* config/sh/sh.h: Delete dead GO_IF_LEGITIMATE_INDEX macro.
[official-gcc.git] / gcc / cse.c
blob6424bb1864f61e3b050ed31adea2c78908827dfe
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, 1))
479 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
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, int);
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 void remove_invalid_refs (unsigned int);
577 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
578 enum machine_mode);
579 static void rehash_using_reg (rtx);
580 static void invalidate_memory (void);
581 static void invalidate_for_call (void);
582 static rtx use_related_value (rtx, struct table_elt *);
584 static inline unsigned canon_hash (rtx, enum machine_mode);
585 static inline unsigned safe_hash (rtx, enum machine_mode);
586 static inline unsigned hash_rtx_string (const char *);
588 static rtx canon_reg (rtx, rtx);
589 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
590 enum machine_mode *,
591 enum machine_mode *);
592 static rtx fold_rtx (rtx, rtx);
593 static rtx equiv_constant (rtx);
594 static void record_jump_equiv (rtx, bool);
595 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
596 int);
597 static void cse_insn (rtx);
598 static void cse_prescan_path (struct cse_basic_block_data *);
599 static void invalidate_from_clobbers (rtx);
600 static rtx cse_process_notes (rtx, rtx, bool *);
601 static void cse_extended_basic_block (struct cse_basic_block_data *);
602 static void count_reg_usage (rtx, int *, rtx, int);
603 static int check_for_label_ref (rtx *, void *);
604 extern void dump_class (struct table_elt*);
605 static void get_cse_reg_info_1 (unsigned int regno);
606 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
607 static int check_dependence (rtx *, void *);
609 static void flush_hash_table (void);
610 static bool insn_live_p (rtx, int *);
611 static bool set_live_p (rtx, rtx, int *);
612 static int cse_change_cc_mode (rtx *, void *);
613 static void cse_change_cc_mode_insn (rtx, rtx);
614 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
615 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
616 bool);
619 #undef RTL_HOOKS_GEN_LOWPART
620 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
622 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
624 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
625 virtual regs here because the simplify_*_operation routines are called
626 by integrate.c, which is called before virtual register instantiation. */
628 static bool
629 fixed_base_plus_p (rtx x)
631 switch (GET_CODE (x))
633 case REG:
634 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
635 return true;
636 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
637 return true;
638 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
639 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
640 return true;
641 return false;
643 case PLUS:
644 if (!CONST_INT_P (XEXP (x, 1)))
645 return false;
646 return fixed_base_plus_p (XEXP (x, 0));
648 default:
649 return false;
653 /* Dump the expressions in the equivalence class indicated by CLASSP.
654 This function is used only for debugging. */
655 void
656 dump_class (struct table_elt *classp)
658 struct table_elt *elt;
660 fprintf (stderr, "Equivalence chain for ");
661 print_rtl (stderr, classp->exp);
662 fprintf (stderr, ": \n");
664 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
666 print_rtl (stderr, elt->exp);
667 fprintf (stderr, "\n");
671 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
673 static int
674 approx_reg_cost_1 (rtx *xp, void *data)
676 rtx x = *xp;
677 int *cost_p = (int *) data;
679 if (x && REG_P (x))
681 unsigned int regno = REGNO (x);
683 if (! CHEAP_REGNO (regno))
685 if (regno < FIRST_PSEUDO_REGISTER)
687 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
688 return 1;
689 *cost_p += 2;
691 else
692 *cost_p += 1;
696 return 0;
699 /* Return an estimate of the cost of the registers used in an rtx.
700 This is mostly the number of different REG expressions in the rtx;
701 however for some exceptions like fixed registers we use a cost of
702 0. If any other hard register reference occurs, return MAX_COST. */
704 static int
705 approx_reg_cost (rtx x)
707 int cost = 0;
709 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
710 return MAX_COST;
712 return cost;
715 /* Return a negative value if an rtx A, whose costs are given by COST_A
716 and REGCOST_A, is more desirable than an rtx B.
717 Return a positive value if A is less desirable, or 0 if the two are
718 equally good. */
719 static int
720 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
722 /* First, get rid of cases involving expressions that are entirely
723 unwanted. */
724 if (cost_a != cost_b)
726 if (cost_a == MAX_COST)
727 return 1;
728 if (cost_b == MAX_COST)
729 return -1;
732 /* Avoid extending lifetimes of hardregs. */
733 if (regcost_a != regcost_b)
735 if (regcost_a == MAX_COST)
736 return 1;
737 if (regcost_b == MAX_COST)
738 return -1;
741 /* Normal operation costs take precedence. */
742 if (cost_a != cost_b)
743 return cost_a - cost_b;
744 /* Only if these are identical consider effects on register pressure. */
745 if (regcost_a != regcost_b)
746 return regcost_a - regcost_b;
747 return 0;
750 /* Internal function, to compute cost when X is not a register; called
751 from COST macro to keep it simple. */
753 static int
754 notreg_cost (rtx x, enum rtx_code outer, int opno)
756 return ((GET_CODE (x) == SUBREG
757 && REG_P (SUBREG_REG (x))
758 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
759 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
760 && (GET_MODE_SIZE (GET_MODE (x))
761 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
762 && subreg_lowpart_p (x)
763 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
764 GET_MODE (SUBREG_REG (x))))
766 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
770 /* Initialize CSE_REG_INFO_TABLE. */
772 static void
773 init_cse_reg_info (unsigned int nregs)
775 /* Do we need to grow the table? */
776 if (nregs > cse_reg_info_table_size)
778 unsigned int new_size;
780 if (cse_reg_info_table_size < 2048)
782 /* Compute a new size that is a power of 2 and no smaller
783 than the large of NREGS and 64. */
784 new_size = (cse_reg_info_table_size
785 ? cse_reg_info_table_size : 64);
787 while (new_size < nregs)
788 new_size *= 2;
790 else
792 /* If we need a big table, allocate just enough to hold
793 NREGS registers. */
794 new_size = nregs;
797 /* Reallocate the table with NEW_SIZE entries. */
798 free (cse_reg_info_table);
799 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
800 cse_reg_info_table_size = new_size;
801 cse_reg_info_table_first_uninitialized = 0;
804 /* Do we have all of the first NREGS entries initialized? */
805 if (cse_reg_info_table_first_uninitialized < nregs)
807 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
808 unsigned int i;
810 /* Put the old timestamp on newly allocated entries so that they
811 will all be considered out of date. We do not touch those
812 entries beyond the first NREGS entries to be nice to the
813 virtual memory. */
814 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
815 cse_reg_info_table[i].timestamp = old_timestamp;
817 cse_reg_info_table_first_uninitialized = nregs;
821 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
823 static void
824 get_cse_reg_info_1 (unsigned int regno)
826 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
827 entry will be considered to have been initialized. */
828 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
830 /* Initialize the rest of the entry. */
831 cse_reg_info_table[regno].reg_tick = 1;
832 cse_reg_info_table[regno].reg_in_table = -1;
833 cse_reg_info_table[regno].subreg_ticked = -1;
834 cse_reg_info_table[regno].reg_qty = -regno - 1;
837 /* Find a cse_reg_info entry for REGNO. */
839 static inline struct cse_reg_info *
840 get_cse_reg_info (unsigned int regno)
842 struct cse_reg_info *p = &cse_reg_info_table[regno];
844 /* If this entry has not been initialized, go ahead and initialize
845 it. */
846 if (p->timestamp != cse_reg_info_timestamp)
847 get_cse_reg_info_1 (regno);
849 return p;
852 /* Clear the hash table and initialize each register with its own quantity,
853 for a new basic block. */
855 static void
856 new_basic_block (void)
858 int i;
860 next_qty = 0;
862 /* Invalidate cse_reg_info_table. */
863 cse_reg_info_timestamp++;
865 /* Clear out hash table state for this pass. */
866 CLEAR_HARD_REG_SET (hard_regs_in_table);
868 /* The per-quantity values used to be initialized here, but it is
869 much faster to initialize each as it is made in `make_new_qty'. */
871 for (i = 0; i < HASH_SIZE; i++)
873 struct table_elt *first;
875 first = table[i];
876 if (first != NULL)
878 struct table_elt *last = first;
880 table[i] = NULL;
882 while (last->next_same_hash != NULL)
883 last = last->next_same_hash;
885 /* Now relink this hash entire chain into
886 the free element list. */
888 last->next_same_hash = free_element_chain;
889 free_element_chain = first;
893 #ifdef HAVE_cc0
894 prev_insn_cc0 = 0;
895 #endif
898 /* Say that register REG contains a quantity in mode MODE not in any
899 register before and initialize that quantity. */
901 static void
902 make_new_qty (unsigned int reg, enum machine_mode mode)
904 int q;
905 struct qty_table_elem *ent;
906 struct reg_eqv_elem *eqv;
908 gcc_assert (next_qty < max_qty);
910 q = REG_QTY (reg) = next_qty++;
911 ent = &qty_table[q];
912 ent->first_reg = reg;
913 ent->last_reg = reg;
914 ent->mode = mode;
915 ent->const_rtx = ent->const_insn = NULL_RTX;
916 ent->comparison_code = UNKNOWN;
918 eqv = &reg_eqv_table[reg];
919 eqv->next = eqv->prev = -1;
922 /* Make reg NEW equivalent to reg OLD.
923 OLD is not changing; NEW is. */
925 static void
926 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
928 unsigned int lastr, firstr;
929 int q = REG_QTY (old_reg);
930 struct qty_table_elem *ent;
932 ent = &qty_table[q];
934 /* Nothing should become eqv until it has a "non-invalid" qty number. */
935 gcc_assert (REGNO_QTY_VALID_P (old_reg));
937 REG_QTY (new_reg) = q;
938 firstr = ent->first_reg;
939 lastr = ent->last_reg;
941 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
942 hard regs. Among pseudos, if NEW will live longer than any other reg
943 of the same qty, and that is beyond the current basic block,
944 make it the new canonical replacement for this qty. */
945 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
946 /* Certain fixed registers might be of the class NO_REGS. This means
947 that not only can they not be allocated by the compiler, but
948 they cannot be used in substitutions or canonicalizations
949 either. */
950 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
951 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
952 || (new_reg >= FIRST_PSEUDO_REGISTER
953 && (firstr < FIRST_PSEUDO_REGISTER
954 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
955 && !bitmap_bit_p (cse_ebb_live_out, firstr))
956 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
957 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
959 reg_eqv_table[firstr].prev = new_reg;
960 reg_eqv_table[new_reg].next = firstr;
961 reg_eqv_table[new_reg].prev = -1;
962 ent->first_reg = new_reg;
964 else
966 /* If NEW is a hard reg (known to be non-fixed), insert at end.
967 Otherwise, insert before any non-fixed hard regs that are at the
968 end. Registers of class NO_REGS cannot be used as an
969 equivalent for anything. */
970 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
971 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
972 && new_reg >= FIRST_PSEUDO_REGISTER)
973 lastr = reg_eqv_table[lastr].prev;
974 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
975 if (reg_eqv_table[lastr].next >= 0)
976 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
977 else
978 qty_table[q].last_reg = new_reg;
979 reg_eqv_table[lastr].next = new_reg;
980 reg_eqv_table[new_reg].prev = lastr;
984 /* Remove REG from its equivalence class. */
986 static void
987 delete_reg_equiv (unsigned int reg)
989 struct qty_table_elem *ent;
990 int q = REG_QTY (reg);
991 int p, n;
993 /* If invalid, do nothing. */
994 if (! REGNO_QTY_VALID_P (reg))
995 return;
997 ent = &qty_table[q];
999 p = reg_eqv_table[reg].prev;
1000 n = reg_eqv_table[reg].next;
1002 if (n != -1)
1003 reg_eqv_table[n].prev = p;
1004 else
1005 ent->last_reg = p;
1006 if (p != -1)
1007 reg_eqv_table[p].next = n;
1008 else
1009 ent->first_reg = n;
1011 REG_QTY (reg) = -reg - 1;
1014 /* Remove any invalid expressions from the hash table
1015 that refer to any of the registers contained in expression X.
1017 Make sure that newly inserted references to those registers
1018 as subexpressions will be considered valid.
1020 mention_regs is not called when a register itself
1021 is being stored in the table.
1023 Return 1 if we have done something that may have changed the hash code
1024 of X. */
1026 static int
1027 mention_regs (rtx x)
1029 enum rtx_code code;
1030 int i, j;
1031 const char *fmt;
1032 int changed = 0;
1034 if (x == 0)
1035 return 0;
1037 code = GET_CODE (x);
1038 if (code == REG)
1040 unsigned int regno = REGNO (x);
1041 unsigned int endregno = END_REGNO (x);
1042 unsigned int i;
1044 for (i = regno; i < endregno; i++)
1046 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1047 remove_invalid_refs (i);
1049 REG_IN_TABLE (i) = REG_TICK (i);
1050 SUBREG_TICKED (i) = -1;
1053 return 0;
1056 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1057 pseudo if they don't use overlapping words. We handle only pseudos
1058 here for simplicity. */
1059 if (code == SUBREG && REG_P (SUBREG_REG (x))
1060 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1062 unsigned int i = REGNO (SUBREG_REG (x));
1064 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1066 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1067 the last store to this register really stored into this
1068 subreg, then remove the memory of this subreg.
