* config/m68k/m68k.md (bungt_rev): New pattern.
[official-gcc.git] / gcc / cse.c
blob0304d6f763f1435bc0f82216b112ff6312cc4967
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 /* Number of elements in the hash table. */
533 static unsigned int table_size;
535 /* Chain of `struct table_elt's made so far for this function
536 but currently removed from the table. */
538 static struct table_elt *free_element_chain;
540 /* Set to the cost of a constant pool reference if one was found for a
541 symbolic constant. If this was found, it means we should try to
542 convert constants into constant pool entries if they don't fit in
543 the insn. */
545 static int constant_pool_entries_cost;
546 static int constant_pool_entries_regcost;
548 /* This data describes a block that will be processed by cse_basic_block. */
550 struct cse_basic_block_data
552 /* Lowest CUID value of insns in block. */
553 int low_cuid;
554 /* Highest CUID value of insns in block. */
555 int high_cuid;
556 /* Total number of SETs in block. */
557 int nsets;
558 /* Last insn in the block. */
559 rtx last;
560 /* Size of current branch path, if any. */
561 int path_size;
562 /* Current branch path, indicating which branches will be taken. */
563 struct branch_path
565 /* The branch insn. */
566 rtx branch;
567 /* Whether it should be taken or not. AROUND is the same as taken
568 except that it is used when the destination label is not preceded
569 by a BARRIER. */
570 enum taken {PATH_TAKEN, PATH_NOT_TAKEN, PATH_AROUND} status;
571 } *path;
574 static bool fixed_base_plus_p (rtx x);
575 static int notreg_cost (rtx, enum rtx_code);
576 static int approx_reg_cost_1 (rtx *, void *);
577 static int approx_reg_cost (rtx);
578 static int preferable (int, int, int, int);
579 static void new_basic_block (void);
580 static void make_new_qty (unsigned int, enum machine_mode);
581 static void make_regs_eqv (unsigned int, unsigned int);
582 static void delete_reg_equiv (unsigned int);
583 static int mention_regs (rtx);
584 static int insert_regs (rtx, struct table_elt *, int);
585 static void remove_from_table (struct table_elt *, unsigned);
586 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
587 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
588 static rtx lookup_as_function (rtx, enum rtx_code);
589 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
590 enum machine_mode);
591 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
592 static void invalidate (rtx, enum machine_mode);
593 static int cse_rtx_varies_p (rtx, int);
594 static void remove_invalid_refs (unsigned int);
595 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
596 enum machine_mode);
597 static void rehash_using_reg (rtx);
598 static void invalidate_memory (void);
599 static void invalidate_for_call (void);
600 static rtx use_related_value (rtx, struct table_elt *);
602 static inline unsigned canon_hash (rtx, enum machine_mode);
603 static inline unsigned safe_hash (rtx, enum machine_mode);
604 static unsigned hash_rtx_string (const char *);
606 static rtx canon_reg (rtx, rtx);
607 static void find_best_addr (rtx, rtx *, enum machine_mode);
608 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
609 enum machine_mode *,
610 enum machine_mode *);
611 static rtx fold_rtx (rtx, rtx);
612 static rtx equiv_constant (rtx);
613 static void record_jump_equiv (rtx, int);
614 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
615 int);
616 static void cse_insn (rtx, rtx);
617 static void cse_end_of_basic_block (rtx, struct cse_basic_block_data *,
618 int, int);
619 static int addr_affects_sp_p (rtx);
620 static void invalidate_from_clobbers (rtx);
621 static rtx cse_process_notes (rtx, rtx);
622 static void invalidate_skipped_set (rtx, rtx, void *);
623 static void invalidate_skipped_block (rtx);
624 static rtx cse_basic_block (rtx, rtx, struct branch_path *);
625 static void count_reg_usage (rtx, int *, rtx, int);
626 static int check_for_label_ref (rtx *, void *);
627 extern void dump_class (struct table_elt*);
628 static void get_cse_reg_info_1 (unsigned int regno);
629 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
630 static int check_dependence (rtx *, void *);
632 static void flush_hash_table (void);
633 static bool insn_live_p (rtx, int *);
634 static bool set_live_p (rtx, rtx, int *);
635 static bool dead_libcall_p (rtx, int *);
636 static int cse_change_cc_mode (rtx *, void *);
637 static void cse_change_cc_mode_insn (rtx, rtx);
638 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
639 static enum machine_mode cse_cc_succs (basic_block, rtx, rtx, bool);
642 #undef RTL_HOOKS_GEN_LOWPART
643 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
645 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
647 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
648 virtual regs here because the simplify_*_operation routines are called
649 by integrate.c, which is called before virtual register instantiation. */
651 static bool
652 fixed_base_plus_p (rtx x)
654 switch (GET_CODE (x))
656 case REG:
657 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
658 return true;
659 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
660 return true;
661 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
662 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
663 return true;
664 return false;
666 case PLUS:
667 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
668 return false;
669 return fixed_base_plus_p (XEXP (x, 0));
671 default:
672 return false;
676 /* Dump the expressions in the equivalence class indicated by CLASSP.
677 This function is used only for debugging. */
678 void
679 dump_class (struct table_elt *classp)
681 struct table_elt *elt;
683 fprintf (stderr, "Equivalence chain for ");
684 print_rtl (stderr, classp->exp);
685 fprintf (stderr, ": \n");
687 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
689 print_rtl (stderr, elt->exp);
690 fprintf (stderr, "\n");
694 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
696 static int
697 approx_reg_cost_1 (rtx *xp, void *data)
699 rtx x = *xp;
700 int *cost_p = data;
702 if (x && REG_P (x))
704 unsigned int regno = REGNO (x);
706 if (! CHEAP_REGNO (regno))
708 if (regno < FIRST_PSEUDO_REGISTER)
710 if (SMALL_REGISTER_CLASSES)
711 return 1;
712 *cost_p += 2;
714 else
715 *cost_p += 1;
719 return 0;
722 /* Return an estimate of the cost of the registers used in an rtx.
723 This is mostly the number of different REG expressions in the rtx;
724 however for some exceptions like fixed registers we use a cost of
725 0. If any other hard register reference occurs, return MAX_COST. */
727 static int
728 approx_reg_cost (rtx x)
730 int cost = 0;
732 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
733 return MAX_COST;
735 return cost;
738 /* Returns a canonical version of X for the address, from the point of view,
739 that all multiplications are represented as MULT instead of the multiply
740 by a power of 2 being represented as ASHIFT. */
742 static rtx
743 canon_for_address (rtx x)
745 enum rtx_code code;
746 enum machine_mode mode;
747 rtx new = 0;
748 int i;
749 const char *fmt;
751 if (!x)
752 return x;
754 code = GET_CODE (x);
755 mode = GET_MODE (x);
757 switch (code)
759 case ASHIFT:
760 if (GET_CODE (XEXP (x, 1)) == CONST_INT
761 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode)
762 && INTVAL (XEXP (x, 1)) >= 0)
764 new = canon_for_address (XEXP (x, 0));
765 new = gen_rtx_MULT (mode, new,
766 gen_int_mode ((HOST_WIDE_INT) 1
767 << INTVAL (XEXP (x, 1)),
768 mode));
770 break;
771 default:
772 break;
775 if (new)
776 return new;
778 /* Now recursively process each operand of this operation. */
779 fmt = GET_RTX_FORMAT (code);
780 for (i = 0; i < GET_RTX_LENGTH (code); i++)
781 if (fmt[i] == 'e')
783 new = canon_for_address (XEXP (x, i));
784 XEXP (x, i) = new;
786 return x;
789 /* Return a negative value if an rtx A, whose costs are given by COST_A
790 and REGCOST_A, is more desirable than an rtx B.
791 Return a positive value if A is less desirable, or 0 if the two are
792 equally good. */
793 static int
794 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
796 /* First, get rid of cases involving expressions that are entirely
797 unwanted. */
798 if (cost_a != cost_b)
800 if (cost_a == MAX_COST)
801 return 1;
802 if (cost_b == MAX_COST)
803 return -1;
806 /* Avoid extending lifetimes of hardregs. */
807 if (regcost_a != regcost_b)
809 if (regcost_a == MAX_COST)
810 return 1;
811 if (regcost_b == MAX_COST)
812 return -1;
815 /* Normal operation costs take precedence. */
816 if (cost_a != cost_b)
817 return cost_a - cost_b;
818 /* Only if these are identical consider effects on register pressure. */
819 if (regcost_a != regcost_b)
820 return regcost_a - regcost_b;
821 return 0;
824 /* Internal function, to compute cost when X is not a register; called
825 from COST macro to keep it simple. */
827 static int
828 notreg_cost (rtx x, enum rtx_code outer)
830 return ((GET_CODE (x) == SUBREG
831 && REG_P (SUBREG_REG (x))
832 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
833 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
834 && (GET_MODE_SIZE (GET_MODE (x))
835 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
836 && subreg_lowpart_p (x)
837 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
838 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
840 : rtx_cost (x, outer) * 2);
844 /* Initialize CSE_REG_INFO_TABLE. */
846 static void
847 init_cse_reg_info (unsigned int nregs)
849 /* Do we need to grow the table? */
850 if (nregs > cse_reg_info_table_size)
852 unsigned int new_size;
854 if (cse_reg_info_table_size < 2048)
856 /* Compute a new size that is a power of 2 and no smaller
857 than the large of NREGS and 64. */
858 new_size = (cse_reg_info_table_size
859 ? cse_reg_info_table_size : 64);
861 while (new_size < nregs)
862 new_size *= 2;
864 else
866 /* If we need a big table, allocate just enough to hold
867 NREGS registers. */
868 new_size = nregs;
871 /* Reallocate the table with NEW_SIZE entries. */
872 if (cse_reg_info_table)
873 free (cse_reg_info_table);
874 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
875 cse_reg_info_table_size = new_size;
876 cse_reg_info_table_first_uninitialized = 0;
879 /* Do we have all of the first NREGS entries initialized? */
880 if (cse_reg_info_table_first_uninitialized < nregs)
882 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
883 unsigned int i;
885 /* Put the old timestamp on newly allocated entries so that they
886 will all be considered out of date. We do not touch those
887 entries beyond the first NREGS entries to be nice to the
888 virtual memory. */
889 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
890 cse_reg_info_table[i].timestamp = old_timestamp;
892 cse_reg_info_table_first_uninitialized = nregs;
896 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
898 static void
899 get_cse_reg_info_1 (unsigned int regno)
901 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
902 entry will be considered to have been initialized. */
903 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
905 /* Initialize the rest of the entry. */
906 cse_reg_info_table[regno].reg_tick = 1;
907 cse_reg_info_table[regno].reg_in_table = -1;
908 cse_reg_info_table[regno].subreg_ticked = -1;
909 cse_reg_info_table[regno].reg_qty = -regno - 1;
912 /* Find a cse_reg_info entry for REGNO. */
914 static inline struct cse_reg_info *
915 get_cse_reg_info (unsigned int regno)
917 struct cse_reg_info *p = &cse_reg_info_table[regno];
919 /* If this entry has not been initialized, go ahead and initialize
920 it. */
921 if (p->timestamp != cse_reg_info_timestamp)
922 get_cse_reg_info_1 (regno);
924 return p;
927 /* Clear the hash table and initialize each register with its own quantity,
928 for a new basic block. */
930 static void
931 new_basic_block (void)
933 int i;
935 next_qty = 0;
937 /* Invalidate cse_reg_info_table. */
938 cse_reg_info_timestamp++;
940 /* Clear out hash table state for this pass. */
941 CLEAR_HARD_REG_SET (hard_regs_in_table);
943 /* The per-quantity values used to be initialized here, but it is
944 much faster to initialize each as it is made in `make_new_qty'. */
946 for (i = 0; i < HASH_SIZE; i++)
948 struct table_elt *first;
950 first = table[i];
951 if (first != NULL)
953 struct table_elt *last = first;
955 table[i] = NULL;
957 while (last->next_same_hash != NULL)
958 last = last->next_same_hash;
960 /* Now relink this hash entire chain into
961 the free element list. */
963 last->next_same_hash = free_element_chain;
964 free_element_chain = first;
968 table_size = 0;
970 #ifdef HAVE_cc0
971 prev_insn = 0;
972 prev_insn_cc0 = 0;
973 #endif
976 /* Say that register REG contains a quantity in mode MODE not in any
977 register before and initialize that quantity. */
979 static void
980 make_new_qty (unsigned int reg, enum machine_mode mode)
982 int q;
983 struct qty_table_elem *ent;
984 struct reg_eqv_elem *eqv;
986 gcc_assert (next_qty < max_qty);
988 q = REG_QTY (reg) = next_qty++;
989 ent = &qty_table[q];
990 ent->first_reg = reg;
991 ent->last_reg = reg;
992 ent->mode = mode;
993 ent->const_rtx = ent->const_insn = NULL_RTX;
994 ent->comparison_code = UNKNOWN;
996 eqv = &reg_eqv_table[reg];
997 eqv->next = eqv->prev = -1;
1000 /* Make reg NEW equivalent to reg OLD.
1001 OLD is not changing; NEW is. */
1003 static void
1004 make_regs_eqv (unsigned int new, unsigned int old)
1006 unsigned int lastr, firstr;
1007 int q = REG_QTY (old);
1008 struct qty_table_elem *ent;
1010 ent = &qty_table[q];
1012 /* Nothing should become eqv until it has a "non-invalid" qty number. */
1013 gcc_assert (REGNO_QTY_VALID_P (old));
1015 REG_QTY (new) = q;
1016 firstr = ent->first_reg;
1017 lastr = ent->last_reg;
1019 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
1020 hard regs. Among pseudos, if NEW will live longer than any other reg
1021 of the same qty, and that is beyond the current basic block,
1022 make it the new canonical replacement for this qty. */
1023 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
1024 /* Certain fixed registers might be of the class NO_REGS. This means
1025 that not only can they not be allocated by the compiler, but
1026 they cannot be used in substitutions or canonicalizations
1027 either. */
1028 && (new >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new) != NO_REGS)
1029 && ((new < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new))
1030 || (new >= FIRST_PSEUDO_REGISTER
1031 && (firstr < FIRST_PSEUDO_REGISTER
1032 || ((uid_cuid[REGNO_LAST_UID (new)] > cse_basic_block_end
1033 || (uid_cuid[REGNO_FIRST_UID (new)]
1034 < cse_basic_block_start))
1035 && (uid_cuid[REGNO_LAST_UID (new)]
1036 > uid_cuid[REGNO_LAST_UID (firstr)]))))))
1038 reg_eqv_table[firstr].prev = new;
1039 reg_eqv_table[new].next = firstr;
1040 reg_eqv_table[new].prev = -1;
1041 ent->first_reg = new;
1043 else
1045 /* If NEW is a hard reg (known to be non-fixed), insert at end.
1046 Otherwise, insert before any non-fixed hard regs that are at the
1047 end. Registers of class NO_REGS cannot be used as an
1048 equivalent for anything. */
1049 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
1050 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
1051 && new >= FIRST_PSEUDO_REGISTER)
1052 lastr = reg_eqv_table[lastr].prev;
1053 reg_eqv_table[new].next = reg_eqv_table[lastr].next;
1054 if (reg_eqv_table[lastr].next >= 0)
1055 reg_eqv_table[reg_eqv_table[lastr].next].prev = new;
1056 else
1057 qty_table[q].last_reg = new;
1058 reg_eqv_table[lastr].next = new;
1059 reg_eqv_table[new].prev = lastr;
1063 /* Remove REG from its equivalence class. */
1065 static void
1066 delete_reg_equiv (unsigned int reg)
1068 struct qty_table_elem *ent;
1069 int q = REG_QTY (reg);
1070 int p, n;
1072 /* If invalid, do nothing. */
1073 if (! REGNO_QTY_VALID_P (reg))
1074 return;
1076 ent = &qty_table[q];
1078 p = reg_eqv_table[reg].prev;
1079 n = reg_eqv_table[reg].next;
1081 if (n != -1)
1082 reg_eqv_table[n].prev = p;
1083 else
1084 ent->last_reg = p;
1085 if (p != -1)
1086 reg_eqv_table[p].next = n;
1087 else
1088 ent->first_reg = n;
1090 REG_QTY (reg) = -reg - 1;
1093 /* Remove any invalid expressions from the hash table
1094 that refer to any of the registers contained in expression X.
1096 Make sure that newly inserted references to those registers
1097 as subexpressions will be considered valid.
1099 mention_regs is not called when a register itself
1100 is being stored in the table.
1102 Return 1 if we have done something that may have changed the hash code
1103 of X. */
1105 static int
1106 mention_regs (rtx x)
1108 enum rtx_code code;
1109 int i, j;
1110 const char *fmt;
1111 int changed = 0;
1113 if (x == 0)
1114 return 0;
1116 code = GET_CODE (x);
1117 if (code == REG)
1119 unsigned int regno = REGNO (x);
1120 unsigned int endregno
1121 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
1122 : hard_regno_nregs[regno][GET_MODE (x)]);
1123 unsigned int i;
1125 for (i = regno; i < endregno; i++)
1127 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1128 remove_invalid_refs (i);
1130 REG_IN_TABLE (i) = REG_TICK (i);
1131 SUBREG_TICKED (i) = -1;
1134 return 0;
1137 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1138 pseudo if they don't use overlapping words. We handle only pseudos
1139 here for simplicity. */
1140 if (code == SUBREG && REG_P (SUBREG_REG (x))
1141 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1143 unsigned int i = REGNO (SUBREG_REG (x));
1145 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1147 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1148 the last store to this register really stored into this
1149 subreg, then remove the memory of this subreg.
1150 Otherwise, remove any memory of the entire register and
1151 all its subregs from the table. */
1152 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1153 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1154 remove_invalid_refs (i);
1155 else
1156 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1159 REG_IN_TABLE (i) = REG_TICK (i);
1160 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1161 return 0;
1164 /* If X is a comparison or a COMPARE and either operand is a register
1165 that does not have a quantity, give it one. This is so that a later
1166 call to record_jump_equiv won't cause X to be assigned a different
1167 hash code and not found in the table after that call.
1169 It is not necessary to do this here, since rehash_using_reg can
1170 fix up the table later, but doing this here eliminates the need to
1171 call that expensive function in the most common case where the only
1172 use of the register is in the comparison. */
1174 if (code == COMPARE || COMPARISON_P (x))
1176 if (REG_P (XEXP (x, 0))
1177 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1178 if (insert_regs (XEXP (x, 0), NULL, 0))
1180 rehash_using_reg (XEXP (x, 0));
1181 changed = 1;
1184 if (REG_P (XEXP (x, 1))
1185 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1186 if (insert_regs (XEXP (x, 1), NULL, 0))
1188 rehash_using_reg (XEXP (x, 1));
1189 changed = 1;
1193 fmt = GET_RTX_FORMAT (code);
1194 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1195 if (fmt[i] == 'e')
1196 changed |= mention_regs (XEXP (x, i));
1197 else if (fmt[i] == 'E')
1198 for (j = 0; j < XVECLEN (x, i); j++)
1199 changed |= mention_regs (XVECEXP (x, i, j));
1201 return changed;
1204 /* Update the register quantities for inserting X into the hash table
1205 with a value equivalent to CLASSP.
1206 (If the class does not contain a REG, it is irrelevant.)
1207 If MODIFIED is nonzero, X is a destination; it is being modified.
1208 Note that delete_reg_equiv should be called on a register
1209 before insert_regs is done on that register with MODIFIED != 0.
1211 Nonzero value means that elements of reg_qty have changed
1212 so X's hash code may be different. */
1214 static int
1215 insert_regs (rtx x, struct table_elt *classp, int modified)
1217 if (REG_P (x))
1219 unsigned int regno = REGNO (x);
1220 int qty_valid;
1222 /* If REGNO is in the equivalence table already but is of the
1223 wrong mode for that equivalence, don't do anything here. */
1225 qty_valid = REGNO_QTY_VALID_P (regno);
1226 if (qty_valid)
1228 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1230 if (ent->mode != GET_MODE (x))
1231 return 0;
1234 if (modified || ! qty_valid)
1236 if (classp)
1237 for (classp = classp->first_same_value;
1238 classp != 0;
1239 classp = classp->next_same_value)
1240 if (REG_P (classp->exp)
1241 && GET_MODE (classp->exp) == GET_MODE (x))
1243 unsigned c_regno = REGNO (classp->exp);
1245 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1247 /* Suppose that 5 is hard reg and 100 and 101 are
1248 pseudos. Consider
1250 (set (reg:si 100) (reg:si 5))
1251 (set (reg:si 5) (reg:si 100))
1252 (set (reg:di 101) (reg:di 5))
1254 We would now set REG_QTY (101) = REG_QTY (5), but the
1255 entry for 5 is in SImode. When we use this later in
1256 copy propagation, we get the register in wrong mode. */
1257 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1258 continue;
1260 make_regs_eqv (regno, c_regno);
1261 return 1;
1264 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1265 than REG_IN_TABLE to find out if there was only a single preceding
1266 invalidation - for the SUBREG - or another one, which would be
1267 for the full register. However, if we find here that REG_TICK
1268 indicates that the register is invalid, it means that it has
1269 been invalidated in a separate operation. The SUBREG might be used
1270 now (then this is a recursive call), or we might use the full REG
1271 now and a SUBREG of it later. So bump up REG_TICK so that
1272 mention_regs will do the right thing. */
1273 if (! modified
1274 && REG_IN_TABLE (regno) >= 0
1275 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1276 REG_TICK (regno)++;
1277 make_new_qty (regno, GET_MODE (x));
1278 return 1;
1281 return 0;
1284 /* If X is a SUBREG, we will likely be inserting the inner register in the
1285 table. If that register doesn't have an assigned quantity number at
1286 this point but does later, the insertion that we will be doing now will
1287 not be accessible because its hash code will have changed. So assign
1288 a quantity number now. */
1290 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1291 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1293 insert_regs (SUBREG_REG (x), NULL, 0);
1294 mention_regs (x);
1295 return 1;
1297 else
1298 return mention_regs (x);
1301 /* Look in or update the hash table. */
1303 /* Remove table element ELT from use in the table.
1304 HASH is its hash code, made using the HASH macro.
1305 It's an argument because often that is known in advance
1306 and we save much time not recomputing it. */
1308 static void
1309 remove_from_table (struct table_elt *elt, unsigned int hash)
1311 if (elt == 0)
1312 return;
1314 /* Mark this element as removed. See cse_insn. */
1315 elt->first_same_value = 0;
1317 /* Remove the table element from its equivalence class. */
1320 struct table_elt *prev = elt->prev_same_value;
1321 struct table_elt *next = elt->next_same_value;
1323 if (next)
1324 next->prev_same_value = prev;
1326 if (prev)
1327 prev->next_same_value = next;
1328 else
1330 struct table_elt *newfirst = next;
1331 while (next)
1333 next->first_same_value = newfirst;
1334 next = next->next_same_value;
1339 /* Remove the table element from its hash bucket. */
1342 struct table_elt *prev = elt->prev_same_hash;
1343 struct table_elt *next = elt->next_same_hash;
1345 if (next)
1346 next->prev_same_hash = prev;
1348 if (prev)
1349 prev->next_same_hash = next;
1350 else if (table[hash] == elt)
1351 table[hash] = next;
1352 else
1354 /* This entry is not in the proper hash bucket. This can happen
1355 when two classes were merged by `merge_equiv_classes'. Search
1356 for the hash bucket that it heads. This happens only very
1357 rarely, so the cost is acceptable. */
1358 for (hash = 0; hash < HASH_SIZE; hash++)
1359 if (table[hash] == elt)
1360 table[hash] = next;
1364 /* Remove the table element from its related-value circular chain. */
1366 if (elt->related_value != 0 && elt->related_value != elt)
1368 struct table_elt *p = elt->related_value;
1370 while (p->related_value != elt)
1371 p = p->related_value;
1372 p->related_value = elt->related_value;
1373 if (p->related_value == p)
1374 p->related_value = 0;
1377 /* Now add it to the free element chain. */
1378 elt->next_same_hash = free_element_chain;
1379 free_element_chain = elt;
1381 table_size--;
1384 /* Look up X in the hash table and return its table element,
1385 or 0 if X is not in the table.
1387 MODE is the machine-mode of X, or if X is an integer constant
1388 with VOIDmode then MODE is the mode with which X will be used.
1390 Here we are satisfied to find an expression whose tree structure
1391 looks like X. */
1393 static struct table_elt *
1394 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1396 struct table_elt *p;
1398 for (p = table[hash]; p; p = p->next_same_hash)
1399 if (mode == p->mode && ((x == p->exp && REG_P (x))
1400 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1401 return p;
1403 return 0;
1406 /* Like `lookup' but don't care whether the table element uses invalid regs.
1407 Also ignore discrepancies in the machine mode of a register. */
1409 static struct table_elt *
1410 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1412 struct table_elt *p;
1414 if (REG_P (x))
1416 unsigned int regno = REGNO (x);
1418 /* Don't check the machine mode when comparing registers;
1419 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1420 for (p = table[hash]; p; p = p->next_same_hash)
1421 if (REG_P (p->exp)
1422 && REGNO (p->exp) == regno)
1423 return p;
1425 else
1427 for (p = table[hash]; p; p = p->next_same_hash)
1428 if (mode == p->mode
1429 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1430 return p;
1433 return 0;
1436 /* Look for an expression equivalent to X and with code CODE.
1437 If one is found, return that expression. */
1439 static rtx
1440 lookup_as_function (rtx x, enum rtx_code code)
1442 struct table_elt *p
1443 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1445 /* If we are looking for a CONST_INT, the mode doesn't really matter, as
1446 long as we are narrowing. So if we looked in vain for a mode narrower
1447 than word_mode before, look for word_mode now. */
1448 if (p == 0 && code == CONST_INT
1449 && GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (word_mode))
1451 x = copy_rtx (x);
1452 PUT_MODE (x, word_mode);
1453 p = lookup (x, SAFE_HASH (x, VOIDmode), word_mode);
1456 if (p == 0)
1457 return 0;
1459 for (p = p->first_same_value; p; p = p->next_same_value)
1460 if (GET_CODE (p->exp) == code
1461 /* Make sure this is a valid entry in the table. */
1462 && exp_equiv_p (p->exp, p->exp, 1, false))
1463 return p->exp;
1465 return 0;
1468 /* Insert X in the hash table, assuming HASH is its hash code
1469 and CLASSP is an element of the class it should go in
1470 (or 0 if a new class should be made).
1471 It is inserted at the proper position to keep the class in
1472 the order cheapest first.
1474 MODE is the machine-mode of X, or if X is an integer constant
1475 with VOIDmode then MODE is the mode with which X will be used.
1477 For elements of equal cheapness, the most recent one
1478 goes in front, except that the first element in the list
1479 remains first unless a cheaper element is added. The order of
1480 pseudo-registers does not matter, as canon_reg will be called to
1481 find the cheapest when a register is retrieved from the table.
1483 The in_memory field in the hash table element is set to 0.
1484 The caller must set it nonzero if appropriate.
1486 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1487 and if insert_regs returns a nonzero value
1488 you must then recompute its hash code before calling here.
1490 If necessary, update table showing constant values of quantities. */
1492 #define CHEAPER(X, Y) \
1493 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
1495 static struct table_elt *
1496 insert (rtx x, struct table_elt *classp, unsigned int hash, enum machine_mode mode)
1498 struct table_elt *elt;
1500 /* If X is a register and we haven't made a quantity for it,
1501 something is wrong. */
1502 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1504 /* If X is a hard register, show it is being put in the table. */
1505 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1507 unsigned int regno = REGNO (x);
1508 unsigned int endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
1509 unsigned int i;
1511 for (i = regno; i < endregno; i++)
1512 SET_HARD_REG_BIT (hard_regs_in_table, i);
1515 /* Put an element for X into the right hash bucket. */
1517 elt = free_element_chain;
1518 if (elt)
1519 free_element_chain = elt->next_same_hash;
1520 else
1521 elt = XNEW (struct table_elt);
1523 elt->exp = x;
1524 elt->canon_exp = NULL_RTX;
1525 elt->cost = COST (x);
1526 elt->regcost = approx_reg_cost (x);
1527 elt->next_same_value = 0;
1528 elt->prev_same_value = 0;
1529 elt->next_same_hash = table[hash];
1530 elt->prev_same_hash = 0;
1531 elt->related_value = 0;
1532 elt->in_memory = 0;
1533 elt->mode = mode;
1534 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1536 if (table[hash])
1537 table[hash]->prev_same_hash = elt;
1538 table[hash] = elt;
1540 /* Put it into the proper value-class. */
1541 if (classp)
1543 classp = classp->first_same_value;
1544 if (CHEAPER (elt, classp))
1545 /* Insert at the head of the class. */
1547 struct table_elt *p;
1548 elt->next_same_value = classp;
1549 classp->prev_same_value = elt;
1550 elt->first_same_value = elt;
1552 for (p = classp; p; p = p->next_same_value)
1553 p->first_same_value = elt;
1555 else
1557 /* Insert not at head of the class. */
1558 /* Put it after the last element cheaper than X. */
1559 struct table_elt *p, *next;
1561 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1562 p = next);
1564 /* Put it after P and before NEXT. */
1565 elt->next_same_value = next;
1566 if (next)
1567 next->prev_same_value = elt;
1569 elt->prev_same_value = p;
1570 p->next_same_value = elt;
1571 elt->first_same_value = classp;
1574 else
1575 elt->first_same_value = elt;
1577 /* If this is a constant being set equivalent to a register or a register
1578 being set equivalent to a constant, note the constant equivalence.
