* dwarf2out.c (dwarf_attr_name): Map DW_AT_GNAT_descriptive_type.
[official-gcc.git] / gcc / cse.c
blobf7b477c60b1fa994f8d76cb0472e63956ffbb805
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, 2006, 2007, 2008, 2009, 2010
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "regs.h"
30 #include "basic-block.h"
31 #include "flags.h"
32 #include "insn-config.h"
33 #include "recog.h"
34 #include "function.h"
35 #include "expr.h"
36 #include "diagnostic-core.h"
37 #include "toplev.h"
38 #include "output.h"
39 #include "ggc.h"
40 #include "timevar.h"
41 #include "except.h"
42 #include "target.h"
43 #include "params.h"
44 #include "rtlhooks-def.h"
45 #include "tree-pass.h"
46 #include "df.h"
47 #include "dbgcnt.h"
49 /* The basic idea of common subexpression elimination is to go
50 through the code, keeping a record of expressions that would
51 have the same value at the current scan point, and replacing
52 expressions encountered with the cheapest equivalent expression.
54 It is too complicated to keep track of the different possibilities
55 when control paths merge in this code; so, at each label, we forget all
56 that is known and start fresh. This can be described as processing each
57 extended basic block separately. We have a separate pass to perform
58 global CSE.
60 Note CSE can turn a conditional or computed jump into a nop or
61 an unconditional jump. When this occurs we arrange to run the jump
62 optimizer after CSE to delete the unreachable code.
64 We use two data structures to record the equivalent expressions:
65 a hash table for most expressions, and a vector of "quantity
66 numbers" to record equivalent (pseudo) registers.
68 The use of the special data structure for registers is desirable
69 because it is faster. It is possible because registers references
70 contain a fairly small number, the register number, taken from
71 a contiguously allocated series, and two register references are
72 identical if they have the same number. General expressions
73 do not have any such thing, so the only way to retrieve the
74 information recorded on an expression other than a register
75 is to keep it in a hash table.
77 Registers and "quantity numbers":
79 At the start of each basic block, all of the (hardware and pseudo)
80 registers used in the function are given distinct quantity
81 numbers to indicate their contents. During scan, when the code
82 copies one register into another, we copy the quantity number.
83 When a register is loaded in any other way, we allocate a new
84 quantity number to describe the value generated by this operation.
85 `REG_QTY (N)' records what quantity register N is currently thought
86 of as containing.
88 All real quantity numbers are greater than or equal to zero.
89 If register N has not been assigned a quantity, `REG_QTY (N)' will
90 equal -N - 1, which is always negative.
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
95 We also maintain a bidirectional chain of registers for each
96 quantity number. The `qty_table` members `first_reg' and `last_reg',
97 and `reg_eqv_table' members `next' and `prev' hold these chains.
99 The first register in a chain is the one whose lifespan is least local.
100 Among equals, it is the one that was seen first.
101 We replace any equivalent register with that one.
103 If two registers have the same quantity number, it must be true that
104 REG expressions with qty_table `mode' must be in the hash table for both
105 registers and must be in the same class.
107 The converse is not true. Since hard registers may be referenced in
108 any mode, two REG expressions might be equivalent in the hash table
109 but not have the same quantity number if the quantity number of one
110 of the registers is not the same mode as those expressions.
112 Constants and quantity numbers
114 When a quantity has a known constant value, that value is stored
115 in the appropriate qty_table `const_rtx'. This is in addition to
116 putting the constant in the hash table as is usual for non-regs.
118 Whether a reg or a constant is preferred is determined by the configuration
119 macro CONST_COSTS and will often depend on the constant value. In any
120 event, expressions containing constants can be simplified, by fold_rtx.
122 When a quantity has a known nearly constant value (such as an address
123 of a stack slot), that value is stored in the appropriate qty_table
124 `const_rtx'.
126 Integer constants don't have a machine mode. However, cse
127 determines the intended machine mode from the destination
128 of the instruction that moves the constant. The machine mode
129 is recorded in the hash table along with the actual RTL
130 constant expression so that different modes are kept separate.
132 Other expressions:
134 To record known equivalences among expressions in general
135 we use a hash table called `table'. It has a fixed number of buckets
136 that contain chains of `struct table_elt' elements for expressions.
137 These chains connect the elements whose expressions have the same
138 hash codes.
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
143 Register references in an expression are canonicalized before hashing
144 the expression. This is done using `reg_qty' and qty_table `first_reg'.
145 The hash code of a register reference is computed using the quantity
146 number, not the register number.
148 When the value of an expression changes, it is necessary to remove from the
149 hash table not just that expression but all expressions whose values
150 could be different as a result.
152 1. If the value changing is in memory, except in special cases
153 ANYTHING referring to memory could be changed. That is because
154 nobody knows where a pointer does not point.
155 The function `invalidate_memory' removes what is necessary.
157 The special cases are when the address is constant or is
158 a constant plus a fixed register such as the frame pointer
159 or a static chain pointer. When such addresses are stored in,
160 we can tell exactly which other such addresses must be invalidated
161 due to overlap. `invalidate' does this.
162 All expressions that refer to non-constant
163 memory addresses are also invalidated. `invalidate_memory' does this.
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
169 Because searching the entire hash table for expressions that contain
170 a register is very slow, we try to figure out when it isn't necessary.
171 Precisely, this is necessary only when expressions have been
172 entered in the hash table using this register, and then the value has
173 changed, and then another expression wants to be added to refer to
174 the register's new value. This sequence of circumstances is rare
175 within any one basic block.
177 `REG_TICK' and `REG_IN_TABLE', accessors for members of
178 cse_reg_info, are used to detect this case. REG_TICK (i) is
179 incremented whenever a value is stored in register i.
180 REG_IN_TABLE (i) holds -1 if no references to register i have been
181 entered in the table; otherwise, it contains the value REG_TICK (i)
182 had when the references were entered. If we want to enter a
183 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
184 remove old references. Until we want to enter a new entry, the
185 mere fact that the two vectors don't match makes the entries be
186 ignored if anyone tries to match them.
188 Registers themselves are entered in the hash table as well as in
189 the equivalent-register chains. However, `REG_TICK' and
190 `REG_IN_TABLE' do not apply to expressions which are simple
191 register references. These expressions are removed from the table
192 immediately when they become invalid, and this can be done even if
193 we do not immediately search for all the expressions that refer to
194 the register.
196 A CLOBBER rtx in an instruction invalidates its operand for further
197 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
198 invalidates everything that resides in memory.
200 Related expressions:
202 Constant expressions that differ only by an additive integer
203 are called related. When a constant expression is put in
204 the table, the related expression with no constant term
205 is also entered. These are made to point at each other
206 so that it is possible to find out if there exists any
207 register equivalent to an expression related to a given expression. */
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
212 static int max_qty;
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
217 static int next_qty;
219 /* Per-qty information tracking.
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
224 `mode' contains the machine mode of this quantity.
226 `const_rtx' holds the rtx of the constant value of this
227 quantity, if known. A summations of the frame/arg pointer
228 and a constant can also be entered here. When this holds
229 a known value, `const_insn' is the insn which stored the
230 constant value.
232 `comparison_{code,const,qty}' are used to track when a
233 comparison between a quantity and some constant or register has
234 been passed. In such a case, we know the results of the comparison
235 in case we see it again. These members record a comparison that
236 is known to be true. `comparison_code' holds the rtx code of such
237 a comparison, else it is set to UNKNOWN and the other two
238 comparison members are undefined. `comparison_const' holds
239 the constant being compared against, or zero if the comparison
240 is not against a constant. `comparison_qty' holds the quantity
241 being compared against when the result is known. If the comparison
242 is not with a register, `comparison_qty' is -1. */
244 struct qty_table_elem
246 rtx const_rtx;
247 rtx const_insn;
248 rtx comparison_const;
249 int comparison_qty;
250 unsigned int first_reg, last_reg;
251 /* The sizes of these fields should match the sizes of the
252 code and mode fields of struct rtx_def (see rtl.h). */
253 ENUM_BITFIELD(rtx_code) comparison_code : 16;
254 ENUM_BITFIELD(machine_mode) mode : 8;
257 /* The table of all qtys, indexed by qty number. */
258 static struct qty_table_elem *qty_table;
260 /* Structure used to pass arguments via for_each_rtx to function
261 cse_change_cc_mode. */
262 struct change_cc_mode_args
264 rtx insn;
265 rtx newreg;
268 #ifdef HAVE_cc0
269 /* For machines that have a CC0, we do not record its value in the hash
270 table since its use is guaranteed to be the insn immediately following
271 its definition and any other insn is presumed to invalidate it.
273 Instead, we store below the current and last value assigned to CC0.
274 If it should happen to be a constant, it is stored in preference
275 to the actual assigned value. In case it is a constant, we store
276 the mode in which the constant should be interpreted. */
278 static rtx this_insn_cc0, prev_insn_cc0;
279 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
280 #endif
282 /* Insn being scanned. */
284 static rtx this_insn;
285 static bool optimize_this_for_speed_p;
287 /* Index by register number, gives the number of the next (or
288 previous) register in the chain of registers sharing the same
289 value.
291 Or -1 if this register is at the end of the chain.
293 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
295 /* Per-register equivalence chain. */
296 struct reg_eqv_elem
298 int next, prev;
301 /* The table of all register equivalence chains. */
302 static struct reg_eqv_elem *reg_eqv_table;
304 struct cse_reg_info
306 /* The timestamp at which this register is initialized. */
307 unsigned int timestamp;
309 /* The quantity number of the register's current contents. */
310 int reg_qty;
312 /* The number of times the register has been altered in the current
313 basic block. */
314 int reg_tick;
316 /* The REG_TICK value at which rtx's containing this register are
317 valid in the hash table. If this does not equal the current
318 reg_tick value, such expressions existing in the hash table are
319 invalid. */
320 int reg_in_table;
322 /* The SUBREG that was set when REG_TICK was last incremented. Set
323 to -1 if the last store was to the whole register, not a subreg. */
324 unsigned int subreg_ticked;
327 /* A table of cse_reg_info indexed by register numbers. */
328 static struct cse_reg_info *cse_reg_info_table;
330 /* The size of the above table. */
331 static unsigned int cse_reg_info_table_size;
333 /* The index of the first entry that has not been initialized. */
334 static unsigned int cse_reg_info_table_first_uninitialized;
336 /* The timestamp at the beginning of the current run of
337 cse_extended_basic_block. We increment this variable at the beginning of
338 the current run of cse_extended_basic_block. The timestamp field of a
339 cse_reg_info entry matches the value of this variable if and only
340 if the entry has been initialized during the current run of
341 cse_extended_basic_block. */
342 static unsigned int cse_reg_info_timestamp;
344 /* A HARD_REG_SET containing all the hard registers for which there is
345 currently a REG expression in the hash table. Note the difference
346 from the above variables, which indicate if the REG is mentioned in some
347 expression in the table. */
349 static HARD_REG_SET hard_regs_in_table;
351 /* True if CSE has altered the CFG. */
352 static bool cse_cfg_altered;
354 /* True if CSE has altered conditional jump insns in such a way
355 that jump optimization should be redone. */
356 static bool cse_jumps_altered;
358 /* True if we put a LABEL_REF into the hash table for an INSN
359 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
360 to put in the note. */
361 static bool recorded_label_ref;
363 /* canon_hash stores 1 in do_not_record
364 if it notices a reference to CC0, PC, or some other volatile
365 subexpression. */
367 static int do_not_record;
369 /* canon_hash stores 1 in hash_arg_in_memory
370 if it notices a reference to memory within the expression being hashed. */
372 static int hash_arg_in_memory;
374 /* The hash table contains buckets which are chains of `struct table_elt's,
375 each recording one expression's information.
376 That expression is in the `exp' field.
378 The canon_exp field contains a canonical (from the point of view of
379 alias analysis) version of the `exp' field.
381 Those elements with the same hash code are chained in both directions
382 through the `next_same_hash' and `prev_same_hash' fields.
384 Each set of expressions with equivalent values
385 are on a two-way chain through the `next_same_value'
386 and `prev_same_value' fields, and all point with
387 the `first_same_value' field at the first element in
388 that chain. The chain is in order of increasing cost.
389 Each element's cost value is in its `cost' field.
391 The `in_memory' field is nonzero for elements that
392 involve any reference to memory. These elements are removed
393 whenever a write is done to an unidentified location in memory.
394 To be safe, we assume that a memory address is unidentified unless
395 the address is either a symbol constant or a constant plus
396 the frame pointer or argument pointer.
398 The `related_value' field is used to connect related expressions
399 (that differ by adding an integer).
400 The related expressions are chained in a circular fashion.
401 `related_value' is zero for expressions for which this
402 chain is not useful.
404 The `cost' field stores the cost of this element's expression.
405 The `regcost' field stores the value returned by approx_reg_cost for
406 this element's expression.
408 The `is_const' flag is set if the element is a constant (including
409 a fixed address).
411 The `flag' field is used as a temporary during some search routines.
413 The `mode' field is usually the same as GET_MODE (`exp'), but
414 if `exp' is a CONST_INT and has no machine mode then the `mode'
415 field is the mode it was being used as. Each constant is
416 recorded separately for each mode it is used with. */
418 struct table_elt
420 rtx exp;
421 rtx canon_exp;
422 struct table_elt *next_same_hash;
423 struct table_elt *prev_same_hash;
424 struct table_elt *next_same_value;
425 struct table_elt *prev_same_value;
426 struct table_elt *first_same_value;
427 struct table_elt *related_value;
428 int cost;
429 int regcost;
430 /* The size of this field should match the size
431 of the mode field of struct rtx_def (see rtl.h). */
432 ENUM_BITFIELD(machine_mode) mode : 8;
433 char in_memory;
434 char is_const;
435 char flag;
438 /* We don't want a lot of buckets, because we rarely have very many
439 things stored in the hash table, and a lot of buckets slows
440 down a lot of loops that happen frequently. */
441 #define HASH_SHIFT 5
442 #define HASH_SIZE (1 << HASH_SHIFT)
443 #define HASH_MASK (HASH_SIZE - 1)
445 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
446 register (hard registers may require `do_not_record' to be set). */
448 #define HASH(X, M) \
449 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
450 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
451 : canon_hash (X, M)) & HASH_MASK)
453 /* Like HASH, but without side-effects. */
454 #define SAFE_HASH(X, M) \
455 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
456 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
457 : safe_hash (X, M)) & HASH_MASK)
459 /* Determine whether register number N is considered a fixed register for the
460 purpose of approximating register costs.
461 It is desirable to replace other regs with fixed regs, to reduce need for
462 non-fixed hard regs.
463 A reg wins if it is either the frame pointer or designated as fixed. */
464 #define FIXED_REGNO_P(N) \
465 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
466 || fixed_regs[N] || global_regs[N])
468 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
469 hard registers and pointers into the frame are the cheapest with a cost
470 of 0. Next come pseudos with a cost of one and other hard registers with
471 a cost of 2. Aside from these special cases, call `rtx_cost'. */
473 #define CHEAP_REGNO(N) \
474 (REGNO_PTR_FRAME_P(N) \
475 || (HARD_REGISTER_NUM_P (N) \
476 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
478 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
479 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
481 /* Get the number of times this register has been updated in this
482 basic block. */
484 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
486 /* Get the point at which REG was recorded in the table. */
488 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
490 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
491 SUBREG). */
493 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
495 /* Get the quantity number for REG. */
497 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
499 /* Determine if the quantity number for register X represents a valid index
500 into the qty_table. */
502 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
504 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
506 #define CHEAPER(X, Y) \
507 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
509 static struct table_elt *table[HASH_SIZE];
511 /* Chain of `struct table_elt's made so far for this function
512 but currently removed from the table. */
514 static struct table_elt *free_element_chain;
516 /* Set to the cost of a constant pool reference if one was found for a
517 symbolic constant. If this was found, it means we should try to
518 convert constants into constant pool entries if they don't fit in
519 the insn. */
521 static int constant_pool_entries_cost;
522 static int constant_pool_entries_regcost;
524 /* Trace a patch through the CFG. */
526 struct branch_path
528 /* The basic block for this path entry. */
529 basic_block bb;
532 /* This data describes a block that will be processed by
533 cse_extended_basic_block. */
535 struct cse_basic_block_data
537 /* Total number of SETs in block. */
538 int nsets;
539 /* Size of current branch path, if any. */
540 int path_size;
541 /* Current path, indicating which basic_blocks will be processed. */
542 struct branch_path *path;
546 /* Pointers to the live in/live out bitmaps for the boundaries of the
547 current EBB. */
548 static bitmap cse_ebb_live_in, cse_ebb_live_out;
550 /* A simple bitmap to track which basic blocks have been visited
551 already as part of an already processed extended basic block. */
552 static sbitmap cse_visited_basic_blocks;
554 static bool fixed_base_plus_p (rtx x);
555 static int notreg_cost (rtx, enum rtx_code);
556 static int approx_reg_cost_1 (rtx *, void *);
557 static int approx_reg_cost (rtx);
558 static int preferable (int, int, int, int);
559 static void new_basic_block (void);
560 static void make_new_qty (unsigned int, enum machine_mode);
561 static void make_regs_eqv (unsigned int, unsigned int);
562 static void delete_reg_equiv (unsigned int);
563 static int mention_regs (rtx);
564 static int insert_regs (rtx, struct table_elt *, int);
565 static void remove_from_table (struct table_elt *, unsigned);
566 static void remove_pseudo_from_table (rtx, unsigned);
567 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
568 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
569 static rtx lookup_as_function (rtx, enum rtx_code);
570 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
571 enum machine_mode, int, int);
572 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
573 enum machine_mode);
574 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
575 static void invalidate (rtx, enum machine_mode);
576 static bool cse_rtx_varies_p (const_rtx, bool);
577 static void remove_invalid_refs (unsigned int);
578 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
579 enum machine_mode);
580 static void rehash_using_reg (rtx);
581 static void invalidate_memory (void);
582 static void invalidate_for_call (void);
583 static rtx use_related_value (rtx, struct table_elt *);
585 static inline unsigned canon_hash (rtx, enum machine_mode);
586 static inline unsigned safe_hash (rtx, enum machine_mode);
587 static inline unsigned hash_rtx_string (const char *);
589 static rtx canon_reg (rtx, rtx);
590 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
591 enum machine_mode *,
592 enum machine_mode *);
593 static rtx fold_rtx (rtx, rtx);
594 static rtx equiv_constant (rtx);
595 static void record_jump_equiv (rtx, bool);
596 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
597 int);
598 static void cse_insn (rtx);
599 static void cse_prescan_path (struct cse_basic_block_data *);
600 static void invalidate_from_clobbers (rtx);
601 static rtx cse_process_notes (rtx, rtx, bool *);
602 static void cse_extended_basic_block (struct cse_basic_block_data *);
603 static void count_reg_usage (rtx, int *, rtx, int);
604 static int check_for_label_ref (rtx *, void *);
605 extern void dump_class (struct table_elt*);
606 static void get_cse_reg_info_1 (unsigned int regno);
607 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
608 static int check_dependence (rtx *, void *);
610 static void flush_hash_table (void);
611 static bool insn_live_p (rtx, int *);
612 static bool set_live_p (rtx, rtx, int *);
613 static int cse_change_cc_mode (rtx *, void *);
614 static void cse_change_cc_mode_insn (rtx, rtx);
615 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
616 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
617 bool);
620 #undef RTL_HOOKS_GEN_LOWPART
621 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
623 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
625 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
626 virtual regs here because the simplify_*_operation routines are called
627 by integrate.c, which is called before virtual register instantiation. */
629 static bool
630 fixed_base_plus_p (rtx x)
632 switch (GET_CODE (x))
634 case REG:
635 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
636 return true;
637 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
638 return true;
639 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
640 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
641 return true;
642 return false;
644 case PLUS:
645 if (!CONST_INT_P (XEXP (x, 1)))
646 return false;
647 return fixed_base_plus_p (XEXP (x, 0));
649 default:
650 return false;
654 /* Dump the expressions in the equivalence class indicated by CLASSP.
655 This function is used only for debugging. */
656 void
657 dump_class (struct table_elt *classp)
659 struct table_elt *elt;
661 fprintf (stderr, "Equivalence chain for ");
662 print_rtl (stderr, classp->exp);
663 fprintf (stderr, ": \n");
665 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
667 print_rtl (stderr, elt->exp);
668 fprintf (stderr, "\n");
672 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
674 static int
675 approx_reg_cost_1 (rtx *xp, void *data)
677 rtx x = *xp;
678 int *cost_p = (int *) data;
680 if (x && REG_P (x))
682 unsigned int regno = REGNO (x);
684 if (! CHEAP_REGNO (regno))
686 if (regno < FIRST_PSEUDO_REGISTER)
688 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
689 return 1;
690 *cost_p += 2;
692 else
693 *cost_p += 1;
697 return 0;
700 /* Return an estimate of the cost of the registers used in an rtx.
701 This is mostly the number of different REG expressions in the rtx;
702 however for some exceptions like fixed registers we use a cost of
703 0. If any other hard register reference occurs, return MAX_COST. */
705 static int
706 approx_reg_cost (rtx x)
708 int cost = 0;
710 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
711 return MAX_COST;
713 return cost;
716 /* Return a negative value if an rtx A, whose costs are given by COST_A
717 and REGCOST_A, is more desirable than an rtx B.
718 Return a positive value if A is less desirable, or 0 if the two are
719 equally good. */
720 static int
721 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
723 /* First, get rid of cases involving expressions that are entirely
724 unwanted. */
725 if (cost_a != cost_b)
727 if (cost_a == MAX_COST)
728 return 1;
729 if (cost_b == MAX_COST)
730 return -1;
733 /* Avoid extending lifetimes of hardregs. */
734 if (regcost_a != regcost_b)
736 if (regcost_a == MAX_COST)
737 return 1;
738 if (regcost_b == MAX_COST)
739 return -1;
742 /* Normal operation costs take precedence. */
743 if (cost_a != cost_b)
744 return cost_a - cost_b;
745 /* Only if these are identical consider effects on register pressure. */
746 if (regcost_a != regcost_b)
747 return regcost_a - regcost_b;
748 return 0;
751 /* Internal function, to compute cost when X is not a register; called
752 from COST macro to keep it simple. */
754 static int
755 notreg_cost (rtx x, enum rtx_code outer)
757 return ((GET_CODE (x) == SUBREG
758 && REG_P (SUBREG_REG (x))
759 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
760 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
761 && (GET_MODE_SIZE (GET_MODE (x))
762 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
763 && subreg_lowpart_p (x)
764 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
765 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
767 : rtx_cost (x, outer, optimize_this_for_speed_p) * 2);
771 /* Initialize CSE_REG_INFO_TABLE. */
773 static void
774 init_cse_reg_info (unsigned int nregs)
776 /* Do we need to grow the table? */
777 if (nregs > cse_reg_info_table_size)
779 unsigned int new_size;
781 if (cse_reg_info_table_size < 2048)
783 /* Compute a new size that is a power of 2 and no smaller
784 than the large of NREGS and 64. */
785 new_size = (cse_reg_info_table_size
786 ? cse_reg_info_table_size : 64);
788 while (new_size < nregs)
789 new_size *= 2;
791 else
793 /* If we need a big table, allocate just enough to hold
794 NREGS registers. */
795 new_size = nregs;
798 /* Reallocate the table with NEW_SIZE entries. */
799 if (cse_reg_info_table)
800 free (cse_reg_info_table);
801 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
802 cse_reg_info_table_size = new_size;
803 cse_reg_info_table_first_uninitialized = 0;
806 /* Do we have all of the first NREGS entries initialized? */
807 if (cse_reg_info_table_first_uninitialized < nregs)
809 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
810 unsigned int i;
812 /* Put the old timestamp on newly allocated entries so that they
813 will all be considered out of date. We do not touch those
814 entries beyond the first NREGS entries to be nice to the
815 virtual memory. */
816 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
817 cse_reg_info_table[i].timestamp = old_timestamp;
819 cse_reg_info_table_first_uninitialized = nregs;
823 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
825 static void
826 get_cse_reg_info_1 (unsigned int regno)
828 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
829 entry will be considered to have been initialized. */
830 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
832 /* Initialize the rest of the entry. */
833 cse_reg_info_table[regno].reg_tick = 1;
834 cse_reg_info_table[regno].reg_in_table = -1;
835 cse_reg_info_table[regno].subreg_ticked = -1;
836 cse_reg_info_table[regno].reg_qty = -regno - 1;
839 /* Find a cse_reg_info entry for REGNO. */
841 static inline struct cse_reg_info *
842 get_cse_reg_info (unsigned int regno)
844 struct cse_reg_info *p = &cse_reg_info_table[regno];
846 /* If this entry has not been initialized, go ahead and initialize
847 it. */
848 if (p->timestamp != cse_reg_info_timestamp)
849 get_cse_reg_info_1 (regno);
851 return p;
854 /* Clear the hash table and initialize each register with its own quantity,
855 for a new basic block. */
857 static void
858 new_basic_block (void)
860 int i;
862 next_qty = 0;
864 /* Invalidate cse_reg_info_table. */
865 cse_reg_info_timestamp++;
867 /* Clear out hash table state for this pass. */
868 CLEAR_HARD_REG_SET (hard_regs_in_table);
870 /* The per-quantity values used to be initialized here, but it is
871 much faster to initialize each as it is made in `make_new_qty'. */
873 for (i = 0; i < HASH_SIZE; i++)
875 struct table_elt *first;
877 first = table[i];
878 if (first != NULL)
880 struct table_elt *last = first;
882 table[i] = NULL;
884 while (last->next_same_hash != NULL)
885 last = last->next_same_hash;
887 /* Now relink this hash entire chain into
888 the free element list. */
890 last->next_same_hash = free_element_chain;
891 free_element_chain = first;
895 #ifdef HAVE_cc0
896 prev_insn_cc0 = 0;
897 #endif
900 /* Say that register REG contains a quantity in mode MODE not in any
901 register before and initialize that quantity. */
903 static void
904 make_new_qty (unsigned int reg, enum machine_mode mode)
906 int q;
907 struct qty_table_elem *ent;
908 struct reg_eqv_elem *eqv;
910 gcc_assert (next_qty < max_qty);
912 q = REG_QTY (reg) = next_qty++;
913 ent = &qty_table[q];
914 ent->first_reg = reg;
915 ent->last_reg = reg;
916 ent->mode = mode;
917 ent->const_rtx = ent->const_insn = NULL_RTX;
918 ent->comparison_code = UNKNOWN;
920 eqv = &reg_eqv_table[reg];
921 eqv->next = eqv->prev = -1;
924 /* Make reg NEW equivalent to reg OLD.