1069 Otherwise, remove any memory of the entire register and
1070 all its subregs from the table. */
1071 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1072 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1073 remove_invalid_refs (i);
1074 else
1075 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1078 REG_IN_TABLE (i) = REG_TICK (i);
1079 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1080 return 0;
1083 /* If X is a comparison or a COMPARE and either operand is a register
1084 that does not have a quantity, give it one. This is so that a later
1085 call to record_jump_equiv won't cause X to be assigned a different
1086 hash code and not found in the table after that call.
1088 It is not necessary to do this here, since rehash_using_reg can
1089 fix up the table later, but doing this here eliminates the need to
1090 call that expensive function in the most common case where the only
1091 use of the register is in the comparison. */
1093 if (code == COMPARE || COMPARISON_P (x))
1095 if (REG_P (XEXP (x, 0))
1096 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1097 if (insert_regs (XEXP (x, 0), NULL, 0))
1099 rehash_using_reg (XEXP (x, 0));
1100 changed = 1;
1103 if (REG_P (XEXP (x, 1))
1104 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1105 if (insert_regs (XEXP (x, 1), NULL, 0))
1107 rehash_using_reg (XEXP (x, 1));
1108 changed = 1;
1112 fmt = GET_RTX_FORMAT (code);
1113 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1114 if (fmt[i] == 'e')
1115 changed |= mention_regs (XEXP (x, i));
1116 else if (fmt[i] == 'E')
1117 for (j = 0; j < XVECLEN (x, i); j++)
1118 changed |= mention_regs (XVECEXP (x, i, j));
1120 return changed;
1123 /* Update the register quantities for inserting X into the hash table
1124 with a value equivalent to CLASSP.
1125 (If the class does not contain a REG, it is irrelevant.)
1126 If MODIFIED is nonzero, X is a destination; it is being modified.
1127 Note that delete_reg_equiv should be called on a register
1128 before insert_regs is done on that register with MODIFIED != 0.
1130 Nonzero value means that elements of reg_qty have changed
1131 so X's hash code may be different. */
1133 static int
1134 insert_regs (rtx x, struct table_elt *classp, int modified)
1136 if (REG_P (x))
1138 unsigned int regno = REGNO (x);
1139 int qty_valid;
1141 /* If REGNO is in the equivalence table already but is of the
1142 wrong mode for that equivalence, don't do anything here. */
1144 qty_valid = REGNO_QTY_VALID_P (regno);
1145 if (qty_valid)
1147 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1149 if (ent->mode != GET_MODE (x))
1150 return 0;
1153 if (modified || ! qty_valid)
1155 if (classp)
1156 for (classp = classp->first_same_value;
1157 classp != 0;
1158 classp = classp->next_same_value)
1159 if (REG_P (classp->exp)
1160 && GET_MODE (classp->exp) == GET_MODE (x))
1162 unsigned c_regno = REGNO (classp->exp);
1164 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1166 /* Suppose that 5 is hard reg and 100 and 101 are
1167 pseudos. Consider
1169 (set (reg:si 100) (reg:si 5))
1170 (set (reg:si 5) (reg:si 100))
1171 (set (reg:di 101) (reg:di 5))
1173 We would now set REG_QTY (101) = REG_QTY (5), but the
1174 entry for 5 is in SImode. When we use this later in
1175 copy propagation, we get the register in wrong mode. */
1176 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1177 continue;
1179 make_regs_eqv (regno, c_regno);
1180 return 1;
1183 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1184 than REG_IN_TABLE to find out if there was only a single preceding
1185 invalidation - for the SUBREG - or another one, which would be
1186 for the full register. However, if we find here that REG_TICK
1187 indicates that the register is invalid, it means that it has
1188 been invalidated in a separate operation. The SUBREG might be used
1189 now (then this is a recursive call), or we might use the full REG
1190 now and a SUBREG of it later. So bump up REG_TICK so that
1191 mention_regs will do the right thing. */
1192 if (! modified
1193 && REG_IN_TABLE (regno) >= 0
1194 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1195 REG_TICK (regno)++;
1196 make_new_qty (regno, GET_MODE (x));
1197 return 1;
1200 return 0;
1203 /* If X is a SUBREG, we will likely be inserting the inner register in the
1204 table. If that register doesn't have an assigned quantity number at
1205 this point but does later, the insertion that we will be doing now will
1206 not be accessible because its hash code will have changed. So assign
1207 a quantity number now. */
1209 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1210 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1212 insert_regs (SUBREG_REG (x), NULL, 0);
1213 mention_regs (x);
1214 return 1;
1216 else
1217 return mention_regs (x);
1221 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1222 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1223 CST is equal to an anchor. */
1225 static bool
1226 compute_const_anchors (rtx cst,
1227 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1228 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1230 HOST_WIDE_INT n = INTVAL (cst);
1232 *lower_base = n & ~(targetm.const_anchor - 1);
1233 if (*lower_base == n)
1234 return false;
1236 *upper_base =
1237 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1238 *upper_offs = n - *upper_base;
1239 *lower_offs = n - *lower_base;
1240 return true;
1243 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1245 static void
1246 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1247 enum machine_mode mode)
1249 struct table_elt *elt;
1250 unsigned hash;
1251 rtx anchor_exp;
1252 rtx exp;
1254 anchor_exp = GEN_INT (anchor);
1255 hash = HASH (anchor_exp, mode);
1256 elt = lookup (anchor_exp, hash, mode);
1257 if (!elt)
1258 elt = insert (anchor_exp, NULL, hash, mode);
1260 exp = plus_constant (reg, offs);
1261 /* REG has just been inserted and the hash codes recomputed. */
1262 mention_regs (exp);
1263 hash = HASH (exp, mode);
1265 /* Use the cost of the register rather than the whole expression. When
1266 looking up constant anchors we will further offset the corresponding
1267 expression therefore it does not make sense to prefer REGs over
1268 reg-immediate additions. Prefer instead the oldest expression. Also
1269 don't prefer pseudos over hard regs so that we derive constants in
1270 argument registers from other argument registers rather than from the
1271 original pseudo that was used to synthesize the constant. */
1272 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1275 /* The constant CST is equivalent to the register REG. Create
1276 equivalences between the two anchors of CST and the corresponding
1277 register-offset expressions using REG. */
1279 static void
1280 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1282 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1284 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1285 &upper_base, &upper_offs))
1286 return;
1288 /* Ignore anchors of value 0. Constants accessible from zero are
1289 simple. */
1290 if (lower_base != 0)
1291 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1293 if (upper_base != 0)
1294 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1297 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1298 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1299 valid expression. Return the cheapest and oldest of such expressions. In
1300 *OLD, return how old the resulting expression is compared to the other
1301 equivalent expressions. */
1303 static rtx
1304 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1305 unsigned *old)
1307 struct table_elt *elt;
1308 unsigned idx;
1309 struct table_elt *match_elt;
1310 rtx match;
1312 /* Find the cheapest and *oldest* expression to maximize the chance of
1313 reusing the same pseudo. */
1315 match_elt = NULL;
1316 match = NULL_RTX;
1317 for (elt = anchor_elt->first_same_value, idx = 0;
1318 elt;
1319 elt = elt->next_same_value, idx++)
1321 if (match_elt && CHEAPER (match_elt, elt))
1322 return match;
1324 if (REG_P (elt->exp)
1325 || (GET_CODE (elt->exp) == PLUS
1326 && REG_P (XEXP (elt->exp, 0))
1327 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1329 rtx x;
1331 /* Ignore expressions that are no longer valid. */
1332 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1333 continue;
1335 x = plus_constant (elt->exp, offs);
1336 if (REG_P (x)
1337 || (GET_CODE (x) == PLUS
1338 && IN_RANGE (INTVAL (XEXP (x, 1)),
1339 -targetm.const_anchor,
1340 targetm.const_anchor - 1)))
1342 match = x;
1343 match_elt = elt;
1344 *old = idx;
1349 return match;
1352 /* Try to express the constant SRC_CONST using a register+offset expression
1353 derived from a constant anchor. Return it if successful or NULL_RTX,
1354 otherwise. */
1356 static rtx
1357 try_const_anchors (rtx src_const, enum machine_mode mode)
1359 struct table_elt *lower_elt, *upper_elt;
1360 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1361 rtx lower_anchor_rtx, upper_anchor_rtx;
1362 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1363 unsigned lower_old, upper_old;
1365 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1366 &upper_base, &upper_offs))
1367 return NULL_RTX;
1369 lower_anchor_rtx = GEN_INT (lower_base);
1370 upper_anchor_rtx = GEN_INT (upper_base);
1371 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1372 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1374 if (lower_elt)
1375 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1376 if (upper_elt)
1377 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1379 if (!lower_exp)
1380 return upper_exp;
1381 if (!upper_exp)
1382 return lower_exp;
1384 /* Return the older expression. */
1385 return (upper_old > lower_old ? upper_exp : lower_exp);
1388 /* Look in or update the hash table. */
1390 /* Remove table element ELT from use in the table.
1391 HASH is its hash code, made using the HASH macro.
1392 It's an argument because often that is known in advance
1393 and we save much time not recomputing it. */
1395 static void
1396 remove_from_table (struct table_elt *elt, unsigned int hash)
1398 if (elt == 0)
1399 return;
1401 /* Mark this element as removed. See cse_insn. */
1402 elt->first_same_value = 0;
1404 /* Remove the table element from its equivalence class. */
1407 struct table_elt *prev = elt->prev_same_value;
1408 struct table_elt *next = elt->next_same_value;
1410 if (next)
1411 next->prev_same_value = prev;
1413 if (prev)
1414 prev->next_same_value = next;
1415 else
1417 struct table_elt *newfirst = next;
1418 while (next)
1420 next->first_same_value = newfirst;
1421 next = next->next_same_value;
1426 /* Remove the table element from its hash bucket. */
1429 struct table_elt *prev = elt->prev_same_hash;
1430 struct table_elt *next = elt->next_same_hash;
1432 if (next)
1433 next->prev_same_hash = prev;
1435 if (prev)
1436 prev->next_same_hash = next;
1437 else if (table[hash] == elt)
1438 table[hash] = next;
1439 else
1441 /* This entry is not in the proper hash bucket. This can happen
1442 when two classes were merged by `merge_equiv_classes'. Search
1443 for the hash bucket that it heads. This happens only very
1444 rarely, so the cost is acceptable. */
1445 for (hash = 0; hash < HASH_SIZE; hash++)
1446 if (table[hash] == elt)
1447 table[hash] = next;
1451 /* Remove the table element from its related-value circular chain. */
1453 if (elt->related_value != 0 && elt->related_value != elt)
1455 struct table_elt *p = elt->related_value;
1457 while (p->related_value != elt)
1458 p = p->related_value;
1459 p->related_value = elt->related_value;
1460 if (p->related_value == p)
1461 p->related_value = 0;
1464 /* Now add it to the free element chain. */
1465 elt->next_same_hash = free_element_chain;
1466 free_element_chain = elt;
1469 /* Same as above, but X is a pseudo-register. */
1471 static void
1472 remove_pseudo_from_table (rtx x, unsigned int hash)
1474 struct table_elt *elt;
1476 /* Because a pseudo-register can be referenced in more than one
1477 mode, we might have to remove more than one table entry. */
1478 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1479 remove_from_table (elt, hash);
1482 /* Look up X in the hash table and return its table element,
1483 or 0 if X is not in the table.
1485 MODE is the machine-mode of X, or if X is an integer constant
1486 with VOIDmode then MODE is the mode with which X will be used.
1488 Here we are satisfied to find an expression whose tree structure
1489 looks like X. */
1491 static struct table_elt *
1492 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1494 struct table_elt *p;
1496 for (p = table[hash]; p; p = p->next_same_hash)
1497 if (mode == p->mode && ((x == p->exp && REG_P (x))
1498 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1499 return p;
1501 return 0;
1504 /* Like `lookup' but don't care whether the table element uses invalid regs.
1505 Also ignore discrepancies in the machine mode of a register. */
1507 static struct table_elt *
1508 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1510 struct table_elt *p;
1512 if (REG_P (x))
1514 unsigned int regno = REGNO (x);
1516 /* Don't check the machine mode when comparing registers;
1517 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1518 for (p = table[hash]; p; p = p->next_same_hash)
1519 if (REG_P (p->exp)
1520 && REGNO (p->exp) == regno)
1521 return p;
1523 else
1525 for (p = table[hash]; p; p = p->next_same_hash)
1526 if (mode == p->mode
1527 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1528 return p;
1531 return 0;
1534 /* Look for an expression equivalent to X and with code CODE.