1580 If this is a constant, it cannot be equivalent to a different constant,
1581 and a constant is the only thing that can be cheaper than a register. So
1582 we know the register is the head of the class (before the constant was
1583 inserted).
1585 If this is a register that is not already known equivalent to a
1586 constant, we must check the entire class.
1588 If this is a register that is already known equivalent to an insn,
1589 update the qtys `const_insn' to show that `this_insn' is the latest
1590 insn making that quantity equivalent to the constant. */
1592 if (elt->is_const && classp && REG_P (classp->exp)
1593 && !REG_P (x))
1595 int exp_q = REG_QTY (REGNO (classp->exp));
1596 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1598 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1599 exp_ent->const_insn = this_insn;
1602 else if (REG_P (x)
1603 && classp
1604 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1605 && ! elt->is_const)
1607 struct table_elt *p;
1609 for (p = classp; p != 0; p = p->next_same_value)
1611 if (p->is_const && !REG_P (p->exp))
1613 int x_q = REG_QTY (REGNO (x));
1614 struct qty_table_elem *x_ent = &qty_table[x_q];
1616 x_ent->const_rtx
1617 = gen_lowpart (GET_MODE (x), p->exp);
1618 x_ent->const_insn = this_insn;
1619 break;
1624 else if (REG_P (x)
1625 && qty_table[REG_QTY (REGNO (x))].const_rtx
1626 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1627 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1629 /* If this is a constant with symbolic value,
1630 and it has a term with an explicit integer value,
1631 link it up with related expressions. */
1632 if (GET_CODE (x) == CONST)
1634 rtx subexp = get_related_value (x);
1635 unsigned subhash;
1636 struct table_elt *subelt, *subelt_prev;
1638 if (subexp != 0)
1640 /* Get the integer-free subexpression in the hash table. */
1641 subhash = SAFE_HASH (subexp, mode);
1642 subelt = lookup (subexp, subhash, mode);
1643 if (subelt == 0)
1644 subelt = insert (subexp, NULL, subhash, mode);
1645 /* Initialize SUBELT's circular chain if it has none. */
1646 if (subelt->related_value == 0)
1647 subelt->related_value = subelt;
1648 /* Find the element in the circular chain that precedes SUBELT. */
1649 subelt_prev = subelt;
1650 while (subelt_prev->related_value != subelt)
1651 subelt_prev = subelt_prev->related_value;
1652 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1653 This way the element that follows SUBELT is the oldest one. */
1654 elt->related_value = subelt_prev->related_value;
1655 subelt_prev->related_value = elt;
1659 table_size++;
1661 return elt;
1664 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1665 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1666 the two classes equivalent.
1668 CLASS1 will be the surviving class; CLASS2 should not be used after this
1669 call.
1671 Any invalid entries in CLASS2 will not be copied. */
1673 static void
1674 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1676 struct table_elt *elt, *next, *new;
1678 /* Ensure we start with the head of the classes. */
1679 class1 = class1->first_same_value;
1680 class2 = class2->first_same_value;
1682 /* If they were already equal, forget it. */
1683 if (class1 == class2)
1684 return;
1686 for (elt = class2; elt; elt = next)
1688 unsigned int hash;
1689 rtx exp = elt->exp;
1690 enum machine_mode mode = elt->mode;
1692 next = elt->next_same_value;
1694 /* Remove old entry, make a new one in CLASS1's class.
1695 Don't do this for invalid entries as we cannot find their
1696 hash code (it also isn't necessary). */
1697 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1699 bool need_rehash = false;
1701 hash_arg_in_memory = 0;
1702 hash = HASH (exp, mode);
1704 if (REG_P (exp))
1706 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1707 delete_reg_equiv (REGNO (exp));
1710 remove_from_table (elt, hash);
1712 if (insert_regs (exp, class1, 0) || need_rehash)
1714 rehash_using_reg (exp);
1715 hash = HASH (exp, mode);
1717 new = insert (exp, class1, hash, mode);
1718 new->in_memory = hash_arg_in_memory;
1723 /* Flush the entire hash table. */
1725 static void
1726 flush_hash_table (void)
1728 int i;
1729 struct table_elt *p;
1731 for (i = 0; i < HASH_SIZE; i++)
1732 for (p = table[i]; p; p = table[i])
1734 /* Note that invalidate can remove elements
1735 after P in the current hash chain. */
1736 if (REG_P (p->exp))
1737 invalidate (p->exp, VOIDmode);
1738 else
1739 remove_from_table (p, i);
1743 /* Function called for each rtx to check whether true dependence exist. */
1744 struct check_dependence_data
1746 enum machine_mode mode;
1747 rtx exp;
1748 rtx addr;
1751 static int
1752 check_dependence (rtx *x, void *data)
1754 struct check_dependence_data *d = (struct check_dependence_data *) data;
1755 if (*x && MEM_P (*x))
1756 return canon_true_dependence (d->exp, d->mode, d->addr, *x,
1757 cse_rtx_varies_p);
1758 else
1759 return 0;
1762 /* Remove from the hash table, or mark as invalid, all expressions whose
1763 values could be altered by storing in X. X is a register, a subreg, or
1764 a memory reference with nonvarying address (because, when a memory
1765 reference with a varying address is stored in, all memory references are
1766 removed by invalidate_memory so specific invalidation is superfluous).
1767 FULL_MODE, if not VOIDmode, indicates that this much should be
1768 invalidated instead of just the amount indicated by the mode of X. This
1769 is only used for bitfield stores into memory.
1771 A nonvarying address may be just a register or just a symbol reference,
1772 or it may be either of those plus a numeric offset. */
1774 static void
1775 invalidate (rtx x, enum machine_mode full_mode)
1777 int i;
1778 struct table_elt *p;
1779 rtx addr;
1781 switch (GET_CODE (x))
1783 case REG:
1785 /* If X is a register, dependencies on its contents are recorded
1786 through the qty number mechanism. Just change the qty number of
1787 the register, mark it as invalid for expressions that refer to it,
1788 and remove it itself. */
1789 unsigned int regno = REGNO (x);
1790 unsigned int hash = HASH (x, GET_MODE (x));
1792 /* Remove REGNO from any quantity list it might be on and indicate
1793 that its value might have changed. If it is a pseudo, remove its
1794 entry from the hash table.
1796 For a hard register, we do the first two actions above for any
1797 additional hard registers corresponding to X. Then, if any of these
1798 registers are in the table, we must remove any REG entries that
1799 overlap these registers. */
1801 delete_reg_equiv (regno);
1802 REG_TICK (regno)++;
1803 SUBREG_TICKED (regno) = -1;
1805 if (regno >= FIRST_PSEUDO_REGISTER)
1807 /* Because a register can be referenced in more than one mode,
1808 we might have to remove more than one table entry. */
1809 struct table_elt *elt;
1811 while ((elt = lookup_for_remove (x, hash, GET_MODE (x))))
1812 remove_from_table (elt, hash);
1814 else
1816 HOST_WIDE_INT in_table
1817 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1818 unsigned int endregno
1819 = regno + hard_regno_nregs[regno][GET_MODE (x)];
1820 unsigned int tregno, tendregno, rn;
1821 struct table_elt *p, *next;
1823 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1825 for (rn = regno + 1; rn < endregno; rn++)
1827 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1828 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1829 delete_reg_equiv (rn);
1830 REG_TICK (rn)++;
1831 SUBREG_TICKED (rn) = -1;
1834 if (in_table)
1835 for (hash = 0; hash < HASH_SIZE; hash++)
1836 for (p = table[hash]; p; p = next)
1838 next = p->next_same_hash;
1840 if (!REG_P (p->exp)
1841 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1842 continue;
1844 tregno = REGNO (p->exp);
1845 tendregno
1846 = tregno + hard_regno_nregs[tregno][GET_MODE (p->exp)];
1847 if (tendregno > regno && tregno < endregno)
1848 remove_from_table (p, hash);
1852 return;
1854 case SUBREG:
1855 invalidate (SUBREG_REG (x), VOIDmode);
1856 return;
1858 case PARALLEL:
1859 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1860 invalidate (XVECEXP (x, 0, i), VOIDmode);
1861 return;
1863 case EXPR_LIST:
1864 /* This is part of a disjoint return value; extract the location in
1865 question ignoring the offset. */
1866 invalidate (XEXP (x, 0), VOIDmode);
1867 return;
1869 case MEM:
1870 addr = canon_rtx (get_addr (XEXP (x, 0)));
1871 /* Calculate the canonical version of X here so that
1872 true_dependence doesn't generate new RTL for X on each call. */
1873 x = canon_rtx (x);
1875 /* Remove all hash table elements that refer to overlapping pieces of
1876 memory. */
1877 if (full_mode == VOIDmode)
1878 full_mode = GET_MODE (x);
1880 for (i = 0; i < HASH_SIZE; i++)
1882 struct table_elt *next;
1884 for (p = table[i]; p; p = next)
1886 next = p->next_same_hash;
1887 if (p->in_memory)
1889 struct check_dependence_data d;
1891 /* Just canonicalize the expression once;
1892 otherwise each time we call invalidate
1893 true_dependence will canonicalize the
1894 expression again. */
1895 if (!p->canon_exp)
1896 p->canon_exp = canon_rtx (p->exp);
1897 d.exp = x;
1898 d.addr = addr;
1899 d.mode = full_mode;
1900 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1901 remove_from_table (p, i);
1905 return;
1907 default:
1908 gcc_unreachable ();
1912 /* Remove all expressions that refer to register REGNO,
1913 since they are already invalid, and we are about to
1914 mark that register valid again and don't want the old
1915 expressions to reappear as valid. */
1917 static void
1918 remove_invalid_refs (unsigned int regno)
1920 unsigned int i;
1921 struct table_elt *p, *next;
1923 for (i = 0; i < HASH_SIZE; i++)
1924 for (p = table[i]; p; p = next)
1926 next = p->next_same_hash;
1927 if (!REG_P (p->exp)
1928 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1929 remove_from_table (p, i);
1933 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1934 and mode MODE. */
1935 static void
1936 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1937 enum machine_mode mode)
1939 unsigned int i;
1940 struct table_elt *p, *next;
1941 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
1943 for (i = 0; i < HASH_SIZE; i++)
1944 for (p = table[i]; p; p = next)
1946 rtx exp = p->exp;
1947 next = p->next_same_hash;
1949 if (!REG_P (exp)
1950 && (GET_CODE (exp) != SUBREG
1951 || !REG_P (SUBREG_REG (exp))
1952 || REGNO (SUBREG_REG (exp)) != regno
1953 || (((SUBREG_BYTE (exp)
1954 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
1955 && SUBREG_BYTE (exp) <= end))
1956 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
1957 remove_from_table (p, i);
1961 /* Recompute the hash codes of any valid entries in the hash table that
1962 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
1964 This is called when we make a jump equivalence. */
1966 static void
1967 rehash_using_reg (rtx x)
1969 unsigned int i;
1970 struct table_elt *p, *next;
1971 unsigned hash;
1973 if (GET_CODE (x) == SUBREG)
1974 x = SUBREG_REG (x);
1976 /* If X is not a register or if the register is known not to be in any
1977 valid entries in the table, we have no work to do. */
1979 if (!REG_P (x)
1980 || REG_IN_TABLE (REGNO (x)) < 0
1981 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
1982 return;
1984 /* Scan all hash chains looking for valid entries that mention X.
1985 If we find one and it is in the wrong hash chain, move it. */
1987 for (i = 0; i < HASH_SIZE; i++)
1988 for (p = table[i]; p; p = next)
1990 next = p->next_same_hash;
1991 if (reg_mentioned_p (x, p->exp)
1992 && exp_equiv_p (p->exp, p->exp, 1, false)
1993 && i != (hash = SAFE_HASH (p->exp, p->mode)))
1995 if (p->next_same_hash)
1996 p->next_same_hash->prev_same_hash = p->prev_same_hash;
1998 if (p->prev_same_hash)
1999 p->prev_same_hash->next_same_hash = p->next_same_hash;
2000 else
2001 table[i] = p->next_same_hash;
2003 p->next_same_hash = table[hash];
2004 p->prev_same_hash = 0;
2005 if (table[hash])
2006 table[hash]->prev_same_hash = p;
2007 table[hash] = p;
2012 /* Remove from the hash table any expression that is a call-clobbered
2013 register. Also update their TICK values. */
2015 static void
2016 invalidate_for_call (void)
2018 unsigned int regno, endregno;
2019 unsigned int i;
2020 unsigned hash;
2021 struct table_elt *p, *next;
2022 int in_table = 0;
2024 /* Go through all the hard registers. For each that is clobbered in
2025 a CALL_INSN, remove the register from quantity chains and update
2026 reg_tick if defined. Also see if any of these registers is currently
2027 in the table. */
2029 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2030 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2032 delete_reg_equiv (regno);
2033 if (REG_TICK (regno) >= 0)
2035 REG_TICK (regno)++;
2036 SUBREG_TICKED (regno) = -1;
2039 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2042 /* In the case where we have no call-clobbered hard registers in the
2043 table, we are done. Otherwise, scan the table and remove any
2044 entry that overlaps a call-clobbered register. */
2046 if (in_table)
2047 for (hash = 0; hash < HASH_SIZE; hash++)
2048 for (p = table[hash]; p; p = next)
2050 next = p->next_same_hash;
2052 if (!REG_P (p->exp)
2053 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2054 continue;
2056 regno = REGNO (p->exp);
2057 endregno = regno + hard_regno_nregs[regno][GET_MODE (p->exp)];
2059 for (i = regno; i < endregno; i++)
2060 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2062 remove_from_table (p, hash);
2063 break;
2068 /* Given an expression X of type CONST,
2069 and ELT which is its table entry (or 0 if it
2070 is not in the hash table),
2071 return an alternate expression for X as a register plus integer.
2072 If none can be found, return 0. */
2074 static rtx
2075 use_related_value (rtx x, struct table_elt *elt)
2077 struct table_elt *relt = 0;
2078 struct table_elt *p, *q;
2079 HOST_WIDE_INT offset;
2081 /* First, is there anything related known?
2082 If we have a table element, we can tell from that.
2083 Otherwise, must look it up. */
2085 if (elt != 0 && elt->related_value != 0)
2086 relt = elt;
2087 else if (elt == 0 && GET_CODE (x) == CONST)
2089 rtx subexp = get_related_value (x);
2090 if (subexp != 0)
2091 relt = lookup (subexp,
2092 SAFE_HASH (subexp, GET_MODE (subexp)),
2093 GET_MODE (subexp));
2096 if (relt == 0)
2097 return 0;
2099 /* Search all related table entries for one that has an
2100 equivalent register. */
2102 p = relt;
2103 while (1)
2105 /* This loop is strange in that it is executed in two different cases.
2106 The first is when X is already in the table. Then it is searching
2107 the RELATED_VALUE list of X's class (RELT). The second case is when
2108 X is not in the table. Then RELT points to a class for the related
2109 value.
2111 Ensure that, whatever case we are in, that we ignore classes that have
2112 the same value as X. */
2114 if (rtx_equal_p (x, p->exp))
2115 q = 0;
2116 else
2117 for (q = p->first_same_value; q; q = q->next_same_value)
2118 if (REG_P (q->exp))
2119 break;
2121 if (q)
2122 break;
2124 p = p->related_value;
2126 /* We went all the way around, so there is nothing to be found.
2127 Alternatively, perhaps RELT was in the table for some other reason
2128 and it has no related values recorded. */
2129 if (p == relt || p == 0)
2130 break;
2133 if (q == 0)
2134 return 0;
2136 offset = (get_integer_term (x) - get_integer_term (p->exp));
2137 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2138 return plus_constant (q->exp, offset);
2141 /* Hash a string. Just add its bytes up. */
2142 static inline unsigned
2143 hash_rtx_string (const char *ps)
2145 unsigned hash = 0;
2146 const unsigned char *p = (const unsigned char *) ps;
2148 if (p)
2149 while (*p)
2150 hash += *p++;
2152 return hash;
2155 /* Hash an rtx. We are careful to make sure the value is never negative.
2156 Equivalent registers hash identically.
2157 MODE is used in hashing for CONST_INTs only;
2158 otherwise the mode of X is used.
2160 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2162 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2163 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2165 Note that cse_insn knows that the hash code of a MEM expression
2166 is just (int) MEM plus the hash code of the address. */
2168 unsigned
2169 hash_rtx (rtx x, enum machine_mode mode, int *do_not_record_p,
2170 int *hash_arg_in_memory_p, bool have_reg_qty)
2172 int i, j;
2173 unsigned hash = 0;
2174 enum rtx_code code;
2175 const char *fmt;
2177 /* Used to turn recursion into iteration. We can't rely on GCC's
2178 tail-recursion elimination since we need to keep accumulating values
2179 in HASH. */
2180 repeat:
2181 if (x == 0)
2182 return hash;
2184 code = GET_CODE (x);
2185 switch (code)
2187 case REG:
2189 unsigned int regno = REGNO (x);
2191 if (!reload_completed)
2193 /* On some machines, we can't record any non-fixed hard register,
2194 because extending its life will cause reload problems. We
2195 consider ap, fp, sp, gp to be fixed for this purpose.
2197 We also consider CCmode registers to be fixed for this purpose;
2198 failure to do so leads to failure to simplify 0<100 type of
2199 conditionals.
2201 On all machines, we can't record any global registers.
2202 Nor should we record any register that is in a small
2203 class, as defined by CLASS_LIKELY_SPILLED_P. */
2204 bool record;
2206 if (regno >= FIRST_PSEUDO_REGISTER)
2207 record = true;
2208 else if (x == frame_pointer_rtx
2209 || x == hard_frame_pointer_rtx
2210 || x == arg_pointer_rtx
2211 || x == stack_pointer_rtx
2212 || x == pic_offset_table_rtx)
2213 record = true;
2214 else if (global_regs[regno])
2215 record = false;
2216 else if (fixed_regs[regno])
2217 record = true;
2218 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2219 record = true;
2220 else if (SMALL_REGISTER_CLASSES)
2221 record = false;
2222 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2223 record = false;
2224 else
2225 record = true;
2227 if (!record)
2229 *do_not_record_p = 1;
2230 return 0;
2234 hash += ((unsigned int) REG << 7);
2235 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2236 return hash;
2239 /* We handle SUBREG of a REG specially because the underlying
2240 reg changes its hash value with every value change; we don't
2241 want to have to forget unrelated subregs when one subreg changes. */
2242 case SUBREG:
2244 if (REG_P (SUBREG_REG (x)))
2246 hash += (((unsigned int) SUBREG << 7)
2247 + REGNO (SUBREG_REG (x))
2248 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2249 return hash;
2251 break;
2254 case CONST_INT:
2255 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2256 + (unsigned int) INTVAL (x));
2257 return hash;
2259 case CONST_DOUBLE:
2260 /* This is like the general case, except that it only counts
2261 the integers representing the constant. */
2262 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2263 if (GET_MODE (x) != VOIDmode)
2264 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2265 else
2266 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2267 + (unsigned int) CONST_DOUBLE_HIGH (x));
2268 return hash;
2270 case CONST_VECTOR:
2272 int units;
2273 rtx elt;
2275 units = CONST_VECTOR_NUNITS (x);
2277 for (i = 0; i < units; ++i)
2279 elt = CONST_VECTOR_ELT (x, i);
2280 hash += hash_rtx (elt, GET_MODE (elt), do_not_record_p,
2281 hash_arg_in_memory_p, have_reg_qty);
2284 return hash;
2287 /* Assume there is only one rtx object for any given label. */
2288 case LABEL_REF:
2289 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2290 differences and differences between each stage's debugging dumps. */
2291 hash += (((unsigned int) LABEL_REF << 7)
2292 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2293 return hash;
2295 case SYMBOL_REF:
2297 /* Don't hash on the symbol's address to avoid bootstrap differences.
2298 Different hash values may cause expressions to be recorded in
2299 different orders and thus different registers to be used in the
2300 final assembler. This also avoids differences in the dump files
2301 between various stages. */
2302 unsigned int h = 0;
2303 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2305 while (*p)
2306 h += (h << 7) + *p++; /* ??? revisit */
2308 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2309 return hash;
2312 case MEM:
2313 /* We don't record if marked volatile or if BLKmode since we don't
2314 know the size of the move. */
2315 if (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode)
2317 *do_not_record_p = 1;
2318 return 0;
2320 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2321 *hash_arg_in_memory_p = 1;
2323 /* Now that we have already found this special case,
2324 might as well speed it up as much as possible. */
2325 hash += (unsigned) MEM;
2326 x = XEXP (x, 0);
2327 goto repeat;
2329 case USE:
2330 /* A USE that mentions non-volatile memory needs special
2331 handling since the MEM may be BLKmode which normally
2332 prevents an entry from being made. Pure calls are
2333 marked by a USE which mentions BLKmode memory.
2334 See calls.c:emit_call_1. */
2335 if (MEM_P (XEXP (x, 0))
2336 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2338 hash += (unsigned) USE;
2339 x = XEXP (x, 0);
2341 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2342 *hash_arg_in_memory_p = 1;
2344 /* Now that we have already found this special case,
2345 might as well speed it up as much as possible. */
2346 hash += (unsigned) MEM;
2347 x = XEXP (x, 0);
2348 goto repeat;
2350 break;
2352 case PRE_DEC:
2353 case PRE_INC:
2354 case POST_DEC:
2355 case POST_INC:
2356 case PRE_MODIFY:
2357 case POST_MODIFY:
2358 case PC:
2359 case CC0:
2360 case CALL:
2361 case UNSPEC_VOLATILE:
2362 *do_not_record_p = 1;
2363 return 0;
2365 case ASM_OPERANDS:
2366 if (MEM_VOLATILE_P (x))
2368 *do_not_record_p = 1;
2369 return 0;
2371 else
2373 /* We don't want to take the filename and line into account. */
2374 hash += (unsigned) code + (unsigned) GET_MODE (x)
2375 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2376 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2377 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2379 if (ASM_OPERANDS_INPUT_LENGTH (x))
2381 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2383 hash += (hash_rtx (ASM_OPERANDS_INPUT (x, i),
2384 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2385 do_not_record_p, hash_arg_in_memory_p,
2386 have_reg_qty)
2387 + hash_rtx_string
2388 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2391 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2392 x = ASM_OPERANDS_INPUT (x, 0);
2393 mode = GET_MODE (x);
2394 goto repeat;
2397 return hash;
2399 break;
2401 default:
2402 break;
2405 i = GET_RTX_LENGTH (code) - 1;
2406 hash += (unsigned) code + (unsigned) GET_MODE (x);
2407 fmt = GET_RTX_FORMAT (code);
2408 for (; i >= 0; i--)
2410 switch (fmt[i])
2412 case 'e':
2413 /* If we are about to do the last recursive call
2414 needed at this level, change it into iteration.
2415 This function is called enough to be worth it. */
2416 if (i == 0)
2418 x = XEXP (x, i);
2419 goto repeat;
2422 hash += hash_rtx (XEXP (x, i), 0, do_not_record_p,
2423 hash_arg_in_memory_p, have_reg_qty);
2424 break;
2426 case 'E':
2427 for (j = 0; j < XVECLEN (x, i); j++)
2428 hash += hash_rtx (XVECEXP (x, i, j), 0, do_not_record_p,
2429 hash_arg_in_memory_p, have_reg_qty);
2430 break;
2432 case 's':
2433 hash += hash_rtx_string (XSTR (x, i));
2434 break;
2436 case 'i':
2437 hash += (unsigned int) XINT (x, i);
2438 break;
2440 case '0': case 't':
2441 /* Unused. */
2442 break;
2444 default:
2445 gcc_unreachable ();
2449 return hash;
2452 /* Hash an rtx X for cse via hash_rtx.
2453 Stores 1 in do_not_record if any subexpression is volatile.
2454 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2455 does not have the RTX_UNCHANGING_P bit set. */
2457 static inline unsigned
2458 canon_hash (rtx x, enum machine_mode mode)
2460 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2463 /* Like canon_hash but with no side effects, i.e. do_not_record
2464 and hash_arg_in_memory are not changed. */
2466 static inline unsigned
2467 safe_hash (rtx x, enum machine_mode mode)
2469 int dummy_do_not_record;
2470 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2473 /* Return 1 iff X and Y would canonicalize into the same thing,
2474 without actually constructing the canonicalization of either one.
2475 If VALIDATE is nonzero,
2476 we assume X is an expression being processed from the rtl
2477 and Y was found in the hash table. We check register refs
2478 in Y for being marked as valid.
2480 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2483 exp_equiv_p (rtx x, rtx y, int validate, bool for_gcse)
2485 int i, j;
2486 enum rtx_code code;
2487 const char *fmt;
2489 /* Note: it is incorrect to assume an expression is equivalent to itself
2490 if VALIDATE is nonzero. */
2491 if (x == y && !validate)
2492 return 1;
2494 if (x == 0 || y == 0)
2495 return x == y;
2497 code = GET_CODE (x);
2498 if (code != GET_CODE (y))
2499 return 0;
2501 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2502 if (GET_MODE (x) != GET_MODE (y))
2503 return 0;
2505 switch (code)
2507 case PC:
2508 case CC0:
2509 case CONST_INT:
2510 case CONST_DOUBLE:
2511 return x == y;
2513 case LABEL_REF:
2514 return XEXP (x, 0) == XEXP (y, 0);
2516 case SYMBOL_REF:
2517 return XSTR (x, 0) == XSTR (y, 0);
2519 case REG:
2520 if (for_gcse)
2521 return REGNO (x) == REGNO (y);
2522 else
2524 unsigned int regno = REGNO (y);
2525 unsigned int i;
2526 unsigned int endregno
2527 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
2528 : hard_regno_nregs[regno][GET_MODE (y)]);
2530 /* If the quantities are not the same, the expressions are not
2531 equivalent. If there are and we are not to validate, they
2532 are equivalent. Otherwise, ensure all regs are up-to-date. */
2534 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2535 return 0;
2537 if (! validate)
2538 return 1;
2540 for (i = regno; i < endregno; i++)
2541 if (REG_IN_TABLE (i) != REG_TICK (i))
2542 return 0;
2544 return 1;
2547 case MEM:
2548 if (for_gcse)
2550 /* A volatile mem should not be considered equivalent to any
2551 other. */
2552 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2553 return 0;
2555 /* Can't merge two expressions in different alias sets, since we
2556 can decide that the expression is transparent in a block when
2557 it isn't, due to it being set with the different alias set.
2559 Also, can't merge two expressions with different MEM_ATTRS.
2560 They could e.g. be two different entities allocated into the
2561 same space on the stack (see e.g. PR25130). In that case, the
2562 MEM addresses can be the same, even though the two MEMs are
2563 absolutely not equivalent.