925 OLD is not changing; NEW is. */
927 static void
928 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
930 unsigned int lastr, firstr;
931 int q = REG_QTY (old_reg);
932 struct qty_table_elem *ent;
934 ent = &qty_table[q];
936 /* Nothing should become eqv until it has a "non-invalid" qty number. */
937 gcc_assert (REGNO_QTY_VALID_P (old_reg));
939 REG_QTY (new_reg) = q;
940 firstr = ent->first_reg;
941 lastr = ent->last_reg;
943 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
944 hard regs. Among pseudos, if NEW will live longer than any other reg
945 of the same qty, and that is beyond the current basic block,
946 make it the new canonical replacement for this qty. */
947 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
948 /* Certain fixed registers might be of the class NO_REGS. This means
949 that not only can they not be allocated by the compiler, but
950 they cannot be used in substitutions or canonicalizations
951 either. */
952 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
953 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
954 || (new_reg >= FIRST_PSEUDO_REGISTER
955 && (firstr < FIRST_PSEUDO_REGISTER
956 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
957 && !bitmap_bit_p (cse_ebb_live_out, firstr))
958 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
959 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
961 reg_eqv_table[firstr].prev = new_reg;
962 reg_eqv_table[new_reg].next = firstr;
963 reg_eqv_table[new_reg].prev = -1;
964 ent->first_reg = new_reg;
966 else
968 /* If NEW is a hard reg (known to be non-fixed), insert at end.
969 Otherwise, insert before any non-fixed hard regs that are at the
970 end. Registers of class NO_REGS cannot be used as an
971 equivalent for anything. */
972 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
973 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
974 && new_reg >= FIRST_PSEUDO_REGISTER)
975 lastr = reg_eqv_table[lastr].prev;
976 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
977 if (reg_eqv_table[lastr].next >= 0)
978 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
979 else
980 qty_table[q].last_reg = new_reg;
981 reg_eqv_table[lastr].next = new_reg;
982 reg_eqv_table[new_reg].prev = lastr;
986 /* Remove REG from its equivalence class. */
988 static void
989 delete_reg_equiv (unsigned int reg)
991 struct qty_table_elem *ent;
992 int q = REG_QTY (reg);
993 int p, n;
995 /* If invalid, do nothing. */
996 if (! REGNO_QTY_VALID_P (reg))
997 return;
999 ent = &qty_table[q];
1001 p = reg_eqv_table[reg].prev;
1002 n = reg_eqv_table[reg].next;
1004 if (n != -1)
1005 reg_eqv_table[n].prev = p;
1006 else
1007 ent->last_reg = p;
1008 if (p != -1)
1009 reg_eqv_table[p].next = n;
1010 else
1011 ent->first_reg = n;
1013 REG_QTY (reg) = -reg - 1;
1016 /* Remove any invalid expressions from the hash table
1017 that refer to any of the registers contained in expression X.
1019 Make sure that newly inserted references to those registers
1020 as subexpressions will be considered valid.
1022 mention_regs is not called when a register itself
1023 is being stored in the table.
1025 Return 1 if we have done something that may have changed the hash code
1026 of X. */
1028 static int
1029 mention_regs (rtx x)
1031 enum rtx_code code;
1032 int i, j;
1033 const char *fmt;
1034 int changed = 0;
1036 if (x == 0)
1037 return 0;
1039 code = GET_CODE (x);
1040 if (code == REG)
1042 unsigned int regno = REGNO (x);
1043 unsigned int endregno = END_REGNO (x);
1044 unsigned int i;
1046 for (i = regno; i < endregno; i++)
1048 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1049 remove_invalid_refs (i);
1051 REG_IN_TABLE (i) = REG_TICK (i);
1052 SUBREG_TICKED (i) = -1;
1055 return 0;
1058 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1059 pseudo if they don't use overlapping words. We handle only pseudos
1060 here for simplicity. */
1061 if (code == SUBREG && REG_P (SUBREG_REG (x))
1062 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1064 unsigned int i = REGNO (SUBREG_REG (x));
1066 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1068 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1069 the last store to this register really stored into this
1070 subreg, then remove the memory of this subreg.
1071 Otherwise, remove any memory of the entire register and
1072 all its subregs from the table. */
1073 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1074 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1075 remove_invalid_refs (i);
1076 else
1077 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1080 REG_IN_TABLE (i) = REG_TICK (i);
1081 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1082 return 0;
1085 /* If X is a comparison or a COMPARE and either operand is a register
1086 that does not have a quantity, give it one. This is so that a later
1087 call to record_jump_equiv won't cause X to be assigned a different
1088 hash code and not found in the table after that call.
1090 It is not necessary to do this here, since rehash_using_reg can
1091 fix up the table later, but doing this here eliminates the need to
1092 call that expensive function in the most common case where the only
1093 use of the register is in the comparison. */
1095 if (code == COMPARE || COMPARISON_P (x))
1097 if (REG_P (XEXP (x, 0))
1098 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1099 if (insert_regs (XEXP (x, 0), NULL, 0))
1101 rehash_using_reg (XEXP (x, 0));
1102 changed = 1;
1105 if (REG_P (XEXP (x, 1))
1106 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1107 if (insert_regs (XEXP (x, 1), NULL, 0))
1109 rehash_using_reg (XEXP (x, 1));
1110 changed = 1;
1114 fmt = GET_RTX_FORMAT (code);
1115 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1116 if (fmt[i] == 'e')
1117 changed |= mention_regs (XEXP (x, i));
1118 else if (fmt[i] == 'E')
1119 for (j = 0; j < XVECLEN (x, i); j++)
1120 changed |= mention_regs (XVECEXP (x, i, j));
1122 return changed;
1125 /* Update the register quantities for inserting X into the hash table
1126 with a value equivalent to CLASSP.
1127 (If the class does not contain a REG, it is irrelevant.)
1128 If MODIFIED is nonzero, X is a destination; it is being modified.
1129 Note that delete_reg_equiv should be called on a register
1130 before insert_regs is done on that register with MODIFIED != 0.
1132 Nonzero value means that elements of reg_qty have changed
1133 so X's hash code may be different. */
1135 static int
1136 insert_regs (rtx x, struct table_elt *classp, int modified)
1138 if (REG_P (x))
1140 unsigned int regno = REGNO (x);
1141 int qty_valid;
1143 /* If REGNO is in the equivalence table already but is of the
1144 wrong mode for that equivalence, don't do anything here. */
1146 qty_valid = REGNO_QTY_VALID_P (regno);
1147 if (qty_valid)
1149 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1151 if (ent->mode != GET_MODE (x))
1152 return 0;
1155 if (modified || ! qty_valid)
1157 if (classp)
1158 for (classp = classp->first_same_value;
1159 classp != 0;
1160 classp = classp->next_same_value)
1161 if (REG_P (classp->exp)
1162 && GET_MODE (classp->exp) == GET_MODE (x))
1164 unsigned c_regno = REGNO (classp->exp);
1166 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1168 /* Suppose that 5 is hard reg and 100 and 101 are
1169 pseudos. Consider
1171 (set (reg:si 100) (reg:si 5))
1172 (set (reg:si 5) (reg:si 100))
1173 (set (reg:di 101) (reg:di 5))
1175 We would now set REG_QTY (101) = REG_QTY (5), but the
1176 entry for 5 is in SImode. When we use this later in
1177 copy propagation, we get the register in wrong mode. */
1178 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1179 continue;
1181 make_regs_eqv (regno, c_regno);
1182 return 1;
1185 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1186 than REG_IN_TABLE to find out if there was only a single preceding
1187 invalidation - for the SUBREG - or another one, which would be
1188 for the full register. However, if we find here that REG_TICK
1189 indicates that the register is invalid, it means that it has
1190 been invalidated in a separate operation. The SUBREG might be used
1191 now (then this is a recursive call), or we might use the full REG
1192 now and a SUBREG of it later. So bump up REG_TICK so that
1193 mention_regs will do the right thing. */
1194 if (! modified
1195 && REG_IN_TABLE (regno) >= 0
1196 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1197 REG_TICK (regno)++;
1198 make_new_qty (regno, GET_MODE (x));
1199 return 1;
1202 return 0;
1205 /* If X is a SUBREG, we will likely be inserting the inner register in the
1206 table. If that register doesn't have an assigned quantity number at
1207 this point but does later, the insertion that we will be doing now will
1208 not be accessible because its hash code will have changed. So assign
1209 a quantity number now. */
1211 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1212 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1214 insert_regs (SUBREG_REG (x), NULL, 0);
1215 mention_regs (x);
1216 return 1;
1218 else
1219 return mention_regs (x);
1223 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1224 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1225 CST is equal to an anchor. */
1227 static bool
1228 compute_const_anchors (rtx cst,
1229 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1230 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1232 HOST_WIDE_INT n = INTVAL (cst);
1234 *lower_base = n & ~(targetm.const_anchor - 1);
1235 if (*lower_base == n)
1236 return false;
1238 *upper_base =
1239 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1240 *upper_offs = n - *upper_base;
1241 *lower_offs = n - *lower_base;
1242 return true;
1245 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1247 static void
1248 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1249 enum machine_mode mode)
1251 struct table_elt *elt;
1252 unsigned hash;
1253 rtx anchor_exp;
1254 rtx exp;
1256 anchor_exp = GEN_INT (anchor);
1257 hash = HASH (anchor_exp, mode);
1258 elt = lookup (anchor_exp, hash, mode);
1259 if (!elt)
1260 elt = insert (anchor_exp, NULL, hash, mode);
1262 exp = plus_constant (reg, offs);
1263 /* REG has just been inserted and the hash codes recomputed. */
1264 mention_regs (exp);
1265 hash = HASH (exp, mode);
1267 /* Use the cost of the register rather than the whole expression. When
1268 looking up constant anchors we will further offset the corresponding
1269 expression therefore it does not make sense to prefer REGs over
1270 reg-immediate additions. Prefer instead the oldest expression. Also
1271 don't prefer pseudos over hard regs so that we derive constants in
1272 argument registers from other argument registers rather than from the
1273 original pseudo that was used to synthesize the constant. */
1274 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1277 /* The constant CST is equivalent to the register REG. Create
1278 equivalences between the two anchors of CST and the corresponding
1279 register-offset expressions using REG. */
1281 static void
1282 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1284 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1286 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1287 &upper_base, &upper_offs))
1288 return;
1290 /* Ignore anchors of value 0. Constants accessible from zero are
1291 simple. */
1292 if (lower_base != 0)
1293 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1295 if (upper_base != 0)
1296 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1299 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1300 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1301 valid expression. Return the cheapest and oldest of such expressions. In
1302 *OLD, return how old the resulting expression is compared to the other
1303 equivalent expressions. */
1305 static rtx
1306 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1307 unsigned *old)
1309 struct table_elt *elt;
1310 unsigned idx;
1311 struct table_elt *match_elt;
1312 rtx match;
1314 /* Find the cheapest and *oldest* expression to maximize the chance of
1315 reusing the same pseudo. */
1317 match_elt = NULL;
1318 match = NULL_RTX;
1319 for (elt = anchor_elt->first_same_value, idx = 0;
1320 elt;
1321 elt = elt->next_same_value, idx++)
1323 if (match_elt && CHEAPER (match_elt, elt))
1324 return match;
1326 if (REG_P (elt->exp)
1327 || (GET_CODE (elt->exp) == PLUS
1328 && REG_P (XEXP (elt->exp, 0))
1329 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1331 rtx x;
1333 /* Ignore expressions that are no longer valid. */
1334 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1335 continue;
1337 x = plus_constant (elt->exp, offs);
1338 if (REG_P (x)
1339 || (GET_CODE (x) == PLUS
1340 && IN_RANGE (INTVAL (XEXP (x, 1)),
1341 -targetm.const_anchor,
1342 targetm.const_anchor - 1)))
1344 match = x;
1345 match_elt = elt;
1346 *old = idx;
1351 return match;
1354 /* Try to express the constant SRC_CONST using a register+offset expression
1355 derived from a constant anchor. Return it if successful or NULL_RTX,
1356 otherwise. */
1358 static rtx
1359 try_const_anchors (rtx src_const, enum machine_mode mode)
1361 struct table_elt *lower_elt, *upper_elt;
1362 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1363 rtx lower_anchor_rtx, upper_anchor_rtx;
1364 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1365 unsigned lower_old, upper_old;
1367 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1368 &upper_base, &upper_offs))
1369 return NULL_RTX;
1371 lower_anchor_rtx = GEN_INT (lower_base);
1372 upper_anchor_rtx = GEN_INT (upper_base);
1373 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1374 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1376 if (lower_elt)
1377 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1378 if (upper_elt)
1379 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1381 if (!lower_exp)
1382 return upper_exp;
1383 if (!upper_exp)
1384 return lower_exp;
1386 /* Return the older expression. */
1387 return (upper_old > lower_old ? upper_exp : lower_exp);
1390 /* Look in or update the hash table. */
1392 /* Remove table element ELT from use in the table.
1393 HASH is its hash code, made using the HASH macro.
1394 It's an argument because often that is known in advance
1395 and we save much time not recomputing it. */
1397 static void
1398 remove_from_table (struct table_elt *elt, unsigned int hash)
1400 if (elt == 0)
1401 return;
1403 /* Mark this element as removed. See cse_insn. */
1404 elt->first_same_value = 0;
1406 /* Remove the table element from its equivalence class. */
1409 struct table_elt *prev = elt->prev_same_value;
1410 struct table_elt *next = elt->next_same_value;
1412 if (next)
1413 next->prev_same_value = prev;
1415 if (prev)
1416 prev->next_same_value = next;
1417 else
1419 struct table_elt *newfirst = next;
1420 while (next)
1422 next->first_same_value = newfirst;
1423 next = next->next_same_value;
1428 /* Remove the table element from its hash bucket. */
1431 struct table_elt *prev = elt->prev_same_hash;
1432 struct table_elt *next = elt->next_same_hash;
1434 if (next)
1435 next->prev_same_hash = prev;
1437 if (prev)
1438 prev->next_same_hash = next;
1439 else if (table[hash] == elt)
1440 table[hash] = next;
1441 else
1443 /* This entry is not in the proper hash bucket. This can happen
1444 when two classes were merged by `merge_equiv_classes'. Search
1445 for the hash bucket that it heads. This happens only very
1446 rarely, so the cost is acceptable. */
1447 for (hash = 0; hash < HASH_SIZE; hash++)
1448 if (table[hash] == elt)
1449 table[hash] = next;
1453 /* Remove the table element from its related-value circular chain. */
1455 if (elt->related_value != 0 && elt->related_value != elt)
1457 struct table_elt *p = elt->related_value;
1459 while (p->related_value != elt)
1460 p = p->related_value;
1461 p->related_value = elt->related_value;
1462 if (p->related_value == p)
1463 p->related_value = 0;
1466 /* Now add it to the free element chain. */
1467 elt->next_same_hash = free_element_chain;
1468 free_element_chain = elt;
1471 /* Same as above, but X is a pseudo-register. */
1473 static void
1474 remove_pseudo_from_table (rtx x, unsigned int hash)
1476 struct table_elt *elt;
1478 /* Because a pseudo-register can be referenced in more than one
1479 mode, we might have to remove more than one table entry. */
1480 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1481 remove_from_table (elt, hash);
1484 /* Look up X in the hash table and return its table element,
1485 or 0 if X is not in the table.
1487 MODE is the machine-mode of X, or if X is an integer constant
1488 with VOIDmode then MODE is the mode with which X will be used.
1490 Here we are satisfied to find an expression whose tree structure
1491 looks like X. */
1493 static struct table_elt *
1494 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1496 struct table_elt *p;
1498 for (p = table[hash]; p; p = p->next_same_hash)
1499 if (mode == p->mode && ((x == p->exp && REG_P (x))
1500 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1501 return p;
1503 return 0;
1506 /* Like `lookup' but don't care whether the table element uses invalid regs.
1507 Also ignore discrepancies in the machine mode of a register. */
1509 static struct table_elt *
1510 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1512 struct table_elt *p;
1514 if (REG_P (x))
1516 unsigned int regno = REGNO (x);
1518 /* Don't check the machine mode when comparing registers;
1519 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1520 for (p = table[hash]; p; p = p->next_same_hash)
1521 if (REG_P (p->exp)
1522 && REGNO (p->exp) == regno)
1523 return p;
1525 else
1527 for (p = table[hash]; p; p = p->next_same_hash)
1528 if (mode == p->mode
1529 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1530 return p;
1533 return 0;
1536 /* Look for an expression equivalent to X and with code CODE.
1537 If one is found, return that expression. */
1539 static rtx
1540 lookup_as_function (rtx x, enum rtx_code code)
1542 struct table_elt *p
1543 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1545 if (p == 0)
1546 return 0;
1548 for (p = p->first_same_value; p; p = p->next_same_value)
1549 if (GET_CODE (p->exp) == code
1550 /* Make sure this is a valid entry in the table. */
1551 && exp_equiv_p (p->exp, p->exp, 1, false))
1552 return p->exp;
1554 return 0;
1557 /* Insert X in the hash table, assuming HASH is its hash code and
1558 CLASSP is an element of the class it should go in (or 0 if a new
1559 class should be made). COST is the code of X and reg_cost is the
1560 cost of registers in X. It is inserted at the proper position to
1561 keep the class in the order cheapest first.
1563 MODE is the machine-mode of X, or if X is an integer constant
1564 with VOIDmode then MODE is the mode with which X will be used.
1566 For elements of equal cheapness, the most recent one
1567 goes in front, except that the first element in the list
1568 remains first unless a cheaper element is added. The order of
1569 pseudo-registers does not matter, as canon_reg will be called to
1570 find the cheapest when a register is retrieved from the table.
1572 The in_memory field in the hash table element is set to 0.
1573 The caller must set it nonzero if appropriate.
1575 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1576 and if insert_regs returns a nonzero value
1577 you must then recompute its hash code before calling here.
1579 If necessary, update table showing constant values of quantities. */
1581 static struct table_elt *
1582 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1583 enum machine_mode mode, int cost, int reg_cost)
1585 struct table_elt *elt;
1587 /* If X is a register and we haven't made a quantity for it,
1588 something is wrong. */
1589 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1591 /* If X is a hard register, show it is being put in the table. */
1592 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1593 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1595 /* Put an element for X into the right hash bucket. */
1597 elt = free_element_chain;
1598 if (elt)
1599 free_element_chain = elt->next_same_hash;
1600 else
1601 elt = XNEW (struct table_elt);
1603 elt->exp = x;
1604 elt->canon_exp = NULL_RTX;
1605 elt->cost = cost;
1606 elt->regcost = reg_cost;
1607 elt->next_same_value = 0;
1608 elt->prev_same_value = 0;
1609 elt->next_same_hash = table[hash];
1610 elt->prev_same_hash = 0;
1611 elt->related_value = 0;
1612 elt->in_memory = 0;
1613 elt->mode = mode;
1614 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1616 if (table[hash])
1617 table[hash]->prev_same_hash = elt;
1618 table[hash] = elt;
1620 /* Put it into the proper value-class. */
1621 if (classp)
1623 classp = classp->first_same_value;
1624 if (CHEAPER (elt, classp))
1625 /* Insert at the head of the class. */
1627 struct table_elt *p;
1628 elt->next_same_value = classp;
1629 classp->prev_same_value = elt;
1630 elt->first_same_value = elt;
1632 for (p = classp; p; p = p->next_same_value)
1633 p->first_same_value = elt;
1635 else
1637 /* Insert not at head of the class. */
1638 /* Put it after the last element cheaper than X. */
1639 struct table_elt *p, *next;
1641 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1642 p = next);
1644 /* Put it after P and before NEXT. */
1645 elt->next_same_value = next;
1646 if (next)
1647 next->prev_same_value = elt;
1649 elt->prev_same_value = p;
1650 p->next_same_value = elt;
1651 elt->first_same_value = classp;
1654 else
1655 elt->first_same_value = elt;
1657 /* If this is a constant being set equivalent to a register or a register
1658 being set equivalent to a constant, note the constant equivalence.
1660 If this is a constant, it cannot be equivalent to a different constant,
1661 and a constant is the only thing that can be cheaper than a register. So
1662 we know the register is the head of the class (before the constant was
1663 inserted).
1665 If this is a register that is not already known equivalent to a
1666 constant, we must check the entire class.
1668 If this is a register that is already known equivalent to an insn,
1669 update the qtys `const_insn' to show that `this_insn' is the latest
1670 insn making that quantity equivalent to the constant. */
1672 if (elt->is_const && classp && REG_P (classp->exp)
1673 && !REG_P (x))
1675 int exp_q = REG_QTY (REGNO (classp->exp));
1676 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1678 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1679 exp_ent->const_insn = this_insn;
1682 else if (REG_P (x)
1683 && classp
1684 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1685 && ! elt->is_const)
1687 struct table_elt *p;
1689 for (p = classp; p != 0; p = p->next_same_value)
1691 if (p->is_const && !REG_P (p->exp))
1693 int x_q = REG_QTY (REGNO (x));
1694 struct qty_table_elem *x_ent = &qty_table[x_q];
1696 x_ent->const_rtx
1697 = gen_lowpart (GET_MODE (x), p->exp);
1698 x_ent->const_insn = this_insn;
1699 break;
1704 else if (REG_P (x)
1705 && qty_table[REG_QTY (REGNO (x))].const_rtx
1706 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1707 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1709 /* If this is a constant with symbolic value,
1710 and it has a term with an explicit integer value,
1711 link it up with related expressions. */
1712 if (GET_CODE (x) == CONST)
1714 rtx subexp = get_related_value (x);
1715 unsigned subhash;
1716 struct table_elt *subelt, *subelt_prev;
1718 if (subexp != 0)
1720 /* Get the integer-free subexpression in the hash table. */
1721 subhash = SAFE_HASH (subexp, mode);
1722 subelt = lookup (subexp, subhash, mode);
1723 if (subelt == 0)
1724 subelt = insert (subexp, NULL, subhash, mode);
1725 /* Initialize SUBELT's circular chain if it has none. */
1726 if (subelt->related_value == 0)
1727 subelt->related_value = subelt;
1728 /* Find the element in the circular chain that precedes SUBELT. */
1729 subelt_prev = subelt;
1730 while (subelt_prev->related_value != subelt)
1731 subelt_prev = subelt_prev->related_value;
1732 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1733 This way the element that follows SUBELT is the oldest one. */
1734 elt->related_value = subelt_prev->related_value;
1735 subelt_prev->related_value = elt;
1739 return elt;
1742 /* Wrap insert_with_costs by passing the default costs. */
1744 static struct table_elt *
1745 insert (rtx x, struct table_elt *classp, unsigned int hash,
1746 enum machine_mode mode)
1748 return
1749 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1753 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1754 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1755 the two classes equivalent.
1757 CLASS1 will be the surviving class; CLASS2 should not be used after this
1758 call.
1760 Any invalid entries in CLASS2 will not be copied. */
1762 static void
1763 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1765 struct table_elt *elt, *next, *new_elt;
1767 /* Ensure we start with the head of the classes. */
1768 class1 = class1->first_same_value;
1769 class2 = class2->first_same_value;
1771 /* If they were already equal, forget it. */
1772 if (class1 == class2)
1773 return;
1775 for (elt = class2; elt; elt = next)
1777 unsigned int hash;
1778 rtx exp = elt->exp;
1779 enum machine_mode mode = elt->mode;
1781 next = elt->next_same_value;
1783 /* Remove old entry, make a new one in CLASS1's class.
1784 Don't do this for invalid entries as we cannot find their
1785 hash code (it also isn't necessary). */
1786 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1788 bool need_rehash = false;
1790 hash_arg_in_memory = 0;
1791 hash = HASH (exp, mode);
1793 if (REG_P (exp))
1795 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1796 delete_reg_equiv (REGNO (exp));
1799 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1800 remove_pseudo_from_table (exp, hash);
1801 else
1802 remove_from_table (elt, hash);
1804 if (insert_regs (exp, class1, 0) || need_rehash)
1806 rehash_using_reg (exp);
1807 hash = HASH (exp, mode);
1809 new_elt = insert (exp, class1, hash, mode);
1810 new_elt->in_memory = hash_arg_in_memory;
1815 /* Flush the entire hash table. */
1817 static void
1818 flush_hash_table (void)
1820 int i;
1821 struct table_elt *p;
1823 for (i = 0; i < HASH_SIZE; i++)
1824 for (p = table[i]; p; p = table[i])
1826 /* Note that invalidate can remove elements
1827 after P in the current hash chain. */
1828 if (REG_P (p->exp))
1829 invalidate (p->exp, VOIDmode);
1830 else
1831 remove_from_table (p, i);
1835 /* Function called for each rtx to check whether true dependence exist. */
1836 struct check_dependence_data
1838 enum machine_mode mode;
1839 rtx exp;
1840 rtx addr;
1843 static int
1844 check_dependence (rtx *x, void *data)
1846 struct check_dependence_data *d = (struct check_dependence_data *) data;
1847 if (*x && MEM_P (*x))
1848 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX,
1849 cse_rtx_varies_p);
1850 else
1851 return 0;
1854 /* Remove from the hash table, or mark as invalid, all expressions whose
1855 values could be altered by storing in X. X is a register, a subreg, or
1856 a memory reference with nonvarying address (because, when a memory
1857 reference with a varying address is stored in, all memory references are
1858 removed by invalidate_memory so specific invalidation is superfluous).
1859 FULL_MODE, if not VOIDmode, indicates that this much should be
1860 invalidated instead of just the amount indicated by the mode of X. This
1861 is only used for bitfield stores into memory.
1863 A nonvarying address may be just a register or just a symbol reference,
1864 or it may be either of those plus a numeric offset. */
1866 static void
1867 invalidate (rtx x, enum machine_mode full_mode)
1869 int i;
1870 struct table_elt *p;
1871 rtx addr;
1873 switch (GET_CODE (x))
1875 case REG:
1877 /* If X is a register, dependencies on its contents are recorded
1878 through the qty number mechanism. Just change the qty number of
1879 the register, mark it as invalid for expressions that refer to it,
1880 and remove it itself. */
1881 unsigned int regno = REGNO (x);
1882 unsigned int hash = HASH (x, GET_MODE (x));
1884 /* Remove REGNO from any quantity list it might be on and indicate
1885 that its value might have changed. If it is a pseudo, remove its
1886 entry from the hash table.