1535 If one is found, return that expression. */
1537 static rtx
1538 lookup_as_function (rtx x, enum rtx_code code)
1540 struct table_elt *p
1541 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1543 if (p == 0)
1544 return 0;
1546 for (p = p->first_same_value; p; p = p->next_same_value)
1547 if (GET_CODE (p->exp) == code
1548 /* Make sure this is a valid entry in the table. */
1549 && exp_equiv_p (p->exp, p->exp, 1, false))
1550 return p->exp;
1552 return 0;
1555 /* Insert X in the hash table, assuming HASH is its hash code and
1556 CLASSP is an element of the class it should go in (or 0 if a new
1557 class should be made). COST is the code of X and reg_cost is the
1558 cost of registers in X. It is inserted at the proper position to
1559 keep the class in the order cheapest first.
1561 MODE is the machine-mode of X, or if X is an integer constant
1562 with VOIDmode then MODE is the mode with which X will be used.
1564 For elements of equal cheapness, the most recent one
1565 goes in front, except that the first element in the list
1566 remains first unless a cheaper element is added. The order of
1567 pseudo-registers does not matter, as canon_reg will be called to
1568 find the cheapest when a register is retrieved from the table.
1570 The in_memory field in the hash table element is set to 0.
1571 The caller must set it nonzero if appropriate.
1573 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1574 and if insert_regs returns a nonzero value
1575 you must then recompute its hash code before calling here.
1577 If necessary, update table showing constant values of quantities. */
1579 static struct table_elt *
1580 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1581 enum machine_mode mode, int cost, int reg_cost)
1583 struct table_elt *elt;
1585 /* If X is a register and we haven't made a quantity for it,
1586 something is wrong. */
1587 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1589 /* If X is a hard register, show it is being put in the table. */
1590 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1591 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1593 /* Put an element for X into the right hash bucket. */
1595 elt = free_element_chain;
1596 if (elt)
1597 free_element_chain = elt->next_same_hash;
1598 else
1599 elt = XNEW (struct table_elt);
1601 elt->exp = x;
1602 elt->canon_exp = NULL_RTX;
1603 elt->cost = cost;
1604 elt->regcost = reg_cost;
1605 elt->next_same_value = 0;
1606 elt->prev_same_value = 0;
1607 elt->next_same_hash = table[hash];
1608 elt->prev_same_hash = 0;
1609 elt->related_value = 0;
1610 elt->in_memory = 0;
1611 elt->mode = mode;
1612 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1614 if (table[hash])
1615 table[hash]->prev_same_hash = elt;
1616 table[hash] = elt;
1618 /* Put it into the proper value-class. */
1619 if (classp)
1621 classp = classp->first_same_value;
1622 if (CHEAPER (elt, classp))
1623 /* Insert at the head of the class. */
1625 struct table_elt *p;
1626 elt->next_same_value = classp;
1627 classp->prev_same_value = elt;
1628 elt->first_same_value = elt;
1630 for (p = classp; p; p = p->next_same_value)
1631 p->first_same_value = elt;
1633 else
1635 /* Insert not at head of the class. */
1636 /* Put it after the last element cheaper than X. */
1637 struct table_elt *p, *next;
1639 for (p = classp;
1640 (next = p->next_same_value) && CHEAPER (next, elt);
1641 p = next)
1644 /* Put it after P and before NEXT. */
1645 elt->next_same_value = next;
1646 if (next)
1647 next->prev_same_value = elt;
1649 elt->prev_same_value = p;
1650 p->next_same_value = elt;
1651 elt->first_same_value = classp;
1654 else
1655 elt->first_same_value = elt;
1657 /* If this is a constant being set equivalent to a register or a register
1658 being set equivalent to a constant, note the constant equivalence.
1660 If this is a constant, it cannot be equivalent to a different constant,
1661 and a constant is the only thing that can be cheaper than a register. So
1662 we know the register is the head of the class (before the constant was
1663 inserted).
1665 If this is a register that is not already known equivalent to a
1666 constant, we must check the entire class.
1668 If this is a register that is already known equivalent to an insn,
1669 update the qtys `const_insn' to show that `this_insn' is the latest
1670 insn making that quantity equivalent to the constant. */
1672 if (elt->is_const && classp && REG_P (classp->exp)
1673 && !REG_P (x))
1675 int exp_q = REG_QTY (REGNO (classp->exp));
1676 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1678 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1679 exp_ent->const_insn = this_insn;
1682 else if (REG_P (x)
1683 && classp
1684 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1685 && ! elt->is_const)
1687 struct table_elt *p;
1689 for (p = classp; p != 0; p = p->next_same_value)
1691 if (p->is_const && !REG_P (p->exp))
1693 int x_q = REG_QTY (REGNO (x));
1694 struct qty_table_elem *x_ent = &qty_table[x_q];
1696 x_ent->const_rtx
1697 = gen_lowpart (GET_MODE (x), p->exp);
1698 x_ent->const_insn = this_insn;
1699 break;
1704 else if (REG_P (x)
1705 && qty_table[REG_QTY (REGNO (x))].const_rtx
1706 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1707 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1709 /* If this is a constant with symbolic value,
1710 and it has a term with an explicit integer value,
1711 link it up with related expressions. */
1712 if (GET_CODE (x) == CONST)
1714 rtx subexp = get_related_value (x);
1715 unsigned subhash;
1716 struct table_elt *subelt, *subelt_prev;
1718 if (subexp != 0)
1720 /* Get the integer-free subexpression in the hash table. */
1721 subhash = SAFE_HASH (subexp, mode);
1722 subelt = lookup (subexp, subhash, mode);
1723 if (subelt == 0)
1724 subelt = insert (subexp, NULL, subhash, mode);
1725 /* Initialize SUBELT's circular chain if it has none. */
1726 if (subelt->related_value == 0)
1727 subelt->related_value = subelt;
1728 /* Find the element in the circular chain that precedes SUBELT. */
1729 subelt_prev = subelt;
1730 while (subelt_prev->related_value != subelt)
1731 subelt_prev = subelt_prev->related_value;
1732 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1733 This way the element that follows SUBELT is the oldest one. */
1734 elt->related_value = subelt_prev->related_value;
1735 subelt_prev->related_value = elt;
1739 return elt;
1742 /* Wrap insert_with_costs by passing the default costs. */
1744 static struct table_elt *
1745 insert (rtx x, struct table_elt *classp, unsigned int hash,
1746 enum machine_mode mode)
1748 return
1749 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1753 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1754 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1755 the two classes equivalent.
1757 CLASS1 will be the surviving class; CLASS2 should not be used after this
1758 call.
1760 Any invalid entries in CLASS2 will not be copied. */
1762 static void
1763 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1765 struct table_elt *elt, *next, *new_elt;
1767 /* Ensure we start with the head of the classes. */
1768 class1 = class1->first_same_value;
1769 class2 = class2->first_same_value;
1771 /* If they were already equal, forget it. */
1772 if (class1 == class2)
1773 return;
1775 for (elt = class2; elt; elt = next)
1777 unsigned int hash;
1778 rtx exp = elt->exp;
1779 enum machine_mode mode = elt->mode;
1781 next = elt->next_same_value;
1783 /* Remove old entry, make a new one in CLASS1's class.
1784 Don't do this for invalid entries as we cannot find their
1785 hash code (it also isn't necessary). */
1786 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1788 bool need_rehash = false;
1790 hash_arg_in_memory = 0;
1791 hash = HASH (exp, mode);
1793 if (REG_P (exp))
1795 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1796 delete_reg_equiv (REGNO (exp));
1799 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1800 remove_pseudo_from_table (exp, hash);
1801 else
1802 remove_from_table (elt, hash);
1804 if (insert_regs (exp, class1, 0) || need_rehash)
1806 rehash_using_reg (exp);
1807 hash = HASH (exp, mode);
1809 new_elt = insert (exp, class1, hash, mode);
1810 new_elt->in_memory = hash_arg_in_memory;
1815 /* Flush the entire hash table. */
1817 static void
1818 flush_hash_table (void)
1820 int i;
1821 struct table_elt *p;
1823 for (i = 0; i < HASH_SIZE; i++)
1824 for (p = table[i]; p; p = table[i])
1826 /* Note that invalidate can remove elements
1827 after P in the current hash chain. */
1828 if (REG_P (p->exp))
1829 invalidate (p->exp, VOIDmode);
1830 else
1831 remove_from_table (p, i);
1835 /* Function called for each rtx to check whether true dependence exist. */
1836 struct check_dependence_data
1838 enum machine_mode mode;
1839 rtx exp;
1840 rtx addr;
1843 static int
1844 check_dependence (rtx *x, void *data)
1846 struct check_dependence_data *d = (struct check_dependence_data *) data;
1847 if (*x && MEM_P (*x))
1848 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX);
1849 else
1850 return 0;
1853 /* Remove from the hash table, or mark as invalid, all expressions whose
1854 values could be altered by storing in X. X is a register, a subreg, or
1855 a memory reference with nonvarying address (because, when a memory
1856 reference with a varying address is stored in, all memory references are
1857 removed by invalidate_memory so specific invalidation is superfluous).
1858 FULL_MODE, if not VOIDmode, indicates that this much should be
1859 invalidated instead of just the amount indicated by the mode of X. This
1860 is only used for bitfield stores into memory.
1862 A nonvarying address may be just a register or just a symbol reference,
1863 or it may be either of those plus a numeric offset. */
1865 static void
1866 invalidate (rtx x, enum machine_mode full_mode)
1868 int i;
1869 struct table_elt *p;
1870 rtx addr;
1872 switch (GET_CODE (x))
1874 case REG:
1876 /* If X is a register, dependencies on its contents are recorded
1877 through the qty number mechanism. Just change the qty number of
1878 the register, mark it as invalid for expressions that refer to it,
1879 and remove it itself. */
1880 unsigned int regno = REGNO (x);
1881 unsigned int hash = HASH (x, GET_MODE (x));
1883 /* Remove REGNO from any quantity list it might be on and indicate
1884 that its value might have changed. If it is a pseudo, remove its
1885 entry from the hash table.
1887 For a hard register, we do the first two actions above for any
1888 additional hard registers corresponding to X. Then, if any of these
1889 registers are in the table, we must remove any REG entries that
1890 overlap these registers. */
1892 delete_reg_equiv (regno);
1893 REG_TICK (regno)++;
1894 SUBREG_TICKED (regno) = -1;
1896 if (regno >= FIRST_PSEUDO_REGISTER)
1897 remove_pseudo_from_table (x, hash);
1898 else
1900 HOST_WIDE_INT in_table
1901 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1902 unsigned int endregno = END_HARD_REGNO (x);
1903 unsigned int tregno, tendregno, rn;
1904 struct table_elt *p, *next;
1906 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1908 for (rn = regno + 1; rn < endregno; rn++)
1910 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1911 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1912 delete_reg_equiv (rn);
1913 REG_TICK (rn)++;
1914 SUBREG_TICKED (rn) = -1;
1917 if (in_table)
1918 for (hash = 0; hash < HASH_SIZE; hash++)
1919 for (p = table[hash]; p; p = next)
1921 next = p->next_same_hash;
1923 if (!REG_P (p->exp)
1924 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1925 continue;
1927 tregno = REGNO (p->exp);
1928 tendregno = END_HARD_REGNO (p->exp);
1929 if (tendregno > regno && tregno < endregno)
1930 remove_from_table (p, hash);
1934 return;
1936 case SUBREG:
1937 invalidate (SUBREG_REG (x), VOIDmode);
1938 return;
1940 case PARALLEL:
1941 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1942 invalidate (XVECEXP (x, 0, i), VOIDmode);
1943 return;
1945 case EXPR_LIST:
1946 /* This is part of a disjoint return value; extract the location in
1947 question ignoring the offset. */
1948 invalidate (XEXP (x, 0), VOIDmode);
1949 return;
1951 case MEM:
1952 addr = canon_rtx (get_addr (XEXP (x, 0)));
1953 /* Calculate the canonical version of X here so that
1954 true_dependence doesn't generate new RTL for X on each call. */
1955 x = canon_rtx (x);
1957 /* Remove all hash table elements that refer to overlapping pieces of
1958 memory. */
1959 if (full_mode == VOIDmode)
1960 full_mode = GET_MODE (x);
1962 for (i = 0; i < HASH_SIZE; i++)
1964 struct table_elt *next;
1966 for (p = table[i]; p; p = next)
1968 next = p->next_same_hash;
1969 if (p->in_memory)
1971 struct check_dependence_data d;
1973 /* Just canonicalize the expression once;
1974 otherwise each time we call invalidate
1975 true_dependence will canonicalize the
1976 expression again. */
1977 if (!p->canon_exp)
1978 p->canon_exp = canon_rtx (p->exp);
1979 d.exp = x;
1980 d.addr = addr;
1981 d.mode = full_mode;
1982 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1983 remove_from_table (p, i);
1987 return;
1989 default:
1990 gcc_unreachable ();
1994 /* Remove all expressions that refer to register REGNO,
1995 since they are already invalid, and we are about to
1996 mark that register valid again and don't want the old
1997 expressions to reappear as valid. */
1999 static void
2000 remove_invalid_refs (unsigned int regno)
2002 unsigned int i;
2003 struct table_elt *p, *next;
2005 for (i = 0; i < HASH_SIZE; i++)
2006 for (p = table[i]; p; p = next)
2008 next = p->next_same_hash;
2009 if (!REG_P (p->exp)
2010 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2011 remove_from_table (p, i);
2015 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2016 and mode MODE. */
2017 static void
2018 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2019 enum machine_mode mode)
2021 unsigned int i;
2022 struct table_elt *p, *next;
2023 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2025 for (i = 0; i < HASH_SIZE; i++)
2026 for (p = table[i]; p; p = next)
2028 rtx exp = p->exp;
2029 next = p->next_same_hash;
2031 if (!REG_P (exp)
2032 && (GET_CODE (exp) != SUBREG
2033 || !REG_P (SUBREG_REG (exp))
2034 || REGNO (SUBREG_REG (exp)) != regno
2035 || (((SUBREG_BYTE (exp)
2036 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2037 && SUBREG_BYTE (exp) <= end))
2038 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2039 remove_from_table (p, i);
2043 /* Recompute the hash codes of any valid entries in the hash table that
2044 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2046 This is called when we make a jump equivalence. */
2048 static void
2049 rehash_using_reg (rtx x)
2051 unsigned int i;
2052 struct table_elt *p, *next;
2053 unsigned hash;
2055 if (GET_CODE (x) == SUBREG)
2056 x = SUBREG_REG (x);
2058 /* If X is not a register or if the register is known not to be in any
2059 valid entries in the table, we have no work to do. */
2061 if (!REG_P (x)
2062 || REG_IN_TABLE (REGNO (x)) < 0
2063 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2064 return;
2066 /* Scan all hash chains looking for valid entries that mention X.
