Fix a bug that broke -freorder-functions
[official-gcc.git] / gcc / cse.c
blob54edf7ee94202345ef679ce60ec386eeca9638af
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 2011 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_MODES_P (GET_MODE (x),
765 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 free (cse_reg_info_table);
800 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
801 cse_reg_info_table_size = new_size;
802 cse_reg_info_table_first_uninitialized = 0;
805 /* Do we have all of the first NREGS entries initialized? */
806 if (cse_reg_info_table_first_uninitialized < nregs)
808 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
809 unsigned int i;
811 /* Put the old timestamp on newly allocated entries so that they
812 will all be considered out of date. We do not touch those
813 entries beyond the first NREGS entries to be nice to the
814 virtual memory. */
815 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
816 cse_reg_info_table[i].timestamp = old_timestamp;
818 cse_reg_info_table_first_uninitialized = nregs;
822 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
824 static void
825 get_cse_reg_info_1 (unsigned int regno)
827 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
828 entry will be considered to have been initialized. */
829 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
831 /* Initialize the rest of the entry. */
832 cse_reg_info_table[regno].reg_tick = 1;
833 cse_reg_info_table[regno].reg_in_table = -1;
834 cse_reg_info_table[regno].subreg_ticked = -1;
835 cse_reg_info_table[regno].reg_qty = -regno - 1;
838 /* Find a cse_reg_info entry for REGNO. */
840 static inline struct cse_reg_info *
841 get_cse_reg_info (unsigned int regno)
843 struct cse_reg_info *p = &cse_reg_info_table[regno];
845 /* If this entry has not been initialized, go ahead and initialize
846 it. */
847 if (p->timestamp != cse_reg_info_timestamp)
848 get_cse_reg_info_1 (regno);
850 return p;
853 /* Clear the hash table and initialize each register with its own quantity,
854 for a new basic block. */
856 static void
857 new_basic_block (void)
859 int i;
861 next_qty = 0;
863 /* Invalidate cse_reg_info_table. */
864 cse_reg_info_timestamp++;
866 /* Clear out hash table state for this pass. */
867 CLEAR_HARD_REG_SET (hard_regs_in_table);
869 /* The per-quantity values used to be initialized here, but it is
870 much faster to initialize each as it is made in `make_new_qty'. */
872 for (i = 0; i < HASH_SIZE; i++)
874 struct table_elt *first;
876 first = table[i];
877 if (first != NULL)
879 struct table_elt *last = first;
881 table[i] = NULL;
883 while (last->next_same_hash != NULL)
884 last = last->next_same_hash;
886 /* Now relink this hash entire chain into
887 the free element list. */
889 last->next_same_hash = free_element_chain;
890 free_element_chain = first;
894 #ifdef HAVE_cc0
895 prev_insn_cc0 = 0;
896 #endif
899 /* Say that register REG contains a quantity in mode MODE not in any
900 register before and initialize that quantity. */
902 static void
903 make_new_qty (unsigned int reg, enum machine_mode mode)
905 int q;
906 struct qty_table_elem *ent;
907 struct reg_eqv_elem *eqv;
909 gcc_assert (next_qty < max_qty);
911 q = REG_QTY (reg) = next_qty++;
912 ent = &qty_table[q];
913 ent->first_reg = reg;
914 ent->last_reg = reg;
915 ent->mode = mode;
916 ent->const_rtx = ent->const_insn = NULL_RTX;
917 ent->comparison_code = UNKNOWN;
919 eqv = &reg_eqv_table[reg];
920 eqv->next = eqv->prev = -1;
923 /* Make reg NEW equivalent to reg OLD.
924 OLD is not changing; NEW is. */
926 static void
927 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
929 unsigned int lastr, firstr;
930 int q = REG_QTY (old_reg);
931 struct qty_table_elem *ent;
933 ent = &qty_table[q];
935 /* Nothing should become eqv until it has a "non-invalid" qty number. */
936 gcc_assert (REGNO_QTY_VALID_P (old_reg));
938 REG_QTY (new_reg) = q;
939 firstr = ent->first_reg;
940 lastr = ent->last_reg;
942 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
943 hard regs. Among pseudos, if NEW will live longer than any other reg
944 of the same qty, and that is beyond the current basic block,
945 make it the new canonical replacement for this qty. */
946 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
947 /* Certain fixed registers might be of the class NO_REGS. This means
948 that not only can they not be allocated by the compiler, but
949 they cannot be used in substitutions or canonicalizations
950 either. */
951 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
952 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
953 || (new_reg >= FIRST_PSEUDO_REGISTER
954 && (firstr < FIRST_PSEUDO_REGISTER
955 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
956 && !bitmap_bit_p (cse_ebb_live_out, firstr))
957 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
958 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
960 reg_eqv_table[firstr].prev = new_reg;
961 reg_eqv_table[new_reg].next = firstr;
962 reg_eqv_table[new_reg].prev = -1;
963 ent->first_reg = new_reg;
965 else
967 /* If NEW is a hard reg (known to be non-fixed), insert at end.
968 Otherwise, insert before any non-fixed hard regs that are at the
969 end. Registers of class NO_REGS cannot be used as an
970 equivalent for anything. */
971 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
972 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
973 && new_reg >= FIRST_PSEUDO_REGISTER)
974 lastr = reg_eqv_table[lastr].prev;
975 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
976 if (reg_eqv_table[lastr].next >= 0)
977 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
978 else
979 qty_table[q].last_reg = new_reg;
980 reg_eqv_table[lastr].next = new_reg;
981 reg_eqv_table[new_reg].prev = lastr;
985 /* Remove REG from its equivalence class. */
987 static void
988 delete_reg_equiv (unsigned int reg)
990 struct qty_table_elem *ent;
991 int q = REG_QTY (reg);
992 int p, n;
994 /* If invalid, do nothing. */
995 if (! REGNO_QTY_VALID_P (reg))
996 return;
998 ent = &qty_table[q];
1000 p = reg_eqv_table[reg].prev;
1001 n = reg_eqv_table[reg].next;
1003 if (n != -1)
1004 reg_eqv_table[n].prev = p;
1005 else
1006 ent->last_reg = p;
1007 if (p != -1)
1008 reg_eqv_table[p].next = n;
1009 else
1010 ent->first_reg = n;
1012 REG_QTY (reg) = -reg - 1;
1015 /* Remove any invalid expressions from the hash table
1016 that refer to any of the registers contained in expression X.
1018 Make sure that newly inserted references to those registers
1019 as subexpressions will be considered valid.
1021 mention_regs is not called when a register itself
1022 is being stored in the table.
1024 Return 1 if we have done something that may have changed the hash code
1025 of X. */
1027 static int
1028 mention_regs (rtx x)
1030 enum rtx_code code;
1031 int i, j;
1032 const char *fmt;
1033 int changed = 0;
1035 if (x == 0)
1036 return 0;
1038 code = GET_CODE (x);
1039 if (code == REG)
1041 unsigned int regno = REGNO (x);
1042 unsigned int endregno = END_REGNO (x);
1043 unsigned int i;
1045 for (i = regno; i < endregno; i++)
1047 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1048 remove_invalid_refs (i);
1050 REG_IN_TABLE (i) = REG_TICK (i);
1051 SUBREG_TICKED (i) = -1;
1054 return 0;
1057 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1058 pseudo if they don't use overlapping words. We handle only pseudos
1059 here for simplicity. */
1060 if (code == SUBREG && REG_P (SUBREG_REG (x))
1061 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1063 unsigned int i = REGNO (SUBREG_REG (x));
1065 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1067 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1068 the last store to this register really stored into this
1069 subreg, then remove the memory of this subreg.
1070 Otherwise, remove any memory of the entire register and
1071 all its subregs from the table. */
1072 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1073 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1074 remove_invalid_refs (i);
1075 else
1076 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1079 REG_IN_TABLE (i) = REG_TICK (i);
1080 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1081 return 0;
1084 /* If X is a comparison or a COMPARE and either operand is a register
1085 that does not have a quantity, give it one. This is so that a later
1086 call to record_jump_equiv won't cause X to be assigned a different
1087 hash code and not found in the table after that call.
1089 It is not necessary to do this here, since rehash_using_reg can
1090 fix up the table later, but doing this here eliminates the need to
1091 call that expensive function in the most common case where the only
1092 use of the register is in the comparison. */
1094 if (code == COMPARE || COMPARISON_P (x))
1096 if (REG_P (XEXP (x, 0))
1097 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1098 if (insert_regs (XEXP (x, 0), NULL, 0))
1100 rehash_using_reg (XEXP (x, 0));
1101 changed = 1;
1104 if (REG_P (XEXP (x, 1))
1105 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1106 if (insert_regs (XEXP (x, 1), NULL, 0))
1108 rehash_using_reg (XEXP (x, 1));
1109 changed = 1;
1113 fmt = GET_RTX_FORMAT (code);
1114 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1115 if (fmt[i] == 'e')
1116 changed |= mention_regs (XEXP (x, i));
1117 else if (fmt[i] == 'E')
1118 for (j = 0; j < XVECLEN (x, i); j++)
1119 changed |= mention_regs (XVECEXP (x, i, j));
1121 return changed;
1124 /* Update the register quantities for inserting X into the hash table
1125 with a value equivalent to CLASSP.
1126 (If the class does not contain a REG, it is irrelevant.)
1127 If MODIFIED is nonzero, X is a destination; it is being modified.
1128 Note that delete_reg_equiv should be called on a register
1129 before insert_regs is done on that register with MODIFIED != 0.
1131 Nonzero value means that elements of reg_qty have changed
1132 so X's hash code may be different. */
1134 static int
1135 insert_regs (rtx x, struct table_elt *classp, int modified)
1137 if (REG_P (x))
1139 unsigned int regno = REGNO (x);
1140 int qty_valid;
1142 /* If REGNO is in the equivalence table already but is of the
1143 wrong mode for that equivalence, don't do anything here. */
1145 qty_valid = REGNO_QTY_VALID_P (regno);
1146 if (qty_valid)
1148 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1150 if (ent->mode != GET_MODE (x))
1151 return 0;
1154 if (modified || ! qty_valid)
1156 if (classp)
1157 for (classp = classp->first_same_value;
1158 classp != 0;
1159 classp = classp->next_same_value)
1160 if (REG_P (classp->exp)
1161 && GET_MODE (classp->exp) == GET_MODE (x))
1163 unsigned c_regno = REGNO (classp->exp);
1165 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1167 /* Suppose that 5 is hard reg and 100 and 101 are
1168 pseudos. Consider
1170 (set (reg:si 100) (reg:si 5))
1171 (set (reg:si 5) (reg:si 100))
1172 (set (reg:di 101) (reg:di 5))
1174 We would now set REG_QTY (101) = REG_QTY (5), but the
1175 entry for 5 is in SImode. When we use this later in
1176 copy propagation, we get the register in wrong mode. */
1177 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1178 continue;
1180 make_regs_eqv (regno, c_regno);
1181 return 1;
1184 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1185 than REG_IN_TABLE to find out if there was only a single preceding
1186 invalidation - for the SUBREG - or another one, which would be
1187 for the full register. However, if we find here that REG_TICK
1188 indicates that the register is invalid, it means that it has
1189 been invalidated in a separate operation. The SUBREG might be used
1190 now (then this is a recursive call), or we might use the full REG
1191 now and a SUBREG of it later. So bump up REG_TICK so that
1192 mention_regs will do the right thing. */
1193 if (! modified
1194 && REG_IN_TABLE (regno) >= 0
1195 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1196 REG_TICK (regno)++;
1197 make_new_qty (regno, GET_MODE (x));
1198 return 1;
1201 return 0;
1204 /* If X is a SUBREG, we will likely be inserting the inner register in the
1205 table. If that register doesn't have an assigned quantity number at
1206 this point but does later, the insertion that we will be doing now will
1207 not be accessible because its hash code will have changed. So assign
1208 a quantity number now. */
1210 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1211 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1213 insert_regs (SUBREG_REG (x), NULL, 0);
1214 mention_regs (x);
1215 return 1;
1217 else
1218 return mention_regs (x);
1222 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1223 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1224 CST is equal to an anchor. */
1226 static bool
1227 compute_const_anchors (rtx cst,
1228 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1229 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1231 HOST_WIDE_INT n = INTVAL (cst);
1233 *lower_base = n & ~(targetm.const_anchor - 1);
1234 if (*lower_base == n)
1235 return false;
1237 *upper_base =
1238 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1239 *upper_offs = n - *upper_base;
1240 *lower_offs = n - *lower_base;
1241 return true;
1244 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1246 static void
1247 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1248 enum machine_mode mode)
1250 struct table_elt *elt;
1251 unsigned hash;
1252 rtx anchor_exp;
1253 rtx exp;
1255 anchor_exp = GEN_INT (anchor);
1256 hash = HASH (anchor_exp, mode);
1257 elt = lookup (anchor_exp, hash, mode);
1258 if (!elt)
1259 elt = insert (anchor_exp, NULL, hash, mode);
1261 exp = plus_constant (reg, offs);
1262 /* REG has just been inserted and the hash codes recomputed. */
1263 mention_regs (exp);
1264 hash = HASH (exp, mode);
1266 /* Use the cost of the register rather than the whole expression. When
1267 looking up constant anchors we will further offset the corresponding
1268 expression therefore it does not make sense to prefer REGs over
1269 reg-immediate additions. Prefer instead the oldest expression. Also
1270 don't prefer pseudos over hard regs so that we derive constants in
1271 argument registers from other argument registers rather than from the
1272 original pseudo that was used to synthesize the constant. */
1273 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1276 /* The constant CST is equivalent to the register REG. Create
1277 equivalences between the two anchors of CST and the corresponding
1278 register-offset expressions using REG. */
1280 static void
1281 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1283 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1285 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1286 &upper_base, &upper_offs))
1287 return;
1289 /* Ignore anchors of value 0. Constants accessible from zero are
1290 simple. */
1291 if (lower_base != 0)
1292 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1294 if (upper_base != 0)
1295 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1298 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1299 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1300 valid expression. Return the cheapest and oldest of such expressions. In
1301 *OLD, return how old the resulting expression is compared to the other
1302 equivalent expressions. */
1304 static rtx
1305 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1306 unsigned *old)
1308 struct table_elt *elt;
1309 unsigned idx;
1310 struct table_elt *match_elt;
1311 rtx match;
1313 /* Find the cheapest and *oldest* expression to maximize the chance of
1314 reusing the same pseudo. */
1316 match_elt = NULL;
1317 match = NULL_RTX;
1318 for (elt = anchor_elt->first_same_value, idx = 0;
1319 elt;
1320 elt = elt->next_same_value, idx++)
1322 if (match_elt && CHEAPER (match_elt, elt))
1323 return match;
1325 if (REG_P (elt->exp)
1326 || (GET_CODE (elt->exp) == PLUS
1327 && REG_P (XEXP (elt->exp, 0))
1328 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1330 rtx x;
1332 /* Ignore expressions that are no longer valid. */
1333 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1334 continue;
1336 x = plus_constant (elt->exp, offs);
1337 if (REG_P (x)
1338 || (GET_CODE (x) == PLUS
1339 && IN_RANGE (INTVAL (XEXP (x, 1)),
1340 -targetm.const_anchor,
1341 targetm.const_anchor - 1)))
1343 match = x;
1344 match_elt = elt;
1345 *old = idx;
1350 return match;
1353 /* Try to express the constant SRC_CONST using a register+offset expression
1354 derived from a constant anchor. Return it if successful or NULL_RTX,
1355 otherwise. */
1357 static rtx
1358 try_const_anchors (rtx src_const, enum machine_mode mode)
1360 struct table_elt *lower_elt, *upper_elt;
1361 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1362 rtx lower_anchor_rtx, upper_anchor_rtx;
1363 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1364 unsigned lower_old, upper_old;
1366 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1367 &upper_base, &upper_offs))
1368 return NULL_RTX;
1370 lower_anchor_rtx = GEN_INT (lower_base);
1371 upper_anchor_rtx = GEN_INT (upper_base);
1372 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1373 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1375 if (lower_elt)
1376 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1377 if (upper_elt)
1378 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1380 if (!lower_exp)
1381 return upper_exp;
1382 if (!upper_exp)
1383 return lower_exp;
1385 /* Return the older expression. */
1386 return (upper_old > lower_old ? upper_exp : lower_exp);
1389 /* Look in or update the hash table. */
1391 /* Remove table element ELT from use in the table.
1392 HASH is its hash code, made using the HASH macro.
1393 It's an argument because often that is known in advance
1394 and we save much time not recomputing it. */
1396 static void
1397 remove_from_table (struct table_elt *elt, unsigned int hash)
1399 if (elt == 0)
1400 return;
1402 /* Mark this element as removed. See cse_insn. */
1403 elt->first_same_value = 0;
1405 /* Remove the table element from its equivalence class. */
1408 struct table_elt *prev = elt->prev_same_value;
1409 struct table_elt *next = elt->next_same_value;
1411 if (next)
1412 next->prev_same_value = prev;
1414 if (prev)
1415 prev->next_same_value = next;
1416 else
1418 struct table_elt *newfirst = next;
1419 while (next)
1421 next->first_same_value = newfirst;
1422 next = next->next_same_value;
1427 /* Remove the table element from its hash bucket. */
1430 struct table_elt *prev = elt->prev_same_hash;
1431 struct table_elt *next = elt->next_same_hash;
1433 if (next)
1434 next->prev_same_hash = prev;
1436 if (prev)
1437 prev->next_same_hash = next;
1438 else if (table[hash] == elt)
1439 table[hash] = next;
1440 else
1442 /* This entry is not in the proper hash bucket. This can happen
1443 when two classes were merged by `merge_equiv_classes'. Search
1444 for the hash bucket that it heads. This happens only very
1445 rarely, so the cost is acceptable. */
1446 for (hash = 0; hash < HASH_SIZE; hash++)
1447 if (table[hash] == elt)
1448 table[hash] = next;
1452 /* Remove the table element from its related-value circular chain. */
1454 if (elt->related_value != 0 && elt->related_value != elt)
1456 struct table_elt *p = elt->related_value;
1458 while (p->related_value != elt)
1459 p = p->related_value;
1460 p->related_value = elt->related_value;
1461 if (p->related_value == p)
1462 p->related_value = 0;
1465 /* Now add it to the free element chain. */
1466 elt->next_same_hash = free_element_chain;
1467 free_element_chain = elt;
1470 /* Same as above, but X is a pseudo-register. */
1472 static void
1473 remove_pseudo_from_table (rtx x, unsigned int hash)
1475 struct table_elt *elt;
1477 /* Because a pseudo-register can be referenced in more than one
1478 mode, we might have to remove more than one table entry. */
1479 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1480 remove_from_table (elt, hash);
1483 /* Look up X in the hash table and return its table element,
1484 or 0 if X is not in the table.
1486 MODE is the machine-mode of X, or if X is an integer constant
1487 with VOIDmode then MODE is the mode with which X will be used.
1489 Here we are satisfied to find an expression whose tree structure
1490 looks like X. */
1492 static struct table_elt *
1493 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1495 struct table_elt *p;
1497 for (p = table[hash]; p; p = p->next_same_hash)
1498 if (mode == p->mode && ((x == p->exp && REG_P (x))
1499 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1500 return p;
1502 return 0;
1505 /* Like `lookup' but don't care whether the table element uses invalid regs.
1506 Also ignore discrepancies in the machine mode of a register. */
1508 static struct table_elt *
1509 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1511 struct table_elt *p;
1513 if (REG_P (x))
1515 unsigned int regno = REGNO (x);
1517 /* Don't check the machine mode when comparing registers;
1518 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1519 for (p = table[hash]; p; p = p->next_same_hash)
1520 if (REG_P (p->exp)
1521 && REGNO (p->exp) == regno)
1522 return p;
1524 else
1526 for (p = table[hash]; p; p = p->next_same_hash)
1527 if (mode == p->mode
1528 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1529 return p;
1532 return 0;
1535 /* Look for an expression equivalent to X and with code CODE.
1536 If one is found, return that expression. */
1538 static rtx
1539 lookup_as_function (rtx x, enum rtx_code code)
1541 struct table_elt *p
1542 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1544 if (p == 0)
1545 return 0;
1547 for (p = p->first_same_value; p; p = p->next_same_value)
1548 if (GET_CODE (p->exp) == code
1549 /* Make sure this is a valid entry in the table. */
1550 && exp_equiv_p (p->exp, p->exp, 1, false))
1551 return p->exp;
1553 return 0;
1556 /* Insert X in the hash table, assuming HASH is its hash code and
1557 CLASSP is an element of the class it should go in (or 0 if a new
1558 class should be made). COST is the code of X and reg_cost is the
1559 cost of registers in X. It is inserted at the proper position to
1560 keep the class in the order cheapest first.
1562 MODE is the machine-mode of X, or if X is an integer constant
1563 with VOIDmode then MODE is the mode with which X will be used.
1565 For elements of equal cheapness, the most recent one
1566 goes in front, except that the first element in the list
1567 remains first unless a cheaper element is added. The order of
1568 pseudo-registers does not matter, as canon_reg will be called to
1569 find the cheapest when a register is retrieved from the table.
1571 The in_memory field in the hash table element is set to 0.
1572 The caller must set it nonzero if appropriate.
1574 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1575 and if insert_regs returns a nonzero value
1576 you must then recompute its hash code before calling here.
1578 If necessary, update table showing constant values of quantities. */
1580 static struct table_elt *
1581 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1582 enum machine_mode mode, int cost, int reg_cost)
1584 struct table_elt *elt;
1586 /* If X is a register and we haven't made a quantity for it,
1587 something is wrong. */
1588 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1590 /* If X is a hard register, show it is being put in the table. */
1591 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1592 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1594 /* Put an element for X into the right hash bucket. */
1596 elt = free_element_chain;
1597 if (elt)
1598 free_element_chain = elt->next_same_hash;
1599 else
1600 elt = XNEW (struct table_elt);
1602 elt->exp = x;
1603 elt->canon_exp = NULL_RTX;
1604 elt->cost = cost;
1605 elt->regcost = reg_cost;
1606 elt->next_same_value = 0;
1607 elt->prev_same_value = 0;
1608 elt->next_same_hash = table[hash];
1609 elt->prev_same_hash = 0;
1610 elt->related_value = 0;
1611 elt->in_memory = 0;
1612 elt->mode = mode;
1613 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1615 if (table[hash])
1616 table[hash]->prev_same_hash = elt;
1617 table[hash] = elt;
1619 /* Put it into the proper value-class. */
1620 if (classp)
1622 classp = classp->first_same_value;
1623 if (CHEAPER (elt, classp))
1624 /* Insert at the head of the class. */
1626 struct table_elt *p;
1627 elt->next_same_value = classp;
1628 classp->prev_same_value = elt;
1629 elt->first_same_value = elt;
1631 for (p = classp; p; p = p->next_same_value)
1632 p->first_same_value = elt;
1634 else
1636 /* Insert not at head of the class. */
1637 /* Put it after the last element cheaper than X. */
1638 struct table_elt *p, *next;
1640 for (p = classp;
1641 (next = p->next_same_value) && CHEAPER (next, elt);
1642 p = next)
1645 /* Put it after P and before NEXT. */
1646 elt->next_same_value = next;
1647 if (next)
1648 next->prev_same_value = elt;
1650 elt->prev_same_value = p;
1651 p->next_same_value = elt;
1652 elt->first_same_value = classp;
1655 else
1656 elt->first_same_value = elt;
1658 /* If this is a constant being set equivalent to a register or a register
1659 being set equivalent to a constant, note the constant equivalence.
