 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6233 / UNITS_PER_WORD)
6234 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6235 && (GET_MODE_SIZE (GET_MODE (dest))
6236 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6237 && sets[i].src_elt != 0)
6239 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6240 struct table_elt *elt, *classp = 0;
6242 for (elt = sets[i].src_elt->first_same_value; elt;
6243 elt = elt->next_same_value)
6245 rtx new_src = 0;
6246 unsigned src_hash;
6247 struct table_elt *src_elt;
6248 int byte = 0;
6250 /* Ignore invalid entries. */
6251 if (!REG_P (elt->exp)
6252 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6253 continue;
6255 /* We may have already been playing subreg games. If the
6256 mode is already correct for the destination, use it. */
6257 if (GET_MODE (elt->exp) == new_mode)
6258 new_src = elt->exp;
6259 else
6261 /* Calculate big endian correction for the SUBREG_BYTE.
6262 We have already checked that M1 (GET_MODE (dest))
6263 is not narrower than M2 (new_mode). */
6264 if (BYTES_BIG_ENDIAN)
6265 byte = (GET_MODE_SIZE (GET_MODE (dest))
6266 - GET_MODE_SIZE (new_mode));
6268 new_src = simplify_gen_subreg (new_mode, elt->exp,
6269 GET_MODE (dest), byte);
6272 /* The call to simplify_gen_subreg fails if the value
6273 is VOIDmode, yet we can't do any simplification, e.g.
6274 for EXPR_LISTs denoting function call results.
6275 It is invalid to construct a SUBREG with a VOIDmode
6276 SUBREG_REG, hence a zero new_src means we can't do
6277 this substitution. */
6278 if (! new_src)
6279 continue;
6281 src_hash = HASH (new_src, new_mode);
6282 src_elt = lookup (new_src, src_hash, new_mode);
6284 /* Put the new source in the hash table is if isn't
6285 already. */
6286 if (src_elt == 0)
6288 if (insert_regs (new_src, classp, 0))
6290 rehash_using_reg (new_src);
6291 src_hash = HASH (new_src, new_mode);
6293 src_elt = insert (new_src, classp, src_hash, new_mode);
6294 src_elt->in_memory = elt->in_memory;
6296 else if (classp && classp != src_elt->first_same_value)
6297 /* Show that two things that we've seen before are
6298 actually the same. */
6299 merge_equiv_classes (src_elt, classp);
6301 classp = src_elt->first_same_value;
6302 /* Ignore invalid entries. */
6303 while (classp
6304 && !REG_P (classp->exp)
6305 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6306 classp = classp->next_same_value;
6311 /* Special handling for (set REG0 REG1) where REG0 is the
6312 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6313 be used in the sequel, so (if easily done) change this insn to
6314 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6315 that computed their value. Then REG1 will become a dead store
6316 and won't cloud the situation for later optimizations.
6318 Do not make this change if REG1 is a hard register, because it will
6319 then be used in the sequel and we may be changing a two-operand insn
6320 into a three-operand insn.
6322 Also do not do this if we are operating on a copy of INSN.
6324 Also don't do this if INSN ends a libcall; this would cause an unrelated
6325 register to be set in the middle of a libcall, and we then get bad code
6326 if the libcall is deleted. */
6328 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6329 && NEXT_INSN (PREV_INSN (insn)) == insn
6330 && REG_P (SET_SRC (sets[0].rtl))
6331 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6332 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6334 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6335 struct qty_table_elem *src_ent = &qty_table[src_q];
6337 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6338 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6340 rtx prev = insn;
6341 /* Scan for the previous nonnote insn, but stop at a basic
6342 block boundary. */
6345 prev = PREV_INSN (prev);
6347 while (prev && NOTE_P (prev)
6348 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6350 /* Do not swap the registers around if the previous instruction
6351 attaches a REG_EQUIV note to REG1.
6353 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6354 from the pseudo that originally shadowed an incoming argument
6355 to another register. Some uses of REG_EQUIV might rely on it
6356 being attached to REG1 rather than REG2.
6358 This section previously turned the REG_EQUIV into a REG_EQUAL
6359 note. We cannot do that because REG_EQUIV may provide an
6360 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6362 if (prev != 0 && NONJUMP_INSN_P (prev)
6363 && GET_CODE (PATTERN (prev)) == SET
6364 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6365 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6367 rtx dest = SET_DEST (sets[0].rtl);
6368 rtx src = SET_SRC (sets[0].rtl);
6369 rtx note;
6371 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6372 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6373 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6374 apply_change_group ();
6376 /* If INSN has a REG_EQUAL note, and this note mentions
6377 REG0, then we must delete it, because the value in
6378 REG0 has changed. If the note's value is REG1, we must
6379 also delete it because that is now this insn's dest. */
6380 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6381 if (note != 0
6382 && (reg_mentioned_p (dest, XEXP (note, 0))
6383 || rtx_equal_p (src, XEXP (note, 0))))
6384 remove_note (insn, note);
6389 /* If this is a conditional jump insn, record any known equivalences due to
6390 the condition being tested. */
6392 if (JUMP_P (insn)
6393 && n_sets == 1 && GET_CODE (x) == SET
6394 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6395 record_jump_equiv (insn, 0);
6397 #ifdef HAVE_cc0
6398 /* If the previous insn set CC0 and this insn no longer references CC0,
6399 delete the previous insn. Here we use the fact that nothing expects CC0
6400 to be valid over an insn, which is true until the final pass. */
6401 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6402 && (tem = single_set (prev_insn)) != 0
6403 && SET_DEST (tem) == cc0_rtx
6404 && ! reg_mentioned_p (cc0_rtx, x))
6405 delete_insn (prev_insn);
6407 prev_insn_cc0 = this_insn_cc0;
6408 prev_insn_cc0_mode = this_insn_cc0_mode;
6409 prev_insn = insn;
6410 #endif
6413 /* Remove from the hash table all expressions that reference memory. */
6415 static void
6416 invalidate_memory (void)
6418 int i;
6419 struct table_elt *p, *next;
6421 for (i = 0; i < HASH_SIZE; i++)
6422 for (p = table[i]; p; p = next)
6424 next = p->next_same_hash;
6425 if (p->in_memory)
6426 remove_from_table (p, i);
6430 /* If ADDR is an address that implicitly affects the stack pointer, return
6431 1 and update the register tables to show the effect. Else, return 0. */
6433 static int
6434 addr_affects_sp_p (rtx addr)
6436 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6437 && REG_P (XEXP (addr, 0))
6438 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6440 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6442 REG_TICK (STACK_POINTER_REGNUM)++;
6443 /* Is it possible to use a subreg of SP? */
6444 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6447 /* This should be *very* rare. */
6448 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6449 invalidate (stack_pointer_rtx, VOIDmode);
6451 return 1;
6454 return 0;
6457 /* Perform invalidation on the basis of everything about an insn
6458 except for invalidating the actual places that are SET in it.