2067 If we find one and it is in the wrong hash chain, move it. */
2069 for (i = 0; i < HASH_SIZE; i++)
2070 for (p = table[i]; p; p = next)
2072 next = p->next_same_hash;
2073 if (reg_mentioned_p (x, p->exp)
2074 && exp_equiv_p (p->exp, p->exp, 1, false)
2075 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2077 if (p->next_same_hash)
2078 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2080 if (p->prev_same_hash)
2081 p->prev_same_hash->next_same_hash = p->next_same_hash;
2082 else
2083 table[i] = p->next_same_hash;
2085 p->next_same_hash = table[hash];
2086 p->prev_same_hash = 0;
2087 if (table[hash])
2088 table[hash]->prev_same_hash = p;
2089 table[hash] = p;
2094 /* Remove from the hash table any expression that is a call-clobbered
2095 register. Also update their TICK values. */
2097 static void
2098 invalidate_for_call (void)
2100 unsigned int regno, endregno;
2101 unsigned int i;
2102 unsigned hash;
2103 struct table_elt *p, *next;
2104 int in_table = 0;
2106 /* Go through all the hard registers. For each that is clobbered in
2107 a CALL_INSN, remove the register from quantity chains and update
2108 reg_tick if defined. Also see if any of these registers is currently
2109 in the table. */
2111 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2112 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2114 delete_reg_equiv (regno);
2115 if (REG_TICK (regno) >= 0)
2117 REG_TICK (regno)++;
2118 SUBREG_TICKED (regno) = -1;
2121 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2124 /* In the case where we have no call-clobbered hard registers in the
2125 table, we are done. Otherwise, scan the table and remove any
2126 entry that overlaps a call-clobbered register. */
2128 if (in_table)
2129 for (hash = 0; hash < HASH_SIZE; hash++)
2130 for (p = table[hash]; p; p = next)
2132 next = p->next_same_hash;
2134 if (!REG_P (p->exp)
2135 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2136 continue;
2138 regno = REGNO (p->exp);
2139 endregno = END_HARD_REGNO (p->exp);
2141 for (i = regno; i < endregno; i++)
2142 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2144 remove_from_table (p, hash);
2145 break;
2150 /* Given an expression X of type CONST,
2151 and ELT which is its table entry (or 0 if it
2152 is not in the hash table),
2153 return an alternate expression for X as a register plus integer.
2154 If none can be found, return 0. */
2156 static rtx
2157 use_related_value (rtx x, struct table_elt *elt)
2159 struct table_elt *relt = 0;
2160 struct table_elt *p, *q;
2161 HOST_WIDE_INT offset;
2163 /* First, is there anything related known?
2164 If we have a table element, we can tell from that.
2165 Otherwise, must look it up. */
2167 if (elt != 0 && elt->related_value != 0)
2168 relt = elt;
2169 else if (elt == 0 && GET_CODE (x) == CONST)
2171 rtx subexp = get_related_value (x);
2172 if (subexp != 0)
2173 relt = lookup (subexp,
2174 SAFE_HASH (subexp, GET_MODE (subexp)),
2175 GET_MODE (subexp));
2178 if (relt == 0)
2179 return 0;
2181 /* Search all related table entries for one that has an
2182 equivalent register. */
2184 p = relt;
2185 while (1)
2187 /* This loop is strange in that it is executed in two different cases.
2188 The first is when X is already in the table. Then it is searching
2189 the RELATED_VALUE list of X's class (RELT). The second case is when
2190 X is not in the table. Then RELT points to a class for the related
2191 value.
2193 Ensure that, whatever case we are in, that we ignore classes that have
2194 the same value as X. */
2196 if (rtx_equal_p (x, p->exp))
2197 q = 0;
2198 else
2199 for (q = p->first_same_value; q; q = q->next_same_value)
2200 if (REG_P (q->exp))
2201 break;
2203 if (q)
2204 break;
2206 p = p->related_value;
2208 /* We went all the way around, so there is nothing to be found.
2209 Alternatively, perhaps RELT was in the table for some other reason
2210 and it has no related values recorded. */
2211 if (p == relt || p == 0)
2212 break;
2215 if (q == 0)
2216 return 0;
2218 offset = (get_integer_term (x) - get_integer_term (p->exp));
2219 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2220 return plus_constant (q->exp, offset);
2224 /* Hash a string. Just add its bytes up. */
2225 static inline unsigned
2226 hash_rtx_string (const char *ps)
2228 unsigned hash = 0;
2229 const unsigned char *p = (const unsigned char *) ps;
2231 if (p)
2232 while (*p)
2233 hash += *p++;
2235 return hash;
2238 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2239 When the callback returns true, we continue with the new rtx. */
2241 unsigned
2242 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2243 int *do_not_record_p, int *hash_arg_in_memory_p,
2244 bool have_reg_qty, hash_rtx_callback_function cb)
2246 int i, j;
2247 unsigned hash = 0;
2248 enum rtx_code code;
2249 const char *fmt;
2250 enum machine_mode newmode;
2251 rtx newx;
2253 /* Used to turn recursion into iteration. We can't rely on GCC's
2254 tail-recursion elimination since we need to keep accumulating values
2255 in HASH. */
2256 repeat:
2257 if (x == 0)
2258 return hash;
2260 /* Invoke the callback first. */
2261 if (cb != NULL
2262 && ((*cb) (x, mode, &newx, &newmode)))
2264 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2265 hash_arg_in_memory_p, have_reg_qty, cb);
2266 return hash;
2269 code = GET_CODE (x);
2270 switch (code)
2272 case REG:
2274 unsigned int regno = REGNO (x);
2276 if (do_not_record_p && !reload_completed)
2278 /* On some machines, we can't record any non-fixed hard register,
2279 because extending its life will cause reload problems. We
2280 consider ap, fp, sp, gp to be fixed for this purpose.
2282 We also consider CCmode registers to be fixed for this purpose;
2283 failure to do so leads to failure to simplify 0<100 type of
2284 conditionals.
2286 On all machines, we can't record any global registers.
2287 Nor should we record any register that is in a small
2288 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2289 bool record;
2291 if (regno >= FIRST_PSEUDO_REGISTER)
2292 record = true;
2293 else if (x == frame_pointer_rtx
2294 || x == hard_frame_pointer_rtx
2295 || x == arg_pointer_rtx
2296 || x == stack_pointer_rtx
2297 || x == pic_offset_table_rtx)
2298 record = true;
2299 else if (global_regs[regno])
2300 record = false;
2301 else if (fixed_regs[regno])
2302 record = true;
2303 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2304 record = true;
2305 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2306 record = false;
2307 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2308 record = false;
2309 else
2310 record = true;
2312 if (!record)
2314 *do_not_record_p = 1;
2315 return 0;
2319 hash += ((unsigned int) REG << 7);
2320 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2321 return hash;
2324 /* We handle SUBREG of a REG specially because the underlying
2325 reg changes its hash value with every value change; we don't
2326 want to have to forget unrelated subregs when one subreg changes. */
2327 case SUBREG:
2329 if (REG_P (SUBREG_REG (x)))
2331 hash += (((unsigned int) SUBREG << 7)
2332 + REGNO (SUBREG_REG (x))
2333 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2334 return hash;
2336 break;
2339 case CONST_INT:
2340 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2341 + (unsigned int) INTVAL (x));
2342 return hash;
2344 case CONST_DOUBLE:
2345 /* This is like the general case, except that it only counts
2346 the integers representing the constant. */
2347 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2348 if (GET_MODE (x) != VOIDmode)
2349 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2350 else
2351 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2352 + (unsigned int) CONST_DOUBLE_HIGH (x));
2353 return hash;
2355 case CONST_FIXED:
2356 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2357 hash += fixed_hash (CONST_FIXED_VALUE (x));
2358 return hash;
2360 case CONST_VECTOR:
2362 int units;
2363 rtx elt;
2365 units = CONST_VECTOR_NUNITS (x);
2367 for (i = 0; i < units; ++i)
2369 elt = CONST_VECTOR_ELT (x, i);
2370 hash += hash_rtx_cb (elt, GET_MODE (elt),
2371 do_not_record_p, hash_arg_in_memory_p,
2372 have_reg_qty, cb);
2375 return hash;
2378 /* Assume there is only one rtx object for any given label. */
2379 case LABEL_REF:
2380 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2381 differences and differences between each stage's debugging dumps. */
2382 hash += (((unsigned int) LABEL_REF << 7)
2383 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2384 return hash;
2386 case SYMBOL_REF:
2388 /* Don't hash on the symbol's address to avoid bootstrap differences.
2389 Different hash values may cause expressions to be recorded in
2390 different orders and thus different registers to be used in the
2391 final assembler. This also avoids differences in the dump files
2392 between various stages. */
2393 unsigned int h = 0;
2394 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2396 while (*p)
2397 h += (h << 7) + *p++; /* ??? revisit */
2399 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2400 return hash;
2403 case MEM:
2404 /* We don't record if marked volatile or if BLKmode since we don't
2405 know the size of the move. */
2406 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2408 *do_not_record_p = 1;
2409 return 0;
2411 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2412 *hash_arg_in_memory_p = 1;
2414 /* Now that we have already found this special case,
2415 might as well speed it up as much as possible. */
2416 hash += (unsigned) MEM;
2417 x = XEXP (x, 0);
2418 goto repeat;
2420 case USE:
2421 /* A USE that mentions non-volatile memory needs special
2422 handling since the MEM may be BLKmode which normally
2423 prevents an entry from being made. Pure calls are
2424 marked by a USE which mentions BLKmode memory.
2425 See calls.c:emit_call_1. */
2426 if (MEM_P (XEXP (x, 0))
2427 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2429 hash += (unsigned) USE;
2430 x = XEXP (x, 0);
2432 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2433 *hash_arg_in_memory_p = 1;
2435 /* Now that we have already found this special case,
2436 might as well speed it up as much as possible. */
2437 hash += (unsigned) MEM;
2438 x = XEXP (x, 0);
2439 goto repeat;
2441 break;
2443 case PRE_DEC:
2444 case PRE_INC:
2445 case POST_DEC:
2446 case POST_INC:
2447 case PRE_MODIFY:
2448 case POST_MODIFY:
2449 case PC:
2450 case CC0:
2451 case CALL:
2452 case UNSPEC_VOLATILE:
2453 if (do_not_record_p) {
2454 *do_not_record_p = 1;
2455 return 0;
2457 else
2458 return hash;
2459 break;
2461 case ASM_OPERANDS:
2462 if (do_not_record_p && MEM_VOLATILE_P (x))
2464 *do_not_record_p = 1;
2465 return 0;
2467 else
2469 /* We don't want to take the filename and line into account. */
2470 hash += (unsigned) code + (unsigned) GET_MODE (x)
2471 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2472 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2473 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2475 if (ASM_OPERANDS_INPUT_LENGTH (x))
2477 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2479 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2480 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2481 do_not_record_p, hash_arg_in_memory_p,
2482 have_reg_qty, cb)
2483 + hash_rtx_string
2484 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2487 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2488 x = ASM_OPERANDS_INPUT (x, 0);
2489 mode = GET_MODE (x);
2490 goto repeat;
2493 return hash;
2495 break;
2497 default:
2498 break;
2501 i = GET_RTX_LENGTH (code) - 1;
2502 hash += (unsigned) code + (unsigned) GET_MODE (x);
2503 fmt = GET_RTX_FORMAT (code);
2504 for (; i >= 0; i--)
2506 switch (fmt[i])
2508 case 'e':
2509 /* If we are about to do the last recursive call
2510 needed at this level, change it into iteration.