1888 For a hard register, we do the first two actions above for any
1889 additional hard registers corresponding to X. Then, if any of these
1890 registers are in the table, we must remove any REG entries that
1891 overlap these registers. */
1893 delete_reg_equiv (regno);
1894 REG_TICK (regno)++;
1895 SUBREG_TICKED (regno) = -1;
1897 if (regno >= FIRST_PSEUDO_REGISTER)
1898 remove_pseudo_from_table (x, hash);
1899 else
1901 HOST_WIDE_INT in_table
1902 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1903 unsigned int endregno = END_HARD_REGNO (x);
1904 unsigned int tregno, tendregno, rn;
1905 struct table_elt *p, *next;
1907 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1909 for (rn = regno + 1; rn < endregno; rn++)
1911 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1912 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1913 delete_reg_equiv (rn);
1914 REG_TICK (rn)++;
1915 SUBREG_TICKED (rn) = -1;
1918 if (in_table)
1919 for (hash = 0; hash < HASH_SIZE; hash++)
1920 for (p = table[hash]; p; p = next)
1922 next = p->next_same_hash;
1924 if (!REG_P (p->exp)
1925 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1926 continue;
1928 tregno = REGNO (p->exp);
1929 tendregno = END_HARD_REGNO (p->exp);
1930 if (tendregno > regno && tregno < endregno)
1931 remove_from_table (p, hash);
1935 return;
1937 case SUBREG:
1938 invalidate (SUBREG_REG (x), VOIDmode);
1939 return;
1941 case PARALLEL:
1942 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1943 invalidate (XVECEXP (x, 0, i), VOIDmode);
1944 return;
1946 case EXPR_LIST:
1947 /* This is part of a disjoint return value; extract the location in
1948 question ignoring the offset. */
1949 invalidate (XEXP (x, 0), VOIDmode);
1950 return;
1952 case MEM:
1953 addr = canon_rtx (get_addr (XEXP (x, 0)));
1954 /* Calculate the canonical version of X here so that
1955 true_dependence doesn't generate new RTL for X on each call. */
1956 x = canon_rtx (x);
1958 /* Remove all hash table elements that refer to overlapping pieces of
1959 memory. */
1960 if (full_mode == VOIDmode)
1961 full_mode = GET_MODE (x);
1963 for (i = 0; i < HASH_SIZE; i++)
1965 struct table_elt *next;
1967 for (p = table[i]; p; p = next)
1969 next = p->next_same_hash;
1970 if (p->in_memory)
1972 struct check_dependence_data d;
1974 /* Just canonicalize the expression once;
1975 otherwise each time we call invalidate
1976 true_dependence will canonicalize the
1977 expression again. */
1978 if (!p->canon_exp)
1979 p->canon_exp = canon_rtx (p->exp);
1980 d.exp = x;
1981 d.addr = addr;
1982 d.mode = full_mode;
1983 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1984 remove_from_table (p, i);
1988 return;
1990 default:
1991 gcc_unreachable ();
1995 /* Remove all expressions that refer to register REGNO,
1996 since they are already invalid, and we are about to
1997 mark that register valid again and don't want the old
1998 expressions to reappear as valid. */
2000 static void
2001 remove_invalid_refs (unsigned int regno)
2003 unsigned int i;
2004 struct table_elt *p, *next;
2006 for (i = 0; i < HASH_SIZE; i++)
2007 for (p = table[i]; p; p = next)
2009 next = p->next_same_hash;
2010 if (!REG_P (p->exp)
2011 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2012 remove_from_table (p, i);
2016 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2017 and mode MODE. */
2018 static void
2019 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2020 enum machine_mode mode)
2022 unsigned int i;
2023 struct table_elt *p, *next;
2024 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2026 for (i = 0; i < HASH_SIZE; i++)
2027 for (p = table[i]; p; p = next)
2029 rtx exp = p->exp;
2030 next = p->next_same_hash;
2032 if (!REG_P (exp)
2033 && (GET_CODE (exp) != SUBREG
2034 || !REG_P (SUBREG_REG (exp))
2035 || REGNO (SUBREG_REG (exp)) != regno
2036 || (((SUBREG_BYTE (exp)
2037 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2038 && SUBREG_BYTE (exp) <= end))
2039 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2040 remove_from_table (p, i);
2044 /* Recompute the hash codes of any valid entries in the hash table that
2045 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2047 This is called when we make a jump equivalence. */
2049 static void
2050 rehash_using_reg (rtx x)
2052 unsigned int i;
2053 struct table_elt *p, *next;
2054 unsigned hash;
2056 if (GET_CODE (x) == SUBREG)
2057 x = SUBREG_REG (x);
2059 /* If X is not a register or if the register is known not to be in any
2060 valid entries in the table, we have no work to do. */
2062 if (!REG_P (x)
2063 || REG_IN_TABLE (REGNO (x)) < 0
2064 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2065 return;
2067 /* Scan all hash chains looking for valid entries that mention X.
2068 If we find one and it is in the wrong hash chain, move it. */
2070 for (i = 0; i < HASH_SIZE; i++)
2071 for (p = table[i]; p; p = next)
2073 next = p->next_same_hash;
2074 if (reg_mentioned_p (x, p->exp)
2075 && exp_equiv_p (p->exp, p->exp, 1, false)
2076 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2078 if (p->next_same_hash)
2079 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2081 if (p->prev_same_hash)
2082 p->prev_same_hash->next_same_hash = p->next_same_hash;
2083 else
2084 table[i] = p->next_same_hash;
2086 p->next_same_hash = table[hash];
2087 p->prev_same_hash = 0;
2088 if (table[hash])
2089 table[hash]->prev_same_hash = p;
2090 table[hash] = p;
2095 /* Remove from the hash table any expression that is a call-clobbered
2096 register. Also update their TICK values. */
2098 static void
2099 invalidate_for_call (void)
2101 unsigned int regno, endregno;
2102 unsigned int i;
2103 unsigned hash;
2104 struct table_elt *p, *next;
2105 int in_table = 0;
2107 /* Go through all the hard registers. For each that is clobbered in
2108 a CALL_INSN, remove the register from quantity chains and update
2109 reg_tick if defined. Also see if any of these registers is currently
2110 in the table. */
2112 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2113 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2115 delete_reg_equiv (regno);
2116 if (REG_TICK (regno) >= 0)
2118 REG_TICK (regno)++;
2119 SUBREG_TICKED (regno) = -1;
2122 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2125 /* In the case where we have no call-clobbered hard registers in the
2126 table, we are done. Otherwise, scan the table and remove any
2127 entry that overlaps a call-clobbered register. */
2129 if (in_table)
2130 for (hash = 0; hash < HASH_SIZE; hash++)
2131 for (p = table[hash]; p; p = next)
2133 next = p->next_same_hash;
2135 if (!REG_P (p->exp)
2136 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2137 continue;
2139 regno = REGNO (p->exp);
2140 endregno = END_HARD_REGNO (p->exp);
2142 for (i = regno; i < endregno; i++)
2143 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2145 remove_from_table (p, hash);
2146 break;
2151 /* Given an expression X of type CONST,
2152 and ELT which is its table entry (or 0 if it
2153 is not in the hash table),
2154 return an alternate expression for X as a register plus integer.
2155 If none can be found, return 0. */
2157 static rtx
2158 use_related_value (rtx x, struct table_elt *elt)
2160 struct table_elt *relt = 0;
2161 struct table_elt *p, *q;
2162 HOST_WIDE_INT offset;
2164 /* First, is there anything related known?
2165 If we have a table element, we can tell from that.
2166 Otherwise, must look it up. */
2168 if (elt != 0 && elt->related_value != 0)
2169 relt = elt;
2170 else if (elt == 0 && GET_CODE (x) == CONST)
2172 rtx subexp = get_related_value (x);
2173 if (subexp != 0)
2174 relt = lookup (subexp,
2175 SAFE_HASH (subexp, GET_MODE (subexp)),
2176 GET_MODE (subexp));
2179 if (relt == 0)
2180 return 0;
2182 /* Search all related table entries for one that has an
2183 equivalent register. */
2185 p = relt;
2186 while (1)
2188 /* This loop is strange in that it is executed in two different cases.
2189 The first is when X is already in the table. Then it is searching
2190 the RELATED_VALUE list of X's class (RELT). The second case is when
2191 X is not in the table. Then RELT points to a class for the related
2192 value.
2194 Ensure that, whatever case we are in, that we ignore classes that have
2195 the same value as X. */
2197 if (rtx_equal_p (x, p->exp))
2198 q = 0;
2199 else
2200 for (q = p->first_same_value; q; q = q->next_same_value)
2201 if (REG_P (q->exp))
2202 break;
2204 if (q)
2205 break;
2207 p = p->related_value;
2209 /* We went all the way around, so there is nothing to be found.
2210 Alternatively, perhaps RELT was in the table for some other reason
2211 and it has no related values recorded. */
2212 if (p == relt || p == 0)
2213 break;
2216 if (q == 0)
2217 return 0;
2219 offset = (get_integer_term (x) - get_integer_term (p->exp));
2220 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2221 return plus_constant (q->exp, offset);
2225 /* Hash a string. Just add its bytes up. */
2226 static inline unsigned
2227 hash_rtx_string (const char *ps)
2229 unsigned hash = 0;
2230 const unsigned char *p = (const unsigned char *) ps;
2232 if (p)
2233 while (*p)
2234 hash += *p++;
2236 return hash;
2239 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2240 When the callback returns true, we continue with the new rtx. */
2242 unsigned
2243 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2244 int *do_not_record_p, int *hash_arg_in_memory_p,
2245 bool have_reg_qty, hash_rtx_callback_function cb)
2247 int i, j;
2248 unsigned hash = 0;
2249 enum rtx_code code;
2250 const char *fmt;
2251 enum machine_mode newmode;
2252 rtx newx;
2254 /* Used to turn recursion into iteration. We can't rely on GCC's
2255 tail-recursion elimination since we need to keep accumulating values
2256 in HASH. */
2257 repeat:
2258 if (x == 0)
2259 return hash;
2261 /* Invoke the callback first. */
2262 if (cb != NULL
2263 && ((*cb) (x, mode, &newx, &newmode)))
2265 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2266 hash_arg_in_memory_p, have_reg_qty, cb);
2267 return hash;
2270 code = GET_CODE (x);
2271 switch (code)
2273 case REG:
2275 unsigned int regno = REGNO (x);
2277 if (do_not_record_p && !reload_completed)
2279 /* On some machines, we can't record any non-fixed hard register,
2280 because extending its life will cause reload problems. We
2281 consider ap, fp, sp, gp to be fixed for this purpose.
2283 We also consider CCmode registers to be fixed for this purpose;
2284 failure to do so leads to failure to simplify 0<100 type of
2285 conditionals.
2287 On all machines, we can't record any global registers.
2288 Nor should we record any register that is in a small
2289 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2290 bool record;
2292 if (regno >= FIRST_PSEUDO_REGISTER)
2293 record = true;
2294 else if (x == frame_pointer_rtx
2295 || x == hard_frame_pointer_rtx
2296 || x == arg_pointer_rtx
2297 || x == stack_pointer_rtx
2298 || x == pic_offset_table_rtx)
2299 record = true;
2300 else if (global_regs[regno])
2301 record = false;
2302 else if (fixed_regs[regno])
2303 record = true;
2304 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2305 record = true;
2306 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2307 record = false;
2308 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2309 record = false;
2310 else
2311 record = true;
2313 if (!record)
2315 *do_not_record_p = 1;
2316 return 0;
2320 hash += ((unsigned int) REG << 7);
2321 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2322 return hash;
2325 /* We handle SUBREG of a REG specially because the underlying
2326 reg changes its hash value with every value change; we don't
2327 want to have to forget unrelated subregs when one subreg changes. */
2328 case SUBREG:
2330 if (REG_P (SUBREG_REG (x)))
2332 hash += (((unsigned int) SUBREG << 7)
2333 + REGNO (SUBREG_REG (x))
2334 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2335 return hash;
2337 break;
2340 case CONST_INT:
2341 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2342 + (unsigned int) INTVAL (x));
2343 return hash;
2345 case CONST_DOUBLE:
2346 /* This is like the general case, except that it only counts
2347 the integers representing the constant. */
2348 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2349 if (GET_MODE (x) != VOIDmode)
2350 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2351 else
2352 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2353 + (unsigned int) CONST_DOUBLE_HIGH (x));
2354 return hash;
2356 case CONST_FIXED:
2357 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2358 hash += fixed_hash (CONST_FIXED_VALUE (x));
2359 return hash;
2361 case CONST_VECTOR:
2363 int units;
2364 rtx elt;
2366 units = CONST_VECTOR_NUNITS (x);
2368 for (i = 0; i < units; ++i)
2370 elt = CONST_VECTOR_ELT (x, i);
2371 hash += hash_rtx_cb (elt, GET_MODE (elt),
2372 do_not_record_p, hash_arg_in_memory_p,
2373 have_reg_qty, cb);
2376 return hash;
2379 /* Assume there is only one rtx object for any given label. */
2380 case LABEL_REF:
2381 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2382 differences and differences between each stage's debugging dumps. */
2383 hash += (((unsigned int) LABEL_REF << 7)
2384 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2385 return hash;
2387 case SYMBOL_REF:
2389 /* Don't hash on the symbol's address to avoid bootstrap differences.
2390 Different hash values may cause expressions to be recorded in
2391 different orders and thus different registers to be used in the
2392 final assembler. This also avoids differences in the dump files
2393 between various stages. */
2394 unsigned int h = 0;
2395 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2397 while (*p)
2398 h += (h << 7) + *p++; /* ??? revisit */
2400 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2401 return hash;
2404 case MEM:
2405 /* We don't record if marked volatile or if BLKmode since we don't
2406 know the size of the move. */
2407 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2409 *do_not_record_p = 1;
2410 return 0;
2412 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2413 *hash_arg_in_memory_p = 1;
2415 /* Now that we have already found this special case,
2416 might as well speed it up as much as possible. */
2417 hash += (unsigned) MEM;
2418 x = XEXP (x, 0);
2419 goto repeat;
2421 case USE:
2422 /* A USE that mentions non-volatile memory needs special
2423 handling since the MEM may be BLKmode which normally
2424 prevents an entry from being made. Pure calls are
2425 marked by a USE which mentions BLKmode memory.
2426 See calls.c:emit_call_1. */
2427 if (MEM_P (XEXP (x, 0))
2428 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2430 hash += (unsigned) USE;
2431 x = XEXP (x, 0);
2433 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2434 *hash_arg_in_memory_p = 1;
2436 /* Now that we have already found this special case,
2437 might as well speed it up as much as possible. */
2438 hash += (unsigned) MEM;
2439 x = XEXP (x, 0);
2440 goto repeat;
2442 break;
2444 case PRE_DEC:
2445 case PRE_INC:
2446 case POST_DEC:
2447 case POST_INC:
2448 case PRE_MODIFY:
2449 case POST_MODIFY:
2450 case PC:
2451 case CC0:
2452 case CALL:
2453 case UNSPEC_VOLATILE:
2454 if (do_not_record_p) {
2455 *do_not_record_p = 1;
2456 return 0;
2458 else
2459 return hash;
2460 break;
2462 case ASM_OPERANDS:
2463 if (do_not_record_p && MEM_VOLATILE_P (x))
2465 *do_not_record_p = 1;
2466 return 0;
2468 else
2470 /* We don't want to take the filename and line into account. */
2471 hash += (unsigned) code + (unsigned) GET_MODE (x)
2472 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2473 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2474 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2476 if (ASM_OPERANDS_INPUT_LENGTH (x))
2478 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2480 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2481 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2482 do_not_record_p, hash_arg_in_memory_p,
2483 have_reg_qty, cb)
2484 + hash_rtx_string
2485 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2488 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2489 x = ASM_OPERANDS_INPUT (x, 0);
2490 mode = GET_MODE (x);
2491 goto repeat;
2494 return hash;
2496 break;
2498 default:
2499 break;
2502 i = GET_RTX_LENGTH (code) - 1;
2503 hash += (unsigned) code + (unsigned) GET_MODE (x);
2504 fmt = GET_RTX_FORMAT (code);
2505 for (; i >= 0; i--)
2507 switch (fmt[i])
2509 case 'e':
2510 /* If we are about to do the last recursive call
2511 needed at this level, change it into iteration.
2512 This function is called enough to be worth it. */
2513 if (i == 0)
2515 x = XEXP (x, i);
2516 goto repeat;
2519 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2520 hash_arg_in_memory_p,
2521 have_reg_qty, cb);
2522 break;
2524 case 'E':
2525 for (j = 0; j < XVECLEN (x, i); j++)
2526 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2527 hash_arg_in_memory_p,
2528 have_reg_qty, cb);
2529 break;
2531 case 's':
2532 hash += hash_rtx_string (XSTR (x, i));
2533 break;
2535 case 'i':
2536 hash += (unsigned int) XINT (x, i);
2537 break;
2539 case '0': case 't':
2540 /* Unused. */
2541 break;
2543 default:
2544 gcc_unreachable ();
2548 return hash;
2551 /* Hash an rtx. We are careful to make sure the value is never negative.
2552 Equivalent registers hash identically.
2553 MODE is used in hashing for CONST_INTs only;
2554 otherwise the mode of X is used.
2556 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2558 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2559 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2561 Note that cse_insn knows that the hash code of a MEM expression
2562 is just (int) MEM plus the hash code of the address. */
2564 unsigned
2565 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2566 int *hash_arg_in_memory_p, bool have_reg_qty)
2568 return hash_rtx_cb (x, mode, do_not_record_p,
2569 hash_arg_in_memory_p, have_reg_qty, NULL);
2572 /* Hash an rtx X for cse via hash_rtx.
2573 Stores 1 in do_not_record if any subexpression is volatile.
2574 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2575 does not have the RTX_UNCHANGING_P bit set. */
2577 static inline unsigned
2578 canon_hash (rtx x, enum machine_mode mode)
2580 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2583 /* Like canon_hash but with no side effects, i.e. do_not_record
2584 and hash_arg_in_memory are not changed. */
2586 static inline unsigned
2587 safe_hash (rtx x, enum machine_mode mode)
2589 int dummy_do_not_record;
2590 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2593 /* Return 1 iff X and Y would canonicalize into the same thing,
2594 without actually constructing the canonicalization of either one.
2595 If VALIDATE is nonzero,
2596 we assume X is an expression being processed from the rtl
2597 and Y was found in the hash table. We check register refs
2598 in Y for being marked as valid.
2600 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2603 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2605 int i, j;
2606 enum rtx_code code;
2607 const char *fmt;
2609 /* Note: it is incorrect to assume an expression is equivalent to itself
2610 if VALIDATE is nonzero. */
2611 if (x == y && !validate)
2612 return 1;
2614 if (x == 0 || y == 0)
2615 return x == y;
2617 code = GET_CODE (x);
2618 if (code != GET_CODE (y))
2619 return 0;
2621 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2622 if (GET_MODE (x) != GET_MODE (y))
2623 return 0;
2625 /* MEMs refering to different address space are not equivalent. */
2626 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2627 return 0;
2629 switch (code)
2631 case PC:
2632 case CC0:
2633 case CONST_INT:
2634 case CONST_DOUBLE:
2635 case CONST_FIXED:
2636 return x == y;
2638 case LABEL_REF:
2639 return XEXP (x, 0) == XEXP (y, 0);
2641 case SYMBOL_REF:
2642 return XSTR (x, 0) == XSTR (y, 0);
2644 case REG:
2645 if (for_gcse)
2646 return REGNO (x) == REGNO (y);
2647 else
2649 unsigned int regno = REGNO (y);
2650 unsigned int i;
2651 unsigned int endregno = END_REGNO (y);
2653 /* If the quantities are not the same, the expressions are not
2654 equivalent. If there are and we are not to validate, they
2655 are equivalent. Otherwise, ensure all regs are up-to-date. */
2657 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2658 return 0;
2660 if (! validate)
2661 return 1;
2663 for (i = regno; i < endregno; i++)
2664 if (REG_IN_TABLE (i) != REG_TICK (i))
2665 return 0;
2667 return 1;
2670 case MEM:
2671 if (for_gcse)
2673 /* Can't merge two expressions in different alias sets, since we
2674 can decide that the expression is transparent in a block when
2675 it isn't, due to it being set with the different alias set. */
2676 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
2677 return 0;
2679 /* A volatile mem should not be considered equivalent to any
2680 other. */
2681 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2682 return 0;
2684 break;
2686 /* For commutative operations, check both orders. */
2687 case PLUS:
2688 case MULT:
2689 case AND:
2690 case IOR:
2691 case XOR:
2692 case NE:
2693 case EQ:
2694 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2695 validate, for_gcse)
2696 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2697 validate, for_gcse))
2698 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2699 validate, for_gcse)
2700 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2701 validate, for_gcse)));
2703 case ASM_OPERANDS:
2704 /* We don't use the generic code below because we want to
2705 disregard filename and line numbers. */
2707 /* A volatile asm isn't equivalent to any other. */
2708 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2709 return 0;
2711 if (GET_MODE (x) != GET_MODE (y)
2712 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2713 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2714 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2715 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2716 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2717 return 0;
2719 if (ASM_OPERANDS_INPUT_LENGTH (x))
2721 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2722 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2723 ASM_OPERANDS_INPUT (y, i),
2724 validate, for_gcse)
2725 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2726 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2727 return 0;
2730 return 1;
2732 default:
2733 break;
2736 /* Compare the elements. If any pair of corresponding elements
2737 fail to match, return 0 for the whole thing. */
2739 fmt = GET_RTX_FORMAT (code);
2740 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2742 switch (fmt[i])
2744 case 'e':
2745 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2746 validate, for_gcse))
2747 return 0;
2748 break;
2750 case 'E':
2751 if (XVECLEN (x, i) != XVECLEN (y, i))
2752 return 0;
2753 for (j = 0; j < XVECLEN (x, i); j++)
2754 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2755 validate, for_gcse))
2756 return 0;
2757 break;
2759 case 's':
2760 if (strcmp (XSTR (x, i), XSTR (y, i)))
2761 return 0;
2762 break;
2764 case 'i':
2765 if (XINT (x, i) != XINT (y, i))
2766 return 0;
2767 break;
2769 case 'w':
2770 if (XWINT (x, i) != XWINT (y, i))
2771 return 0;
2772 break;
2774 case '0':
2775 case 't':
2776 break;
2778 default:
2779 gcc_unreachable ();
2783 return 1;
2786 /* Return 1 if X has a value that can vary even between two
2787 executions of the program. 0 means X can be compared reliably
2788 against certain constants or near-constants. */
2790 static bool
2791 cse_rtx_varies_p (const_rtx x, bool from_alias)
2793 /* We need not check for X and the equivalence class being of the same
2794 mode because if X is equivalent to a constant in some mode, it
2795 doesn't vary in any mode. */
2797 if (REG_P (x)
2798 && REGNO_QTY_VALID_P (REGNO (x)))
2800 int x_q = REG_QTY (REGNO (x));
2801 struct qty_table_elem *x_ent = &qty_table[x_q];
2803 if (GET_MODE (x) == x_ent->mode
2804 && x_ent->const_rtx != NULL_RTX)
2805 return 0;
2808 if (GET_CODE (x) == PLUS
2809 && CONST_INT_P (XEXP (x, 1))
2810 && REG_P (XEXP (x, 0))
2811 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2813 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2814 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2816 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2817 && x0_ent->const_rtx != NULL_RTX)
2818 return 0;
2821 /* This can happen as the result of virtual register instantiation, if
2822 the initial constant is too large to be a valid address. This gives
2823 us a three instruction sequence, load large offset into a register,
2824 load fp minus a constant into a register, then a MEM which is the
2825 sum of the two `constant' registers. */
2826 if (GET_CODE (x) == PLUS
2827 && REG_P (XEXP (x, 0))
2828 && REG_P (XEXP (x, 1))
2829 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2830 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2832 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2833 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2834 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2835 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2837 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2838 && x0_ent->const_rtx != NULL_RTX
2839 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2840 && x1_ent->const_rtx != NULL_RTX)
2841 return 0;
2844 return rtx_varies_p (x, from_alias);
2847 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2848 the result if necessary. INSN is as for canon_reg. */
2850 static void
2851 validate_canon_reg (rtx *xloc, rtx insn)
2853 if (*xloc)
2855 rtx new_rtx = canon_reg (*xloc, insn);
2857 /* If replacing pseudo with hard reg or vice versa, ensure the
2858 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2859 gcc_assert (insn && new_rtx);
2860 validate_change (insn, xloc, new_rtx, 1);
2864 /* Canonicalize an expression:
2865 replace each register reference inside it
2866 with the "oldest" equivalent register.
2868 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2869 after we make our substitution. The calls are made with IN_GROUP nonzero
2870 so apply_change_group must be called upon the outermost return from this
2871 function (unless INSN is zero). The result of apply_change_group can
2872 generally be discarded since the changes we are making are optional. */
2874 static rtx
2875 canon_reg (rtx x, rtx insn)
2877 int i;
2878 enum rtx_code code;
2879 const char *fmt;
2881 if (x == 0)
2882 return x;
2884 code = GET_CODE (x);
2885 switch (code)
2887 case PC:
2888 case CC0:
2889 case CONST:
2890 case CONST_INT:
2891 case CONST_DOUBLE:
2892 case CONST_FIXED:
2893 case CONST_VECTOR:
2894 case SYMBOL_REF:
2895 case LABEL_REF:
2896 case ADDR_VEC:
2897 case ADDR_DIFF_VEC:
2898 return x;
2900 case REG:
2902 int first;
2903 int q;
2904 struct qty_table_elem *ent;
2906 /* Never replace a hard reg, because hard regs can appear
2907 in more than one machine mode, and we must preserve the mode
2908 of each occurrence. Also, some hard regs appear in
2909 MEMs that are shared and mustn't be altered. Don't try to
2910 replace any reg that maps to a reg of class NO_REGS. */
2911 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2912 || ! REGNO_QTY_VALID_P (REGNO (x)))
2913 return x;
2915 q = REG_QTY (REGNO (x));
2916 ent = &qty_table[q];
2917 first = ent->first_reg;
2918 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2919 : REGNO_REG_CLASS (first) == NO_REGS ? x
2920 : gen_rtx_REG (ent->mode, first));
2923 default:
2924 break;
2927 fmt = GET_RTX_FORMAT (code);
2928 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2930 int j;
2932 if (fmt[i] == 'e')
2933 validate_canon_reg (&XEXP (x, i), insn);
2934 else if (fmt[i] == 'E')
2935 for (j = 0; j < XVECLEN (x, i); j++)
2936 validate_canon_reg (&XVECEXP (x, i, j), insn);
2939 return x;
2942 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2943 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2944 what values are being compared.