6459 This includes the places CLOBBERed, and anything that might
6460 alias with something that is SET or CLOBBERed.
6462 X is the pattern of the insn. */
6464 static void
6465 invalidate_from_clobbers (rtx x)
6467 if (GET_CODE (x) == CLOBBER)
6469 rtx ref = XEXP (x, 0);
6470 if (ref)
6472 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6473 || MEM_P (ref))
6474 invalidate (ref, VOIDmode);
6475 else if (GET_CODE (ref) == STRICT_LOW_PART
6476 || GET_CODE (ref) == ZERO_EXTRACT)
6477 invalidate (XEXP (ref, 0), GET_MODE (ref));
6480 else if (GET_CODE (x) == PARALLEL)
6482 int i;
6483 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6485 rtx y = XVECEXP (x, 0, i);
6486 if (GET_CODE (y) == CLOBBER)
6488 rtx ref = XEXP (y, 0);
6489 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6490 || MEM_P (ref))
6491 invalidate (ref, VOIDmode);
6492 else if (GET_CODE (ref) == STRICT_LOW_PART
6493 || GET_CODE (ref) == ZERO_EXTRACT)
6494 invalidate (XEXP (ref, 0), GET_MODE (ref));
6500 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6501 and replace any registers in them with either an equivalent constant
6502 or the canonical form of the register. If we are inside an address,
6503 only do this if the address remains valid.
6505 OBJECT is 0 except when within a MEM in which case it is the MEM.
6507 Return the replacement for X. */
6509 static rtx
6510 cse_process_notes (rtx x, rtx object)
6512 enum rtx_code code = GET_CODE (x);
6513 const char *fmt = GET_RTX_FORMAT (code);
6514 int i;
6516 switch (code)
6518 case CONST_INT:
6519 case CONST:
6520 case SYMBOL_REF:
6521 case LABEL_REF:
6522 case CONST_DOUBLE:
6523 case CONST_VECTOR:
6524 case PC:
6525 case CC0:
6526 case LO_SUM:
6527 return x;
6529 case MEM:
6530 validate_change (x, &XEXP (x, 0),
6531 cse_process_notes (XEXP (x, 0), x), 0);
6532 return x;
6534 case EXPR_LIST:
6535 case INSN_LIST:
6536 if (REG_NOTE_KIND (x) == REG_EQUAL)
6537 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6538 if (XEXP (x, 1))
6539 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6540 return x;
6542 case SIGN_EXTEND:
6543 case ZERO_EXTEND:
6544 case SUBREG:
6546 rtx new = cse_process_notes (XEXP (x, 0), object);
6547 /* We don't substitute VOIDmode constants into these rtx,
6548 since they would impede folding. */
6549 if (GET_MODE (new) != VOIDmode)
6550 validate_change (object, &XEXP (x, 0), new, 0);
6551 return x;
6554 case REG:
6555 i = REG_QTY (REGNO (x));
6557 /* Return a constant or a constant register. */
6558 if (REGNO_QTY_VALID_P (REGNO (x)))
6560 struct qty_table_elem *ent = &qty_table[i];
6562 if (ent->const_rtx != NULL_RTX
6563 && (CONSTANT_P (ent->const_rtx)
6564 || REG_P (ent->const_rtx)))
6566 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6567 if (new)
6568 return new;
6572 /* Otherwise, canonicalize this register. */
6573 return canon_reg (x, NULL_RTX);
6575 default:
6576 break;
6579 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6580 if (fmt[i] == 'e')
6581 validate_change (object, &XEXP (x, i),
6582 cse_process_notes (XEXP (x, i), object), 0);
6584 return x;
6587 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6588 since they are done elsewhere. This function is called via note_stores. */
6590 static void
6591 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6593 enum rtx_code code = GET_CODE (dest);
6595 if (code == MEM
6596 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6597 /* There are times when an address can appear varying and be a PLUS
6598 during this scan when it would be a fixed address were we to know
6599 the proper equivalences. So invalidate all memory if there is
6600 a BLKmode or nonscalar memory reference or a reference to a
6601 variable address. */
6602 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6603 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6605 invalidate_memory ();
6606 return;
6609 if (GET_CODE (set) == CLOBBER
6610 || CC0_P (dest)
6611 || dest == pc_rtx)
6612 return;
6614 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6615 invalidate (XEXP (dest, 0), GET_MODE (dest));
6616 else if (code == REG || code == SUBREG || code == MEM)
6617 invalidate (dest, VOIDmode);
6620 /* Invalidate all insns from START up to the end of the function or the
6621 next label. This called when we wish to CSE around a block that is
6622 conditionally executed. */
6624 static void
6625 invalidate_skipped_block (rtx start)
6627 rtx insn;
6629 for (insn = start; insn && !LABEL_P (insn);
6630 insn = NEXT_INSN (insn))
6632 if (! INSN_P (insn))
6633 continue;
6635 if (CALL_P (insn))
6637 if (! CONST_OR_PURE_CALL_P (insn))
6638 invalidate_memory ();
6639 invalidate_for_call ();
6642 invalidate_from_clobbers (PATTERN (insn));
6643 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6647 /* Find the end of INSN's basic block and return its range,
6648 the total number of SETs in all the insns of the block, the last insn of the
6649 block, and the branch path.