2511 This function is called enough to be worth it. */
2512 if (i == 0)
2514 x = XEXP (x, i);
2515 goto repeat;
2518 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2519 hash_arg_in_memory_p,
2520 have_reg_qty, cb);
2521 break;
2523 case 'E':
2524 for (j = 0; j < XVECLEN (x, i); j++)
2525 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2526 hash_arg_in_memory_p,
2527 have_reg_qty, cb);
2528 break;
2530 case 's':
2531 hash += hash_rtx_string (XSTR (x, i));
2532 break;
2534 case 'i':
2535 hash += (unsigned int) XINT (x, i);
2536 break;
2538 case '0': case 't':
2539 /* Unused. */
2540 break;
2542 default:
2543 gcc_unreachable ();
2547 return hash;
2550 /* Hash an rtx. We are careful to make sure the value is never negative.
2551 Equivalent registers hash identically.
2552 MODE is used in hashing for CONST_INTs only;
2553 otherwise the mode of X is used.
2555 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2557 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2558 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2560 Note that cse_insn knows that the hash code of a MEM expression
2561 is just (int) MEM plus the hash code of the address. */
2563 unsigned
2564 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2565 int *hash_arg_in_memory_p, bool have_reg_qty)
2567 return hash_rtx_cb (x, mode, do_not_record_p,
2568 hash_arg_in_memory_p, have_reg_qty, NULL);
2571 /* Hash an rtx X for cse via hash_rtx.
2572 Stores 1 in do_not_record if any subexpression is volatile.
2573 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2574 does not have the RTX_UNCHANGING_P bit set. */
2576 static inline unsigned
2577 canon_hash (rtx x, enum machine_mode mode)
2579 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2582 /* Like canon_hash but with no side effects, i.e. do_not_record
2583 and hash_arg_in_memory are not changed. */
2585 static inline unsigned
2586 safe_hash (rtx x, enum machine_mode mode)
2588 int dummy_do_not_record;
2589 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2592 /* Return 1 iff X and Y would canonicalize into the same thing,
2593 without actually constructing the canonicalization of either one.
2594 If VALIDATE is nonzero,
2595 we assume X is an expression being processed from the rtl
2596 and Y was found in the hash table. We check register refs
2597 in Y for being marked as valid.
2599 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2602 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2604 int i, j;
2605 enum rtx_code code;
2606 const char *fmt;
2608 /* Note: it is incorrect to assume an expression is equivalent to itself
2609 if VALIDATE is nonzero. */
2610 if (x == y && !validate)
2611 return 1;
2613 if (x == 0 || y == 0)
2614 return x == y;
2616 code = GET_CODE (x);
2617 if (code != GET_CODE (y))
2618 return 0;
2620 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2621 if (GET_MODE (x) != GET_MODE (y))
2622 return 0;
2624 /* MEMs refering to different address space are not equivalent. */
2625 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2626 return 0;
2628 switch (code)
2630 case PC:
2631 case CC0:
2632 case CONST_INT:
2633 case CONST_DOUBLE:
2634 case CONST_FIXED:
2635 return x == y;
2637 case LABEL_REF:
2638 return XEXP (x, 0) == XEXP (y, 0);
2640 case SYMBOL_REF:
2641 return XSTR (x, 0) == XSTR (y, 0);
2643 case REG:
2644 if (for_gcse)
2645 return REGNO (x) == REGNO (y);
2646 else
2648 unsigned int regno = REGNO (y);
2649 unsigned int i;
2650 unsigned int endregno = END_REGNO (y);
2652 /* If the quantities are not the same, the expressions are not
2653 equivalent. If there are and we are not to validate, they
2654 are equivalent. Otherwise, ensure all regs are up-to-date. */
2656 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2657 return 0;
2659 if (! validate)
2660 return 1;
2662 for (i = regno; i < endregno; i++)
2663 if (REG_IN_TABLE (i) != REG_TICK (i))
2664 return 0;
2666 return 1;
2669 case MEM:
2670 if (for_gcse)
2672 /* A volatile mem should not be considered equivalent to any
2673 other. */
2674 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2675 return 0;
2677 /* Can't merge two expressions in different alias sets, since we
2678 can decide that the expression is transparent in a block when
2679 it isn't, due to it being set with the different alias set.
2681 Also, can't merge two expressions with different MEM_ATTRS.
2682 They could e.g. be two different entities allocated into the
2683 same space on the stack (see e.g. PR25130). In that case, the
2684 MEM addresses can be the same, even though the two MEMs are
2685 absolutely not equivalent.
2687 But because really all MEM attributes should be the same for
2688 equivalent MEMs, we just use the invariant that MEMs that have
2689 the same attributes share the same mem_attrs data structure. */
2690 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2691 return 0;
2693 break;
2695 /* For commutative operations, check both orders. */
2696 case PLUS:
2697 case MULT:
2698 case AND:
2699 case IOR:
2700 case XOR:
2701 case NE:
2702 case EQ:
2703 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2704 validate, for_gcse)
2705 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2706 validate, for_gcse))
2707 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2708 validate, for_gcse)
2709 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2710 validate, for_gcse)));
2712 case ASM_OPERANDS:
2713 /* We don't use the generic code below because we want to
2714 disregard filename and line numbers. */
2716 /* A volatile asm isn't equivalent to any other. */
2717 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2718 return 0;
2720 if (GET_MODE (x) != GET_MODE (y)
2721 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2722 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2723 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2724 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2725 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2726 return 0;
2728 if (ASM_OPERANDS_INPUT_LENGTH (x))
2730 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2731 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2732 ASM_OPERANDS_INPUT (y, i),
2733 validate, for_gcse)
2734 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2735 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2736 return 0;
2739 return 1;
2741 default:
2742 break;
2745 /* Compare the elements. If any pair of corresponding elements
2746 fail to match, return 0 for the whole thing. */
2748 fmt = GET_RTX_FORMAT (code);
2749 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2751 switch (fmt[i])
2753 case 'e':
2754 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2755 validate, for_gcse))
2756 return 0;
2757 break;
2759 case 'E':
2760 if (XVECLEN (x, i) != XVECLEN (y, i))
2761 return 0;
2762 for (j = 0; j < XVECLEN (x, i); j++)
2763 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2764 validate, for_gcse))
2765 return 0;
2766 break;
2768 case 's':
2769 if (strcmp (XSTR (x, i), XSTR (y, i)))
2770 return 0;
2771 break;
2773 case 'i':
2774 if (XINT (x, i) != XINT (y, i))
2775 return 0;
2776 break;
2778 case 'w':
2779 if (XWINT (x, i) != XWINT (y, i))
2780 return 0;
2781 break;
2783 case '0':
2784 case 't':
2785 break;
2787 default:
2788 gcc_unreachable ();
2792 return 1;
2795 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2796 the result if necessary. INSN is as for canon_reg. */
2798 static void
2799 validate_canon_reg (rtx *xloc, rtx insn)
2801 if (*xloc)
2803 rtx new_rtx = canon_reg (*xloc, insn);
2805 /* If replacing pseudo with hard reg or vice versa, ensure the
2806 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2807 gcc_assert (insn && new_rtx);
2808 validate_change (insn, xloc, new_rtx, 1);
2812 /* Canonicalize an expression:
2813 replace each register reference inside it
2814 with the "oldest" equivalent register.
2816 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2817 after we make our substitution. The calls are made with IN_GROUP nonzero
2818 so apply_change_group must be called upon the outermost return from this
2819 function (unless INSN is zero). The result of apply_change_group can
2820 generally be discarded since the changes we are making are optional. */
2822 static rtx
2823 canon_reg (rtx x, rtx insn)
2825 int i;
2826 enum rtx_code code;
2827 const char *fmt;
2829 if (x == 0)
2830 return x;
2832 code = GET_CODE (x);
2833 switch (code)
2835 case PC:
2836 case CC0:
2837 case CONST:
2838 case CONST_INT:
2839 case CONST_DOUBLE:
2840 case CONST_FIXED:
2841 case CONST_VECTOR:
2842 case SYMBOL_REF:
2843 case LABEL_REF:
2844 case ADDR_VEC:
2845 case ADDR_DIFF_VEC:
2846 return x;
2848 case REG:
2850 int first;
2851 int q;
2852 struct qty_table_elem *ent;
2854 /* Never replace a hard reg, because hard regs can appear
2855 in more than one machine mode, and we must preserve the mode
2856 of each occurrence. Also, some hard regs appear in
2857 MEMs that are shared and mustn't be altered. Don't try to
2858 replace any reg that maps to a reg of class NO_REGS. */
2859 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2860 || ! REGNO_QTY_VALID_P (REGNO (x)))
2861 return x;
2863 q = REG_QTY (REGNO (x));
2864 ent = &qty_table[q];
2865 first = ent->first_reg;
2866 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2867 : REGNO_REG_CLASS (first) == NO_REGS ? x
2868 : gen_rtx_REG (ent->mode, first));
2871 default:
2872 break;
2875 fmt = GET_RTX_FORMAT (code);
2876 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2878 int j;
2880 if (fmt[i] == 'e')
2881 validate_canon_reg (&XEXP (x, i), insn);
2882 else if (fmt[i] == 'E')
2883 for (j = 0; j < XVECLEN (x, i); j++)
2884 validate_canon_reg (&XVECEXP (x, i, j), insn);
2887 return x;
2890 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2891 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2892 what values are being compared.
2894 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2895 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2896 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2897 compared to produce cc0.
2899 The return value is the comparison operator and is either the code of
2900 A or the code corresponding to the inverse of the comparison. */
2902 static enum rtx_code
2903 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2904 enum machine_mode *pmode1, enum machine_mode *pmode2)
2906 rtx arg1, arg2;
2908 arg1 = *parg1, arg2 = *parg2;
2910 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2912 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2914 /* Set nonzero when we find something of interest. */
2915 rtx x = 0;
2916 int reverse_code = 0;
2917 struct table_elt *p = 0;
2919 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2920 On machines with CC0, this is the only case that can occur, since
2921 fold_rtx will return the COMPARE or item being compared with zero
2922 when given CC0. */
2924 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2925 x = arg1;
2927 /* If ARG1 is a comparison operator and CODE is testing for
2928 STORE_FLAG_VALUE, get the inner arguments. */
2930 else if (COMPARISON_P (arg1))
2932 #ifdef FLOAT_STORE_FLAG_VALUE
2933 REAL_VALUE_TYPE fsfv;
2934 #endif
2936 if (code == NE
2937 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2938 && code == LT && STORE_FLAG_VALUE == -1)
2939 #ifdef FLOAT_STORE_FLAG_VALUE
2940 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2941 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2942 REAL_VALUE_NEGATIVE (fsfv)))
2943 #endif
2945 x = arg1;
2946 else if (code == EQ
2947 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2948 && code == GE && STORE_FLAG_VALUE == -1)
2949 #ifdef FLOAT_STORE_FLAG_VALUE
2950 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2951 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2952 REAL_VALUE_NEGATIVE (fsfv)))
2953 #endif
2955 x = arg1, reverse_code = 1;
2958 /* ??? We could also check for
2960 (ne (and (eq (...) (const_int 1))) (const_int 0))
2962 and related forms, but let's wait until we see them occurring. */
2964 if (x == 0)
2965 /* Look up ARG1 in the hash table and see if it has an equivalence
2966 that lets us see what is being compared. */
2967 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2968 if (p)
2970 p = p->first_same_value;
2972 /* If what we compare is already known to be constant, that is as
2973 good as it gets.
2974 We need to break the loop in this case, because otherwise we
2975 can have an infinite loop when looking at a reg that is known
2976 to be a constant which is the same as a comparison of a reg
2977 against zero which appears later in the insn stream, which in
2978 turn is constant and the same as the comparison of the first reg
2979 against zero... */
2980 if (p->is_const)
2981 break;
2984 for (; p; p = p->next_same_value)
2986 enum machine_mode inner_mode = GET_MODE (p->exp);
2987 #ifdef FLOAT_STORE_FLAG_VALUE
2988 REAL_VALUE_TYPE fsfv;
2989 #endif
2991 /* If the entry isn't valid, skip it. */
2992 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2993 continue;
2995 /* If it's the same comparison we're already looking at, skip it. */
2996 if (COMPARISON_P (p->exp)
2997 && XEXP (p->exp, 0) == arg1
2998 && XEXP (p->exp, 1) == arg2)
2999 continue;
3001 if (GET_CODE (p->exp) == COMPARE
3002 /* Another possibility is that this machine has a compare insn
3003 that includes the comparison code. In that case, ARG1 would
3004 be equivalent to a comparison operation that would set ARG1 to
3005 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3006 ORIG_CODE is the actual comparison being done; if it is an EQ,
3007 we must reverse ORIG_CODE. On machine with a negative value
3008 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3009 || ((code == NE
3010 || (code == LT
3011 && val_signbit_known_set_p (inner_mode,
3012 STORE_FLAG_VALUE))
3013 #ifdef FLOAT_STORE_FLAG_VALUE
3014 || (code == LT
3015 && SCALAR_FLOAT_MODE_P (inner_mode)
3016 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3017 REAL_VALUE_NEGATIVE (fsfv)))
3018 #endif
3020 && COMPARISON_P (p->exp)))
3022 x = p->exp;
3023 break;
3025 else if ((code == EQ
3026 || (code == GE
3027 && val_signbit_known_set_p (inner_mode,
3028 STORE_FLAG_VALUE))
3029 #ifdef FLOAT_STORE_FLAG_VALUE
3030 || (code == GE
3031 && SCALAR_FLOAT_MODE_P (inner_mode)
3032 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3033 REAL_VALUE_NEGATIVE (fsfv)))
3034 #endif
3036 && COMPARISON_P (p->exp))
3038 reverse_code = 1;
3039 x = p->exp;
3040 break;
3043 /* If this non-trapping address, e.g. fp + constant, the
3044 equivalent is a better operand since it may let us predict
3045 the value of the comparison. */
3046 else if (!rtx_addr_can_trap_p (p->exp))
3048 arg1 = p->exp;
3049 continue;
3053 /* If we didn't find a useful equivalence for ARG1, we are done.