2565 But because really all MEM attributes should be the same for
2566 equivalent MEMs, we just use the invariant that MEMs that have
2567 the same attributes share the same mem_attrs data structure. */
2568 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2569 return 0;
2571 break;
2573 /* For commutative operations, check both orders. */
2574 case PLUS:
2575 case MULT:
2576 case AND:
2577 case IOR:
2578 case XOR:
2579 case NE:
2580 case EQ:
2581 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2582 validate, for_gcse)
2583 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2584 validate, for_gcse))
2585 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2586 validate, for_gcse)
2587 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2588 validate, for_gcse)));
2590 case ASM_OPERANDS:
2591 /* We don't use the generic code below because we want to
2592 disregard filename and line numbers. */
2594 /* A volatile asm isn't equivalent to any other. */
2595 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2596 return 0;
2598 if (GET_MODE (x) != GET_MODE (y)
2599 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2600 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2601 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2602 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2603 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2604 return 0;
2606 if (ASM_OPERANDS_INPUT_LENGTH (x))
2608 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2609 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2610 ASM_OPERANDS_INPUT (y, i),
2611 validate, for_gcse)
2612 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2613 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2614 return 0;
2617 return 1;
2619 default:
2620 break;
2623 /* Compare the elements. If any pair of corresponding elements
2624 fail to match, return 0 for the whole thing. */
2626 fmt = GET_RTX_FORMAT (code);
2627 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2629 switch (fmt[i])
2631 case 'e':
2632 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2633 validate, for_gcse))
2634 return 0;
2635 break;
2637 case 'E':
2638 if (XVECLEN (x, i) != XVECLEN (y, i))
2639 return 0;
2640 for (j = 0; j < XVECLEN (x, i); j++)
2641 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2642 validate, for_gcse))
2643 return 0;
2644 break;
2646 case 's':
2647 if (strcmp (XSTR (x, i), XSTR (y, i)))
2648 return 0;
2649 break;
2651 case 'i':
2652 if (XINT (x, i) != XINT (y, i))
2653 return 0;
2654 break;
2656 case 'w':
2657 if (XWINT (x, i) != XWINT (y, i))
2658 return 0;
2659 break;
2661 case '0':
2662 case 't':
2663 break;
2665 default:
2666 gcc_unreachable ();
2670 return 1;
2673 /* Return 1 if X has a value that can vary even between two
2674 executions of the program. 0 means X can be compared reliably
2675 against certain constants or near-constants. */
2677 static int
2678 cse_rtx_varies_p (rtx x, int from_alias)
2680 /* We need not check for X and the equivalence class being of the same
2681 mode because if X is equivalent to a constant in some mode, it
2682 doesn't vary in any mode. */
2684 if (REG_P (x)
2685 && REGNO_QTY_VALID_P (REGNO (x)))
2687 int x_q = REG_QTY (REGNO (x));
2688 struct qty_table_elem *x_ent = &qty_table[x_q];
2690 if (GET_MODE (x) == x_ent->mode
2691 && x_ent->const_rtx != NULL_RTX)
2692 return 0;
2695 if (GET_CODE (x) == PLUS
2696 && GET_CODE (XEXP (x, 1)) == CONST_INT
2697 && REG_P (XEXP (x, 0))
2698 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2700 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2701 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2703 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2704 && x0_ent->const_rtx != NULL_RTX)
2705 return 0;
2708 /* This can happen as the result of virtual register instantiation, if
2709 the initial constant is too large to be a valid address. This gives
2710 us a three instruction sequence, load large offset into a register,
2711 load fp minus a constant into a register, then a MEM which is the
2712 sum of the two `constant' registers. */
2713 if (GET_CODE (x) == PLUS
2714 && REG_P (XEXP (x, 0))
2715 && REG_P (XEXP (x, 1))
2716 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2717 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2719 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2720 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2721 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2722 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2724 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2725 && x0_ent->const_rtx != NULL_RTX
2726 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2727 && x1_ent->const_rtx != NULL_RTX)
2728 return 0;
2731 return rtx_varies_p (x, from_alias);
2734 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2735 the result if necessary. INSN is as for canon_reg. */
2737 static void
2738 validate_canon_reg (rtx *xloc, rtx insn)
2740 rtx new = canon_reg (*xloc, insn);
2742 /* If replacing pseudo with hard reg or vice versa, ensure the
2743 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2744 if (insn != 0 && new != 0)
2745 validate_change (insn, xloc, new, 1);
2746 else
2747 *xloc = new;
2750 /* Canonicalize an expression:
2751 replace each register reference inside it
2752 with the "oldest" equivalent register.
2754 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2755 after we make our substitution. The calls are made with IN_GROUP nonzero
2756 so apply_change_group must be called upon the outermost return from this
2757 function (unless INSN is zero). The result of apply_change_group can
2758 generally be discarded since the changes we are making are optional. */
2760 static rtx
2761 canon_reg (rtx x, rtx insn)
2763 int i;
2764 enum rtx_code code;
2765 const char *fmt;
2767 if (x == 0)
2768 return x;
2770 code = GET_CODE (x);
2771 switch (code)
2773 case PC:
2774 case CC0:
2775 case CONST:
2776 case CONST_INT:
2777 case CONST_DOUBLE:
2778 case CONST_VECTOR:
2779 case SYMBOL_REF:
2780 case LABEL_REF:
2781 case ADDR_VEC:
2782 case ADDR_DIFF_VEC:
2783 return x;
2785 case REG:
2787 int first;
2788 int q;
2789 struct qty_table_elem *ent;
2791 /* Never replace a hard reg, because hard regs can appear
2792 in more than one machine mode, and we must preserve the mode
2793 of each occurrence. Also, some hard regs appear in
2794 MEMs that are shared and mustn't be altered. Don't try to
2795 replace any reg that maps to a reg of class NO_REGS. */
2796 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2797 || ! REGNO_QTY_VALID_P (REGNO (x)))
2798 return x;
2800 q = REG_QTY (REGNO (x));
2801 ent = &qty_table[q];
2802 first = ent->first_reg;
2803 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2804 : REGNO_REG_CLASS (first) == NO_REGS ? x
2805 : gen_rtx_REG (ent->mode, first));
2808 default:
2809 break;
2812 fmt = GET_RTX_FORMAT (code);
2813 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2815 int j;
2817 if (fmt[i] == 'e')
2818 validate_canon_reg (&XEXP (x, i), insn);
2819 else if (fmt[i] == 'E')
2820 for (j = 0; j < XVECLEN (x, i); j++)
2821 validate_canon_reg (&XVECEXP (x, i, j), insn);
2824 return x;
2827 /* LOC is a location within INSN that is an operand address (the contents of
2828 a MEM). Find the best equivalent address to use that is valid for this
2829 insn.
2831 On most CISC machines, complicated address modes are costly, and rtx_cost
2832 is a good approximation for that cost. However, most RISC machines have
2833 only a few (usually only one) memory reference formats. If an address is
2834 valid at all, it is often just as cheap as any other address. Hence, for
2835 RISC machines, we use `address_cost' to compare the costs of various
2836 addresses. For two addresses of equal cost, choose the one with the
2837 highest `rtx_cost' value as that has the potential of eliminating the
2838 most insns. For equal costs, we choose the first in the equivalence
2839 class. Note that we ignore the fact that pseudo registers are cheaper than
2840 hard registers here because we would also prefer the pseudo registers. */
2842 static void
2843 find_best_addr (rtx insn, rtx *loc, enum machine_mode mode)
2845 struct table_elt *elt;
2846 rtx addr = *loc;
2847 struct table_elt *p;
2848 int found_better = 1;
2849 int save_do_not_record = do_not_record;
2850 int save_hash_arg_in_memory = hash_arg_in_memory;
2851 int addr_volatile;
2852 int regno;
2853 unsigned hash;
2855 /* Do not try to replace constant addresses or addresses of local and
2856 argument slots. These MEM expressions are made only once and inserted
2857 in many instructions, as well as being used to control symbol table
2858 output. It is not safe to clobber them.
2860 There are some uncommon cases where the address is already in a register
2861 for some reason, but we cannot take advantage of that because we have
2862 no easy way to unshare the MEM. In addition, looking up all stack
2863 addresses is costly. */
2864 if ((GET_CODE (addr) == PLUS
2865 && REG_P (XEXP (addr, 0))
2866 && GET_CODE (XEXP (addr, 1)) == CONST_INT
2867 && (regno = REGNO (XEXP (addr, 0)),
2868 regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM
2869 || regno == ARG_POINTER_REGNUM))
2870 || (REG_P (addr)
2871 && (regno = REGNO (addr), regno == FRAME_POINTER_REGNUM
2872 || regno == HARD_FRAME_POINTER_REGNUM
2873 || regno == ARG_POINTER_REGNUM))
2874 || CONSTANT_ADDRESS_P (addr))
2875 return;
2877 /* If this address is not simply a register, try to fold it. This will
2878 sometimes simplify the expression. Many simplifications
2879 will not be valid, but some, usually applying the associative rule, will
2880 be valid and produce better code. */
2881 if (!REG_P (addr))
2883 rtx folded = canon_for_address (fold_rtx (addr, NULL_RTX));
2885 if (folded != addr)
2887 int addr_folded_cost = address_cost (folded, mode);
2888 int addr_cost = address_cost (addr, mode);
2890 if ((addr_folded_cost < addr_cost
2891 || (addr_folded_cost == addr_cost
2892 /* ??? The rtx_cost comparison is left over from an older
2893 version of this code. It is probably no longer helpful.*/
2894 && (rtx_cost (folded, MEM) > rtx_cost (addr, MEM)
2895 || approx_reg_cost (folded) < approx_reg_cost (addr))))
2896 && validate_change (insn, loc, folded, 0))
2897 addr = folded;
2901 /* If this address is not in the hash table, we can't look for equivalences
2902 of the whole address. Also, ignore if volatile. */
2904 do_not_record = 0;
2905 hash = HASH (addr, Pmode);
2906 addr_volatile = do_not_record;
2907 do_not_record = save_do_not_record;
2908 hash_arg_in_memory = save_hash_arg_in_memory;
2910 if (addr_volatile)
2911 return;
2913 elt = lookup (addr, hash, Pmode);
2915 if (elt)
2917 /* We need to find the best (under the criteria documented above) entry
2918 in the class that is valid. We use the `flag' field to indicate
2919 choices that were invalid and iterate until we can't find a better
2920 one that hasn't already been tried. */
2922 for (p = elt->first_same_value; p; p = p->next_same_value)
2923 p->flag = 0;
2925 while (found_better)
2927 int best_addr_cost = address_cost (*loc, mode);
2928 int best_rtx_cost = (elt->cost + 1) >> 1;
2929 int exp_cost;
2930 struct table_elt *best_elt = elt;
2932 found_better = 0;
2933 for (p = elt->first_same_value; p; p = p->next_same_value)
2934 if (! p->flag)
2936 if ((REG_P (p->exp)
2937 || exp_equiv_p (p->exp, p->exp, 1, false))
2938 && ((exp_cost = address_cost (p->exp, mode)) < best_addr_cost
2939 || (exp_cost == best_addr_cost
2940 && ((p->cost + 1) >> 1) > best_rtx_cost)))
2942 found_better = 1;
2943 best_addr_cost = exp_cost;
2944 best_rtx_cost = (p->cost + 1) >> 1;
2945 best_elt = p;
2949 if (found_better)
2951 if (validate_change (insn, loc,
2952 canon_reg (copy_rtx (best_elt->exp),
2953 NULL_RTX), 0))
2954 return;
2955 else
2956 best_elt->flag = 1;
2961 /* If the address is a binary operation with the first operand a register
2962 and the second a constant, do the same as above, but looking for
2963 equivalences of the register. Then try to simplify before checking for
2964 the best address to use. This catches a few cases: First is when we
2965 have REG+const and the register is another REG+const. We can often merge
2966 the constants and eliminate one insn and one register. It may also be
2967 that a machine has a cheap REG+REG+const. Finally, this improves the
2968 code on the Alpha for unaligned byte stores. */
2970 if (flag_expensive_optimizations
2971 && ARITHMETIC_P (*loc)
2972 && REG_P (XEXP (*loc, 0)))
2974 rtx op1 = XEXP (*loc, 1);
2976 do_not_record = 0;
2977 hash = HASH (XEXP (*loc, 0), Pmode);
2978 do_not_record = save_do_not_record;
2979 hash_arg_in_memory = save_hash_arg_in_memory;
2981 elt = lookup (XEXP (*loc, 0), hash, Pmode);
2982 if (elt == 0)
2983 return;
2985 /* We need to find the best (under the criteria documented above) entry
2986 in the class that is valid. We use the `flag' field to indicate
2987 choices that were invalid and iterate until we can't find a better
2988 one that hasn't already been tried. */
2990 for (p = elt->first_same_value; p; p = p->next_same_value)
2991 p->flag = 0;
2993 while (found_better)
2995 int best_addr_cost = address_cost (*loc, mode);
2996 int best_rtx_cost = (COST (*loc) + 1) >> 1;
2997 struct table_elt *best_elt = elt;
2998 rtx best_rtx = *loc;
2999 int count;
3001 /* This is at worst case an O(n^2) algorithm, so limit our search
3002 to the first 32 elements on the list. This avoids trouble
3003 compiling code with very long basic blocks that can easily
3004 call simplify_gen_binary so many times that we run out of
3005 memory. */
3007 found_better = 0;
3008 for (p = elt->first_same_value, count = 0;
3009 p && count < 32;
3010 p = p->next_same_value, count++)
3011 if (! p->flag
3012 && (REG_P (p->exp)
3013 || (GET_CODE (p->exp) != EXPR_LIST
3014 && exp_equiv_p (p->exp, p->exp, 1, false))))
3017 rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
3018 p->exp, op1);
3019 int new_cost;
3021 /* Get the canonical version of the address so we can accept
3022 more. */
3023 new = canon_for_address (new);
3025 new_cost = address_cost (new, mode);
3027 if (new_cost < best_addr_cost
3028 || (new_cost == best_addr_cost
3029 && (COST (new) + 1) >> 1 > best_rtx_cost))
3031 found_better = 1;
3032 best_addr_cost = new_cost;
3033 best_rtx_cost = (COST (new) + 1) >> 1;
3034 best_elt = p;
3035 best_rtx = new;
3039 if (found_better)
3041 if (validate_change (insn, loc,
3042 canon_reg (copy_rtx (best_rtx),
3043 NULL_RTX), 0))
3044 return;
3045 else
3046 best_elt->flag = 1;
3052 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
3053 operation (EQ, NE, GT, etc.), follow it back through the hash table and
3054 what values are being compared.
3056 *PARG1 and *PARG2 are updated to contain the rtx representing the values
3057 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
3058 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
3059 compared to produce cc0.
3061 The return value is the comparison operator and is either the code of
3062 A or the code corresponding to the inverse of the comparison. */
3064 static enum rtx_code
3065 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
3066 enum machine_mode *pmode1, enum machine_mode *pmode2)
3068 rtx arg1, arg2;
3070 arg1 = *parg1, arg2 = *parg2;
3072 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
3074 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
3076 /* Set nonzero when we find something of interest. */
3077 rtx x = 0;
3078 int reverse_code = 0;
3079 struct table_elt *p = 0;
3081 /* If arg1 is a COMPARE, extract the comparison arguments from it.
3082 On machines with CC0, this is the only case that can occur, since
3083 fold_rtx will return the COMPARE or item being compared with zero
3084 when given CC0. */
3086 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
3087 x = arg1;
3089 /* If ARG1 is a comparison operator and CODE is testing for
3090 STORE_FLAG_VALUE, get the inner arguments. */
3092 else if (COMPARISON_P (arg1))
3094 #ifdef FLOAT_STORE_FLAG_VALUE
3095 REAL_VALUE_TYPE fsfv;
3096 #endif
3098 if (code == NE
3099 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3100 && code == LT && 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;
3108 else if (code == EQ
3109 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3110 && code == GE && STORE_FLAG_VALUE == -1)
3111 #ifdef FLOAT_STORE_FLAG_VALUE
3112 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3113 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3114 REAL_VALUE_NEGATIVE (fsfv)))
3115 #endif
3117 x = arg1, reverse_code = 1;
3120 /* ??? We could also check for
3122 (ne (and (eq (...) (const_int 1))) (const_int 0))
3124 and related forms, but let's wait until we see them occurring. */
3126 if (x == 0)
3127 /* Look up ARG1 in the hash table and see if it has an equivalence
3128 that lets us see what is being compared. */
3129 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3130 if (p)
3132 p = p->first_same_value;
3134 /* If what we compare is already known to be constant, that is as
3135 good as it gets.
3136 We need to break the loop in this case, because otherwise we
3137 can have an infinite loop when looking at a reg that is known
3138 to be a constant which is the same as a comparison of a reg
3139 against zero which appears later in the insn stream, which in
3140 turn is constant and the same as the comparison of the first reg
3141 against zero... */
3142 if (p->is_const)
3143 break;
3146 for (; p; p = p->next_same_value)
3148 enum machine_mode inner_mode = GET_MODE (p->exp);
3149 #ifdef FLOAT_STORE_FLAG_VALUE
3150 REAL_VALUE_TYPE fsfv;
3151 #endif
3153 /* If the entry isn't valid, skip it. */
3154 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3155 continue;
3157 if (GET_CODE (p->exp) == COMPARE
3158 /* Another possibility is that this machine has a compare insn
3159 that includes the comparison code. In that case, ARG1 would
3160 be equivalent to a comparison operation that would set ARG1 to
3161 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3162 ORIG_CODE is the actual comparison being done; if it is an EQ,
3163 we must reverse ORIG_CODE. On machine with a negative value
3164 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3165 || ((code == NE
3166 || (code == LT
3167 && GET_MODE_CLASS (inner_mode) == MODE_INT
3168 && (GET_MODE_BITSIZE (inner_mode)
3169 <= HOST_BITS_PER_WIDE_INT)
3170 && (STORE_FLAG_VALUE
3171 & ((HOST_WIDE_INT) 1
3172 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3173 #ifdef FLOAT_STORE_FLAG_VALUE
3174 || (code == LT
3175 && SCALAR_FLOAT_MODE_P (inner_mode)
3176 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3177 REAL_VALUE_NEGATIVE (fsfv)))
3178 #endif
3180 && COMPARISON_P (p->exp)))
3182 x = p->exp;
3183 break;
3185 else if ((code == EQ
3186 || (code == GE
3187 && GET_MODE_CLASS (inner_mode) == MODE_INT
3188 && (GET_MODE_BITSIZE (inner_mode)
3189 <= HOST_BITS_PER_WIDE_INT)
3190 && (STORE_FLAG_VALUE
3191 & ((HOST_WIDE_INT) 1
3192 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3193 #ifdef FLOAT_STORE_FLAG_VALUE
3194 || (code == GE
3195 && SCALAR_FLOAT_MODE_P (inner_mode)
3196 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3197 REAL_VALUE_NEGATIVE (fsfv)))
3198 #endif
3200 && COMPARISON_P (p->exp))
3202 reverse_code = 1;
3203 x = p->exp;
3204 break;
3207 /* If this non-trapping address, e.g. fp + constant, the
3208 equivalent is a better operand since it may let us predict
3209 the value of the comparison. */
3210 else if (!rtx_addr_can_trap_p (p->exp))
3212 arg1 = p->exp;
3213 continue;
3217 /* If we didn't find a useful equivalence for ARG1, we are done.
3218 Otherwise, set up for the next iteration. */
3219 if (x == 0)
3220 break;
3222 /* If we need to reverse the comparison, make sure that that is
3223 possible -- we can't necessarily infer the value of GE from LT
3224 with floating-point operands. */
3225 if (reverse_code)
3227 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3228 if (reversed == UNKNOWN)
3229 break;
3230 else
3231 code = reversed;
3233 else if (COMPARISON_P (x))
3234 code = GET_CODE (x);
3235 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3238 /* Return our results. Return the modes from before fold_rtx
3239 because fold_rtx might produce const_int, and then it's too late. */
3240 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3241 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3243 return code;
3246 /* Fold SUBREG. */
3248 static rtx
3249 fold_rtx_subreg (rtx x, rtx insn)
3251 enum machine_mode mode = GET_MODE (x);
3252 rtx folded_arg0;
3253 rtx const_arg0;
3254 rtx new;
3256 /* See if we previously assigned a constant value to this SUBREG. */
3257 if ((new = lookup_as_function (x, CONST_INT)) != 0
3258 || (new = lookup_as_function (x, CONST_DOUBLE)) != 0)
3259 return new;
3261 /* If this is a paradoxical SUBREG, we have no idea what value the
3262 extra bits would have. However, if the operand is equivalent to
3263 a SUBREG whose operand is the same as our mode, and all the modes
3264 are within a word, we can just use the inner operand because
3265 these SUBREGs just say how to treat the register.
3267 Similarly if we find an integer constant. */
3269 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
3271 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3272 struct table_elt *elt;
3274 if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
3275 && GET_MODE_SIZE (imode) <= UNITS_PER_WORD
3276 && (elt = lookup (SUBREG_REG (x), HASH (SUBREG_REG (x), imode),
3277 imode)) != 0)
3278 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3280 if (CONSTANT_P (elt->exp)
3281 && GET_MODE (elt->exp) == VOIDmode)
3282 return elt->exp;
3284 if (GET_CODE (elt->exp) == SUBREG
3285 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3286 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3287 return copy_rtx (SUBREG_REG (elt->exp));
3290 return x;
3293 /* Fold SUBREG_REG. If it changed, see if we can simplify the
3294 SUBREG. We might be able to if the SUBREG is extracting a single
3295 word in an integral mode or extracting the low part. */
3297 folded_arg0 = fold_rtx (SUBREG_REG (x), insn);
3298 const_arg0 = equiv_constant (folded_arg0);
3299 if (const_arg0)
3300 folded_arg0 = const_arg0;
3302 if (folded_arg0 != SUBREG_REG (x))
3304 new = simplify_subreg (mode, folded_arg0,
3305 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
3306 if (new)
3307 return new;
3310 if (REG_P (folded_arg0)
3311 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (folded_arg0)))
3313 struct table_elt *elt;
3315 elt = lookup (folded_arg0,
3316 HASH (folded_arg0, GET_MODE (folded_arg0)),
3317 GET_MODE (folded_arg0));
3319 if (elt)
3320 elt = elt->first_same_value;
3322 if (subreg_lowpart_p (x))
3323 /* If this is a narrowing SUBREG and our operand is a REG, see
3324 if we can find an equivalence for REG that is an arithmetic
3325 operation in a wider mode where both operands are
3326 paradoxical SUBREGs from objects of our result mode. In
3327 that case, we couldn-t report an equivalent value for that
3328 operation, since we don't know what the extra bits will be.
3329 But we can find an equivalence for this SUBREG by folding
3330 that operation in the narrow mode. This allows us to fold
3331 arithmetic in narrow modes when the machine only supports
3332 word-sized arithmetic.
3334 Also look for a case where we have a SUBREG whose operand
3335 is the same as our result. If both modes are smaller than
3336 a word, we are simply interpreting a register in different
3337 modes and we can use the inner value. */
3339 for (; elt; elt = elt->next_same_value)
3341 enum rtx_code eltcode = GET_CODE (elt->exp);
3343 /* Just check for unary and binary operations. */
3344 if (UNARY_P (elt->exp)
3345 && eltcode != SIGN_EXTEND
3346 && eltcode != ZERO_EXTEND
3347 && GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3348 && GET_MODE (SUBREG_REG (XEXP (elt->exp, 0))) == mode
3349 && (GET_MODE_CLASS (mode)
3350 == GET_MODE_CLASS (GET_MODE (XEXP (elt->exp, 0)))))
3352 rtx op0 = SUBREG_REG (XEXP (elt->exp, 0));
3354 if (!REG_P (op0) && ! CONSTANT_P (op0))
3355 op0 = fold_rtx (op0, NULL_RTX);
3357 op0 = equiv_constant (op0);
3358 if (op0)
3359 new = simplify_unary_operation (GET_CODE (elt->exp), mode,
3360 op0, mode);
3362 else if (ARITHMETIC_P (elt->exp)
3363 && eltcode != DIV && eltcode != MOD
3364 && eltcode != UDIV && eltcode != UMOD
3365 && eltcode != ASHIFTRT && eltcode != LSHIFTRT
3366 && eltcode != ROTATE && eltcode != ROTATERT
3367 && ((GET_CODE (XEXP (elt->exp, 0)) == SUBREG
3368 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 0)))
3369 == mode))
3370 || CONSTANT_P (XEXP (elt->exp, 0)))
3371 && ((GET_CODE (XEXP (elt->exp, 1)) == SUBREG
3372 && (GET_MODE (SUBREG_REG (XEXP (elt->exp, 1)))
3373 == mode))
3374 || CONSTANT_P (XEXP (elt->exp, 1))))
3376 rtx op0 = gen_lowpart_common (mode, XEXP (elt->exp, 0));
3377 rtx op1 = gen_lowpart_common (mode, XEXP (elt->exp, 1));
3379 if (op0 && !REG_P (op0) && ! CONSTANT_P (op0))
3380 op0 = fold_rtx (op0, NULL_RTX);
3382 if (op0)
3383 op0 = equiv_constant (op0);
3385 if (op1 && !REG_P (op1) && ! CONSTANT_P (op1))
3386 op1 = fold_rtx (op1, NULL_RTX);
3388 if (op1)
3389 op1 = equiv_constant (op1);
3391 /* If we are looking for the low SImode part of
3392 (ashift:DI c (const_int 32)), it doesn't work to
3393 compute that in SImode, because a 32-bit shift in
3394 SImode is unpredictable. We know the value is
3395 0. */
3396 if (op0 && op1
3397 && GET_CODE (elt->exp) == ASHIFT
3398 && GET_CODE (op1) == CONST_INT
3399 && INTVAL (op1) >= GET_MODE_BITSIZE (mode))
3401 if (INTVAL (op1)
3402 < GET_MODE_BITSIZE (GET_MODE (elt->exp)))
3403 /* If the count fits in the inner mode's width,
3404 but exceeds the outer mode's width, the value
3405 will get truncated to 0 by the subreg. */
3406 new = CONST0_RTX (mode);
3407 else
3408 /* If the count exceeds even the inner mode's width,
3409 don't fold this expression. */
3410 new = 0;
3412 else if (op0 && op1)
3413 new = simplify_binary_operation (GET_CODE (elt->exp),
3414 mode, op0, op1);
3417 else if (GET_CODE (elt->exp) == SUBREG
3418 && GET_MODE (SUBREG_REG (elt->exp)) == mode
3419 && (GET_MODE_SIZE (GET_MODE (folded_arg0))
3420 <= UNITS_PER_WORD)
3421 && exp_equiv_p (elt->exp, elt->exp, 1, false))
3422 new = copy_rtx (SUBREG_REG (elt->exp));
3424 if (new)
3425 return new;
3427 else
3428 /* A SUBREG resulting from a zero extension may fold to zero
3429 if it extracts higher bits than the ZERO_EXTEND's source
3430 bits. FIXME: if combine tried to, er, combine these
3431 instructions, this transformation may be moved to
3432 simplify_subreg. */
3433 for (; elt; elt = elt->next_same_value)
3435 if (GET_CODE (elt->exp) == ZERO_EXTEND
3436 && subreg_lsb (x)
3437 >= GET_MODE_BITSIZE (GET_MODE (XEXP (elt->exp, 0))))
3438 return CONST0_RTX (mode);
3442 return x;
3445 /* Fold MEM. Not to be called directly, see fold_rtx_mem instead. */
3447 static rtx
3448 fold_rtx_mem_1 (rtx x, rtx insn)
3450 enum machine_mode mode = GET_MODE (x);
3451 rtx new;
3453 /* If we are not actually processing an insn, don't try to find the
3454 best address. Not only don't we care, but we could modify the
3455 MEM in an invalid way since we have no insn to validate
3456 against. */
3457 if (insn != 0)
3458 find_best_addr (insn, &XEXP (x, 0), mode);
3461 /* Even if we don't fold in the insn itself, we can safely do so
3462 here, in hopes of getting a constant. */
3463 rtx addr = fold_rtx (XEXP (x, 0), NULL_RTX);
3464 rtx base = 0;
3465 HOST_WIDE_INT offset = 0;
3467 if (REG_P (addr)
3468 && REGNO_QTY_VALID_P (REGNO (addr)))
3470 int addr_q = REG_QTY (REGNO (addr));
3471 struct qty_table_elem *addr_ent = &qty_table[addr_q];
3473 if (GET_MODE (addr) == addr_ent->mode
3474 && addr_ent->const_rtx != NULL_RTX)
3475 addr = addr_ent->const_rtx;
3478 /* Call target hook to avoid the effects of -fpic etc.... */
3479 addr = targetm.delegitimize_address (addr);
3481 /* If address is constant, split it into a base and integer
3482 offset. */
3483 if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
3484 base = addr;
3485 else if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
3486 && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT)
3488 base = XEXP (XEXP (addr, 0), 0);
3489 offset = INTVAL (XEXP (XEXP (addr, 0), 1));
3491 else if (GET_CODE (addr) == LO_SUM
3492 && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
3493 base = XEXP (addr, 1);
3495 /* If this is a constant pool reference, we can fold it into its
3496 constant to allow better value tracking. */
3497 if (base && GET_CODE (base) == SYMBOL_REF
3498 && CONSTANT_POOL_ADDRESS_P (base))
3500 rtx constant = get_pool_constant (base);
3501 enum machine_mode const_mode = get_pool_mode (base);
3502 rtx new;
3504 if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
3506 constant_pool_entries_cost = COST (constant);
3507 constant_pool_entries_regcost = approx_reg_cost (constant);
3510 /* If we are loading the full constant, we have an
3511 equivalence. */
3512 if (offset == 0 && mode == const_mode)
3513 return constant;
3515 /* If this actually isn't a constant (weird!), we can't do
3516 anything. Otherwise, handle the two most common cases:
3517 extracting a word from a multi-word constant, and
3518 extracting the low-order bits. Other cases don't seem
3519 common enough to worry about. */
3520 if (! CONSTANT_P (constant))
3521 return x;
3523 if (GET_MODE_CLASS (mode) == MODE_INT
3524 && GET_MODE_SIZE (mode) == UNITS_PER_WORD
3525 && offset % UNITS_PER_WORD == 0
3526 && (new = operand_subword (constant,
3527 offset / UNITS_PER_WORD,
3528 0, const_mode)) != 0)
3529 return new;
3531 if (((BYTES_BIG_ENDIAN
3532 && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
3533 || (! BYTES_BIG_ENDIAN && offset == 0))
3534 && (new = gen_lowpart (mode, constant)) != 0)
3535 return new;
3538 /* If this is a reference to a label at a known position in a jump
3539 table, we also know its value. */
3540 if (base && GET_CODE (base) == LABEL_REF)
3542 rtx label = XEXP (base, 0);
3543 rtx table_insn = NEXT_INSN (label);
3545 if (table_insn && JUMP_P (table_insn)
3546 && GET_CODE (PATTERN (table_insn)) == ADDR_VEC)
3548 rtx table = PATTERN (table_insn);
3550 if (offset >= 0
3551 && (offset / GET_MODE_SIZE (GET_MODE (table))
3552 < XVECLEN (table, 0)))
3554 rtx label = XVECEXP
3555 (table, 0, offset / GET_MODE_SIZE (GET_MODE (table)));
3556 rtx set;
3558 /* If we have an insn that loads the label from the
3559 jumptable into a reg, we don't want to set the reg
3560 to the label, because this may cause a reference to
3561 the label to remain after the label is removed in
3562 some very obscure cases (PR middle-end/18628). */
3563 if (!insn)
3564 return label;
3566 set = single_set (insn);
3568 if (! set || SET_SRC (set) != x)
3569 return x;
3571 /* If it's a jump, it's safe to reference the label. */
3572 if (SET_DEST (set) == pc_rtx)
3573 return label;
3575 return x;
3578 if (table_insn && JUMP_P (table_insn)
3579 && GET_CODE (PATTERN (table_insn)) == ADDR_DIFF_VEC)
3581 rtx table = PATTERN (table_insn);
3583 if (offset >= 0
3584 && (offset / GET_MODE_SIZE (GET_MODE (table))
3585 < XVECLEN (table, 1)))
3587 offset /= GET_MODE_SIZE (GET_MODE (table));
3588 new = gen_rtx_MINUS (Pmode, XVECEXP (table, 1, offset),
3589 XEXP (table, 0));
3591 if (GET_MODE (table) != Pmode)
3592 new = gen_rtx_TRUNCATE (GET_MODE (table), new);
3594 /* Indicate this is a constant. This isn't a valid
3595 form of CONST, but it will only be used to fold the
3596 next insns and then discarded, so it should be
3597 safe.