6651 The branch path indicates which branches should be followed. If a nonzero
6652 path size is specified, the block should be rescanned and a different set
6653 of branches will be taken. The branch path is only used if
6654 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6656 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6657 used to describe the block. It is filled in with the information about
6658 the current block. The incoming structure's branch path, if any, is used
6659 to construct the output branch path. */
6661 static void
6662 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6663 int follow_jumps, int skip_blocks)
6665 rtx p = insn, q;
6666 int nsets = 0;
6667 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6668 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6669 int path_size = data->path_size;
6670 int path_entry = 0;
6671 int i;
6673 /* Update the previous branch path, if any. If the last branch was
6674 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6675 If it was previously PATH_NOT_TAKEN,
6676 shorten the path by one and look at the previous branch. We know that
6677 at least one branch must have been taken if PATH_SIZE is nonzero. */
6678 while (path_size > 0)
6680 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6682 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6683 break;
6685 else
6686 path_size--;
6689 /* If the first instruction is marked with QImode, that means we've
6690 already processed this block. Our caller will look at DATA->LAST
6691 to figure out where to go next. We want to return the next block
6692 in the instruction stream, not some branched-to block somewhere
6693 else. We accomplish this by pretending our called forbid us to
6694 follow jumps, or skip blocks. */
6695 if (GET_MODE (insn) == QImode)
6696 follow_jumps = skip_blocks = 0;
6698 /* Scan to end of this basic block. */
6699 while (p && !LABEL_P (p))
6701 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6702 the regs restored by the longjmp come from
6703 a later time than the setjmp. */
6704 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6705 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6706 break;
6708 /* A PARALLEL can have lots of SETs in it,
6709 especially if it is really an ASM_OPERANDS. */
6710 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6711 nsets += XVECLEN (PATTERN (p), 0);
6712 else if (!NOTE_P (p))
6713 nsets += 1;
6715 /* Ignore insns made by CSE; they cannot affect the boundaries of
6716 the basic block. */
6718 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6719 high_cuid = INSN_CUID (p);
6720 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6721 low_cuid = INSN_CUID (p);
6723 /* See if this insn is in our branch path. If it is and we are to
6724 take it, do so. */
6725 if (path_entry < path_size && data->path[path_entry].branch == p)
6727 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6728 p = JUMP_LABEL (p);
6730 /* Point to next entry in path, if any. */
6731 path_entry++;
6734 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6735 was specified, we haven't reached our maximum path length, there are
6736 insns following the target of the jump, this is the only use of the
6737 jump label, and the target label is preceded by a BARRIER.
6739 Alternatively, we can follow the jump if it branches around a
6740 block of code and there are no other branches into the block.
6741 In this case invalidate_skipped_block will be called to invalidate any
6742 registers set in the block when following the jump. */
6744 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6745 && JUMP_P (p)
6746 && GET_CODE (PATTERN (p)) == SET
6747 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6748 && JUMP_LABEL (p) != 0
6749 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6750 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6752 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6753 if ((!NOTE_P (q)
6754 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6755 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6756 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6757 break;
6759 /* If we ran into a BARRIER, this code is an extension of the
6760 basic block when the branch is taken. */
6761 if (follow_jumps && q != 0 && BARRIER_P (q))
6763 /* Don't allow ourself to keep walking around an
6764 always-executed loop. */
6765 if (next_real_insn (q) == next)
6767 p = NEXT_INSN (p);
6768 continue;
6771 /* Similarly, don't put a branch in our path more than once. */
6772 for (i = 0; i < path_entry; i++)
6773 if (data->path[i].branch == p)
6774 break;
6776 if (i != path_entry)
6777 break;
6779 data->path[path_entry].branch = p;
6780 data->path[path_entry++].status = PATH_TAKEN;
6782 /* This branch now ends our path. It was possible that we
6783 didn't see this branch the last time around (when the
6784 insn in front of the target was a JUMP_INSN that was
6785 turned into a no-op). */
6786 path_size = path_entry;
6788 p = JUMP_LABEL (p);
6789 /* Mark block so we won't scan it again later. */
6790 PUT_MODE (NEXT_INSN (p), QImode);
6792 /* Detect a branch around a block of code. */
6793 else if (skip_blocks && q != 0 && !LABEL_P (q))
6795 rtx tmp;
6797 if (next_real_insn (q) == next)
6799 p = NEXT_INSN (p);
6800 continue;
6803 for (i = 0; i < path_entry; i++)
6804 if (data->path[i].branch == p)
6805 break;
6807 if (i != path_entry)
6808 break;
6810 /* This is no_labels_between_p (p, q) with an added check for
6811 reaching the end of a function (in case Q precedes P). */
6812 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6813 if (LABEL_P (tmp))
6814 break;
6816 if (tmp == q)
6818 data->path[path_entry].branch = p;
6819 data->path[path_entry++].status = PATH_AROUND;
6821 path_size = path_entry;
6823 p = JUMP_LABEL (p);
6824 /* Mark block so we won't scan it again later. */
6825 PUT_MODE (NEXT_INSN (p), QImode);
6829 p = NEXT_INSN (p);
6832 data->low_cuid = low_cuid;
6833 data->high_cuid = high_cuid;
6834 data->nsets = nsets;
6835 data->last = p;
6837 /* If all jumps in the path are not taken, set our path length to zero
6838 so a rescan won't be done. */
6839 for (i = path_size - 1; i >= 0; i--)
6840 if (data->path[i].status != PATH_NOT_TAKEN)
6841 break;
6843 if (i == -1)
6844 data->path_size = 0;
6845 else
6846 data->path_size = path_size;
6848 /* End the current branch path. */
6849 data->path[path_size].branch = 0;
6852 /* Perform cse on the instructions of a function.
6853 F is the first instruction.
6854 NREGS is one plus the highest pseudo-reg number used in the instruction.
6856 Returns 1 if jump_optimize should be redone due to simplifications
6857 in conditional jump instructions. */
6860 cse_main (rtx f, int nregs)
6862 struct cse_basic_block_data val;
6863 rtx insn = f;
6864 int i;
6866 init_cse_reg_info (nregs);
6868 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6870 cse_jumps_altered = 0;
6871 recorded_label_ref = 0;
6872 constant_pool_entries_cost = 0;
6873 constant_pool_entries_regcost = 0;
6874 val.path_size = 0;
6875 rtl_hooks = cse_rtl_hooks;
6877 init_recog ();
6878 init_alias_analysis ();
6880 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6882 /* Find the largest uid. */
6884 max_uid = get_max_uid ();
6885 uid_cuid = XCNEWVEC (int, max_uid + 1);
6887 /* Compute the mapping from uids to cuids.