3054 Otherwise, set up for the next iteration. */
3055 if (x == 0)
3056 break;
3058 /* If we need to reverse the comparison, make sure that that is
3059 possible -- we can't necessarily infer the value of GE from LT
3060 with floating-point operands. */
3061 if (reverse_code)
3063 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3064 if (reversed == UNKNOWN)
3065 break;
3066 else
3067 code = reversed;
3069 else if (COMPARISON_P (x))
3070 code = GET_CODE (x);
3071 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3074 /* Return our results. Return the modes from before fold_rtx
3075 because fold_rtx might produce const_int, and then it's too late. */
3076 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3077 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3079 return code;
3082 /* If X is a nontrivial arithmetic operation on an argument for which
3083 a constant value can be determined, return the result of operating
3084 on that value, as a constant. Otherwise, return X, possibly with
3085 one or more operands changed to a forward-propagated constant.
3087 If X is a register whose contents are known, we do NOT return
3088 those contents here; equiv_constant is called to perform that task.
3089 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3091 INSN is the insn that we may be modifying. If it is 0, make a copy
3092 of X before modifying it. */
3094 static rtx
3095 fold_rtx (rtx x, rtx insn)
3097 enum rtx_code code;
3098 enum machine_mode mode;
3099 const char *fmt;
3100 int i;
3101 rtx new_rtx = 0;
3102 int changed = 0;
3104 /* Operands of X. */
3105 rtx folded_arg0;
3106 rtx folded_arg1;
3108 /* Constant equivalents of first three operands of X;
3109 0 when no such equivalent is known. */
3110 rtx const_arg0;
3111 rtx const_arg1;
3112 rtx const_arg2;
3114 /* The mode of the first operand of X. We need this for sign and zero
3115 extends. */
3116 enum machine_mode mode_arg0;
3118 if (x == 0)
3119 return x;
3121 /* Try to perform some initial simplifications on X. */
3122 code = GET_CODE (x);
3123 switch (code)
3125 case MEM:
3126 case SUBREG:
3127 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3128 return new_rtx;
3129 return x;
3131 case CONST:
3132 case CONST_INT:
3133 case CONST_DOUBLE:
3134 case CONST_FIXED:
3135 case CONST_VECTOR:
3136 case SYMBOL_REF:
3137 case LABEL_REF:
3138 case REG:
3139 case PC:
3140 /* No use simplifying an EXPR_LIST
3141 since they are used only for lists of args
3142 in a function call's REG_EQUAL note. */
3143 case EXPR_LIST:
3144 return x;
3146 #ifdef HAVE_cc0
3147 case CC0:
3148 return prev_insn_cc0;
3149 #endif
3151 case ASM_OPERANDS:
3152 if (insn)
3154 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3155 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3156 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3158 return x;
3160 #ifdef NO_FUNCTION_CSE
3161 case CALL:
3162 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3163 return x;
3164 break;
3165 #endif
3167 /* Anything else goes through the loop below. */
3168 default:
3169 break;
3172 mode = GET_MODE (x);
3173 const_arg0 = 0;
3174 const_arg1 = 0;
3175 const_arg2 = 0;
3176 mode_arg0 = VOIDmode;
3178 /* Try folding our operands.
3179 Then see which ones have constant values known. */
3181 fmt = GET_RTX_FORMAT (code);
3182 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3183 if (fmt[i] == 'e')
3185 rtx folded_arg = XEXP (x, i), const_arg;
3186 enum machine_mode mode_arg = GET_MODE (folded_arg);
3188 switch (GET_CODE (folded_arg))
3190 case MEM:
3191 case REG:
3192 case SUBREG:
3193 const_arg = equiv_constant (folded_arg);
3194 break;
3196 case CONST:
3197 case CONST_INT:
3198 case SYMBOL_REF:
3199 case LABEL_REF:
3200 case CONST_DOUBLE:
3201 case CONST_FIXED:
3202 case CONST_VECTOR:
3203 const_arg = folded_arg;
3204 break;
3206 #ifdef HAVE_cc0
3207 case CC0:
3208 folded_arg = prev_insn_cc0;
3209 mode_arg = prev_insn_cc0_mode;
3210 const_arg = equiv_constant (folded_arg);
3211 break;
3212 #endif
3214 default:
3215 folded_arg = fold_rtx (folded_arg, insn);
3216 const_arg = equiv_constant (folded_arg);
3217 break;
3220 /* For the first three operands, see if the operand
3221 is constant or equivalent to a constant. */
3222 switch (i)
3224 case 0:
3225 folded_arg0 = folded_arg;
3226 const_arg0 = const_arg;
3227 mode_arg0 = mode_arg;
3228 break;
3229 case 1:
3230 folded_arg1 = folded_arg;
3231 const_arg1 = const_arg;
3232 break;
3233 case 2:
3234 const_arg2 = const_arg;
3235 break;
3238 /* Pick the least expensive of the argument and an equivalent constant
3239 argument. */
3240 if (const_arg != 0
3241 && const_arg != folded_arg
3242 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3244 /* It's not safe to substitute the operand of a conversion
3245 operator with a constant, as the conversion's identity
3246 depends upon the mode of its operand. This optimization
3247 is handled by the call to simplify_unary_operation. */
3248 && (GET_RTX_CLASS (code) != RTX_UNARY
3249 || GET_MODE (const_arg) == mode_arg0
3250 || (code != ZERO_EXTEND
3251 && code != SIGN_EXTEND
3252 && code != TRUNCATE
3253 && code != FLOAT_TRUNCATE
3254 && code != FLOAT_EXTEND
3255 && code != FLOAT
3256 && code != FIX
3257 && code != UNSIGNED_FLOAT
3258 && code != UNSIGNED_FIX)))
3259 folded_arg = const_arg;
3261 if (folded_arg == XEXP (x, i))
3262 continue;
3264 if (insn == NULL_RTX && !changed)
3265 x = copy_rtx (x);
3266 changed = 1;
3267 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3270 if (changed)
3272 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3273 consistent with the order in X. */
3274 if (canonicalize_change_group (insn, x))
3276 rtx tem;
3277 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3278 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3281 apply_change_group ();
3284 /* If X is an arithmetic operation, see if we can simplify it. */
3286 switch (GET_RTX_CLASS (code))
3288 case RTX_UNARY:
3290 /* We can't simplify extension ops unless we know the
3291 original mode. */
3292 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3293 && mode_arg0 == VOIDmode)
3294 break;
3296 new_rtx = simplify_unary_operation (code, mode,
3297 const_arg0 ? const_arg0 : folded_arg0,
3298 mode_arg0);
3300 break;
3302 case RTX_COMPARE:
3303 case RTX_COMM_COMPARE:
3304 /* See what items are actually being compared and set FOLDED_ARG[01]
3305 to those values and CODE to the actual comparison code. If any are
3306 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3307 do anything if both operands are already known to be constant. */
3309 /* ??? Vector mode comparisons are not supported yet. */
3310 if (VECTOR_MODE_P (mode))
3311 break;
3313 if (const_arg0 == 0 || const_arg1 == 0)
3315 struct table_elt *p0, *p1;
3316 rtx true_rtx, false_rtx;
3317 enum machine_mode mode_arg1;
3319 if (SCALAR_FLOAT_MODE_P (mode))
3321 #ifdef FLOAT_STORE_FLAG_VALUE
3322 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3323 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3324 #else
3325 true_rtx = NULL_RTX;
3326 #endif
3327 false_rtx = CONST0_RTX (mode);
3329 else
3331 true_rtx = const_true_rtx;
3332 false_rtx = const0_rtx;
3335 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3336 &mode_arg0, &mode_arg1);
3338 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3339 what kinds of things are being compared, so we can't do
3340 anything with this comparison. */
3342 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3343 break;
3345 const_arg0 = equiv_constant (folded_arg0);
3346 const_arg1 = equiv_constant (folded_arg1);
3348 /* If we do not now have two constants being compared, see
3349 if we can nevertheless deduce some things about the
3350 comparison. */
3351 if (const_arg0 == 0 || const_arg1 == 0)
3353 if (const_arg1 != NULL)
3355 rtx cheapest_simplification;
3356 int cheapest_cost;
3357 rtx simp_result;
3358 struct table_elt *p;
3360 /* See if we can find an equivalent of folded_arg0
3361 that gets us a cheaper expression, possibly a
3362 constant through simplifications. */
3363 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3364 mode_arg0);
3366 if (p != NULL)
3368 cheapest_simplification = x;
3369 cheapest_cost = COST (x);
3371 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3373 int cost;
3375 /* If the entry isn't valid, skip it. */
3376 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3377 continue;
3379 /* Try to simplify using this equivalence. */
3380 simp_result
3381 = simplify_relational_operation (code, mode,
3382 mode_arg0,
3383 p->exp,
3384 const_arg1);
3386 if (simp_result == NULL)
3387 continue;
3389 cost = COST (simp_result);
3390 if (cost < cheapest_cost)
3392 cheapest_cost = cost;
3393 cheapest_simplification = simp_result;
3397 /* If we have a cheaper expression now, use that
3398 and try folding it further, from the top. */
3399 if (cheapest_simplification != x)
3400 return fold_rtx (copy_rtx (cheapest_simplification),
3401 insn);
3405 /* See if the two operands are the same. */
3407 if ((REG_P (folded_arg0)
3408 && REG_P (folded_arg1)
3409 && (REG_QTY (REGNO (folded_arg0))
3410 == REG_QTY (REGNO (folded_arg1))))
3411 || ((p0 = lookup (folded_arg0,
3412 SAFE_HASH (folded_arg0, mode_arg0),
3413 mode_arg0))
3414 && (p1 = lookup (folded_arg1,
3415 SAFE_HASH (folded_arg1, mode_arg0),
3416 mode_arg0))
3417 && p0->first_same_value == p1->first_same_value))
3418 folded_arg1 = folded_arg0;
3420 /* If FOLDED_ARG0 is a register, see if the comparison we are
3421 doing now is either the same as we did before or the reverse
3422 (we only check the reverse if not floating-point). */
3423 else if (REG_P (folded_arg0))
3425 int qty = REG_QTY (REGNO (folded_arg0));
3427 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3429 struct qty_table_elem *ent = &qty_table[qty];
3431 if ((comparison_dominates_p (ent->comparison_code, code)
3432 || (! FLOAT_MODE_P (mode_arg0)
3433 && comparison_dominates_p (ent->comparison_code,
3434 reverse_condition (code))))
3435 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3436 || (const_arg1
3437 && rtx_equal_p (ent->comparison_const,
3438 const_arg1))
3439 || (REG_P (folded_arg1)
3440 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3442 if (comparison_dominates_p (ent->comparison_code, code))
3444 if (true_rtx)
3445 return true_rtx;
3446 else
3447 break;
3449 else
3450 return false_rtx;
3457 /* If we are comparing against zero, see if the first operand is
3458 equivalent to an IOR with a constant. If so, we may be able to
3459 determine the result of this comparison. */
3460 if (const_arg1 == const0_rtx && !const_arg0)
3462 rtx y = lookup_as_function (folded_arg0, IOR);
3463 rtx inner_const;
3465 if (y != 0
3466 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3467 && CONST_INT_P (inner_const)
3468 && INTVAL (inner_const) != 0)
3469 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3473 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3474 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3475 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3477 break;
3479 case RTX_BIN_ARITH:
3480 case RTX_COMM_ARITH:
3481 switch (code)
3483 case PLUS:
3484 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3485 with that LABEL_REF as its second operand. If so, the result is
3486 the first operand of that MINUS. This handles switches with an
3487 ADDR_DIFF_VEC table. */
3488 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3490 rtx y
3491 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3492 : lookup_as_function (folded_arg0, MINUS);
3494 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3495 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3496 return XEXP (y, 0);
3498 /* Now try for a CONST of a MINUS like the above. */
3499 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3500 : lookup_as_function (folded_arg0, CONST))) != 0
3501 && GET_CODE (XEXP (y, 0)) == MINUS
3502 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3503 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3504 return XEXP (XEXP (y, 0), 0);
3507 /* Likewise if the operands are in the other order. */
3508 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3510 rtx y
3511 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3512 : lookup_as_function (folded_arg1, MINUS);
3514 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3515 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3516 return XEXP (y, 0);
3518 /* Now try for a CONST of a MINUS like the above. */
3519 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3520 : lookup_as_function (folded_arg1, CONST))) != 0
3521 && GET_CODE (XEXP (y, 0)) == MINUS
3522 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3523 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3524 return XEXP (XEXP (y, 0), 0);
3527 /* If second operand is a register equivalent to a negative
3528 CONST_INT, see if we can find a register equivalent to the
3529 positive constant. Make a MINUS if so. Don't do this for
3530 a non-negative constant since we might then alternate between
3531 choosing positive and negative constants. Having the positive
3532 constant previously-used is the more common case. Be sure
3533 the resulting constant is non-negative; if const_arg1 were
3534 the smallest negative number this would overflow: depending
3535 on the mode, this would either just be the same value (and
3536 hence not save anything) or be incorrect. */
3537 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3538 && INTVAL (const_arg1) < 0
3539 /* This used to test
3541 -INTVAL (const_arg1) >= 0
3543 But The Sun V5.0 compilers mis-compiled that test. So
3544 instead we test for the problematic value in a more direct
3545 manner and hope the Sun compilers get it correct. */
3546 && INTVAL (const_arg1) !=
3547 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3548 && REG_P (folded_arg1))
3550 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3551 struct table_elt *p
3552 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3554 if (p)
3555 for (p = p->first_same_value; p; p = p->next_same_value)
3556 if (REG_P (p->exp))
3557 return simplify_gen_binary (MINUS, mode, folded_arg0,
3558 canon_reg (p->exp, NULL_RTX));
3560 goto from_plus;
3562 case MINUS:
3563 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3564 If so, produce (PLUS Z C2-C). */
3565 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3567 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3568 if (y && CONST_INT_P (XEXP (y, 1)))
3569 return fold_rtx (plus_constant (copy_rtx (y),
3570 -INTVAL (const_arg1)),
3571 NULL_RTX);
3574 /* Fall through. */
3576 from_plus:
3577 case SMIN: case SMAX: case UMIN: case UMAX:
3578 case IOR: case AND: case XOR:
3579 case MULT:
3580 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3581 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3582 is known to be of similar form, we may be able to replace the
3583 operation with a combined operation. This may eliminate the
3584 intermediate operation if every use is simplified in this way.