1661 If this is a constant, it cannot be equivalent to a different constant,
1662 and a constant is the only thing that can be cheaper than a register. So
1663 we know the register is the head of the class (before the constant was
1664 inserted).
1666 If this is a register that is not already known equivalent to a
1667 constant, we must check the entire class.
1669 If this is a register that is already known equivalent to an insn,
1670 update the qtys `const_insn' to show that `this_insn' is the latest
1671 insn making that quantity equivalent to the constant. */
1673 if (elt->is_const && classp && REG_P (classp->exp)
1674 && !REG_P (x))
1676 int exp_q = REG_QTY (REGNO (classp->exp));
1677 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1679 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1680 exp_ent->const_insn = this_insn;
1683 else if (REG_P (x)
1684 && classp
1685 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1686 && ! elt->is_const)
1688 struct table_elt *p;
1690 for (p = classp; p != 0; p = p->next_same_value)
1692 if (p->is_const && !REG_P (p->exp))
1694 int x_q = REG_QTY (REGNO (x));
1695 struct qty_table_elem *x_ent = &qty_table[x_q];
1697 x_ent->const_rtx
1698 = gen_lowpart (GET_MODE (x), p->exp);
1699 x_ent->const_insn = this_insn;
1700 break;
1705 else if (REG_P (x)
1706 && qty_table[REG_QTY (REGNO (x))].const_rtx
1707 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1708 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1710 /* If this is a constant with symbolic value,
1711 and it has a term with an explicit integer value,
1712 link it up with related expressions. */
1713 if (GET_CODE (x) == CONST)
1715 rtx subexp = get_related_value (x);
1716 unsigned subhash;
1717 struct table_elt *subelt, *subelt_prev;
1719 if (subexp != 0)
1721 /* Get the integer-free subexpression in the hash table. */
1722 subhash = SAFE_HASH (subexp, mode);
1723 subelt = lookup (subexp, subhash, mode);
1724 if (subelt == 0)
1725 subelt = insert (subexp, NULL, subhash, mode);
1726 /* Initialize SUBELT's circular chain if it has none. */
1727 if (subelt->related_value == 0)
1728 subelt->related_value = subelt;
1729 /* Find the element in the circular chain that precedes SUBELT. */
1730 subelt_prev = subelt;
1731 while (subelt_prev->related_value != subelt)
1732 subelt_prev = subelt_prev->related_value;
1733 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1734 This way the element that follows SUBELT is the oldest one. */
1735 elt->related_value = subelt_prev->related_value;
1736 subelt_prev->related_value = elt;
1740 return elt;
1743 /* Wrap insert_with_costs by passing the default costs. */
1745 static struct table_elt *
1746 insert (rtx x, struct table_elt *classp, unsigned int hash,
1747 enum machine_mode mode)
1749 return
1750 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1754 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1755 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1756 the two classes equivalent.
1758 CLASS1 will be the surviving class; CLASS2 should not be used after this
1759 call.
1761 Any invalid entries in CLASS2 will not be copied. */
1763 static void
1764 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1766 struct table_elt *elt, *next, *new_elt;
1768 /* Ensure we start with the head of the classes. */
1769 class1 = class1->first_same_value;
1770 class2 = class2->first_same_value;
1772 /* If they were already equal, forget it. */
1773 if (class1 == class2)
1774 return;
1776 for (elt = class2; elt; elt = next)
1778 unsigned int hash;
1779 rtx exp = elt->exp;
1780 enum machine_mode mode = elt->mode;
1782 next = elt->next_same_value;
1784 /* Remove old entry, make a new one in CLASS1's class.
1785 Don't do this for invalid entries as we cannot find their
1786 hash code (it also isn't necessary). */
1787 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1789 bool need_rehash = false;
1791 hash_arg_in_memory = 0;
1792 hash = HASH (exp, mode);
1794 if (REG_P (exp))
1796 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1797 delete_reg_equiv (REGNO (exp));
1800 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1801 remove_pseudo_from_table (exp, hash);
1802 else
1803 remove_from_table (elt, hash);
1805 if (insert_regs (exp, class1, 0) || need_rehash)
1807 rehash_using_reg (exp);
1808 hash = HASH (exp, mode);
1810 new_elt = insert (exp, class1, hash, mode);
1811 new_elt->in_memory = hash_arg_in_memory;
1816 /* Flush the entire hash table. */
1818 static void
1819 flush_hash_table (void)
1821 int i;
1822 struct table_elt *p;
1824 for (i = 0; i < HASH_SIZE; i++)
1825 for (p = table[i]; p; p = table[i])
1827 /* Note that invalidate can remove elements
1828 after P in the current hash chain. */
1829 if (REG_P (p->exp))
1830 invalidate (p->exp, VOIDmode);
1831 else
1832 remove_from_table (p, i);
1836 /* Function called for each rtx to check whether true dependence exist. */
1837 struct check_dependence_data
1839 enum machine_mode mode;
1840 rtx exp;
1841 rtx addr;
1844 static int
1845 check_dependence (rtx *x, void *data)
1847 struct check_dependence_data *d = (struct check_dependence_data *) data;
1848 if (*x && MEM_P (*x))
1849 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX,
1850 cse_rtx_varies_p);
1851 else
1852 return 0;
1855 /* Remove from the hash table, or mark as invalid, all expressions whose
1856 values could be altered by storing in X. X is a register, a subreg, or
1857 a memory reference with nonvarying address (because, when a memory
1858 reference with a varying address is stored in, all memory references are
1859 removed by invalidate_memory so specific invalidation is superfluous).
1860 FULL_MODE, if not VOIDmode, indicates that this much should be
1861 invalidated instead of just the amount indicated by the mode of X. This
1862 is only used for bitfield stores into memory.
1864 A nonvarying address may be just a register or just a symbol reference,
1865 or it may be either of those plus a numeric offset. */
1867 static void
1868 invalidate (rtx x, enum machine_mode full_mode)
1870 int i;
1871 struct table_elt *p;
1872 rtx addr;
1874 switch (GET_CODE (x))
1876 case REG:
1878 /* If X is a register, dependencies on its contents are recorded
1879 through the qty number mechanism. Just change the qty number of
1880 the register, mark it as invalid for expressions that refer to it,
1881 and remove it itself. */
1882 unsigned int regno = REGNO (x);
1883 unsigned int hash = HASH (x, GET_MODE (x));
1885 /* Remove REGNO from any quantity list it might be on and indicate
1886 that its value might have changed. If it is a pseudo, remove its
1887 entry from the hash table.
1889 For a hard register, we do the first two actions above for any
1890 additional hard registers corresponding to X. Then, if any of these
1891 registers are in the table, we must remove any REG entries that
1892 overlap these registers. */
1894 delete_reg_equiv (regno);
1895 REG_TICK (regno)++;
1896 SUBREG_TICKED (regno) = -1;
1898 if (regno >= FIRST_PSEUDO_REGISTER)
1899 remove_pseudo_from_table (x, hash);
1900 else
1902 HOST_WIDE_INT in_table
1903 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1904 unsigned int endregno = END_HARD_REGNO (x);
1905 unsigned int tregno, tendregno, rn;
1906 struct table_elt *p, *next;
1908 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1910 for (rn = regno + 1; rn < endregno; rn++)
1912 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1913 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1914 delete_reg_equiv (rn);
1915 REG_TICK (rn)++;
1916 SUBREG_TICKED (rn) = -1;
1919 if (in_table)
1920 for (hash = 0; hash < HASH_SIZE; hash++)
1921 for (p = table[hash]; p; p = next)
1923 next = p->next_same_hash;
1925 if (!REG_P (p->exp)
1926 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1927 continue;
1929 tregno = REGNO (p->exp);
1930 tendregno = END_HARD_REGNO (p->exp);
1931 if (tendregno > regno && tregno < endregno)
1932 remove_from_table (p, hash);
1936 return;
1938 case SUBREG:
1939 invalidate (SUBREG_REG (x), VOIDmode);
1940 return;
1942 case PARALLEL:
1943 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1944 invalidate (XVECEXP (x, 0, i), VOIDmode);
1945 return;
1947 case EXPR_LIST:
1948 /* This is part of a disjoint return value; extract the location in
1949 question ignoring the offset. */
1950 invalidate (XEXP (x, 0), VOIDmode);
1951 return;
1953 case MEM:
1954 addr = canon_rtx (get_addr (XEXP (x, 0)));
1955 /* Calculate the canonical version of X here so that
1956 true_dependence doesn't generate new RTL for X on each call. */
1957 x = canon_rtx (x);
1959 /* Remove all hash table elements that refer to overlapping pieces of
1960 memory. */
1961 if (full_mode == VOIDmode)
1962 full_mode = GET_MODE (x);
1964 for (i = 0; i < HASH_SIZE; i++)
1966 struct table_elt *next;
1968 for (p = table[i]; p; p = next)
1970 next = p->next_same_hash;
1971 if (p->in_memory)
1973 struct check_dependence_data d;
1975 /* Just canonicalize the expression once;
1976 otherwise each time we call invalidate
1977 true_dependence will canonicalize the
1978 expression again. */
1979 if (!p->canon_exp)
1980 p->canon_exp = canon_rtx (p->exp);
1981 d.exp = x;
1982 d.addr = addr;
1983 d.mode = full_mode;
1984 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1985 remove_from_table (p, i);
1989 return;
1991 default:
1992 gcc_unreachable ();
1996 /* Remove all expressions that refer to register REGNO,
1997 since they are already invalid, and we are about to
1998 mark that register valid again and don't want the old
1999 expressions to reappear as valid. */
2001 static void
2002 remove_invalid_refs (unsigned int regno)
2004 unsigned int i;
2005 struct table_elt *p, *next;
2007 for (i = 0; i < HASH_SIZE; i++)
2008 for (p = table[i]; p; p = next)
2010 next = p->next_same_hash;
2011 if (!REG_P (p->exp)
2012 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2013 remove_from_table (p, i);
2017 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2018 and mode MODE. */
2019 static void
2020 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2021 enum machine_mode mode)
2023 unsigned int i;
2024 struct table_elt *p, *next;
2025 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2027 for (i = 0; i < HASH_SIZE; i++)
2028 for (p = table[i]; p; p = next)
2030 rtx exp = p->exp;
2031 next = p->next_same_hash;
2033 if (!REG_P (exp)
2034 && (GET_CODE (exp) != SUBREG
2035 || !REG_P (SUBREG_REG (exp))
2036 || REGNO (SUBREG_REG (exp)) != regno
2037 || (((SUBREG_BYTE (exp)
2038 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2039 && SUBREG_BYTE (exp) <= end))
2040 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2041 remove_from_table (p, i);
2045 /* Recompute the hash codes of any valid entries in the hash table that
2046 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2048 This is called when we make a jump equivalence. */
2050 static void
2051 rehash_using_reg (rtx x)
2053 unsigned int i;
2054 struct table_elt *p, *next;
2055 unsigned hash;
2057 if (GET_CODE (x) == SUBREG)
2058 x = SUBREG_REG (x);
2060 /* If X is not a register or if the register is known not to be in any
2061 valid entries in the table, we have no work to do. */
2063 if (!REG_P (x)
2064 || REG_IN_TABLE (REGNO (x)) < 0
2065 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2066 return;
2068 /* Scan all hash chains looking for valid entries that mention X.
2069 If we find one and it is in the wrong hash chain, move it. */
2071 for (i = 0; i < HASH_SIZE; i++)
2072 for (p = table[i]; p; p = next)
2074 next = p->next_same_hash;
2075 if (reg_mentioned_p (x, p->exp)
2076 && exp_equiv_p (p->exp, p->exp, 1, false)
2077 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2079 if (p->next_same_hash)
2080 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2082 if (p->prev_same_hash)
2083 p->prev_same_hash->next_same_hash = p->next_same_hash;
2084 else
2085 table[i] = p->next_same_hash;
2087 p->next_same_hash = table[hash];
2088 p->prev_same_hash = 0;
2089 if (table[hash])
2090 table[hash]->prev_same_hash = p;
2091 table[hash] = p;
2096 /* Remove from the hash table any expression that is a call-clobbered
2097 register. Also update their TICK values. */
2099 static void
2100 invalidate_for_call (void)
2102 unsigned int regno, endregno;
2103 unsigned int i;
2104 unsigned hash;
2105 struct table_elt *p, *next;
2106 int in_table = 0;
2108 /* Go through all the hard registers. For each that is clobbered in
2109 a CALL_INSN, remove the register from quantity chains and update
2110 reg_tick if defined. Also see if any of these registers is currently
2111 in the table. */
2113 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2114 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2116 delete_reg_equiv (regno);
2117 if (REG_TICK (regno) >= 0)
2119 REG_TICK (regno)++;
2120 SUBREG_TICKED (regno) = -1;
2123 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2126 /* In the case where we have no call-clobbered hard registers in the
2127 table, we are done. Otherwise, scan the table and remove any
2128 entry that overlaps a call-clobbered register. */
2130 if (in_table)
2131 for (hash = 0; hash < HASH_SIZE; hash++)
2132 for (p = table[hash]; p; p = next)
2134 next = p->next_same_hash;
2136 if (!REG_P (p->exp)
2137 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2138 continue;
2140 regno = REGNO (p->exp);
2141 endregno = END_HARD_REGNO (p->exp);
2143 for (i = regno; i < endregno; i++)
2144 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2146 remove_from_table (p, hash);
2147 break;
2152 /* Given an expression X of type CONST,
2153 and ELT which is its table entry (or 0 if it
2154 is not in the hash table),
2155 return an alternate expression for X as a register plus integer.
2156 If none can be found, return 0. */
2158 static rtx
2159 use_related_value (rtx x, struct table_elt *elt)
2161 struct table_elt *relt = 0;
2162 struct table_elt *p, *q;
2163 HOST_WIDE_INT offset;
2165 /* First, is there anything related known?
2166 If we have a table element, we can tell from that.
2167 Otherwise, must look it up. */
2169 if (elt != 0 && elt->related_value != 0)
2170 relt = elt;
2171 else if (elt == 0 && GET_CODE (x) == CONST)
2173 rtx subexp = get_related_value (x);
2174 if (subexp != 0)
2175 relt = lookup (subexp,
2176 SAFE_HASH (subexp, GET_MODE (subexp)),
2177 GET_MODE (subexp));
2180 if (relt == 0)
2181 return 0;
2183 /* Search all related table entries for one that has an
2184 equivalent register. */
2186 p = relt;
2187 while (1)
2189 /* This loop is strange in that it is executed in two different cases.
2190 The first is when X is already in the table. Then it is searching
2191 the RELATED_VALUE list of X's class (RELT). The second case is when
2192 X is not in the table. Then RELT points to a class for the related
2193 value.
2195 Ensure that, whatever case we are in, that we ignore classes that have
2196 the same value as X. */
2198 if (rtx_equal_p (x, p->exp))
2199 q = 0;
2200 else
2201 for (q = p->first_same_value; q; q = q->next_same_value)
2202 if (REG_P (q->exp))
2203 break;
2205 if (q)
2206 break;
2208 p = p->related_value;
2210 /* We went all the way around, so there is nothing to be found.
2211 Alternatively, perhaps RELT was in the table for some other reason
2212 and it has no related values recorded. */
2213 if (p == relt || p == 0)
2214 break;
2217 if (q == 0)
2218 return 0;
2220 offset = (get_integer_term (x) - get_integer_term (p->exp));
2221 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2222 return plus_constant (q->exp, offset);
2226 /* Hash a string. Just add its bytes up. */
2227 static inline unsigned
2228 hash_rtx_string (const char *ps)
2230 unsigned hash = 0;
2231 const unsigned char *p = (const unsigned char *) ps;
2233 if (p)
2234 while (*p)
2235 hash += *p++;
2237 return hash;
2240 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2241 When the callback returns true, we continue with the new rtx. */
2243 unsigned
2244 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2245 int *do_not_record_p, int *hash_arg_in_memory_p,
2246 bool have_reg_qty, hash_rtx_callback_function cb)
2248 int i, j;
2249 unsigned hash = 0;
2250 enum rtx_code code;
2251 const char *fmt;
2252 enum machine_mode newmode;
2253 rtx newx;
2255 /* Used to turn recursion into iteration. We can't rely on GCC's
2256 tail-recursion elimination since we need to keep accumulating values
2257 in HASH. */
2258 repeat:
2259 if (x == 0)
2260 return hash;
2262 /* Invoke the callback first. */
2263 if (cb != NULL
2264 && ((*cb) (x, mode, &newx, &newmode)))
2266 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2267 hash_arg_in_memory_p, have_reg_qty, cb);
2268 return hash;
2271 code = GET_CODE (x);
2272 switch (code)
2274 case REG:
2276 unsigned int regno = REGNO (x);
2278 if (do_not_record_p && !reload_completed)
2280 /* On some machines, we can't record any non-fixed hard register,
2281 because extending its life will cause reload problems. We
2282 consider ap, fp, sp, gp to be fixed for this purpose.
2284 We also consider CCmode registers to be fixed for this purpose;
2285 failure to do so leads to failure to simplify 0<100 type of
2286 conditionals.
2288 On all machines, we can't record any global registers.
2289 Nor should we record any register that is in a small
2290 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2291 bool record;
2293 if (regno >= FIRST_PSEUDO_REGISTER)
2294 record = true;
2295 else if (x == frame_pointer_rtx
2296 || x == hard_frame_pointer_rtx
2297 || x == arg_pointer_rtx
2298 || x == stack_pointer_rtx
2299 || x == pic_offset_table_rtx)
2300 record = true;
2301 else if (global_regs[regno])
2302 record = false;
2303 else if (fixed_regs[regno])
2304 record = true;
2305 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2306 record = true;
2307 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2308 record = false;
2309 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2310 record = false;
2311 else
2312 record = true;
2314 if (!record)
2316 *do_not_record_p = 1;
2317 return 0;
2321 hash += ((unsigned int) REG << 7);
2322 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2323 return hash;
2326 /* We handle SUBREG of a REG specially because the underlying
2327 reg changes its hash value with every value change; we don't
2328 want to have to forget unrelated subregs when one subreg changes. */
2329 case SUBREG:
2331 if (REG_P (SUBREG_REG (x)))
2333 hash += (((unsigned int) SUBREG << 7)
2334 + REGNO (SUBREG_REG (x))
2335 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2336 return hash;
2338 break;
2341 case CONST_INT:
2342 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2343 + (unsigned int) INTVAL (x));
2344 return hash;
2346 case CONST_DOUBLE:
2347 /* This is like the general case, except that it only counts
2348 the integers representing the constant. */
2349 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2350 if (GET_MODE (x) != VOIDmode)
2351 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2352 else
2353 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2354 + (unsigned int) CONST_DOUBLE_HIGH (x));
2355 return hash;
2357 case CONST_FIXED:
2358 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2359 hash += fixed_hash (CONST_FIXED_VALUE (x));
2360 return hash;
2362 case CONST_VECTOR:
2364 int units;
2365 rtx elt;
2367 units = CONST_VECTOR_NUNITS (x);
2369 for (i = 0; i < units; ++i)
2371 elt = CONST_VECTOR_ELT (x, i);
2372 hash += hash_rtx_cb (elt, GET_MODE (elt),
2373 do_not_record_p, hash_arg_in_memory_p,
2374 have_reg_qty, cb);
2377 return hash;
2380 /* Assume there is only one rtx object for any given label. */
2381 case LABEL_REF:
2382 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2383 differences and differences between each stage's debugging dumps. */
2384 hash += (((unsigned int) LABEL_REF << 7)
2385 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2386 return hash;
2388 case SYMBOL_REF:
2390 /* Don't hash on the symbol's address to avoid bootstrap differences.
2391 Different hash values may cause expressions to be recorded in
2392 different orders and thus different registers to be used in the
2393 final assembler. This also avoids differences in the dump files
2394 between various stages. */
2395 unsigned int h = 0;
2396 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2398 while (*p)
2399 h += (h << 7) + *p++; /* ??? revisit */
2401 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2402 return hash;
2405 case MEM:
2406 /* We don't record if marked volatile or if BLKmode since we don't
2407 know the size of the move. */
2408 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2410 *do_not_record_p = 1;
2411 return 0;
2413 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2414 *hash_arg_in_memory_p = 1;
2416 /* Now that we have already found this special case,
2417 might as well speed it up as much as possible. */
2418 hash += (unsigned) MEM;
2419 x = XEXP (x, 0);
2420 goto repeat;
2422 case USE:
2423 /* A USE that mentions non-volatile memory needs special
2424 handling since the MEM may be BLKmode which normally
2425 prevents an entry from being made. Pure calls are
2426 marked by a USE which mentions BLKmode memory.
2427 See calls.c:emit_call_1. */
2428 if (MEM_P (XEXP (x, 0))
2429 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2431 hash += (unsigned) USE;
2432 x = XEXP (x, 0);
2434 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2435 *hash_arg_in_memory_p = 1;
2437 /* Now that we have already found this special case,
2438 might as well speed it up as much as possible. */
2439 hash += (unsigned) MEM;
2440 x = XEXP (x, 0);
2441 goto repeat;
2443 break;
2445 case PRE_DEC:
2446 case PRE_INC:
2447 case POST_DEC:
2448 case POST_INC:
2449 case PRE_MODIFY:
2450 case POST_MODIFY:
2451 case PC:
2452 case CC0:
2453 case CALL:
2454 case UNSPEC_VOLATILE:
2455 if (do_not_record_p) {
2456 *do_not_record_p = 1;
2457 return 0;
2459 else
2460 return hash;
2461 break;
2463 case ASM_OPERANDS:
2464 if (do_not_record_p && MEM_VOLATILE_P (x))
2466 *do_not_record_p = 1;
2467 return 0;
2469 else
2471 /* We don't want to take the filename and line into account. */
2472 hash += (unsigned) code + (unsigned) GET_MODE (x)
2473 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2474 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2475 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2477 if (ASM_OPERANDS_INPUT_LENGTH (x))
2479 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2481 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2482 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2483 do_not_record_p, hash_arg_in_memory_p,
2484 have_reg_qty, cb)
2485 + hash_rtx_string
2486 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2489 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2490 x = ASM_OPERANDS_INPUT (x, 0);
2491 mode = GET_MODE (x);
2492 goto repeat;
2495 return hash;
2497 break;
2499 default:
2500 break;
2503 i = GET_RTX_LENGTH (code) - 1;
2504 hash += (unsigned) code + (unsigned) GET_MODE (x);
2505 fmt = GET_RTX_FORMAT (code);
2506 for (; i >= 0; i--)
2508 switch (fmt[i])
2510 case 'e':
2511 /* If we are about to do the last recursive call
2512 needed at this level, change it into iteration.
2513 This function is called enough to be worth it. */
2514 if (i == 0)
2516 x = XEXP (x, i);
2517 goto repeat;
2520 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2521 hash_arg_in_memory_p,
2522 have_reg_qty, cb);
2523 break;
2525 case 'E':
2526 for (j = 0; j < XVECLEN (x, i); j++)
2527 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2528 hash_arg_in_memory_p,
2529 have_reg_qty, cb);
2530 break;
2532 case 's':
2533 hash += hash_rtx_string (XSTR (x, i));
2534 break;
2536 case 'i':
2537 hash += (unsigned int) XINT (x, i);
2538 break;
2540 case '0': case 't':
2541 /* Unused. */
2542 break;
2544 default:
2545 gcc_unreachable ();
2549 return hash;
2552 /* Hash an rtx. We are careful to make sure the value is never negative.
2553 Equivalent registers hash identically.
2554 MODE is used in hashing for CONST_INTs only;
2555 otherwise the mode of X is used.