6888 CUIDs are numbers assigned to insns, like uids,
6889 except that cuids increase monotonically through the code.
6890 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6891 between two insns is not affected by -g. */
6893 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6895 if (!NOTE_P (insn)
6896 || NOTE_LINE_NUMBER (insn) < 0)
6897 INSN_CUID (insn) = ++i;
6898 else
6899 /* Give a line number note the same cuid as preceding insn. */
6900 INSN_CUID (insn) = i;
6903 /* Loop over basic blocks.
6904 Compute the maximum number of qty's needed for each basic block
6905 (which is 2 for each SET). */
6906 insn = f;
6907 while (insn)
6909 cse_altered = 0;
6910 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6911 flag_cse_skip_blocks);
6913 /* If this basic block was already processed or has no sets, skip it. */
6914 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6916 PUT_MODE (insn, VOIDmode);
6917 insn = (val.last ? NEXT_INSN (val.last) : 0);
6918 val.path_size = 0;
6919 continue;
6922 cse_basic_block_start = val.low_cuid;
6923 cse_basic_block_end = val.high_cuid;
6924 max_qty = val.nsets * 2;
6926 if (dump_file)
6927 fprintf (dump_file, ";; Processing block from %d to %d, %d sets.\n",
6928 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6929 val.nsets);
6931 /* Make MAX_QTY bigger to give us room to optimize
6932 past the end of this basic block, if that should prove useful. */
6933 if (max_qty < 500)
6934 max_qty = 500;
6936 /* If this basic block is being extended by following certain jumps,
6937 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6938 Otherwise, we start after this basic block. */
6939 if (val.path_size > 0)
6940 cse_basic_block (insn, val.last, val.path);
6941 else
6943 int old_cse_jumps_altered = cse_jumps_altered;
6944 rtx temp;
6946 /* When cse changes a conditional jump to an unconditional
6947 jump, we want to reprocess the block, since it will give
6948 us a new branch path to investigate. */
6949 cse_jumps_altered = 0;
6950 temp = cse_basic_block (insn, val.last, val.path);
6951 if (cse_jumps_altered == 0
6952 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
6953 insn = temp;
6955 cse_jumps_altered |= old_cse_jumps_altered;
6958 if (cse_altered)
6959 ggc_collect ();
6961 #ifdef USE_C_ALLOCA
6962 alloca (0);
6963 #endif
6966 /* Clean up. */
6967 end_alias_analysis ();
6968 free (uid_cuid);
6969 free (reg_eqv_table);
6970 free (val.path);
6971 rtl_hooks = general_rtl_hooks;
6973 return cse_jumps_altered || recorded_label_ref;
6976 /* Process a single basic block. FROM and TO and the limits of the basic
6977 block. NEXT_BRANCH points to the branch path when following jumps or
6978 a null path when not following jumps. */
6980 static rtx
6981 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
6983 rtx insn;
6984 int to_usage = 0;
6985 rtx libcall_insn = NULL_RTX;
6986 int num_insns = 0;
6987 int no_conflict = 0;
6989 /* Allocate the space needed by qty_table. */
6990 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6992 new_basic_block ();
6994 /* TO might be a label. If so, protect it from being deleted. */
6995 if (to != 0 && LABEL_P (to))
6996 ++LABEL_NUSES (to);
6998 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7000 enum rtx_code code = GET_CODE (insn);
7002 /* If we have processed 1,000 insns, flush the hash table to
7003 avoid extreme quadratic behavior. We must not include NOTEs
7004 in the count since there may be more of them when generating
7005 debugging information. If we clear the table at different
7006 times, code generated with -g -O might be different than code
7007 generated with -O but not -g.
7009 ??? This is a real kludge and needs to be done some other way.
7010 Perhaps for 2.9. */
7011 if (code != NOTE && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
7013 flush_hash_table ();
7014 num_insns = 0;
7017 /* See if this is a branch that is part of the path. If so, and it is
7018 to be taken, do so. */
7019 if (next_branch->branch == insn)
7021 enum taken status = next_branch++->status;
7022 if (status != PATH_NOT_TAKEN)
7024 if (status == PATH_TAKEN)
7025 record_jump_equiv (insn, 1);
7026 else
7027 invalidate_skipped_block (NEXT_INSN (insn));
7029 /* Set the last insn as the jump insn; it doesn't affect cc0.
7030 Then follow this branch. */
7031 #ifdef HAVE_cc0
7032 prev_insn_cc0 = 0;
7033 prev_insn = insn;
7034 #endif
7035 insn = JUMP_LABEL (insn);
7036 continue;
7040 if (GET_MODE (insn) == QImode)
7041 PUT_MODE (insn, VOIDmode);
7043 if (GET_RTX_CLASS (code) == RTX_INSN)
7045 rtx p;
7047 /* Process notes first so we have all notes in canonical forms when
7048 looking for duplicate operations. */
7050 if (REG_NOTES (insn))
7051 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7053 /* Track when we are inside in LIBCALL block. Inside such a block,
7054 we do not want to record destinations. The last insn of a
7055 LIBCALL block is not considered to be part of the block, since
7056 its destination is the result of the block and hence should be
7057 recorded. */
7059 if (REG_NOTES (insn) != 0)
7061 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7062 libcall_insn = XEXP (p, 0);
7063 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7065 /* Keep libcall_insn for the last SET insn of a no-conflict
7066 block to prevent changing the destination. */
7067 if (! no_conflict)
7068 libcall_insn = 0;
7069 else
7070 no_conflict = -1;
7072 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
7073 no_conflict = 1;
7076 cse_insn (insn, libcall_insn);
7078 if (no_conflict == -1)
7080 libcall_insn = 0;
7081 no_conflict = 0;
7084 /* If we haven't already found an insn where we added a LABEL_REF,
7085 check this one. */
7086 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
7087 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
7088 (void *) insn))
7089 recorded_label_ref = 1;
7092 /* If INSN is now an unconditional jump, skip to the end of our
7093 basic block by pretending that we just did the last insn in the
7094 basic block. If we are jumping to the end of our block, show
7095 that we can have one usage of TO. */
7097 if (any_uncondjump_p (insn))
7099 if (to == 0)
7101 free (qty_table);
7102 return 0;
7105 if (JUMP_LABEL (insn) == to)
7106 to_usage = 1;
7108 /* Maybe TO was deleted because the jump is unconditional.