3585 Note that the similar optimization done by combine.c only works
3586 if the intermediate operation's result has only one reference. */
3588 if (REG_P (folded_arg0)
3589 && const_arg1 && CONST_INT_P (const_arg1))
3591 int is_shift
3592 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3593 rtx y, inner_const, new_const;
3594 rtx canon_const_arg1 = const_arg1;
3595 enum rtx_code associate_code;
3597 if (is_shift
3598 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3599 || INTVAL (const_arg1) < 0))
3601 if (SHIFT_COUNT_TRUNCATED)
3602 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3603 & (GET_MODE_BITSIZE (mode)
3604 - 1));
3605 else
3606 break;
3609 y = lookup_as_function (folded_arg0, code);
3610 if (y == 0)
3611 break;
3613 /* If we have compiled a statement like
3614 "if (x == (x & mask1))", and now are looking at
3615 "x & mask2", we will have a case where the first operand
3616 of Y is the same as our first operand. Unless we detect
3617 this case, an infinite loop will result. */
3618 if (XEXP (y, 0) == folded_arg0)
3619 break;
3621 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3622 if (!inner_const || !CONST_INT_P (inner_const))
3623 break;
3625 /* Don't associate these operations if they are a PLUS with the
3626 same constant and it is a power of two. These might be doable
3627 with a pre- or post-increment. Similarly for two subtracts of
3628 identical powers of two with post decrement. */
3630 if (code == PLUS && const_arg1 == inner_const
3631 && ((HAVE_PRE_INCREMENT
3632 && exact_log2 (INTVAL (const_arg1)) >= 0)
3633 || (HAVE_POST_INCREMENT
3634 && exact_log2 (INTVAL (const_arg1)) >= 0)
3635 || (HAVE_PRE_DECREMENT
3636 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3637 || (HAVE_POST_DECREMENT
3638 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3639 break;
3641 /* ??? Vector mode shifts by scalar
3642 shift operand are not supported yet. */
3643 if (is_shift && VECTOR_MODE_P (mode))
3644 break;
3646 if (is_shift
3647 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3648 || INTVAL (inner_const) < 0))
3650 if (SHIFT_COUNT_TRUNCATED)
3651 inner_const = GEN_INT (INTVAL (inner_const)
3652 & (GET_MODE_BITSIZE (mode) - 1));
3653 else
3654 break;
3657 /* Compute the code used to compose the constants. For example,
3658 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3660 associate_code = (is_shift || code == MINUS ? PLUS : code);
3662 new_const = simplify_binary_operation (associate_code, mode,
3663 canon_const_arg1,
3664 inner_const);
3666 if (new_const == 0)
3667 break;
3669 /* If we are associating shift operations, don't let this
3670 produce a shift of the size of the object or larger.
3671 This could occur when we follow a sign-extend by a right
3672 shift on a machine that does a sign-extend as a pair
3673 of shifts. */
3675 if (is_shift
3676 && CONST_INT_P (new_const)
3677 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3679 /* As an exception, we can turn an ASHIFTRT of this
3680 form into a shift of the number of bits - 1. */
3681 if (code == ASHIFTRT)
3682 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3683 else if (!side_effects_p (XEXP (y, 0)))
3684 return CONST0_RTX (mode);
3685 else
3686 break;
3689 y = copy_rtx (XEXP (y, 0));
3691 /* If Y contains our first operand (the most common way this
3692 can happen is if Y is a MEM), we would do into an infinite
3693 loop if we tried to fold it. So don't in that case. */
3695 if (! reg_mentioned_p (folded_arg0, y))
3696 y = fold_rtx (y, insn);
3698 return simplify_gen_binary (code, mode, y, new_const);
3700 break;
3702 case DIV: case UDIV:
3703 /* ??? The associative optimization performed immediately above is
3704 also possible for DIV and UDIV using associate_code of MULT.
3705 However, we would need extra code to verify that the
3706 multiplication does not overflow, that is, there is no overflow
3707 in the calculation of new_const. */
3708 break;
3710 default:
3711 break;
3714 new_rtx = simplify_binary_operation (code, mode,
3715 const_arg0 ? const_arg0 : folded_arg0,
3716 const_arg1 ? const_arg1 : folded_arg1);
3717 break;
3719 case RTX_OBJ:
3720 /* (lo_sum (high X) X) is simply X. */
3721 if (code == LO_SUM && const_arg0 != 0
3722 && GET_CODE (const_arg0) == HIGH
3723 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3724 return const_arg1;
3725 break;
3727 case RTX_TERNARY:
3728 case RTX_BITFIELD_OPS:
3729 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3730 const_arg0 ? const_arg0 : folded_arg0,
3731 const_arg1 ? const_arg1 : folded_arg1,
3732 const_arg2 ? const_arg2 : XEXP (x, 2));
3733 break;
3735 default:
3736 break;
3739 return new_rtx ? new_rtx : x;
3742 /* Return a constant value currently equivalent to X.
3743 Return 0 if we don't know one. */
3745 static rtx
3746 equiv_constant (rtx x)
3748 if (REG_P (x)
3749 && REGNO_QTY_VALID_P (REGNO (x)))
3751 int x_q = REG_QTY (REGNO (x));
3752 struct qty_table_elem *x_ent = &qty_table[x_q];
3754 if (x_ent->const_rtx)
3755 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3758 if (x == 0 || CONSTANT_P (x))
3759 return x;
3761 if (GET_CODE (x) == SUBREG)
3763 enum machine_mode mode = GET_MODE (x);
3764 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3765 rtx new_rtx;
3767 /* See if we previously assigned a constant value to this SUBREG. */
3768 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3769 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3770 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3771 return new_rtx;
3773 /* If we didn't and if doing so makes sense, see if we previously
3774 assigned a constant value to the enclosing word mode SUBREG. */
3775 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3776 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3778 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3779 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3781 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3782 new_rtx = lookup_as_function (y, CONST_INT);
3783 if (new_rtx)
3784 return gen_lowpart (mode, new_rtx);
3788 /* Otherwise see if we already have a constant for the inner REG. */
3789 if (REG_P (SUBREG_REG (x))
3790 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3791 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3793 return 0;
3796 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3797 the hash table in case its value was seen before. */
3799 if (MEM_P (x))
3801 struct table_elt *elt;
3803 x = avoid_constant_pool_reference (x);
3804 if (CONSTANT_P (x))
3805 return x;
3807 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3808 if (elt == 0)
3809 return 0;
3811 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3812 if (elt->is_const && CONSTANT_P (elt->exp))
3813 return elt->exp;
3816 return 0;
3819 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3820 "taken" branch.
3822 In certain cases, this can cause us to add an equivalence. For example,
3823 if we are following the taken case of
3824 if (i == 2)
3825 we can add the fact that `i' and '2' are now equivalent.
3827 In any case, we can record that this comparison was passed. If the same
3828 comparison is seen later, we will know its value. */
3830 static void
3831 record_jump_equiv (rtx insn, bool taken)
3833 int cond_known_true;
3834 rtx op0, op1;
3835 rtx set;
3836 enum machine_mode mode, mode0, mode1;
3837 int reversed_nonequality = 0;
3838 enum rtx_code code;
3840 /* Ensure this is the right kind of insn. */
3841 gcc_assert (any_condjump_p (insn));
3843 set = pc_set (insn);
3845 /* See if this jump condition is known true or false. */
3846 if (taken)
3847 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3848 else
3849 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3851 /* Get the type of comparison being done and the operands being compared.
3852 If we had to reverse a non-equality condition, record that fact so we
3853 know that it isn't valid for floating-point. */
3854 code = GET_CODE (XEXP (SET_SRC (set), 0));
3855 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3856 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3858 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3859 if (! cond_known_true)
3861 code = reversed_comparison_code_parts (code, op0, op1, insn);
3863 /* Don't remember if we can't find the inverse. */
3864 if (code == UNKNOWN)
3865 return;
3868 /* The mode is the mode of the non-constant. */
3869 mode = mode0;
3870 if (mode1 != VOIDmode)
3871 mode = mode1;
3873 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3876 /* Yet another form of subreg creation. In this case, we want something in
3877 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3879 static rtx
3880 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3882 enum machine_mode op_mode = GET_MODE (op);
3883 if (op_mode == mode || op_mode == VOIDmode)
3884 return op;
3885 return lowpart_subreg (mode, op, op_mode);
3888 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3889 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3890 Make any useful entries we can with that information. Called from
3891 above function and called recursively. */
3893 static void
3894 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3895 rtx op1, int reversed_nonequality)
3897 unsigned op0_hash, op1_hash;
3898 int op0_in_memory, op1_in_memory;
3899 struct table_elt *op0_elt, *op1_elt;
3901 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3902 we know that they are also equal in the smaller mode (this is also
3903 true for all smaller modes whether or not there is a SUBREG, but
3904 is not worth testing for with no SUBREG). */
3906 /* Note that GET_MODE (op0) may not equal MODE. */
3907 if (code == EQ && paradoxical_subreg_p (op0))
3909 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3910 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3911 if (tem)
3912 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3913 reversed_nonequality);
3916 if (code == EQ && paradoxical_subreg_p (op1))
3918 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3919 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3920 if (tem)
3921 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3922 reversed_nonequality);
3925 /* Similarly, if this is an NE comparison, and either is a SUBREG
3926 making a smaller mode, we know the whole thing is also NE. */
3928 /* Note that GET_MODE (op0) may not equal MODE;
3929 if we test MODE instead, we can get an infinite recursion
3930 alternating between two modes each wider than MODE. */
3932 if (code == NE && GET_CODE (op0) == SUBREG
3933 && subreg_lowpart_p (op0)
3934 && (GET_MODE_SIZE (GET_MODE (op0))
3935 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3937 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3938 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3939 if (tem)
3940 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3941 reversed_nonequality);
3944 if (code == NE && GET_CODE (op1) == SUBREG
3945 && subreg_lowpart_p (op1)
3946 && (GET_MODE_SIZE (GET_MODE (op1))
3947 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3949 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3950 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3951 if (tem)
3952 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3953 reversed_nonequality);
3956 /* Hash both operands. */
3958 do_not_record = 0;
3959 hash_arg_in_memory = 0;
3960 op0_hash = HASH (op0, mode);
3961 op0_in_memory = hash_arg_in_memory;
3963 if (do_not_record)
3964 return;
3966 do_not_record = 0;
3967 hash_arg_in_memory = 0;
3968 op1_hash = HASH (op1, mode);
3969 op1_in_memory = hash_arg_in_memory;
3971 if (do_not_record)
3972 return;
3974 /* Look up both operands. */
3975 op0_elt = lookup (op0, op0_hash, mode);
3976 op1_elt = lookup (op1, op1_hash, mode);
3978 /* If both operands are already equivalent or if they are not in the
3979 table but are identical, do nothing. */
3980 if ((op0_elt != 0 && op1_elt != 0
3981 && op0_elt->first_same_value == op1_elt->first_same_value)
3982 || op0 == op1 || rtx_equal_p (op0, op1))
3983 return;
3985 /* If we aren't setting two things equal all we can do is save this
3986 comparison. Similarly if this is floating-point. In the latter
3987 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
3988 If we record the equality, we might inadvertently delete code
3989 whose intent was to change -0 to +0. */
3991 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
3993 struct qty_table_elem *ent;
3994 int qty;
3996 /* If we reversed a floating-point comparison, if OP0 is not a
3997 register, or if OP1 is neither a register or constant, we can't
3998 do anything. */
4000 if (!REG_P (op1))
4001 op1 = equiv_constant (op1);
4003 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4004 || !REG_P (op0) || op1 == 0)
4005 return;
4007 /* Put OP0 in the hash table if it isn't already. This gives it a
4008 new quantity number. */
4009 if (op0_elt == 0)
4011 if (insert_regs (op0, NULL, 0))
4013 rehash_using_reg (op0);
4014 op0_hash = HASH (op0, mode);
4016 /* If OP0 is contained in OP1, this changes its hash code
4017 as well. Faster to rehash than to check, except
4018 for the simple case of a constant. */
4019 if (! CONSTANT_P (op1))
4020 op1_hash = HASH (op1,mode);
4023 op0_elt = insert (op0, NULL, op0_hash, mode);
4024 op0_elt->in_memory = op0_in_memory;
4027 qty = REG_QTY (REGNO (op0));
4028 ent = &qty_table[qty];
4030 ent->comparison_code = code;
4031 if (REG_P (op1))
4033 /* Look it up again--in case op0 and op1 are the same. */
4034 op1_elt = lookup (op1, op1_hash, mode);
4036 /* Put OP1 in the hash table so it gets a new quantity number. */
4037 if (op1_elt == 0)
4039 if (insert_regs (op1, NULL, 0))
4041 rehash_using_reg (op1);
4042 op1_hash = HASH (op1, mode);
4045 op1_elt = insert (op1, NULL, op1_hash, mode);
4046 op1_elt->in_memory = op1_in_memory;
4049 ent->comparison_const = NULL_RTX;
4050 ent->comparison_qty = REG_QTY (REGNO (op1));
4052 else
4054 ent->comparison_const = op1;
4055 ent->comparison_qty = -1;
4058 return;
4061 /* If either side is still missing an equivalence, make it now,
4062 then merge the equivalences. */
4064 if (op0_elt == 0)
4066 if (insert_regs (op0, NULL, 0))
4068 rehash_using_reg (op0);
4069 op0_hash = HASH (op0, mode);
4072 op0_elt = insert (op0, NULL, op0_hash, mode);
4073 op0_elt->in_memory = op0_in_memory;
4076 if (op1_elt == 0)
4078 if (insert_regs (op1, NULL, 0))
4080 rehash_using_reg (op1);
4081 op1_hash = HASH (op1, mode);
4084 op1_elt = insert (op1, NULL, op1_hash, mode);
4085 op1_elt->in_memory = op1_in_memory;
4088 merge_equiv_classes (op0_elt, op1_elt);
4091 /* CSE processing for one instruction.