2557 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2559 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2560 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2562 Note that cse_insn knows that the hash code of a MEM expression
2563 is just (int) MEM plus the hash code of the address. */
2565 unsigned
2566 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2567 int *hash_arg_in_memory_p, bool have_reg_qty)
2569 return hash_rtx_cb (x, mode, do_not_record_p,
2570 hash_arg_in_memory_p, have_reg_qty, NULL);
2573 /* Hash an rtx X for cse via hash_rtx.
2574 Stores 1 in do_not_record if any subexpression is volatile.
2575 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2576 does not have the RTX_UNCHANGING_P bit set. */
2578 static inline unsigned
2579 canon_hash (rtx x, enum machine_mode mode)
2581 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2584 /* Like canon_hash but with no side effects, i.e. do_not_record
2585 and hash_arg_in_memory are not changed. */
2587 static inline unsigned
2588 safe_hash (rtx x, enum machine_mode mode)
2590 int dummy_do_not_record;
2591 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2594 /* Return 1 iff X and Y would canonicalize into the same thing,
2595 without actually constructing the canonicalization of either one.
2596 If VALIDATE is nonzero,
2597 we assume X is an expression being processed from the rtl
2598 and Y was found in the hash table. We check register refs
2599 in Y for being marked as valid.
2601 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2604 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2606 int i, j;
2607 enum rtx_code code;
2608 const char *fmt;
2610 /* Note: it is incorrect to assume an expression is equivalent to itself
2611 if VALIDATE is nonzero. */
2612 if (x == y && !validate)
2613 return 1;
2615 if (x == 0 || y == 0)
2616 return x == y;
2618 code = GET_CODE (x);
2619 if (code != GET_CODE (y))
2620 return 0;
2622 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2623 if (GET_MODE (x) != GET_MODE (y))
2624 return 0;
2626 /* MEMs refering to different address space are not equivalent. */
2627 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2628 return 0;
2630 switch (code)
2632 case PC:
2633 case CC0:
2634 case CONST_INT:
2635 case CONST_DOUBLE:
2636 case CONST_FIXED:
2637 return x == y;
2639 case LABEL_REF:
2640 return XEXP (x, 0) == XEXP (y, 0);
2642 case SYMBOL_REF:
2643 return XSTR (x, 0) == XSTR (y, 0);
2645 case REG:
2646 if (for_gcse)
2647 return REGNO (x) == REGNO (y);
2648 else
2650 unsigned int regno = REGNO (y);
2651 unsigned int i;
2652 unsigned int endregno = END_REGNO (y);
2654 /* If the quantities are not the same, the expressions are not
2655 equivalent. If there are and we are not to validate, they
2656 are equivalent. Otherwise, ensure all regs are up-to-date. */
2658 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2659 return 0;
2661 if (! validate)
2662 return 1;
2664 for (i = regno; i < endregno; i++)
2665 if (REG_IN_TABLE (i) != REG_TICK (i))
2666 return 0;
2668 return 1;
2671 case MEM:
2672 if (for_gcse)
2674 /* A volatile mem should not be considered equivalent to any
2675 other. */
2676 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2677 return 0;
2679 /* Can't merge two expressions in different alias sets, since we
2680 can decide that the expression is transparent in a block when
2681 it isn't, due to it being set with the different alias set.
2683 Also, can't merge two expressions with different MEM_ATTRS.
2684 They could e.g. be two different entities allocated into the
2685 same space on the stack (see e.g. PR25130). In that case, the
2686 MEM addresses can be the same, even though the two MEMs are
2687 absolutely not equivalent.
2689 But because really all MEM attributes should be the same for
2690 equivalent MEMs, we just use the invariant that MEMs that have
2691 the same attributes share the same mem_attrs data structure. */
2692 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2693 return 0;
2695 break;
2697 /* For commutative operations, check both orders. */
2698 case PLUS:
2699 case MULT:
2700 case AND:
2701 case IOR:
2702 case XOR:
2703 case NE:
2704 case EQ:
2705 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2706 validate, for_gcse)
2707 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2708 validate, for_gcse))
2709 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2710 validate, for_gcse)
2711 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2712 validate, for_gcse)));
2714 case ASM_OPERANDS:
2715 /* We don't use the generic code below because we want to
2716 disregard filename and line numbers. */
2718 /* A volatile asm isn't equivalent to any other. */
2719 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2720 return 0;
2722 if (GET_MODE (x) != GET_MODE (y)
2723 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2724 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2725 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2726 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2727 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2728 return 0;
2730 if (ASM_OPERANDS_INPUT_LENGTH (x))
2732 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2733 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2734 ASM_OPERANDS_INPUT (y, i),
2735 validate, for_gcse)
2736 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2737 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2738 return 0;
2741 return 1;
2743 default:
2744 break;
2747 /* Compare the elements. If any pair of corresponding elements
2748 fail to match, return 0 for the whole thing. */
2750 fmt = GET_RTX_FORMAT (code);
2751 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2753 switch (fmt[i])
2755 case 'e':
2756 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2757 validate, for_gcse))
2758 return 0;
2759 break;
2761 case 'E':
2762 if (XVECLEN (x, i) != XVECLEN (y, i))
2763 return 0;
2764 for (j = 0; j < XVECLEN (x, i); j++)
2765 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2766 validate, for_gcse))
2767 return 0;
2768 break;
2770 case 's':
2771 if (strcmp (XSTR (x, i), XSTR (y, i)))
2772 return 0;
2773 break;
2775 case 'i':
2776 if (XINT (x, i) != XINT (y, i))
2777 return 0;
2778 break;
2780 case 'w':
2781 if (XWINT (x, i) != XWINT (y, i))
2782 return 0;
2783 break;
2785 case '0':
2786 case 't':
2787 break;
2789 default:
2790 gcc_unreachable ();
2794 return 1;
2797 /* Return 1 if X has a value that can vary even between two
2798 executions of the program. 0 means X can be compared reliably
2799 against certain constants or near-constants. */
2801 static bool
2802 cse_rtx_varies_p (const_rtx x, bool from_alias)
2804 /* We need not check for X and the equivalence class being of the same
2805 mode because if X is equivalent to a constant in some mode, it
2806 doesn't vary in any mode. */
2808 if (REG_P (x)
2809 && REGNO_QTY_VALID_P (REGNO (x)))
2811 int x_q = REG_QTY (REGNO (x));
2812 struct qty_table_elem *x_ent = &qty_table[x_q];
2814 if (GET_MODE (x) == x_ent->mode
2815 && x_ent->const_rtx != NULL_RTX)
2816 return 0;
2819 if (GET_CODE (x) == PLUS
2820 && CONST_INT_P (XEXP (x, 1))
2821 && REG_P (XEXP (x, 0))
2822 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2824 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2825 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2827 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2828 && x0_ent->const_rtx != NULL_RTX)
2829 return 0;
2832 /* This can happen as the result of virtual register instantiation, if
2833 the initial constant is too large to be a valid address. This gives
2834 us a three instruction sequence, load large offset into a register,
2835 load fp minus a constant into a register, then a MEM which is the
2836 sum of the two `constant' registers. */
2837 if (GET_CODE (x) == PLUS
2838 && REG_P (XEXP (x, 0))
2839 && REG_P (XEXP (x, 1))
2840 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2841 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2843 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2844 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2845 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2846 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2848 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2849 && x0_ent->const_rtx != NULL_RTX
2850 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2851 && x1_ent->const_rtx != NULL_RTX)
2852 return 0;
2855 return rtx_varies_p (x, from_alias);
2858 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2859 the result if necessary. INSN is as for canon_reg. */
2861 static void
2862 validate_canon_reg (rtx *xloc, rtx insn)
2864 if (*xloc)
2866 rtx new_rtx = canon_reg (*xloc, insn);
2868 /* If replacing pseudo with hard reg or vice versa, ensure the
2869 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2870 gcc_assert (insn && new_rtx);
2871 validate_change (insn, xloc, new_rtx, 1);
2875 /* Canonicalize an expression:
2876 replace each register reference inside it
2877 with the "oldest" equivalent register.
2879 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2880 after we make our substitution. The calls are made with IN_GROUP nonzero
2881 so apply_change_group must be called upon the outermost return from this
2882 function (unless INSN is zero). The result of apply_change_group can
2883 generally be discarded since the changes we are making are optional. */
2885 static rtx
2886 canon_reg (rtx x, rtx insn)
2888 int i;
2889 enum rtx_code code;
2890 const char *fmt;
2892 if (x == 0)
2893 return x;
2895 code = GET_CODE (x);
2896 switch (code)
2898 case PC:
2899 case CC0:
2900 case CONST:
2901 case CONST_INT:
2902 case CONST_DOUBLE:
2903 case CONST_FIXED:
2904 case CONST_VECTOR:
2905 case SYMBOL_REF:
2906 case LABEL_REF:
2907 case ADDR_VEC:
2908 case ADDR_DIFF_VEC:
2909 return x;
2911 case REG:
2913 int first;
2914 int q;
2915 struct qty_table_elem *ent;
2917 /* Never replace a hard reg, because hard regs can appear
2918 in more than one machine mode, and we must preserve the mode
2919 of each occurrence. Also, some hard regs appear in
2920 MEMs that are shared and mustn't be altered. Don't try to
2921 replace any reg that maps to a reg of class NO_REGS. */
2922 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2923 || ! REGNO_QTY_VALID_P (REGNO (x)))
2924 return x;
2926 q = REG_QTY (REGNO (x));
2927 ent = &qty_table[q];
2928 first = ent->first_reg;
2929 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2930 : REGNO_REG_CLASS (first) == NO_REGS ? x
2931 : gen_rtx_REG (ent->mode, first));
2934 default:
2935 break;
2938 fmt = GET_RTX_FORMAT (code);
2939 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2941 int j;
2943 if (fmt[i] == 'e')
2944 validate_canon_reg (&XEXP (x, i), insn);
2945 else if (fmt[i] == 'E')
2946 for (j = 0; j < XVECLEN (x, i); j++)
2947 validate_canon_reg (&XVECEXP (x, i, j), insn);
2950 return x;
2953 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2954 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2955 what values are being compared.
2957 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2958 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2959 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2960 compared to produce cc0.
2962 The return value is the comparison operator and is either the code of
2963 A or the code corresponding to the inverse of the comparison. */
2965 static enum rtx_code
2966 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2967 enum machine_mode *pmode1, enum machine_mode *pmode2)
2969 rtx arg1, arg2;
2971 arg1 = *parg1, arg2 = *parg2;
2973 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2975 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2977 /* Set nonzero when we find something of interest. */
2978 rtx x = 0;
2979 int reverse_code = 0;
2980 struct table_elt *p = 0;
2982 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2983 On machines with CC0, this is the only case that can occur, since
2984 fold_rtx will return the COMPARE or item being compared with zero
2985 when given CC0. */
2987 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2988 x = arg1;
2990 /* If ARG1 is a comparison operator and CODE is testing for
2991 STORE_FLAG_VALUE, get the inner arguments. */
2993 else if (COMPARISON_P (arg1))
2995 #ifdef FLOAT_STORE_FLAG_VALUE
2996 REAL_VALUE_TYPE fsfv;
2997 #endif
2999 if (code == NE
3000 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3001 && code == LT && STORE_FLAG_VALUE == -1)
3002 #ifdef FLOAT_STORE_FLAG_VALUE
3003 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3004 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3005 REAL_VALUE_NEGATIVE (fsfv)))
3006 #endif
3008 x = arg1;
3009 else if (code == EQ
3010 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3011 && code == GE && STORE_FLAG_VALUE == -1)
3012 #ifdef FLOAT_STORE_FLAG_VALUE
3013 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3014 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3015 REAL_VALUE_NEGATIVE (fsfv)))
3016 #endif
3018 x = arg1, reverse_code = 1;
3021 /* ??? We could also check for
3023 (ne (and (eq (...) (const_int 1))) (const_int 0))
3025 and related forms, but let's wait until we see them occurring. */
3027 if (x == 0)
3028 /* Look up ARG1 in the hash table and see if it has an equivalence
3029 that lets us see what is being compared. */
3030 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3031 if (p)
3033 p = p->first_same_value;
3035 /* If what we compare is already known to be constant, that is as
3036 good as it gets.
3037 We need to break the loop in this case, because otherwise we
3038 can have an infinite loop when looking at a reg that is known
3039 to be a constant which is the same as a comparison of a reg
3040 against zero which appears later in the insn stream, which in
3041 turn is constant and the same as the comparison of the first reg
3042 against zero... */
3043 if (p->is_const)
3044 break;
3047 for (; p; p = p->next_same_value)
3049 enum machine_mode inner_mode = GET_MODE (p->exp);
3050 #ifdef FLOAT_STORE_FLAG_VALUE
3051 REAL_VALUE_TYPE fsfv;
3052 #endif
3054 /* If the entry isn't valid, skip it. */
3055 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3056 continue;
3058 if (GET_CODE (p->exp) == COMPARE
3059 /* Another possibility is that this machine has a compare insn
3060 that includes the comparison code. In that case, ARG1 would
3061 be equivalent to a comparison operation that would set ARG1 to
3062 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3063 ORIG_CODE is the actual comparison being done; if it is an EQ,
3064 we must reverse ORIG_CODE. On machine with a negative value
3065 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3066 || ((code == NE
3067 || (code == LT
3068 && val_signbit_known_set_p (inner_mode,
3069 STORE_FLAG_VALUE))
3070 #ifdef FLOAT_STORE_FLAG_VALUE
3071 || (code == LT
3072 && SCALAR_FLOAT_MODE_P (inner_mode)
3073 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3074 REAL_VALUE_NEGATIVE (fsfv)))
3075 #endif
3077 && COMPARISON_P (p->exp)))
3079 x = p->exp;
3080 break;
3082 else if ((code == EQ
3083 || (code == GE
3084 && val_signbit_known_set_p (inner_mode,
3085 STORE_FLAG_VALUE))
3086 #ifdef FLOAT_STORE_FLAG_VALUE
3087 || (code == GE
3088 && SCALAR_FLOAT_MODE_P (inner_mode)
3089 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3090 REAL_VALUE_NEGATIVE (fsfv)))
3091 #endif
3093 && COMPARISON_P (p->exp))
3095 reverse_code = 1;
3096 x = p->exp;
3097 break;
3100 /* If this non-trapping address, e.g. fp + constant, the
3101 equivalent is a better operand since it may let us predict
3102 the value of the comparison. */
3103 else if (!rtx_addr_can_trap_p (p->exp))
3105 arg1 = p->exp;
3106 continue;
3110 /* If we didn't find a useful equivalence for ARG1, we are done.
3111 Otherwise, set up for the next iteration. */
3112 if (x == 0)
3113 break;
3115 /* If we need to reverse the comparison, make sure that that is
3116 possible -- we can't necessarily infer the value of GE from LT
3117 with floating-point operands. */
3118 if (reverse_code)
3120 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3121 if (reversed == UNKNOWN)
3122 break;
3123 else
3124 code = reversed;
3126 else if (COMPARISON_P (x))
3127 code = GET_CODE (x);
3128 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3131 /* Return our results. Return the modes from before fold_rtx
3132 because fold_rtx might produce const_int, and then it's too late. */
3133 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3134 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3136 return code;
3139 /* If X is a nontrivial arithmetic operation on an argument for which
3140 a constant value can be determined, return the result of operating
3141 on that value, as a constant. Otherwise, return X, possibly with
3142 one or more operands changed to a forward-propagated constant.
3144 If X is a register whose contents are known, we do NOT return
3145 those contents here; equiv_constant is called to perform that task.
3146 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3148 INSN is the insn that we may be modifying. If it is 0, make a copy
3149 of X before modifying it. */
3151 static rtx
3152 fold_rtx (rtx x, rtx insn)
3154 enum rtx_code code;
3155 enum machine_mode mode;
3156 const char *fmt;
3157 int i;
3158 rtx new_rtx = 0;
3159 int changed = 0;
3161 /* Operands of X. */
3162 rtx folded_arg0;
3163 rtx folded_arg1;
3165 /* Constant equivalents of first three operands of X;
3166 0 when no such equivalent is known. */
3167 rtx const_arg0;
3168 rtx const_arg1;
3169 rtx const_arg2;
3171 /* The mode of the first operand of X. We need this for sign and zero
3172 extends. */
3173 enum machine_mode mode_arg0;
3175 if (x == 0)
3176 return x;
3178 /* Try to perform some initial simplifications on X. */
3179 code = GET_CODE (x);
3180 switch (code)
3182 case MEM:
3183 case SUBREG:
3184 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3185 return new_rtx;
3186 return x;
3188 case CONST:
3189 case CONST_INT:
3190 case CONST_DOUBLE:
3191 case CONST_FIXED:
3192 case CONST_VECTOR:
3193 case SYMBOL_REF:
3194 case LABEL_REF:
3195 case REG:
3196 case PC:
3197 /* No use simplifying an EXPR_LIST
3198 since they are used only for lists of args
3199 in a function call's REG_EQUAL note. */
3200 case EXPR_LIST:
3201 return x;
3203 #ifdef HAVE_cc0
3204 case CC0:
3205 return prev_insn_cc0;
3206 #endif
3208 case ASM_OPERANDS:
3209 if (insn)
3211 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3212 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3213 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3215 return x;
3217 #ifdef NO_FUNCTION_CSE
3218 case CALL:
3219 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3220 return x;
3221 break;
3222 #endif
3224 /* Anything else goes through the loop below. */
3225 default:
3226 break;
3229 mode = GET_MODE (x);
3230 const_arg0 = 0;
3231 const_arg1 = 0;
3232 const_arg2 = 0;
3233 mode_arg0 = VOIDmode;
3235 /* Try folding our operands.
3236 Then see which ones have constant values known. */
3238 fmt = GET_RTX_FORMAT (code);
3239 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3240 if (fmt[i] == 'e')
3242 rtx folded_arg = XEXP (x, i), const_arg;
3243 enum machine_mode mode_arg = GET_MODE (folded_arg);
3245 switch (GET_CODE (folded_arg))
3247 case MEM:
3248 case REG:
3249 case SUBREG:
3250 const_arg = equiv_constant (folded_arg);
3251 break;
3253 case CONST:
3254 case CONST_INT:
3255 case SYMBOL_REF:
3256 case LABEL_REF:
3257 case CONST_DOUBLE:
3258 case CONST_FIXED:
3259 case CONST_VECTOR:
3260 const_arg = folded_arg;
3261 break;
3263 #ifdef HAVE_cc0
3264 case CC0:
3265 folded_arg = prev_insn_cc0;
3266 mode_arg = prev_insn_cc0_mode;
3267 const_arg = equiv_constant (folded_arg);
3268 break;
3269 #endif
3271 default:
3272 folded_arg = fold_rtx (folded_arg, insn);
3273 const_arg = equiv_constant (folded_arg);
3274 break;
3277 /* For the first three operands, see if the operand
3278 is constant or equivalent to a constant. */
3279 switch (i)
3281 case 0:
3282 folded_arg0 = folded_arg;
3283 const_arg0 = const_arg;
3284 mode_arg0 = mode_arg;
3285 break;
3286 case 1:
3287 folded_arg1 = folded_arg;
3288 const_arg1 = const_arg;
3289 break;
3290 case 2:
3291 const_arg2 = const_arg;
3292 break;
3295 /* Pick the least expensive of the argument and an equivalent constant
3296 argument. */
3297 if (const_arg != 0
3298 && const_arg != folded_arg
3299 && COST_IN (const_arg, code) <= COST_IN (folded_arg, code)
3301 /* It's not safe to substitute the operand of a conversion
3302 operator with a constant, as the conversion's identity
3303 depends upon the mode of its operand. This optimization
3304 is handled by the call to simplify_unary_operation. */
3305 && (GET_RTX_CLASS (code) != RTX_UNARY
3306 || GET_MODE (const_arg) == mode_arg0
3307 || (code != ZERO_EXTEND
3308 && code != SIGN_EXTEND
3309 && code != TRUNCATE
3310 && code != FLOAT_TRUNCATE
3311 && code != FLOAT_EXTEND
3312 && code != FLOAT
3313 && code != FIX
3314 && code != UNSIGNED_FLOAT
3315 && code != UNSIGNED_FIX)))
3316 folded_arg = const_arg;
3318 if (folded_arg == XEXP (x, i))
3319 continue;
3321 if (insn == NULL_RTX && !changed)
3322 x = copy_rtx (x);
3323 changed = 1;
3324 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3327 if (changed)
3329 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3330 consistent with the order in X. */
3331 if (canonicalize_change_group (insn, x))
3333 rtx tem;
3334 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3335 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3338 apply_change_group ();
3341 /* If X is an arithmetic operation, see if we can simplify it. */
3343 switch (GET_RTX_CLASS (code))
3345 case RTX_UNARY:
3347 /* We can't simplify extension ops unless we know the
3348 original mode. */
3349 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3350 && mode_arg0 == VOIDmode)
3351 break;
3353 new_rtx = simplify_unary_operation (code, mode,
3354 const_arg0 ? const_arg0 : folded_arg0,
3355 mode_arg0);
3357 break;
3359 case RTX_COMPARE:
3360 case RTX_COMM_COMPARE:
3361 /* See what items are actually being compared and set FOLDED_ARG[01]
3362 to those values and CODE to the actual comparison code. If any are
3363 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3364 do anything if both operands are already known to be constant. */
3366 /* ??? Vector mode comparisons are not supported yet. */
3367 if (VECTOR_MODE_P (mode))
3368 break;
3370 if (const_arg0 == 0 || const_arg1 == 0)
3372 struct table_elt *p0, *p1;
3373 rtx true_rtx, false_rtx;
3374 enum machine_mode mode_arg1;
3376 if (SCALAR_FLOAT_MODE_P (mode))
3378 #ifdef FLOAT_STORE_FLAG_VALUE
3379 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3380 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3381 #else
3382 true_rtx = NULL_RTX;
3383 #endif
3384 false_rtx = CONST0_RTX (mode);
3386 else
3388 true_rtx = const_true_rtx;
3389 false_rtx = const0_rtx;
3392 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3393 &mode_arg0, &mode_arg1);
3395 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3396 what kinds of things are being compared, so we can't do
3397 anything with this comparison. */
3399 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3400 break;
3402 const_arg0 = equiv_constant (folded_arg0);
3403 const_arg1 = equiv_constant (folded_arg1);
3405 /* If we do not now have two constants being compared, see
3406 if we can nevertheless deduce some things about the
3407 comparison. */
3408 if (const_arg0 == 0 || const_arg1 == 0)
3410 if (const_arg1 != NULL)
3412 rtx cheapest_simplification;
3413 int cheapest_cost;
3414 rtx simp_result;
3415 struct table_elt *p;
3417 /* See if we can find an equivalent of folded_arg0
3418 that gets us a cheaper expression, possibly a
3419 constant through simplifications. */
3420 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3421 mode_arg0);
3423 if (p != NULL)
3425 cheapest_simplification = x;
3426 cheapest_cost = COST (x);
3428 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3430 int cost;
3432 /* If the entry isn't valid, skip it. */
3433 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3434 continue;
3436 /* Try to simplify using this equivalence. */
3437 simp_result
3438 = simplify_relational_operation (code, mode,
3439 mode_arg0,
3440 p->exp,
3441 const_arg1);
3443 if (simp_result == NULL)
3444 continue;
3446 cost = COST (simp_result);
3447 if (cost < cheapest_cost)
3449 cheapest_cost = cost;
3450 cheapest_simplification = simp_result;
3454 /* If we have a cheaper expression now, use that
3455 and try folding it further, from the top. */
3456 if (cheapest_simplification != x)
3457 return fold_rtx (copy_rtx (cheapest_simplification),
3458 insn);
3462 /* See if the two operands are the same. */
3464 if ((REG_P (folded_arg0)
3465 && REG_P (folded_arg1)
3466 && (REG_QTY (REGNO (folded_arg0))
3467 == REG_QTY (REGNO (folded_arg1))))
3468 || ((p0 = lookup (folded_arg0,
3469 SAFE_HASH (folded_arg0, mode_arg0),
3470 mode_arg0))
3471 && (p1 = lookup (folded_arg1,
3472 SAFE_HASH (folded_arg1, mode_arg0),
3473 mode_arg0))
3474 && p0->first_same_value == p1->first_same_value))
3475 folded_arg1 = folded_arg0;
3477 /* If FOLDED_ARG0 is a register, see if the comparison we are
3478 doing now is either the same as we did before or the reverse
3479 (we only check the reverse if not floating-point). */
3480 else if (REG_P (folded_arg0))
3482 int qty = REG_QTY (REGNO (folded_arg0));
3484 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3486 struct qty_table_elem *ent = &qty_table[qty];
3488 if ((comparison_dominates_p (ent->comparison_code, code)
3489 || (! FLOAT_MODE_P (mode_arg0)
3490 && comparison_dominates_p (ent->comparison_code,
3491 reverse_condition (code))))
3492 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3493 || (const_arg1
3494 && rtx_equal_p (ent->comparison_const,
3495 const_arg1))
3496 || (REG_P (folded_arg1)
3497 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3499 if (comparison_dominates_p (ent->comparison_code, code))
3501 if (true_rtx)
3502 return true_rtx;
3503 else
3504 break;
3506 else
3507 return false_rtx;
3514 /* If we are comparing against zero, see if the first operand is
3515 equivalent to an IOR with a constant. If so, we may be able to
3516 determine the result of this comparison. */
3517 if (const_arg1 == const0_rtx && !const_arg0)
3519 rtx y = lookup_as_function (folded_arg0, IOR);
3520 rtx inner_const;
3522 if (y != 0
3523 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3524 && CONST_INT_P (inner_const)
3525 && INTVAL (inner_const) != 0)
3526 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3530 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3531 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3532 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3534 break;
3536 case RTX_BIN_ARITH:
3537 case RTX_COMM_ARITH:
3538 switch (code)
3540 case PLUS:
3541 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3542 with that LABEL_REF as its second operand. If so, the result is
3543 the first operand of that MINUS. This handles switches with an
3544 ADDR_DIFF_VEC table. */
3545 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3547 rtx y
3548 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3549 : lookup_as_function (folded_arg0, MINUS);
3551 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3552 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3553 return XEXP (y, 0);
3555 /* Now try for a CONST of a MINUS like the above. */
3556 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3557 : lookup_as_function (folded_arg0, CONST))) != 0
3558 && GET_CODE (XEXP (y, 0)) == MINUS
3559 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3560 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3561 return XEXP (XEXP (y, 0), 0);
3564 /* Likewise if the operands are in the other order. */
3565 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3567 rtx y
3568 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3569 : lookup_as_function (folded_arg1, MINUS);
3571 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3572 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3573 return XEXP (y, 0);
3575 /* Now try for a CONST of a MINUS like the above. */
3576 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3577 : lookup_as_function (folded_arg1, CONST))) != 0
3578 && GET_CODE (XEXP (y, 0)) == MINUS
3579 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3580 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3581 return XEXP (XEXP (y, 0), 0);
3584 /* If second operand is a register equivalent to a negative
3585 CONST_INT, see if we can find a register equivalent to the
3586 positive constant. Make a MINUS if so. Don't do this for
3587 a non-negative constant since we might then alternate between
3588 choosing positive and negative constants. Having the positive
3589 constant previously-used is the more common case. Be sure
3590 the resulting constant is non-negative; if const_arg1 were
3591 the smallest negative number this would overflow: depending
3592 on the mode, this would either just be the same value (and
3593 hence not save anything) or be incorrect. */
3594 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3595 && INTVAL (const_arg1) < 0
3596 /* This used to test
3598 -INTVAL (const_arg1) >= 0
3600 But The Sun V5.0 compilers mis-compiled that test. So
3601 instead we test for the problematic value in a more direct
3602 manner and hope the Sun compilers get it correct. */
3603 && INTVAL (const_arg1) !=
3604 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3605 && REG_P (folded_arg1))
3607 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3608 struct table_elt *p
3609 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3611 if (p)
3612 for (p = p->first_same_value; p; p = p->next_same_value)
3613 if (REG_P (p->exp))
3614 return simplify_gen_binary (MINUS, mode, folded_arg0,
3615 canon_reg (p->exp, NULL_RTX));
3617 goto from_plus;
3619 case MINUS:
3620 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3621 If so, produce (PLUS Z C2-C). */
3622 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3624 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3625 if (y && CONST_INT_P (XEXP (y, 1)))
3626 return fold_rtx (plus_constant (copy_rtx (y),
3627 -INTVAL (const_arg1)),
3628 NULL_RTX);
3631 /* Fall through. */
3633 from_plus:
3634 case SMIN: case SMAX: case UMIN: case UMAX:
3635 case IOR: case AND: case XOR:
3636 case MULT:
3637 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3638 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3639 is known to be of similar form, we may be able to replace the
3640 operation with a combined operation. This may eliminate the
3641 intermediate operation if every use is simplified in this way.