7109 If so, there is nothing left in this basic block. */
7110 /* ??? Perhaps it would be smarter to set TO
7111 to whatever follows this insn,
7112 and pretend the basic block had always ended here. */
7113 if (INSN_DELETED_P (to))
7114 break;
7116 insn = PREV_INSN (to);
7119 /* See if it is ok to keep on going past the label
7120 which used to end our basic block. Remember that we incremented
7121 the count of that label, so we decrement it here. If we made
7122 a jump unconditional, TO_USAGE will be one; in that case, we don't
7123 want to count the use in that jump. */
7125 if (to != 0 && NEXT_INSN (insn) == to
7126 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7128 struct cse_basic_block_data val;
7129 rtx prev;
7131 insn = NEXT_INSN (to);
7133 /* If TO was the last insn in the function, we are done. */
7134 if (insn == 0)
7136 free (qty_table);
7137 return 0;
7140 /* If TO was preceded by a BARRIER we are done with this block
7141 because it has no continuation. */
7142 prev = prev_nonnote_insn (to);
7143 if (prev && BARRIER_P (prev))
7145 free (qty_table);
7146 return insn;
7149 /* Find the end of the following block. Note that we won't be
7150 following branches in this case. */
7151 to_usage = 0;
7152 val.path_size = 0;
7153 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7154 cse_end_of_basic_block (insn, &val, 0, 0);
7155 free (val.path);
7157 /* If the tables we allocated have enough space left
7158 to handle all the SETs in the next basic block,
7159 continue through it. Otherwise, return,
7160 and that block will be scanned individually. */
7161 if (val.nsets * 2 + next_qty > max_qty)
7162 break;
7164 cse_basic_block_start = val.low_cuid;
7165 cse_basic_block_end = val.high_cuid;
7166 to = val.last;
7168 /* Prevent TO from being deleted if it is a label. */
7169 if (to != 0 && LABEL_P (to))
7170 ++LABEL_NUSES (to);
7172 /* Back up so we process the first insn in the extension. */
7173 insn = PREV_INSN (insn);
7177 gcc_assert (next_qty <= max_qty);
7179 free (qty_table);
7181 return to ? NEXT_INSN (to) : 0;
7184 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7185 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7187 static int
7188 check_for_label_ref (rtx *rtl, void *data)
7190 rtx insn = (rtx) data;
7192 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7193 we must rerun jump since it needs to place the note. If this is a
7194 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7195 since no REG_LABEL will be added. */
7196 return (GET_CODE (*rtl) == LABEL_REF
7197 && ! LABEL_REF_NONLOCAL_P (*rtl)
7198 && LABEL_P (XEXP (*rtl, 0))
7199 && INSN_UID (XEXP (*rtl, 0)) != 0
7200 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7203 /* Count the number of times registers are used (not set) in X.
7204 COUNTS is an array in which we accumulate the count, INCR is how much
7205 we count each register usage.
7207 Don't count a usage of DEST, which is the SET_DEST of a SET which
7208 contains X in its SET_SRC. This is because such a SET does not
7209 modify the liveness of DEST.
7210 DEST is set to pc_rtx for a trapping insn, which means that we must count
7211 uses of a SET_DEST regardless because the insn can't be deleted here. */
7213 static void
7214 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
7216 enum rtx_code code;
7217 rtx note;
7218 const char *fmt;
7219 int i, j;
7221 if (x == 0)
7222 return;
7224 switch (code = GET_CODE (x))
7226 case REG:
7227 if (x != dest)
7228 counts[REGNO (x)] += incr;
7229 return;
7231 case PC:
7232 case CC0:
7233 case CONST:
7234 case CONST_INT:
7235 case CONST_DOUBLE:
7236 case CONST_VECTOR:
7237 case SYMBOL_REF:
7238 case LABEL_REF:
7239 return;
7241 case CLOBBER:
7242 /* If we are clobbering a MEM, mark any registers inside the address
7243 as being used. */
7244 if (MEM_P (XEXP (x, 0)))
7245 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7246 return;
7248 case SET:
7249 /* Unless we are setting a REG, count everything in SET_DEST. */
7250 if (!REG_P (SET_DEST (x)))
7251 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7252 count_reg_usage (SET_SRC (x), counts,
7253 dest ? dest : SET_DEST (x),
7254 incr);
7255 return;
7257 case CALL_INSN:
7258 case INSN:
7259 case JUMP_INSN:
7260 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
7261 this fact by setting DEST to pc_rtx. */
7262 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
7263 dest = pc_rtx;
7264 if (code == CALL_INSN)
7265 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
7266 count_reg_usage (PATTERN (x), counts, dest, incr);
7268 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7269 use them. */
7271 note = find_reg_equal_equiv_note (x);
7272 if (note)
7274 rtx eqv = XEXP (note, 0);
7276 if (GET_CODE (eqv) == EXPR_LIST)
7277 /* This REG_EQUAL note describes the result of a function call.