2946 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2947 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2948 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2949 compared to produce cc0.
2951 The return value is the comparison operator and is either the code of
2952 A or the code corresponding to the inverse of the comparison. */
2954 static enum rtx_code
2955 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2956 enum machine_mode *pmode1, enum machine_mode *pmode2)
2958 rtx arg1, arg2;
2960 arg1 = *parg1, arg2 = *parg2;
2962 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2964 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2966 /* Set nonzero when we find something of interest. */
2967 rtx x = 0;
2968 int reverse_code = 0;
2969 struct table_elt *p = 0;
2971 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2972 On machines with CC0, this is the only case that can occur, since
2973 fold_rtx will return the COMPARE or item being compared with zero
2974 when given CC0. */
2976 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2977 x = arg1;
2979 /* If ARG1 is a comparison operator and CODE is testing for
2980 STORE_FLAG_VALUE, get the inner arguments. */
2982 else if (COMPARISON_P (arg1))
2984 #ifdef FLOAT_STORE_FLAG_VALUE
2985 REAL_VALUE_TYPE fsfv;
2986 #endif
2988 if (code == NE
2989 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2990 && code == LT && STORE_FLAG_VALUE == -1)
2991 #ifdef FLOAT_STORE_FLAG_VALUE
2992 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2993 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2994 REAL_VALUE_NEGATIVE (fsfv)))
2995 #endif
2997 x = arg1;
2998 else if (code == EQ
2999 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3000 && code == GE && STORE_FLAG_VALUE == -1)
3001 #ifdef FLOAT_STORE_FLAG_VALUE
3002 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3003 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3004 REAL_VALUE_NEGATIVE (fsfv)))
3005 #endif
3007 x = arg1, reverse_code = 1;
3010 /* ??? We could also check for
3012 (ne (and (eq (...) (const_int 1))) (const_int 0))
3014 and related forms, but let's wait until we see them occurring. */
3016 if (x == 0)
3017 /* Look up ARG1 in the hash table and see if it has an equivalence
3018 that lets us see what is being compared. */
3019 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3020 if (p)
3022 p = p->first_same_value;
3024 /* If what we compare is already known to be constant, that is as
3025 good as it gets.
3026 We need to break the loop in this case, because otherwise we
3027 can have an infinite loop when looking at a reg that is known
3028 to be a constant which is the same as a comparison of a reg
3029 against zero which appears later in the insn stream, which in
3030 turn is constant and the same as the comparison of the first reg
3031 against zero... */
3032 if (p->is_const)
3033 break;
3036 for (; p; p = p->next_same_value)
3038 enum machine_mode inner_mode = GET_MODE (p->exp);
3039 #ifdef FLOAT_STORE_FLAG_VALUE
3040 REAL_VALUE_TYPE fsfv;
3041 #endif
3043 /* If the entry isn't valid, skip it. */
3044 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3045 continue;
3047 if (GET_CODE (p->exp) == COMPARE
3048 /* Another possibility is that this machine has a compare insn
3049 that includes the comparison code. In that case, ARG1 would
3050 be equivalent to a comparison operation that would set ARG1 to
3051 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3052 ORIG_CODE is the actual comparison being done; if it is an EQ,
3053 we must reverse ORIG_CODE. On machine with a negative value
3054 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3055 || ((code == NE
3056 || (code == LT
3057 && GET_MODE_CLASS (inner_mode) == MODE_INT
3058 && (GET_MODE_BITSIZE (inner_mode)
3059 <= HOST_BITS_PER_WIDE_INT)
3060 && (STORE_FLAG_VALUE
3061 & ((HOST_WIDE_INT) 1
3062 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3063 #ifdef FLOAT_STORE_FLAG_VALUE
3064 || (code == LT
3065 && SCALAR_FLOAT_MODE_P (inner_mode)
3066 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3067 REAL_VALUE_NEGATIVE (fsfv)))
3068 #endif
3070 && COMPARISON_P (p->exp)))
3072 x = p->exp;
3073 break;
3075 else if ((code == EQ
3076 || (code == GE
3077 && GET_MODE_CLASS (inner_mode) == MODE_INT
3078 && (GET_MODE_BITSIZE (inner_mode)
3079 <= HOST_BITS_PER_WIDE_INT)
3080 && (STORE_FLAG_VALUE
3081 & ((HOST_WIDE_INT) 1
3082 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3083 #ifdef FLOAT_STORE_FLAG_VALUE
3084 || (code == GE
3085 && SCALAR_FLOAT_MODE_P (inner_mode)
3086 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3087 REAL_VALUE_NEGATIVE (fsfv)))
3088 #endif
3090 && COMPARISON_P (p->exp))
3092 reverse_code = 1;
3093 x = p->exp;
3094 break;
3097 /* If this non-trapping address, e.g. fp + constant, the
3098 equivalent is a better operand since it may let us predict
3099 the value of the comparison. */
3100 else if (!rtx_addr_can_trap_p (p->exp))
3102 arg1 = p->exp;
3103 continue;
3107 /* If we didn't find a useful equivalence for ARG1, we are done.
3108 Otherwise, set up for the next iteration. */
3109 if (x == 0)
3110 break;
3112 /* If we need to reverse the comparison, make sure that that is
3113 possible -- we can't necessarily infer the value of GE from LT
3114 with floating-point operands. */
3115 if (reverse_code)
3117 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3118 if (reversed == UNKNOWN)
3119 break;
3120 else
3121 code = reversed;
3123 else if (COMPARISON_P (x))
3124 code = GET_CODE (x);
3125 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3128 /* Return our results. Return the modes from before fold_rtx
3129 because fold_rtx might produce const_int, and then it's too late. */
3130 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3131 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3133 return code;
3136 /* If X is a nontrivial arithmetic operation on an argument for which
3137 a constant value can be determined, return the result of operating
3138 on that value, as a constant. Otherwise, return X, possibly with
3139 one or more operands changed to a forward-propagated constant.
3141 If X is a register whose contents are known, we do NOT return
3142 those contents here; equiv_constant is called to perform that task.
3143 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3145 INSN is the insn that we may be modifying. If it is 0, make a copy
3146 of X before modifying it. */
3148 static rtx
3149 fold_rtx (rtx x, rtx insn)
3151 enum rtx_code code;
3152 enum machine_mode mode;
3153 const char *fmt;
3154 int i;
3155 rtx new_rtx = 0;
3156 int changed = 0;
3158 /* Operands of X. */
3159 rtx folded_arg0;
3160 rtx folded_arg1;
3162 /* Constant equivalents of first three operands of X;
3163 0 when no such equivalent is known. */
3164 rtx const_arg0;
3165 rtx const_arg1;
3166 rtx const_arg2;
3168 /* The mode of the first operand of X. We need this for sign and zero
3169 extends. */
3170 enum machine_mode mode_arg0;
3172 if (x == 0)
3173 return x;
3175 /* Try to perform some initial simplifications on X. */
3176 code = GET_CODE (x);
3177 switch (code)
3179 case MEM:
3180 case SUBREG:
3181 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3182 return new_rtx;
3183 return x;
3185 case CONST:
3186 case CONST_INT:
3187 case CONST_DOUBLE:
3188 case CONST_FIXED:
3189 case CONST_VECTOR:
3190 case SYMBOL_REF:
3191 case LABEL_REF:
3192 case REG:
3193 case PC:
3194 /* No use simplifying an EXPR_LIST
3195 since they are used only for lists of args
3196 in a function call's REG_EQUAL note. */
3197 case EXPR_LIST:
3198 return x;
3200 #ifdef HAVE_cc0
3201 case CC0:
3202 return prev_insn_cc0;
3203 #endif
3205 case ASM_OPERANDS:
3206 if (insn)
3208 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3209 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3210 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3212 return x;
3214 #ifdef NO_FUNCTION_CSE
3215 case CALL:
3216 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3217 return x;
3218 break;
3219 #endif
3221 /* Anything else goes through the loop below. */
3222 default:
3223 break;
3226 mode = GET_MODE (x);
3227 const_arg0 = 0;
3228 const_arg1 = 0;
3229 const_arg2 = 0;
3230 mode_arg0 = VOIDmode;
3232 /* Try folding our operands.
3233 Then see which ones have constant values known. */
3235 fmt = GET_RTX_FORMAT (code);
3236 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3237 if (fmt[i] == 'e')
3239 rtx folded_arg = XEXP (x, i), const_arg;
3240 enum machine_mode mode_arg = GET_MODE (folded_arg);
3242 switch (GET_CODE (folded_arg))
3244 case MEM:
3245 case REG:
3246 case SUBREG:
3247 const_arg = equiv_constant (folded_arg);
3248 break;
3250 case CONST:
3251 case CONST_INT:
3252 case SYMBOL_REF:
3253 case LABEL_REF:
3254 case CONST_DOUBLE:
3255 case CONST_FIXED:
3256 case CONST_VECTOR:
3257 const_arg = folded_arg;
3258 break;
3260 #ifdef HAVE_cc0
3261 case CC0:
3262 folded_arg = prev_insn_cc0;
3263 mode_arg = prev_insn_cc0_mode;
3264 const_arg = equiv_constant (folded_arg);
3265 break;
3266 #endif
3268 default:
3269 folded_arg = fold_rtx (folded_arg, insn);
3270 const_arg = equiv_constant (folded_arg);
3271 break;
3274 /* For the first three operands, see if the operand
3275 is constant or equivalent to a constant. */
3276 switch (i)
3278 case 0:
3279 folded_arg0 = folded_arg;
3280 const_arg0 = const_arg;
3281 mode_arg0 = mode_arg;
3282 break;
3283 case 1:
3284 folded_arg1 = folded_arg;
3285 const_arg1 = const_arg;
3286 break;
3287 case 2:
3288 const_arg2 = const_arg;
3289 break;
3292 /* Pick the least expensive of the argument and an equivalent constant
3293 argument. */
3294 if (const_arg != 0
3295 && const_arg != folded_arg
3296 && COST_IN (const_arg, code) <= COST_IN (folded_arg, code)
3298 /* It's not safe to substitute the operand of a conversion
3299 operator with a constant, as the conversion's identity
3300 depends upon the mode of its operand. This optimization
3301 is handled by the call to simplify_unary_operation. */
3302 && (GET_RTX_CLASS (code) != RTX_UNARY
3303 || GET_MODE (const_arg) == mode_arg0
3304 || (code != ZERO_EXTEND
3305 && code != SIGN_EXTEND
3306 && code != TRUNCATE
3307 && code != FLOAT_TRUNCATE
3308 && code != FLOAT_EXTEND
3309 && code != FLOAT
3310 && code != FIX
3311 && code != UNSIGNED_FLOAT
3312 && code != UNSIGNED_FIX)))
3313 folded_arg = const_arg;
3315 if (folded_arg == XEXP (x, i))
3316 continue;
3318 if (insn == NULL_RTX && !changed)
3319 x = copy_rtx (x);
3320 changed = 1;
3321 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3324 if (changed)
3326 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3327 consistent with the order in X. */
3328 if (canonicalize_change_group (insn, x))
3330 rtx tem;
3331 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3332 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3335 apply_change_group ();
3338 /* If X is an arithmetic operation, see if we can simplify it. */
3340 switch (GET_RTX_CLASS (code))
3342 case RTX_UNARY:
3344 /* We can't simplify extension ops unless we know the
3345 original mode. */
3346 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3347 && mode_arg0 == VOIDmode)
3348 break;
3350 new_rtx = simplify_unary_operation (code, mode,
3351 const_arg0 ? const_arg0 : folded_arg0,
3352 mode_arg0);
3354 break;
3356 case RTX_COMPARE:
3357 case RTX_COMM_COMPARE:
3358 /* See what items are actually being compared and set FOLDED_ARG[01]
3359 to those values and CODE to the actual comparison code. If any are
3360 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3361 do anything if both operands are already known to be constant. */
3363 /* ??? Vector mode comparisons are not supported yet. */
3364 if (VECTOR_MODE_P (mode))
3365 break;
3367 if (const_arg0 == 0 || const_arg1 == 0)
3369 struct table_elt *p0, *p1;
3370 rtx true_rtx, false_rtx;
3371 enum machine_mode mode_arg1;
3373 if (SCALAR_FLOAT_MODE_P (mode))
3375 #ifdef FLOAT_STORE_FLAG_VALUE
3376 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3377 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3378 #else
3379 true_rtx = NULL_RTX;
3380 #endif
3381 false_rtx = CONST0_RTX (mode);
3383 else
3385 true_rtx = const_true_rtx;
3386 false_rtx = const0_rtx;
3389 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3390 &mode_arg0, &mode_arg1);
3392 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3393 what kinds of things are being compared, so we can't do
3394 anything with this comparison. */
3396 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3397 break;
3399 const_arg0 = equiv_constant (folded_arg0);
3400 const_arg1 = equiv_constant (folded_arg1);
3402 /* If we do not now have two constants being compared, see
3403 if we can nevertheless deduce some things about the
3404 comparison. */
3405 if (const_arg0 == 0 || const_arg1 == 0)
3407 if (const_arg1 != NULL)
3409 rtx cheapest_simplification;
3410 int cheapest_cost;
3411 rtx simp_result;
3412 struct table_elt *p;
3414 /* See if we can find an equivalent of folded_arg0
3415 that gets us a cheaper expression, possibly a
3416 constant through simplifications. */
3417 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3418 mode_arg0);
3420 if (p != NULL)
3422 cheapest_simplification = x;
3423 cheapest_cost = COST (x);
3425 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3427 int cost;
3429 /* If the entry isn't valid, skip it. */
3430 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3431 continue;
3433 /* Try to simplify using this equivalence. */
3434 simp_result
3435 = simplify_relational_operation (code, mode,
3436 mode_arg0,
3437 p->exp,
3438 const_arg1);
3440 if (simp_result == NULL)
3441 continue;
3443 cost = COST (simp_result);
3444 if (cost < cheapest_cost)
3446 cheapest_cost = cost;
3447 cheapest_simplification = simp_result;
3451 /* If we have a cheaper expression now, use that
3452 and try folding it further, from the top. */
3453 if (cheapest_simplification != x)
3454 return fold_rtx (copy_rtx (cheapest_simplification),
3455 insn);
3459 /* See if the two operands are the same. */
3461 if ((REG_P (folded_arg0)
3462 && REG_P (folded_arg1)
3463 && (REG_QTY (REGNO (folded_arg0))
3464 == REG_QTY (REGNO (folded_arg1))))
3465 || ((p0 = lookup (folded_arg0,
3466 SAFE_HASH (folded_arg0, mode_arg0),
3467 mode_arg0))
3468 && (p1 = lookup (folded_arg1,
3469 SAFE_HASH (folded_arg1, mode_arg0),
3470 mode_arg0))
3471 && p0->first_same_value == p1->first_same_value))
3472 folded_arg1 = folded_arg0;
3474 /* If FOLDED_ARG0 is a register, see if the comparison we are
3475 doing now is either the same as we did before or the reverse
3476 (we only check the reverse if not floating-point). */
3477 else if (REG_P (folded_arg0))
3479 int qty = REG_QTY (REGNO (folded_arg0));
3481 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3483 struct qty_table_elem *ent = &qty_table[qty];
3485 if ((comparison_dominates_p (ent->comparison_code, code)
3486 || (! FLOAT_MODE_P (mode_arg0)
3487 && comparison_dominates_p (ent->comparison_code,
3488 reverse_condition (code))))
3489 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3490 || (const_arg1
3491 && rtx_equal_p (ent->comparison_const,
3492 const_arg1))
3493 || (REG_P (folded_arg1)
3494 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3496 if (comparison_dominates_p (ent->comparison_code, code))
3498 if (true_rtx)
3499 return true_rtx;
3500 else
3501 break;
3503 else
3504 return false_rtx;
3511 /* If we are comparing against zero, see if the first operand is
3512 equivalent to an IOR with a constant. If so, we may be able to
3513 determine the result of this comparison. */
3514 if (const_arg1 == const0_rtx && !const_arg0)
3516 rtx y = lookup_as_function (folded_arg0, IOR);
3517 rtx inner_const;
3519 if (y != 0
3520 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3521 && CONST_INT_P (inner_const)
3522 && INTVAL (inner_const) != 0)
3523 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3527 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3528 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3529 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3531 break;
3533 case RTX_BIN_ARITH:
3534 case RTX_COMM_ARITH:
3535 switch (code)
3537 case PLUS:
3538 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3539 with that LABEL_REF as its second operand. If so, the result is
3540 the first operand of that MINUS. This handles switches with an
3541 ADDR_DIFF_VEC table. */
3542 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3544 rtx y
3545 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3546 : lookup_as_function (folded_arg0, MINUS);
3548 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3549 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3550 return XEXP (y, 0);
3552 /* Now try for a CONST of a MINUS like the above. */
3553 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3554 : lookup_as_function (folded_arg0, CONST))) != 0
3555 && GET_CODE (XEXP (y, 0)) == MINUS
3556 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3557 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3558 return XEXP (XEXP (y, 0), 0);
3561 /* Likewise if the operands are in the other order. */
3562 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3564 rtx y
3565 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3566 : lookup_as_function (folded_arg1, MINUS);
3568 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3569 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3570 return XEXP (y, 0);
3572 /* Now try for a CONST of a MINUS like the above. */
3573 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3574 : lookup_as_function (folded_arg1, CONST))) != 0
3575 && GET_CODE (XEXP (y, 0)) == MINUS
3576 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3577 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3578 return XEXP (XEXP (y, 0), 0);
3581 /* If second operand is a register equivalent to a negative
3582 CONST_INT, see if we can find a register equivalent to the
3583 positive constant. Make a MINUS if so. Don't do this for
3584 a non-negative constant since we might then alternate between
3585 choosing positive and negative constants. Having the positive
3586 constant previously-used is the more common case. Be sure
3587 the resulting constant is non-negative; if const_arg1 were
3588 the smallest negative number this would overflow: depending
3589 on the mode, this would either just be the same value (and
3590 hence not save anything) or be incorrect. */
3591 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3592 && INTVAL (const_arg1) < 0
3593 /* This used to test
3595 -INTVAL (const_arg1) >= 0
3597 But The Sun V5.0 compilers mis-compiled that test. So
3598 instead we test for the problematic value in a more direct
3599 manner and hope the Sun compilers get it correct. */
3600 && INTVAL (const_arg1) !=
3601 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3602 && REG_P (folded_arg1))
3604 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3605 struct table_elt *p
3606 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3608 if (p)
3609 for (p = p->first_same_value; p; p = p->next_same_value)
3610 if (REG_P (p->exp))
3611 return simplify_gen_binary (MINUS, mode, folded_arg0,
3612 canon_reg (p->exp, NULL_RTX));
3614 goto from_plus;
3616 case MINUS:
3617 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3618 If so, produce (PLUS Z C2-C). */
3619 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3621 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3622 if (y && CONST_INT_P (XEXP (y, 1)))
3623 return fold_rtx (plus_constant (copy_rtx (y),
3624 -INTVAL (const_arg1)),
3625 NULL_RTX);
3628 /* Fall through. */
3630 from_plus:
3631 case SMIN: case SMAX: case UMIN: case UMAX:
3632 case IOR: case AND: case XOR:
3633 case MULT:
3634 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3635 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3636 is known to be of similar form, we may be able to replace the
3637 operation with a combined operation. This may eliminate the
3638 intermediate operation if every use is simplified in this way.
3639 Note that the similar optimization done by combine.c only works
3640 if the intermediate operation's result has only one reference. */
3642 if (REG_P (folded_arg0)
3643 && const_arg1 && CONST_INT_P (const_arg1))
3645 int is_shift
3646 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3647 rtx y, inner_const, new_const;
3648 rtx canon_const_arg1 = const_arg1;
3649 enum rtx_code associate_code;
3651 if (is_shift
3652 && (INTVAL (const_arg1) >= GET_MODE_BITSIZE (mode)
3653 || INTVAL (const_arg1) < 0))
3655 if (SHIFT_COUNT_TRUNCATED)
3656 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3657 & (GET_MODE_BITSIZE (mode)
3658 - 1));
3659 else
3660 break;
3663 y = lookup_as_function (folded_arg0, code);
3664 if (y == 0)
3665 break;
3667 /* If we have compiled a statement like
3668 "if (x == (x & mask1))", and now are looking at
3669 "x & mask2", we will have a case where the first operand
3670 of Y is the same as our first operand. Unless we detect
3671 this case, an infinite loop will result. */
3672 if (XEXP (y, 0) == folded_arg0)
3673 break;
3675 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3676 if (!inner_const || !CONST_INT_P (inner_const))
3677 break;
3679 /* Don't associate these operations if they are a PLUS with the
3680 same constant and it is a power of two. These might be doable
3681 with a pre- or post-increment. Similarly for two subtracts of
3682 identical powers of two with post decrement. */
3684 if (code == PLUS && const_arg1 == inner_const
3685 && ((HAVE_PRE_INCREMENT
3686 && exact_log2 (INTVAL (const_arg1)) >= 0)
3687 || (HAVE_POST_INCREMENT
3688 && exact_log2 (INTVAL (const_arg1)) >= 0)
3689 || (HAVE_PRE_DECREMENT
3690 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3691 || (HAVE_POST_DECREMENT
3692 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3693 break;
3695 /* ??? Vector mode shifts by scalar
3696 shift operand are not supported yet. */
3697 if (is_shift && VECTOR_MODE_P (mode))
3698 break;
3700 if (is_shift
3701 && (INTVAL (inner_const) >= GET_MODE_BITSIZE (mode)
3702 || INTVAL (inner_const) < 0))
3704 if (SHIFT_COUNT_TRUNCATED)
3705 inner_const = GEN_INT (INTVAL (inner_const)
3706 & (GET_MODE_BITSIZE (mode) - 1));
3707 else
3708 break;
3711 /* Compute the code used to compose the constants. For example,
3712 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3714 associate_code = (is_shift || code == MINUS ? PLUS : code);
3716 new_const = simplify_binary_operation (associate_code, mode,
3717 canon_const_arg1,
3718 inner_const);
3720 if (new_const == 0)
3721 break;
3723 /* If we are associating shift operations, don't let this
3724 produce a shift of the size of the object or larger.
3725 This could occur when we follow a sign-extend by a right
3726 shift on a machine that does a sign-extend as a pair
3727 of shifts. */
3729 if (is_shift
3730 && CONST_INT_P (new_const)
3731 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
3733 /* As an exception, we can turn an ASHIFTRT of this
3734 form into a shift of the number of bits - 1. */
3735 if (code == ASHIFTRT)
3736 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3737 else if (!side_effects_p (XEXP (y, 0)))
3738 return CONST0_RTX (mode);
3739 else
3740 break;
3743 y = copy_rtx (XEXP (y, 0));
3745 /* If Y contains our first operand (the most common way this
3746 can happen is if Y is a MEM), we would do into an infinite
3747 loop if we tried to fold it. So don't in that case. */
3749 if (! reg_mentioned_p (folded_arg0, y))
3750 y = fold_rtx (y, insn);
3752 return simplify_gen_binary (code, mode, y, new_const);
3754 break;
3756 case DIV: case UDIV:
3757 /* ??? The associative optimization performed immediately above is
3758 also possible for DIV and UDIV using associate_code of MULT.
3759 However, we would need extra code to verify that the
3760 multiplication does not overflow, that is, there is no overflow
3761 in the calculation of new_const. */
3762 break;
3764 default:
3765 break;
3768 new_rtx = simplify_binary_operation (code, mode,
3769 const_arg0 ? const_arg0 : folded_arg0,
3770 const_arg1 ? const_arg1 : folded_arg1);
3771 break;
3773 case RTX_OBJ:
3774 /* (lo_sum (high X) X) is simply X. */
3775 if (code == LO_SUM && const_arg0 != 0
3776 && GET_CODE (const_arg0) == HIGH
3777 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3778 return const_arg1;
3779 break;
3781 case RTX_TERNARY:
3782 case RTX_BITFIELD_OPS:
3783 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3784 const_arg0 ? const_arg0 : folded_arg0,
3785 const_arg1 ? const_arg1 : folded_arg1,
3786 const_arg2 ? const_arg2 : XEXP (x, 2));
3787 break;
3789 default:
3790 break;
3793 return new_rtx ? new_rtx : x;
3796 /* Return a constant value currently equivalent to X.
3797 Return 0 if we don't know one. */
3799 static rtx
3800 equiv_constant (rtx x)
3802 if (REG_P (x)
3803 && REGNO_QTY_VALID_P (REGNO (x)))
3805 int x_q = REG_QTY (REGNO (x));
3806 struct qty_table_elem *x_ent = &qty_table[x_q];
3808 if (x_ent->const_rtx)
3809 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3812 if (x == 0 || CONSTANT_P (x))
3813 return x;
3815 if (GET_CODE (x) == SUBREG)
3817 enum machine_mode mode = GET_MODE (x);
3818 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3819 rtx new_rtx;
3821 /* See if we previously assigned a constant value to this SUBREG. */
3822 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3823 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3824 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3825 return new_rtx;
3827 /* If we didn't and if doing so makes sense, see if we previously
3828 assigned a constant value to the enclosing word mode SUBREG. */
3829 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3830 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3832 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3833 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3835 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3836 new_rtx = lookup_as_function (y, CONST_INT);
3837 if (new_rtx)
3838 return gen_lowpart (mode, new_rtx);
3842 /* Otherwise see if we already have a constant for the inner REG. */
3843 if (REG_P (SUBREG_REG (x))
3844 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3845 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3847 return 0;
3850 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3851 the hash table in case its value was seen before. */
3853 if (MEM_P (x))
3855 struct table_elt *elt;
3857 x = avoid_constant_pool_reference (x);
3858 if (CONSTANT_P (x))
3859 return x;
3861 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3862 if (elt == 0)
3863 return 0;
3865 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3866 if (elt->is_const && CONSTANT_P (elt->exp))
3867 return elt->exp;
3870 return 0;
3873 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3874 "taken" branch.
3876 In certain cases, this can cause us to add an equivalence. For example,
3877 if we are following the taken case of
3878 if (i == 2)
3879 we can add the fact that `i' and '2' are now equivalent.