3642 Note that the similar optimization done by combine.c only works
3643 if the intermediate operation's result has only one reference. */
3645 if (REG_P (folded_arg0)
3646 && const_arg1 && CONST_INT_P (const_arg1))
3648 int is_shift
3649 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3650 rtx y, inner_const, new_const;
3651 rtx canon_const_arg1 = const_arg1;
3652 enum rtx_code associate_code;
3654 if (is_shift
3655 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3656 || INTVAL (const_arg1) < 0))
3658 if (SHIFT_COUNT_TRUNCATED)
3659 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3660 & (GET_MODE_BITSIZE (mode)
3661 - 1));
3662 else
3663 break;
3666 y = lookup_as_function (folded_arg0, code);
3667 if (y == 0)
3668 break;
3670 /* If we have compiled a statement like
3671 "if (x == (x & mask1))", and now are looking at
3672 "x & mask2", we will have a case where the first operand
3673 of Y is the same as our first operand. Unless we detect
3674 this case, an infinite loop will result. */
3675 if (XEXP (y, 0) == folded_arg0)
3676 break;
3678 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3679 if (!inner_const || !CONST_INT_P (inner_const))
3680 break;
3682 /* Don't associate these operations if they are a PLUS with the
3683 same constant and it is a power of two. These might be doable
3684 with a pre- or post-increment. Similarly for two subtracts of
3685 identical powers of two with post decrement. */
3687 if (code == PLUS && const_arg1 == inner_const
3688 && ((HAVE_PRE_INCREMENT
3689 && exact_log2 (INTVAL (const_arg1)) >= 0)
3690 || (HAVE_POST_INCREMENT
3691 && exact_log2 (INTVAL (const_arg1)) >= 0)
3692 || (HAVE_PRE_DECREMENT
3693 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3694 || (HAVE_POST_DECREMENT
3695 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3696 break;
3698 /* ??? Vector mode shifts by scalar
3699 shift operand are not supported yet. */
3700 if (is_shift && VECTOR_MODE_P (mode))
3701 break;
3703 if (is_shift
3704 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3705 || INTVAL (inner_const) < 0))
3707 if (SHIFT_COUNT_TRUNCATED)
3708 inner_const = GEN_INT (INTVAL (inner_const)
3709 & (GET_MODE_BITSIZE (mode) - 1));
3710 else
3711 break;
3714 /* Compute the code used to compose the constants. For example,
3715 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3717 associate_code = (is_shift || code == MINUS ? PLUS : code);
3719 new_const = simplify_binary_operation (associate_code, mode,
3720 canon_const_arg1,
3721 inner_const);
3723 if (new_const == 0)
3724 break;
3726 /* If we are associating shift operations, don't let this
3727 produce a shift of the size of the object or larger.
3728 This could occur when we follow a sign-extend by a right
3729 shift on a machine that does a sign-extend as a pair
3730 of shifts. */
3732 if (is_shift
3733 && CONST_INT_P (new_const)
3734 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3736 /* As an exception, we can turn an ASHIFTRT of this
3737 form into a shift of the number of bits - 1. */
3738 if (code == ASHIFTRT)
3739 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3740 else if (!side_effects_p (XEXP (y, 0)))
3741 return CONST0_RTX (mode);
3742 else
3743 break;
3746 y = copy_rtx (XEXP (y, 0));
3748 /* If Y contains our first operand (the most common way this
3749 can happen is if Y is a MEM), we would do into an infinite
3750 loop if we tried to fold it. So don't in that case. */
3752 if (! reg_mentioned_p (folded_arg0, y))
3753 y = fold_rtx (y, insn);
3755 return simplify_gen_binary (code, mode, y, new_const);
3757 break;
3759 case DIV: case UDIV:
3760 /* ??? The associative optimization performed immediately above is
3761 also possible for DIV and UDIV using associate_code of MULT.
3762 However, we would need extra code to verify that the
3763 multiplication does not overflow, that is, there is no overflow
3764 in the calculation of new_const. */
3765 break;
3767 default:
3768 break;
3771 new_rtx = simplify_binary_operation (code, mode,
3772 const_arg0 ? const_arg0 : folded_arg0,
3773 const_arg1 ? const_arg1 : folded_arg1);
3774 break;
3776 case RTX_OBJ:
3777 /* (lo_sum (high X) X) is simply X. */
3778 if (code == LO_SUM && const_arg0 != 0
3779 && GET_CODE (const_arg0) == HIGH
3780 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3781 return const_arg1;
3782 break;
3784 case RTX_TERNARY:
3785 case RTX_BITFIELD_OPS:
3786 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3787 const_arg0 ? const_arg0 : folded_arg0,
3788 const_arg1 ? const_arg1 : folded_arg1,
3789 const_arg2 ? const_arg2 : XEXP (x, 2));
3790 break;
3792 default:
3793 break;
3796 return new_rtx ? new_rtx : x;
3799 /* Return a constant value currently equivalent to X.
3800 Return 0 if we don't know one. */
3802 static rtx
3803 equiv_constant (rtx x)
3805 if (REG_P (x)
3806 && REGNO_QTY_VALID_P (REGNO (x)))
3808 int x_q = REG_QTY (REGNO (x));
3809 struct qty_table_elem *x_ent = &qty_table[x_q];
3811 if (x_ent->const_rtx)
3812 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3815 if (x == 0 || CONSTANT_P (x))
3816 return x;
3818 if (GET_CODE (x) == SUBREG)
3820 enum machine_mode mode = GET_MODE (x);
3821 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3822 rtx new_rtx;
3824 /* See if we previously assigned a constant value to this SUBREG. */
3825 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3826 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3827 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3828 return new_rtx;
3830 /* If we didn't and if doing so makes sense, see if we previously
3831 assigned a constant value to the enclosing word mode SUBREG. */
3832 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3833 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3835 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3836 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3838 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3839 new_rtx = lookup_as_function (y, CONST_INT);
3840 if (new_rtx)
3841 return gen_lowpart (mode, new_rtx);
3845 /* Otherwise see if we already have a constant for the inner REG. */
3846 if (REG_P (SUBREG_REG (x))
3847 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3848 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3850 return 0;
3853 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3854 the hash table in case its value was seen before. */
3856 if (MEM_P (x))
3858 struct table_elt *elt;
3860 x = avoid_constant_pool_reference (x);
3861 if (CONSTANT_P (x))
3862 return x;
3864 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3865 if (elt == 0)
3866 return 0;
3868 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3869 if (elt->is_const && CONSTANT_P (elt->exp))
3870 return elt->exp;
3873 return 0;
3876 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3877 "taken" branch.
3879 In certain cases, this can cause us to add an equivalence. For example,
3880 if we are following the taken case of
3881 if (i == 2)
3882 we can add the fact that `i' and '2' are now equivalent.
3884 In any case, we can record that this comparison was passed. If the same
3885 comparison is seen later, we will know its value. */
3887 static void
3888 record_jump_equiv (rtx insn, bool taken)
3890 int cond_known_true;
3891 rtx op0, op1;
3892 rtx set;
3893 enum machine_mode mode, mode0, mode1;
3894 int reversed_nonequality = 0;
3895 enum rtx_code code;
3897 /* Ensure this is the right kind of insn. */
3898 gcc_assert (any_condjump_p (insn));
3900 set = pc_set (insn);
3902 /* See if this jump condition is known true or false. */
3903 if (taken)
3904 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3905 else
3906 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3908 /* Get the type of comparison being done and the operands being compared.
3909 If we had to reverse a non-equality condition, record that fact so we
3910 know that it isn't valid for floating-point. */
3911 code = GET_CODE (XEXP (SET_SRC (set), 0));
3912 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3913 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3915 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3916 if (! cond_known_true)
3918 code = reversed_comparison_code_parts (code, op0, op1, insn);
3920 /* Don't remember if we can't find the inverse. */
3921 if (code == UNKNOWN)
3922 return;
3925 /* The mode is the mode of the non-constant. */
3926 mode = mode0;
3927 if (mode1 != VOIDmode)
3928 mode = mode1;
3930 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3933 /* Yet another form of subreg creation. In this case, we want something in
3934 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3936 static rtx
3937 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3939 enum machine_mode op_mode = GET_MODE (op);
3940 if (op_mode == mode || op_mode == VOIDmode)
3941 return op;
3942 return lowpart_subreg (mode, op, op_mode);
3945 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3946 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3947 Make any useful entries we can with that information. Called from
3948 above function and called recursively. */
3950 static void
3951 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3952 rtx op1, int reversed_nonequality)
3954 unsigned op0_hash, op1_hash;
3955 int op0_in_memory, op1_in_memory;
3956 struct table_elt *op0_elt, *op1_elt;
3958 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3959 we know that they are also equal in the smaller mode (this is also
3960 true for all smaller modes whether or not there is a SUBREG, but
3961 is not worth testing for with no SUBREG). */
3963 /* Note that GET_MODE (op0) may not equal MODE. */
3964 if (code == EQ && paradoxical_subreg_p (op0))
3966 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3967 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3968 if (tem)
3969 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3970 reversed_nonequality);
3973 if (code == EQ && paradoxical_subreg_p (op1))
3975 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3976 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3977 if (tem)
3978 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3979 reversed_nonequality);
3982 /* Similarly, if this is an NE comparison, and either is a SUBREG
3983 making a smaller mode, we know the whole thing is also NE. */
3985 /* Note that GET_MODE (op0) may not equal MODE;
3986 if we test MODE instead, we can get an infinite recursion
3987 alternating between two modes each wider than MODE. */
3989 if (code == NE && GET_CODE (op0) == SUBREG
3990 && subreg_lowpart_p (op0)
3991 && (GET_MODE_SIZE (GET_MODE (op0))
3992 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3994 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3995 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3996 if (tem)
3997 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3998 reversed_nonequality);
4001 if (code == NE && GET_CODE (op1) == SUBREG
4002 && subreg_lowpart_p (op1)
4003 && (GET_MODE_SIZE (GET_MODE (op1))
4004 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4006 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4007 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4008 if (tem)
4009 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4010 reversed_nonequality);
4013 /* Hash both operands. */
4015 do_not_record = 0;
4016 hash_arg_in_memory = 0;
4017 op0_hash = HASH (op0, mode);
4018 op0_in_memory = hash_arg_in_memory;
4020 if (do_not_record)
4021 return;
4023 do_not_record = 0;
4024 hash_arg_in_memory = 0;
4025 op1_hash = HASH (op1, mode);
4026 op1_in_memory = hash_arg_in_memory;
4028 if (do_not_record)
4029 return;
4031 /* Look up both operands. */
4032 op0_elt = lookup (op0, op0_hash, mode);
4033 op1_elt = lookup (op1, op1_hash, mode);
4035 /* If both operands are already equivalent or if they are not in the
4036 table but are identical, do nothing. */
4037 if ((op0_elt != 0 && op1_elt != 0
4038 && op0_elt->first_same_value == op1_elt->first_same_value)
4039 || op0 == op1 || rtx_equal_p (op0, op1))
4040 return;
4042 /* If we aren't setting two things equal all we can do is save this
4043 comparison. Similarly if this is floating-point. In the latter
4044 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4045 If we record the equality, we might inadvertently delete code
4046 whose intent was to change -0 to +0. */
4048 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4050 struct qty_table_elem *ent;
4051 int qty;
4053 /* If we reversed a floating-point comparison, if OP0 is not a
4054 register, or if OP1 is neither a register or constant, we can't
4055 do anything. */
4057 if (!REG_P (op1))
4058 op1 = equiv_constant (op1);
4060 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4061 || !REG_P (op0) || op1 == 0)
4062 return;
4064 /* Put OP0 in the hash table if it isn't already. This gives it a
4065 new quantity number. */
4066 if (op0_elt == 0)
4068 if (insert_regs (op0, NULL, 0))
4070 rehash_using_reg (op0);
4071 op0_hash = HASH (op0, mode);
4073 /* If OP0 is contained in OP1, this changes its hash code
4074 as well. Faster to rehash than to check, except
4075 for the simple case of a constant. */
4076 if (! CONSTANT_P (op1))
4077 op1_hash = HASH (op1,mode);
4080 op0_elt = insert (op0, NULL, op0_hash, mode);
4081 op0_elt->in_memory = op0_in_memory;
4084 qty = REG_QTY (REGNO (op0));
4085 ent = &qty_table[qty];
4087 ent->comparison_code = code;
4088 if (REG_P (op1))
4090 /* Look it up again--in case op0 and op1 are the same. */
4091 op1_elt = lookup (op1, op1_hash, mode);
4093 /* Put OP1 in the hash table so it gets a new quantity number. */
4094 if (op1_elt == 0)
4096 if (insert_regs (op1, NULL, 0))
4098 rehash_using_reg (op1);
4099 op1_hash = HASH (op1, mode);
4102 op1_elt = insert (op1, NULL, op1_hash, mode);
4103 op1_elt->in_memory = op1_in_memory;
4106 ent->comparison_const = NULL_RTX;
4107 ent->comparison_qty = REG_QTY (REGNO (op1));
4109 else
4111 ent->comparison_const = op1;
4112 ent->comparison_qty = -1;
4115 return;
4118 /* If either side is still missing an equivalence, make it now,
4119 then merge the equivalences. */
4121 if (op0_elt == 0)
4123 if (insert_regs (op0, NULL, 0))
4125 rehash_using_reg (op0);
4126 op0_hash = HASH (op0, mode);
4129 op0_elt = insert (op0, NULL, op0_hash, mode);
4130 op0_elt->in_memory = op0_in_memory;
4133 if (op1_elt == 0)
4135 if (insert_regs (op1, NULL, 0))
4137 rehash_using_reg (op1);
4138 op1_hash = HASH (op1, mode);
4141 op1_elt = insert (op1, NULL, op1_hash, mode);
4142 op1_elt->in_memory = op1_in_memory;
4145 merge_equiv_classes (op0_elt, op1_elt);
4148 /* CSE processing for one instruction.
4149 First simplify sources and addresses of all assignments
4150 in the instruction, using previously-computed equivalents values.
4151 Then install the new sources and destinations in the table
4152 of available values. */
4154 /* Data on one SET contained in the instruction. */
4156 struct set
4158 /* The SET rtx itself. */
4159 rtx rtl;
4160 /* The SET_SRC of the rtx (the original value, if it is changing). */
4161 rtx src;
4162 /* The hash-table element for the SET_SRC of the SET. */
4163 struct table_elt *src_elt;
4164 /* Hash value for the SET_SRC. */
4165 unsigned src_hash;
4166 /* Hash value for the SET_DEST. */
4167 unsigned dest_hash;
4168 /* The SET_DEST, with SUBREG, etc., stripped. */
4169 rtx inner_dest;
4170 /* Nonzero if the SET_SRC is in memory. */
4171 char src_in_memory;
4172 /* Nonzero if the SET_SRC contains something
4173 whose value cannot be predicted and understood. */
4174 char src_volatile;
4175 /* Original machine mode, in case it becomes a CONST_INT.
4176 The size of this field should match the size of the mode
4177 field of struct rtx_def (see rtl.h). */
4178 ENUM_BITFIELD(machine_mode) mode : 8;
4179 /* A constant equivalent for SET_SRC, if any. */
4180 rtx src_const;
4181 /* Hash value of constant equivalent for SET_SRC. */
4182 unsigned src_const_hash;
4183 /* Table entry for constant equivalent for SET_SRC, if any. */
4184 struct table_elt *src_const_elt;
4185 /* Table entry for the destination address. */
4186 struct table_elt *dest_addr_elt;
4189 static void
4190 cse_insn (rtx insn)
4192 rtx x = PATTERN (insn);
4193 int i;
4194 rtx tem;
4195 int n_sets = 0;
4197 rtx src_eqv = 0;
4198 struct table_elt *src_eqv_elt = 0;
4199 int src_eqv_volatile = 0;
4200 int src_eqv_in_memory = 0;
4201 unsigned src_eqv_hash = 0;
4203 struct set *sets = (struct set *) 0;
4205 this_insn = insn;
4206 #ifdef HAVE_cc0
4207 /* Records what this insn does to set CC0. */
4208 this_insn_cc0 = 0;
4209 this_insn_cc0_mode = VOIDmode;
4210 #endif
4212 /* Find all the SETs and CLOBBERs in this instruction.
4213 Record all the SETs in the array `set' and count them.
4214 Also determine whether there is a CLOBBER that invalidates
4215 all memory references, or all references at varying addresses. */
4217 if (CALL_P (insn))
4219 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4221 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4222 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4223 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4227 if (GET_CODE (x) == SET)
4229 sets = XALLOCA (struct set);
4230 sets[0].rtl = x;
4232 /* Ignore SETs that are unconditional jumps.
4233 They never need cse processing, so this does not hurt.
4234 The reason is not efficiency but rather
4235 so that we can test at the end for instructions
4236 that have been simplified to unconditional jumps
4237 and not be misled by unchanged instructions
4238 that were unconditional jumps to begin with. */
4239 if (SET_DEST (x) == pc_rtx
4240 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4243 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4244 The hard function value register is used only once, to copy to
4245 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4246 Ensure we invalidate the destination register. On the 80386 no
4247 other code would invalidate it since it is a fixed_reg.
4248 We need not check the return of apply_change_group; see canon_reg. */
4250 else if (GET_CODE (SET_SRC (x)) == CALL)
4252 canon_reg (SET_SRC (x), insn);
4253 apply_change_group ();
4254 fold_rtx (SET_SRC (x), insn);
4255 invalidate (SET_DEST (x), VOIDmode);
4257 else
4258 n_sets = 1;
4260 else if (GET_CODE (x) == PARALLEL)
4262 int lim = XVECLEN (x, 0);
4264 sets = XALLOCAVEC (struct set, lim);
4266 /* Find all regs explicitly clobbered in this insn,
4267 and ensure they are not replaced with any other regs
4268 elsewhere in this insn.
4269 When a reg that is clobbered is also used for input,
4270 we should presume that that is for a reason,
4271 and we should not substitute some other register
4272 which is not supposed to be clobbered.