7278 Process all the arguments. */
7281 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
7282 eqv = XEXP (eqv, 1);
7284 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7285 else
7286 count_reg_usage (eqv, counts, dest, incr);
7288 return;
7290 case EXPR_LIST:
7291 if (REG_NOTE_KIND (x) == REG_EQUAL
7292 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7293 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7294 involving registers in the address. */
7295 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7296 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7298 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7299 return;
7301 case ASM_OPERANDS:
7302 /* If the asm is volatile, then this insn cannot be deleted,
7303 and so the inputs *must* be live. */
7304 if (MEM_VOLATILE_P (x))
7305 dest = NULL_RTX;
7306 /* Iterate over just the inputs, not the constraints as well. */
7307 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7308 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
7309 return;
7311 case INSN_LIST:
7312 gcc_unreachable ();
7314 default:
7315 break;
7318 fmt = GET_RTX_FORMAT (code);
7319 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7321 if (fmt[i] == 'e')
7322 count_reg_usage (XEXP (x, i), counts, dest, incr);
7323 else if (fmt[i] == 'E')
7324 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7325 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7329 /* Return true if set is live. */
7330 static bool
7331 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7332 int *counts)
7334 #ifdef HAVE_cc0
7335 rtx tem;
7336 #endif
7338 if (set_noop_p (set))
7341 #ifdef HAVE_cc0
7342 else if (GET_CODE (SET_DEST (set)) == CC0
7343 && !side_effects_p (SET_SRC (set))
7344 && ((tem = next_nonnote_insn (insn)) == 0
7345 || !INSN_P (tem)
7346 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7347 return false;
7348 #endif
7349 else if (!REG_P (SET_DEST (set))
7350 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7351 || counts[REGNO (SET_DEST (set))] != 0
7352 || side_effects_p (SET_SRC (set)))
7353 return true;
7354 return false;
7357 /* Return true if insn is live. */
7359 static bool
7360 insn_live_p (rtx insn, int *counts)
7362 int i;
7363 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7364 return true;
7365 else if (GET_CODE (PATTERN (insn)) == SET)
7366 return set_live_p (PATTERN (insn), insn, counts);
7367 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7369 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7371 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7373 if (GET_CODE (elt) == SET)
7375 if (set_live_p (elt, insn, counts))
7376 return true;
7378 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7379 return true;
7381 return false;
7383 else
7384 return true;
7387 /* Return true if libcall is dead as a whole. */
7389 static bool
7390 dead_libcall_p (rtx insn, int *counts)
7392 rtx note, set, new;
7394 /* See if there's a REG_EQUAL note on this insn and try to
7395 replace the source with the REG_EQUAL expression.
7397 We assume that insns with REG_RETVALs can only be reg->reg
7398 copies at this point. */
7399 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7400 if (!note)
7401 return false;
7403 set = single_set (insn);
7404 if (!set)
7405 return false;
7407 new = simplify_rtx (XEXP (note, 0));
7408 if (!new)
7409 new = XEXP (note, 0);
7411 /* While changing insn, we must update the counts accordingly. */
7412 count_reg_usage (insn, counts, NULL_RTX, -1);
7414 if (validate_change (insn, &SET_SRC (set), new, 0))
7416 count_reg_usage (insn, counts, NULL_RTX, 1);
7417 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7418 remove_note (insn, note);
7419 return true;
7422 if (CONSTANT_P (new))
7424 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7425 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7427 count_reg_usage (insn, counts, NULL_RTX, 1);
7428 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7429 remove_note (insn, note);
7430 return true;
7434 count_reg_usage (insn, counts, NULL_RTX, 1);
7435 return false;
7438 /* Scan all the insns and delete any that are dead; i.e., they store a register
7439 that is never used or they copy a register to itself.
7441 This is used to remove insns made obviously dead by cse, loop or other
7442 optimizations. It improves the heuristics in loop since it won't try to
7443 move dead invariants out of loops or make givs for dead quantities. The
7444 remaining passes of the compilation are also sped up. */
7447 delete_trivially_dead_insns (rtx insns, int nreg)
7449 int *counts;
7450 rtx insn, prev;
7451 int in_libcall = 0, dead_libcall = 0;
7452 int ndead = 0;
7454 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7455 /* First count the number of times each register is used. */
7456 counts = XCNEWVEC (int, nreg);
7457 for (insn = insns; insn; insn = NEXT_INSN (insn))
7458 if (INSN_P (insn))
7459 count_reg_usage (insn, counts, NULL_RTX, 1);
7461 /* Go from the last insn to the first and delete insns that only set unused
7462 registers or copy a register to itself. As we delete an insn, remove
7463 usage counts for registers it uses.
7465 The first jump optimization pass may leave a real insn as the last
7466 insn in the function. We must not skip that insn or we may end
7467 up deleting code that is not really dead. */
7468 for (insn = get_last_insn (); insn; insn = prev)
7470 int live_insn = 0;
7472 prev = PREV_INSN (insn);
7473 if (!INSN_P (insn))
7474 continue;
7476 /* Don't delete any insns that are part of a libcall block unless
7477 we can delete the whole libcall block.
7479 Flow or loop might get confused if we did that. Remember
7480 that we are scanning backwards. */
7481 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7483 in_libcall = 1;
7484 live_insn = 1;
7485 dead_libcall = dead_libcall_p (insn, counts);
7487 else if (in_libcall)
7488 live_insn = ! dead_libcall;
7489 else
7490 live_insn = insn_live_p (insn, counts);
7492 /* If this is a dead insn, delete it and show registers in it aren't
7493 being used. */
7495 if (! live_insn)
7497 count_reg_usage (insn, counts, NULL_RTX, -1);
7498 delete_insn_and_edges (insn);
7499 ndead++;
7502 if (in_libcall && find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7504 in_libcall = 0;
7505 dead_libcall = 0;
7509 if (dump_file && ndead)
7510 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7511 ndead);
7512 /* Clean up. */
7513 free (counts);
7514 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7515 return ndead;
7518 /* This function is called via for_each_rtx. The argument, NEWREG, is
7519 a condition code register with the desired mode. If we are looking
7520 at the same register in a different mode, replace it with
7521 NEWREG. */
7523 static int
7524 cse_change_cc_mode (rtx *loc, void *data)
7526 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7528 if (*loc
7529 && REG_P (*loc)
7530 && REGNO (*loc) == REGNO (args->newreg)
7531 && GET_MODE (*loc) != GET_MODE (args->newreg))
7533 validate_change (args->insn, loc, args->newreg, 1);
7535 return -1;
7537 return 0;
7540 /* Change the mode of any reference to the register REGNO (NEWREG) to
7541 GET_MODE (NEWREG) in INSN. */
7543 static void
7544 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7546 struct change_cc_mode_args args;
7547 int success;
7549 if (!INSN_P (insn))
7550 return;
7552 args.insn = insn;
7553 args.newreg = newreg;
7555 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7556 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7558 /* If the following assertion was triggered, there is most probably
7559 something wrong with the cc_modes_compatible back end function.