3599 Note this expression must be explicitly discarded,
3600 by cse_insn, else it may end up in a REG_EQUAL note
3601 and "escape" to cause problems elsewhere. */
3602 return gen_rtx_CONST (GET_MODE (new), new);
3607 return x;
3611 /* Fold MEM. */
3613 static rtx
3614 fold_rtx_mem (rtx x, rtx insn)
3616 /* To avoid infinite oscillations between fold_rtx and fold_rtx_mem,
3617 refuse to allow recursion of the latter past n levels. This can
3618 happen because fold_rtx_mem will try to fold the address of the
3619 memory reference it is passed, i.e. conceptually throwing away
3620 the MEM and reinjecting the bare address into fold_rtx. As a
3621 result, patterns like
3623 set (reg1)
3624 (plus (reg)
3625 (mem (plus (reg2) (const_int))))
3627 set (reg2)
3628 (plus (reg)
3629 (mem (plus (reg1) (const_int))))
3631 will defeat any "first-order" short-circuit put in either
3632 function to prevent these infinite oscillations.
3634 The heuristics for determining n is as follows: since each time
3635 it is invoked fold_rtx_mem throws away a MEM, and since MEMs
3636 are generically not nested, we assume that each invocation of
3637 fold_rtx_mem corresponds to a new "top-level" operand, i.e.
3638 the source or the destination of a SET. So fold_rtx_mem is
3639 bound to stop or cycle before n recursions, n being the number
3640 of expressions recorded in the hash table. We also leave some
3641 play to account for the initial steps. */
3643 static unsigned int depth;
3644 rtx ret;
3646 if (depth > 3 + table_size)
3647 return x;
3649 depth++;
3650 ret = fold_rtx_mem_1 (x, insn);
3651 depth--;
3653 return ret;
3656 /* If X is a nontrivial arithmetic operation on an argument
3657 for which a constant value can be determined, return
3658 the result of operating on that value, as a constant.
3659 Otherwise, return X, possibly with one or more operands
3660 modified by recursive calls to this function.
3662 If X is a register whose contents are known, we do NOT
3663 return those contents here. equiv_constant is called to
3664 perform that task.
3666 INSN is the insn that we may be modifying. If it is 0, make a copy
3667 of X before modifying it. */
3669 static rtx
3670 fold_rtx (rtx x, rtx insn)
3672 enum rtx_code code;
3673 enum machine_mode mode;
3674 const char *fmt;
3675 int i;
3676 rtx new = 0;
3677 int copied = 0;
3678 int must_swap = 0;
3680 /* Folded equivalents of first two operands of X. */
3681 rtx folded_arg0;
3682 rtx folded_arg1;
3684 /* Constant equivalents of first three operands of X;
3685 0 when no such equivalent is known. */
3686 rtx const_arg0;
3687 rtx const_arg1;
3688 rtx const_arg2;
3690 /* The mode of the first operand of X. We need this for sign and zero
3691 extends. */
3692 enum machine_mode mode_arg0;
3694 if (x == 0)
3695 return x;
3697 mode = GET_MODE (x);
3698 code = GET_CODE (x);
3699 switch (code)
3701 case CONST:
3702 case CONST_INT:
3703 case CONST_DOUBLE:
3704 case CONST_VECTOR:
3705 case SYMBOL_REF:
3706 case LABEL_REF:
3707 case REG:
3708 case PC:
3709 /* No use simplifying an EXPR_LIST
3710 since they are used only for lists of args
3711 in a function call's REG_EQUAL note. */
3712 case EXPR_LIST:
3713 return x;
3715 #ifdef HAVE_cc0
3716 case CC0:
3717 return prev_insn_cc0;
3718 #endif
3720 case SUBREG:
3721 return fold_rtx_subreg (x, insn);
3723 case NOT:
3724 case NEG:
3725 /* If we have (NOT Y), see if Y is known to be (NOT Z).
3726 If so, (NOT Y) simplifies to Z. Similarly for NEG. */
3727 new = lookup_as_function (XEXP (x, 0), code);
3728 if (new)
3729 return fold_rtx (copy_rtx (XEXP (new, 0)), insn);
3730 break;
3732 case MEM:
3733 return fold_rtx_mem (x, insn);
3735 #ifdef NO_FUNCTION_CSE
3736 case CALL:
3737 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3738 return x;
3739 break;
3740 #endif
3742 case ASM_OPERANDS:
3743 if (insn)
3745 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3746 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3747 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3749 break;
3751 default:
3752 break;
3755 const_arg0 = 0;
3756 const_arg1 = 0;
3757 const_arg2 = 0;
3758 mode_arg0 = VOIDmode;
3760 /* Try folding our operands.
3761 Then see which ones have constant values known. */
3763 fmt = GET_RTX_FORMAT (code);
3764 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3765 if (fmt[i] == 'e')
3767 rtx arg = XEXP (x, i);
3768 rtx folded_arg = arg, const_arg = 0;
3769 enum machine_mode mode_arg = GET_MODE (arg);
3770 rtx cheap_arg, expensive_arg;
3771 rtx replacements[2];
3772 int j;
3773 int old_cost = COST_IN (XEXP (x, i), code);
3775 /* Most arguments are cheap, so handle them specially. */
3776 switch (GET_CODE (arg))
3778 case REG:
3779 /* This is the same as calling equiv_constant; it is duplicated
3780 here for speed. */
3781 if (REGNO_QTY_VALID_P (REGNO (arg)))
3783 int arg_q = REG_QTY (REGNO (arg));
3784 struct qty_table_elem *arg_ent = &qty_table[arg_q];
3786 if (arg_ent->const_rtx != NULL_RTX
3787 && !REG_P (arg_ent->const_rtx)
3788 && GET_CODE (arg_ent->const_rtx) != PLUS)
3789 const_arg
3790 = gen_lowpart (GET_MODE (arg),
3791 arg_ent->const_rtx);
3793 break;
3795 case CONST:
3796 case CONST_INT:
3797 case SYMBOL_REF:
3798 case LABEL_REF:
3799 case CONST_DOUBLE:
3800 case CONST_VECTOR:
3801 const_arg = arg;
3802 break;
3804 #ifdef HAVE_cc0
3805 case CC0:
3806 folded_arg = prev_insn_cc0;
3807 mode_arg = prev_insn_cc0_mode;
3808 const_arg = equiv_constant (folded_arg);
3809 break;
3810 #endif
3812 default:
3813 folded_arg = fold_rtx (arg, insn);
3814 const_arg = equiv_constant (folded_arg);
3817 /* For the first three operands, see if the operand
3818 is constant or equivalent to a constant. */
3819 switch (i)
3821 case 0:
3822 folded_arg0 = folded_arg;
3823 const_arg0 = const_arg;
3824 mode_arg0 = mode_arg;
3825 break;
3826 case 1:
3827 folded_arg1 = folded_arg;
3828 const_arg1 = const_arg;
3829 break;
3830 case 2:
3831 const_arg2 = const_arg;
3832 break;
3835 /* Pick the least expensive of the folded argument and an
3836 equivalent constant argument. */
3837 if (const_arg == 0 || const_arg == folded_arg
3838 || COST_IN (const_arg, code) > COST_IN (folded_arg, code))
3839 cheap_arg = folded_arg, expensive_arg = const_arg;
3840 else
3841 cheap_arg = const_arg, expensive_arg = folded_arg;
3843 /* Try to replace the operand with the cheapest of the two
3844 possibilities. If it doesn't work and this is either of the first
3845 two operands of a commutative operation, try swapping them.
3846 If THAT fails, try the more expensive, provided it is cheaper
3847 than what is already there. */
3849 if (cheap_arg == XEXP (x, i))
3850 continue;
3852 if (insn == 0 && ! copied)
3854 x = copy_rtx (x);
3855 copied = 1;
3858 /* Order the replacements from cheapest to most expensive. */
3859 replacements[0] = cheap_arg;
3860 replacements[1] = expensive_arg;
3862 for (j = 0; j < 2 && replacements[j]; j++)
3864 int new_cost = COST_IN (replacements[j], code);
3866 /* Stop if what existed before was cheaper. Prefer constants
3867 in the case of a tie. */
3868 if (new_cost > old_cost
3869 || (new_cost == old_cost && CONSTANT_P (XEXP (x, i))))
3870 break;
3872 /* It's not safe to substitute the operand of a conversion
3873 operator with a constant, as the conversion's identity
3874 depends upon the mode of its operand. This optimization
3875 is handled by the call to simplify_unary_operation. */
3876 if (GET_RTX_CLASS (code) == RTX_UNARY
3877 && GET_MODE (replacements[j]) != mode_arg0
3878 && (code == ZERO_EXTEND
3879 || code == SIGN_EXTEND
3880 || code == TRUNCATE
3881 || code == FLOAT_TRUNCATE
3882 || code == FLOAT_EXTEND
3883 || code == FLOAT
3884 || code == FIX
3885 || code == UNSIGNED_FLOAT
3886 || code == UNSIGNED_FIX))
3887 continue;
3889 if (validate_change (insn, &XEXP (x, i), replacements[j], 0))
3890 break;
3892 if (GET_RTX_CLASS (code) == RTX_COMM_COMPARE
3893 || GET_RTX_CLASS (code) == RTX_COMM_ARITH)
3895 validate_change (insn, &XEXP (x, i), XEXP (x, 1 - i), 1);
3896 validate_change (insn, &XEXP (x, 1 - i), replacements[j], 1);
3898 if (apply_change_group ())
3900 /* Swap them back to be invalid so that this loop can
3901 continue and flag them to be swapped back later. */
3902 rtx tem;
3904 tem = XEXP (x, 0); XEXP (x, 0) = XEXP (x, 1);
3905 XEXP (x, 1) = tem;
3906 must_swap = 1;
3907 break;
3913 else
3915 if (fmt[i] == 'E')
3916 /* Don't try to fold inside of a vector of expressions.
3917 Doing nothing is harmless. */
3921 /* If a commutative operation, place a constant integer as the second
3922 operand unless the first operand is also a constant integer. Otherwise,
3923 place any constant second unless the first operand is also a constant. */
3925 if (COMMUTATIVE_P (x))
3927 if (must_swap
3928 || swap_commutative_operands_p (const_arg0 ? const_arg0
3929 : XEXP (x, 0),
3930 const_arg1 ? const_arg1
3931 : XEXP (x, 1)))
3933 rtx tem = XEXP (x, 0);
3935 if (insn == 0 && ! copied)
3937 x = copy_rtx (x);
3938 copied = 1;
3941 validate_change (insn, &XEXP (x, 0), XEXP (x, 1), 1);
3942 validate_change (insn, &XEXP (x, 1), tem, 1);
3943 if (apply_change_group ())
3945 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3946 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3951 /* If X is an arithmetic operation, see if we can simplify it. */
3953 switch (GET_RTX_CLASS (code))
3955 case RTX_UNARY:
3957 int is_const = 0;
3959 /* We can't simplify extension ops unless we know the
3960 original mode. */
3961 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3962 && mode_arg0 == VOIDmode)
3963 break;
3965 /* If we had a CONST, strip it off and put it back later if we
3966 fold. */
3967 if (const_arg0 != 0 && GET_CODE (const_arg0) == CONST)
3968 is_const = 1, const_arg0 = XEXP (const_arg0, 0);
3970 new = simplify_unary_operation (code, mode,
3971 const_arg0 ? const_arg0 : folded_arg0,
3972 mode_arg0);
3973 /* NEG of PLUS could be converted into MINUS, but that causes
3974 expressions of the form
3975 (CONST (MINUS (CONST_INT) (SYMBOL_REF)))
3976 which many ports mistakenly treat as LEGITIMATE_CONSTANT_P.
3977 FIXME: those ports should be fixed. */
3978 if (new != 0 && is_const
3979 && GET_CODE (new) == PLUS
3980 && (GET_CODE (XEXP (new, 0)) == SYMBOL_REF
3981 || GET_CODE (XEXP (new, 0)) == LABEL_REF)
3982 && GET_CODE (XEXP (new, 1)) == CONST_INT)
3983 new = gen_rtx_CONST (mode, new);
3985 break;
3987 case RTX_COMPARE:
3988 case RTX_COMM_COMPARE:
3989 /* See what items are actually being compared and set FOLDED_ARG[01]
3990 to those values and CODE to the actual comparison code. If any are
3991 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3992 do anything if both operands are already known to be constant. */
3994 /* ??? Vector mode comparisons are not supported yet. */
3995 if (VECTOR_MODE_P (mode))
3996 break;
3998 if (const_arg0 == 0 || const_arg1 == 0)
4000 struct table_elt *p0, *p1;
4001 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4002 enum machine_mode mode_arg1;
4004 #ifdef FLOAT_STORE_FLAG_VALUE
4005 if (SCALAR_FLOAT_MODE_P (mode))
4007 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4008 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4009 false_rtx = CONST0_RTX (mode);
4011 #endif
4013 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
4014 &mode_arg0, &mode_arg1);
4016 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
4017 what kinds of things are being compared, so we can't do
4018 anything with this comparison. */
4020 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
4021 break;
4023 const_arg0 = equiv_constant (folded_arg0);
4024 const_arg1 = equiv_constant (folded_arg1);
4026 /* If we do not now have two constants being compared, see
4027 if we can nevertheless deduce some things about the
4028 comparison. */
4029 if (const_arg0 == 0 || const_arg1 == 0)
4031 if (const_arg1 != NULL)
4033 rtx cheapest_simplification;
4034 int cheapest_cost;
4035 rtx simp_result;
4036 struct table_elt *p;
4038 /* See if we can find an equivalent of folded_arg0
4039 that gets us a cheaper expression, possibly a
4040 constant through simplifications. */
4041 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
4042 mode_arg0);
4044 if (p != NULL)
4046 cheapest_simplification = x;
4047 cheapest_cost = COST (x);
4049 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
4051 int cost;
4053 /* If the entry isn't valid, skip it. */
4054 if (! exp_equiv_p (p->exp, p->exp, 1, false))
4055 continue;
4057 /* Try to simplify using this equivalence. */
4058 simp_result
4059 = simplify_relational_operation (code, mode,
4060 mode_arg0,
4061 p->exp,
4062 const_arg1);
4064 if (simp_result == NULL)
4065 continue;
4067 cost = COST (simp_result);
4068 if (cost < cheapest_cost)
4070 cheapest_cost = cost;
4071 cheapest_simplification = simp_result;
4075 /* If we have a cheaper expression now, use that
4076 and try folding it further, from the top. */
4077 if (cheapest_simplification != x)
4078 return fold_rtx (cheapest_simplification, insn);
4082 /* Some addresses are known to be nonzero. We don't know
4083 their sign, but equality comparisons are known. */
4084 if (const_arg1 == const0_rtx
4085 && nonzero_address_p (folded_arg0))
4087 if (code == EQ)
4088 return false_rtx;
4089 else if (code == NE)
4090 return true_rtx;
4093 /* See if the two operands are the same. */
4095 if (folded_arg0 == folded_arg1
4096 || (REG_P (folded_arg0)
4097 && REG_P (folded_arg1)
4098 && (REG_QTY (REGNO (folded_arg0))
4099 == REG_QTY (REGNO (folded_arg1))))
4100 || ((p0 = lookup (folded_arg0,
4101 SAFE_HASH (folded_arg0, mode_arg0),
4102 mode_arg0))
4103 && (p1 = lookup (folded_arg1,
4104 SAFE_HASH (folded_arg1, mode_arg0),
4105 mode_arg0))
4106 && p0->first_same_value == p1->first_same_value))
4108 /* Sadly two equal NaNs are not equivalent. */
4109 if (!HONOR_NANS (mode_arg0))
4110 return ((code == EQ || code == LE || code == GE
4111 || code == LEU || code == GEU || code == UNEQ
4112 || code == UNLE || code == UNGE
4113 || code == ORDERED)
4114 ? true_rtx : false_rtx);
4115 /* Take care for the FP compares we can resolve. */
4116 if (code == UNEQ || code == UNLE || code == UNGE)
4117 return true_rtx;
4118 if (code == LTGT || code == LT || code == GT)
4119 return false_rtx;
4122 /* If FOLDED_ARG0 is a register, see if the comparison we are
4123 doing now is either the same as we did before or the reverse
4124 (we only check the reverse if not floating-point). */
4125 else if (REG_P (folded_arg0))
4127 int qty = REG_QTY (REGNO (folded_arg0));
4129 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
4131 struct qty_table_elem *ent = &qty_table[qty];
4133 if ((comparison_dominates_p (ent->comparison_code, code)
4134 || (! FLOAT_MODE_P (mode_arg0)
4135 && comparison_dominates_p (ent->comparison_code,
4136 reverse_condition (code))))
4137 && (rtx_equal_p (ent->comparison_const, folded_arg1)
4138 || (const_arg1
4139 && rtx_equal_p (ent->comparison_const,
4140 const_arg1))
4141 || (REG_P (folded_arg1)
4142 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
4143 return (comparison_dominates_p (ent->comparison_code, code)
4144 ? true_rtx : false_rtx);
4150 /* If we are comparing against zero, see if the first operand is
4151 equivalent to an IOR with a constant. If so, we may be able to
4152 determine the result of this comparison. */
4154 if (const_arg1 == const0_rtx)
4156 rtx y = lookup_as_function (folded_arg0, IOR);
4157 rtx inner_const;
4159 if (y != 0
4160 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
4161 && GET_CODE (inner_const) == CONST_INT
4162 && INTVAL (inner_const) != 0)
4164 int sign_bitnum = GET_MODE_BITSIZE (mode_arg0) - 1;
4165 int has_sign = (HOST_BITS_PER_WIDE_INT >= sign_bitnum
4166 && (INTVAL (inner_const)
4167 & ((HOST_WIDE_INT) 1 << sign_bitnum)));
4168 rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
4170 #ifdef FLOAT_STORE_FLAG_VALUE
4171 if (SCALAR_FLOAT_MODE_P (mode))
4173 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
4174 (FLOAT_STORE_FLAG_VALUE (mode), mode));
4175 false_rtx = CONST0_RTX (mode);
4177 #endif
4179 switch (code)
4181 case EQ:
4182 return false_rtx;
4183 case NE:
4184 return true_rtx;
4185 case LT: case LE:
4186 if (has_sign)
4187 return true_rtx;
4188 break;
4189 case GT: case GE:
4190 if (has_sign)
4191 return false_rtx;
4192 break;
4193 default:
4194 break;
4200 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
4201 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
4202 new = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
4204 break;
4206 case RTX_BIN_ARITH:
4207 case RTX_COMM_ARITH:
4208 switch (code)
4210 case PLUS:
4211 /* If the second operand is a LABEL_REF, see if the first is a MINUS
4212 with that LABEL_REF as its second operand. If so, the result is
4213 the first operand of that MINUS. This handles switches with an
4214 ADDR_DIFF_VEC table. */
4215 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
4217 rtx y
4218 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
4219 : lookup_as_function (folded_arg0, MINUS);
4221 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4222 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
4223 return XEXP (y, 0);
4225 /* Now try for a CONST of a MINUS like the above. */
4226 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
4227 : lookup_as_function (folded_arg0, CONST))) != 0
4228 && GET_CODE (XEXP (y, 0)) == MINUS
4229 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4230 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
4231 return XEXP (XEXP (y, 0), 0);
4234 /* Likewise if the operands are in the other order. */
4235 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
4237 rtx y
4238 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
4239 : lookup_as_function (folded_arg1, MINUS);
4241 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
4242 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
4243 return XEXP (y, 0);
4245 /* Now try for a CONST of a MINUS like the above. */
4246 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
4247 : lookup_as_function (folded_arg1, CONST))) != 0
4248 && GET_CODE (XEXP (y, 0)) == MINUS
4249 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
4250 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
4251 return XEXP (XEXP (y, 0), 0);
4254 /* If second operand is a register equivalent to a negative
4255 CONST_INT, see if we can find a register equivalent to the
4256 positive constant. Make a MINUS if so. Don't do this for
4257 a non-negative constant since we might then alternate between
4258 choosing positive and negative constants. Having the positive
4259 constant previously-used is the more common case. Be sure
4260 the resulting constant is non-negative; if const_arg1 were
4261 the smallest negative number this would overflow: depending
4262 on the mode, this would either just be the same value (and
4263 hence not save anything) or be incorrect. */
4264 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT
4265 && INTVAL (const_arg1) < 0
4266 /* This used to test
4268 -INTVAL (const_arg1) >= 0
4270 But The Sun V5.0 compilers mis-compiled that test. So
4271 instead we test for the problematic value in a more direct
4272 manner and hope the Sun compilers get it correct. */
4273 && INTVAL (const_arg1) !=
4274 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
4275 && REG_P (folded_arg1))
4277 rtx new_const = GEN_INT (-INTVAL (const_arg1));
4278 struct table_elt *p
4279 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
4281 if (p)
4282 for (p = p->first_same_value; p; p = p->next_same_value)
4283 if (REG_P (p->exp))
4284 return simplify_gen_binary (MINUS, mode, folded_arg0,
4285 canon_reg (p->exp, NULL_RTX));
4287 goto from_plus;
4289 case MINUS:
4290 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
4291 If so, produce (PLUS Z C2-C). */
4292 if (const_arg1 != 0 && GET_CODE (const_arg1) == CONST_INT)
4294 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
4295 if (y && GET_CODE (XEXP (y, 1)) == CONST_INT)
4296 return fold_rtx (plus_constant (copy_rtx (y),
4297 -INTVAL (const_arg1)),
4298 NULL_RTX);
4301 /* Fall through. */
4303 from_plus:
4304 case SMIN: case SMAX: case UMIN: case UMAX:
4305 case IOR: case AND: case XOR:
4306 case MULT:
4307 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
4308 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
4309 is known to be of similar form, we may be able to replace the
4310 operation with a combined operation. This may eliminate the
4311 intermediate operation if every use is simplified in this way.
4312 Note that the similar optimization done by combine.c only works
4313 if the intermediate operation's result has only one reference. */
4315 if (REG_P (folded_arg0)
4316 && const_arg1 && GET_CODE (const_arg1) == CONST_INT)
4318 int is_shift
4319 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
4320 rtx y, inner_const, new_const;
4321 enum rtx_code associate_code;
4323 if (is_shift
4324 && (INTVAL (const_arg1) >= GET_MODE_BITSIZE (mode)
4325 || INTVAL (const_arg1) < 0))
4327 if (SHIFT_COUNT_TRUNCATED)
4328 const_arg1 = GEN_INT (INTVAL (const_arg1)
4329 & (GET_MODE_BITSIZE (mode) - 1));
4330 else
4331 break;
4334 y = lookup_as_function (folded_arg0, code);
4335 if (y == 0)
4336 break;
4338 /* If we have compiled a statement like
4339 "if (x == (x & mask1))", and now are looking at
4340 "x & mask2", we will have a case where the first operand
4341 of Y is the same as our first operand. Unless we detect
4342 this case, an infinite loop will result. */
4343 if (XEXP (y, 0) == folded_arg0)
4344 break;
4346 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
4347 if (!inner_const || GET_CODE (inner_const) != CONST_INT)
4348 break;
4350 /* Don't associate these operations if they are a PLUS with the
4351 same constant and it is a power of two. These might be doable
4352 with a pre- or post-increment. Similarly for two subtracts of
4353 identical powers of two with post decrement. */
4355 if (code == PLUS && const_arg1 == inner_const
4356 && ((HAVE_PRE_INCREMENT
4357 && exact_log2 (INTVAL (const_arg1)) >= 0)
4358 || (HAVE_POST_INCREMENT
4359 && exact_log2 (INTVAL (const_arg1)) >= 0)
4360 || (HAVE_PRE_DECREMENT
4361 && exact_log2 (- INTVAL (const_arg1)) >= 0)
4362 || (HAVE_POST_DECREMENT
4363 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
4364 break;
4366 if (is_shift
4367 && (INTVAL (inner_const) >= GET_MODE_BITSIZE (mode)
4368 || INTVAL (inner_const) < 0))
4370 if (SHIFT_COUNT_TRUNCATED)
4371 inner_const = GEN_INT (INTVAL (inner_const)
4372 & (GET_MODE_BITSIZE (mode) - 1));
4373 else
4374 break;
4377 /* Compute the code used to compose the constants. For example,
4378 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
4380 associate_code = (is_shift || code == MINUS ? PLUS : code);
4382 new_const = simplify_binary_operation (associate_code, mode,
4383 const_arg1, inner_const);
4385 if (new_const == 0)
4386 break;
4388 /* If we are associating shift operations, don't let this
4389 produce a shift of the size of the object or larger.