4273 Therefore, this loop cannot be merged into the one below
4274 because a CALL may precede a CLOBBER and refer to the
4275 value clobbered. We must not let a canonicalization do
4276 anything in that case. */
4277 for (i = 0; i < lim; i++)
4279 rtx y = XVECEXP (x, 0, i);
4280 if (GET_CODE (y) == CLOBBER)
4282 rtx clobbered = XEXP (y, 0);
4284 if (REG_P (clobbered)
4285 || GET_CODE (clobbered) == SUBREG)
4286 invalidate (clobbered, VOIDmode);
4287 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4288 || GET_CODE (clobbered) == ZERO_EXTRACT)
4289 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4293 for (i = 0; i < lim; i++)
4295 rtx y = XVECEXP (x, 0, i);
4296 if (GET_CODE (y) == SET)
4298 /* As above, we ignore unconditional jumps and call-insns and
4299 ignore the result of apply_change_group. */
4300 if (GET_CODE (SET_SRC (y)) == CALL)
4302 canon_reg (SET_SRC (y), insn);
4303 apply_change_group ();
4304 fold_rtx (SET_SRC (y), insn);
4305 invalidate (SET_DEST (y), VOIDmode);
4307 else if (SET_DEST (y) == pc_rtx
4308 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4310 else
4311 sets[n_sets++].rtl = y;
4313 else if (GET_CODE (y) == CLOBBER)
4315 /* If we clobber memory, canon the address.
4316 This does nothing when a register is clobbered
4317 because we have already invalidated the reg. */
4318 if (MEM_P (XEXP (y, 0)))
4319 canon_reg (XEXP (y, 0), insn);
4321 else if (GET_CODE (y) == USE
4322 && ! (REG_P (XEXP (y, 0))
4323 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4324 canon_reg (y, insn);
4325 else if (GET_CODE (y) == CALL)
4327 /* The result of apply_change_group can be ignored; see
4328 canon_reg. */
4329 canon_reg (y, insn);
4330 apply_change_group ();
4331 fold_rtx (y, insn);
4335 else if (GET_CODE (x) == CLOBBER)
4337 if (MEM_P (XEXP (x, 0)))
4338 canon_reg (XEXP (x, 0), insn);
4340 /* Canonicalize a USE of a pseudo register or memory location. */
4341 else if (GET_CODE (x) == USE
4342 && ! (REG_P (XEXP (x, 0))
4343 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4344 canon_reg (x, insn);
4345 else if (GET_CODE (x) == ASM_OPERANDS)
4347 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4349 rtx input = ASM_OPERANDS_INPUT (x, i);
4350 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4352 input = canon_reg (input, insn);
4353 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4357 else if (GET_CODE (x) == CALL)
4359 /* The result of apply_change_group can be ignored; see canon_reg. */
4360 canon_reg (x, insn);
4361 apply_change_group ();
4362 fold_rtx (x, insn);
4364 else if (DEBUG_INSN_P (insn))
4365 canon_reg (PATTERN (insn), insn);
4367 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4368 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4369 is handled specially for this case, and if it isn't set, then there will
4370 be no equivalence for the destination. */
4371 if (n_sets == 1 && REG_NOTES (insn) != 0
4372 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4373 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4374 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4376 /* The result of apply_change_group can be ignored; see canon_reg. */
4377 canon_reg (XEXP (tem, 0), insn);
4378 apply_change_group ();
4379 src_eqv = fold_rtx (XEXP (tem, 0), insn);
4380 XEXP (tem, 0) = copy_rtx (src_eqv);
4381 df_notes_rescan (insn);
4384 /* Canonicalize sources and addresses of destinations.
4385 We do this in a separate pass to avoid problems when a MATCH_DUP is
4386 present in the insn pattern. In that case, we want to ensure that
4387 we don't break the duplicate nature of the pattern. So we will replace
4388 both operands at the same time. Otherwise, we would fail to find an
4389 equivalent substitution in the loop calling validate_change below.
4391 We used to suppress canonicalization of DEST if it appears in SRC,
4392 but we don't do this any more. */
4394 for (i = 0; i < n_sets; i++)
4396 rtx dest = SET_DEST (sets[i].rtl);
4397 rtx src = SET_SRC (sets[i].rtl);
4398 rtx new_rtx = canon_reg (src, insn);
4400 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4402 if (GET_CODE (dest) == ZERO_EXTRACT)
4404 validate_change (insn, &XEXP (dest, 1),
4405 canon_reg (XEXP (dest, 1), insn), 1);
4406 validate_change (insn, &XEXP (dest, 2),
4407 canon_reg (XEXP (dest, 2), insn), 1);
4410 while (GET_CODE (dest) == SUBREG
4411 || GET_CODE (dest) == ZERO_EXTRACT
4412 || GET_CODE (dest) == STRICT_LOW_PART)
4413 dest = XEXP (dest, 0);
4415 if (MEM_P (dest))
4416 canon_reg (dest, insn);
4419 /* Now that we have done all the replacements, we can apply the change
4420 group and see if they all work. Note that this will cause some
4421 canonicalizations that would have worked individually not to be applied
4422 because some other canonicalization didn't work, but this should not
4423 occur often.
4425 The result of apply_change_group can be ignored; see canon_reg. */
4427 apply_change_group ();
4429 /* Set sets[i].src_elt to the class each source belongs to.
4430 Detect assignments from or to volatile things
4431 and set set[i] to zero so they will be ignored
4432 in the rest of this function.
4434 Nothing in this loop changes the hash table or the register chains. */
4436 for (i = 0; i < n_sets; i++)
4438 bool repeat = false;
4439 rtx src, dest;
4440 rtx src_folded;
4441 struct table_elt *elt = 0, *p;
4442 enum machine_mode mode;
4443 rtx src_eqv_here;
4444 rtx src_const = 0;
4445 rtx src_related = 0;
4446 bool src_related_is_const_anchor = false;
4447 struct table_elt *src_const_elt = 0;
4448 int src_cost = MAX_COST;
4449 int src_eqv_cost = MAX_COST;
4450 int src_folded_cost = MAX_COST;
4451 int src_related_cost = MAX_COST;
4452 int src_elt_cost = MAX_COST;
4453 int src_regcost = MAX_COST;
4454 int src_eqv_regcost = MAX_COST;
4455 int src_folded_regcost = MAX_COST;
4456 int src_related_regcost = MAX_COST;
4457 int src_elt_regcost = MAX_COST;
4458 /* Set nonzero if we need to call force_const_mem on with the
4459 contents of src_folded before using it. */
4460 int src_folded_force_flag = 0;
4462 dest = SET_DEST (sets[i].rtl);
4463 src = SET_SRC (sets[i].rtl);
4465 /* If SRC is a constant that has no machine mode,
4466 hash it with the destination's machine mode.
4467 This way we can keep different modes separate. */
4469 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4470 sets[i].mode = mode;
4472 if (src_eqv)
4474 enum machine_mode eqvmode = mode;
4475 if (GET_CODE (dest) == STRICT_LOW_PART)
4476 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4477 do_not_record = 0;
4478 hash_arg_in_memory = 0;
4479 src_eqv_hash = HASH (src_eqv, eqvmode);
4481 /* Find the equivalence class for the equivalent expression. */
4483 if (!do_not_record)
4484 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4486 src_eqv_volatile = do_not_record;
4487 src_eqv_in_memory = hash_arg_in_memory;
4490 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4491 value of the INNER register, not the destination. So it is not
4492 a valid substitution for the source. But save it for later. */
4493 if (GET_CODE (dest) == STRICT_LOW_PART)
4494 src_eqv_here = 0;
4495 else
4496 src_eqv_here = src_eqv;
4498 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4499 simplified result, which may not necessarily be valid. */
4500 src_folded = fold_rtx (src, insn);
4502 #if 0
4503 /* ??? This caused bad code to be generated for the m68k port with -O2.
4504 Suppose src is (CONST_INT -1), and that after truncation src_folded
4505 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4506 At the end we will add src and src_const to the same equivalence
4507 class. We now have 3 and -1 on the same equivalence class. This
4508 causes later instructions to be mis-optimized. */
4509 /* If storing a constant in a bitfield, pre-truncate the constant
4510 so we will be able to record it later. */
4511 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4513 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4515 if (CONST_INT_P (src)
4516 && CONST_INT_P (width)
4517 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4518 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4519 src_folded
4520 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4521 << INTVAL (width)) - 1));
4523 #endif
4525 /* Compute SRC's hash code, and also notice if it
4526 should not be recorded at all. In that case,
4527 prevent any further processing of this assignment. */
4528 do_not_record = 0;
4529 hash_arg_in_memory = 0;
4531 sets[i].src = src;
4532 sets[i].src_hash = HASH (src, mode);
4533 sets[i].src_volatile = do_not_record;
4534 sets[i].src_in_memory = hash_arg_in_memory;
4536 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4537 a pseudo, do not record SRC. Using SRC as a replacement for
4538 anything else will be incorrect in that situation. Note that
4539 this usually occurs only for stack slots, in which case all the
4540 RTL would be referring to SRC, so we don't lose any optimization
4541 opportunities by not having SRC in the hash table. */
4543 if (MEM_P (src)
4544 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4545 && REG_P (dest)
4546 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4547 sets[i].src_volatile = 1;
4549 #if 0
4550 /* It is no longer clear why we used to do this, but it doesn't
4551 appear to still be needed. So let's try without it since this
4552 code hurts cse'ing widened ops. */
4553 /* If source is a paradoxical subreg (such as QI treated as an SI),
4554 treat it as volatile. It may do the work of an SI in one context
4555 where the extra bits are not being used, but cannot replace an SI
4556 in general. */
4557 if (paradoxical_subreg_p (src))
4558 sets[i].src_volatile = 1;
4559 #endif
4561 /* Locate all possible equivalent forms for SRC. Try to replace
4562 SRC in the insn with each cheaper equivalent.
4564 We have the following types of equivalents: SRC itself, a folded
4565 version, a value given in a REG_EQUAL note, or a value related
4566 to a constant.
4568 Each of these equivalents may be part of an additional class
4569 of equivalents (if more than one is in the table, they must be in
4570 the same class; we check for this).
4572 If the source is volatile, we don't do any table lookups.
4574 We note any constant equivalent for possible later use in a
4575 REG_NOTE. */
4577 if (!sets[i].src_volatile)
4578 elt = lookup (src, sets[i].src_hash, mode);
4580 sets[i].src_elt = elt;
4582 if (elt && src_eqv_here && src_eqv_elt)
4584 if (elt->first_same_value != src_eqv_elt->first_same_value)
4586 /* The REG_EQUAL is indicating that two formerly distinct
4587 classes are now equivalent. So merge them. */
4588 merge_equiv_classes (elt, src_eqv_elt);
4589 src_eqv_hash = HASH (src_eqv, elt->mode);
4590 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4593 src_eqv_here = 0;
4596 else if (src_eqv_elt)
4597 elt = src_eqv_elt;
4599 /* Try to find a constant somewhere and record it in `src_const'.
4600 Record its table element, if any, in `src_const_elt'. Look in
4601 any known equivalences first. (If the constant is not in the
4602 table, also set `sets[i].src_const_hash'). */
4603 if (elt)
4604 for (p = elt->first_same_value; p; p = p->next_same_value)
4605 if (p->is_const)
4607 src_const = p->exp;
4608 src_const_elt = elt;
4609 break;
4612 if (src_const == 0
4613 && (CONSTANT_P (src_folded)
4614 /* Consider (minus (label_ref L1) (label_ref L2)) as
4615 "constant" here so we will record it. This allows us
4616 to fold switch statements when an ADDR_DIFF_VEC is used. */
4617 || (GET_CODE (src_folded) == MINUS
4618 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4619 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4620 src_const = src_folded, src_const_elt = elt;
4621 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4622 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4624 /* If we don't know if the constant is in the table, get its
4625 hash code and look it up. */
4626 if (src_const && src_const_elt == 0)
4628 sets[i].src_const_hash = HASH (src_const, mode);
4629 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4632 sets[i].src_const = src_const;
4633 sets[i].src_const_elt = src_const_elt;
4635 /* If the constant and our source are both in the table, mark them as
4636 equivalent. Otherwise, if a constant is in the table but the source
4637 isn't, set ELT to it. */
4638 if (src_const_elt && elt
4639 && src_const_elt->first_same_value != elt->first_same_value)
4640 merge_equiv_classes (elt, src_const_elt);
4641 else if (src_const_elt && elt == 0)
4642 elt = src_const_elt;
4644 /* See if there is a register linearly related to a constant
4645 equivalent of SRC. */
4646 if (src_const
4647 && (GET_CODE (src_const) == CONST
4648 || (src_const_elt && src_const_elt->related_value != 0)))
4650 src_related = use_related_value (src_const, src_const_elt);
4651 if (src_related)
4653 struct table_elt *src_related_elt
4654 = lookup (src_related, HASH (src_related, mode), mode);
4655 if (src_related_elt && elt)
4657 if (elt->first_same_value
4658 != src_related_elt->first_same_value)
4659 /* This can occur when we previously saw a CONST
4660 involving a SYMBOL_REF and then see the SYMBOL_REF
4661 twice. Merge the involved classes. */
4662 merge_equiv_classes (elt, src_related_elt);
4664 src_related = 0;
4665 src_related_elt = 0;
4667 else if (src_related_elt && elt == 0)
4668 elt = src_related_elt;
4672 /* See if we have a CONST_INT that is already in a register in a
4673 wider mode. */
4675 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4676 && GET_MODE_CLASS (mode) == MODE_INT
4677 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4679 enum machine_mode wider_mode;
4681 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4682 wider_mode != VOIDmode
4683 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4684 && src_related == 0;
4685 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4687 struct table_elt *const_elt
4688 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4690 if (const_elt == 0)
4691 continue;
4693 for (const_elt = const_elt->first_same_value;
4694 const_elt; const_elt = const_elt->next_same_value)
4695 if (REG_P (const_elt->exp))
4697 src_related = gen_lowpart (mode, const_elt->exp);
4698 break;
4703 /* Another possibility is that we have an AND with a constant in
4704 a mode narrower than a word. If so, it might have been generated
4705 as part of an "if" which would narrow the AND. If we already
4706 have done the AND in a wider mode, we can use a SUBREG of that
4707 value. */
4709 if (flag_expensive_optimizations && ! src_related
4710 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4711 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4713 enum machine_mode tmode;
4714 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4716 for (tmode = GET_MODE_WIDER_MODE (mode);
4717 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4718 tmode = GET_MODE_WIDER_MODE (tmode))
4720 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4721 struct table_elt *larger_elt;
4723 if (inner)
4725 PUT_MODE (new_and, tmode);
4726 XEXP (new_and, 0) = inner;
4727 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4728 if (larger_elt == 0)
4729 continue;
4731 for (larger_elt = larger_elt->first_same_value;
4732 larger_elt; larger_elt = larger_elt->next_same_value)
4733 if (REG_P (larger_elt->exp))
4735 src_related
4736 = gen_lowpart (mode, larger_elt->exp);
4737 break;
4740 if (src_related)
4741 break;
4746 #ifdef LOAD_EXTEND_OP
4747 /* See if a MEM has already been loaded with a widening operation;
4748 if it has, we can use a subreg of that. Many CISC machines
4749 also have such operations, but this is only likely to be
4750 beneficial on these machines. */
4752 if (flag_expensive_optimizations && src_related == 0
4753 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4754 && GET_MODE_CLASS (mode) == MODE_INT
4755 && MEM_P (src) && ! do_not_record
4756 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4758 struct rtx_def memory_extend_buf;
4759 rtx memory_extend_rtx = &memory_extend_buf;
4760 enum machine_mode tmode;
4762 /* Set what we are trying to extend and the operation it might
4763 have been extended with. */
4764 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4765 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4766 XEXP (memory_extend_rtx, 0) = src;
4768 for (tmode = GET_MODE_WIDER_MODE (mode);
4769 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4770 tmode = GET_MODE_WIDER_MODE (tmode))
4772 struct table_elt *larger_elt;
4774 PUT_MODE (memory_extend_rtx, tmode);
4775 larger_elt = lookup (memory_extend_rtx,
4776 HASH (memory_extend_rtx, tmode), tmode);
4777 if (larger_elt == 0)
4778 continue;
4780 for (larger_elt = larger_elt->first_same_value;
4781 larger_elt; larger_elt = larger_elt->next_same_value)
4782 if (REG_P (larger_elt->exp))
4784 src_related = gen_lowpart (mode, larger_elt->exp);
4785 break;
4788 if (src_related)
4789 break;
4792 #endif /* LOAD_EXTEND_OP */
4794 /* Try to express the constant using a register+offset expression
4795 derived from a constant anchor. */
4797 if (targetm.const_anchor
4798 && !src_related
4799 && src_const
4800 && GET_CODE (src_const) == CONST_INT)
4802 src_related = try_const_anchors (src_const, mode);
4803 src_related_is_const_anchor = src_related != NULL_RTX;
4807 if (src == src_folded)
4808 src_folded = 0;
4810 /* At this point, ELT, if nonzero, points to a class of expressions
4811 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4812 and SRC_RELATED, if nonzero, each contain additional equivalent
4813 expressions. Prune these latter expressions by deleting expressions
4814 already in the equivalence class.
4816 Check for an equivalent identical to the destination. If found,
4817 this is the preferred equivalent since it will likely lead to
4818 elimination of the insn. Indicate this by placing it in
4819 `src_related'. */
4821 if (elt)
4822 elt = elt->first_same_value;
4823 for (p = elt; p; p = p->next_same_value)
4825 enum rtx_code code = GET_CODE (p->exp);
4827 /* If the expression is not valid, ignore it. Then we do not
4828 have to check for validity below. In most cases, we can use
4829 `rtx_equal_p', since canonicalization has already been done. */
4830 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4831 continue;
4833 /* Also skip paradoxical subregs, unless that's what we're
4834 looking for. */
4835 if (paradoxical_subreg_p (p->exp)
4836 && ! (src != 0
4837 && GET_CODE (src) == SUBREG
4838 && GET_MODE (src) == GET_MODE (p->exp)
4839 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4840 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4841 continue;
4843 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4844 src = 0;
4845 else if (src_folded && GET_CODE (src_folded) == code
4846 && rtx_equal_p (src_folded, p->exp))
4847 src_folded = 0;
4848 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4849 && rtx_equal_p (src_eqv_here, p->exp))
4850 src_eqv_here = 0;
4851 else if (src_related && GET_CODE (src_related) == code
4852 && rtx_equal_p (src_related, p->exp))
4853 src_related = 0;
4855 /* This is the same as the destination of the insns, we want
4856 to prefer it. Copy it to src_related. The code below will
4857 then give it a negative cost. */
4858 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4859 src_related = dest;
4862 /* Find the cheapest valid equivalent, trying all the available
4863 possibilities. Prefer items not in the hash table to ones
4864 that are when they are equal cost. Note that we can never
4865 worsen an insn as the current contents will also succeed.
4866 If we find an equivalent identical to the destination, use it as best,
4867 since this insn will probably be eliminated in that case. */
4868 if (src)
4870 if (rtx_equal_p (src, dest))
4871 src_cost = src_regcost = -1;
4872 else
4874 src_cost = COST (src);
4875 src_regcost = approx_reg_cost (src);
4879 if (src_eqv_here)
4881 if (rtx_equal_p (src_eqv_here, dest))
4882 src_eqv_cost = src_eqv_regcost = -1;
4883 else
4885 src_eqv_cost = COST (src_eqv_here);
4886 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4890 if (src_folded)
4892 if (rtx_equal_p (src_folded, dest))
4893 src_folded_cost = src_folded_regcost = -1;
4894 else
4896 src_folded_cost = COST (src_folded);
4897 src_folded_regcost = approx_reg_cost (src_folded);
4901 if (src_related)
4903 if (rtx_equal_p (src_related, dest))
4904 src_related_cost = src_related_regcost = -1;
4905 else
4907 src_related_cost = COST (src_related);
4908 src_related_regcost = approx_reg_cost (src_related);
4910 /* If a const-anchor is used to synthesize a constant that
4911 normally requires multiple instructions then slightly prefer
4912 it over the original sequence. These instructions are likely
4913 to become redundant now. We can't compare against the cost
4914 of src_eqv_here because, on MIPS for example, multi-insn
4915 constants have zero cost; they are assumed to be hoisted from
4916 loops. */
4917 if (src_related_is_const_anchor
4918 && src_related_cost == src_cost
4919 && src_eqv_here)
4920 src_related_cost--;
4924 /* If this was an indirect jump insn, a known label will really be
4925 cheaper even though it looks more expensive. */
4926 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
4927 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
4929 /* Terminate loop when replacement made. This must terminate since
4930 the current contents will be tested and will always be valid. */
4931 while (1)
4933 rtx trial;
4935 /* Skip invalid entries. */
4936 while (elt && !REG_P (elt->exp)
4937 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
4938 elt = elt->next_same_value;
4940 /* A paradoxical subreg would be bad here: it'll be the right
4941 size, but later may be adjusted so that the upper bits aren't
4942 what we want. So reject it. */
4943 if (elt != 0
4944 && paradoxical_subreg_p (elt->exp)
4945 /* It is okay, though, if the rtx we're trying to match
4946 will ignore any of the bits we can't predict. */
4947 && ! (src != 0
4948 && GET_CODE (src) == SUBREG
4949 && GET_MODE (src) == GET_MODE (elt->exp)
4950 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4951 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
4953 elt = elt->next_same_value;
4954 continue;
4957 if (elt)
4959 src_elt_cost = elt->cost;
4960 src_elt_regcost = elt->regcost;
4963 /* Find cheapest and skip it for the next time. For items
4964 of equal cost, use this order:
4965 src_folded, src, src_eqv, src_related and hash table entry. */
4966 if (src_folded
4967 && preferable (src_folded_cost, src_folded_regcost,
4968 src_cost, src_regcost) <= 0
4969 && preferable (src_folded_cost, src_folded_regcost,
4970 src_eqv_cost, src_eqv_regcost) <= 0
4971 && preferable (src_folded_cost, src_folded_regcost,
4972 src_related_cost, src_related_regcost) <= 0
4973 && preferable (src_folded_cost, src_folded_regcost,
4974 src_elt_cost, src_elt_regcost) <= 0)
4976 trial = src_folded, src_folded_cost = MAX_COST;
4977 if (src_folded_force_flag)
4979 rtx forced = force_const_mem (mode, trial);
4980 if (forced)
4981 trial = forced;
4984 else if (src
4985 && preferable (src_cost, src_regcost,
4986 src_eqv_cost, src_eqv_regcost) <= 0
4987 && preferable (src_cost, src_regcost,
4988 src_related_cost, src_related_regcost) <= 0
4989 && preferable (src_cost, src_regcost,
4990 src_elt_cost, src_elt_regcost) <= 0)
4991 trial = src, src_cost = MAX_COST;
4992 else if (src_eqv_here
4993 && preferable (src_eqv_cost, src_eqv_regcost,
4994 src_related_cost, src_related_regcost) <= 0
4995 && preferable (src_eqv_cost, src_eqv_regcost,
4996 src_elt_cost, src_elt_regcost) <= 0)
4997 trial = src_eqv_here, src_eqv_cost = MAX_COST;
4998 else if (src_related
4999 && preferable (src_related_cost, src_related_regcost,
5000 src_elt_cost, src_elt_regcost) <= 0)
5001 trial = src_related, src_related_cost = MAX_COST;
5002 else
5004 trial = elt->exp;
5005 elt = elt->next_same_value;
5006 src_elt_cost = MAX_COST;
5009 /* Avoid creation of overlapping memory moves. */
5010 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5012 rtx src, dest;
5014 /* BLKmode moves are not handled by cse anyway. */
5015 if (GET_MODE (trial) == BLKmode)
5016 break;
5018 src = canon_rtx (trial);
5019 dest = canon_rtx (SET_DEST (sets[i].rtl));
5021 if (!MEM_P (src) || !MEM_P (dest)
5022 || !nonoverlapping_memrefs_p (src, dest, false))
5023 break;
5026 /* Try to optimize
5027 (set (reg:M N) (const_int A))
5028 (set (reg:M2 O) (const_int B))
5029 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5030 (reg:M2 O)). */
5031 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5032 && CONST_INT_P (trial)
5033 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5034 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5035 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5036 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5037 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5038 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5039 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5040 <= HOST_BITS_PER_WIDE_INT))
5042 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5043 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5044 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5045 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5046 struct table_elt *dest_elt
5047 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5048 rtx dest_cst = NULL;
5050 if (dest_elt)
5051 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5052 if (p->is_const && CONST_INT_P (p->exp))
5054 dest_cst = p->exp;
5055 break;
5057 if (dest_cst)
5059 HOST_WIDE_INT val = INTVAL (dest_cst);
5060 HOST_WIDE_INT mask;
5061 unsigned int shift;
5062 if (BITS_BIG_ENDIAN)
5063 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5064 - INTVAL (pos) - INTVAL (width);
5065 else
5066 shift = INTVAL (pos);
5067 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5068 mask = ~(HOST_WIDE_INT) 0;
5069 else
5070 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5071 val &= ~(mask << shift);
5072 val |= (INTVAL (trial) & mask) << shift;
5073 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5074 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5075 dest_reg, 1);
5076 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5077 GEN_INT (val), 1);
5078 if (apply_change_group ())
5080 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5081 if (note)
5083 remove_note (insn, note);
5084 df_notes_rescan (insn);
5086 src_eqv = NULL_RTX;
5087 src_eqv_elt = NULL;
5088 src_eqv_volatile = 0;
5089 src_eqv_in_memory = 0;
5090 src_eqv_hash = 0;
5091 repeat = true;
5092 break;
5097 /* We don't normally have an insn matching (set (pc) (pc)), so
5098 check for this separately here. We will delete such an
5099 insn below.