7560 CC modes only can be considered compatible if the insn - with the mode
7561 replaced by any of the compatible modes - can still be recognized. */
7562 success = apply_change_group ();
7563 gcc_assert (success);
7566 /* Change the mode of any reference to the register REGNO (NEWREG) to
7567 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7568 any instruction which modifies NEWREG. */
7570 static void
7571 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7573 rtx insn;
7575 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7577 if (! INSN_P (insn))
7578 continue;
7580 if (reg_set_p (newreg, insn))
7581 return;
7583 cse_change_cc_mode_insn (insn, newreg);
7587 /* BB is a basic block which finishes with CC_REG as a condition code
7588 register which is set to CC_SRC. Look through the successors of BB
7589 to find blocks which have a single predecessor (i.e., this one),
7590 and look through those blocks for an assignment to CC_REG which is
7591 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7592 permitted to change the mode of CC_SRC to a compatible mode. This
7593 returns VOIDmode if no equivalent assignments were found.
7594 Otherwise it returns the mode which CC_SRC should wind up with.
7596 The main complexity in this function is handling the mode issues.
7597 We may have more than one duplicate which we can eliminate, and we
7598 try to find a mode which will work for multiple duplicates. */
7600 static enum machine_mode
7601 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7603 bool found_equiv;
7604 enum machine_mode mode;
7605 unsigned int insn_count;
7606 edge e;
7607 rtx insns[2];
7608 enum machine_mode modes[2];
7609 rtx last_insns[2];
7610 unsigned int i;
7611 rtx newreg;
7612 edge_iterator ei;
7614 /* We expect to have two successors. Look at both before picking
7615 the final mode for the comparison. If we have more successors
7616 (i.e., some sort of table jump, although that seems unlikely),
7617 then we require all beyond the first two to use the same
7618 mode. */
7620 found_equiv = false;
7621 mode = GET_MODE (cc_src);
7622 insn_count = 0;
7623 FOR_EACH_EDGE (e, ei, bb->succs)
7625 rtx insn;
7626 rtx end;
7628 if (e->flags & EDGE_COMPLEX)
7629 continue;
7631 if (EDGE_COUNT (e->dest->preds) != 1
7632 || e->dest == EXIT_BLOCK_PTR)
7633 continue;
7635 end = NEXT_INSN (BB_END (e->dest));
7636 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7638 rtx set;
7640 if (! INSN_P (insn))
7641 continue;
7643 /* If CC_SRC is modified, we have to stop looking for
7644 something which uses it. */
7645 if (modified_in_p (cc_src, insn))
7646 break;
7648 /* Check whether INSN sets CC_REG to CC_SRC. */
7649 set = single_set (insn);
7650 if (set
7651 && REG_P (SET_DEST (set))
7652 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7654 bool found;
7655 enum machine_mode set_mode;
7656 enum machine_mode comp_mode;
7658 found = false;
7659 set_mode = GET_MODE (SET_SRC (set));
7660 comp_mode = set_mode;
7661 if (rtx_equal_p (cc_src, SET_SRC (set)))
7662 found = true;
7663 else if (GET_CODE (cc_src) == COMPARE
7664 && GET_CODE (SET_SRC (set)) == COMPARE
7665 && mode != set_mode
7666 && rtx_equal_p (XEXP (cc_src, 0),
7667 XEXP (SET_SRC (set), 0))
7668 && rtx_equal_p (XEXP (cc_src, 1),
7669 XEXP (SET_SRC (set), 1)))
7672 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7673 if (comp_mode != VOIDmode
7674 && (can_change_mode || comp_mode == mode))
7675 found = true;
7678 if (found)
7680 found_equiv = true;
7681 if (insn_count < ARRAY_SIZE (insns))
7683 insns[insn_count] = insn;
7684 modes[insn_count] = set_mode;
7685 last_insns[insn_count] = end;
7686 ++insn_count;
7688 if (mode != comp_mode)
7690 gcc_assert (can_change_mode);
7691 mode = comp_mode;
7693 /* The modified insn will be re-recognized later. */
7694 PUT_MODE (cc_src, mode);
7697 else
7699 if (set_mode != mode)
7701 /* We found a matching expression in the
7702 wrong mode, but we don't have room to
7703 store it in the array. Punt. This case
7704 should be rare. */
7705 break;
7707 /* INSN sets CC_REG to a value equal to CC_SRC
7708 with the right mode. We can simply delete
7709 it. */
7710 delete_insn (insn);
7713 /* We found an instruction to delete. Keep looking,
7714 in the hopes of finding a three-way jump. */
7715 continue;
7718 /* We found an instruction which sets the condition
7719 code, so don't look any farther. */
7720 break;
7723 /* If INSN sets CC_REG in some other way, don't look any
7724 farther. */
7725 if (reg_set_p (cc_reg, insn))
7726 break;
7729 /* If we fell off the bottom of the block, we can keep looking
7730 through successors. We pass CAN_CHANGE_MODE as false because
7731 we aren't prepared to handle compatibility between the
7732 further blocks and this block. */
7733 if (insn == end)
7735 enum machine_mode submode;
7737 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7738 if (submode != VOIDmode)
7740 gcc_assert (submode == mode);
7741 found_equiv = true;
7742 can_change_mode = false;
7747 if (! found_equiv)
7748 return VOIDmode;
7750 /* Now INSN_COUNT is the number of instructions we found which set
7751 CC_REG to a value equivalent to CC_SRC. The instructions are in
7752 INSNS. The modes used by those instructions are in MODES. */
7754 newreg = NULL_RTX;
7755 for (i = 0; i < insn_count; ++i)
7757 if (modes[i] != mode)
7759 /* We need to change the mode of CC_REG in INSNS[i] and
7760 subsequent instructions. */
7761 if (! newreg)
7763 if (GET_MODE (cc_reg) == mode)
7764 newreg = cc_reg;
7765 else
7766 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7768 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7769 newreg);
7772 delete_insn (insns[i]);
7775 return mode;