4390 This could occur when we follow a sign-extend by a right
4391 shift on a machine that does a sign-extend as a pair
4392 of shifts. */
4394 if (is_shift
4395 && GET_CODE (new_const) == CONST_INT
4396 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
4398 /* As an exception, we can turn an ASHIFTRT of this
4399 form into a shift of the number of bits - 1. */
4400 if (code == ASHIFTRT)
4401 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
4402 else if (!side_effects_p (XEXP (y, 0)))
4403 return CONST0_RTX (mode);
4404 else
4405 break;
4408 y = copy_rtx (XEXP (y, 0));
4410 /* If Y contains our first operand (the most common way this
4411 can happen is if Y is a MEM), we would do into an infinite
4412 loop if we tried to fold it. So don't in that case. */
4414 if (! reg_mentioned_p (folded_arg0, y))
4415 y = fold_rtx (y, insn);
4417 return simplify_gen_binary (code, mode, y, new_const);
4419 break;
4421 case DIV: case UDIV:
4422 /* ??? The associative optimization performed immediately above is
4423 also possible for DIV and UDIV using associate_code of MULT.
4424 However, we would need extra code to verify that the
4425 multiplication does not overflow, that is, there is no overflow
4426 in the calculation of new_const. */
4427 break;
4429 default:
4430 break;
4433 new = simplify_binary_operation (code, mode,
4434 const_arg0 ? const_arg0 : folded_arg0,
4435 const_arg1 ? const_arg1 : folded_arg1);
4436 break;
4438 case RTX_OBJ:
4439 /* (lo_sum (high X) X) is simply X. */
4440 if (code == LO_SUM && const_arg0 != 0
4441 && GET_CODE (const_arg0) == HIGH
4442 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
4443 return const_arg1;
4444 break;
4446 case RTX_TERNARY:
4447 case RTX_BITFIELD_OPS:
4448 new = simplify_ternary_operation (code, mode, mode_arg0,
4449 const_arg0 ? const_arg0 : folded_arg0,
4450 const_arg1 ? const_arg1 : folded_arg1,
4451 const_arg2 ? const_arg2 : XEXP (x, 2));
4452 break;
4454 default:
4455 break;
4458 return new ? new : x;
4461 /* Return a constant value currently equivalent to X.
4462 Return 0 if we don't know one. */
4464 static rtx
4465 equiv_constant (rtx x)
4467 if (REG_P (x)
4468 && REGNO_QTY_VALID_P (REGNO (x)))
4470 int x_q = REG_QTY (REGNO (x));
4471 struct qty_table_elem *x_ent = &qty_table[x_q];
4473 if (x_ent->const_rtx)
4474 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
4477 if (x == 0 || CONSTANT_P (x))
4478 return x;
4480 /* If X is a MEM, try to fold it outside the context of any insn to see if
4481 it might be equivalent to a constant. That handles the case where it
4482 is a constant-pool reference. Then try to look it up in the hash table
4483 in case it is something whose value we have seen before. */
4485 if (MEM_P (x))
4487 struct table_elt *elt;
4489 x = fold_rtx (x, NULL_RTX);
4490 if (CONSTANT_P (x))
4491 return x;
4493 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
4494 if (elt == 0)
4495 return 0;
4497 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
4498 if (elt->is_const && CONSTANT_P (elt->exp))
4499 return elt->exp;
4502 return 0;
4505 /* Given INSN, a jump insn, PATH_TAKEN indicates if we are following the "taken"
4506 branch. It will be zero if not.
4508 In certain cases, this can cause us to add an equivalence. For example,
4509 if we are following the taken case of
4510 if (i == 2)
4511 we can add the fact that `i' and '2' are now equivalent.
4513 In any case, we can record that this comparison was passed. If the same
4514 comparison is seen later, we will know its value. */
4516 static void
4517 record_jump_equiv (rtx insn, int taken)
4519 int cond_known_true;
4520 rtx op0, op1;
4521 rtx set;
4522 enum machine_mode mode, mode0, mode1;
4523 int reversed_nonequality = 0;
4524 enum rtx_code code;
4526 /* Ensure this is the right kind of insn. */
4527 if (! any_condjump_p (insn))
4528 return;
4529 set = pc_set (insn);
4531 /* See if this jump condition is known true or false. */
4532 if (taken)
4533 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
4534 else
4535 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
4537 /* Get the type of comparison being done and the operands being compared.
4538 If we had to reverse a non-equality condition, record that fact so we
4539 know that it isn't valid for floating-point. */
4540 code = GET_CODE (XEXP (SET_SRC (set), 0));
4541 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
4542 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
4544 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
4545 if (! cond_known_true)
4547 code = reversed_comparison_code_parts (code, op0, op1, insn);
4549 /* Don't remember if we can't find the inverse. */
4550 if (code == UNKNOWN)
4551 return;
4554 /* The mode is the mode of the non-constant. */
4555 mode = mode0;
4556 if (mode1 != VOIDmode)
4557 mode = mode1;
4559 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
4562 /* Yet another form of subreg creation. In this case, we want something in
4563 MODE, and we should assume OP has MODE iff it is naturally modeless. */
4565 static rtx
4566 record_jump_cond_subreg (enum machine_mode mode, rtx op)
4568 enum machine_mode op_mode = GET_MODE (op);
4569 if (op_mode == mode || op_mode == VOIDmode)
4570 return op;
4571 return lowpart_subreg (mode, op, op_mode);
4574 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
4575 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
4576 Make any useful entries we can with that information. Called from
4577 above function and called recursively. */
4579 static void
4580 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
4581 rtx op1, int reversed_nonequality)
4583 unsigned op0_hash, op1_hash;
4584 int op0_in_memory, op1_in_memory;
4585 struct table_elt *op0_elt, *op1_elt;
4587 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
4588 we know that they are also equal in the smaller mode (this is also
4589 true for all smaller modes whether or not there is a SUBREG, but
4590 is not worth testing for with no SUBREG). */
4592 /* Note that GET_MODE (op0) may not equal MODE. */
4593 if (code == EQ && GET_CODE (op0) == SUBREG
4594 && (GET_MODE_SIZE (GET_MODE (op0))
4595 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4597 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4598 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4599 if (tem)
4600 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4601 reversed_nonequality);
4604 if (code == EQ && GET_CODE (op1) == SUBREG
4605 && (GET_MODE_SIZE (GET_MODE (op1))
4606 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4608 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4609 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4610 if (tem)
4611 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4612 reversed_nonequality);
4615 /* Similarly, if this is an NE comparison, and either is a SUBREG
4616 making a smaller mode, we know the whole thing is also NE. */
4618 /* Note that GET_MODE (op0) may not equal MODE;
4619 if we test MODE instead, we can get an infinite recursion
4620 alternating between two modes each wider than MODE. */
4622 if (code == NE && GET_CODE (op0) == SUBREG
4623 && subreg_lowpart_p (op0)
4624 && (GET_MODE_SIZE (GET_MODE (op0))
4625 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4627 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4628 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4629 if (tem)
4630 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4631 reversed_nonequality);
4634 if (code == NE && GET_CODE (op1) == SUBREG
4635 && subreg_lowpart_p (op1)
4636 && (GET_MODE_SIZE (GET_MODE (op1))
4637 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4639 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4640 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4641 if (tem)
4642 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4643 reversed_nonequality);
4646 /* Hash both operands. */
4648 do_not_record = 0;
4649 hash_arg_in_memory = 0;
4650 op0_hash = HASH (op0, mode);
4651 op0_in_memory = hash_arg_in_memory;
4653 if (do_not_record)
4654 return;
4656 do_not_record = 0;
4657 hash_arg_in_memory = 0;
4658 op1_hash = HASH (op1, mode);
4659 op1_in_memory = hash_arg_in_memory;
4661 if (do_not_record)
4662 return;
4664 /* Look up both operands. */
4665 op0_elt = lookup (op0, op0_hash, mode);
4666 op1_elt = lookup (op1, op1_hash, mode);
4668 /* If both operands are already equivalent or if they are not in the
4669 table but are identical, do nothing. */
4670 if ((op0_elt != 0 && op1_elt != 0
4671 && op0_elt->first_same_value == op1_elt->first_same_value)
4672 || op0 == op1 || rtx_equal_p (op0, op1))
4673 return;
4675 /* If we aren't setting two things equal all we can do is save this
4676 comparison. Similarly if this is floating-point. In the latter
4677 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4678 If we record the equality, we might inadvertently delete code
4679 whose intent was to change -0 to +0. */
4681 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4683 struct qty_table_elem *ent;
4684 int qty;
4686 /* If we reversed a floating-point comparison, if OP0 is not a
4687 register, or if OP1 is neither a register or constant, we can't
4688 do anything. */
4690 if (!REG_P (op1))
4691 op1 = equiv_constant (op1);
4693 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4694 || !REG_P (op0) || op1 == 0)
4695 return;
4697 /* Put OP0 in the hash table if it isn't already. This gives it a
4698 new quantity number. */
4699 if (op0_elt == 0)
4701 if (insert_regs (op0, NULL, 0))
4703 rehash_using_reg (op0);
4704 op0_hash = HASH (op0, mode);
4706 /* If OP0 is contained in OP1, this changes its hash code
4707 as well. Faster to rehash than to check, except
4708 for the simple case of a constant. */
4709 if (! CONSTANT_P (op1))
4710 op1_hash = HASH (op1,mode);
4713 op0_elt = insert (op0, NULL, op0_hash, mode);
4714 op0_elt->in_memory = op0_in_memory;
4717 qty = REG_QTY (REGNO (op0));
4718 ent = &qty_table[qty];
4720 ent->comparison_code = code;
4721 if (REG_P (op1))
4723 /* Look it up again--in case op0 and op1 are the same. */
4724 op1_elt = lookup (op1, op1_hash, mode);
4726 /* Put OP1 in the hash table so it gets a new quantity number. */
4727 if (op1_elt == 0)
4729 if (insert_regs (op1, NULL, 0))
4731 rehash_using_reg (op1);
4732 op1_hash = HASH (op1, mode);
4735 op1_elt = insert (op1, NULL, op1_hash, mode);
4736 op1_elt->in_memory = op1_in_memory;
4739 ent->comparison_const = NULL_RTX;
4740 ent->comparison_qty = REG_QTY (REGNO (op1));
4742 else
4744 ent->comparison_const = op1;
4745 ent->comparison_qty = -1;
4748 return;
4751 /* If either side is still missing an equivalence, make it now,
4752 then merge the equivalences. */
4754 if (op0_elt == 0)
4756 if (insert_regs (op0, NULL, 0))
4758 rehash_using_reg (op0);
4759 op0_hash = HASH (op0, mode);
4762 op0_elt = insert (op0, NULL, op0_hash, mode);
4763 op0_elt->in_memory = op0_in_memory;
4766 if (op1_elt == 0)
4768 if (insert_regs (op1, NULL, 0))
4770 rehash_using_reg (op1);
4771 op1_hash = HASH (op1, mode);
4774 op1_elt = insert (op1, NULL, op1_hash, mode);
4775 op1_elt->in_memory = op1_in_memory;
4778 merge_equiv_classes (op0_elt, op1_elt);
4781 /* CSE processing for one instruction.
4782 First simplify sources and addresses of all assignments
4783 in the instruction, using previously-computed equivalents values.
4784 Then install the new sources and destinations in the table
4785 of available values.
4787 If LIBCALL_INSN is nonzero, don't record any equivalence made in
4788 the insn. It means that INSN is inside libcall block. In this
4789 case LIBCALL_INSN is the corresponding insn with REG_LIBCALL. */
4791 /* Data on one SET contained in the instruction. */
4793 struct set
4795 /* The SET rtx itself. */
4796 rtx rtl;
4797 /* The SET_SRC of the rtx (the original value, if it is changing). */
4798 rtx src;
4799 /* The hash-table element for the SET_SRC of the SET. */
4800 struct table_elt *src_elt;
4801 /* Hash value for the SET_SRC. */
4802 unsigned src_hash;
4803 /* Hash value for the SET_DEST. */
4804 unsigned dest_hash;
4805 /* The SET_DEST, with SUBREG, etc., stripped. */
4806 rtx inner_dest;
4807 /* Nonzero if the SET_SRC is in memory. */
4808 char src_in_memory;
4809 /* Nonzero if the SET_SRC contains something
4810 whose value cannot be predicted and understood. */
4811 char src_volatile;
4812 /* Original machine mode, in case it becomes a CONST_INT.
4813 The size of this field should match the size of the mode
4814 field of struct rtx_def (see rtl.h). */
4815 ENUM_BITFIELD(machine_mode) mode : 8;
4816 /* A constant equivalent for SET_SRC, if any. */
4817 rtx src_const;
4818 /* Original SET_SRC value used for libcall notes. */
4819 rtx orig_src;
4820 /* Hash value of constant equivalent for SET_SRC. */
4821 unsigned src_const_hash;
4822 /* Table entry for constant equivalent for SET_SRC, if any. */
4823 struct table_elt *src_const_elt;
4824 /* Table entry for the destination address. */
4825 struct table_elt *dest_addr_elt;
4828 static void
4829 cse_insn (rtx insn, rtx libcall_insn)
4831 rtx x = PATTERN (insn);
4832 int i;
4833 rtx tem;
4834 int n_sets = 0;
4836 #ifdef HAVE_cc0
4837 /* Records what this insn does to set CC0. */
4838 rtx this_insn_cc0 = 0;
4839 enum machine_mode this_insn_cc0_mode = VOIDmode;
4840 #endif
4842 rtx src_eqv = 0;
4843 struct table_elt *src_eqv_elt = 0;
4844 int src_eqv_volatile = 0;
4845 int src_eqv_in_memory = 0;
4846 unsigned src_eqv_hash = 0;
4848 struct set *sets = (struct set *) 0;
4850 this_insn = insn;
4852 /* Find all the SETs and CLOBBERs in this instruction.
4853 Record all the SETs in the array `set' and count them.
4854 Also determine whether there is a CLOBBER that invalidates
4855 all memory references, or all references at varying addresses. */
4857 if (CALL_P (insn))
4859 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4861 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4862 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4863 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4867 if (GET_CODE (x) == SET)
4869 sets = alloca (sizeof (struct set));
4870 sets[0].rtl = x;
4872 /* Ignore SETs that are unconditional jumps.
4873 They never need cse processing, so this does not hurt.
4874 The reason is not efficiency but rather
4875 so that we can test at the end for instructions
4876 that have been simplified to unconditional jumps
4877 and not be misled by unchanged instructions
4878 that were unconditional jumps to begin with. */
4879 if (SET_DEST (x) == pc_rtx
4880 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4883 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4884 The hard function value register is used only once, to copy to
4885 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4886 Ensure we invalidate the destination register. On the 80386 no
4887 other code would invalidate it since it is a fixed_reg.
4888 We need not check the return of apply_change_group; see canon_reg. */
4890 else if (GET_CODE (SET_SRC (x)) == CALL)
4892 canon_reg (SET_SRC (x), insn);
4893 apply_change_group ();
4894 fold_rtx (SET_SRC (x), insn);
4895 invalidate (SET_DEST (x), VOIDmode);
4897 else
4898 n_sets = 1;
4900 else if (GET_CODE (x) == PARALLEL)
4902 int lim = XVECLEN (x, 0);
4904 sets = alloca (lim * sizeof (struct set));
4906 /* Find all regs explicitly clobbered in this insn,
4907 and ensure they are not replaced with any other regs
4908 elsewhere in this insn.
4909 When a reg that is clobbered is also used for input,
4910 we should presume that that is for a reason,
4911 and we should not substitute some other register
4912 which is not supposed to be clobbered.
4913 Therefore, this loop cannot be merged into the one below
4914 because a CALL may precede a CLOBBER and refer to the
4915 value clobbered. We must not let a canonicalization do
4916 anything in that case. */
4917 for (i = 0; i < lim; i++)
4919 rtx y = XVECEXP (x, 0, i);
4920 if (GET_CODE (y) == CLOBBER)
4922 rtx clobbered = XEXP (y, 0);
4924 if (REG_P (clobbered)
4925 || GET_CODE (clobbered) == SUBREG)
4926 invalidate (clobbered, VOIDmode);
4927 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4928 || GET_CODE (clobbered) == ZERO_EXTRACT)
4929 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4933 for (i = 0; i < lim; i++)
4935 rtx y = XVECEXP (x, 0, i);
4936 if (GET_CODE (y) == SET)
4938 /* As above, we ignore unconditional jumps and call-insns and
4939 ignore the result of apply_change_group. */
4940 if (GET_CODE (SET_SRC (y)) == CALL)
4942 canon_reg (SET_SRC (y), insn);
4943 apply_change_group ();
4944 fold_rtx (SET_SRC (y), insn);
4945 invalidate (SET_DEST (y), VOIDmode);
4947 else if (SET_DEST (y) == pc_rtx
4948 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4950 else
4951 sets[n_sets++].rtl = y;
4953 else if (GET_CODE (y) == CLOBBER)
4955 /* If we clobber memory, canon the address.
4956 This does nothing when a register is clobbered
4957 because we have already invalidated the reg. */
4958 if (MEM_P (XEXP (y, 0)))
4959 canon_reg (XEXP (y, 0), NULL_RTX);
4961 else if (GET_CODE (y) == USE
4962 && ! (REG_P (XEXP (y, 0))
4963 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4964 canon_reg (y, NULL_RTX);
4965 else if (GET_CODE (y) == CALL)
4967 /* The result of apply_change_group can be ignored; see
4968 canon_reg. */
4969 canon_reg (y, insn);
4970 apply_change_group ();
4971 fold_rtx (y, insn);
4975 else if (GET_CODE (x) == CLOBBER)
4977 if (MEM_P (XEXP (x, 0)))
4978 canon_reg (XEXP (x, 0), NULL_RTX);
4981 /* Canonicalize a USE of a pseudo register or memory location. */
4982 else if (GET_CODE (x) == USE
4983 && ! (REG_P (XEXP (x, 0))
4984 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4985 canon_reg (XEXP (x, 0), NULL_RTX);
4986 else if (GET_CODE (x) == CALL)
4988 /* The result of apply_change_group can be ignored; see canon_reg. */
4989 canon_reg (x, insn);
4990 apply_change_group ();
4991 fold_rtx (x, insn);
4994 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4995 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4996 is handled specially for this case, and if it isn't set, then there will
4997 be no equivalence for the destination. */
4998 if (n_sets == 1 && REG_NOTES (insn) != 0
4999 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
5000 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
5001 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
5003 src_eqv = fold_rtx (canon_reg (XEXP (tem, 0), NULL_RTX), insn);
5004 XEXP (tem, 0) = src_eqv;
5007 /* Canonicalize sources and addresses of destinations.
5008 We do this in a separate pass to avoid problems when a MATCH_DUP is
5009 present in the insn pattern. In that case, we want to ensure that
5010 we don't break the duplicate nature of the pattern. So we will replace
5011 both operands at the same time. Otherwise, we would fail to find an
5012 equivalent substitution in the loop calling validate_change below.
5014 We used to suppress canonicalization of DEST if it appears in SRC,
5015 but we don't do this any more. */
5017 for (i = 0; i < n_sets; i++)
5019 rtx dest = SET_DEST (sets[i].rtl);
5020 rtx src = SET_SRC (sets[i].rtl);
5021 rtx new = canon_reg (src, insn);
5023 sets[i].orig_src = src;
5024 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5026 if (GET_CODE (dest) == ZERO_EXTRACT)
5028 validate_change (insn, &XEXP (dest, 1),
5029 canon_reg (XEXP (dest, 1), insn), 1);
5030 validate_change (insn, &XEXP (dest, 2),
5031 canon_reg (XEXP (dest, 2), insn), 1);
5034 while (GET_CODE (dest) == SUBREG
5035 || GET_CODE (dest) == ZERO_EXTRACT
5036 || GET_CODE (dest) == STRICT_LOW_PART)
5037 dest = XEXP (dest, 0);
5039 if (MEM_P (dest))
5040 canon_reg (dest, insn);
5043 /* Now that we have done all the replacements, we can apply the change
5044 group and see if they all work. Note that this will cause some
5045 canonicalizations that would have worked individually not to be applied
5046 because some other canonicalization didn't work, but this should not
5047 occur often.
5049 The result of apply_change_group can be ignored; see canon_reg. */
5051 apply_change_group ();
5053 /* Set sets[i].src_elt to the class each source belongs to.
5054 Detect assignments from or to volatile things
5055 and set set[i] to zero so they will be ignored
5056 in the rest of this function.
5058 Nothing in this loop changes the hash table or the register chains. */
5060 for (i = 0; i < n_sets; i++)
5062 rtx src, dest;
5063 rtx src_folded;
5064 struct table_elt *elt = 0, *p;
5065 enum machine_mode mode;
5066 rtx src_eqv_here;
5067 rtx src_const = 0;
5068 rtx src_related = 0;
5069 struct table_elt *src_const_elt = 0;
5070 int src_cost = MAX_COST;
5071 int src_eqv_cost = MAX_COST;
5072 int src_folded_cost = MAX_COST;
5073 int src_related_cost = MAX_COST;
5074 int src_elt_cost = MAX_COST;
5075 int src_regcost = MAX_COST;
5076 int src_eqv_regcost = MAX_COST;
5077 int src_folded_regcost = MAX_COST;
5078 int src_related_regcost = MAX_COST;
5079 int src_elt_regcost = MAX_COST;
5080 /* Set nonzero if we need to call force_const_mem on with the
5081 contents of src_folded before using it. */
5082 int src_folded_force_flag = 0;
5084 dest = SET_DEST (sets[i].rtl);
5085 src = SET_SRC (sets[i].rtl);
5087 /* If SRC is a constant that has no machine mode,
5088 hash it with the destination's machine mode.
5089 This way we can keep different modes separate. */
5091 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5092 sets[i].mode = mode;
5094 if (src_eqv)
5096 enum machine_mode eqvmode = mode;
5097 if (GET_CODE (dest) == STRICT_LOW_PART)
5098 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5099 do_not_record = 0;
5100 hash_arg_in_memory = 0;
5101 src_eqv_hash = HASH (src_eqv, eqvmode);
5103 /* Find the equivalence class for the equivalent expression. */
5105 if (!do_not_record)
5106 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
5108 src_eqv_volatile = do_not_record;
5109 src_eqv_in_memory = hash_arg_in_memory;
5112 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
5113 value of the INNER register, not the destination. So it is not
5114 a valid substitution for the source. But save it for later. */
5115 if (GET_CODE (dest) == STRICT_LOW_PART)
5116 src_eqv_here = 0;
5117 else
5118 src_eqv_here = src_eqv;
5120 /* Simplify and foldable subexpressions in SRC. Then get the fully-
5121 simplified result, which may not necessarily be valid. */
5122 src_folded = fold_rtx (src, insn);
5124 #if 0
5125 /* ??? This caused bad code to be generated for the m68k port with -O2.
5126 Suppose src is (CONST_INT -1), and that after truncation src_folded
5127 is (CONST_INT 3). Suppose src_folded is then used for src_const.
5128 At the end we will add src and src_const to the same equivalence
5129 class. We now have 3 and -1 on the same equivalence class. This
5130 causes later instructions to be mis-optimized. */
5131 /* If storing a constant in a bitfield, pre-truncate the constant
5132 so we will be able to record it later. */
5133 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5135 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5137 if (GET_CODE (src) == CONST_INT
5138 && GET_CODE (width) == CONST_INT
5139 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5140 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5141 src_folded
5142 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
5143 << INTVAL (width)) - 1));
5145 #endif
5147 /* Compute SRC's hash code, and also notice if it
5148 should not be recorded at all. In that case,
5149 prevent any further processing of this assignment. */
5150 do_not_record = 0;
5151 hash_arg_in_memory = 0;
5153 sets[i].src = src;
5154 sets[i].src_hash = HASH (src, mode);
5155 sets[i].src_volatile = do_not_record;
5156 sets[i].src_in_memory = hash_arg_in_memory;
5158 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
5159 a pseudo, do not record SRC. Using SRC as a replacement for
5160 anything else will be incorrect in that situation. Note that
5161 this usually occurs only for stack slots, in which case all the
5162 RTL would be referring to SRC, so we don't lose any optimization
5163 opportunities by not having SRC in the hash table. */
5165 if (MEM_P (src)
5166 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
5167 && REG_P (dest)
5168 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
5169 sets[i].src_volatile = 1;
5171 #if 0
5172 /* It is no longer clear why we used to do this, but it doesn't
5173 appear to still be needed. So let's try without it since this
5174 code hurts cse'ing widened ops. */
5175 /* If source is a paradoxical subreg (such as QI treated as an SI),
5176 treat it as volatile. It may do the work of an SI in one context
5177 where the extra bits are not being used, but cannot replace an SI
5178 in general. */
5179 if (GET_CODE (src) == SUBREG
5180 && (GET_MODE_SIZE (GET_MODE (src))
5181 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
5182 sets[i].src_volatile = 1;
5183 #endif
5185 /* Locate all possible equivalent forms for SRC. Try to replace
5186 SRC in the insn with each cheaper equivalent.
5188 We have the following types of equivalents: SRC itself, a folded
5189 version, a value given in a REG_EQUAL note, or a value related
5190 to a constant.
5192 Each of these equivalents may be part of an additional class
5193 of equivalents (if more than one is in the table, they must be in
5194 the same class; we check for this).
5196 If the source is volatile, we don't do any table lookups.
5198 We note any constant equivalent for possible later use in a
5199 REG_NOTE. */
5201 if (!sets[i].src_volatile)
5202 elt = lookup (src, sets[i].src_hash, mode);
5204 sets[i].src_elt = elt;
5206 if (elt && src_eqv_here && src_eqv_elt)
5208 if (elt->first_same_value != src_eqv_elt->first_same_value)
5210 /* The REG_EQUAL is indicating that two formerly distinct
5211 classes are now equivalent. So merge them. */
5212 merge_equiv_classes (elt, src_eqv_elt);
5213 src_eqv_hash = HASH (src_eqv, elt->mode);
5214 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
5217 src_eqv_here = 0;
5220 else if (src_eqv_elt)
5221 elt = src_eqv_elt;
5223 /* Try to find a constant somewhere and record it in `src_const'.