5101 For other cases such as a table jump or conditional jump
5102 where we know the ultimate target, go ahead and replace the
5103 operand. While that may not make a valid insn, we will
5104 reemit the jump below (and also insert any necessary
5105 barriers). */
5106 if (n_sets == 1 && dest == pc_rtx
5107 && (trial == pc_rtx
5108 || (GET_CODE (trial) == LABEL_REF
5109 && ! condjump_p (insn))))
5111 /* Don't substitute non-local labels, this confuses CFG. */
5112 if (GET_CODE (trial) == LABEL_REF
5113 && LABEL_REF_NONLOCAL_P (trial))
5114 continue;
5116 SET_SRC (sets[i].rtl) = trial;
5117 cse_jumps_altered = true;
5118 break;
5121 /* Reject certain invalid forms of CONST that we create. */
5122 else if (CONSTANT_P (trial)
5123 && GET_CODE (trial) == CONST
5124 /* Reject cases that will cause decode_rtx_const to
5125 die. On the alpha when simplifying a switch, we
5126 get (const (truncate (minus (label_ref)
5127 (label_ref)))). */
5128 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5129 /* Likewise on IA-64, except without the
5130 truncate. */
5131 || (GET_CODE (XEXP (trial, 0)) == MINUS
5132 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5133 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5134 /* Do nothing for this case. */
5137 /* Look for a substitution that makes a valid insn. */
5138 else if (validate_unshare_change
5139 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5141 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5143 /* The result of apply_change_group can be ignored; see
5144 canon_reg. */
5146 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5147 apply_change_group ();
5149 break;
5152 /* If we previously found constant pool entries for
5153 constants and this is a constant, try making a
5154 pool entry. Put it in src_folded unless we already have done
5155 this since that is where it likely came from. */
5157 else if (constant_pool_entries_cost
5158 && CONSTANT_P (trial)
5159 && (src_folded == 0
5160 || (!MEM_P (src_folded)
5161 && ! src_folded_force_flag))
5162 && GET_MODE_CLASS (mode) != MODE_CC
5163 && mode != VOIDmode)
5165 src_folded_force_flag = 1;
5166 src_folded = trial;
5167 src_folded_cost = constant_pool_entries_cost;
5168 src_folded_regcost = constant_pool_entries_regcost;
5172 /* If we changed the insn too much, handle this set from scratch. */
5173 if (repeat)
5175 i--;
5176 continue;
5179 src = SET_SRC (sets[i].rtl);
5181 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5182 However, there is an important exception: If both are registers
5183 that are not the head of their equivalence class, replace SET_SRC
5184 with the head of the class. If we do not do this, we will have
5185 both registers live over a portion of the basic block. This way,
5186 their lifetimes will likely abut instead of overlapping. */
5187 if (REG_P (dest)
5188 && REGNO_QTY_VALID_P (REGNO (dest)))
5190 int dest_q = REG_QTY (REGNO (dest));
5191 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5193 if (dest_ent->mode == GET_MODE (dest)
5194 && dest_ent->first_reg != REGNO (dest)
5195 && REG_P (src) && REGNO (src) == REGNO (dest)
5196 /* Don't do this if the original insn had a hard reg as
5197 SET_SRC or SET_DEST. */
5198 && (!REG_P (sets[i].src)
5199 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5200 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5201 /* We can't call canon_reg here because it won't do anything if
5202 SRC is a hard register. */
5204 int src_q = REG_QTY (REGNO (src));
5205 struct qty_table_elem *src_ent = &qty_table[src_q];
5206 int first = src_ent->first_reg;
5207 rtx new_src
5208 = (first >= FIRST_PSEUDO_REGISTER
5209 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5211 /* We must use validate-change even for this, because this
5212 might be a special no-op instruction, suitable only to
5213 tag notes onto. */
5214 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5216 src = new_src;
5217 /* If we had a constant that is cheaper than what we are now
5218 setting SRC to, use that constant. We ignored it when we
5219 thought we could make this into a no-op. */
5220 if (src_const && COST (src_const) < COST (src)
5221 && validate_change (insn, &SET_SRC (sets[i].rtl),
5222 src_const, 0))
5223 src = src_const;
5228 /* If we made a change, recompute SRC values. */
5229 if (src != sets[i].src)
5231 do_not_record = 0;
5232 hash_arg_in_memory = 0;
5233 sets[i].src = src;
5234 sets[i].src_hash = HASH (src, mode);
5235 sets[i].src_volatile = do_not_record;
5236 sets[i].src_in_memory = hash_arg_in_memory;
5237 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5240 /* If this is a single SET, we are setting a register, and we have an
5241 equivalent constant, we want to add a REG_NOTE. We don't want
5242 to write a REG_EQUAL note for a constant pseudo since verifying that
5243 that pseudo hasn't been eliminated is a pain. Such a note also
5244 won't help anything.
5246 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5247 which can be created for a reference to a compile time computable
5248 entry in a jump table. */
5250 if (n_sets == 1 && src_const && REG_P (dest)
5251 && !REG_P (src_const)
5252 && ! (GET_CODE (src_const) == CONST
5253 && GET_CODE (XEXP (src_const, 0)) == MINUS
5254 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5255 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5257 /* We only want a REG_EQUAL note if src_const != src. */
5258 if (! rtx_equal_p (src, src_const))
5260 /* Make sure that the rtx is not shared. */
5261 src_const = copy_rtx (src_const);
5263 /* Record the actual constant value in a REG_EQUAL note,
5264 making a new one if one does not already exist. */
5265 set_unique_reg_note (insn, REG_EQUAL, src_const);
5266 df_notes_rescan (insn);
5270 /* Now deal with the destination. */
5271 do_not_record = 0;
5273 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5274 while (GET_CODE (dest) == SUBREG
5275 || GET_CODE (dest) == ZERO_EXTRACT
5276 || GET_CODE (dest) == STRICT_LOW_PART)
5277 dest = XEXP (dest, 0);
5279 sets[i].inner_dest = dest;
5281 if (MEM_P (dest))
5283 #ifdef PUSH_ROUNDING
5284 /* Stack pushes invalidate the stack pointer. */
5285 rtx addr = XEXP (dest, 0);
5286 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5287 && XEXP (addr, 0) == stack_pointer_rtx)
5288 invalidate (stack_pointer_rtx, VOIDmode);
5289 #endif
5290 dest = fold_rtx (dest, insn);
5293 /* Compute the hash code of the destination now,
5294 before the effects of this instruction are recorded,
5295 since the register values used in the address computation
5296 are those before this instruction. */
5297 sets[i].dest_hash = HASH (dest, mode);
5299 /* Don't enter a bit-field in the hash table
5300 because the value in it after the store
5301 may not equal what was stored, due to truncation. */
5303 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5305 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5307 if (src_const != 0 && CONST_INT_P (src_const)
5308 && CONST_INT_P (width)
5309 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5310 && ! (INTVAL (src_const)
5311 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5312 /* Exception: if the value is constant,
5313 and it won't be truncated, record it. */
5315 else
5317 /* This is chosen so that the destination will be invalidated
5318 but no new value will be recorded.
5319 We must invalidate because sometimes constant
5320 values can be recorded for bitfields. */
5321 sets[i].src_elt = 0;
5322 sets[i].src_volatile = 1;
5323 src_eqv = 0;
5324 src_eqv_elt = 0;
5328 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5329 the insn. */
5330 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5332 /* One less use of the label this insn used to jump to. */
5333 delete_insn_and_edges (insn);
5334 cse_jumps_altered = true;
5335 /* No more processing for this set. */
5336 sets[i].rtl = 0;
5339 /* If this SET is now setting PC to a label, we know it used to
5340 be a conditional or computed branch. */
5341 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5342 && !LABEL_REF_NONLOCAL_P (src))
5344 /* We reemit the jump in as many cases as possible just in
5345 case the form of an unconditional jump is significantly
5346 different than a computed jump or conditional jump.
5348 If this insn has multiple sets, then reemitting the
5349 jump is nontrivial. So instead we just force rerecognition
5350 and hope for the best. */
5351 if (n_sets == 1)
5353 rtx new_rtx, note;
5355 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5356 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5357 LABEL_NUSES (XEXP (src, 0))++;
5359 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5360 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5361 if (note)
5363 XEXP (note, 1) = NULL_RTX;
5364 REG_NOTES (new_rtx) = note;
5367 delete_insn_and_edges (insn);
5368 insn = new_rtx;
5370 else
5371 INSN_CODE (insn) = -1;
5373 /* Do not bother deleting any unreachable code, let jump do it. */
5374 cse_jumps_altered = true;
5375 sets[i].rtl = 0;
5378 /* If destination is volatile, invalidate it and then do no further
5379 processing for this assignment. */
5381 else if (do_not_record)
5383 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5384 invalidate (dest, VOIDmode);
5385 else if (MEM_P (dest))
5386 invalidate (dest, VOIDmode);
5387 else if (GET_CODE (dest) == STRICT_LOW_PART
5388 || GET_CODE (dest) == ZERO_EXTRACT)
5389 invalidate (XEXP (dest, 0), GET_MODE (dest));
5390 sets[i].rtl = 0;
5393 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5394 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5396 #ifdef HAVE_cc0
5397 /* If setting CC0, record what it was set to, or a constant, if it
5398 is equivalent to a constant. If it is being set to a floating-point
5399 value, make a COMPARE with the appropriate constant of 0. If we
5400 don't do this, later code can interpret this as a test against
5401 const0_rtx, which can cause problems if we try to put it into an
5402 insn as a floating-point operand. */
5403 if (dest == cc0_rtx)
5405 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5406 this_insn_cc0_mode = mode;
5407 if (FLOAT_MODE_P (mode))
5408 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5409 CONST0_RTX (mode));
5411 #endif
5414 /* Now enter all non-volatile source expressions in the hash table
5415 if they are not already present.
5416 Record their equivalence classes in src_elt.
5417 This way we can insert the corresponding destinations into
5418 the same classes even if the actual sources are no longer in them
5419 (having been invalidated). */
5421 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5422 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5424 struct table_elt *elt;
5425 struct table_elt *classp = sets[0].src_elt;
5426 rtx dest = SET_DEST (sets[0].rtl);
5427 enum machine_mode eqvmode = GET_MODE (dest);
5429 if (GET_CODE (dest) == STRICT_LOW_PART)
5431 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5432 classp = 0;
5434 if (insert_regs (src_eqv, classp, 0))
5436 rehash_using_reg (src_eqv);
5437 src_eqv_hash = HASH (src_eqv, eqvmode);
5439 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5440 elt->in_memory = src_eqv_in_memory;
5441 src_eqv_elt = elt;
5443 /* Check to see if src_eqv_elt is the same as a set source which
5444 does not yet have an elt, and if so set the elt of the set source
5445 to src_eqv_elt. */
5446 for (i = 0; i < n_sets; i++)
5447 if (sets[i].rtl && sets[i].src_elt == 0
5448 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5449 sets[i].src_elt = src_eqv_elt;
5452 for (i = 0; i < n_sets; i++)
5453 if (sets[i].rtl && ! sets[i].src_volatile
5454 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5456 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5458 /* REG_EQUAL in setting a STRICT_LOW_PART
5459 gives an equivalent for the entire destination register,
5460 not just for the subreg being stored in now.
5461 This is a more interesting equivalence, so we arrange later
5462 to treat the entire reg as the destination. */
5463 sets[i].src_elt = src_eqv_elt;
5464 sets[i].src_hash = src_eqv_hash;
5466 else
5468 /* Insert source and constant equivalent into hash table, if not
5469 already present. */
5470 struct table_elt *classp = src_eqv_elt;
5471 rtx src = sets[i].src;
5472 rtx dest = SET_DEST (sets[i].rtl);
5473 enum machine_mode mode
5474 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5476 /* It's possible that we have a source value known to be
5477 constant but don't have a REG_EQUAL note on the insn.
5478 Lack of a note will mean src_eqv_elt will be NULL. This
5479 can happen where we've generated a SUBREG to access a
5480 CONST_INT that is already in a register in a wider mode.
5481 Ensure that the source expression is put in the proper
5482 constant class. */
5483 if (!classp)
5484 classp = sets[i].src_const_elt;
5486 if (sets[i].src_elt == 0)
5488 struct table_elt *elt;
5490 /* Note that these insert_regs calls cannot remove
5491 any of the src_elt's, because they would have failed to
5492 match if not still valid. */
5493 if (insert_regs (src, classp, 0))
5495 rehash_using_reg (src);
5496 sets[i].src_hash = HASH (src, mode);
5498 elt = insert (src, classp, sets[i].src_hash, mode);
5499 elt->in_memory = sets[i].src_in_memory;
5500 sets[i].src_elt = classp = elt;
5502 if (sets[i].src_const && sets[i].src_const_elt == 0
5503 && src != sets[i].src_const
5504 && ! rtx_equal_p (sets[i].src_const, src))
5505 sets[i].src_elt = insert (sets[i].src_const, classp,
5506 sets[i].src_const_hash, mode);
5509 else if (sets[i].src_elt == 0)
5510 /* If we did not insert the source into the hash table (e.g., it was
5511 volatile), note the equivalence class for the REG_EQUAL value, if any,
5512 so that the destination goes into that class. */
5513 sets[i].src_elt = src_eqv_elt;
5515 /* Record destination addresses in the hash table. This allows us to
5516 check if they are invalidated by other sets. */
5517 for (i = 0; i < n_sets; i++)
5519 if (sets[i].rtl)
5521 rtx x = sets[i].inner_dest;
5522 struct table_elt *elt;
5523 enum machine_mode mode;
5524 unsigned hash;
5526 if (MEM_P (x))
5528 x = XEXP (x, 0);
5529 mode = GET_MODE (x);
5530 hash = HASH (x, mode);
5531 elt = lookup (x, hash, mode);
5532 if (!elt)
5534 if (insert_regs (x, NULL, 0))
5536 rtx dest = SET_DEST (sets[i].rtl);
5538 rehash_using_reg (x);
5539 hash = HASH (x, mode);
5540 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5542 elt = insert (x, NULL, hash, mode);
5545 sets[i].dest_addr_elt = elt;
5547 else
5548 sets[i].dest_addr_elt = NULL;
5552 invalidate_from_clobbers (x);
5554 /* Some registers are invalidated by subroutine calls. Memory is
5555 invalidated by non-constant calls. */
5557 if (CALL_P (insn))
5559 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5560 invalidate_memory ();
5561 invalidate_for_call ();
5564 /* Now invalidate everything set by this instruction.
5565 If a SUBREG or other funny destination is being set,
5566 sets[i].rtl is still nonzero, so here we invalidate the reg
5567 a part of which is being set. */
5569 for (i = 0; i < n_sets; i++)
5570 if (sets[i].rtl)
5572 /* We can't use the inner dest, because the mode associated with
5573 a ZERO_EXTRACT is significant. */
5574 rtx dest = SET_DEST (sets[i].rtl);
5576 /* Needed for registers to remove the register from its
5577 previous quantity's chain.
5578 Needed for memory if this is a nonvarying address, unless
5579 we have just done an invalidate_memory that covers even those. */
5580 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5581 invalidate (dest, VOIDmode);
5582 else if (MEM_P (dest))
5583 invalidate (dest, VOIDmode);
5584 else if (GET_CODE (dest) == STRICT_LOW_PART
5585 || GET_CODE (dest) == ZERO_EXTRACT)
5586 invalidate (XEXP (dest, 0), GET_MODE (dest));
5589 /* A volatile ASM invalidates everything. */
5590 if (NONJUMP_INSN_P (insn)
5591 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5592 && MEM_VOLATILE_P (PATTERN (insn)))
5593 flush_hash_table ();
5595 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5596 the regs restored by the longjmp come from a later time
5597 than the setjmp. */
5598 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5600 flush_hash_table ();
5601 goto done;
5604 /* Make sure registers mentioned in destinations
5605 are safe for use in an expression to be inserted.
5606 This removes from the hash table
5607 any invalid entry that refers to one of these registers.
5609 We don't care about the return value from mention_regs because
5610 we are going to hash the SET_DEST values unconditionally. */
5612 for (i = 0; i < n_sets; i++)
5614 if (sets[i].rtl)
5616 rtx x = SET_DEST (sets[i].rtl);
5618 if (!REG_P (x))
5619 mention_regs (x);
5620 else
5622 /* We used to rely on all references to a register becoming
5623 inaccessible when a register changes to a new quantity,
5624 since that changes the hash code. However, that is not
5625 safe, since after HASH_SIZE new quantities we get a
5626 hash 'collision' of a register with its own invalid
5627 entries. And since SUBREGs have been changed not to
5628 change their hash code with the hash code of the register,
5629 it wouldn't work any longer at all. So we have to check
5630 for any invalid references lying around now.
5631 This code is similar to the REG case in mention_regs,
5632 but it knows that reg_tick has been incremented, and
5633 it leaves reg_in_table as -1 . */
5634 unsigned int regno = REGNO (x);
5635 unsigned int endregno = END_REGNO (x);
5636 unsigned int i;
5638 for (i = regno; i < endregno; i++)
5640 if (REG_IN_TABLE (i) >= 0)
5642 remove_invalid_refs (i);
5643 REG_IN_TABLE (i) = -1;
5650 /* We may have just removed some of the src_elt's from the hash table.
5651 So replace each one with the current head of the same class.
5652 Also check if destination addresses have been removed. */
5654 for (i = 0; i < n_sets; i++)
5655 if (sets[i].rtl)
5657 if (sets[i].dest_addr_elt
5658 && sets[i].dest_addr_elt->first_same_value == 0)
5660 /* The elt was removed, which means this destination is not
5661 valid after this instruction. */
5662 sets[i].rtl = NULL_RTX;
5664 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5665 /* If elt was removed, find current head of same class,
5666 or 0 if nothing remains of that class. */
5668 struct table_elt *elt = sets[i].src_elt;
5670 while (elt && elt->prev_same_value)
5671 elt = elt->prev_same_value;
5673 while (elt && elt->first_same_value == 0)
5674 elt = elt->next_same_value;
5675 sets[i].src_elt = elt ? elt->first_same_value : 0;
5679 /* Now insert the destinations into their equivalence classes. */
5681 for (i = 0; i < n_sets; i++)
5682 if (sets[i].rtl)
5684 rtx dest = SET_DEST (sets[i].rtl);
5685 struct table_elt *elt;
5687 /* Don't record value if we are not supposed to risk allocating
5688 floating-point values in registers that might be wider than
5689 memory. */
5690 if ((flag_float_store
5691 && MEM_P (dest)
5692 && FLOAT_MODE_P (GET_MODE (dest)))
5693 /* Don't record BLKmode values, because we don't know the
5694 size of it, and can't be sure that other BLKmode values
5695 have the same or smaller size. */
5696 || GET_MODE (dest) == BLKmode
5697 /* If we didn't put a REG_EQUAL value or a source into the hash
5698 table, there is no point is recording DEST. */
5699 || sets[i].src_elt == 0
5700 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5701 or SIGN_EXTEND, don't record DEST since it can cause
5702 some tracking to be wrong.
5704 ??? Think about this more later. */
5705 || (paradoxical_subreg_p (dest)
5706 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5707 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5708 continue;
5710 /* STRICT_LOW_PART isn't part of the value BEING set,
5711 and neither is the SUBREG inside it.
5712 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5713 if (GET_CODE (dest) == STRICT_LOW_PART)
5714 dest = SUBREG_REG (XEXP (dest, 0));
5716 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5717 /* Registers must also be inserted into chains for quantities. */
5718 if (insert_regs (dest, sets[i].src_elt, 1))
5720 /* If `insert_regs' changes something, the hash code must be
5721 recalculated. */
5722 rehash_using_reg (dest);
5723 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5726 elt = insert (dest, sets[i].src_elt,
5727 sets[i].dest_hash, GET_MODE (dest));
5729 /* If this is a constant, insert the constant anchors with the
5730 equivalent register-offset expressions using register DEST. */
5731 if (targetm.const_anchor
5732 && REG_P (dest)
5733 && SCALAR_INT_MODE_P (GET_MODE (dest))
5734 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5735 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5737 elt->in_memory = (MEM_P (sets[i].inner_dest)
5738 && !MEM_READONLY_P (sets[i].inner_dest));
5740 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5741 narrower than M2, and both M1 and M2 are the same number of words,
5742 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5743 make that equivalence as well.
5745 However, BAR may have equivalences for which gen_lowpart
5746 will produce a simpler value than gen_lowpart applied to
5747 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5748 BAR's equivalences. If we don't get a simplified form, make
5749 the SUBREG. It will not be used in an equivalence, but will
5750 cause two similar assignments to be detected.
5752 Note the loop below will find SUBREG_REG (DEST) since we have
5753 already entered SRC and DEST of the SET in the table. */
5755 if (GET_CODE (dest) == SUBREG
5756 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5757 / UNITS_PER_WORD)
5758 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5759 && (GET_MODE_SIZE (GET_MODE (dest))
5760 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5761 && sets[i].src_elt != 0)
5763 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5764 struct table_elt *elt, *classp = 0;
5766 for (elt = sets[i].src_elt->first_same_value; elt;
5767 elt = elt->next_same_value)
5769 rtx new_src = 0;
5770 unsigned src_hash;
5771 struct table_elt *src_elt;
5772 int byte = 0;
5774 /* Ignore invalid entries. */
5775 if (!REG_P (elt->exp)
5776 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5777 continue;
5779 /* We may have already been playing subreg games. If the
5780 mode is already correct for the destination, use it. */
5781 if (GET_MODE (elt->exp) == new_mode)
5782 new_src = elt->exp;
5783 else
5785 /* Calculate big endian correction for the SUBREG_BYTE.