3881 In any case, we can record that this comparison was passed. If the same
3882 comparison is seen later, we will know its value. */
3884 static void
3885 record_jump_equiv (rtx insn, bool taken)
3887 int cond_known_true;
3888 rtx op0, op1;
3889 rtx set;
3890 enum machine_mode mode, mode0, mode1;
3891 int reversed_nonequality = 0;
3892 enum rtx_code code;
3894 /* Ensure this is the right kind of insn. */
3895 gcc_assert (any_condjump_p (insn));
3897 set = pc_set (insn);
3899 /* See if this jump condition is known true or false. */
3900 if (taken)
3901 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3902 else
3903 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3905 /* Get the type of comparison being done and the operands being compared.
3906 If we had to reverse a non-equality condition, record that fact so we
3907 know that it isn't valid for floating-point. */
3908 code = GET_CODE (XEXP (SET_SRC (set), 0));
3909 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3910 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3912 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3913 if (! cond_known_true)
3915 code = reversed_comparison_code_parts (code, op0, op1, insn);
3917 /* Don't remember if we can't find the inverse. */
3918 if (code == UNKNOWN)
3919 return;
3922 /* The mode is the mode of the non-constant. */
3923 mode = mode0;
3924 if (mode1 != VOIDmode)
3925 mode = mode1;
3927 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3930 /* Yet another form of subreg creation. In this case, we want something in
3931 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3933 static rtx
3934 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3936 enum machine_mode op_mode = GET_MODE (op);
3937 if (op_mode == mode || op_mode == VOIDmode)
3938 return op;
3939 return lowpart_subreg (mode, op, op_mode);
3942 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3943 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3944 Make any useful entries we can with that information. Called from
3945 above function and called recursively. */
3947 static void
3948 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3949 rtx op1, int reversed_nonequality)
3951 unsigned op0_hash, op1_hash;
3952 int op0_in_memory, op1_in_memory;
3953 struct table_elt *op0_elt, *op1_elt;
3955 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3956 we know that they are also equal in the smaller mode (this is also
3957 true for all smaller modes whether or not there is a SUBREG, but
3958 is not worth testing for with no SUBREG). */
3960 /* Note that GET_MODE (op0) may not equal MODE. */
3961 if (code == EQ && GET_CODE (op0) == SUBREG
3962 && (GET_MODE_SIZE (GET_MODE (op0))
3963 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3965 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3966 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3967 if (tem)
3968 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3969 reversed_nonequality);
3972 if (code == EQ && GET_CODE (op1) == SUBREG
3973 && (GET_MODE_SIZE (GET_MODE (op1))
3974 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3976 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3977 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3978 if (tem)
3979 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3980 reversed_nonequality);
3983 /* Similarly, if this is an NE comparison, and either is a SUBREG
3984 making a smaller mode, we know the whole thing is also NE. */
3986 /* Note that GET_MODE (op0) may not equal MODE;
3987 if we test MODE instead, we can get an infinite recursion
3988 alternating between two modes each wider than MODE. */
3990 if (code == NE && GET_CODE (op0) == SUBREG
3991 && subreg_lowpart_p (op0)
3992 && (GET_MODE_SIZE (GET_MODE (op0))
3993 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3995 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3996 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3997 if (tem)
3998 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3999 reversed_nonequality);
4002 if (code == NE && GET_CODE (op1) == SUBREG
4003 && subreg_lowpart_p (op1)
4004 && (GET_MODE_SIZE (GET_MODE (op1))
4005 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4007 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4008 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4009 if (tem)
4010 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4011 reversed_nonequality);
4014 /* Hash both operands. */
4016 do_not_record = 0;
4017 hash_arg_in_memory = 0;
4018 op0_hash = HASH (op0, mode);
4019 op0_in_memory = hash_arg_in_memory;
4021 if (do_not_record)
4022 return;
4024 do_not_record = 0;
4025 hash_arg_in_memory = 0;
4026 op1_hash = HASH (op1, mode);
4027 op1_in_memory = hash_arg_in_memory;
4029 if (do_not_record)
4030 return;
4032 /* Look up both operands. */
4033 op0_elt = lookup (op0, op0_hash, mode);
4034 op1_elt = lookup (op1, op1_hash, mode);
4036 /* If both operands are already equivalent or if they are not in the
4037 table but are identical, do nothing. */
4038 if ((op0_elt != 0 && op1_elt != 0
4039 && op0_elt->first_same_value == op1_elt->first_same_value)
4040 || op0 == op1 || rtx_equal_p (op0, op1))
4041 return;
4043 /* If we aren't setting two things equal all we can do is save this
4044 comparison. Similarly if this is floating-point. In the latter
4045 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4046 If we record the equality, we might inadvertently delete code
4047 whose intent was to change -0 to +0. */
4049 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4051 struct qty_table_elem *ent;
4052 int qty;
4054 /* If we reversed a floating-point comparison, if OP0 is not a
4055 register, or if OP1 is neither a register or constant, we can't
4056 do anything. */
4058 if (!REG_P (op1))
4059 op1 = equiv_constant (op1);
4061 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4062 || !REG_P (op0) || op1 == 0)
4063 return;
4065 /* Put OP0 in the hash table if it isn't already. This gives it a
4066 new quantity number. */
4067 if (op0_elt == 0)
4069 if (insert_regs (op0, NULL, 0))
4071 rehash_using_reg (op0);
4072 op0_hash = HASH (op0, mode);
4074 /* If OP0 is contained in OP1, this changes its hash code
4075 as well. Faster to rehash than to check, except
4076 for the simple case of a constant. */
4077 if (! CONSTANT_P (op1))
4078 op1_hash = HASH (op1,mode);
4081 op0_elt = insert (op0, NULL, op0_hash, mode);
4082 op0_elt->in_memory = op0_in_memory;
4085 qty = REG_QTY (REGNO (op0));
4086 ent = &qty_table[qty];
4088 ent->comparison_code = code;
4089 if (REG_P (op1))
4091 /* Look it up again--in case op0 and op1 are the same. */
4092 op1_elt = lookup (op1, op1_hash, mode);
4094 /* Put OP1 in the hash table so it gets a new quantity number. */
4095 if (op1_elt == 0)
4097 if (insert_regs (op1, NULL, 0))
4099 rehash_using_reg (op1);
4100 op1_hash = HASH (op1, mode);
4103 op1_elt = insert (op1, NULL, op1_hash, mode);
4104 op1_elt->in_memory = op1_in_memory;
4107 ent->comparison_const = NULL_RTX;
4108 ent->comparison_qty = REG_QTY (REGNO (op1));
4110 else
4112 ent->comparison_const = op1;
4113 ent->comparison_qty = -1;
4116 return;
4119 /* If either side is still missing an equivalence, make it now,
4120 then merge the equivalences. */
4122 if (op0_elt == 0)
4124 if (insert_regs (op0, NULL, 0))
4126 rehash_using_reg (op0);
4127 op0_hash = HASH (op0, mode);
4130 op0_elt = insert (op0, NULL, op0_hash, mode);
4131 op0_elt->in_memory = op0_in_memory;
4134 if (op1_elt == 0)
4136 if (insert_regs (op1, NULL, 0))
4138 rehash_using_reg (op1);
4139 op1_hash = HASH (op1, mode);
4142 op1_elt = insert (op1, NULL, op1_hash, mode);
4143 op1_elt->in_memory = op1_in_memory;
4146 merge_equiv_classes (op0_elt, op1_elt);
4149 /* CSE processing for one instruction.
4150 First simplify sources and addresses of all assignments
4151 in the instruction, using previously-computed equivalents values.
4152 Then install the new sources and destinations in the table
4153 of available values. */
4155 /* Data on one SET contained in the instruction. */
4157 struct set
4159 /* The SET rtx itself. */
4160 rtx rtl;
4161 /* The SET_SRC of the rtx (the original value, if it is changing). */
4162 rtx src;
4163 /* The hash-table element for the SET_SRC of the SET. */
4164 struct table_elt *src_elt;
4165 /* Hash value for the SET_SRC. */
4166 unsigned src_hash;
4167 /* Hash value for the SET_DEST. */
4168 unsigned dest_hash;
4169 /* The SET_DEST, with SUBREG, etc., stripped. */
4170 rtx inner_dest;
4171 /* Nonzero if the SET_SRC is in memory. */
4172 char src_in_memory;
4173 /* Nonzero if the SET_SRC contains something
4174 whose value cannot be predicted and understood. */
4175 char src_volatile;
4176 /* Original machine mode, in case it becomes a CONST_INT.
4177 The size of this field should match the size of the mode
4178 field of struct rtx_def (see rtl.h). */
4179 ENUM_BITFIELD(machine_mode) mode : 8;
4180 /* A constant equivalent for SET_SRC, if any. */
4181 rtx src_const;
4182 /* Hash value of constant equivalent for SET_SRC. */
4183 unsigned src_const_hash;
4184 /* Table entry for constant equivalent for SET_SRC, if any. */
4185 struct table_elt *src_const_elt;
4186 /* Table entry for the destination address. */
4187 struct table_elt *dest_addr_elt;
4190 static void
4191 cse_insn (rtx insn)
4193 rtx x = PATTERN (insn);
4194 int i;
4195 rtx tem;
4196 int n_sets = 0;
4198 rtx src_eqv = 0;
4199 struct table_elt *src_eqv_elt = 0;
4200 int src_eqv_volatile = 0;
4201 int src_eqv_in_memory = 0;
4202 unsigned src_eqv_hash = 0;
4204 struct set *sets = (struct set *) 0;
4206 this_insn = insn;
4207 #ifdef HAVE_cc0
4208 /* Records what this insn does to set CC0. */
4209 this_insn_cc0 = 0;
4210 this_insn_cc0_mode = VOIDmode;
4211 #endif
4213 /* Find all the SETs and CLOBBERs in this instruction.
4214 Record all the SETs in the array `set' and count them.
4215 Also determine whether there is a CLOBBER that invalidates
4216 all memory references, or all references at varying addresses. */
4218 if (CALL_P (insn))
4220 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4222 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4223 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4224 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4228 if (GET_CODE (x) == SET)
4230 sets = XALLOCA (struct set);
4231 sets[0].rtl = x;
4233 /* Ignore SETs that are unconditional jumps.
4234 They never need cse processing, so this does not hurt.
4235 The reason is not efficiency but rather
4236 so that we can test at the end for instructions
4237 that have been simplified to unconditional jumps
4238 and not be misled by unchanged instructions
4239 that were unconditional jumps to begin with. */
4240 if (SET_DEST (x) == pc_rtx
4241 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4244 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4245 The hard function value register is used only once, to copy to
4246 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4247 Ensure we invalidate the destination register. On the 80386 no
4248 other code would invalidate it since it is a fixed_reg.
4249 We need not check the return of apply_change_group; see canon_reg. */
4251 else if (GET_CODE (SET_SRC (x)) == CALL)
4253 canon_reg (SET_SRC (x), insn);
4254 apply_change_group ();
4255 fold_rtx (SET_SRC (x), insn);
4256 invalidate (SET_DEST (x), VOIDmode);
4258 else
4259 n_sets = 1;
4261 else if (GET_CODE (x) == PARALLEL)
4263 int lim = XVECLEN (x, 0);
4265 sets = XALLOCAVEC (struct set, lim);
4267 /* Find all regs explicitly clobbered in this insn,
4268 and ensure they are not replaced with any other regs
4269 elsewhere in this insn.
4270 When a reg that is clobbered is also used for input,
4271 we should presume that that is for a reason,
4272 and we should not substitute some other register
4273 which is not supposed to be clobbered.
4274 Therefore, this loop cannot be merged into the one below
4275 because a CALL may precede a CLOBBER and refer to the
4276 value clobbered. We must not let a canonicalization do
4277 anything in that case. */
4278 for (i = 0; i < lim; i++)
4280 rtx y = XVECEXP (x, 0, i);
4281 if (GET_CODE (y) == CLOBBER)
4283 rtx clobbered = XEXP (y, 0);
4285 if (REG_P (clobbered)
4286 || GET_CODE (clobbered) == SUBREG)
4287 invalidate (clobbered, VOIDmode);
4288 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4289 || GET_CODE (clobbered) == ZERO_EXTRACT)
4290 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4294 for (i = 0; i < lim; i++)
4296 rtx y = XVECEXP (x, 0, i);
4297 if (GET_CODE (y) == SET)
4299 /* As above, we ignore unconditional jumps and call-insns and
4300 ignore the result of apply_change_group. */
4301 if (GET_CODE (SET_SRC (y)) == CALL)
4303 canon_reg (SET_SRC (y), insn);
4304 apply_change_group ();
4305 fold_rtx (SET_SRC (y), insn);
4306 invalidate (SET_DEST (y), VOIDmode);
4308 else if (SET_DEST (y) == pc_rtx
4309 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4311 else
4312 sets[n_sets++].rtl = y;
4314 else if (GET_CODE (y) == CLOBBER)
4316 /* If we clobber memory, canon the address.
4317 This does nothing when a register is clobbered
4318 because we have already invalidated the reg. */
4319 if (MEM_P (XEXP (y, 0)))
4320 canon_reg (XEXP (y, 0), insn);
4322 else if (GET_CODE (y) == USE
4323 && ! (REG_P (XEXP (y, 0))
4324 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4325 canon_reg (y, insn);
4326 else if (GET_CODE (y) == CALL)
4328 /* The result of apply_change_group can be ignored; see
4329 canon_reg. */
4330 canon_reg (y, insn);
4331 apply_change_group ();
4332 fold_rtx (y, insn);
4336 else if (GET_CODE (x) == CLOBBER)
4338 if (MEM_P (XEXP (x, 0)))
4339 canon_reg (XEXP (x, 0), insn);
4341 /* Canonicalize a USE of a pseudo register or memory location. */
4342 else if (GET_CODE (x) == USE
4343 && ! (REG_P (XEXP (x, 0))
4344 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4345 canon_reg (x, insn);
4346 else if (GET_CODE (x) == ASM_OPERANDS)
4348 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4350 rtx input = ASM_OPERANDS_INPUT (x, i);
4351 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4353 input = canon_reg (input, insn);
4354 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4358 else if (GET_CODE (x) == CALL)
4360 /* The result of apply_change_group can be ignored; see canon_reg. */
4361 canon_reg (x, insn);
4362 apply_change_group ();
4363 fold_rtx (x, insn);
4365 else if (DEBUG_INSN_P (insn))
4366 canon_reg (PATTERN (insn), insn);
4368 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4369 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4370 is handled specially for this case, and if it isn't set, then there will
4371 be no equivalence for the destination. */
4372 if (n_sets == 1 && REG_NOTES (insn) != 0
4373 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4374 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4375 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4377 /* The result of apply_change_group can be ignored; see canon_reg. */
4378 canon_reg (XEXP (tem, 0), insn);
4379 apply_change_group ();
4380 src_eqv = fold_rtx (XEXP (tem, 0), insn);
4381 XEXP (tem, 0) = copy_rtx (src_eqv);
4382 df_notes_rescan (insn);
4385 /* Canonicalize sources and addresses of destinations.
4386 We do this in a separate pass to avoid problems when a MATCH_DUP is
4387 present in the insn pattern. In that case, we want to ensure that
4388 we don't break the duplicate nature of the pattern. So we will replace
4389 both operands at the same time. Otherwise, we would fail to find an
4390 equivalent substitution in the loop calling validate_change below.
4392 We used to suppress canonicalization of DEST if it appears in SRC,
4393 but we don't do this any more. */
4395 for (i = 0; i < n_sets; i++)
4397 rtx dest = SET_DEST (sets[i].rtl);
4398 rtx src = SET_SRC (sets[i].rtl);
4399 rtx new_rtx = canon_reg (src, insn);
4401 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4403 if (GET_CODE (dest) == ZERO_EXTRACT)
4405 validate_change (insn, &XEXP (dest, 1),
4406 canon_reg (XEXP (dest, 1), insn), 1);
4407 validate_change (insn, &XEXP (dest, 2),
4408 canon_reg (XEXP (dest, 2), insn), 1);
4411 while (GET_CODE (dest) == SUBREG
4412 || GET_CODE (dest) == ZERO_EXTRACT
4413 || GET_CODE (dest) == STRICT_LOW_PART)
4414 dest = XEXP (dest, 0);
4416 if (MEM_P (dest))
4417 canon_reg (dest, insn);
4420 /* Now that we have done all the replacements, we can apply the change
4421 group and see if they all work. Note that this will cause some
4422 canonicalizations that would have worked individually not to be applied
4423 because some other canonicalization didn't work, but this should not
4424 occur often.
4426 The result of apply_change_group can be ignored; see canon_reg. */
4428 apply_change_group ();
4430 /* Set sets[i].src_elt to the class each source belongs to.
4431 Detect assignments from or to volatile things
4432 and set set[i] to zero so they will be ignored
4433 in the rest of this function.
4435 Nothing in this loop changes the hash table or the register chains. */
4437 for (i = 0; i < n_sets; i++)
4439 bool repeat = false;
4440 rtx src, dest;
4441 rtx src_folded;
4442 struct table_elt *elt = 0, *p;
4443 enum machine_mode mode;
4444 rtx src_eqv_here;
4445 rtx src_const = 0;
4446 rtx src_related = 0;
4447 bool src_related_is_const_anchor = false;
4448 struct table_elt *src_const_elt = 0;
4449 int src_cost = MAX_COST;
4450 int src_eqv_cost = MAX_COST;
4451 int src_folded_cost = MAX_COST;
4452 int src_related_cost = MAX_COST;
4453 int src_elt_cost = MAX_COST;
4454 int src_regcost = MAX_COST;
4455 int src_eqv_regcost = MAX_COST;
4456 int src_folded_regcost = MAX_COST;
4457 int src_related_regcost = MAX_COST;
4458 int src_elt_regcost = MAX_COST;
4459 /* Set nonzero if we need to call force_const_mem on with the
4460 contents of src_folded before using it. */
4461 int src_folded_force_flag = 0;
4463 dest = SET_DEST (sets[i].rtl);
4464 src = SET_SRC (sets[i].rtl);
4466 /* If SRC is a constant that has no machine mode,
4467 hash it with the destination's machine mode.
4468 This way we can keep different modes separate. */
4470 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4471 sets[i].mode = mode;
4473 if (src_eqv)
4475 enum machine_mode eqvmode = mode;
4476 if (GET_CODE (dest) == STRICT_LOW_PART)
4477 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4478 do_not_record = 0;
4479 hash_arg_in_memory = 0;
4480 src_eqv_hash = HASH (src_eqv, eqvmode);
4482 /* Find the equivalence class for the equivalent expression. */
4484 if (!do_not_record)
4485 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4487 src_eqv_volatile = do_not_record;
4488 src_eqv_in_memory = hash_arg_in_memory;
4491 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4492 value of the INNER register, not the destination. So it is not
4493 a valid substitution for the source. But save it for later. */
4494 if (GET_CODE (dest) == STRICT_LOW_PART)
4495 src_eqv_here = 0;
4496 else
4497 src_eqv_here = src_eqv;
4499 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4500 simplified result, which may not necessarily be valid. */
4501 src_folded = fold_rtx (src, insn);
4503 #if 0
4504 /* ??? This caused bad code to be generated for the m68k port with -O2.
4505 Suppose src is (CONST_INT -1), and that after truncation src_folded
4506 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4507 At the end we will add src and src_const to the same equivalence
4508 class. We now have 3 and -1 on the same equivalence class. This
4509 causes later instructions to be mis-optimized. */
4510 /* If storing a constant in a bitfield, pre-truncate the constant
4511 so we will be able to record it later. */
4512 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4514 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4516 if (CONST_INT_P (src)
4517 && CONST_INT_P (width)
4518 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4519 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4520 src_folded
4521 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4522 << INTVAL (width)) - 1));
4524 #endif
4526 /* Compute SRC's hash code, and also notice if it
4527 should not be recorded at all. In that case,
4528 prevent any further processing of this assignment. */
4529 do_not_record = 0;
4530 hash_arg_in_memory = 0;
4532 sets[i].src = src;
4533 sets[i].src_hash = HASH (src, mode);
4534 sets[i].src_volatile = do_not_record;
4535 sets[i].src_in_memory = hash_arg_in_memory;
4537 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4538 a pseudo, do not record SRC. Using SRC as a replacement for
4539 anything else will be incorrect in that situation. Note that
4540 this usually occurs only for stack slots, in which case all the
4541 RTL would be referring to SRC, so we don't lose any optimization
4542 opportunities by not having SRC in the hash table. */
4544 if (MEM_P (src)
4545 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4546 && REG_P (dest)
4547 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4548 sets[i].src_volatile = 1;
4550 #if 0
4551 /* It is no longer clear why we used to do this, but it doesn't
4552 appear to still be needed. So let's try without it since this
4553 code hurts cse'ing widened ops. */
4554 /* If source is a paradoxical subreg (such as QI treated as an SI),
4555 treat it as volatile. It may do the work of an SI in one context
4556 where the extra bits are not being used, but cannot replace an SI
4557 in general. */
4558 if (GET_CODE (src) == SUBREG
4559 && (GET_MODE_SIZE (GET_MODE (src))
4560 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
4561 sets[i].src_volatile = 1;
4562 #endif
4564 /* Locate all possible equivalent forms for SRC. Try to replace
4565 SRC in the insn with each cheaper equivalent.
4567 We have the following types of equivalents: SRC itself, a folded
4568 version, a value given in a REG_EQUAL note, or a value related
4569 to a constant.
4571 Each of these equivalents may be part of an additional class
4572 of equivalents (if more than one is in the table, they must be in
4573 the same class; we check for this).
4575 If the source is volatile, we don't do any table lookups.
4577 We note any constant equivalent for possible later use in a
4578 REG_NOTE. */
4580 if (!sets[i].src_volatile)
4581 elt = lookup (src, sets[i].src_hash, mode);
4583 sets[i].src_elt = elt;
4585 if (elt && src_eqv_here && src_eqv_elt)
4587 if (elt->first_same_value != src_eqv_elt->first_same_value)
4589 /* The REG_EQUAL is indicating that two formerly distinct
4590 classes are now equivalent. So merge them. */
4591 merge_equiv_classes (elt, src_eqv_elt);
4592 src_eqv_hash = HASH (src_eqv, elt->mode);
4593 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4596 src_eqv_here = 0;
4599 else if (src_eqv_elt)
4600 elt = src_eqv_elt;
4602 /* Try to find a constant somewhere and record it in `src_const'.
4603 Record its table element, if any, in `src_const_elt'. Look in
4604 any known equivalences first. (If the constant is not in the
4605 table, also set `sets[i].src_const_hash'). */
4606 if (elt)
4607 for (p = elt->first_same_value; p; p = p->next_same_value)
4608 if (p->is_const)
4610 src_const = p->exp;
4611 src_const_elt = elt;
4612 break;
4615 if (src_const == 0
4616 && (CONSTANT_P (src_folded)
4617 /* Consider (minus (label_ref L1) (label_ref L2)) as
4618 "constant" here so we will record it. This allows us
4619 to fold switch statements when an ADDR_DIFF_VEC is used. */
4620 || (GET_CODE (src_folded) == MINUS
4621 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4622 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4623 src_const = src_folded, src_const_elt = elt;
4624 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4625 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4627 /* If we don't know if the constant is in the table, get its
4628 hash code and look it up. */
4629 if (src_const && src_const_elt == 0)
4631 sets[i].src_const_hash = HASH (src_const, mode);
4632 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4635 sets[i].src_const = src_const;
4636 sets[i].src_const_elt = src_const_elt;
4638 /* If the constant and our source are both in the table, mark them as
4639 equivalent. Otherwise, if a constant is in the table but the source
4640 isn't, set ELT to it. */
4641 if (src_const_elt && elt
4642 && src_const_elt->first_same_value != elt->first_same_value)
4643 merge_equiv_classes (elt, src_const_elt);
4644 else if (src_const_elt && elt == 0)
4645 elt = src_const_elt;
4647 /* See if there is a register linearly related to a constant
4648 equivalent of SRC. */
4649 if (src_const
4650 && (GET_CODE (src_const) == CONST
4651 || (src_const_elt && src_const_elt->related_value != 0)))
4653 src_related = use_related_value (src_const, src_const_elt);
4654 if (src_related)
4656 struct table_elt *src_related_elt
4657 = lookup (src_related, HASH (src_related, mode), mode);
4658 if (src_related_elt && elt)
4660 if (elt->first_same_value
4661 != src_related_elt->first_same_value)
4662 /* This can occur when we previously saw a CONST
4663 involving a SYMBOL_REF and then see the SYMBOL_REF
4664 twice. Merge the involved classes. */
4665 merge_equiv_classes (elt, src_related_elt);
4667 src_related = 0;
4668 src_related_elt = 0;
4670 else if (src_related_elt && elt == 0)
4671 elt = src_related_elt;
4675 /* See if we have a CONST_INT that is already in a register in a
4676 wider mode. */
4678 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4679 && GET_MODE_CLASS (mode) == MODE_INT
4680 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
4682 enum machine_mode wider_mode;
4684 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4685 wider_mode != VOIDmode
4686 && GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
4687 && src_related == 0;
4688 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4690 struct table_elt *const_elt
4691 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4693 if (const_elt == 0)
4694 continue;
4696 for (const_elt = const_elt->first_same_value;
4697 const_elt; const_elt = const_elt->next_same_value)
4698 if (REG_P (const_elt->exp))
4700 src_related = gen_lowpart (mode, const_elt->exp);
4701 break;
4706 /* Another possibility is that we have an AND with a constant in
4707 a mode narrower than a word. If so, it might have been generated
4708 as part of an "if" which would narrow the AND. If we already
4709 have done the AND in a wider mode, we can use a SUBREG of that
4710 value. */
4712 if (flag_expensive_optimizations && ! src_related
4713 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4714 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4716 enum machine_mode tmode;
4717 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4719 for (tmode = GET_MODE_WIDER_MODE (mode);
4720 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4721 tmode = GET_MODE_WIDER_MODE (tmode))
4723 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4724 struct table_elt *larger_elt;
4726 if (inner)
4728 PUT_MODE (new_and, tmode);
4729 XEXP (new_and, 0) = inner;
4730 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4731 if (larger_elt == 0)
4732 continue;
4734 for (larger_elt = larger_elt->first_same_value;
4735 larger_elt; larger_elt = larger_elt->next_same_value)
4736 if (REG_P (larger_elt->exp))
4738 src_related
4739 = gen_lowpart (mode, larger_elt->exp);
4740 break;
4743 if (src_related)
4744 break;
4749 #ifdef LOAD_EXTEND_OP
4750 /* See if a MEM has already been loaded with a widening operation;
4751 if it has, we can use a subreg of that. Many CISC machines
4752 also have such operations, but this is only likely to be
4753 beneficial on these machines. */
4755 if (flag_expensive_optimizations && src_related == 0
4756 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4757 && GET_MODE_CLASS (mode) == MODE_INT
4758 && MEM_P (src) && ! do_not_record
4759 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4761 struct rtx_def memory_extend_buf;
4762 rtx memory_extend_rtx = &memory_extend_buf;
4763 enum machine_mode tmode;
4765 /* Set what we are trying to extend and the operation it might
4766 have been extended with. */
4767 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4768 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4769 XEXP (memory_extend_rtx, 0) = src;
4771 for (tmode = GET_MODE_WIDER_MODE (mode);
4772 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4773 tmode = GET_MODE_WIDER_MODE (tmode))
4775 struct table_elt *larger_elt;
4777 PUT_MODE (memory_extend_rtx, tmode);
4778 larger_elt = lookup (memory_extend_rtx,
4779 HASH (memory_extend_rtx, tmode), tmode);
4780 if (larger_elt == 0)
4781 continue;
4783 for (larger_elt = larger_elt->first_same_value;
4784 larger_elt; larger_elt = larger_elt->next_same_value)
4785 if (REG_P (larger_elt->exp))
4787 src_related = gen_lowpart (mode, larger_elt->exp);
4788 break;
4791 if (src_related)
4792 break;
4795 #endif /* LOAD_EXTEND_OP */
4797 /* Try to express the constant using a register+offset expression
4798 derived from a constant anchor. */
4800 if (targetm.const_anchor
4801 && !src_related
4802 && src_const
4803 && GET_CODE (src_const) == CONST_INT)
4805 src_related = try_const_anchors (src_const, mode);
4806 src_related_is_const_anchor = src_related != NULL_RTX;
4810 if (src == src_folded)
4811 src_folded = 0;
4813 /* At this point, ELT, if nonzero, points to a class of expressions
4814 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4815 and SRC_RELATED, if nonzero, each contain additional equivalent
4816 expressions. Prune these latter expressions by deleting expressions
4817 already in the equivalence class.