5224 Record its table element, if any, in `src_const_elt'. Look in
5225 any known equivalences first. (If the constant is not in the
5226 table, also set `sets[i].src_const_hash'). */
5227 if (elt)
5228 for (p = elt->first_same_value; p; p = p->next_same_value)
5229 if (p->is_const)
5231 src_const = p->exp;
5232 src_const_elt = elt;
5233 break;
5236 if (src_const == 0
5237 && (CONSTANT_P (src_folded)
5238 /* Consider (minus (label_ref L1) (label_ref L2)) as
5239 "constant" here so we will record it. This allows us
5240 to fold switch statements when an ADDR_DIFF_VEC is used. */
5241 || (GET_CODE (src_folded) == MINUS
5242 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
5243 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
5244 src_const = src_folded, src_const_elt = elt;
5245 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
5246 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
5248 /* If we don't know if the constant is in the table, get its
5249 hash code and look it up. */
5250 if (src_const && src_const_elt == 0)
5252 sets[i].src_const_hash = HASH (src_const, mode);
5253 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
5256 sets[i].src_const = src_const;
5257 sets[i].src_const_elt = src_const_elt;
5259 /* If the constant and our source are both in the table, mark them as
5260 equivalent. Otherwise, if a constant is in the table but the source
5261 isn't, set ELT to it. */
5262 if (src_const_elt && elt
5263 && src_const_elt->first_same_value != elt->first_same_value)
5264 merge_equiv_classes (elt, src_const_elt);
5265 else if (src_const_elt && elt == 0)
5266 elt = src_const_elt;
5268 /* See if there is a register linearly related to a constant
5269 equivalent of SRC. */
5270 if (src_const
5271 && (GET_CODE (src_const) == CONST
5272 || (src_const_elt && src_const_elt->related_value != 0)))
5274 src_related = use_related_value (src_const, src_const_elt);
5275 if (src_related)
5277 struct table_elt *src_related_elt
5278 = lookup (src_related, HASH (src_related, mode), mode);
5279 if (src_related_elt && elt)
5281 if (elt->first_same_value
5282 != src_related_elt->first_same_value)
5283 /* This can occur when we previously saw a CONST
5284 involving a SYMBOL_REF and then see the SYMBOL_REF
5285 twice. Merge the involved classes. */
5286 merge_equiv_classes (elt, src_related_elt);
5288 src_related = 0;
5289 src_related_elt = 0;
5291 else if (src_related_elt && elt == 0)
5292 elt = src_related_elt;
5296 /* See if we have a CONST_INT that is already in a register in a
5297 wider mode. */
5299 if (src_const && src_related == 0 && GET_CODE (src_const) == CONST_INT
5300 && GET_MODE_CLASS (mode) == MODE_INT
5301 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
5303 enum machine_mode wider_mode;
5305 for (wider_mode = GET_MODE_WIDER_MODE (mode);
5306 GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
5307 && src_related == 0;
5308 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
5310 struct table_elt *const_elt
5311 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
5313 if (const_elt == 0)
5314 continue;
5316 for (const_elt = const_elt->first_same_value;
5317 const_elt; const_elt = const_elt->next_same_value)
5318 if (REG_P (const_elt->exp))
5320 src_related = gen_lowpart (mode,
5321 const_elt->exp);
5322 break;
5327 /* Another possibility is that we have an AND with a constant in
5328 a mode narrower than a word. If so, it might have been generated
5329 as part of an "if" which would narrow the AND. If we already
5330 have done the AND in a wider mode, we can use a SUBREG of that
5331 value. */
5333 if (flag_expensive_optimizations && ! src_related
5334 && GET_CODE (src) == AND && GET_CODE (XEXP (src, 1)) == CONST_INT
5335 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5337 enum machine_mode tmode;
5338 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
5340 for (tmode = GET_MODE_WIDER_MODE (mode);
5341 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5342 tmode = GET_MODE_WIDER_MODE (tmode))
5344 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
5345 struct table_elt *larger_elt;
5347 if (inner)
5349 PUT_MODE (new_and, tmode);
5350 XEXP (new_and, 0) = inner;
5351 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
5352 if (larger_elt == 0)
5353 continue;
5355 for (larger_elt = larger_elt->first_same_value;
5356 larger_elt; larger_elt = larger_elt->next_same_value)
5357 if (REG_P (larger_elt->exp))
5359 src_related
5360 = gen_lowpart (mode, larger_elt->exp);
5361 break;
5364 if (src_related)
5365 break;
5370 #ifdef LOAD_EXTEND_OP
5371 /* See if a MEM has already been loaded with a widening operation;
5372 if it has, we can use a subreg of that. Many CISC machines
5373 also have such operations, but this is only likely to be
5374 beneficial on these machines. */
5376 if (flag_expensive_optimizations && src_related == 0
5377 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
5378 && GET_MODE_CLASS (mode) == MODE_INT
5379 && MEM_P (src) && ! do_not_record
5380 && LOAD_EXTEND_OP (mode) != UNKNOWN)
5382 struct rtx_def memory_extend_buf;
5383 rtx memory_extend_rtx = &memory_extend_buf;
5384 enum machine_mode tmode;
5386 /* Set what we are trying to extend and the operation it might
5387 have been extended with. */
5388 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
5389 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
5390 XEXP (memory_extend_rtx, 0) = src;
5392 for (tmode = GET_MODE_WIDER_MODE (mode);
5393 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
5394 tmode = GET_MODE_WIDER_MODE (tmode))
5396 struct table_elt *larger_elt;
5398 PUT_MODE (memory_extend_rtx, tmode);
5399 larger_elt = lookup (memory_extend_rtx,
5400 HASH (memory_extend_rtx, tmode), tmode);
5401 if (larger_elt == 0)
5402 continue;
5404 for (larger_elt = larger_elt->first_same_value;
5405 larger_elt; larger_elt = larger_elt->next_same_value)
5406 if (REG_P (larger_elt->exp))
5408 src_related = gen_lowpart (mode,
5409 larger_elt->exp);
5410 break;
5413 if (src_related)
5414 break;
5417 #endif /* LOAD_EXTEND_OP */
5419 if (src == src_folded)
5420 src_folded = 0;
5422 /* At this point, ELT, if nonzero, points to a class of expressions
5423 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
5424 and SRC_RELATED, if nonzero, each contain additional equivalent
5425 expressions. Prune these latter expressions by deleting expressions
5426 already in the equivalence class.
5428 Check for an equivalent identical to the destination. If found,
5429 this is the preferred equivalent since it will likely lead to
5430 elimination of the insn. Indicate this by placing it in
5431 `src_related'. */
5433 if (elt)
5434 elt = elt->first_same_value;
5435 for (p = elt; p; p = p->next_same_value)
5437 enum rtx_code code = GET_CODE (p->exp);
5439 /* If the expression is not valid, ignore it. Then we do not
5440 have to check for validity below. In most cases, we can use
5441 `rtx_equal_p', since canonicalization has already been done. */
5442 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5443 continue;
5445 /* Also skip paradoxical subregs, unless that's what we're
5446 looking for. */
5447 if (code == SUBREG
5448 && (GET_MODE_SIZE (GET_MODE (p->exp))
5449 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
5450 && ! (src != 0
5451 && GET_CODE (src) == SUBREG
5452 && GET_MODE (src) == GET_MODE (p->exp)
5453 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5454 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5455 continue;
5457 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5458 src = 0;
5459 else if (src_folded && GET_CODE (src_folded) == code
5460 && rtx_equal_p (src_folded, p->exp))
5461 src_folded = 0;
5462 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5463 && rtx_equal_p (src_eqv_here, p->exp))
5464 src_eqv_here = 0;
5465 else if (src_related && GET_CODE (src_related) == code
5466 && rtx_equal_p (src_related, p->exp))
5467 src_related = 0;
5469 /* This is the same as the destination of the insns, we want
5470 to prefer it. Copy it to src_related. The code below will
5471 then give it a negative cost. */
5472 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5473 src_related = dest;
5476 /* Find the cheapest valid equivalent, trying all the available
5477 possibilities. Prefer items not in the hash table to ones
5478 that are when they are equal cost. Note that we can never
5479 worsen an insn as the current contents will also succeed.
5480 If we find an equivalent identical to the destination, use it as best,
5481 since this insn will probably be eliminated in that case. */
5482 if (src)
5484 if (rtx_equal_p (src, dest))
5485 src_cost = src_regcost = -1;
5486 else
5488 src_cost = COST (src);
5489 src_regcost = approx_reg_cost (src);
5493 if (src_eqv_here)
5495 if (rtx_equal_p (src_eqv_here, dest))
5496 src_eqv_cost = src_eqv_regcost = -1;
5497 else
5499 src_eqv_cost = COST (src_eqv_here);
5500 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5504 if (src_folded)
5506 if (rtx_equal_p (src_folded, dest))
5507 src_folded_cost = src_folded_regcost = -1;
5508 else
5510 src_folded_cost = COST (src_folded);
5511 src_folded_regcost = approx_reg_cost (src_folded);
5515 if (src_related)
5517 if (rtx_equal_p (src_related, dest))
5518 src_related_cost = src_related_regcost = -1;
5519 else
5521 src_related_cost = COST (src_related);
5522 src_related_regcost = approx_reg_cost (src_related);
5526 /* If this was an indirect jump insn, a known label will really be
5527 cheaper even though it looks more expensive. */
5528 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5529 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5531 /* Terminate loop when replacement made. This must terminate since
5532 the current contents will be tested and will always be valid. */
5533 while (1)
5535 rtx trial;
5537 /* Skip invalid entries. */
5538 while (elt && !REG_P (elt->exp)
5539 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5540 elt = elt->next_same_value;
5542 /* A paradoxical subreg would be bad here: it'll be the right
5543 size, but later may be adjusted so that the upper bits aren't
5544 what we want. So reject it. */
5545 if (elt != 0
5546 && GET_CODE (elt->exp) == SUBREG
5547 && (GET_MODE_SIZE (GET_MODE (elt->exp))
5548 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
5549 /* It is okay, though, if the rtx we're trying to match
5550 will ignore any of the bits we can't predict. */
5551 && ! (src != 0
5552 && GET_CODE (src) == SUBREG
5553 && GET_MODE (src) == GET_MODE (elt->exp)
5554 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5555 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5557 elt = elt->next_same_value;
5558 continue;
5561 if (elt)
5563 src_elt_cost = elt->cost;
5564 src_elt_regcost = elt->regcost;
5567 /* Find cheapest and skip it for the next time. For items
5568 of equal cost, use this order:
5569 src_folded, src, src_eqv, src_related and hash table entry. */
5570 if (src_folded
5571 && preferable (src_folded_cost, src_folded_regcost,
5572 src_cost, src_regcost) <= 0
5573 && preferable (src_folded_cost, src_folded_regcost,
5574 src_eqv_cost, src_eqv_regcost) <= 0
5575 && preferable (src_folded_cost, src_folded_regcost,
5576 src_related_cost, src_related_regcost) <= 0
5577 && preferable (src_folded_cost, src_folded_regcost,
5578 src_elt_cost, src_elt_regcost) <= 0)
5580 trial = src_folded, src_folded_cost = MAX_COST;
5581 if (src_folded_force_flag)
5583 rtx forced = force_const_mem (mode, trial);
5584 if (forced)
5585 trial = forced;
5588 else if (src
5589 && preferable (src_cost, src_regcost,
5590 src_eqv_cost, src_eqv_regcost) <= 0
5591 && preferable (src_cost, src_regcost,
5592 src_related_cost, src_related_regcost) <= 0
5593 && preferable (src_cost, src_regcost,
5594 src_elt_cost, src_elt_regcost) <= 0)
5595 trial = src, src_cost = MAX_COST;
5596 else if (src_eqv_here
5597 && preferable (src_eqv_cost, src_eqv_regcost,
5598 src_related_cost, src_related_regcost) <= 0
5599 && preferable (src_eqv_cost, src_eqv_regcost,
5600 src_elt_cost, src_elt_regcost) <= 0)
5601 trial = copy_rtx (src_eqv_here), src_eqv_cost = MAX_COST;
5602 else if (src_related
5603 && preferable (src_related_cost, src_related_regcost,
5604 src_elt_cost, src_elt_regcost) <= 0)
5605 trial = copy_rtx (src_related), src_related_cost = MAX_COST;
5606 else
5608 trial = copy_rtx (elt->exp);
5609 elt = elt->next_same_value;
5610 src_elt_cost = MAX_COST;
5613 /* We don't normally have an insn matching (set (pc) (pc)), so
5614 check for this separately here. We will delete such an
5615 insn below.
5617 For other cases such as a table jump or conditional jump
5618 where we know the ultimate target, go ahead and replace the
5619 operand. While that may not make a valid insn, we will
5620 reemit the jump below (and also insert any necessary
5621 barriers). */
5622 if (n_sets == 1 && dest == pc_rtx
5623 && (trial == pc_rtx
5624 || (GET_CODE (trial) == LABEL_REF
5625 && ! condjump_p (insn))))
5627 /* Don't substitute non-local labels, this confuses CFG. */
5628 if (GET_CODE (trial) == LABEL_REF
5629 && LABEL_REF_NONLOCAL_P (trial))
5630 continue;
5632 SET_SRC (sets[i].rtl) = trial;
5633 cse_jumps_altered = 1;
5634 break;
5637 /* Reject certain invalid forms of CONST that we create. */
5638 else if (CONSTANT_P (trial)
5639 && GET_CODE (trial) == CONST
5640 /* Reject cases that will cause decode_rtx_const to
5641 die. On the alpha when simplifying a switch, we
5642 get (const (truncate (minus (label_ref)
5643 (label_ref)))). */
5644 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5645 /* Likewise on IA-64, except without the
5646 truncate. */
5647 || (GET_CODE (XEXP (trial, 0)) == MINUS
5648 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5649 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5650 /* Do nothing for this case. */
5653 /* Look for a substitution that makes a valid insn. */
5654 else if (validate_change (insn, &SET_SRC (sets[i].rtl), trial, 0))
5656 rtx new = canon_reg (SET_SRC (sets[i].rtl), insn);
5658 /* If we just made a substitution inside a libcall, then we
5659 need to make the same substitution in any notes attached
5660 to the RETVAL insn. */
5661 if (libcall_insn
5662 && (REG_P (sets[i].orig_src)
5663 || GET_CODE (sets[i].orig_src) == SUBREG
5664 || MEM_P (sets[i].orig_src)))
5666 rtx note = find_reg_equal_equiv_note (libcall_insn);
5667 if (note != 0)
5668 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0),
5669 sets[i].orig_src,
5670 copy_rtx (new));
5673 /* The result of apply_change_group can be ignored; see
5674 canon_reg. */
5676 validate_change (insn, &SET_SRC (sets[i].rtl), new, 1);
5677 apply_change_group ();
5678 break;
5681 /* If we previously found constant pool entries for
5682 constants and this is a constant, try making a
5683 pool entry. Put it in src_folded unless we already have done
5684 this since that is where it likely came from. */
5686 else if (constant_pool_entries_cost
5687 && CONSTANT_P (trial)
5688 && (src_folded == 0
5689 || (!MEM_P (src_folded)
5690 && ! src_folded_force_flag))
5691 && GET_MODE_CLASS (mode) != MODE_CC
5692 && mode != VOIDmode)
5694 src_folded_force_flag = 1;
5695 src_folded = trial;
5696 src_folded_cost = constant_pool_entries_cost;
5697 src_folded_regcost = constant_pool_entries_regcost;
5701 src = SET_SRC (sets[i].rtl);
5703 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5704 However, there is an important exception: If both are registers
5705 that are not the head of their equivalence class, replace SET_SRC
5706 with the head of the class. If we do not do this, we will have
5707 both registers live over a portion of the basic block. This way,
5708 their lifetimes will likely abut instead of overlapping. */
5709 if (REG_P (dest)
5710 && REGNO_QTY_VALID_P (REGNO (dest)))
5712 int dest_q = REG_QTY (REGNO (dest));
5713 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5715 if (dest_ent->mode == GET_MODE (dest)
5716 && dest_ent->first_reg != REGNO (dest)
5717 && REG_P (src) && REGNO (src) == REGNO (dest)
5718 /* Don't do this if the original insn had a hard reg as
5719 SET_SRC or SET_DEST. */
5720 && (!REG_P (sets[i].src)
5721 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5722 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5723 /* We can't call canon_reg here because it won't do anything if
5724 SRC is a hard register. */
5726 int src_q = REG_QTY (REGNO (src));
5727 struct qty_table_elem *src_ent = &qty_table[src_q];
5728 int first = src_ent->first_reg;
5729 rtx new_src
5730 = (first >= FIRST_PSEUDO_REGISTER
5731 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5733 /* We must use validate-change even for this, because this
5734 might be a special no-op instruction, suitable only to
5735 tag notes onto. */
5736 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5738 src = new_src;
5739 /* If we had a constant that is cheaper than what we are now
5740 setting SRC to, use that constant. We ignored it when we
5741 thought we could make this into a no-op. */
5742 if (src_const && COST (src_const) < COST (src)
5743 && validate_change (insn, &SET_SRC (sets[i].rtl),
5744 src_const, 0))
5745 src = src_const;
5750 /* If we made a change, recompute SRC values. */
5751 if (src != sets[i].src)
5753 cse_altered = 1;
5754 do_not_record = 0;
5755 hash_arg_in_memory = 0;
5756 sets[i].src = src;
5757 sets[i].src_hash = HASH (src, mode);
5758 sets[i].src_volatile = do_not_record;
5759 sets[i].src_in_memory = hash_arg_in_memory;
5760 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5763 /* If this is a single SET, we are setting a register, and we have an
5764 equivalent constant, we want to add a REG_NOTE. We don't want
5765 to write a REG_EQUAL note for a constant pseudo since verifying that
5766 that pseudo hasn't been eliminated is a pain. Such a note also
5767 won't help anything.
5769 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5770 which can be created for a reference to a compile time computable
5771 entry in a jump table. */
5773 if (n_sets == 1 && src_const && REG_P (dest)
5774 && !REG_P (src_const)
5775 && ! (GET_CODE (src_const) == CONST
5776 && GET_CODE (XEXP (src_const, 0)) == MINUS
5777 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5778 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5780 /* We only want a REG_EQUAL note if src_const != src. */
5781 if (! rtx_equal_p (src, src_const))
5783 /* Make sure that the rtx is not shared. */
5784 src_const = copy_rtx (src_const);
5786 /* Record the actual constant value in a REG_EQUAL note,
5787 making a new one if one does not already exist. */
5788 set_unique_reg_note (insn, REG_EQUAL, src_const);
5792 /* Now deal with the destination. */
5793 do_not_record = 0;
5795 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5796 while (GET_CODE (dest) == SUBREG
5797 || GET_CODE (dest) == ZERO_EXTRACT
5798 || GET_CODE (dest) == STRICT_LOW_PART)
5799 dest = XEXP (dest, 0);
5801 sets[i].inner_dest = dest;
5803 if (MEM_P (dest))
5805 #ifdef PUSH_ROUNDING
5806 /* Stack pushes invalidate the stack pointer. */
5807 rtx addr = XEXP (dest, 0);
5808 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5809 && XEXP (addr, 0) == stack_pointer_rtx)
5810 invalidate (stack_pointer_rtx, VOIDmode);
5811 #endif
5812 dest = fold_rtx (dest, insn);
5815 /* Compute the hash code of the destination now,
5816 before the effects of this instruction are recorded,
5817 since the register values used in the address computation
5818 are those before this instruction. */
5819 sets[i].dest_hash = HASH (dest, mode);
5821 /* Don't enter a bit-field in the hash table
5822 because the value in it after the store
5823 may not equal what was stored, due to truncation. */
5825 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5827 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5829 if (src_const != 0 && GET_CODE (src_const) == CONST_INT
5830 && GET_CODE (width) == CONST_INT
5831 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5832 && ! (INTVAL (src_const)
5833 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5834 /* Exception: if the value is constant,
5835 and it won't be truncated, record it. */
5837 else
5839 /* This is chosen so that the destination will be invalidated
5840 but no new value will be recorded.
5841 We must invalidate because sometimes constant
5842 values can be recorded for bitfields. */
5843 sets[i].src_elt = 0;
5844 sets[i].src_volatile = 1;
5845 src_eqv = 0;
5846 src_eqv_elt = 0;
5850 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5851 the insn. */
5852 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5854 /* One less use of the label this insn used to jump to. */
5855 delete_insn (insn);
5856 cse_jumps_altered = 1;
5857 /* No more processing for this set. */
5858 sets[i].rtl = 0;
5861 /* If this SET is now setting PC to a label, we know it used to
5862 be a conditional or computed branch. */
5863 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5864 && !LABEL_REF_NONLOCAL_P (src))
5866 /* Now emit a BARRIER after the unconditional jump. */
5867 if (NEXT_INSN (insn) == 0
5868 || !BARRIER_P (NEXT_INSN (insn)))
5869 emit_barrier_after (insn);
5871 /* We reemit the jump in as many cases as possible just in
5872 case the form of an unconditional jump is significantly
5873 different than a computed jump or conditional jump.
5875 If this insn has multiple sets, then reemitting the
5876 jump is nontrivial. So instead we just force rerecognition
5877 and hope for the best. */
5878 if (n_sets == 1)
5880 rtx new, note;
5882 new = emit_jump_insn_after (gen_jump (XEXP (src, 0)), insn);
5883 JUMP_LABEL (new) = XEXP (src, 0);
5884 LABEL_NUSES (XEXP (src, 0))++;
5886 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5887 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5888 if (note)
5890 XEXP (note, 1) = NULL_RTX;
5891 REG_NOTES (new) = note;
5894 delete_insn (insn);
5895 insn = new;
5897 /* Now emit a BARRIER after the unconditional jump. */
5898 if (NEXT_INSN (insn) == 0
5899 || !BARRIER_P (NEXT_INSN (insn)))
5900 emit_barrier_after (insn);
5902 else
5903 INSN_CODE (insn) = -1;
5905 /* Do not bother deleting any unreachable code,
5906 let jump/flow do that. */
5908 cse_jumps_altered = 1;
5909 sets[i].rtl = 0;
5912 /* If destination is volatile, invalidate it and then do no further
5913 processing for this assignment. */
5915 else if (do_not_record)
5917 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5918 invalidate (dest, VOIDmode);
5919 else if (MEM_P (dest))
5920 invalidate (dest, VOIDmode);
5921 else if (GET_CODE (dest) == STRICT_LOW_PART
5922 || GET_CODE (dest) == ZERO_EXTRACT)
5923 invalidate (XEXP (dest, 0), GET_MODE (dest));
5924 sets[i].rtl = 0;
5927 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5928 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5930 #ifdef HAVE_cc0
5931 /* If setting CC0, record what it was set to, or a constant, if it
5932 is equivalent to a constant. If it is being set to a floating-point
5933 value, make a COMPARE with the appropriate constant of 0. If we
5934 don't do this, later code can interpret this as a test against
5935 const0_rtx, which can cause problems if we try to put it into an
5936 insn as a floating-point operand. */
5937 if (dest == cc0_rtx)
5939 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5940 this_insn_cc0_mode = mode;
5941 if (FLOAT_MODE_P (mode))
5942 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5943 CONST0_RTX (mode));
5945 #endif
5948 /* Now enter all non-volatile source expressions in the hash table
5949 if they are not already present.
5950 Record their equivalence classes in src_elt.
5951 This way we can insert the corresponding destinations into
5952 the same classes even if the actual sources are no longer in them
5953 (having been invalidated). */
5955 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5956 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5958 struct table_elt *elt;
5959 struct table_elt *classp = sets[0].src_elt;
5960 rtx dest = SET_DEST (sets[0].rtl);
5961 enum machine_mode eqvmode = GET_MODE (dest);
5963 if (GET_CODE (dest) == STRICT_LOW_PART)
5965 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5966 classp = 0;
5968 if (insert_regs (src_eqv, classp, 0))
5970 rehash_using_reg (src_eqv);
5971 src_eqv_hash = HASH (src_eqv, eqvmode);
5973 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5974 elt->in_memory = src_eqv_in_memory;
5975 src_eqv_elt = elt;
5977 /* Check to see if src_eqv_elt is the same as a set source which
5978 does not yet have an elt, and if so set the elt of the set source
5979 to src_eqv_elt. */
5980 for (i = 0; i < n_sets; i++)
5981 if (sets[i].rtl && sets[i].src_elt == 0
5982 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5983 sets[i].src_elt = src_eqv_elt;
5986 for (i = 0; i < n_sets; i++)
5987 if (sets[i].rtl && ! sets[i].src_volatile
5988 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5990 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5992 /* REG_EQUAL in setting a STRICT_LOW_PART
5993 gives an equivalent for the entire destination register,
5994 not just for the subreg being stored in now.
5995 This is a more interesting equivalence, so we arrange later
5996 to treat the entire reg as the destination. */
5997 sets[i].src_elt = src_eqv_elt;
5998 sets[i].src_hash = src_eqv_hash;
6000 else
6002 /* Insert source and constant equivalent into hash table, if not
6003 already present. */
6004 struct table_elt *classp = src_eqv_elt;
6005 rtx src = sets[i].src;
6006 rtx dest = SET_DEST (sets[i].rtl);
6007 enum machine_mode mode
6008 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
6010 /* It's possible that we have a source value known to be
6011 constant but don't have a REG_EQUAL note on the insn.
6012 Lack of a note will mean src_eqv_elt will be NULL. This
6013 can happen where we've generated a SUBREG to access a
6014 CONST_INT that is already in a register in a wider mode.
6015 Ensure that the source expression is put in the proper
6016 constant class. */
6017 if (!classp)
6018 classp = sets[i].src_const_elt;
6020 if (sets[i].src_elt == 0)
6022 /* Don't put a hard register source into the table if this is
6023 the last insn of a libcall. In this case, we only need
6024 to put src_eqv_elt in src_elt. */
6025 if (! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6027 struct table_elt *elt;
6029 /* Note that these insert_regs calls cannot remove
6030 any of the src_elt's, because they would have failed to
6031 match if not still valid. */
6032 if (insert_regs (src, classp, 0))
6034 rehash_using_reg (src);
6035 sets[i].src_hash = HASH (src, mode);
6037 elt = insert (src, classp, sets[i].src_hash, mode);
6038 elt->in_memory = sets[i].src_in_memory;
6039 sets[i].src_elt = classp = elt;
6041 else
6042 sets[i].src_elt = classp;
6044 if (sets[i].src_const && sets[i].src_const_elt == 0
6045 && src != sets[i].src_const
6046 && ! rtx_equal_p (sets[i].src_const, src))
6047 sets[i].src_elt = insert (sets[i].src_const, classp,
6048 sets[i].src_const_hash, mode);
6051 else if (sets[i].src_elt == 0)
6052 /* If we did not insert the source into the hash table (e.g., it was
6053 volatile), note the equivalence class for the REG_EQUAL value, if any,
6054 so that the destination goes into that class. */
6055 sets[i].src_elt = src_eqv_elt;
6057 /* Record destination addresses in the hash table. This allows us to
6058 check if they are invalidated by other sets. */
6059 for (i = 0; i < n_sets; i++)
6061 if (sets[i].rtl)
6063 rtx x = sets[i].inner_dest;
6064 struct table_elt *elt;
6065 enum machine_mode mode;
6066 unsigned hash;
6068 if (MEM_P (x))
6070 x = XEXP (x, 0);
6071 mode = GET_MODE (x);
6072 hash = HASH (x, mode);
6073 elt = lookup (x, hash, mode);
6074 if (!elt)
6076 if (insert_regs (x, NULL, 0))
6078 rehash_using_reg (x);
6079 hash = HASH (x, mode);
6081 elt = insert (x, NULL, hash, mode);
6084 sets[i].dest_addr_elt = elt;
6086 else
6087 sets[i].dest_addr_elt = NULL;
6091 invalidate_from_clobbers (x);
6093 /* Some registers are invalidated by subroutine calls. Memory is
6094 invalidated by non-constant calls. */
6096 if (CALL_P (insn))
6098 if (! CONST_OR_PURE_CALL_P (insn))
6099 invalidate_memory ();
6100 invalidate_for_call ();
6103 /* Now invalidate everything set by this instruction.
6104 If a SUBREG or other funny destination is being set,
6105 sets[i].rtl is still nonzero, so here we invalidate the reg
6106 a part of which is being set. */
6108 for (i = 0; i < n_sets; i++)
6109 if (sets[i].rtl)
6111 /* We can't use the inner dest, because the mode associated with
6112 a ZERO_EXTRACT is significant. */
6113 rtx dest = SET_DEST (sets[i].rtl);
6115 /* Needed for registers to remove the register from its
6116 previous quantity's chain.
6117 Needed for memory if this is a nonvarying address, unless
6118 we have just done an invalidate_memory that covers even those. */
6119 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6120 invalidate (dest, VOIDmode);
6121 else if (MEM_P (dest))
6122 invalidate (dest, VOIDmode);
6123 else if (GET_CODE (dest) == STRICT_LOW_PART
6124 || GET_CODE (dest) == ZERO_EXTRACT)
6125 invalidate (XEXP (dest, 0), GET_MODE (dest));
6128 /* A volatile ASM invalidates everything. */
6129 if (NONJUMP_INSN_P (insn)
6130 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
6131 && MEM_VOLATILE_P (PATTERN (insn)))
6132 flush_hash_table ();
6134 /* Make sure registers mentioned in destinations
6135 are safe for use in an expression to be inserted.
6136 This removes from the hash table
6137 any invalid entry that refers to one of these registers.
6139 We don't care about the return value from mention_regs because
6140 we are going to hash the SET_DEST values unconditionally. */
6142 for (i = 0; i < n_sets; i++)
6144 if (sets[i].rtl)
6146 rtx x = SET_DEST (sets[i].rtl);
6148 if (!REG_P (x))
6149 mention_regs (x);
6150 else
6152 /* We used to rely on all references to a register becoming
6153 inaccessible when a register changes to a new quantity,
6154 since that changes the hash code. However, that is not
6155 safe, since after HASH_SIZE new quantities we get a
6156 hash 'collision' of a register with its own invalid
6157 entries. And since SUBREGs have been changed not to
6158 change their hash code with the hash code of the register,
6159 it wouldn't work any longer at all. So we have to check
6160 for any invalid references lying around now.
6161 This code is similar to the REG case in mention_regs,
6162 but it knows that reg_tick has been incremented, and
6163 it leaves reg_in_table as -1 . */
6164 unsigned int regno = REGNO (x);
6165 unsigned int endregno
6166 = regno + (regno >= FIRST_PSEUDO_REGISTER ? 1
6167 : hard_regno_nregs[regno][GET_MODE (x)]);
6168 unsigned int i;
6170 for (i = regno; i < endregno; i++)
6172 if (REG_IN_TABLE (i) >= 0)
6174 remove_invalid_refs (i);
6175 REG_IN_TABLE (i) = -1;
6182 /* We may have just removed some of the src_elt's from the hash table.