5786 We have already checked that M1 (GET_MODE (dest))
5787 is not narrower than M2 (new_mode). */
5788 if (BYTES_BIG_ENDIAN)
5789 byte = (GET_MODE_SIZE (GET_MODE (dest))
5790 - GET_MODE_SIZE (new_mode));
5792 new_src = simplify_gen_subreg (new_mode, elt->exp,
5793 GET_MODE (dest), byte);
5796 /* The call to simplify_gen_subreg fails if the value
5797 is VOIDmode, yet we can't do any simplification, e.g.
5798 for EXPR_LISTs denoting function call results.
5799 It is invalid to construct a SUBREG with a VOIDmode
5800 SUBREG_REG, hence a zero new_src means we can't do
5801 this substitution. */
5802 if (! new_src)
5803 continue;
5805 src_hash = HASH (new_src, new_mode);
5806 src_elt = lookup (new_src, src_hash, new_mode);
5808 /* Put the new source in the hash table is if isn't
5809 already. */
5810 if (src_elt == 0)
5812 if (insert_regs (new_src, classp, 0))
5814 rehash_using_reg (new_src);
5815 src_hash = HASH (new_src, new_mode);
5817 src_elt = insert (new_src, classp, src_hash, new_mode);
5818 src_elt->in_memory = elt->in_memory;
5820 else if (classp && classp != src_elt->first_same_value)
5821 /* Show that two things that we've seen before are
5822 actually the same. */
5823 merge_equiv_classes (src_elt, classp);
5825 classp = src_elt->first_same_value;
5826 /* Ignore invalid entries. */
5827 while (classp
5828 && !REG_P (classp->exp)
5829 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5830 classp = classp->next_same_value;
5835 /* Special handling for (set REG0 REG1) where REG0 is the
5836 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5837 be used in the sequel, so (if easily done) change this insn to
5838 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5839 that computed their value. Then REG1 will become a dead store
5840 and won't cloud the situation for later optimizations.
5842 Do not make this change if REG1 is a hard register, because it will
5843 then be used in the sequel and we may be changing a two-operand insn
5844 into a three-operand insn.
5846 Also do not do this if we are operating on a copy of INSN. */
5848 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
5849 && NEXT_INSN (PREV_INSN (insn)) == insn
5850 && REG_P (SET_SRC (sets[0].rtl))
5851 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
5852 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
5854 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
5855 struct qty_table_elem *src_ent = &qty_table[src_q];
5857 if (src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
5859 /* Scan for the previous nonnote insn, but stop at a basic
5860 block boundary. */
5861 rtx prev = insn;
5862 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
5865 prev = PREV_INSN (prev);
5867 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
5869 /* Do not swap the registers around if the previous instruction
5870 attaches a REG_EQUIV note to REG1.
5872 ??? It's not entirely clear whether we can transfer a REG_EQUIV
5873 from the pseudo that originally shadowed an incoming argument
5874 to another register. Some uses of REG_EQUIV might rely on it
5875 being attached to REG1 rather than REG2.
5877 This section previously turned the REG_EQUIV into a REG_EQUAL
5878 note. We cannot do that because REG_EQUIV may provide an
5879 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
5880 if (NONJUMP_INSN_P (prev)
5881 && GET_CODE (PATTERN (prev)) == SET
5882 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
5883 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
5885 rtx dest = SET_DEST (sets[0].rtl);
5886 rtx src = SET_SRC (sets[0].rtl);
5887 rtx note;
5889 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
5890 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
5891 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
5892 apply_change_group ();
5894 /* If INSN has a REG_EQUAL note, and this note mentions
5895 REG0, then we must delete it, because the value in
5896 REG0 has changed. If the note's value is REG1, we must
5897 also delete it because that is now this insn's dest. */
5898 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5899 if (note != 0
5900 && (reg_mentioned_p (dest, XEXP (note, 0))
5901 || rtx_equal_p (src, XEXP (note, 0))))
5902 remove_note (insn, note);
5907 done:;
5910 /* Remove from the hash table all expressions that reference memory. */
5912 static void
5913 invalidate_memory (void)
5915 int i;
5916 struct table_elt *p, *next;
5918 for (i = 0; i < HASH_SIZE; i++)
5919 for (p = table[i]; p; p = next)
5921 next = p->next_same_hash;
5922 if (p->in_memory)
5923 remove_from_table (p, i);
5927 /* Perform invalidation on the basis of everything about an insn
5928 except for invalidating the actual places that are SET in it.
5929 This includes the places CLOBBERed, and anything that might
5930 alias with something that is SET or CLOBBERed.
5932 X is the pattern of the insn. */
5934 static void
5935 invalidate_from_clobbers (rtx x)
5937 if (GET_CODE (x) == CLOBBER)
5939 rtx ref = XEXP (x, 0);
5940 if (ref)
5942 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5943 || MEM_P (ref))
5944 invalidate (ref, VOIDmode);
5945 else if (GET_CODE (ref) == STRICT_LOW_PART
5946 || GET_CODE (ref) == ZERO_EXTRACT)
5947 invalidate (XEXP (ref, 0), GET_MODE (ref));
5950 else if (GET_CODE (x) == PARALLEL)
5952 int i;
5953 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5955 rtx y = XVECEXP (x, 0, i);
5956 if (GET_CODE (y) == CLOBBER)
5958 rtx ref = XEXP (y, 0);
5959 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5960 || MEM_P (ref))
5961 invalidate (ref, VOIDmode);
5962 else if (GET_CODE (ref) == STRICT_LOW_PART
5963 || GET_CODE (ref) == ZERO_EXTRACT)
5964 invalidate (XEXP (ref, 0), GET_MODE (ref));
5970 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
5971 and replace any registers in them with either an equivalent constant
5972 or the canonical form of the register. If we are inside an address,
5973 only do this if the address remains valid.
5975 OBJECT is 0 except when within a MEM in which case it is the MEM.
5977 Return the replacement for X. */
5979 static rtx
5980 cse_process_notes_1 (rtx x, rtx object, bool *changed)
5982 enum rtx_code code = GET_CODE (x);
5983 const char *fmt = GET_RTX_FORMAT (code);
5984 int i;
5986 switch (code)
5988 case CONST_INT:
5989 case CONST:
5990 case SYMBOL_REF:
5991 case LABEL_REF:
5992 case CONST_DOUBLE:
5993 case CONST_FIXED:
5994 case CONST_VECTOR:
5995 case PC:
5996 case CC0:
5997 case LO_SUM:
5998 return x;
6000 case MEM:
6001 validate_change (x, &XEXP (x, 0),
6002 cse_process_notes (XEXP (x, 0), x, changed), 0);
6003 return x;
6005 case EXPR_LIST:
6006 case INSN_LIST:
6007 if (REG_NOTE_KIND (x) == REG_EQUAL)
6008 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6009 if (XEXP (x, 1))
6010 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6011 return x;
6013 case SIGN_EXTEND:
6014 case ZERO_EXTEND:
6015 case SUBREG:
6017 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6018 /* We don't substitute VOIDmode constants into these rtx,
6019 since they would impede folding. */
6020 if (GET_MODE (new_rtx) != VOIDmode)
6021 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6022 return x;
6025 case REG:
6026 i = REG_QTY (REGNO (x));
6028 /* Return a constant or a constant register. */
6029 if (REGNO_QTY_VALID_P (REGNO (x)))
6031 struct qty_table_elem *ent = &qty_table[i];
6033 if (ent->const_rtx != NULL_RTX
6034 && (CONSTANT_P (ent->const_rtx)
6035 || REG_P (ent->const_rtx)))
6037 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6038 if (new_rtx)
6039 return copy_rtx (new_rtx);
6043 /* Otherwise, canonicalize this register. */
6044 return canon_reg (x, NULL_RTX);
6046 default:
6047 break;
6050 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6051 if (fmt[i] == 'e')
6052 validate_change (object, &XEXP (x, i),
6053 cse_process_notes (XEXP (x, i), object, changed), 0);
6055 return x;
6058 static rtx
6059 cse_process_notes (rtx x, rtx object, bool *changed)
6061 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6062 if (new_rtx != x)
6063 *changed = true;
6064 return new_rtx;
6068 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6070 DATA is a pointer to a struct cse_basic_block_data, that is used to
6071 describe the path.
6072 It is filled with a queue of basic blocks, starting with FIRST_BB
6073 and following a trace through the CFG.
6075 If all paths starting at FIRST_BB have been followed, or no new path
6076 starting at FIRST_BB can be constructed, this function returns FALSE.
6077 Otherwise, DATA->path is filled and the function returns TRUE indicating
6078 that a path to follow was found.
6080 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6081 block in the path will be FIRST_BB. */
6083 static bool
6084 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6085 int follow_jumps)
6087 basic_block bb;
6088 edge e;
6089 int path_size;
6091 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6093 /* See if there is a previous path. */
6094 path_size = data->path_size;
6096 /* There is a previous path. Make sure it started with FIRST_BB. */
6097 if (path_size)
6098 gcc_assert (data->path[0].bb == first_bb);
6100 /* There was only one basic block in the last path. Clear the path and
6101 return, so that paths starting at another basic block can be tried. */
6102 if (path_size == 1)
6104 path_size = 0;
6105 goto done;
6108 /* If the path was empty from the beginning, construct a new path. */
6109 if (path_size == 0)
6110 data->path[path_size++].bb = first_bb;
6111 else
6113 /* Otherwise, path_size must be equal to or greater than 2, because
6114 a previous path exists that is at least two basic blocks long.
6116 Update the previous branch path, if any. If the last branch was
6117 previously along the branch edge, take the fallthrough edge now. */
6118 while (path_size >= 2)
6120 basic_block last_bb_in_path, previous_bb_in_path;
6121 edge e;
6123 --path_size;
6124 last_bb_in_path = data->path[path_size].bb;
6125 previous_bb_in_path = data->path[path_size - 1].bb;
6127 /* If we previously followed a path along the branch edge, try
6128 the fallthru edge now. */
6129 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6130 && any_condjump_p (BB_END (previous_bb_in_path))
6131 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6132 && e == BRANCH_EDGE (previous_bb_in_path))
6134 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6135 if (bb != EXIT_BLOCK_PTR
6136 && single_pred_p (bb)
6137 /* We used to assert here that we would only see blocks
6138 that we have not visited yet. But we may end up
6139 visiting basic blocks twice if the CFG has changed
6140 in this run of cse_main, because when the CFG changes
6141 the topological sort of the CFG also changes. A basic
6142 blocks that previously had more than two predecessors
6143 may now have a single predecessor, and become part of
6144 a path that starts at another basic block.
6146 We still want to visit each basic block only once, so
6147 halt the path here if we have already visited BB. */
6148 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6150 SET_BIT (cse_visited_basic_blocks, bb->index);
6151 data->path[path_size++].bb = bb;
6152 break;
6156 data->path[path_size].bb = NULL;
6159 /* If only one block remains in the path, bail. */
6160 if (path_size == 1)
6162 path_size = 0;
6163 goto done;
6167 /* Extend the path if possible. */
6168 if (follow_jumps)
6170 bb = data->path[path_size - 1].bb;
6171 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6173 if (single_succ_p (bb))
6174 e = single_succ_edge (bb);
6175 else if (EDGE_COUNT (bb->succs) == 2
6176 && any_condjump_p (BB_END (bb)))
6178 /* First try to follow the branch. If that doesn't lead
6179 to a useful path, follow the fallthru edge. */
6180 e = BRANCH_EDGE (bb);
6181 if (!single_pred_p (e->dest))
6182 e = FALLTHRU_EDGE (bb);
6184 else
6185 e = NULL;
6187 if (e
6188 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6189 && e->dest != EXIT_BLOCK_PTR
6190 && single_pred_p (e->dest)
6191 /* Avoid visiting basic blocks twice. The large comment
6192 above explains why this can happen. */
6193 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6195 basic_block bb2 = e->dest;
6196 SET_BIT (cse_visited_basic_blocks, bb2->index);
6197 data->path[path_size++].bb = bb2;
6198 bb = bb2;
6200 else
6201 bb = NULL;
6205 done:
6206 data->path_size = path_size;
6207 return path_size != 0;
6210 /* Dump the path in DATA to file F. NSETS is the number of sets
6211 in the path. */
6213 static void
6214 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6216 int path_entry;
6218 fprintf (f, ";; Following path with %d sets: ", nsets);
6219 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6220 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6221 fputc ('\n', dump_file);
6222 fflush (f);
6226 /* Return true if BB has exception handling successor edges. */
6228 static bool
6229 have_eh_succ_edges (basic_block bb)
6231 edge e;
6232 edge_iterator ei;
6234 FOR_EACH_EDGE (e, ei, bb->succs)
6235 if (e->flags & EDGE_EH)
6236 return true;
6238 return false;
6242 /* Scan to the end of the path described by DATA. Return an estimate of
6243 the total number of SETs of all insns in the path. */
6245 static void
6246 cse_prescan_path (struct cse_basic_block_data *data)
6248 int nsets = 0;
6249 int path_size = data->path_size;
6250 int path_entry;
6252 /* Scan to end of each basic block in the path. */
6253 for (path_entry = 0; path_entry < path_size; path_entry++)
6255 basic_block bb;
6256 rtx insn;
6258 bb = data->path[path_entry].bb;
6260 FOR_BB_INSNS (bb, insn)
6262 if (!INSN_P (insn))
6263 continue;
6265 /* A PARALLEL can have lots of SETs in it,
6266 especially if it is really an ASM_OPERANDS. */
6267 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6268 nsets += XVECLEN (PATTERN (insn), 0);
6269 else
6270 nsets += 1;
6274 data->nsets = nsets;
6277 /* Process a single extended basic block described by EBB_DATA. */
6279 static void
6280 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6282 int path_size = ebb_data->path_size;
6283 int path_entry;
6284 int num_insns = 0;
6286 /* Allocate the space needed by qty_table. */
6287 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6289 new_basic_block ();
6290 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6291 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6292 for (path_entry = 0; path_entry < path_size; path_entry++)
6294 basic_block bb;
6295 rtx insn;
6297 bb = ebb_data->path[path_entry].bb;
6299 /* Invalidate recorded information for eh regs if there is an EH
6300 edge pointing to that bb. */
6301 if (bb_has_eh_pred (bb))
6303 df_ref *def_rec;
6305 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6307 df_ref def = *def_rec;
6308 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6309 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6313 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6314 FOR_BB_INSNS (bb, insn)
6316 /* If we have processed 1,000 insns, flush the hash table to
6317 avoid extreme quadratic behavior. We must not include NOTEs
6318 in the count since there may be more of them when generating
6319 debugging information. If we clear the table at different
6320 times, code generated with -g -O might be different than code
6321 generated with -O but not -g.
6323 FIXME: This is a real kludge and needs to be done some other
6324 way. */
6325 if (NONDEBUG_INSN_P (insn)
6326 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6328 flush_hash_table ();
6329 num_insns = 0;
6332 if (INSN_P (insn))
6334 /* Process notes first so we have all notes in canonical forms
6335 when looking for duplicate operations. */
6336 if (REG_NOTES (insn))
6338 bool changed = false;
6339 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6340 NULL_RTX, &changed);
6341 if (changed)
6342 df_notes_rescan (insn);
6345 cse_insn (insn);
6347 /* If we haven't already found an insn where we added a LABEL_REF,
6348 check this one. */
6349 if (INSN_P (insn) && !recorded_label_ref
6350 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6351 (void *) insn))
6352 recorded_label_ref = true;
6354 #ifdef HAVE_cc0
6355 if (NONDEBUG_INSN_P (insn))
6357 /* If the previous insn sets CC0 and this insn no
6358 longer references CC0, delete the previous insn.
6359 Here we use fact that nothing expects CC0 to be
6360 valid over an insn, which is true until the final
6361 pass. */
6362 rtx prev_insn, tem;
6364 prev_insn = prev_nonnote_nondebug_insn (insn);
6365 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6366 && (tem = single_set (prev_insn)) != NULL_RTX
6367 && SET_DEST (tem) == cc0_rtx
6368 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6369 delete_insn (prev_insn);
6371 /* If this insn is not the last insn in the basic
6372 block, it will be PREV_INSN(insn) in the next
6373 iteration. If we recorded any CC0-related
6374 information for this insn, remember it. */
6375 if (insn != BB_END (bb))
6377 prev_insn_cc0 = this_insn_cc0;
6378 prev_insn_cc0_mode = this_insn_cc0_mode;
6381 #endif
6385 /* With non-call exceptions, we are not always able to update
6386 the CFG properly inside cse_insn. So clean up possibly
6387 redundant EH edges here. */
6388 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6389 cse_cfg_altered |= purge_dead_edges (bb);
6391 /* If we changed a conditional jump, we may have terminated
6392 the path we are following. Check that by verifying that
6393 the edge we would take still exists. If the edge does
6394 not exist anymore, purge the remainder of the path.
6395 Note that this will cause us to return to the caller. */
6396 if (path_entry < path_size - 1)
6398 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6399 if (!find_edge (bb, next_bb))
6403 path_size--;
6405 /* If we truncate the path, we must also reset the
6406 visited bit on the remaining blocks in the path,
6407 or we will never visit them at all. */
6408 RESET_BIT (cse_visited_basic_blocks,
6409 ebb_data->path[path_size].bb->index);
6410 ebb_data->path[path_size].bb = NULL;
6412 while (path_size - 1 != path_entry);
6413 ebb_data->path_size = path_size;
6417 /* If this is a conditional jump insn, record any known
6418 equivalences due to the condition being tested. */
6419 insn = BB_END (bb);
6420 if (path_entry < path_size - 1
6421 && JUMP_P (insn)
6422 && single_set (insn)
6423 && any_condjump_p (insn))
6425 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6426 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6427 record_jump_equiv (insn, taken);
6430 #ifdef HAVE_cc0
6431 /* Clear the CC0-tracking related insns, they can't provide
6432 useful information across basic block boundaries. */
6433 prev_insn_cc0 = 0;
6434 #endif
6437 gcc_assert (next_qty <= max_qty);
6439 free (qty_table);
6443 /* Perform cse on the instructions of a function.
6444 F is the first instruction.
6445 NREGS is one plus the highest pseudo-reg number used in the instruction.
6447 Return 2 if jump optimizations should be redone due to simplifications
6448 in conditional jump instructions.
6449 Return 1 if the CFG should be cleaned up because it has been modified.
6450 Return 0 otherwise. */
6453 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6455 struct cse_basic_block_data ebb_data;
6456 basic_block bb;
6457 int *rc_order = XNEWVEC (int, last_basic_block);
6458 int i, n_blocks;
6460 df_set_flags (DF_LR_RUN_DCE);
6461 df_analyze ();
6462 df_set_flags (DF_DEFER_INSN_RESCAN);
6464 reg_scan (get_insns (), max_reg_num ());
6465 init_cse_reg_info (nregs);
6467 ebb_data.path = XNEWVEC (struct branch_path,
6468 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6470 cse_cfg_altered = false;
6471 cse_jumps_altered = false;
6472 recorded_label_ref = false;
6473 constant_pool_entries_cost = 0;
6474 constant_pool_entries_regcost = 0;
6475 ebb_data.path_size = 0;
6476 ebb_data.nsets = 0;
6477 rtl_hooks = cse_rtl_hooks;
6479 init_recog ();
6480 init_alias_analysis ();
6482 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6484 /* Set up the table of already visited basic blocks. */
6485 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6486 sbitmap_zero (cse_visited_basic_blocks);
6488 /* Loop over basic blocks in reverse completion order (RPO),
6489 excluding the ENTRY and EXIT blocks. */
6490 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6491 i = 0;
6492 while (i < n_blocks)
6494 /* Find the first block in the RPO queue that we have not yet
6495 processed before. */
6498 bb = BASIC_BLOCK (rc_order[i++]);
6500 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6501 && i < n_blocks);
6503 /* Find all paths starting with BB, and process them. */
6504 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6506 /* Pre-scan the path. */
6507 cse_prescan_path (&ebb_data);
6509 /* If this basic block has no sets, skip it. */
6510 if (ebb_data.nsets == 0)
6511 continue;
6513 /* Get a reasonable estimate for the maximum number of qty's
6514 needed for this path. For this, we take the number of sets
6515 and multiply that by MAX_RECOG_OPERANDS. */
6516 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6518 /* Dump the path we're about to process. */
6519 if (dump_file)
6520 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6522 cse_extended_basic_block (&ebb_data);
6526 /* Clean up. */
6527 end_alias_analysis ();
6528 free (reg_eqv_table);
6529 free (ebb_data.path);
6530 sbitmap_free (cse_visited_basic_blocks);
6531 free (rc_order);
6532 rtl_hooks = general_rtl_hooks;
6534 if (cse_jumps_altered || recorded_label_ref)
6535 return 2;
6536 else if (cse_cfg_altered)
6537 return 1;
6538 else
6539 return 0;
6542 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6543 which there isn't a REG_LABEL_OPERAND note.
6544 Return one if so. DATA is the insn. */
6546 static int
6547 check_for_label_ref (rtx *rtl, void *data)
6549 rtx insn = (rtx) data;
6551 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6552 note for it, we must rerun jump since it needs to place the note. If
6553 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6554 don't do this since no REG_LABEL_OPERAND will be added. */
6555 return (GET_CODE (*rtl) == LABEL_REF
6556 && ! LABEL_REF_NONLOCAL_P (*rtl)
6557 && (!JUMP_P (insn)
6558 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6559 && LABEL_P (XEXP (*rtl, 0))
6560 && INSN_UID (XEXP (*rtl, 0)) != 0
6561 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6564 /* Count the number of times registers are used (not set) in X.
6565 COUNTS is an array in which we accumulate the count, INCR is how much
6566 we count each register usage.
6568 Don't count a usage of DEST, which is the SET_DEST of a SET which
6569 contains X in its SET_SRC. This is because such a SET does not
6570 modify the liveness of DEST.
6571 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6572 We must then count uses of a SET_DEST regardless, because the insn can't be
6573 deleted here. */
6575 static void
6576 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6578 enum rtx_code code;
6579 rtx note;
6580 const char *fmt;
6581 int i, j;
6583 if (x == 0)
6584 return;
6586 switch (code = GET_CODE (x))
6588 case REG:
6589 if (x != dest)
6590 counts[REGNO (x)] += incr;
6591 return;
6593 case PC:
6594 case CC0:
6595 case CONST:
6596 case CONST_INT:
6597 case CONST_DOUBLE:
6598 case CONST_FIXED:
6599 case CONST_VECTOR:
6600 case SYMBOL_REF:
6601 case LABEL_REF:
6602 return;
6604 case CLOBBER:
6605 /* If we are clobbering a MEM, mark any registers inside the address
6606 as being used. */
6607 if (MEM_P (XEXP (x, 0)))
6608 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6609 return;
6611 case SET:
6612 /* Unless we are setting a REG, count everything in SET_DEST. */
6613 if (!REG_P (SET_DEST (x)))
6614 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6615 count_reg_usage (SET_SRC (x), counts,
6616 dest ? dest : SET_DEST (x),
6617 incr);
6618 return;
6620 case DEBUG_INSN:
6621 return;
6623 case CALL_INSN:
6624 case INSN:
6625 case JUMP_INSN:
6626 /* We expect dest to be NULL_RTX here. If the insn may trap,
6627 or if it cannot be deleted due to side-effects, mark this fact
6628 by setting DEST to pc_rtx. */
6629 if (insn_could_throw_p (x) || side_effects_p (PATTERN (x)))
6630 dest = pc_rtx;
6631 if (code == CALL_INSN)
6632 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6633 count_reg_usage (PATTERN (x), counts, dest, incr);
6635 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6636 use them. */
6638 note = find_reg_equal_equiv_note (x);
6639 if (note)
6641 rtx eqv = XEXP (note, 0);
6643 if (GET_CODE (eqv) == EXPR_LIST)
6644 /* This REG_EQUAL note describes the result of a function call.