4819 Check for an equivalent identical to the destination. If found,
4820 this is the preferred equivalent since it will likely lead to
4821 elimination of the insn. Indicate this by placing it in
4822 `src_related'. */
4824 if (elt)
4825 elt = elt->first_same_value;
4826 for (p = elt; p; p = p->next_same_value)
4828 enum rtx_code code = GET_CODE (p->exp);
4830 /* If the expression is not valid, ignore it. Then we do not
4831 have to check for validity below. In most cases, we can use
4832 `rtx_equal_p', since canonicalization has already been done. */
4833 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4834 continue;
4836 /* Also skip paradoxical subregs, unless that's what we're
4837 looking for. */
4838 if (code == SUBREG
4839 && (GET_MODE_SIZE (GET_MODE (p->exp))
4840 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
4841 && ! (src != 0
4842 && GET_CODE (src) == SUBREG
4843 && GET_MODE (src) == GET_MODE (p->exp)
4844 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4845 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4846 continue;
4848 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4849 src = 0;
4850 else if (src_folded && GET_CODE (src_folded) == code
4851 && rtx_equal_p (src_folded, p->exp))
4852 src_folded = 0;
4853 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4854 && rtx_equal_p (src_eqv_here, p->exp))
4855 src_eqv_here = 0;
4856 else if (src_related && GET_CODE (src_related) == code
4857 && rtx_equal_p (src_related, p->exp))
4858 src_related = 0;
4860 /* This is the same as the destination of the insns, we want
4861 to prefer it. Copy it to src_related. The code below will
4862 then give it a negative cost. */
4863 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4864 src_related = dest;
4867 /* Find the cheapest valid equivalent, trying all the available
4868 possibilities. Prefer items not in the hash table to ones
4869 that are when they are equal cost. Note that we can never
4870 worsen an insn as the current contents will also succeed.
4871 If we find an equivalent identical to the destination, use it as best,
4872 since this insn will probably be eliminated in that case. */
4873 if (src)
4875 if (rtx_equal_p (src, dest))
4876 src_cost = src_regcost = -1;
4877 else
4879 src_cost = COST (src);
4880 src_regcost = approx_reg_cost (src);
4884 if (src_eqv_here)
4886 if (rtx_equal_p (src_eqv_here, dest))
4887 src_eqv_cost = src_eqv_regcost = -1;
4888 else
4890 src_eqv_cost = COST (src_eqv_here);
4891 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4895 if (src_folded)
4897 if (rtx_equal_p (src_folded, dest))
4898 src_folded_cost = src_folded_regcost = -1;
4899 else
4901 src_folded_cost = COST (src_folded);
4902 src_folded_regcost = approx_reg_cost (src_folded);
4906 if (src_related)
4908 if (rtx_equal_p (src_related, dest))
4909 src_related_cost = src_related_regcost = -1;
4910 else
4912 src_related_cost = COST (src_related);
4913 src_related_regcost = approx_reg_cost (src_related);
4915 /* If a const-anchor is used to synthesize a constant that
4916 normally requires multiple instructions then slightly prefer
4917 it over the original sequence. These instructions are likely
4918 to become redundant now. We can't compare against the cost
4919 of src_eqv_here because, on MIPS for example, multi-insn
4920 constants have zero cost; they are assumed to be hoisted from
4921 loops. */
4922 if (src_related_is_const_anchor
4923 && src_related_cost == src_cost
4924 && src_eqv_here)
4925 src_related_cost--;
4929 /* If this was an indirect jump insn, a known label will really be
4930 cheaper even though it looks more expensive. */
4931 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
4932 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
4934 /* Terminate loop when replacement made. This must terminate since
4935 the current contents will be tested and will always be valid. */
4936 while (1)
4938 rtx trial;
4940 /* Skip invalid entries. */
4941 while (elt && !REG_P (elt->exp)
4942 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
4943 elt = elt->next_same_value;
4945 /* A paradoxical subreg would be bad here: it'll be the right
4946 size, but later may be adjusted so that the upper bits aren't
4947 what we want. So reject it. */
4948 if (elt != 0
4949 && GET_CODE (elt->exp) == SUBREG
4950 && (GET_MODE_SIZE (GET_MODE (elt->exp))
4951 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
4952 /* It is okay, though, if the rtx we're trying to match
4953 will ignore any of the bits we can't predict. */
4954 && ! (src != 0
4955 && GET_CODE (src) == SUBREG
4956 && GET_MODE (src) == GET_MODE (elt->exp)
4957 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4958 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
4960 elt = elt->next_same_value;
4961 continue;
4964 if (elt)
4966 src_elt_cost = elt->cost;
4967 src_elt_regcost = elt->regcost;
4970 /* Find cheapest and skip it for the next time. For items
4971 of equal cost, use this order:
4972 src_folded, src, src_eqv, src_related and hash table entry. */
4973 if (src_folded
4974 && preferable (src_folded_cost, src_folded_regcost,
4975 src_cost, src_regcost) <= 0
4976 && preferable (src_folded_cost, src_folded_regcost,
4977 src_eqv_cost, src_eqv_regcost) <= 0
4978 && preferable (src_folded_cost, src_folded_regcost,
4979 src_related_cost, src_related_regcost) <= 0
4980 && preferable (src_folded_cost, src_folded_regcost,
4981 src_elt_cost, src_elt_regcost) <= 0)
4983 trial = src_folded, src_folded_cost = MAX_COST;
4984 if (src_folded_force_flag)
4986 rtx forced = force_const_mem (mode, trial);
4987 if (forced)
4988 trial = forced;
4991 else if (src
4992 && preferable (src_cost, src_regcost,
4993 src_eqv_cost, src_eqv_regcost) <= 0
4994 && preferable (src_cost, src_regcost,
4995 src_related_cost, src_related_regcost) <= 0
4996 && preferable (src_cost, src_regcost,
4997 src_elt_cost, src_elt_regcost) <= 0)
4998 trial = src, src_cost = MAX_COST;
4999 else if (src_eqv_here
5000 && preferable (src_eqv_cost, src_eqv_regcost,
5001 src_related_cost, src_related_regcost) <= 0
5002 && preferable (src_eqv_cost, src_eqv_regcost,
5003 src_elt_cost, src_elt_regcost) <= 0)
5004 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5005 else if (src_related
5006 && preferable (src_related_cost, src_related_regcost,
5007 src_elt_cost, src_elt_regcost) <= 0)
5008 trial = src_related, src_related_cost = MAX_COST;
5009 else
5011 trial = elt->exp;
5012 elt = elt->next_same_value;
5013 src_elt_cost = MAX_COST;
5016 /* Avoid creation of overlapping memory moves. */
5017 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5019 rtx src, dest;
5021 /* BLKmode moves are not handled by cse anyway. */
5022 if (GET_MODE (trial) == BLKmode)
5023 break;
5025 src = canon_rtx (trial);
5026 dest = canon_rtx (SET_DEST (sets[i].rtl));
5028 if (!MEM_P (src) || !MEM_P (dest)
5029 || !nonoverlapping_memrefs_p (src, dest, false))
5030 break;
5033 /* Try to optimize
5034 (set (reg:M N) (const_int A))
5035 (set (reg:M2 O) (const_int B))
5036 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5037 (reg:M2 O)). */
5038 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5039 && CONST_INT_P (trial)
5040 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5041 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5042 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5043 && (GET_MODE_BITSIZE (GET_MODE (SET_DEST (sets[i].rtl)))
5044 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5045 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5046 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5047 <= HOST_BITS_PER_WIDE_INT))
5049 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5050 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5051 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5052 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5053 struct table_elt *dest_elt
5054 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5055 rtx dest_cst = NULL;
5057 if (dest_elt)
5058 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5059 if (p->is_const && CONST_INT_P (p->exp))
5061 dest_cst = p->exp;
5062 break;
5064 if (dest_cst)
5066 HOST_WIDE_INT val = INTVAL (dest_cst);
5067 HOST_WIDE_INT mask;
5068 unsigned int shift;
5069 if (BITS_BIG_ENDIAN)
5070 shift = GET_MODE_BITSIZE (GET_MODE (dest_reg))
5071 - INTVAL (pos) - INTVAL (width);
5072 else
5073 shift = INTVAL (pos);
5074 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5075 mask = ~(HOST_WIDE_INT) 0;
5076 else
5077 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5078 val &= ~(mask << shift);
5079 val |= (INTVAL (trial) & mask) << shift;
5080 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5081 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5082 dest_reg, 1);
5083 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5084 GEN_INT (val), 1);
5085 if (apply_change_group ())
5087 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5088 if (note)
5090 remove_note (insn, note);
5091 df_notes_rescan (insn);
5093 src_eqv = NULL_RTX;
5094 src_eqv_elt = NULL;
5095 src_eqv_volatile = 0;
5096 src_eqv_in_memory = 0;
5097 src_eqv_hash = 0;
5098 repeat = true;
5099 break;
5104 /* We don't normally have an insn matching (set (pc) (pc)), so
5105 check for this separately here. We will delete such an
5106 insn below.
5108 For other cases such as a table jump or conditional jump
5109 where we know the ultimate target, go ahead and replace the
5110 operand. While that may not make a valid insn, we will
5111 reemit the jump below (and also insert any necessary
5112 barriers). */
5113 if (n_sets == 1 && dest == pc_rtx
5114 && (trial == pc_rtx
5115 || (GET_CODE (trial) == LABEL_REF
5116 && ! condjump_p (insn))))
5118 /* Don't substitute non-local labels, this confuses CFG. */
5119 if (GET_CODE (trial) == LABEL_REF
5120 && LABEL_REF_NONLOCAL_P (trial))
5121 continue;
5123 SET_SRC (sets[i].rtl) = trial;
5124 cse_jumps_altered = true;
5125 break;
5128 /* Reject certain invalid forms of CONST that we create. */
5129 else if (CONSTANT_P (trial)
5130 && GET_CODE (trial) == CONST
5131 /* Reject cases that will cause decode_rtx_const to
5132 die. On the alpha when simplifying a switch, we
5133 get (const (truncate (minus (label_ref)
5134 (label_ref)))). */
5135 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5136 /* Likewise on IA-64, except without the
5137 truncate. */
5138 || (GET_CODE (XEXP (trial, 0)) == MINUS
5139 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5140 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5141 /* Do nothing for this case. */
5144 /* Look for a substitution that makes a valid insn. */
5145 else if (validate_unshare_change
5146 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5148 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5150 /* The result of apply_change_group can be ignored; see
5151 canon_reg. */
5153 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5154 apply_change_group ();
5156 break;
5159 /* If we previously found constant pool entries for
5160 constants and this is a constant, try making a
5161 pool entry. Put it in src_folded unless we already have done
5162 this since that is where it likely came from. */
5164 else if (constant_pool_entries_cost
5165 && CONSTANT_P (trial)
5166 && (src_folded == 0
5167 || (!MEM_P (src_folded)
5168 && ! src_folded_force_flag))
5169 && GET_MODE_CLASS (mode) != MODE_CC
5170 && mode != VOIDmode)
5172 src_folded_force_flag = 1;
5173 src_folded = trial;
5174 src_folded_cost = constant_pool_entries_cost;
5175 src_folded_regcost = constant_pool_entries_regcost;
5179 /* If we changed the insn too much, handle this set from scratch. */
5180 if (repeat)
5182 i--;
5183 continue;
5186 src = SET_SRC (sets[i].rtl);
5188 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5189 However, there is an important exception: If both are registers
5190 that are not the head of their equivalence class, replace SET_SRC
5191 with the head of the class. If we do not do this, we will have
5192 both registers live over a portion of the basic block. This way,
5193 their lifetimes will likely abut instead of overlapping. */
5194 if (REG_P (dest)
5195 && REGNO_QTY_VALID_P (REGNO (dest)))
5197 int dest_q = REG_QTY (REGNO (dest));
5198 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5200 if (dest_ent->mode == GET_MODE (dest)
5201 && dest_ent->first_reg != REGNO (dest)
5202 && REG_P (src) && REGNO (src) == REGNO (dest)
5203 /* Don't do this if the original insn had a hard reg as
5204 SET_SRC or SET_DEST. */
5205 && (!REG_P (sets[i].src)
5206 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5207 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5208 /* We can't call canon_reg here because it won't do anything if
5209 SRC is a hard register. */
5211 int src_q = REG_QTY (REGNO (src));
5212 struct qty_table_elem *src_ent = &qty_table[src_q];
5213 int first = src_ent->first_reg;
5214 rtx new_src
5215 = (first >= FIRST_PSEUDO_REGISTER
5216 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5218 /* We must use validate-change even for this, because this
5219 might be a special no-op instruction, suitable only to
5220 tag notes onto. */
5221 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5223 src = new_src;
5224 /* If we had a constant that is cheaper than what we are now
5225 setting SRC to, use that constant. We ignored it when we
5226 thought we could make this into a no-op. */
5227 if (src_const && COST (src_const) < COST (src)
5228 && validate_change (insn, &SET_SRC (sets[i].rtl),
5229 src_const, 0))
5230 src = src_const;
5235 /* If we made a change, recompute SRC values. */
5236 if (src != sets[i].src)
5238 do_not_record = 0;
5239 hash_arg_in_memory = 0;
5240 sets[i].src = src;
5241 sets[i].src_hash = HASH (src, mode);
5242 sets[i].src_volatile = do_not_record;
5243 sets[i].src_in_memory = hash_arg_in_memory;
5244 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5247 /* If this is a single SET, we are setting a register, and we have an
5248 equivalent constant, we want to add a REG_NOTE. We don't want
5249 to write a REG_EQUAL note for a constant pseudo since verifying that
5250 that pseudo hasn't been eliminated is a pain. Such a note also
5251 won't help anything.
5253 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5254 which can be created for a reference to a compile time computable
5255 entry in a jump table. */
5257 if (n_sets == 1 && src_const && REG_P (dest)
5258 && !REG_P (src_const)
5259 && ! (GET_CODE (src_const) == CONST
5260 && GET_CODE (XEXP (src_const, 0)) == MINUS
5261 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5262 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5264 /* We only want a REG_EQUAL note if src_const != src. */
5265 if (! rtx_equal_p (src, src_const))
5267 /* Make sure that the rtx is not shared. */
5268 src_const = copy_rtx (src_const);
5270 /* Record the actual constant value in a REG_EQUAL note,
5271 making a new one if one does not already exist. */
5272 set_unique_reg_note (insn, REG_EQUAL, src_const);
5273 df_notes_rescan (insn);
5277 /* Now deal with the destination. */
5278 do_not_record = 0;
5280 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5281 while (GET_CODE (dest) == SUBREG
5282 || GET_CODE (dest) == ZERO_EXTRACT
5283 || GET_CODE (dest) == STRICT_LOW_PART)
5284 dest = XEXP (dest, 0);
5286 sets[i].inner_dest = dest;
5288 if (MEM_P (dest))
5290 #ifdef PUSH_ROUNDING
5291 /* Stack pushes invalidate the stack pointer. */
5292 rtx addr = XEXP (dest, 0);
5293 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5294 && XEXP (addr, 0) == stack_pointer_rtx)
5295 invalidate (stack_pointer_rtx, VOIDmode);
5296 #endif
5297 dest = fold_rtx (dest, insn);
5300 /* Compute the hash code of the destination now,
5301 before the effects of this instruction are recorded,
5302 since the register values used in the address computation
5303 are those before this instruction. */
5304 sets[i].dest_hash = HASH (dest, mode);
5306 /* Don't enter a bit-field in the hash table
5307 because the value in it after the store
5308 may not equal what was stored, due to truncation. */
5310 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5312 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5314 if (src_const != 0 && CONST_INT_P (src_const)
5315 && CONST_INT_P (width)
5316 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5317 && ! (INTVAL (src_const)
5318 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5319 /* Exception: if the value is constant,
5320 and it won't be truncated, record it. */
5322 else
5324 /* This is chosen so that the destination will be invalidated
5325 but no new value will be recorded.
5326 We must invalidate because sometimes constant
5327 values can be recorded for bitfields. */
5328 sets[i].src_elt = 0;
5329 sets[i].src_volatile = 1;
5330 src_eqv = 0;
5331 src_eqv_elt = 0;
5335 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5336 the insn. */
5337 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5339 /* One less use of the label this insn used to jump to. */
5340 delete_insn_and_edges (insn);
5341 cse_jumps_altered = true;
5342 /* No more processing for this set. */
5343 sets[i].rtl = 0;
5346 /* If this SET is now setting PC to a label, we know it used to
5347 be a conditional or computed branch. */
5348 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5349 && !LABEL_REF_NONLOCAL_P (src))
5351 /* We reemit the jump in as many cases as possible just in
5352 case the form of an unconditional jump is significantly
5353 different than a computed jump or conditional jump.
5355 If this insn has multiple sets, then reemitting the
5356 jump is nontrivial. So instead we just force rerecognition
5357 and hope for the best. */
5358 if (n_sets == 1)
5360 rtx new_rtx, note;
5362 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5363 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5364 LABEL_NUSES (XEXP (src, 0))++;
5366 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5367 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5368 if (note)
5370 XEXP (note, 1) = NULL_RTX;
5371 REG_NOTES (new_rtx) = note;
5374 delete_insn_and_edges (insn);
5375 insn = new_rtx;
5377 else
5378 INSN_CODE (insn) = -1;
5380 /* Do not bother deleting any unreachable code, let jump do it. */
5381 cse_jumps_altered = true;
5382 sets[i].rtl = 0;
5385 /* If destination is volatile, invalidate it and then do no further
5386 processing for this assignment. */
5388 else if (do_not_record)
5390 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5391 invalidate (dest, VOIDmode);
5392 else if (MEM_P (dest))
5393 invalidate (dest, VOIDmode);
5394 else if (GET_CODE (dest) == STRICT_LOW_PART
5395 || GET_CODE (dest) == ZERO_EXTRACT)
5396 invalidate (XEXP (dest, 0), GET_MODE (dest));
5397 sets[i].rtl = 0;
5400 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5401 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5403 #ifdef HAVE_cc0
5404 /* If setting CC0, record what it was set to, or a constant, if it
5405 is equivalent to a constant. If it is being set to a floating-point
5406 value, make a COMPARE with the appropriate constant of 0. If we
5407 don't do this, later code can interpret this as a test against
5408 const0_rtx, which can cause problems if we try to put it into an
5409 insn as a floating-point operand. */
5410 if (dest == cc0_rtx)
5412 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5413 this_insn_cc0_mode = mode;
5414 if (FLOAT_MODE_P (mode))
5415 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5416 CONST0_RTX (mode));
5418 #endif
5421 /* Now enter all non-volatile source expressions in the hash table
5422 if they are not already present.
5423 Record their equivalence classes in src_elt.
5424 This way we can insert the corresponding destinations into
5425 the same classes even if the actual sources are no longer in them
5426 (having been invalidated). */
5428 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5429 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5431 struct table_elt *elt;
5432 struct table_elt *classp = sets[0].src_elt;
5433 rtx dest = SET_DEST (sets[0].rtl);
5434 enum machine_mode eqvmode = GET_MODE (dest);
5436 if (GET_CODE (dest) == STRICT_LOW_PART)
5438 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5439 classp = 0;
5441 if (insert_regs (src_eqv, classp, 0))
5443 rehash_using_reg (src_eqv);
5444 src_eqv_hash = HASH (src_eqv, eqvmode);
5446 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5447 elt->in_memory = src_eqv_in_memory;
5448 src_eqv_elt = elt;
5450 /* Check to see if src_eqv_elt is the same as a set source which
5451 does not yet have an elt, and if so set the elt of the set source
5452 to src_eqv_elt. */
5453 for (i = 0; i < n_sets; i++)
5454 if (sets[i].rtl && sets[i].src_elt == 0
5455 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5456 sets[i].src_elt = src_eqv_elt;
5459 for (i = 0; i < n_sets; i++)
5460 if (sets[i].rtl && ! sets[i].src_volatile
5461 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5463 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5465 /* REG_EQUAL in setting a STRICT_LOW_PART
5466 gives an equivalent for the entire destination register,
5467 not just for the subreg being stored in now.
5468 This is a more interesting equivalence, so we arrange later
5469 to treat the entire reg as the destination. */
5470 sets[i].src_elt = src_eqv_elt;
5471 sets[i].src_hash = src_eqv_hash;
5473 else
5475 /* Insert source and constant equivalent into hash table, if not
5476 already present. */
5477 struct table_elt *classp = src_eqv_elt;
5478 rtx src = sets[i].src;
5479 rtx dest = SET_DEST (sets[i].rtl);
5480 enum machine_mode mode
5481 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5483 /* It's possible that we have a source value known to be
5484 constant but don't have a REG_EQUAL note on the insn.
5485 Lack of a note will mean src_eqv_elt will be NULL. This
5486 can happen where we've generated a SUBREG to access a
5487 CONST_INT that is already in a register in a wider mode.
5488 Ensure that the source expression is put in the proper
5489 constant class. */
5490 if (!classp)
5491 classp = sets[i].src_const_elt;
5493 if (sets[i].src_elt == 0)
5495 struct table_elt *elt;
5497 /* Note that these insert_regs calls cannot remove
5498 any of the src_elt's, because they would have failed to
5499 match if not still valid. */
5500 if (insert_regs (src, classp, 0))
5502 rehash_using_reg (src);
5503 sets[i].src_hash = HASH (src, mode);
5505 elt = insert (src, classp, sets[i].src_hash, mode);
5506 elt->in_memory = sets[i].src_in_memory;
5507 sets[i].src_elt = classp = elt;
5509 if (sets[i].src_const && sets[i].src_const_elt == 0
5510 && src != sets[i].src_const
5511 && ! rtx_equal_p (sets[i].src_const, src))
5512 sets[i].src_elt = insert (sets[i].src_const, classp,
5513 sets[i].src_const_hash, mode);
5516 else if (sets[i].src_elt == 0)
5517 /* If we did not insert the source into the hash table (e.g., it was
5518 volatile), note the equivalence class for the REG_EQUAL value, if any,
5519 so that the destination goes into that class. */
5520 sets[i].src_elt = src_eqv_elt;
5522 /* Record destination addresses in the hash table. This allows us to
5523 check if they are invalidated by other sets. */
5524 for (i = 0; i < n_sets; i++)
5526 if (sets[i].rtl)
5528 rtx x = sets[i].inner_dest;
5529 struct table_elt *elt;
5530 enum machine_mode mode;
5531 unsigned hash;
5533 if (MEM_P (x))
5535 x = XEXP (x, 0);
5536 mode = GET_MODE (x);
5537 hash = HASH (x, mode);
5538 elt = lookup (x, hash, mode);
5539 if (!elt)
5541 if (insert_regs (x, NULL, 0))
5543 rtx dest = SET_DEST (sets[i].rtl);
5545 rehash_using_reg (x);
5546 hash = HASH (x, mode);
5547 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5549 elt = insert (x, NULL, hash, mode);
5552 sets[i].dest_addr_elt = elt;
5554 else
5555 sets[i].dest_addr_elt = NULL;
5559 invalidate_from_clobbers (x);
5561 /* Some registers are invalidated by subroutine calls. Memory is
5562 invalidated by non-constant calls. */
5564 if (CALL_P (insn))
5566 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5567 invalidate_memory ();
5568 invalidate_for_call ();
5571 /* Now invalidate everything set by this instruction.
5572 If a SUBREG or other funny destination is being set,
5573 sets[i].rtl is still nonzero, so here we invalidate the reg
5574 a part of which is being set. */
5576 for (i = 0; i < n_sets; i++)
5577 if (sets[i].rtl)
5579 /* We can't use the inner dest, because the mode associated with
5580 a ZERO_EXTRACT is significant. */
5581 rtx dest = SET_DEST (sets[i].rtl);
5583 /* Needed for registers to remove the register from its
5584 previous quantity's chain.
5585 Needed for memory if this is a nonvarying address, unless
5586 we have just done an invalidate_memory that covers even those. */
5587 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5588 invalidate (dest, VOIDmode);
5589 else if (MEM_P (dest))
5590 invalidate (dest, VOIDmode);
5591 else if (GET_CODE (dest) == STRICT_LOW_PART
5592 || GET_CODE (dest) == ZERO_EXTRACT)
5593 invalidate (XEXP (dest, 0), GET_MODE (dest));
5596 /* A volatile ASM invalidates everything. */
5597 if (NONJUMP_INSN_P (insn)
5598 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5599 && MEM_VOLATILE_P (PATTERN (insn)))
5600 flush_hash_table ();
5602 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5603 the regs restored by the longjmp come from a later time
5604 than the setjmp. */
5605 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5607 flush_hash_table ();
5608 goto done;
5611 /* Make sure registers mentioned in destinations
5612 are safe for use in an expression to be inserted.