6183 So replace each one with the current head of the same class.
6184 Also check if destination addresses have been removed. */
6186 for (i = 0; i < n_sets; i++)
6187 if (sets[i].rtl)
6189 if (sets[i].dest_addr_elt
6190 && sets[i].dest_addr_elt->first_same_value == 0)
6192 /* The elt was removed, which means this destination is not
6193 valid after this instruction. */
6194 sets[i].rtl = NULL_RTX;
6196 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
6197 /* If elt was removed, find current head of same class,
6198 or 0 if nothing remains of that class. */
6200 struct table_elt *elt = sets[i].src_elt;
6202 while (elt && elt->prev_same_value)
6203 elt = elt->prev_same_value;
6205 while (elt && elt->first_same_value == 0)
6206 elt = elt->next_same_value;
6207 sets[i].src_elt = elt ? elt->first_same_value : 0;
6211 /* Now insert the destinations into their equivalence classes. */
6213 for (i = 0; i < n_sets; i++)
6214 if (sets[i].rtl)
6216 rtx dest = SET_DEST (sets[i].rtl);
6217 struct table_elt *elt;
6219 /* Don't record value if we are not supposed to risk allocating
6220 floating-point values in registers that might be wider than
6221 memory. */
6222 if ((flag_float_store
6223 && MEM_P (dest)
6224 && FLOAT_MODE_P (GET_MODE (dest)))
6225 /* Don't record BLKmode values, because we don't know the
6226 size of it, and can't be sure that other BLKmode values
6227 have the same or smaller size. */
6228 || GET_MODE (dest) == BLKmode
6229 /* Don't record values of destinations set inside a libcall block
6230 since we might delete the libcall. Things should have been set
6231 up so we won't want to reuse such a value, but we play it safe
6232 here. */
6233 || libcall_insn
6234 /* If we didn't put a REG_EQUAL value or a source into the hash
6235 table, there is no point is recording DEST. */
6236 || sets[i].src_elt == 0
6237 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
6238 or SIGN_EXTEND, don't record DEST since it can cause
6239 some tracking to be wrong.
6241 ??? Think about this more later. */
6242 || (GET_CODE (dest) == SUBREG
6243 && (GET_MODE_SIZE (GET_MODE (dest))
6244 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6245 && (GET_CODE (sets[i].src) == SIGN_EXTEND
6246 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
6247 continue;
6249 /* STRICT_LOW_PART isn't part of the value BEING set,
6250 and neither is the SUBREG inside it.
6251 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
6252 if (GET_CODE (dest) == STRICT_LOW_PART)
6253 dest = SUBREG_REG (XEXP (dest, 0));
6255 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
6256 /* Registers must also be inserted into chains for quantities. */
6257 if (insert_regs (dest, sets[i].src_elt, 1))
6259 /* If `insert_regs' changes something, the hash code must be
6260 recalculated. */
6261 rehash_using_reg (dest);
6262 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
6265 elt = insert (dest, sets[i].src_elt,
6266 sets[i].dest_hash, GET_MODE (dest));
6268 elt->in_memory = (MEM_P (sets[i].inner_dest)
6269 && !MEM_READONLY_P (sets[i].inner_dest));
6271 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
6272 narrower than M2, and both M1 and M2 are the same number of words,
6273 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
6274 make that equivalence as well.
6276 However, BAR may have equivalences for which gen_lowpart
6277 will produce a simpler value than gen_lowpart applied to
6278 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
6279 BAR's equivalences. If we don't get a simplified form, make
6280 the SUBREG. It will not be used in an equivalence, but will
6281 cause two similar assignments to be detected.
6283 Note the loop below will find SUBREG_REG (DEST) since we have
6284 already entered SRC and DEST of the SET in the table. */
6286 if (GET_CODE (dest) == SUBREG
6287 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
6288 / UNITS_PER_WORD)
6289 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
6290 && (GET_MODE_SIZE (GET_MODE (dest))
6291 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
6292 && sets[i].src_elt != 0)
6294 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
6295 struct table_elt *elt, *classp = 0;
6297 for (elt = sets[i].src_elt->first_same_value; elt;
6298 elt = elt->next_same_value)
6300 rtx new_src = 0;
6301 unsigned src_hash;
6302 struct table_elt *src_elt;
6303 int byte = 0;
6305 /* Ignore invalid entries. */
6306 if (!REG_P (elt->exp)
6307 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
6308 continue;
6310 /* We may have already been playing subreg games. If the
6311 mode is already correct for the destination, use it. */
6312 if (GET_MODE (elt->exp) == new_mode)
6313 new_src = elt->exp;
6314 else
6316 /* Calculate big endian correction for the SUBREG_BYTE.
6317 We have already checked that M1 (GET_MODE (dest))
6318 is not narrower than M2 (new_mode). */
6319 if (BYTES_BIG_ENDIAN)
6320 byte = (GET_MODE_SIZE (GET_MODE (dest))
6321 - GET_MODE_SIZE (new_mode));
6323 new_src = simplify_gen_subreg (new_mode, elt->exp,
6324 GET_MODE (dest), byte);
6327 /* The call to simplify_gen_subreg fails if the value
6328 is VOIDmode, yet we can't do any simplification, e.g.
6329 for EXPR_LISTs denoting function call results.
6330 It is invalid to construct a SUBREG with a VOIDmode
6331 SUBREG_REG, hence a zero new_src means we can't do
6332 this substitution. */
6333 if (! new_src)
6334 continue;
6336 src_hash = HASH (new_src, new_mode);
6337 src_elt = lookup (new_src, src_hash, new_mode);
6339 /* Put the new source in the hash table is if isn't
6340 already. */
6341 if (src_elt == 0)
6343 if (insert_regs (new_src, classp, 0))
6345 rehash_using_reg (new_src);
6346 src_hash = HASH (new_src, new_mode);
6348 src_elt = insert (new_src, classp, src_hash, new_mode);
6349 src_elt->in_memory = elt->in_memory;
6351 else if (classp && classp != src_elt->first_same_value)
6352 /* Show that two things that we've seen before are
6353 actually the same. */
6354 merge_equiv_classes (src_elt, classp);
6356 classp = src_elt->first_same_value;
6357 /* Ignore invalid entries. */
6358 while (classp
6359 && !REG_P (classp->exp)
6360 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6361 classp = classp->next_same_value;
6366 /* Special handling for (set REG0 REG1) where REG0 is the
6367 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6368 be used in the sequel, so (if easily done) change this insn to
6369 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6370 that computed their value. Then REG1 will become a dead store
6371 and won't cloud the situation for later optimizations.
6373 Do not make this change if REG1 is a hard register, because it will
6374 then be used in the sequel and we may be changing a two-operand insn
6375 into a three-operand insn.
6377 Also do not do this if we are operating on a copy of INSN.
6379 Also don't do this if INSN ends a libcall; this would cause an unrelated
6380 register to be set in the middle of a libcall, and we then get bad code
6381 if the libcall is deleted. */
6383 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
6384 && NEXT_INSN (PREV_INSN (insn)) == insn
6385 && REG_P (SET_SRC (sets[0].rtl))
6386 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
6387 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
6389 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
6390 struct qty_table_elem *src_ent = &qty_table[src_q];
6392 if ((src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
6393 && ! find_reg_note (insn, REG_RETVAL, NULL_RTX))
6395 rtx prev = insn;
6396 /* Scan for the previous nonnote insn, but stop at a basic
6397 block boundary. */
6400 prev = PREV_INSN (prev);
6402 while (prev && NOTE_P (prev)
6403 && NOTE_LINE_NUMBER (prev) != NOTE_INSN_BASIC_BLOCK);
6405 /* Do not swap the registers around if the previous instruction
6406 attaches a REG_EQUIV note to REG1.
6408 ??? It's not entirely clear whether we can transfer a REG_EQUIV
6409 from the pseudo that originally shadowed an incoming argument
6410 to another register. Some uses of REG_EQUIV might rely on it
6411 being attached to REG1 rather than REG2.
6413 This section previously turned the REG_EQUIV into a REG_EQUAL
6414 note. We cannot do that because REG_EQUIV may provide an
6415 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
6417 if (prev != 0 && NONJUMP_INSN_P (prev)
6418 && GET_CODE (PATTERN (prev)) == SET
6419 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
6420 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
6422 rtx dest = SET_DEST (sets[0].rtl);
6423 rtx src = SET_SRC (sets[0].rtl);
6424 rtx note;
6426 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
6427 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
6428 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
6429 apply_change_group ();
6431 /* If INSN has a REG_EQUAL note, and this note mentions
6432 REG0, then we must delete it, because the value in
6433 REG0 has changed. If the note's value is REG1, we must
6434 also delete it because that is now this insn's dest. */
6435 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
6436 if (note != 0
6437 && (reg_mentioned_p (dest, XEXP (note, 0))
6438 || rtx_equal_p (src, XEXP (note, 0))))
6439 remove_note (insn, note);
6444 /* If this is a conditional jump insn, record any known equivalences due to
6445 the condition being tested. */
6447 if (JUMP_P (insn)
6448 && n_sets == 1 && GET_CODE (x) == SET
6449 && GET_CODE (SET_SRC (x)) == IF_THEN_ELSE)
6450 record_jump_equiv (insn, 0);
6452 #ifdef HAVE_cc0
6453 /* If the previous insn set CC0 and this insn no longer references CC0,
6454 delete the previous insn. Here we use the fact that nothing expects CC0
6455 to be valid over an insn, which is true until the final pass. */
6456 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6457 && (tem = single_set (prev_insn)) != 0
6458 && SET_DEST (tem) == cc0_rtx
6459 && ! reg_mentioned_p (cc0_rtx, x))
6460 delete_insn (prev_insn);
6462 prev_insn_cc0 = this_insn_cc0;
6463 prev_insn_cc0_mode = this_insn_cc0_mode;
6464 prev_insn = insn;
6465 #endif
6468 /* Remove from the hash table all expressions that reference memory. */
6470 static void
6471 invalidate_memory (void)
6473 int i;
6474 struct table_elt *p, *next;
6476 for (i = 0; i < HASH_SIZE; i++)
6477 for (p = table[i]; p; p = next)
6479 next = p->next_same_hash;
6480 if (p->in_memory)
6481 remove_from_table (p, i);
6485 /* If ADDR is an address that implicitly affects the stack pointer, return
6486 1 and update the register tables to show the effect. Else, return 0. */
6488 static int
6489 addr_affects_sp_p (rtx addr)
6491 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
6492 && REG_P (XEXP (addr, 0))
6493 && REGNO (XEXP (addr, 0)) == STACK_POINTER_REGNUM)
6495 if (REG_TICK (STACK_POINTER_REGNUM) >= 0)
6497 REG_TICK (STACK_POINTER_REGNUM)++;
6498 /* Is it possible to use a subreg of SP? */
6499 SUBREG_TICKED (STACK_POINTER_REGNUM) = -1;
6502 /* This should be *very* rare. */
6503 if (TEST_HARD_REG_BIT (hard_regs_in_table, STACK_POINTER_REGNUM))
6504 invalidate (stack_pointer_rtx, VOIDmode);
6506 return 1;
6509 return 0;
6512 /* Perform invalidation on the basis of everything about an insn
6513 except for invalidating the actual places that are SET in it.
6514 This includes the places CLOBBERed, and anything that might
6515 alias with something that is SET or CLOBBERed.
6517 X is the pattern of the insn. */
6519 static void
6520 invalidate_from_clobbers (rtx x)
6522 if (GET_CODE (x) == CLOBBER)
6524 rtx ref = XEXP (x, 0);
6525 if (ref)
6527 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6528 || MEM_P (ref))
6529 invalidate (ref, VOIDmode);
6530 else if (GET_CODE (ref) == STRICT_LOW_PART
6531 || GET_CODE (ref) == ZERO_EXTRACT)
6532 invalidate (XEXP (ref, 0), GET_MODE (ref));
6535 else if (GET_CODE (x) == PARALLEL)
6537 int i;
6538 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6540 rtx y = XVECEXP (x, 0, i);
6541 if (GET_CODE (y) == CLOBBER)
6543 rtx ref = XEXP (y, 0);
6544 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6545 || MEM_P (ref))
6546 invalidate (ref, VOIDmode);
6547 else if (GET_CODE (ref) == STRICT_LOW_PART
6548 || GET_CODE (ref) == ZERO_EXTRACT)
6549 invalidate (XEXP (ref, 0), GET_MODE (ref));
6555 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6556 and replace any registers in them with either an equivalent constant
6557 or the canonical form of the register. If we are inside an address,
6558 only do this if the address remains valid.
6560 OBJECT is 0 except when within a MEM in which case it is the MEM.
6562 Return the replacement for X. */
6564 static rtx
6565 cse_process_notes (rtx x, rtx object)
6567 enum rtx_code code = GET_CODE (x);
6568 const char *fmt = GET_RTX_FORMAT (code);
6569 int i;
6571 switch (code)
6573 case CONST_INT:
6574 case CONST:
6575 case SYMBOL_REF:
6576 case LABEL_REF:
6577 case CONST_DOUBLE:
6578 case CONST_VECTOR:
6579 case PC:
6580 case CC0:
6581 case LO_SUM:
6582 return x;
6584 case MEM:
6585 validate_change (x, &XEXP (x, 0),
6586 cse_process_notes (XEXP (x, 0), x), 0);
6587 return x;
6589 case EXPR_LIST:
6590 case INSN_LIST:
6591 if (REG_NOTE_KIND (x) == REG_EQUAL)
6592 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX);
6593 if (XEXP (x, 1))
6594 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX);
6595 return x;
6597 case SIGN_EXTEND:
6598 case ZERO_EXTEND:
6599 case SUBREG:
6601 rtx new = cse_process_notes (XEXP (x, 0), object);
6602 /* We don't substitute VOIDmode constants into these rtx,
6603 since they would impede folding. */
6604 if (GET_MODE (new) != VOIDmode)
6605 validate_change (object, &XEXP (x, 0), new, 0);
6606 return x;
6609 case REG:
6610 i = REG_QTY (REGNO (x));
6612 /* Return a constant or a constant register. */
6613 if (REGNO_QTY_VALID_P (REGNO (x)))
6615 struct qty_table_elem *ent = &qty_table[i];
6617 if (ent->const_rtx != NULL_RTX
6618 && (CONSTANT_P (ent->const_rtx)
6619 || REG_P (ent->const_rtx)))
6621 rtx new = gen_lowpart (GET_MODE (x), ent->const_rtx);
6622 if (new)
6623 return new;
6627 /* Otherwise, canonicalize this register. */
6628 return canon_reg (x, NULL_RTX);
6630 default:
6631 break;
6634 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6635 if (fmt[i] == 'e')
6636 validate_change (object, &XEXP (x, i),
6637 cse_process_notes (XEXP (x, i), object), 0);
6639 return x;
6642 /* Process one SET of an insn that was skipped. We ignore CLOBBERs
6643 since they are done elsewhere. This function is called via note_stores. */
6645 static void
6646 invalidate_skipped_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
6648 enum rtx_code code = GET_CODE (dest);
6650 if (code == MEM
6651 && ! addr_affects_sp_p (dest) /* If this is not a stack push ... */
6652 /* There are times when an address can appear varying and be a PLUS
6653 during this scan when it would be a fixed address were we to know
6654 the proper equivalences. So invalidate all memory if there is
6655 a BLKmode or nonscalar memory reference or a reference to a
6656 variable address. */
6657 && (MEM_IN_STRUCT_P (dest) || GET_MODE (dest) == BLKmode
6658 || cse_rtx_varies_p (XEXP (dest, 0), 0)))
6660 invalidate_memory ();
6661 return;
6664 if (GET_CODE (set) == CLOBBER
6665 || CC0_P (dest)
6666 || dest == pc_rtx)
6667 return;
6669 if (code == STRICT_LOW_PART || code == ZERO_EXTRACT)
6670 invalidate (XEXP (dest, 0), GET_MODE (dest));
6671 else if (code == REG || code == SUBREG || code == MEM)
6672 invalidate (dest, VOIDmode);
6675 /* Invalidate all insns from START up to the end of the function or the
6676 next label. This called when we wish to CSE around a block that is
6677 conditionally executed. */
6679 static void
6680 invalidate_skipped_block (rtx start)
6682 rtx insn;
6684 for (insn = start; insn && !LABEL_P (insn);
6685 insn = NEXT_INSN (insn))
6687 if (! INSN_P (insn))
6688 continue;
6690 if (CALL_P (insn))
6692 if (! CONST_OR_PURE_CALL_P (insn))
6693 invalidate_memory ();
6694 invalidate_for_call ();
6697 invalidate_from_clobbers (PATTERN (insn));
6698 note_stores (PATTERN (insn), invalidate_skipped_set, NULL);
6702 /* Find the end of INSN's basic block and return its range,
6703 the total number of SETs in all the insns of the block, the last insn of the
6704 block, and the branch path.
6706 The branch path indicates which branches should be followed. If a nonzero
6707 path size is specified, the block should be rescanned and a different set
6708 of branches will be taken. The branch path is only used if
6709 FLAG_CSE_FOLLOW_JUMPS or FLAG_CSE_SKIP_BLOCKS is nonzero.
6711 DATA is a pointer to a struct cse_basic_block_data, defined below, that is
6712 used to describe the block. It is filled in with the information about
6713 the current block. The incoming structure's branch path, if any, is used
6714 to construct the output branch path. */
6716 static void
6717 cse_end_of_basic_block (rtx insn, struct cse_basic_block_data *data,
6718 int follow_jumps, int skip_blocks)
6720 rtx p = insn, q;
6721 int nsets = 0;
6722 int low_cuid = INSN_CUID (insn), high_cuid = INSN_CUID (insn);
6723 rtx next = INSN_P (insn) ? insn : next_real_insn (insn);
6724 int path_size = data->path_size;
6725 int path_entry = 0;
6726 int i;
6728 /* Update the previous branch path, if any. If the last branch was
6729 previously PATH_TAKEN, mark it PATH_NOT_TAKEN.
6730 If it was previously PATH_NOT_TAKEN,
6731 shorten the path by one and look at the previous branch. We know that
6732 at least one branch must have been taken if PATH_SIZE is nonzero. */
6733 while (path_size > 0)
6735 if (data->path[path_size - 1].status != PATH_NOT_TAKEN)
6737 data->path[path_size - 1].status = PATH_NOT_TAKEN;
6738 break;
6740 else
6741 path_size--;
6744 /* If the first instruction is marked with QImode, that means we've
6745 already processed this block. Our caller will look at DATA->LAST
6746 to figure out where to go next. We want to return the next block
6747 in the instruction stream, not some branched-to block somewhere
6748 else. We accomplish this by pretending our called forbid us to
6749 follow jumps, or skip blocks. */
6750 if (GET_MODE (insn) == QImode)
6751 follow_jumps = skip_blocks = 0;
6753 /* Scan to end of this basic block. */
6754 while (p && !LABEL_P (p))
6756 /* Don't cse over a call to setjmp; on some machines (eg VAX)
6757 the regs restored by the longjmp come from
6758 a later time than the setjmp. */
6759 if (PREV_INSN (p) && CALL_P (PREV_INSN (p))
6760 && find_reg_note (PREV_INSN (p), REG_SETJMP, NULL))
6761 break;
6763 /* A PARALLEL can have lots of SETs in it,
6764 especially if it is really an ASM_OPERANDS. */
6765 if (INSN_P (p) && GET_CODE (PATTERN (p)) == PARALLEL)
6766 nsets += XVECLEN (PATTERN (p), 0);
6767 else if (!NOTE_P (p))
6768 nsets += 1;
6770 /* Ignore insns made by CSE; they cannot affect the boundaries of
6771 the basic block. */
6773 if (INSN_UID (p) <= max_uid && INSN_CUID (p) > high_cuid)
6774 high_cuid = INSN_CUID (p);
6775 if (INSN_UID (p) <= max_uid && INSN_CUID (p) < low_cuid)
6776 low_cuid = INSN_CUID (p);
6778 /* See if this insn is in our branch path. If it is and we are to
6779 take it, do so. */
6780 if (path_entry < path_size && data->path[path_entry].branch == p)
6782 if (data->path[path_entry].status != PATH_NOT_TAKEN)
6783 p = JUMP_LABEL (p);
6785 /* Point to next entry in path, if any. */
6786 path_entry++;
6789 /* If this is a conditional jump, we can follow it if -fcse-follow-jumps
6790 was specified, we haven't reached our maximum path length, there are
6791 insns following the target of the jump, this is the only use of the
6792 jump label, and the target label is preceded by a BARRIER.
6794 Alternatively, we can follow the jump if it branches around a
6795 block of code and there are no other branches into the block.
6796 In this case invalidate_skipped_block will be called to invalidate any
6797 registers set in the block when following the jump. */
6799 else if ((follow_jumps || skip_blocks) && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH) - 1
6800 && JUMP_P (p)
6801 && GET_CODE (PATTERN (p)) == SET
6802 && GET_CODE (SET_SRC (PATTERN (p))) == IF_THEN_ELSE
6803 && JUMP_LABEL (p) != 0
6804 && LABEL_NUSES (JUMP_LABEL (p)) == 1
6805 && NEXT_INSN (JUMP_LABEL (p)) != 0)
6807 for (q = PREV_INSN (JUMP_LABEL (p)); q; q = PREV_INSN (q))
6808 if ((!NOTE_P (q)
6809 || (PREV_INSN (q) && CALL_P (PREV_INSN (q))
6810 && find_reg_note (PREV_INSN (q), REG_SETJMP, NULL)))
6811 && (!LABEL_P (q) || LABEL_NUSES (q) != 0))
6812 break;
6814 /* If we ran into a BARRIER, this code is an extension of the
6815 basic block when the branch is taken. */
6816 if (follow_jumps && q != 0 && BARRIER_P (q))
6818 /* Don't allow ourself to keep walking around an
6819 always-executed loop. */
6820 if (next_real_insn (q) == next)
6822 p = NEXT_INSN (p);
6823 continue;
6826 /* Similarly, don't put a branch in our path more than once. */
6827 for (i = 0; i < path_entry; i++)
6828 if (data->path[i].branch == p)
6829 break;
6831 if (i != path_entry)
6832 break;
6834 data->path[path_entry].branch = p;
6835 data->path[path_entry++].status = PATH_TAKEN;
6837 /* This branch now ends our path. It was possible that we
6838 didn't see this branch the last time around (when the
6839 insn in front of the target was a JUMP_INSN that was
6840 turned into a no-op). */
6841 path_size = path_entry;
6843 p = JUMP_LABEL (p);
6844 /* Mark block so we won't scan it again later. */
6845 PUT_MODE (NEXT_INSN (p), QImode);
6847 /* Detect a branch around a block of code. */
6848 else if (skip_blocks && q != 0 && !LABEL_P (q))
6850 rtx tmp;
6852 if (next_real_insn (q) == next)
6854 p = NEXT_INSN (p);
6855 continue;
6858 for (i = 0; i < path_entry; i++)
6859 if (data->path[i].branch == p)
6860 break;
6862 if (i != path_entry)
6863 break;
6865 /* This is no_labels_between_p (p, q) with an added check for
6866 reaching the end of a function (in case Q precedes P). */
6867 for (tmp = NEXT_INSN (p); tmp && tmp != q; tmp = NEXT_INSN (tmp))
6868 if (LABEL_P (tmp))
6869 break;
6871 if (tmp == q)
6873 data->path[path_entry].branch = p;
6874 data->path[path_entry++].status = PATH_AROUND;
6876 path_size = path_entry;
6878 p = JUMP_LABEL (p);
6879 /* Mark block so we won't scan it again later. */
6880 PUT_MODE (NEXT_INSN (p), QImode);
6884 p = NEXT_INSN (p);
6887 data->low_cuid = low_cuid;
6888 data->high_cuid = high_cuid;
6889 data->nsets = nsets;
6890 data->last = p;
6892 /* If all jumps in the path are not taken, set our path length to zero
6893 so a rescan won't be done. */
6894 for (i = path_size - 1; i >= 0; i--)
6895 if (data->path[i].status != PATH_NOT_TAKEN)
6896 break;
6898 if (i == -1)
6899 data->path_size = 0;
6900 else
6901 data->path_size = path_size;
6903 /* End the current branch path. */
6904 data->path[path_size].branch = 0;
6907 /* Perform cse on the instructions of a function.
6908 F is the first instruction.
6909 NREGS is one plus the highest pseudo-reg number used in the instruction.
6911 Returns 1 if jump_optimize should be redone due to simplifications
6912 in conditional jump instructions. */
6915 cse_main (rtx f, int nregs)
6917 struct cse_basic_block_data val;
6918 rtx insn = f;
6919 int i;
6921 init_cse_reg_info (nregs);
6923 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6925 cse_jumps_altered = 0;
6926 recorded_label_ref = 0;
6927 constant_pool_entries_cost = 0;
6928 constant_pool_entries_regcost = 0;
6929 val.path_size = 0;
6930 rtl_hooks = cse_rtl_hooks;
6932 init_recog ();
6933 init_alias_analysis ();
6935 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6937 /* Find the largest uid. */
6939 max_uid = get_max_uid ();
6940 uid_cuid = XCNEWVEC (int, max_uid + 1);
6942 /* Compute the mapping from uids to cuids.
6943 CUIDs are numbers assigned to insns, like uids,
6944 except that cuids increase monotonically through the code.
6945 Don't assign cuids to line-number NOTEs, so that the distance in cuids
6946 between two insns is not affected by -g. */
6948 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
6950 if (!NOTE_P (insn)
6951 || NOTE_LINE_NUMBER (insn) < 0)
6952 INSN_CUID (insn) = ++i;
6953 else
6954 /* Give a line number note the same cuid as preceding insn. */
6955 INSN_CUID (insn) = i;
6958 /* Loop over basic blocks.
6959 Compute the maximum number of qty's needed for each basic block
6960 (which is 2 for each SET). */
6961 insn = f;
6962 while (insn)
6964 cse_altered = 0;
6965 cse_end_of_basic_block (insn, &val, flag_cse_follow_jumps,
6966 flag_cse_skip_blocks);
6968 /* If this basic block was already processed or has no sets, skip it. */
6969 if (val.nsets == 0 || GET_MODE (insn) == QImode)
6971 PUT_MODE (insn, VOIDmode);
6972 insn = (val.last ? NEXT_INSN (val.last) : 0);
6973 val.path_size = 0;
6974 continue;
6977 cse_basic_block_start = val.low_cuid;
6978 cse_basic_block_end = val.high_cuid;
6979 max_qty = val.nsets * 2;
6981 if (dump_file)
6982 fprintf (dump_file, ";; Processing block from %d to %d, %d sets.\n",
6983 INSN_UID (insn), val.last ? INSN_UID (val.last) : 0,
6984 val.nsets);
6986 /* Make MAX_QTY bigger to give us room to optimize
6987 past the end of this basic block, if that should prove useful. */
6988 if (max_qty < 500)
6989 max_qty = 500;
6991 /* If this basic block is being extended by following certain jumps,
6992 (see `cse_end_of_basic_block'), we reprocess the code from the start.
6993 Otherwise, we start after this basic block. */
6994 if (val.path_size > 0)
6995 cse_basic_block (insn, val.last, val.path);
6996 else
6998 int old_cse_jumps_altered = cse_jumps_altered;
6999 rtx temp;
7001 /* When cse changes a conditional jump to an unconditional
7002 jump, we want to reprocess the block, since it will give
7003 us a new branch path to investigate. */
7004 cse_jumps_altered = 0;
7005 temp = cse_basic_block (insn, val.last, val.path);
7006 if (cse_jumps_altered == 0
7007 || (flag_cse_follow_jumps == 0 && flag_cse_skip_blocks == 0))
7008 insn = temp;
7010 cse_jumps_altered |= old_cse_jumps_altered;
7013 if (cse_altered)
7014 ggc_collect ();
7016 #ifdef USE_C_ALLOCA
7017 alloca (0);
7018 #endif
7021 /* Clean up. */
7022 end_alias_analysis ();
7023 free (uid_cuid);
7024 free (reg_eqv_table);
7025 free (val.path);
7026 rtl_hooks = general_rtl_hooks;
7028 return cse_jumps_altered || recorded_label_ref;
7031 /* Process a single basic block. FROM and TO and the limits of the basic
7032 block. NEXT_BRANCH points to the branch path when following jumps or
7033 a null path when not following jumps. */
7035 static rtx
7036 cse_basic_block (rtx from, rtx to, struct branch_path *next_branch)
7038 rtx insn;
7039 int to_usage = 0;
7040 rtx libcall_insn = NULL_RTX;
7041 int num_insns = 0;
7042 int no_conflict = 0;
7044 /* Allocate the space needed by qty_table. */
7045 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
7047 new_basic_block ();
7049 /* TO might be a label. If so, protect it from being deleted. */
7050 if (to != 0 && LABEL_P (to))
7051 ++LABEL_NUSES (to);
7053 for (insn = from; insn != to; insn = NEXT_INSN (insn))
7055 enum rtx_code code = GET_CODE (insn);
7057 /* If we have processed 1,000 insns, flush the hash table to
7058 avoid extreme quadratic behavior. We must not include NOTEs
7059 in the count since there may be more of them when generating
7060 debugging information. If we clear the table at different
7061 times, code generated with -g -O might be different than code
7062 generated with -O but not -g.