7778 /* If we have a fixed condition code register (or two), walk through
7779 the instructions and try to eliminate duplicate assignments. */
7781 static void
7782 cse_condition_code_reg (void)
7784 unsigned int cc_regno_1;
7785 unsigned int cc_regno_2;
7786 rtx cc_reg_1;
7787 rtx cc_reg_2;
7788 basic_block bb;
7790 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7791 return;
7793 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7794 if (cc_regno_2 != INVALID_REGNUM)
7795 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7796 else
7797 cc_reg_2 = NULL_RTX;
7799 FOR_EACH_BB (bb)
7801 rtx last_insn;
7802 rtx cc_reg;
7803 rtx insn;
7804 rtx cc_src_insn;
7805 rtx cc_src;
7806 enum machine_mode mode;
7807 enum machine_mode orig_mode;
7809 /* Look for blocks which end with a conditional jump based on a
7810 condition code register. Then look for the instruction which
7811 sets the condition code register. Then look through the
7812 successor blocks for instructions which set the condition
7813 code register to the same value. There are other possible
7814 uses of the condition code register, but these are by far the
7815 most common and the ones which we are most likely to be able
7816 to optimize. */
7818 last_insn = BB_END (bb);
7819 if (!JUMP_P (last_insn))
7820 continue;
7822 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7823 cc_reg = cc_reg_1;
7824 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7825 cc_reg = cc_reg_2;
7826 else
7827 continue;
7829 cc_src_insn = NULL_RTX;
7830 cc_src = NULL_RTX;
7831 for (insn = PREV_INSN (last_insn);
7832 insn && insn != PREV_INSN (BB_HEAD (bb));
7833 insn = PREV_INSN (insn))
7835 rtx set;
7837 if (! INSN_P (insn))
7838 continue;
7839 set = single_set (insn);
7840 if (set
7841 && REG_P (SET_DEST (set))
7842 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7844 cc_src_insn = insn;
7845 cc_src = SET_SRC (set);
7846 break;
7848 else if (reg_set_p (cc_reg, insn))
7849 break;
7852 if (! cc_src_insn)
7853 continue;
7855 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7856 continue;
7858 /* Now CC_REG is a condition code register used for a
7859 conditional jump at the end of the block, and CC_SRC, in
7860 CC_SRC_INSN, is the value to which that condition code
7861 register is set, and CC_SRC is still meaningful at the end of
7862 the basic block. */
7864 orig_mode = GET_MODE (cc_src);
7865 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7866 if (mode != VOIDmode)
7868 gcc_assert (mode == GET_MODE (cc_src));
7869 if (mode != orig_mode)
7871 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7873 cse_change_cc_mode_insn (cc_src_insn, newreg);
7875 /* Do the same in the following insns that use the
7876 current value of CC_REG within BB. */
7877 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7878 NEXT_INSN (last_insn),
7879 newreg);
7886 /* Perform common subexpression elimination. Nonzero value from
7887 `cse_main' means that jumps were simplified and some code may now
7888 be unreachable, so do jump optimization again. */
7889 static bool
7890 gate_handle_cse (void)
7892 return optimize > 0;
7895 static unsigned int
7896 rest_of_handle_cse (void)
7898 int tem;
7900 if (dump_file)
7901 dump_flow_info (dump_file, dump_flags);
7903 reg_scan (get_insns (), max_reg_num ());
7905 tem = cse_main (get_insns (), max_reg_num ());
7906 if (tem)
7907 rebuild_jump_labels (get_insns ());
7908 if (purge_all_dead_edges ())
7909 delete_unreachable_blocks ();
7911 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7913 /* If we are not running more CSE passes, then we are no longer
7914 expecting CSE to be run. But always rerun it in a cheap mode. */
7915 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7917 if (tem)
7918 delete_dead_jumptables ();
7920 if (tem || optimize > 1)
7921 cleanup_cfg (CLEANUP_EXPENSIVE);
7922 return 0;
7925 struct tree_opt_pass pass_cse =
7927 "cse1", /* name */
7928 gate_handle_cse, /* gate */
7929 rest_of_handle_cse, /* execute */
7930 NULL, /* sub */
7931 NULL, /* next */
7932 0, /* static_pass_number */
7933 TV_CSE, /* tv_id */
7934 0, /* properties_required */
7935 0, /* properties_provided */
7936 0, /* properties_destroyed */
7937 0, /* todo_flags_start */
7938 TODO_dump_func |
7939 TODO_ggc_collect, /* todo_flags_finish */
7940 's' /* letter */
7944 static bool
7945 gate_handle_cse2 (void)
7947 return optimize > 0 && flag_rerun_cse_after_loop;
7950 /* Run second CSE pass after loop optimizations. */
7951 static unsigned int
7952 rest_of_handle_cse2 (void)
7954 int tem;
7956 if (dump_file)
7957 dump_flow_info (dump_file, dump_flags);
7959 tem = cse_main (get_insns (), max_reg_num ());
7961 /* Run a pass to eliminate duplicated assignments to condition code
7962 registers. We have to run this after bypass_jumps, because it
7963 makes it harder for that pass to determine whether a jump can be
7964 bypassed safely. */
7965 cse_condition_code_reg ();
7967 purge_all_dead_edges ();
7968 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7970 if (tem)
7972 timevar_push (TV_JUMP);
7973 rebuild_jump_labels (get_insns ());
7974 delete_dead_jumptables ();
7975 cleanup_cfg (CLEANUP_EXPENSIVE);
7976 timevar_pop (TV_JUMP);
7978 reg_scan (get_insns (), max_reg_num ());
7979 cse_not_expected = 1;
7980 return 0;
7984 struct tree_opt_pass pass_cse2 =
7986 "cse2", /* name */
7987 gate_handle_cse2, /* gate */
7988 rest_of_handle_cse2, /* execute */
7989 NULL, /* sub */
7990 NULL, /* next */
7991 0, /* static_pass_number */
7992 TV_CSE2, /* tv_id */
7993 0, /* properties_required */
7994 0, /* properties_provided */
7995 0, /* properties_destroyed */
7996 0, /* todo_flags_start */
7997 TODO_dump_func |
7998 TODO_ggc_collect, /* todo_flags_finish */
7999 't' /* letter */