6645 Process all the arguments. */
6648 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6649 eqv = XEXP (eqv, 1);
6651 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6652 else
6653 count_reg_usage (eqv, counts, dest, incr);
6655 return;
6657 case EXPR_LIST:
6658 if (REG_NOTE_KIND (x) == REG_EQUAL
6659 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6660 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6661 involving registers in the address. */
6662 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6663 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6665 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6666 return;
6668 case ASM_OPERANDS:
6669 /* Iterate over just the inputs, not the constraints as well. */
6670 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6671 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6672 return;
6674 case INSN_LIST:
6675 gcc_unreachable ();
6677 default:
6678 break;
6681 fmt = GET_RTX_FORMAT (code);
6682 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6684 if (fmt[i] == 'e')
6685 count_reg_usage (XEXP (x, i), counts, dest, incr);
6686 else if (fmt[i] == 'E')
6687 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6688 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6692 /* Return true if X is a dead register. */
6694 static inline int
6695 is_dead_reg (rtx x, int *counts)
6697 return (REG_P (x)
6698 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6699 && counts[REGNO (x)] == 0);
6702 /* Return true if set is live. */
6703 static bool
6704 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6705 int *counts)
6707 #ifdef HAVE_cc0
6708 rtx tem;
6709 #endif
6711 if (set_noop_p (set))
6714 #ifdef HAVE_cc0
6715 else if (GET_CODE (SET_DEST (set)) == CC0
6716 && !side_effects_p (SET_SRC (set))
6717 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6718 || !INSN_P (tem)
6719 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6720 return false;
6721 #endif
6722 else if (!is_dead_reg (SET_DEST (set), counts)
6723 || side_effects_p (SET_SRC (set)))
6724 return true;
6725 return false;
6728 /* Return true if insn is live. */
6730 static bool
6731 insn_live_p (rtx insn, int *counts)
6733 int i;
6734 if (insn_could_throw_p (insn))
6735 return true;
6736 else if (GET_CODE (PATTERN (insn)) == SET)
6737 return set_live_p (PATTERN (insn), insn, counts);
6738 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6740 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6742 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6744 if (GET_CODE (elt) == SET)
6746 if (set_live_p (elt, insn, counts))
6747 return true;
6749 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6750 return true;
6752 return false;
6754 else if (DEBUG_INSN_P (insn))
6756 rtx next;
6758 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6759 if (NOTE_P (next))
6760 continue;
6761 else if (!DEBUG_INSN_P (next))
6762 return true;
6763 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6764 return false;
6766 return true;
6768 else
6769 return true;
6772 /* Count the number of stores into pseudo. Callback for note_stores. */
6774 static void
6775 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6777 int *counts = (int *) data;
6778 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6779 counts[REGNO (x)]++;
6782 struct dead_debug_insn_data
6784 int *counts;
6785 rtx *replacements;
6786 bool seen_repl;
6789 /* Return if a DEBUG_INSN needs to be reset because some dead
6790 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6792 static int
6793 is_dead_debug_insn (rtx *loc, void *data)
6795 rtx x = *loc;
6796 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6798 if (is_dead_reg (x, ddid->counts))
6800 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6801 ddid->seen_repl = true;
6802 else
6803 return 1;
6805 return 0;
6808 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6809 Callback for simplify_replace_fn_rtx. */
6811 static rtx
6812 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6814 rtx *replacements = (rtx *) data;
6816 if (REG_P (x)
6817 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6818 && replacements[REGNO (x)] != NULL_RTX)
6820 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6821 return replacements[REGNO (x)];
6822 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6823 GET_MODE (replacements[REGNO (x)]));
6825 return NULL_RTX;
6828 /* Scan all the insns and delete any that are dead; i.e., they store a register
6829 that is never used or they copy a register to itself.
6831 This is used to remove insns made obviously dead by cse, loop or other
6832 optimizations. It improves the heuristics in loop since it won't try to
6833 move dead invariants out of loops or make givs for dead quantities. The
6834 remaining passes of the compilation are also sped up. */
6837 delete_trivially_dead_insns (rtx insns, int nreg)
6839 int *counts;
6840 rtx insn, prev;
6841 rtx *replacements = NULL;
6842 int ndead = 0;
6844 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6845 /* First count the number of times each register is used. */
6846 if (MAY_HAVE_DEBUG_INSNS)
6848 counts = XCNEWVEC (int, nreg * 3);
6849 for (insn = insns; insn; insn = NEXT_INSN (insn))
6850 if (DEBUG_INSN_P (insn))
6851 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6852 NULL_RTX, 1);
6853 else if (INSN_P (insn))
6855 count_reg_usage (insn, counts, NULL_RTX, 1);
6856 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6858 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6859 First one counts how many times each pseudo is used outside
6860 of debug insns, second counts how many times each pseudo is
6861 used in debug insns and third counts how many times a pseudo
6862 is stored. */
6864 else
6866 counts = XCNEWVEC (int, nreg);
6867 for (insn = insns; insn; insn = NEXT_INSN (insn))
6868 if (INSN_P (insn))
6869 count_reg_usage (insn, counts, NULL_RTX, 1);
6870 /* If no debug insns can be present, COUNTS is just an array
6871 which counts how many times each pseudo is used. */
6873 /* Go from the last insn to the first and delete insns that only set unused
6874 registers or copy a register to itself. As we delete an insn, remove
6875 usage counts for registers it uses.
6877 The first jump optimization pass may leave a real insn as the last
6878 insn in the function. We must not skip that insn or we may end
6879 up deleting code that is not really dead.
6881 If some otherwise unused register is only used in DEBUG_INSNs,
6882 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6883 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6884 has been created for the unused register, replace it with
6885 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6886 for (insn = get_last_insn (); insn; insn = prev)
6888 int live_insn = 0;
6890 prev = PREV_INSN (insn);
6891 if (!INSN_P (insn))
6892 continue;
6894 live_insn = insn_live_p (insn, counts);
6896 /* If this is a dead insn, delete it and show registers in it aren't
6897 being used. */
6899 if (! live_insn && dbg_cnt (delete_trivial_dead))
6901 if (DEBUG_INSN_P (insn))
6902 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6903 NULL_RTX, -1);
6904 else
6906 rtx set;
6907 if (MAY_HAVE_DEBUG_INSNS
6908 && (set = single_set (insn)) != NULL_RTX
6909 && is_dead_reg (SET_DEST (set), counts)
6910 /* Used at least once in some DEBUG_INSN. */
6911 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6912 /* And set exactly once. */
6913 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6914 && !side_effects_p (SET_SRC (set))
6915 && asm_noperands (PATTERN (insn)) < 0)
6917 rtx dval, bind;
6919 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6920 dval = make_debug_expr_from_rtl (SET_DEST (set));
6922 /* Emit a debug bind insn before the insn in which
6923 reg dies. */
6924 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
6925 DEBUG_EXPR_TREE_DECL (dval),
6926 SET_SRC (set),
6927 VAR_INIT_STATUS_INITIALIZED);
6928 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
6930 bind = emit_debug_insn_before (bind, insn);
6931 df_insn_rescan (bind);
6933 if (replacements == NULL)
6934 replacements = XCNEWVEC (rtx, nreg);
6935 replacements[REGNO (SET_DEST (set))] = dval;
6938 count_reg_usage (insn, counts, NULL_RTX, -1);
6939 ndead++;
6941 delete_insn_and_edges (insn);
6945 if (MAY_HAVE_DEBUG_INSNS)
6947 struct dead_debug_insn_data ddid;
6948 ddid.counts = counts;
6949 ddid.replacements = replacements;
6950 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
6951 if (DEBUG_INSN_P (insn))
6953 /* If this debug insn references a dead register that wasn't replaced
6954 with an DEBUG_EXPR, reset the DEBUG_INSN. */
6955 ddid.seen_repl = false;
6956 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
6957 is_dead_debug_insn, &ddid))
6959 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
6960 df_insn_rescan (insn);
6962 else if (ddid.seen_repl)
6964 INSN_VAR_LOCATION_LOC (insn)
6965 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
6966 NULL_RTX, replace_dead_reg,
6967 replacements);
6968 df_insn_rescan (insn);
6971 free (replacements);
6974 if (dump_file && ndead)
6975 fprintf (dump_file, "Deleted %i trivially dead insns\n",
6976 ndead);
6977 /* Clean up. */
6978 free (counts);
6979 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
6980 return ndead;
6983 /* This function is called via for_each_rtx. The argument, NEWREG, is
6984 a condition code register with the desired mode. If we are looking
6985 at the same register in a different mode, replace it with
6986 NEWREG. */
6988 static int
6989 cse_change_cc_mode (rtx *loc, void *data)
6991 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
6993 if (*loc
6994 && REG_P (*loc)
6995 && REGNO (*loc) == REGNO (args->newreg)
6996 && GET_MODE (*loc) != GET_MODE (args->newreg))
6998 validate_change (args->insn, loc, args->newreg, 1);
7000 return -1;
7002 return 0;
7005 /* Change the mode of any reference to the register REGNO (NEWREG) to
7006 GET_MODE (NEWREG) in INSN. */
7008 static void
7009 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7011 struct change_cc_mode_args args;
7012 int success;
7014 if (!INSN_P (insn))
7015 return;
7017 args.insn = insn;
7018 args.newreg = newreg;
7020 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7021 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7023 /* If the following assertion was triggered, there is most probably
7024 something wrong with the cc_modes_compatible back end function.
7025 CC modes only can be considered compatible if the insn - with the mode
7026 replaced by any of the compatible modes - can still be recognized. */
7027 success = apply_change_group ();
7028 gcc_assert (success);
7031 /* Change the mode of any reference to the register REGNO (NEWREG) to
7032 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7033 any instruction which modifies NEWREG. */
7035 static void
7036 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7038 rtx insn;
7040 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7042 if (! INSN_P (insn))
7043 continue;
7045 if (reg_set_p (newreg, insn))
7046 return;
7048 cse_change_cc_mode_insn (insn, newreg);
7052 /* BB is a basic block which finishes with CC_REG as a condition code
7053 register which is set to CC_SRC. Look through the successors of BB
7054 to find blocks which have a single predecessor (i.e., this one),
7055 and look through those blocks for an assignment to CC_REG which is
7056 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7057 permitted to change the mode of CC_SRC to a compatible mode. This
7058 returns VOIDmode if no equivalent assignments were found.
7059 Otherwise it returns the mode which CC_SRC should wind up with.
7060 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7061 but is passed unmodified down to recursive calls in order to prevent
7062 endless recursion.
7064 The main complexity in this function is handling the mode issues.
7065 We may have more than one duplicate which we can eliminate, and we
7066 try to find a mode which will work for multiple duplicates. */
7068 static enum machine_mode
7069 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7070 bool can_change_mode)
7072 bool found_equiv;
7073 enum machine_mode mode;
7074 unsigned int insn_count;
7075 edge e;
7076 rtx insns[2];
7077 enum machine_mode modes[2];
7078 rtx last_insns[2];
7079 unsigned int i;
7080 rtx newreg;
7081 edge_iterator ei;
7083 /* We expect to have two successors. Look at both before picking
7084 the final mode for the comparison. If we have more successors
7085 (i.e., some sort of table jump, although that seems unlikely),
7086 then we require all beyond the first two to use the same
7087 mode. */
7089 found_equiv = false;
7090 mode = GET_MODE (cc_src);
7091 insn_count = 0;
7092 FOR_EACH_EDGE (e, ei, bb->succs)
7094 rtx insn;
7095 rtx end;
7097 if (e->flags & EDGE_COMPLEX)
7098 continue;
7100 if (EDGE_COUNT (e->dest->preds) != 1
7101 || e->dest == EXIT_BLOCK_PTR
7102 /* Avoid endless recursion on unreachable blocks. */
7103 || e->dest == orig_bb)
7104 continue;
7106 end = NEXT_INSN (BB_END (e->dest));
7107 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7109 rtx set;
7111 if (! INSN_P (insn))
7112 continue;
7114 /* If CC_SRC is modified, we have to stop looking for
7115 something which uses it. */
7116 if (modified_in_p (cc_src, insn))
7117 break;
7119 /* Check whether INSN sets CC_REG to CC_SRC. */
7120 set = single_set (insn);
7121 if (set
7122 && REG_P (SET_DEST (set))
7123 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7125 bool found;
7126 enum machine_mode set_mode;
7127 enum machine_mode comp_mode;
7129 found = false;
7130 set_mode = GET_MODE (SET_SRC (set));
7131 comp_mode = set_mode;
7132 if (rtx_equal_p (cc_src, SET_SRC (set)))
7133 found = true;
7134 else if (GET_CODE (cc_src) == COMPARE
7135 && GET_CODE (SET_SRC (set)) == COMPARE
7136 && mode != set_mode
7137 && rtx_equal_p (XEXP (cc_src, 0),
7138 XEXP (SET_SRC (set), 0))
7139 && rtx_equal_p (XEXP (cc_src, 1),
7140 XEXP (SET_SRC (set), 1)))
7143 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7144 if (comp_mode != VOIDmode
7145 && (can_change_mode || comp_mode == mode))
7146 found = true;
7149 if (found)
7151 found_equiv = true;
7152 if (insn_count < ARRAY_SIZE (insns))
7154 insns[insn_count] = insn;
7155 modes[insn_count] = set_mode;
7156 last_insns[insn_count] = end;
7157 ++insn_count;
7159 if (mode != comp_mode)
7161 gcc_assert (can_change_mode);
7162 mode = comp_mode;
7164 /* The modified insn will be re-recognized later. */
7165 PUT_MODE (cc_src, mode);
7168 else
7170 if (set_mode != mode)
7172 /* We found a matching expression in the
7173 wrong mode, but we don't have room to
7174 store it in the array. Punt. This case
7175 should be rare. */
7176 break;
7178 /* INSN sets CC_REG to a value equal to CC_SRC
7179 with the right mode. We can simply delete
7180 it. */
7181 delete_insn (insn);
7184 /* We found an instruction to delete. Keep looking,
7185 in the hopes of finding a three-way jump. */
7186 continue;
7189 /* We found an instruction which sets the condition
7190 code, so don't look any farther. */
7191 break;
7194 /* If INSN sets CC_REG in some other way, don't look any
7195 farther. */
7196 if (reg_set_p (cc_reg, insn))
7197 break;
7200 /* If we fell off the bottom of the block, we can keep looking
7201 through successors. We pass CAN_CHANGE_MODE as false because
7202 we aren't prepared to handle compatibility between the
7203 further blocks and this block. */
7204 if (insn == end)
7206 enum machine_mode submode;
7208 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7209 if (submode != VOIDmode)
7211 gcc_assert (submode == mode);
7212 found_equiv = true;
7213 can_change_mode = false;
7218 if (! found_equiv)
7219 return VOIDmode;
7221 /* Now INSN_COUNT is the number of instructions we found which set
7222 CC_REG to a value equivalent to CC_SRC. The instructions are in
7223 INSNS. The modes used by those instructions are in MODES. */
7225 newreg = NULL_RTX;
7226 for (i = 0; i < insn_count; ++i)
7228 if (modes[i] != mode)
7230 /* We need to change the mode of CC_REG in INSNS[i] and
7231 subsequent instructions. */
7232 if (! newreg)
7234 if (GET_MODE (cc_reg) == mode)
7235 newreg = cc_reg;
7236 else
7237 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7239 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7240 newreg);
7243 delete_insn_and_edges (insns[i]);
7246 return mode;
7249 /* If we have a fixed condition code register (or two), walk through
7250 the instructions and try to eliminate duplicate assignments. */
7252 static void
7253 cse_condition_code_reg (void)
7255 unsigned int cc_regno_1;
7256 unsigned int cc_regno_2;
7257 rtx cc_reg_1;
7258 rtx cc_reg_2;
7259 basic_block bb;
7261 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7262 return;
7264 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7265 if (cc_regno_2 != INVALID_REGNUM)
7266 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7267 else
7268 cc_reg_2 = NULL_RTX;
7270 FOR_EACH_BB (bb)
7272 rtx last_insn;
7273 rtx cc_reg;
7274 rtx insn;
7275 rtx cc_src_insn;
7276 rtx cc_src;
7277 enum machine_mode mode;
7278 enum machine_mode orig_mode;
7280 /* Look for blocks which end with a conditional jump based on a
7281 condition code register. Then look for the instruction which
7282 sets the condition code register. Then look through the
7283 successor blocks for instructions which set the condition
7284 code register to the same value. There are other possible
7285 uses of the condition code register, but these are by far the
7286 most common and the ones which we are most likely to be able
7287 to optimize. */
7289 last_insn = BB_END (bb);
7290 if (!JUMP_P (last_insn))
7291 continue;
7293 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7294 cc_reg = cc_reg_1;
7295 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7296 cc_reg = cc_reg_2;
7297 else
7298 continue;
7300 cc_src_insn = NULL_RTX;
7301 cc_src = NULL_RTX;
7302 for (insn = PREV_INSN (last_insn);
7303 insn && insn != PREV_INSN (BB_HEAD (bb));
7304 insn = PREV_INSN (insn))
7306 rtx set;
7308 if (! INSN_P (insn))
7309 continue;
7310 set = single_set (insn);
7311 if (set
7312 && REG_P (SET_DEST (set))
7313 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7315 cc_src_insn = insn;
7316 cc_src = SET_SRC (set);
7317 break;
7319 else if (reg_set_p (cc_reg, insn))
7320 break;
7323 if (! cc_src_insn)
7324 continue;
7326 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7327 continue;
7329 /* Now CC_REG is a condition code register used for a
7330 conditional jump at the end of the block, and CC_SRC, in
7331 CC_SRC_INSN, is the value to which that condition code
7332 register is set, and CC_SRC is still meaningful at the end of
7333 the basic block. */
7335 orig_mode = GET_MODE (cc_src);
7336 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7337 if (mode != VOIDmode)
7339 gcc_assert (mode == GET_MODE (cc_src));
7340 if (mode != orig_mode)
7342 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7344 cse_change_cc_mode_insn (cc_src_insn, newreg);
7346 /* Do the same in the following insns that use the
7347 current value of CC_REG within BB. */
7348 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7349 NEXT_INSN (last_insn),
7350 newreg);
7357 /* Perform common subexpression elimination. Nonzero value from
7358 `cse_main' means that jumps were simplified and some code may now
7359 be unreachable, so do jump optimization again. */
7360 static bool
7361 gate_handle_cse (void)
7363 return optimize > 0;
7366 static unsigned int
7367 rest_of_handle_cse (void)
7369 int tem;
7371 if (dump_file)
7372 dump_flow_info (dump_file, dump_flags);
7374 tem = cse_main (get_insns (), max_reg_num ());
7376 /* If we are not running more CSE passes, then we are no longer
7377 expecting CSE to be run. But always rerun it in a cheap mode. */
7378 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7380 if (tem == 2)
7382 timevar_push (TV_JUMP);
7383 rebuild_jump_labels (get_insns ());
7384 cleanup_cfg (0);
7385 timevar_pop (TV_JUMP);
7387 else if (tem == 1 || optimize > 1)
7388 cleanup_cfg (0);
7390 return 0;
7393 struct rtl_opt_pass pass_cse =
7396 RTL_PASS,
7397 "cse1", /* name */
7398 gate_handle_cse, /* gate */
7399 rest_of_handle_cse, /* execute */
7400 NULL, /* sub */
7401 NULL, /* next */
7402 0, /* static_pass_number */
7403 TV_CSE, /* tv_id */
7404 0, /* properties_required */
7405 0, /* properties_provided */
7406 0, /* properties_destroyed */
7407 0, /* todo_flags_start */
7408 TODO_df_finish | TODO_verify_rtl_sharing |
7409 TODO_ggc_collect |
7410 TODO_verify_flow, /* todo_flags_finish */
7415 static bool
7416 gate_handle_cse2 (void)
7418 return optimize > 0 && flag_rerun_cse_after_loop;
7421 /* Run second CSE pass after loop optimizations. */
7422 static unsigned int
7423 rest_of_handle_cse2 (void)
7425 int tem;
7427 if (dump_file)
7428 dump_flow_info (dump_file, dump_flags);
7430 tem = cse_main (get_insns (), max_reg_num ());
7432 /* Run a pass to eliminate duplicated assignments to condition code
7433 registers. We have to run this after bypass_jumps, because it
7434 makes it harder for that pass to determine whether a jump can be
7435 bypassed safely. */
7436 cse_condition_code_reg ();
7438 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7440 if (tem == 2)
7442 timevar_push (TV_JUMP);
7443 rebuild_jump_labels (get_insns ());
7444 cleanup_cfg (0);
7445 timevar_pop (TV_JUMP);
7447 else if (tem == 1)
7448 cleanup_cfg (0);
7450 cse_not_expected = 1;
7451 return 0;
7455 struct rtl_opt_pass pass_cse2 =
7458 RTL_PASS,
7459 "cse2", /* name */
7460 gate_handle_cse2, /* gate */
7461 rest_of_handle_cse2, /* execute */
7462 NULL, /* sub */
7463 NULL, /* next */
7464 0, /* static_pass_number */
7465 TV_CSE2, /* tv_id */
7466 0, /* properties_required */
7467 0, /* properties_provided */
7468 0, /* properties_destroyed */
7469 0, /* todo_flags_start */
7470 TODO_df_finish | TODO_verify_rtl_sharing |
7471 TODO_ggc_collect |
7472 TODO_verify_flow /* todo_flags_finish */
7476 static bool
7477 gate_handle_cse_after_global_opts (void)
7479 return optimize > 0 && flag_rerun_cse_after_global_opts;
7482 /* Run second CSE pass after loop optimizations. */
7483 static unsigned int
7484 rest_of_handle_cse_after_global_opts (void)
7486 int save_cfj;
7487 int tem;
7489 /* We only want to do local CSE, so don't follow jumps. */
7490 save_cfj = flag_cse_follow_jumps;
7491 flag_cse_follow_jumps = 0;
7493 rebuild_jump_labels (get_insns ());
7494 tem = cse_main (get_insns (), max_reg_num ());
7495 purge_all_dead_edges ();
7496 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7498 cse_not_expected = !flag_rerun_cse_after_loop;
7500 /* If cse altered any jumps, rerun jump opts to clean things up. */
7501 if (tem == 2)
7503 timevar_push (TV_JUMP);
7504 rebuild_jump_labels (get_insns ());
7505 cleanup_cfg (0);
7506 timevar_pop (TV_JUMP);
7508 else if (tem == 1)
7509 cleanup_cfg (0);
7511 flag_cse_follow_jumps = save_cfj;
7512 return 0;
7515 struct rtl_opt_pass pass_cse_after_global_opts =
7518 RTL_PASS,
7519 "cse_local", /* name */
7520 gate_handle_cse_after_global_opts, /* gate */
7521 rest_of_handle_cse_after_global_opts, /* execute */
7522 NULL, /* sub */
7523 NULL, /* next */
7524 0, /* static_pass_number */
7525 TV_CSE, /* tv_id */
7526 0, /* properties_required */
7527 0, /* properties_provided */
7528 0, /* properties_destroyed */
7529 0, /* todo_flags_start */
7530 TODO_df_finish | TODO_verify_rtl_sharing |
7531 TODO_ggc_collect |
7532 TODO_verify_flow /* todo_flags_finish */