5613 This removes from the hash table
5614 any invalid entry that refers to one of these registers.
5616 We don't care about the return value from mention_regs because
5617 we are going to hash the SET_DEST values unconditionally. */
5619 for (i = 0; i < n_sets; i++)
5621 if (sets[i].rtl)
5623 rtx x = SET_DEST (sets[i].rtl);
5625 if (!REG_P (x))
5626 mention_regs (x);
5627 else
5629 /* We used to rely on all references to a register becoming
5630 inaccessible when a register changes to a new quantity,
5631 since that changes the hash code. However, that is not
5632 safe, since after HASH_SIZE new quantities we get a
5633 hash 'collision' of a register with its own invalid
5634 entries. And since SUBREGs have been changed not to
5635 change their hash code with the hash code of the register,
5636 it wouldn't work any longer at all. So we have to check
5637 for any invalid references lying around now.
5638 This code is similar to the REG case in mention_regs,
5639 but it knows that reg_tick has been incremented, and
5640 it leaves reg_in_table as -1 . */
5641 unsigned int regno = REGNO (x);
5642 unsigned int endregno = END_REGNO (x);
5643 unsigned int i;
5645 for (i = regno; i < endregno; i++)
5647 if (REG_IN_TABLE (i) >= 0)
5649 remove_invalid_refs (i);
5650 REG_IN_TABLE (i) = -1;
5657 /* We may have just removed some of the src_elt's from the hash table.
5658 So replace each one with the current head of the same class.
5659 Also check if destination addresses have been removed. */
5661 for (i = 0; i < n_sets; i++)
5662 if (sets[i].rtl)
5664 if (sets[i].dest_addr_elt
5665 && sets[i].dest_addr_elt->first_same_value == 0)
5667 /* The elt was removed, which means this destination is not
5668 valid after this instruction. */
5669 sets[i].rtl = NULL_RTX;
5671 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5672 /* If elt was removed, find current head of same class,
5673 or 0 if nothing remains of that class. */
5675 struct table_elt *elt = sets[i].src_elt;
5677 while (elt && elt->prev_same_value)
5678 elt = elt->prev_same_value;
5680 while (elt && elt->first_same_value == 0)
5681 elt = elt->next_same_value;
5682 sets[i].src_elt = elt ? elt->first_same_value : 0;
5686 /* Now insert the destinations into their equivalence classes. */
5688 for (i = 0; i < n_sets; i++)
5689 if (sets[i].rtl)
5691 rtx dest = SET_DEST (sets[i].rtl);
5692 struct table_elt *elt;
5694 /* Don't record value if we are not supposed to risk allocating
5695 floating-point values in registers that might be wider than
5696 memory. */
5697 if ((flag_float_store
5698 && MEM_P (dest)
5699 && FLOAT_MODE_P (GET_MODE (dest)))
5700 /* Don't record BLKmode values, because we don't know the
5701 size of it, and can't be sure that other BLKmode values
5702 have the same or smaller size. */
5703 || GET_MODE (dest) == BLKmode
5704 /* If we didn't put a REG_EQUAL value or a source into the hash
5705 table, there is no point is recording DEST. */
5706 || sets[i].src_elt == 0
5707 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5708 or SIGN_EXTEND, don't record DEST since it can cause
5709 some tracking to be wrong.
5711 ??? Think about this more later. */
5712 || (GET_CODE (dest) == SUBREG
5713 && (GET_MODE_SIZE (GET_MODE (dest))
5714 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5715 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5716 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5717 continue;
5719 /* STRICT_LOW_PART isn't part of the value BEING set,
5720 and neither is the SUBREG inside it.
5721 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5722 if (GET_CODE (dest) == STRICT_LOW_PART)
5723 dest = SUBREG_REG (XEXP (dest, 0));
5725 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5726 /* Registers must also be inserted into chains for quantities. */
5727 if (insert_regs (dest, sets[i].src_elt, 1))
5729 /* If `insert_regs' changes something, the hash code must be
5730 recalculated. */
5731 rehash_using_reg (dest);
5732 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5735 elt = insert (dest, sets[i].src_elt,
5736 sets[i].dest_hash, GET_MODE (dest));
5738 /* If this is a constant, insert the constant anchors with the
5739 equivalent register-offset expressions using register DEST. */
5740 if (targetm.const_anchor
5741 && REG_P (dest)
5742 && SCALAR_INT_MODE_P (GET_MODE (dest))
5743 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5744 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5746 elt->in_memory = (MEM_P (sets[i].inner_dest)
5747 && !MEM_READONLY_P (sets[i].inner_dest));
5749 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5750 narrower than M2, and both M1 and M2 are the same number of words,
5751 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5752 make that equivalence as well.
5754 However, BAR may have equivalences for which gen_lowpart
5755 will produce a simpler value than gen_lowpart applied to
5756 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5757 BAR's equivalences. If we don't get a simplified form, make
5758 the SUBREG. It will not be used in an equivalence, but will
5759 cause two similar assignments to be detected.
5761 Note the loop below will find SUBREG_REG (DEST) since we have
5762 already entered SRC and DEST of the SET in the table. */
5764 if (GET_CODE (dest) == SUBREG
5765 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5766 / UNITS_PER_WORD)
5767 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5768 && (GET_MODE_SIZE (GET_MODE (dest))
5769 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5770 && sets[i].src_elt != 0)
5772 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5773 struct table_elt *elt, *classp = 0;
5775 for (elt = sets[i].src_elt->first_same_value; elt;
5776 elt = elt->next_same_value)
5778 rtx new_src = 0;
5779 unsigned src_hash;
5780 struct table_elt *src_elt;
5781 int byte = 0;
5783 /* Ignore invalid entries. */
5784 if (!REG_P (elt->exp)
5785 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5786 continue;
5788 /* We may have already been playing subreg games. If the
5789 mode is already correct for the destination, use it. */
5790 if (GET_MODE (elt->exp) == new_mode)
5791 new_src = elt->exp;
5792 else
5794 /* Calculate big endian correction for the SUBREG_BYTE.
5795 We have already checked that M1 (GET_MODE (dest))
5796 is not narrower than M2 (new_mode). */
5797 if (BYTES_BIG_ENDIAN)
5798 byte = (GET_MODE_SIZE (GET_MODE (dest))
5799 - GET_MODE_SIZE (new_mode));
5801 new_src = simplify_gen_subreg (new_mode, elt->exp,
5802 GET_MODE (dest), byte);
5805 /* The call to simplify_gen_subreg fails if the value
5806 is VOIDmode, yet we can't do any simplification, e.g.
5807 for EXPR_LISTs denoting function call results.
5808 It is invalid to construct a SUBREG with a VOIDmode
5809 SUBREG_REG, hence a zero new_src means we can't do
5810 this substitution. */
5811 if (! new_src)
5812 continue;
5814 src_hash = HASH (new_src, new_mode);
5815 src_elt = lookup (new_src, src_hash, new_mode);
5817 /* Put the new source in the hash table is if isn't
5818 already. */
5819 if (src_elt == 0)
5821 if (insert_regs (new_src, classp, 0))
5823 rehash_using_reg (new_src);
5824 src_hash = HASH (new_src, new_mode);
5826 src_elt = insert (new_src, classp, src_hash, new_mode);
5827 src_elt->in_memory = elt->in_memory;
5829 else if (classp && classp != src_elt->first_same_value)
5830 /* Show that two things that we've seen before are
5831 actually the same. */
5832 merge_equiv_classes (src_elt, classp);
5834 classp = src_elt->first_same_value;
5835 /* Ignore invalid entries. */
5836 while (classp
5837 && !REG_P (classp->exp)
5838 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5839 classp = classp->next_same_value;
5844 /* Special handling for (set REG0 REG1) where REG0 is the
5845 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5846 be used in the sequel, so (if easily done) change this insn to
5847 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5848 that computed their value. Then REG1 will become a dead store
5849 and won't cloud the situation for later optimizations.
5851 Do not make this change if REG1 is a hard register, because it will
5852 then be used in the sequel and we may be changing a two-operand insn
5853 into a three-operand insn.
5855 Also do not do this if we are operating on a copy of INSN. */
5857 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
5858 && NEXT_INSN (PREV_INSN (insn)) == insn
5859 && REG_P (SET_SRC (sets[0].rtl))
5860 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
5861 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
5863 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
5864 struct qty_table_elem *src_ent = &qty_table[src_q];
5866 if (src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
5868 /* Scan for the previous nonnote insn, but stop at a basic
5869 block boundary. */
5870 rtx prev = insn;
5871 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
5874 prev = PREV_INSN (prev);
5876 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
5878 /* Do not swap the registers around if the previous instruction
5879 attaches a REG_EQUIV note to REG1.
5881 ??? It's not entirely clear whether we can transfer a REG_EQUIV
5882 from the pseudo that originally shadowed an incoming argument
5883 to another register. Some uses of REG_EQUIV might rely on it
5884 being attached to REG1 rather than REG2.
5886 This section previously turned the REG_EQUIV into a REG_EQUAL
5887 note. We cannot do that because REG_EQUIV may provide an
5888 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
5889 if (NONJUMP_INSN_P (prev)
5890 && GET_CODE (PATTERN (prev)) == SET
5891 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
5892 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
5894 rtx dest = SET_DEST (sets[0].rtl);
5895 rtx src = SET_SRC (sets[0].rtl);
5896 rtx note;
5898 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
5899 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
5900 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
5901 apply_change_group ();
5903 /* If INSN has a REG_EQUAL note, and this note mentions
5904 REG0, then we must delete it, because the value in
5905 REG0 has changed. If the note's value is REG1, we must
5906 also delete it because that is now this insn's dest. */
5907 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5908 if (note != 0
5909 && (reg_mentioned_p (dest, XEXP (note, 0))
5910 || rtx_equal_p (src, XEXP (note, 0))))
5911 remove_note (insn, note);
5916 done:;
5919 /* Remove from the hash table all expressions that reference memory. */
5921 static void
5922 invalidate_memory (void)
5924 int i;
5925 struct table_elt *p, *next;
5927 for (i = 0; i < HASH_SIZE; i++)
5928 for (p = table[i]; p; p = next)
5930 next = p->next_same_hash;
5931 if (p->in_memory)
5932 remove_from_table (p, i);
5936 /* Perform invalidation on the basis of everything about an insn
5937 except for invalidating the actual places that are SET in it.
5938 This includes the places CLOBBERed, and anything that might
5939 alias with something that is SET or CLOBBERed.
5941 X is the pattern of the insn. */
5943 static void
5944 invalidate_from_clobbers (rtx x)
5946 if (GET_CODE (x) == CLOBBER)
5948 rtx ref = XEXP (x, 0);
5949 if (ref)
5951 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5952 || MEM_P (ref))
5953 invalidate (ref, VOIDmode);
5954 else if (GET_CODE (ref) == STRICT_LOW_PART
5955 || GET_CODE (ref) == ZERO_EXTRACT)
5956 invalidate (XEXP (ref, 0), GET_MODE (ref));
5959 else if (GET_CODE (x) == PARALLEL)
5961 int i;
5962 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5964 rtx y = XVECEXP (x, 0, i);
5965 if (GET_CODE (y) == CLOBBER)
5967 rtx ref = XEXP (y, 0);
5968 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5969 || MEM_P (ref))
5970 invalidate (ref, VOIDmode);
5971 else if (GET_CODE (ref) == STRICT_LOW_PART
5972 || GET_CODE (ref) == ZERO_EXTRACT)
5973 invalidate (XEXP (ref, 0), GET_MODE (ref));
5979 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
5980 and replace any registers in them with either an equivalent constant
5981 or the canonical form of the register. If we are inside an address,
5982 only do this if the address remains valid.
5984 OBJECT is 0 except when within a MEM in which case it is the MEM.
5986 Return the replacement for X. */
5988 static rtx
5989 cse_process_notes_1 (rtx x, rtx object, bool *changed)
5991 enum rtx_code code = GET_CODE (x);
5992 const char *fmt = GET_RTX_FORMAT (code);
5993 int i;
5995 switch (code)
5997 case CONST_INT:
5998 case CONST:
5999 case SYMBOL_REF:
6000 case LABEL_REF:
6001 case CONST_DOUBLE:
6002 case CONST_FIXED:
6003 case CONST_VECTOR:
6004 case PC:
6005 case CC0:
6006 case LO_SUM:
6007 return x;
6009 case MEM:
6010 validate_change (x, &XEXP (x, 0),
6011 cse_process_notes (XEXP (x, 0), x, changed), 0);
6012 return x;
6014 case EXPR_LIST:
6015 case INSN_LIST:
6016 if (REG_NOTE_KIND (x) == REG_EQUAL)
6017 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6018 if (XEXP (x, 1))
6019 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6020 return x;
6022 case SIGN_EXTEND:
6023 case ZERO_EXTEND:
6024 case SUBREG:
6026 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6027 /* We don't substitute VOIDmode constants into these rtx,
6028 since they would impede folding. */
6029 if (GET_MODE (new_rtx) != VOIDmode)
6030 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6031 return x;
6034 case REG:
6035 i = REG_QTY (REGNO (x));
6037 /* Return a constant or a constant register. */
6038 if (REGNO_QTY_VALID_P (REGNO (x)))
6040 struct qty_table_elem *ent = &qty_table[i];
6042 if (ent->const_rtx != NULL_RTX
6043 && (CONSTANT_P (ent->const_rtx)
6044 || REG_P (ent->const_rtx)))
6046 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6047 if (new_rtx)
6048 return copy_rtx (new_rtx);
6052 /* Otherwise, canonicalize this register. */
6053 return canon_reg (x, NULL_RTX);
6055 default:
6056 break;
6059 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6060 if (fmt[i] == 'e')
6061 validate_change (object, &XEXP (x, i),
6062 cse_process_notes (XEXP (x, i), object, changed), 0);
6064 return x;
6067 static rtx
6068 cse_process_notes (rtx x, rtx object, bool *changed)
6070 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6071 if (new_rtx != x)
6072 *changed = true;
6073 return new_rtx;
6077 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6079 DATA is a pointer to a struct cse_basic_block_data, that is used to
6080 describe the path.
6081 It is filled with a queue of basic blocks, starting with FIRST_BB
6082 and following a trace through the CFG.
6084 If all paths starting at FIRST_BB have been followed, or no new path
6085 starting at FIRST_BB can be constructed, this function returns FALSE.
6086 Otherwise, DATA->path is filled and the function returns TRUE indicating
6087 that a path to follow was found.
6089 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6090 block in the path will be FIRST_BB. */
6092 static bool
6093 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6094 int follow_jumps)
6096 basic_block bb;
6097 edge e;
6098 int path_size;
6100 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6102 /* See if there is a previous path. */
6103 path_size = data->path_size;
6105 /* There is a previous path. Make sure it started with FIRST_BB. */
6106 if (path_size)
6107 gcc_assert (data->path[0].bb == first_bb);
6109 /* There was only one basic block in the last path. Clear the path and
6110 return, so that paths starting at another basic block can be tried. */
6111 if (path_size == 1)
6113 path_size = 0;
6114 goto done;
6117 /* If the path was empty from the beginning, construct a new path. */
6118 if (path_size == 0)
6119 data->path[path_size++].bb = first_bb;
6120 else
6122 /* Otherwise, path_size must be equal to or greater than 2, because
6123 a previous path exists that is at least two basic blocks long.
6125 Update the previous branch path, if any. If the last branch was
6126 previously along the branch edge, take the fallthrough edge now. */
6127 while (path_size >= 2)
6129 basic_block last_bb_in_path, previous_bb_in_path;
6130 edge e;
6132 --path_size;
6133 last_bb_in_path = data->path[path_size].bb;
6134 previous_bb_in_path = data->path[path_size - 1].bb;
6136 /* If we previously followed a path along the branch edge, try
6137 the fallthru edge now. */
6138 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6139 && any_condjump_p (BB_END (previous_bb_in_path))
6140 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6141 && e == BRANCH_EDGE (previous_bb_in_path))
6143 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6144 if (bb != EXIT_BLOCK_PTR
6145 && single_pred_p (bb)
6146 /* We used to assert here that we would only see blocks
6147 that we have not visited yet. But we may end up
6148 visiting basic blocks twice if the CFG has changed
6149 in this run of cse_main, because when the CFG changes
6150 the topological sort of the CFG also changes. A basic
6151 blocks that previously had more than two predecessors
6152 may now have a single predecessor, and become part of
6153 a path that starts at another basic block.
6155 We still want to visit each basic block only once, so
6156 halt the path here if we have already visited BB. */
6157 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6159 SET_BIT (cse_visited_basic_blocks, bb->index);
6160 data->path[path_size++].bb = bb;
6161 break;
6165 data->path[path_size].bb = NULL;
6168 /* If only one block remains in the path, bail. */
6169 if (path_size == 1)
6171 path_size = 0;
6172 goto done;
6176 /* Extend the path if possible. */
6177 if (follow_jumps)
6179 bb = data->path[path_size - 1].bb;
6180 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6182 if (single_succ_p (bb))
6183 e = single_succ_edge (bb);
6184 else if (EDGE_COUNT (bb->succs) == 2
6185 && any_condjump_p (BB_END (bb)))
6187 /* First try to follow the branch. If that doesn't lead
6188 to a useful path, follow the fallthru edge. */
6189 e = BRANCH_EDGE (bb);
6190 if (!single_pred_p (e->dest))
6191 e = FALLTHRU_EDGE (bb);
6193 else
6194 e = NULL;
6196 if (e && e->dest != EXIT_BLOCK_PTR
6197 && single_pred_p (e->dest)
6198 /* Avoid visiting basic blocks twice. The large comment
6199 above explains why this can happen. */
6200 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6202 basic_block bb2 = e->dest;
6203 SET_BIT (cse_visited_basic_blocks, bb2->index);
6204 data->path[path_size++].bb = bb2;
6205 bb = bb2;
6207 else
6208 bb = NULL;
6212 done:
6213 data->path_size = path_size;
6214 return path_size != 0;
6217 /* Dump the path in DATA to file F. NSETS is the number of sets
6218 in the path. */
6220 static void
6221 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6223 int path_entry;
6225 fprintf (f, ";; Following path with %d sets: ", nsets);
6226 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6227 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6228 fputc ('\n', dump_file);
6229 fflush (f);
6233 /* Return true if BB has exception handling successor edges. */
6235 static bool
6236 have_eh_succ_edges (basic_block bb)
6238 edge e;
6239 edge_iterator ei;
6241 FOR_EACH_EDGE (e, ei, bb->succs)
6242 if (e->flags & EDGE_EH)
6243 return true;
6245 return false;
6249 /* Scan to the end of the path described by DATA. Return an estimate of
6250 the total number of SETs of all insns in the path. */
6252 static void
6253 cse_prescan_path (struct cse_basic_block_data *data)
6255 int nsets = 0;
6256 int path_size = data->path_size;
6257 int path_entry;
6259 /* Scan to end of each basic block in the path. */
6260 for (path_entry = 0; path_entry < path_size; path_entry++)
6262 basic_block bb;
6263 rtx insn;
6265 bb = data->path[path_entry].bb;
6267 FOR_BB_INSNS (bb, insn)
6269 if (!INSN_P (insn))
6270 continue;
6272 /* A PARALLEL can have lots of SETs in it,
6273 especially if it is really an ASM_OPERANDS. */
6274 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6275 nsets += XVECLEN (PATTERN (insn), 0);
6276 else
6277 nsets += 1;
6281 data->nsets = nsets;
6284 /* Process a single extended basic block described by EBB_DATA. */
6286 static void
6287 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6289 int path_size = ebb_data->path_size;
6290 int path_entry;
6291 int num_insns = 0;
6293 /* Allocate the space needed by qty_table. */
6294 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6296 new_basic_block ();
6297 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6298 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6299 for (path_entry = 0; path_entry < path_size; path_entry++)
6301 basic_block bb;
6302 rtx insn;
6304 bb = ebb_data->path[path_entry].bb;
6306 /* Invalidate recorded information for eh regs if there is an EH
6307 edge pointing to that bb. */
6308 if (bb_has_eh_pred (bb))
6310 df_ref *def_rec;
6312 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6314 df_ref def = *def_rec;
6315 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6316 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6320 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6321 FOR_BB_INSNS (bb, insn)
6323 /* If we have processed 1,000 insns, flush the hash table to
6324 avoid extreme quadratic behavior. We must not include NOTEs
6325 in the count since there may be more of them when generating
6326 debugging information. If we clear the table at different
6327 times, code generated with -g -O might be different than code
6328 generated with -O but not -g.
6330 FIXME: This is a real kludge and needs to be done some other
6331 way. */
6332 if (NONDEBUG_INSN_P (insn)
6333 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6335 flush_hash_table ();
6336 num_insns = 0;
6339 if (INSN_P (insn))
6341 /* Process notes first so we have all notes in canonical forms
6342 when looking for duplicate operations. */
6343 if (REG_NOTES (insn))
6345 bool changed = false;
6346 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6347 NULL_RTX, &changed);
6348 if (changed)
6349 df_notes_rescan (insn);
6352 cse_insn (insn);
6354 /* If we haven't already found an insn where we added a LABEL_REF,
6355 check this one. */
6356 if (INSN_P (insn) && !recorded_label_ref
6357 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6358 (void *) insn))
6359 recorded_label_ref = true;
6361 #ifdef HAVE_cc0
6362 if (NONDEBUG_INSN_P (insn))
6364 /* If the previous insn sets CC0 and this insn no
6365 longer references CC0, delete the previous insn.
6366 Here we use fact that nothing expects CC0 to be
6367 valid over an insn, which is true until the final
6368 pass. */
6369 rtx prev_insn, tem;
6371 prev_insn = prev_nonnote_nondebug_insn (insn);
6372 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6373 && (tem = single_set (prev_insn)) != NULL_RTX
6374 && SET_DEST (tem) == cc0_rtx
6375 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6376 delete_insn (prev_insn);
6378 /* If this insn is not the last insn in the basic
6379 block, it will be PREV_INSN(insn) in the next
6380 iteration. If we recorded any CC0-related
6381 information for this insn, remember it. */
6382 if (insn != BB_END (bb))
6384 prev_insn_cc0 = this_insn_cc0;
6385 prev_insn_cc0_mode = this_insn_cc0_mode;
6388 #endif
6392 /* With non-call exceptions, we are not always able to update
6393 the CFG properly inside cse_insn. So clean up possibly
6394 redundant EH edges here. */
6395 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6396 cse_cfg_altered |= purge_dead_edges (bb);
6398 /* If we changed a conditional jump, we may have terminated
6399 the path we are following. Check that by verifying that
6400 the edge we would take still exists. If the edge does
6401 not exist anymore, purge the remainder of the path.
6402 Note that this will cause us to return to the caller. */
6403 if (path_entry < path_size - 1)
6405 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6406 if (!find_edge (bb, next_bb))
6410 path_size--;
6412 /* If we truncate the path, we must also reset the
6413 visited bit on the remaining blocks in the path,
6414 or we will never visit them at all. */
6415 RESET_BIT (cse_visited_basic_blocks,
6416 ebb_data->path[path_size].bb->index);
6417 ebb_data->path[path_size].bb = NULL;
6419 while (path_size - 1 != path_entry);
6420 ebb_data->path_size = path_size;
6424 /* If this is a conditional jump insn, record any known
6425 equivalences due to the condition being tested. */
6426 insn = BB_END (bb);
6427 if (path_entry < path_size - 1
6428 && JUMP_P (insn)
6429 && single_set (insn)
6430 && any_condjump_p (insn))
6432 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6433 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6434 record_jump_equiv (insn, taken);
6437 #ifdef HAVE_cc0
6438 /* Clear the CC0-tracking related insns, they can't provide
6439 useful information across basic block boundaries. */
6440 prev_insn_cc0 = 0;
6441 #endif
6444 gcc_assert (next_qty <= max_qty);
6446 free (qty_table);
6450 /* Perform cse on the instructions of a function.
6451 F is the first instruction.
6452 NREGS is one plus the highest pseudo-reg number used in the instruction.
6454 Return 2 if jump optimizations should be redone due to simplifications
6455 in conditional jump instructions.
6456 Return 1 if the CFG should be cleaned up because it has been modified.
6457 Return 0 otherwise. */
6460 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6462 struct cse_basic_block_data ebb_data;
6463 basic_block bb;
6464 int *rc_order = XNEWVEC (int, last_basic_block);
6465 int i, n_blocks;
6467 df_set_flags (DF_LR_RUN_DCE);
6468 df_analyze ();
6469 df_set_flags (DF_DEFER_INSN_RESCAN);
6471 reg_scan (get_insns (), max_reg_num ());
6472 init_cse_reg_info (nregs);
6474 ebb_data.path = XNEWVEC (struct branch_path,
6475 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6477 cse_cfg_altered = false;
6478 cse_jumps_altered = false;
6479 recorded_label_ref = false;
6480 constant_pool_entries_cost = 0;
6481 constant_pool_entries_regcost = 0;
6482 ebb_data.path_size = 0;
6483 ebb_data.nsets = 0;
6484 rtl_hooks = cse_rtl_hooks;
6486 init_recog ();
6487 init_alias_analysis ();
6489 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6491 /* Set up the table of already visited basic blocks. */
6492 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6493 sbitmap_zero (cse_visited_basic_blocks);
6495 /* Loop over basic blocks in reverse completion order (RPO),
6496 excluding the ENTRY and EXIT blocks. */
6497 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6498 i = 0;
6499 while (i < n_blocks)
6501 /* Find the first block in the RPO queue that we have not yet
6502 processed before. */
6505 bb = BASIC_BLOCK (rc_order[i++]);
6507 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6508 && i < n_blocks);
6510 /* Find all paths starting with BB, and process them. */
6511 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6513 /* Pre-scan the path. */
6514 cse_prescan_path (&ebb_data);
6516 /* If this basic block has no sets, skip it. */
6517 if (ebb_data.nsets == 0)
6518 continue;
6520 /* Get a reasonable estimate for the maximum number of qty's
6521 needed for this path. For this, we take the number of sets
6522 and multiply that by MAX_RECOG_OPERANDS. */
6523 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6525 /* Dump the path we're about to process. */
6526 if (dump_file)
6527 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6529 cse_extended_basic_block (&ebb_data);
6533 /* Clean up. */
6534 end_alias_analysis ();
6535 free (reg_eqv_table);
6536 free (ebb_data.path);
6537 sbitmap_free (cse_visited_basic_blocks);
6538 free (rc_order);
6539 rtl_hooks = general_rtl_hooks;
6541 if (cse_jumps_altered || recorded_label_ref)
6542 return 2;
6543 else if (cse_cfg_altered)
6544 return 1;
6545 else
6546 return 0;
6549 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6550 which there isn't a REG_LABEL_OPERAND note.