7064 ??? This is a real kludge and needs to be done some other way.
7065 Perhaps for 2.9. */
7066 if (code != NOTE && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
7068 flush_hash_table ();
7069 num_insns = 0;
7072 /* See if this is a branch that is part of the path. If so, and it is
7073 to be taken, do so. */
7074 if (next_branch->branch == insn)
7076 enum taken status = next_branch++->status;
7077 if (status != PATH_NOT_TAKEN)
7079 if (status == PATH_TAKEN)
7080 record_jump_equiv (insn, 1);
7081 else
7082 invalidate_skipped_block (NEXT_INSN (insn));
7084 /* Set the last insn as the jump insn; it doesn't affect cc0.
7085 Then follow this branch. */
7086 #ifdef HAVE_cc0
7087 prev_insn_cc0 = 0;
7088 prev_insn = insn;
7089 #endif
7090 insn = JUMP_LABEL (insn);
7091 continue;
7095 if (GET_MODE (insn) == QImode)
7096 PUT_MODE (insn, VOIDmode);
7098 if (GET_RTX_CLASS (code) == RTX_INSN)
7100 rtx p;
7102 /* Process notes first so we have all notes in canonical forms when
7103 looking for duplicate operations. */
7105 if (REG_NOTES (insn))
7106 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn), NULL_RTX);
7108 /* Track when we are inside in LIBCALL block. Inside such a block,
7109 we do not want to record destinations. The last insn of a
7110 LIBCALL block is not considered to be part of the block, since
7111 its destination is the result of the block and hence should be
7112 recorded. */
7114 if (REG_NOTES (insn) != 0)
7116 if ((p = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
7117 libcall_insn = XEXP (p, 0);
7118 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7120 /* Keep libcall_insn for the last SET insn of a no-conflict
7121 block to prevent changing the destination. */
7122 if (! no_conflict)
7123 libcall_insn = 0;
7124 else
7125 no_conflict = -1;
7127 else if (find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
7128 no_conflict = 1;
7131 cse_insn (insn, libcall_insn);
7133 if (no_conflict == -1)
7135 libcall_insn = 0;
7136 no_conflict = 0;
7139 /* If we haven't already found an insn where we added a LABEL_REF,
7140 check this one. */
7141 if (NONJUMP_INSN_P (insn) && ! recorded_label_ref
7142 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
7143 (void *) insn))
7144 recorded_label_ref = 1;
7147 /* If INSN is now an unconditional jump, skip to the end of our
7148 basic block by pretending that we just did the last insn in the
7149 basic block. If we are jumping to the end of our block, show
7150 that we can have one usage of TO. */
7152 if (any_uncondjump_p (insn))
7154 if (to == 0)
7156 free (qty_table);
7157 return 0;
7160 if (JUMP_LABEL (insn) == to)
7161 to_usage = 1;
7163 /* Maybe TO was deleted because the jump is unconditional.
7164 If so, there is nothing left in this basic block. */
7165 /* ??? Perhaps it would be smarter to set TO
7166 to whatever follows this insn,
7167 and pretend the basic block had always ended here. */
7168 if (INSN_DELETED_P (to))
7169 break;
7171 insn = PREV_INSN (to);
7174 /* See if it is ok to keep on going past the label
7175 which used to end our basic block. Remember that we incremented
7176 the count of that label, so we decrement it here. If we made
7177 a jump unconditional, TO_USAGE will be one; in that case, we don't
7178 want to count the use in that jump. */
7180 if (to != 0 && NEXT_INSN (insn) == to
7181 && LABEL_P (to) && --LABEL_NUSES (to) == to_usage)
7183 struct cse_basic_block_data val;
7184 rtx prev;
7186 insn = NEXT_INSN (to);
7188 /* If TO was the last insn in the function, we are done. */
7189 if (insn == 0)
7191 free (qty_table);
7192 return 0;
7195 /* If TO was preceded by a BARRIER we are done with this block
7196 because it has no continuation. */
7197 prev = prev_nonnote_insn (to);
7198 if (prev && BARRIER_P (prev))
7200 free (qty_table);
7201 return insn;
7204 /* Find the end of the following block. Note that we won't be
7205 following branches in this case. */
7206 to_usage = 0;
7207 val.path_size = 0;
7208 val.path = XNEWVEC (struct branch_path, PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
7209 cse_end_of_basic_block (insn, &val, 0, 0);
7210 free (val.path);
7212 /* If the tables we allocated have enough space left
7213 to handle all the SETs in the next basic block,
7214 continue through it. Otherwise, return,
7215 and that block will be scanned individually. */
7216 if (val.nsets * 2 + next_qty > max_qty)
7217 break;
7219 cse_basic_block_start = val.low_cuid;
7220 cse_basic_block_end = val.high_cuid;
7221 to = val.last;
7223 /* Prevent TO from being deleted if it is a label. */
7224 if (to != 0 && LABEL_P (to))
7225 ++LABEL_NUSES (to);
7227 /* Back up so we process the first insn in the extension. */
7228 insn = PREV_INSN (insn);
7232 gcc_assert (next_qty <= max_qty);
7234 free (qty_table);
7236 return to ? NEXT_INSN (to) : 0;
7239 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for which
7240 there isn't a REG_LABEL note. Return one if so. DATA is the insn. */
7242 static int
7243 check_for_label_ref (rtx *rtl, void *data)
7245 rtx insn = (rtx) data;
7247 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL note for it,
7248 we must rerun jump since it needs to place the note. If this is a
7249 LABEL_REF for a CODE_LABEL that isn't in the insn chain, don't do this
7250 since no REG_LABEL will be added. */
7251 return (GET_CODE (*rtl) == LABEL_REF
7252 && ! LABEL_REF_NONLOCAL_P (*rtl)
7253 && LABEL_P (XEXP (*rtl, 0))
7254 && INSN_UID (XEXP (*rtl, 0)) != 0
7255 && ! find_reg_note (insn, REG_LABEL, XEXP (*rtl, 0)));
7258 /* Count the number of times registers are used (not set) in X.
7259 COUNTS is an array in which we accumulate the count, INCR is how much
7260 we count each register usage.
7262 Don't count a usage of DEST, which is the SET_DEST of a SET which
7263 contains X in its SET_SRC. This is because such a SET does not
7264 modify the liveness of DEST.
7265 DEST is set to pc_rtx for a trapping insn, which means that we must count
7266 uses of a SET_DEST regardless because the insn can't be deleted here. */
7268 static void
7269 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
7271 enum rtx_code code;
7272 rtx note;
7273 const char *fmt;
7274 int i, j;
7276 if (x == 0)
7277 return;
7279 switch (code = GET_CODE (x))
7281 case REG:
7282 if (x != dest)
7283 counts[REGNO (x)] += incr;
7284 return;
7286 case PC:
7287 case CC0:
7288 case CONST:
7289 case CONST_INT:
7290 case CONST_DOUBLE:
7291 case CONST_VECTOR:
7292 case SYMBOL_REF:
7293 case LABEL_REF:
7294 return;
7296 case CLOBBER:
7297 /* If we are clobbering a MEM, mark any registers inside the address
7298 as being used. */
7299 if (MEM_P (XEXP (x, 0)))
7300 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
7301 return;
7303 case SET:
7304 /* Unless we are setting a REG, count everything in SET_DEST. */
7305 if (!REG_P (SET_DEST (x)))
7306 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
7307 count_reg_usage (SET_SRC (x), counts,
7308 dest ? dest : SET_DEST (x),
7309 incr);
7310 return;
7312 case CALL_INSN:
7313 case INSN:
7314 case JUMP_INSN:
7315 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
7316 this fact by setting DEST to pc_rtx. */
7317 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
7318 dest = pc_rtx;
7319 if (code == CALL_INSN)
7320 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
7321 count_reg_usage (PATTERN (x), counts, dest, incr);
7323 /* Things used in a REG_EQUAL note aren't dead since loop may try to
7324 use them. */
7326 note = find_reg_equal_equiv_note (x);
7327 if (note)
7329 rtx eqv = XEXP (note, 0);
7331 if (GET_CODE (eqv) == EXPR_LIST)
7332 /* This REG_EQUAL note describes the result of a function call.
7333 Process all the arguments. */
7336 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
7337 eqv = XEXP (eqv, 1);
7339 while (eqv && GET_CODE (eqv) == EXPR_LIST);
7340 else
7341 count_reg_usage (eqv, counts, dest, incr);
7343 return;
7345 case EXPR_LIST:
7346 if (REG_NOTE_KIND (x) == REG_EQUAL
7347 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
7348 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
7349 involving registers in the address. */
7350 || GET_CODE (XEXP (x, 0)) == CLOBBER)
7351 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
7353 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
7354 return;
7356 case ASM_OPERANDS:
7357 /* If the asm is volatile, then this insn cannot be deleted,
7358 and so the inputs *must* be live. */
7359 if (MEM_VOLATILE_P (x))
7360 dest = NULL_RTX;
7361 /* Iterate over just the inputs, not the constraints as well. */
7362 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
7363 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
7364 return;
7366 case INSN_LIST:
7367 gcc_unreachable ();
7369 default:
7370 break;
7373 fmt = GET_RTX_FORMAT (code);
7374 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7376 if (fmt[i] == 'e')
7377 count_reg_usage (XEXP (x, i), counts, dest, incr);
7378 else if (fmt[i] == 'E')
7379 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7380 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
7384 /* Return true if set is live. */
7385 static bool
7386 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
7387 int *counts)
7389 #ifdef HAVE_cc0
7390 rtx tem;
7391 #endif
7393 if (set_noop_p (set))
7396 #ifdef HAVE_cc0
7397 else if (GET_CODE (SET_DEST (set)) == CC0
7398 && !side_effects_p (SET_SRC (set))
7399 && ((tem = next_nonnote_insn (insn)) == 0
7400 || !INSN_P (tem)
7401 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
7402 return false;
7403 #endif
7404 else if (!REG_P (SET_DEST (set))
7405 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
7406 || counts[REGNO (SET_DEST (set))] != 0
7407 || side_effects_p (SET_SRC (set)))
7408 return true;
7409 return false;
7412 /* Return true if insn is live. */
7414 static bool
7415 insn_live_p (rtx insn, int *counts)
7417 int i;
7418 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
7419 return true;
7420 else if (GET_CODE (PATTERN (insn)) == SET)
7421 return set_live_p (PATTERN (insn), insn, counts);
7422 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
7424 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
7426 rtx elt = XVECEXP (PATTERN (insn), 0, i);
7428 if (GET_CODE (elt) == SET)
7430 if (set_live_p (elt, insn, counts))
7431 return true;
7433 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
7434 return true;
7436 return false;
7438 else
7439 return true;
7442 /* Return true if libcall is dead as a whole. */
7444 static bool
7445 dead_libcall_p (rtx insn, int *counts)
7447 rtx note, set, new;
7449 /* See if there's a REG_EQUAL note on this insn and try to
7450 replace the source with the REG_EQUAL expression.
7452 We assume that insns with REG_RETVALs can only be reg->reg
7453 copies at this point. */
7454 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
7455 if (!note)
7456 return false;
7458 set = single_set (insn);
7459 if (!set)
7460 return false;
7462 new = simplify_rtx (XEXP (note, 0));
7463 if (!new)
7464 new = XEXP (note, 0);
7466 /* While changing insn, we must update the counts accordingly. */
7467 count_reg_usage (insn, counts, NULL_RTX, -1);
7469 if (validate_change (insn, &SET_SRC (set), new, 0))
7471 count_reg_usage (insn, counts, NULL_RTX, 1);
7472 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7473 remove_note (insn, note);
7474 return true;
7477 if (CONSTANT_P (new))
7479 new = force_const_mem (GET_MODE (SET_DEST (set)), new);
7480 if (new && validate_change (insn, &SET_SRC (set), new, 0))
7482 count_reg_usage (insn, counts, NULL_RTX, 1);
7483 remove_note (insn, find_reg_note (insn, REG_RETVAL, NULL_RTX));
7484 remove_note (insn, note);
7485 return true;
7489 count_reg_usage (insn, counts, NULL_RTX, 1);
7490 return false;
7493 /* Scan all the insns and delete any that are dead; i.e., they store a register
7494 that is never used or they copy a register to itself.
7496 This is used to remove insns made obviously dead by cse, loop or other
7497 optimizations. It improves the heuristics in loop since it won't try to
7498 move dead invariants out of loops or make givs for dead quantities. The
7499 remaining passes of the compilation are also sped up. */
7502 delete_trivially_dead_insns (rtx insns, int nreg)
7504 int *counts;
7505 rtx insn, prev;
7506 int in_libcall = 0, dead_libcall = 0;
7507 int ndead = 0;
7509 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7510 /* First count the number of times each register is used. */
7511 counts = XCNEWVEC (int, nreg);
7512 for (insn = insns; insn; insn = NEXT_INSN (insn))
7513 if (INSN_P (insn))
7514 count_reg_usage (insn, counts, NULL_RTX, 1);
7516 /* Go from the last insn to the first and delete insns that only set unused
7517 registers or copy a register to itself. As we delete an insn, remove
7518 usage counts for registers it uses.
7520 The first jump optimization pass may leave a real insn as the last
7521 insn in the function. We must not skip that insn or we may end
7522 up deleting code that is not really dead. */
7523 for (insn = get_last_insn (); insn; insn = prev)
7525 int live_insn = 0;
7527 prev = PREV_INSN (insn);
7528 if (!INSN_P (insn))
7529 continue;
7531 /* Don't delete any insns that are part of a libcall block unless
7532 we can delete the whole libcall block.
7534 Flow or loop might get confused if we did that. Remember
7535 that we are scanning backwards. */
7536 if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
7538 in_libcall = 1;
7539 live_insn = 1;
7540 dead_libcall = dead_libcall_p (insn, counts);
7542 else if (in_libcall)
7543 live_insn = ! dead_libcall;
7544 else
7545 live_insn = insn_live_p (insn, counts);
7547 /* If this is a dead insn, delete it and show registers in it aren't
7548 being used. */
7550 if (! live_insn)
7552 count_reg_usage (insn, counts, NULL_RTX, -1);
7553 delete_insn_and_edges (insn);
7554 ndead++;
7557 if (in_libcall && find_reg_note (insn, REG_LIBCALL, NULL_RTX))
7559 in_libcall = 0;
7560 dead_libcall = 0;
7564 if (dump_file && ndead)
7565 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7566 ndead);
7567 /* Clean up. */
7568 free (counts);
7569 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7570 return ndead;
7573 /* This function is called via for_each_rtx. The argument, NEWREG, is
7574 a condition code register with the desired mode. If we are looking
7575 at the same register in a different mode, replace it with
7576 NEWREG. */
7578 static int
7579 cse_change_cc_mode (rtx *loc, void *data)
7581 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7583 if (*loc
7584 && REG_P (*loc)
7585 && REGNO (*loc) == REGNO (args->newreg)
7586 && GET_MODE (*loc) != GET_MODE (args->newreg))
7588 validate_change (args->insn, loc, args->newreg, 1);
7590 return -1;
7592 return 0;
7595 /* Change the mode of any reference to the register REGNO (NEWREG) to
7596 GET_MODE (NEWREG) in INSN. */
7598 static void
7599 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7601 struct change_cc_mode_args args;
7602 int success;
7604 if (!INSN_P (insn))
7605 return;
7607 args.insn = insn;
7608 args.newreg = newreg;
7610 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7611 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7613 /* If the following assertion was triggered, there is most probably
7614 something wrong with the cc_modes_compatible back end function.
7615 CC modes only can be considered compatible if the insn - with the mode
7616 replaced by any of the compatible modes - can still be recognized. */
7617 success = apply_change_group ();
7618 gcc_assert (success);
7621 /* Change the mode of any reference to the register REGNO (NEWREG) to
7622 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7623 any instruction which modifies NEWREG. */
7625 static void
7626 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7628 rtx insn;
7630 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7632 if (! INSN_P (insn))
7633 continue;
7635 if (reg_set_p (newreg, insn))
7636 return;
7638 cse_change_cc_mode_insn (insn, newreg);
7642 /* BB is a basic block which finishes with CC_REG as a condition code
7643 register which is set to CC_SRC. Look through the successors of BB
7644 to find blocks which have a single predecessor (i.e., this one),
7645 and look through those blocks for an assignment to CC_REG which is
7646 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7647 permitted to change the mode of CC_SRC to a compatible mode. This
7648 returns VOIDmode if no equivalent assignments were found.
7649 Otherwise it returns the mode which CC_SRC should wind up with.
7651 The main complexity in this function is handling the mode issues.
7652 We may have more than one duplicate which we can eliminate, and we
7653 try to find a mode which will work for multiple duplicates. */
7655 static enum machine_mode
7656 cse_cc_succs (basic_block bb, rtx cc_reg, rtx cc_src, bool can_change_mode)
7658 bool found_equiv;
7659 enum machine_mode mode;
7660 unsigned int insn_count;
7661 edge e;
7662 rtx insns[2];
7663 enum machine_mode modes[2];
7664 rtx last_insns[2];
7665 unsigned int i;
7666 rtx newreg;
7667 edge_iterator ei;
7669 /* We expect to have two successors. Look at both before picking
7670 the final mode for the comparison. If we have more successors
7671 (i.e., some sort of table jump, although that seems unlikely),
7672 then we require all beyond the first two to use the same
7673 mode. */
7675 found_equiv = false;
7676 mode = GET_MODE (cc_src);
7677 insn_count = 0;
7678 FOR_EACH_EDGE (e, ei, bb->succs)
7680 rtx insn;
7681 rtx end;
7683 if (e->flags & EDGE_COMPLEX)
7684 continue;
7686 if (EDGE_COUNT (e->dest->preds) != 1
7687 || e->dest == EXIT_BLOCK_PTR)
7688 continue;
7690 end = NEXT_INSN (BB_END (e->dest));
7691 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7693 rtx set;
7695 if (! INSN_P (insn))
7696 continue;
7698 /* If CC_SRC is modified, we have to stop looking for
7699 something which uses it. */
7700 if (modified_in_p (cc_src, insn))
7701 break;
7703 /* Check whether INSN sets CC_REG to CC_SRC. */
7704 set = single_set (insn);
7705 if (set
7706 && REG_P (SET_DEST (set))
7707 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7709 bool found;
7710 enum machine_mode set_mode;
7711 enum machine_mode comp_mode;
7713 found = false;
7714 set_mode = GET_MODE (SET_SRC (set));
7715 comp_mode = set_mode;
7716 if (rtx_equal_p (cc_src, SET_SRC (set)))
7717 found = true;
7718 else if (GET_CODE (cc_src) == COMPARE
7719 && GET_CODE (SET_SRC (set)) == COMPARE
7720 && mode != set_mode
7721 && rtx_equal_p (XEXP (cc_src, 0),
7722 XEXP (SET_SRC (set), 0))
7723 && rtx_equal_p (XEXP (cc_src, 1),
7724 XEXP (SET_SRC (set), 1)))
7727 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7728 if (comp_mode != VOIDmode
7729 && (can_change_mode || comp_mode == mode))
7730 found = true;
7733 if (found)
7735 found_equiv = true;
7736 if (insn_count < ARRAY_SIZE (insns))
7738 insns[insn_count] = insn;
7739 modes[insn_count] = set_mode;
7740 last_insns[insn_count] = end;
7741 ++insn_count;
7743 if (mode != comp_mode)
7745 gcc_assert (can_change_mode);
7746 mode = comp_mode;
7748 /* The modified insn will be re-recognized later. */
7749 PUT_MODE (cc_src, mode);
7752 else
7754 if (set_mode != mode)
7756 /* We found a matching expression in the
7757 wrong mode, but we don't have room to
7758 store it in the array. Punt. This case
7759 should be rare. */
7760 break;
7762 /* INSN sets CC_REG to a value equal to CC_SRC
7763 with the right mode. We can simply delete
7764 it. */
7765 delete_insn (insn);
7768 /* We found an instruction to delete. Keep looking,
7769 in the hopes of finding a three-way jump. */
7770 continue;
7773 /* We found an instruction which sets the condition
7774 code, so don't look any farther. */
7775 break;
7778 /* If INSN sets CC_REG in some other way, don't look any
7779 farther. */
7780 if (reg_set_p (cc_reg, insn))
7781 break;
7784 /* If we fell off the bottom of the block, we can keep looking
7785 through successors. We pass CAN_CHANGE_MODE as false because
7786 we aren't prepared to handle compatibility between the
7787 further blocks and this block. */
7788 if (insn == end)
7790 enum machine_mode submode;
7792 submode = cse_cc_succs (e->dest, cc_reg, cc_src, false);
7793 if (submode != VOIDmode)
7795 gcc_assert (submode == mode);
7796 found_equiv = true;
7797 can_change_mode = false;
7802 if (! found_equiv)
7803 return VOIDmode;
7805 /* Now INSN_COUNT is the number of instructions we found which set
7806 CC_REG to a value equivalent to CC_SRC. The instructions are in
7807 INSNS. The modes used by those instructions are in MODES. */
7809 newreg = NULL_RTX;
7810 for (i = 0; i < insn_count; ++i)
7812 if (modes[i] != mode)
7814 /* We need to change the mode of CC_REG in INSNS[i] and
7815 subsequent instructions. */
7816 if (! newreg)
7818 if (GET_MODE (cc_reg) == mode)
7819 newreg = cc_reg;
7820 else
7821 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7823 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7824 newreg);
7827 delete_insn (insns[i]);
7830 return mode;
7833 /* If we have a fixed condition code register (or two), walk through
7834 the instructions and try to eliminate duplicate assignments. */
7836 static void
7837 cse_condition_code_reg (void)
7839 unsigned int cc_regno_1;
7840 unsigned int cc_regno_2;
7841 rtx cc_reg_1;
7842 rtx cc_reg_2;
7843 basic_block bb;
7845 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7846 return;
7848 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7849 if (cc_regno_2 != INVALID_REGNUM)
7850 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7851 else
7852 cc_reg_2 = NULL_RTX;
7854 FOR_EACH_BB (bb)
7856 rtx last_insn;
7857 rtx cc_reg;
7858 rtx insn;
7859 rtx cc_src_insn;
7860 rtx cc_src;
7861 enum machine_mode mode;
7862 enum machine_mode orig_mode;
7864 /* Look for blocks which end with a conditional jump based on a
7865 condition code register. Then look for the instruction which
7866 sets the condition code register. Then look through the
7867 successor blocks for instructions which set the condition
7868 code register to the same value. There are other possible
7869 uses of the condition code register, but these are by far the
7870 most common and the ones which we are most likely to be able
7871 to optimize. */
7873 last_insn = BB_END (bb);
7874 if (!JUMP_P (last_insn))
7875 continue;
7877 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7878 cc_reg = cc_reg_1;
7879 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7880 cc_reg = cc_reg_2;
7881 else
7882 continue;
7884 cc_src_insn = NULL_RTX;
7885 cc_src = NULL_RTX;
7886 for (insn = PREV_INSN (last_insn);
7887 insn && insn != PREV_INSN (BB_HEAD (bb));
7888 insn = PREV_INSN (insn))
7890 rtx set;
7892 if (! INSN_P (insn))
7893 continue;
7894 set = single_set (insn);
7895 if (set
7896 && REG_P (SET_DEST (set))
7897 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7899 cc_src_insn = insn;
7900 cc_src = SET_SRC (set);
7901 break;
7903 else if (reg_set_p (cc_reg, insn))
7904 break;
7907 if (! cc_src_insn)
7908 continue;
7910 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7911 continue;
7913 /* Now CC_REG is a condition code register used for a
7914 conditional jump at the end of the block, and CC_SRC, in
7915 CC_SRC_INSN, is the value to which that condition code
7916 register is set, and CC_SRC is still meaningful at the end of
7917 the basic block. */
7919 orig_mode = GET_MODE (cc_src);
7920 mode = cse_cc_succs (bb, cc_reg, cc_src, true);
7921 if (mode != VOIDmode)
7923 gcc_assert (mode == GET_MODE (cc_src));
7924 if (mode != orig_mode)
7926 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7928 cse_change_cc_mode_insn (cc_src_insn, newreg);
7930 /* Do the same in the following insns that use the
7931 current value of CC_REG within BB. */
7932 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7933 NEXT_INSN (last_insn),
7934 newreg);
7941 /* Perform common subexpression elimination. Nonzero value from
7942 `cse_main' means that jumps were simplified and some code may now
7943 be unreachable, so do jump optimization again. */
7944 static bool
7945 gate_handle_cse (void)
7947 return optimize > 0;
7950 static unsigned int
7951 rest_of_handle_cse (void)
7953 int tem;
7955 if (dump_file)
7956 dump_flow_info (dump_file, dump_flags);
7958 reg_scan (get_insns (), max_reg_num ());
7960 tem = cse_main (get_insns (), max_reg_num ());
7961 if (tem)
7962 rebuild_jump_labels (get_insns ());
7963 if (purge_all_dead_edges ())
7964 delete_unreachable_blocks ();
7966 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7968 /* If we are not running more CSE passes, then we are no longer
7969 expecting CSE to be run. But always rerun it in a cheap mode. */
7970 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7972 if (tem)
7973 delete_dead_jumptables ();
7975 if (tem || optimize > 1)
7976 cleanup_cfg (CLEANUP_EXPENSIVE);
7977 return 0;
7980 struct tree_opt_pass pass_cse =
7982 "cse1", /* name */
7983 gate_handle_cse, /* gate */
7984 rest_of_handle_cse, /* execute */
7985 NULL, /* sub */
7986 NULL, /* next */
7987 0, /* static_pass_number */
7988 TV_CSE, /* tv_id */
7989 0, /* properties_required */
7990 0, /* properties_provided */
7991 0, /* properties_destroyed */
7992 0, /* todo_flags_start */
7993 TODO_dump_func |
7994 TODO_ggc_collect, /* todo_flags_finish */
7995 's' /* letter */
7999 static bool
8000 gate_handle_cse2 (void)
8002 return optimize > 0 && flag_rerun_cse_after_loop;
8005 /* Run second CSE pass after loop optimizations. */
8006 static unsigned int
8007 rest_of_handle_cse2 (void)
8009 int tem;
8011 if (dump_file)
8012 dump_flow_info (dump_file, dump_flags);
8014 tem = cse_main (get_insns (), max_reg_num ());
8016 /* Run a pass to eliminate duplicated assignments to condition code
8017 registers. We have to run this after bypass_jumps, because it
8018 makes it harder for that pass to determine whether a jump can be
8019 bypassed safely. */
8020 cse_condition_code_reg ();
8022 purge_all_dead_edges ();
8023 delete_trivially_dead_insns (get_insns (), max_reg_num ());
8025 if (tem)
8027 timevar_push (TV_JUMP);
8028 rebuild_jump_labels (get_insns ());
8029 delete_dead_jumptables ();
8030 cleanup_cfg (CLEANUP_EXPENSIVE);
8031 timevar_pop (TV_JUMP);
8033 reg_scan (get_insns (), max_reg_num ());
8034 cse_not_expected = 1;
8035 return 0;
8039 struct tree_opt_pass pass_cse2 =
8041 "cse2", /* name */
8042 gate_handle_cse2, /* gate */
8043 rest_of_handle_cse2, /* execute */
8044 NULL, /* sub */
8045 NULL, /* next */
8046 0, /* static_pass_number */
8047 TV_CSE2, /* tv_id */
8048 0, /* properties_required */
8049 0, /* properties_provided */
8050 0, /* properties_destroyed */
8051 0, /* todo_flags_start */
8052 TODO_dump_func |
8053 TODO_ggc_collect, /* todo_flags_finish */
8054 't' /* letter */