6551 Return one if so. DATA is the insn. */
6553 static int
6554 check_for_label_ref (rtx *rtl, void *data)
6556 rtx insn = (rtx) data;
6558 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6559 note for it, we must rerun jump since it needs to place the note. If
6560 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6561 don't do this since no REG_LABEL_OPERAND will be added. */
6562 return (GET_CODE (*rtl) == LABEL_REF
6563 && ! LABEL_REF_NONLOCAL_P (*rtl)
6564 && (!JUMP_P (insn)
6565 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6566 && LABEL_P (XEXP (*rtl, 0))
6567 && INSN_UID (XEXP (*rtl, 0)) != 0
6568 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6571 /* Count the number of times registers are used (not set) in X.
6572 COUNTS is an array in which we accumulate the count, INCR is how much
6573 we count each register usage.
6575 Don't count a usage of DEST, which is the SET_DEST of a SET which
6576 contains X in its SET_SRC. This is because such a SET does not
6577 modify the liveness of DEST.
6578 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6579 We must then count uses of a SET_DEST regardless, because the insn can't be
6580 deleted here. */
6582 static void
6583 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6585 enum rtx_code code;
6586 rtx note;
6587 const char *fmt;
6588 int i, j;
6590 if (x == 0)
6591 return;
6593 switch (code = GET_CODE (x))
6595 case REG:
6596 if (x != dest)
6597 counts[REGNO (x)] += incr;
6598 return;
6600 case PC:
6601 case CC0:
6602 case CONST:
6603 case CONST_INT:
6604 case CONST_DOUBLE:
6605 case CONST_FIXED:
6606 case CONST_VECTOR:
6607 case SYMBOL_REF:
6608 case LABEL_REF:
6609 return;
6611 case CLOBBER:
6612 /* If we are clobbering a MEM, mark any registers inside the address
6613 as being used. */
6614 if (MEM_P (XEXP (x, 0)))
6615 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6616 return;
6618 case SET:
6619 /* Unless we are setting a REG, count everything in SET_DEST. */
6620 if (!REG_P (SET_DEST (x)))
6621 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6622 count_reg_usage (SET_SRC (x), counts,
6623 dest ? dest : SET_DEST (x),
6624 incr);
6625 return;
6627 case DEBUG_INSN:
6628 return;
6630 case CALL_INSN:
6631 case INSN:
6632 case JUMP_INSN:
6633 /* We expect dest to be NULL_RTX here. If the insn may trap,
6634 or if it cannot be deleted due to side-effects, mark this fact
6635 by setting DEST to pc_rtx. */
6636 if (insn_could_throw_p (x) || side_effects_p (PATTERN (x)))
6637 dest = pc_rtx;
6638 if (code == CALL_INSN)
6639 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6640 count_reg_usage (PATTERN (x), counts, dest, incr);
6642 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6643 use them. */
6645 note = find_reg_equal_equiv_note (x);
6646 if (note)
6648 rtx eqv = XEXP (note, 0);
6650 if (GET_CODE (eqv) == EXPR_LIST)
6651 /* This REG_EQUAL note describes the result of a function call.
6652 Process all the arguments. */
6655 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6656 eqv = XEXP (eqv, 1);
6658 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6659 else
6660 count_reg_usage (eqv, counts, dest, incr);
6662 return;
6664 case EXPR_LIST:
6665 if (REG_NOTE_KIND (x) == REG_EQUAL
6666 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6667 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6668 involving registers in the address. */
6669 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6670 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6672 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6673 return;
6675 case ASM_OPERANDS:
6676 /* Iterate over just the inputs, not the constraints as well. */
6677 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6678 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6679 return;
6681 case INSN_LIST:
6682 gcc_unreachable ();
6684 default:
6685 break;
6688 fmt = GET_RTX_FORMAT (code);
6689 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6691 if (fmt[i] == 'e')
6692 count_reg_usage (XEXP (x, i), counts, dest, incr);
6693 else if (fmt[i] == 'E')
6694 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6695 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6699 /* Return true if X is a dead register. */
6701 static inline int
6702 is_dead_reg (rtx x, int *counts)
6704 return (REG_P (x)
6705 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6706 && counts[REGNO (x)] == 0);
6709 /* Return true if set is live. */
6710 static bool
6711 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6712 int *counts)
6714 #ifdef HAVE_cc0
6715 rtx tem;
6716 #endif
6718 if (set_noop_p (set))
6721 #ifdef HAVE_cc0
6722 else if (GET_CODE (SET_DEST (set)) == CC0
6723 && !side_effects_p (SET_SRC (set))
6724 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6725 || !INSN_P (tem)
6726 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6727 return false;
6728 #endif
6729 else if (!is_dead_reg (SET_DEST (set), counts)
6730 || side_effects_p (SET_SRC (set)))
6731 return true;
6732 return false;
6735 /* Return true if insn is live. */
6737 static bool
6738 insn_live_p (rtx insn, int *counts)
6740 int i;
6741 if (insn_could_throw_p (insn))
6742 return true;
6743 else if (GET_CODE (PATTERN (insn)) == SET)
6744 return set_live_p (PATTERN (insn), insn, counts);
6745 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6747 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6749 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6751 if (GET_CODE (elt) == SET)
6753 if (set_live_p (elt, insn, counts))
6754 return true;
6756 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6757 return true;
6759 return false;
6761 else if (DEBUG_INSN_P (insn))
6763 rtx next;
6765 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6766 if (NOTE_P (next))
6767 continue;
6768 else if (!DEBUG_INSN_P (next))
6769 return true;
6770 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6771 return false;
6773 return true;
6775 else
6776 return true;
6779 /* Count the number of stores into pseudo. Callback for note_stores. */
6781 static void
6782 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6784 int *counts = (int *) data;
6785 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6786 counts[REGNO (x)]++;
6789 struct dead_debug_insn_data
6791 int *counts;
6792 rtx *replacements;
6793 bool seen_repl;
6796 /* Return if a DEBUG_INSN needs to be reset because some dead
6797 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6799 static int
6800 is_dead_debug_insn (rtx *loc, void *data)
6802 rtx x = *loc;
6803 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6805 if (is_dead_reg (x, ddid->counts))
6807 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6808 ddid->seen_repl = true;
6809 else
6810 return 1;
6812 return 0;
6815 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6816 Callback for simplify_replace_fn_rtx. */
6818 static rtx
6819 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6821 rtx *replacements = (rtx *) data;
6823 if (REG_P (x)
6824 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6825 && replacements[REGNO (x)] != NULL_RTX)
6827 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6828 return replacements[REGNO (x)];
6829 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6830 GET_MODE (replacements[REGNO (x)]));
6832 return NULL_RTX;
6835 /* Scan all the insns and delete any that are dead; i.e., they store a register
6836 that is never used or they copy a register to itself.
6838 This is used to remove insns made obviously dead by cse, loop or other
6839 optimizations. It improves the heuristics in loop since it won't try to
6840 move dead invariants out of loops or make givs for dead quantities. The
6841 remaining passes of the compilation are also sped up. */
6844 delete_trivially_dead_insns (rtx insns, int nreg)
6846 int *counts;
6847 rtx insn, prev;
6848 rtx *replacements = NULL;
6849 int ndead = 0;
6851 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6852 /* First count the number of times each register is used. */
6853 if (MAY_HAVE_DEBUG_INSNS)
6855 counts = XCNEWVEC (int, nreg * 3);
6856 for (insn = insns; insn; insn = NEXT_INSN (insn))
6857 if (DEBUG_INSN_P (insn))
6858 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6859 NULL_RTX, 1);
6860 else if (INSN_P (insn))
6862 count_reg_usage (insn, counts, NULL_RTX, 1);
6863 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6865 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6866 First one counts how many times each pseudo is used outside
6867 of debug insns, second counts how many times each pseudo is
6868 used in debug insns and third counts how many times a pseudo
6869 is stored. */
6871 else
6873 counts = XCNEWVEC (int, nreg);
6874 for (insn = insns; insn; insn = NEXT_INSN (insn))
6875 if (INSN_P (insn))
6876 count_reg_usage (insn, counts, NULL_RTX, 1);
6877 /* If no debug insns can be present, COUNTS is just an array
6878 which counts how many times each pseudo is used. */
6880 /* Go from the last insn to the first and delete insns that only set unused
6881 registers or copy a register to itself. As we delete an insn, remove
6882 usage counts for registers it uses.
6884 The first jump optimization pass may leave a real insn as the last
6885 insn in the function. We must not skip that insn or we may end
6886 up deleting code that is not really dead.
6888 If some otherwise unused register is only used in DEBUG_INSNs,
6889 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6890 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6891 has been created for the unused register, replace it with
6892 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6893 for (insn = get_last_insn (); insn; insn = prev)
6895 int live_insn = 0;
6897 prev = PREV_INSN (insn);
6898 if (!INSN_P (insn))
6899 continue;
6901 live_insn = insn_live_p (insn, counts);
6903 /* If this is a dead insn, delete it and show registers in it aren't
6904 being used. */
6906 if (! live_insn && dbg_cnt (delete_trivial_dead))
6908 if (DEBUG_INSN_P (insn))
6909 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6910 NULL_RTX, -1);
6911 else
6913 rtx set;
6914 if (MAY_HAVE_DEBUG_INSNS
6915 && (set = single_set (insn)) != NULL_RTX
6916 && is_dead_reg (SET_DEST (set), counts)
6917 /* Used at least once in some DEBUG_INSN. */
6918 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6919 /* And set exactly once. */
6920 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6921 && !side_effects_p (SET_SRC (set))
6922 && asm_noperands (PATTERN (insn)) < 0)
6924 rtx dval, bind;
6926 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6927 dval = make_debug_expr_from_rtl (SET_DEST (set));
6929 /* Emit a debug bind insn before the insn in which
6930 reg dies. */
6931 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
6932 DEBUG_EXPR_TREE_DECL (dval),
6933 SET_SRC (set),
6934 VAR_INIT_STATUS_INITIALIZED);
6935 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
6937 bind = emit_debug_insn_before (bind, insn);
6938 df_insn_rescan (bind);
6940 if (replacements == NULL)
6941 replacements = XCNEWVEC (rtx, nreg);
6942 replacements[REGNO (SET_DEST (set))] = dval;
6945 count_reg_usage (insn, counts, NULL_RTX, -1);
6946 ndead++;
6948 delete_insn_and_edges (insn);
6952 if (MAY_HAVE_DEBUG_INSNS)
6954 struct dead_debug_insn_data ddid;
6955 ddid.counts = counts;
6956 ddid.replacements = replacements;
6957 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
6958 if (DEBUG_INSN_P (insn))
6960 /* If this debug insn references a dead register that wasn't replaced
6961 with an DEBUG_EXPR, reset the DEBUG_INSN. */
6962 ddid.seen_repl = false;
6963 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
6964 is_dead_debug_insn, &ddid))
6966 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
6967 df_insn_rescan (insn);
6969 else if (ddid.seen_repl)
6971 INSN_VAR_LOCATION_LOC (insn)
6972 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
6973 NULL_RTX, replace_dead_reg,
6974 replacements);
6975 df_insn_rescan (insn);
6978 if (replacements)
6979 free (replacements);
6982 if (dump_file && ndead)
6983 fprintf (dump_file, "Deleted %i trivially dead insns\n",
6984 ndead);
6985 /* Clean up. */
6986 free (counts);
6987 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
6988 return ndead;
6991 /* This function is called via for_each_rtx. The argument, NEWREG, is
6992 a condition code register with the desired mode. If we are looking
6993 at the same register in a different mode, replace it with
6994 NEWREG. */
6996 static int
6997 cse_change_cc_mode (rtx *loc, void *data)
6999 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7001 if (*loc
7002 && REG_P (*loc)
7003 && REGNO (*loc) == REGNO (args->newreg)
7004 && GET_MODE (*loc) != GET_MODE (args->newreg))
7006 validate_change (args->insn, loc, args->newreg, 1);
7008 return -1;
7010 return 0;
7013 /* Change the mode of any reference to the register REGNO (NEWREG) to
7014 GET_MODE (NEWREG) in INSN. */
7016 static void
7017 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7019 struct change_cc_mode_args args;
7020 int success;
7022 if (!INSN_P (insn))
7023 return;
7025 args.insn = insn;
7026 args.newreg = newreg;
7028 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7029 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7031 /* If the following assertion was triggered, there is most probably
7032 something wrong with the cc_modes_compatible back end function.
7033 CC modes only can be considered compatible if the insn - with the mode
7034 replaced by any of the compatible modes - can still be recognized. */
7035 success = apply_change_group ();
7036 gcc_assert (success);
7039 /* Change the mode of any reference to the register REGNO (NEWREG) to
7040 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7041 any instruction which modifies NEWREG. */
7043 static void
7044 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7046 rtx insn;
7048 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7050 if (! INSN_P (insn))
7051 continue;
7053 if (reg_set_p (newreg, insn))
7054 return;
7056 cse_change_cc_mode_insn (insn, newreg);
7060 /* BB is a basic block which finishes with CC_REG as a condition code
7061 register which is set to CC_SRC. Look through the successors of BB
7062 to find blocks which have a single predecessor (i.e., this one),
7063 and look through those blocks for an assignment to CC_REG which is
7064 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7065 permitted to change the mode of CC_SRC to a compatible mode. This
7066 returns VOIDmode if no equivalent assignments were found.
7067 Otherwise it returns the mode which CC_SRC should wind up with.
7068 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7069 but is passed unmodified down to recursive calls in order to prevent
7070 endless recursion.
7072 The main complexity in this function is handling the mode issues.
7073 We may have more than one duplicate which we can eliminate, and we
7074 try to find a mode which will work for multiple duplicates. */
7076 static enum machine_mode
7077 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7078 bool can_change_mode)
7080 bool found_equiv;
7081 enum machine_mode mode;
7082 unsigned int insn_count;
7083 edge e;
7084 rtx insns[2];
7085 enum machine_mode modes[2];
7086 rtx last_insns[2];
7087 unsigned int i;
7088 rtx newreg;
7089 edge_iterator ei;
7091 /* We expect to have two successors. Look at both before picking
7092 the final mode for the comparison. If we have more successors
7093 (i.e., some sort of table jump, although that seems unlikely),
7094 then we require all beyond the first two to use the same
7095 mode. */
7097 found_equiv = false;
7098 mode = GET_MODE (cc_src);
7099 insn_count = 0;
7100 FOR_EACH_EDGE (e, ei, bb->succs)
7102 rtx insn;
7103 rtx end;
7105 if (e->flags & EDGE_COMPLEX)
7106 continue;
7108 if (EDGE_COUNT (e->dest->preds) != 1
7109 || e->dest == EXIT_BLOCK_PTR
7110 /* Avoid endless recursion on unreachable blocks. */
7111 || e->dest == orig_bb)
7112 continue;
7114 end = NEXT_INSN (BB_END (e->dest));
7115 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7117 rtx set;
7119 if (! INSN_P (insn))
7120 continue;
7122 /* If CC_SRC is modified, we have to stop looking for
7123 something which uses it. */
7124 if (modified_in_p (cc_src, insn))
7125 break;
7127 /* Check whether INSN sets CC_REG to CC_SRC. */
7128 set = single_set (insn);
7129 if (set
7130 && REG_P (SET_DEST (set))
7131 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7133 bool found;
7134 enum machine_mode set_mode;
7135 enum machine_mode comp_mode;
7137 found = false;
7138 set_mode = GET_MODE (SET_SRC (set));
7139 comp_mode = set_mode;
7140 if (rtx_equal_p (cc_src, SET_SRC (set)))
7141 found = true;
7142 else if (GET_CODE (cc_src) == COMPARE
7143 && GET_CODE (SET_SRC (set)) == COMPARE
7144 && mode != set_mode
7145 && rtx_equal_p (XEXP (cc_src, 0),
7146 XEXP (SET_SRC (set), 0))
7147 && rtx_equal_p (XEXP (cc_src, 1),
7148 XEXP (SET_SRC (set), 1)))
7151 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7152 if (comp_mode != VOIDmode
7153 && (can_change_mode || comp_mode == mode))
7154 found = true;
7157 if (found)
7159 found_equiv = true;
7160 if (insn_count < ARRAY_SIZE (insns))
7162 insns[insn_count] = insn;
7163 modes[insn_count] = set_mode;
7164 last_insns[insn_count] = end;
7165 ++insn_count;
7167 if (mode != comp_mode)
7169 gcc_assert (can_change_mode);
7170 mode = comp_mode;
7172 /* The modified insn will be re-recognized later. */
7173 PUT_MODE (cc_src, mode);
7176 else
7178 if (set_mode != mode)
7180 /* We found a matching expression in the
7181 wrong mode, but we don't have room to
7182 store it in the array. Punt. This case
7183 should be rare. */
7184 break;
7186 /* INSN sets CC_REG to a value equal to CC_SRC
7187 with the right mode. We can simply delete
7188 it. */
7189 delete_insn (insn);
7192 /* We found an instruction to delete. Keep looking,
7193 in the hopes of finding a three-way jump. */
7194 continue;
7197 /* We found an instruction which sets the condition
7198 code, so don't look any farther. */
7199 break;
7202 /* If INSN sets CC_REG in some other way, don't look any
7203 farther. */
7204 if (reg_set_p (cc_reg, insn))
7205 break;
7208 /* If we fell off the bottom of the block, we can keep looking
7209 through successors. We pass CAN_CHANGE_MODE as false because
7210 we aren't prepared to handle compatibility between the
7211 further blocks and this block. */
7212 if (insn == end)
7214 enum machine_mode submode;
7216 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7217 if (submode != VOIDmode)
7219 gcc_assert (submode == mode);
7220 found_equiv = true;
7221 can_change_mode = false;
7226 if (! found_equiv)
7227 return VOIDmode;
7229 /* Now INSN_COUNT is the number of instructions we found which set
7230 CC_REG to a value equivalent to CC_SRC. The instructions are in
7231 INSNS. The modes used by those instructions are in MODES. */
7233 newreg = NULL_RTX;
7234 for (i = 0; i < insn_count; ++i)
7236 if (modes[i] != mode)
7238 /* We need to change the mode of CC_REG in INSNS[i] and
7239 subsequent instructions. */
7240 if (! newreg)
7242 if (GET_MODE (cc_reg) == mode)
7243 newreg = cc_reg;
7244 else
7245 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7247 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7248 newreg);
7251 delete_insn_and_edges (insns[i]);
7254 return mode;
7257 /* If we have a fixed condition code register (or two), walk through
7258 the instructions and try to eliminate duplicate assignments. */
7260 static void
7261 cse_condition_code_reg (void)
7263 unsigned int cc_regno_1;
7264 unsigned int cc_regno_2;
7265 rtx cc_reg_1;
7266 rtx cc_reg_2;
7267 basic_block bb;
7269 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7270 return;
7272 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7273 if (cc_regno_2 != INVALID_REGNUM)
7274 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7275 else
7276 cc_reg_2 = NULL_RTX;
7278 FOR_EACH_BB (bb)
7280 rtx last_insn;
7281 rtx cc_reg;
7282 rtx insn;
7283 rtx cc_src_insn;
7284 rtx cc_src;
7285 enum machine_mode mode;
7286 enum machine_mode orig_mode;
7288 /* Look for blocks which end with a conditional jump based on a
7289 condition code register. Then look for the instruction which
7290 sets the condition code register. Then look through the
7291 successor blocks for instructions which set the condition
7292 code register to the same value. There are other possible
7293 uses of the condition code register, but these are by far the
7294 most common and the ones which we are most likely to be able
7295 to optimize. */
7297 last_insn = BB_END (bb);
7298 if (!JUMP_P (last_insn))
7299 continue;
7301 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7302 cc_reg = cc_reg_1;
7303 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7304 cc_reg = cc_reg_2;
7305 else
7306 continue;
7308 cc_src_insn = NULL_RTX;
7309 cc_src = NULL_RTX;
7310 for (insn = PREV_INSN (last_insn);
7311 insn && insn != PREV_INSN (BB_HEAD (bb));
7312 insn = PREV_INSN (insn))
7314 rtx set;
7316 if (! INSN_P (insn))
7317 continue;
7318 set = single_set (insn);
7319 if (set
7320 && REG_P (SET_DEST (set))
7321 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7323 cc_src_insn = insn;
7324 cc_src = SET_SRC (set);
7325 break;
7327 else if (reg_set_p (cc_reg, insn))
7328 break;
7331 if (! cc_src_insn)
7332 continue;
7334 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7335 continue;
7337 /* Now CC_REG is a condition code register used for a
7338 conditional jump at the end of the block, and CC_SRC, in
7339 CC_SRC_INSN, is the value to which that condition code
7340 register is set, and CC_SRC is still meaningful at the end of
7341 the basic block. */
7343 orig_mode = GET_MODE (cc_src);
7344 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7345 if (mode != VOIDmode)
7347 gcc_assert (mode == GET_MODE (cc_src));
7348 if (mode != orig_mode)
7350 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7352 cse_change_cc_mode_insn (cc_src_insn, newreg);
7354 /* Do the same in the following insns that use the
7355 current value of CC_REG within BB. */
7356 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7357 NEXT_INSN (last_insn),
7358 newreg);
7365 /* Perform common subexpression elimination. Nonzero value from
7366 `cse_main' means that jumps were simplified and some code may now
7367 be unreachable, so do jump optimization again. */
7368 static bool
7369 gate_handle_cse (void)
7371 return optimize > 0;
7374 static unsigned int
7375 rest_of_handle_cse (void)
7377 int tem;
7379 if (dump_file)
7380 dump_flow_info (dump_file, dump_flags);
7382 tem = cse_main (get_insns (), max_reg_num ());
7384 /* If we are not running more CSE passes, then we are no longer
7385 expecting CSE to be run. But always rerun it in a cheap mode. */
7386 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7388 if (tem == 2)
7390 timevar_push (TV_JUMP);
7391 rebuild_jump_labels (get_insns ());
7392 cleanup_cfg (0);
7393 timevar_pop (TV_JUMP);
7395 else if (tem == 1 || optimize > 1)
7396 cleanup_cfg (0);
7398 return 0;
7401 struct rtl_opt_pass pass_cse =
7404 RTL_PASS,
7405 "cse1", /* name */
7406 gate_handle_cse, /* gate */
7407 rest_of_handle_cse, /* execute */
7408 NULL, /* sub */
7409 NULL, /* next */
7410 0, /* static_pass_number */
7411 TV_CSE, /* tv_id */
7412 0, /* properties_required */
7413 0, /* properties_provided */
7414 0, /* properties_destroyed */
7415 0, /* todo_flags_start */
7416 TODO_df_finish | TODO_verify_rtl_sharing |
7417 TODO_dump_func |
7418 TODO_ggc_collect |
7419 TODO_verify_flow, /* todo_flags_finish */
7424 static bool
7425 gate_handle_cse2 (void)
7427 return optimize > 0 && flag_rerun_cse_after_loop;
7430 /* Run second CSE pass after loop optimizations. */
7431 static unsigned int
7432 rest_of_handle_cse2 (void)
7434 int tem;
7436 if (dump_file)
7437 dump_flow_info (dump_file, dump_flags);
7439 tem = cse_main (get_insns (), max_reg_num ());
7441 /* Run a pass to eliminate duplicated assignments to condition code
7442 registers. We have to run this after bypass_jumps, because it
7443 makes it harder for that pass to determine whether a jump can be
7444 bypassed safely. */
7445 cse_condition_code_reg ();
7447 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7449 if (tem == 2)
7451 timevar_push (TV_JUMP);
7452 rebuild_jump_labels (get_insns ());
7453 cleanup_cfg (0);
7454 timevar_pop (TV_JUMP);
7456 else if (tem == 1)
7457 cleanup_cfg (0);
7459 cse_not_expected = 1;
7460 return 0;
7464 struct rtl_opt_pass pass_cse2 =
7467 RTL_PASS,
7468 "cse2", /* name */
7469 gate_handle_cse2, /* gate */
7470 rest_of_handle_cse2, /* execute */
7471 NULL, /* sub */
7472 NULL, /* next */
7473 0, /* static_pass_number */
7474 TV_CSE2, /* tv_id */
7475 0, /* properties_required */
7476 0, /* properties_provided */
7477 0, /* properties_destroyed */
7478 0, /* todo_flags_start */
7479 TODO_df_finish | TODO_verify_rtl_sharing |
7480 TODO_dump_func |
7481 TODO_ggc_collect |
7482 TODO_verify_flow /* todo_flags_finish */
7486 static bool
7487 gate_handle_cse_after_global_opts (void)
7489 return optimize > 0 && flag_rerun_cse_after_global_opts;
7492 /* Run second CSE pass after loop optimizations. */
7493 static unsigned int
7494 rest_of_handle_cse_after_global_opts (void)
7496 int save_cfj;
7497 int tem;
7499 /* We only want to do local CSE, so don't follow jumps. */
7500 save_cfj = flag_cse_follow_jumps;
7501 flag_cse_follow_jumps = 0;
7503 rebuild_jump_labels (get_insns ());
7504 tem = cse_main (get_insns (), max_reg_num ());
7505 purge_all_dead_edges ();
7506 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7508 cse_not_expected = !flag_rerun_cse_after_loop;
7510 /* If cse altered any jumps, rerun jump opts to clean things up. */
7511 if (tem == 2)
7513 timevar_push (TV_JUMP);
7514 rebuild_jump_labels (get_insns ());
7515 cleanup_cfg (0);
7516 timevar_pop (TV_JUMP);
7518 else if (tem == 1)
7519 cleanup_cfg (0);
7521 flag_cse_follow_jumps = save_cfj;
7522 return 0;
7525 struct rtl_opt_pass pass_cse_after_global_opts =
7528 RTL_PASS,
7529 "cse_local", /* name */
7530 gate_handle_cse_after_global_opts, /* gate */
7531 rest_of_handle_cse_after_global_opts, /* execute */
7532 NULL, /* sub */
7533 NULL, /* next */
7534 0, /* static_pass_number */
7535 TV_CSE, /* tv_id */
7536 0, /* properties_required */
7537 0, /* properties_provided */
7538 0, /* properties_destroyed */
7539 0, /* todo_flags_start */
7540 TODO_df_finish | TODO_verify_rtl_sharing |
7541 TODO_dump_func |
7542 TODO_ggc_collect |
7543 TODO_verify_flow /* todo_flags_finish */