4092 First simplify sources and addresses of all assignments
4093 in the instruction, using previously-computed equivalents values.
4094 Then install the new sources and destinations in the table
4095 of available values. */
4097 /* Data on one SET contained in the instruction. */
4099 struct set
4101 /* The SET rtx itself. */
4102 rtx rtl;
4103 /* The SET_SRC of the rtx (the original value, if it is changing). */
4104 rtx src;
4105 /* The hash-table element for the SET_SRC of the SET. */
4106 struct table_elt *src_elt;
4107 /* Hash value for the SET_SRC. */
4108 unsigned src_hash;
4109 /* Hash value for the SET_DEST. */
4110 unsigned dest_hash;
4111 /* The SET_DEST, with SUBREG, etc., stripped. */
4112 rtx inner_dest;
4113 /* Nonzero if the SET_SRC is in memory. */
4114 char src_in_memory;
4115 /* Nonzero if the SET_SRC contains something
4116 whose value cannot be predicted and understood. */
4117 char src_volatile;
4118 /* Original machine mode, in case it becomes a CONST_INT.
4119 The size of this field should match the size of the mode
4120 field of struct rtx_def (see rtl.h). */
4121 ENUM_BITFIELD(machine_mode) mode : 8;
4122 /* A constant equivalent for SET_SRC, if any. */
4123 rtx src_const;
4124 /* Hash value of constant equivalent for SET_SRC. */
4125 unsigned src_const_hash;
4126 /* Table entry for constant equivalent for SET_SRC, if any. */
4127 struct table_elt *src_const_elt;
4128 /* Table entry for the destination address. */
4129 struct table_elt *dest_addr_elt;
4132 static void
4133 cse_insn (rtx insn)
4135 rtx x = PATTERN (insn);
4136 int i;
4137 rtx tem;
4138 int n_sets = 0;
4140 rtx src_eqv = 0;
4141 struct table_elt *src_eqv_elt = 0;
4142 int src_eqv_volatile = 0;
4143 int src_eqv_in_memory = 0;
4144 unsigned src_eqv_hash = 0;
4146 struct set *sets = (struct set *) 0;
4148 this_insn = insn;
4149 #ifdef HAVE_cc0
4150 /* Records what this insn does to set CC0. */
4151 this_insn_cc0 = 0;
4152 this_insn_cc0_mode = VOIDmode;
4153 #endif
4155 /* Find all the SETs and CLOBBERs in this instruction.
4156 Record all the SETs in the array `set' and count them.
4157 Also determine whether there is a CLOBBER that invalidates
4158 all memory references, or all references at varying addresses. */
4160 if (CALL_P (insn))
4162 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4164 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4165 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4166 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4170 if (GET_CODE (x) == SET)
4172 sets = XALLOCA (struct set);
4173 sets[0].rtl = x;
4175 /* Ignore SETs that are unconditional jumps.
4176 They never need cse processing, so this does not hurt.
4177 The reason is not efficiency but rather
4178 so that we can test at the end for instructions
4179 that have been simplified to unconditional jumps
4180 and not be misled by unchanged instructions
4181 that were unconditional jumps to begin with. */
4182 if (SET_DEST (x) == pc_rtx
4183 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4186 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4187 The hard function value register is used only once, to copy to
4188 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4189 Ensure we invalidate the destination register. On the 80386 no
4190 other code would invalidate it since it is a fixed_reg.
4191 We need not check the return of apply_change_group; see canon_reg. */
4193 else if (GET_CODE (SET_SRC (x)) == CALL)
4195 canon_reg (SET_SRC (x), insn);
4196 apply_change_group ();
4197 fold_rtx (SET_SRC (x), insn);
4198 invalidate (SET_DEST (x), VOIDmode);
4200 else
4201 n_sets = 1;
4203 else if (GET_CODE (x) == PARALLEL)
4205 int lim = XVECLEN (x, 0);
4207 sets = XALLOCAVEC (struct set, lim);
4209 /* Find all regs explicitly clobbered in this insn,
4210 and ensure they are not replaced with any other regs
4211 elsewhere in this insn.
4212 When a reg that is clobbered is also used for input,
4213 we should presume that that is for a reason,
4214 and we should not substitute some other register
4215 which is not supposed to be clobbered.
4216 Therefore, this loop cannot be merged into the one below
4217 because a CALL may precede a CLOBBER and refer to the
4218 value clobbered. We must not let a canonicalization do
4219 anything in that case. */
4220 for (i = 0; i < lim; i++)
4222 rtx y = XVECEXP (x, 0, i);
4223 if (GET_CODE (y) == CLOBBER)
4225 rtx clobbered = XEXP (y, 0);
4227 if (REG_P (clobbered)
4228 || GET_CODE (clobbered) == SUBREG)
4229 invalidate (clobbered, VOIDmode);
4230 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4231 || GET_CODE (clobbered) == ZERO_EXTRACT)
4232 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4236 for (i = 0; i < lim; i++)
4238 rtx y = XVECEXP (x, 0, i);
4239 if (GET_CODE (y) == SET)
4241 /* As above, we ignore unconditional jumps and call-insns and
4242 ignore the result of apply_change_group. */
4243 if (GET_CODE (SET_SRC (y)) == CALL)
4245 canon_reg (SET_SRC (y), insn);
4246 apply_change_group ();
4247 fold_rtx (SET_SRC (y), insn);
4248 invalidate (SET_DEST (y), VOIDmode);
4250 else if (SET_DEST (y) == pc_rtx
4251 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4253 else
4254 sets[n_sets++].rtl = y;
4256 else if (GET_CODE (y) == CLOBBER)
4258 /* If we clobber memory, canon the address.
4259 This does nothing when a register is clobbered
4260 because we have already invalidated the reg. */
4261 if (MEM_P (XEXP (y, 0)))
4262 canon_reg (XEXP (y, 0), insn);
4264 else if (GET_CODE (y) == USE
4265 && ! (REG_P (XEXP (y, 0))
4266 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4267 canon_reg (y, insn);
4268 else if (GET_CODE (y) == CALL)
4270 /* The result of apply_change_group can be ignored; see
4271 canon_reg. */
4272 canon_reg (y, insn);
4273 apply_change_group ();
4274 fold_rtx (y, insn);
4278 else if (GET_CODE (x) == CLOBBER)
4280 if (MEM_P (XEXP (x, 0)))
4281 canon_reg (XEXP (x, 0), insn);
4283 /* Canonicalize a USE of a pseudo register or memory location. */
4284 else if (GET_CODE (x) == USE
4285 && ! (REG_P (XEXP (x, 0))
4286 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4287 canon_reg (x, insn);
4288 else if (GET_CODE (x) == ASM_OPERANDS)
4290 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4292 rtx input = ASM_OPERANDS_INPUT (x, i);
4293 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4295 input = canon_reg (input, insn);
4296 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4300 else if (GET_CODE (x) == CALL)
4302 /* The result of apply_change_group can be ignored; see canon_reg. */
4303 canon_reg (x, insn);
4304 apply_change_group ();
4305 fold_rtx (x, insn);
4307 else if (DEBUG_INSN_P (insn))
4308 canon_reg (PATTERN (insn), insn);
4310 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4311 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4312 is handled specially for this case, and if it isn't set, then there will
4313 be no equivalence for the destination. */
4314 if (n_sets == 1 && REG_NOTES (insn) != 0
4315 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4316 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4317 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4319 /* The result of apply_change_group can be ignored; see canon_reg. */
4320 canon_reg (XEXP (tem, 0), insn);
4321 apply_change_group ();
4322 src_eqv = fold_rtx (XEXP (tem, 0), insn);
4323 XEXP (tem, 0) = copy_rtx (src_eqv);
4324 df_notes_rescan (insn);
4327 /* Canonicalize sources and addresses of destinations.
4328 We do this in a separate pass to avoid problems when a MATCH_DUP is
4329 present in the insn pattern. In that case, we want to ensure that
4330 we don't break the duplicate nature of the pattern. So we will replace
4331 both operands at the same time. Otherwise, we would fail to find an
4332 equivalent substitution in the loop calling validate_change below.
4334 We used to suppress canonicalization of DEST if it appears in SRC,
4335 but we don't do this any more. */
4337 for (i = 0; i < n_sets; i++)
4339 rtx dest = SET_DEST (sets[i].rtl);
4340 rtx src = SET_SRC (sets[i].rtl);
4341 rtx new_rtx = canon_reg (src, insn);
4343 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4345 if (GET_CODE (dest) == ZERO_EXTRACT)
4347 validate_change (insn, &XEXP (dest, 1),
4348 canon_reg (XEXP (dest, 1), insn), 1);
4349 validate_change (insn, &XEXP (dest, 2),
4350 canon_reg (XEXP (dest, 2), insn), 1);
4353 while (GET_CODE (dest) == SUBREG
4354 || GET_CODE (dest) == ZERO_EXTRACT
4355 || GET_CODE (dest) == STRICT_LOW_PART)
4356 dest = XEXP (dest, 0);
4358 if (MEM_P (dest))
4359 canon_reg (dest, insn);
4362 /* Now that we have done all the replacements, we can apply the change
4363 group and see if they all work. Note that this will cause some
4364 canonicalizations that would have worked individually not to be applied
4365 because some other canonicalization didn't work, but this should not
4366 occur often.
4368 The result of apply_change_group can be ignored; see canon_reg. */
4370 apply_change_group ();
4372 /* Set sets[i].src_elt to the class each source belongs to.
4373 Detect assignments from or to volatile things
4374 and set set[i] to zero so they will be ignored
4375 in the rest of this function.
4377 Nothing in this loop changes the hash table or the register chains. */
4379 for (i = 0; i < n_sets; i++)
4381 bool repeat = false;
4382 rtx src, dest;
4383 rtx src_folded;
4384 struct table_elt *elt = 0, *p;
4385 enum machine_mode mode;
4386 rtx src_eqv_here;
4387 rtx src_const = 0;
4388 rtx src_related = 0;
4389 bool src_related_is_const_anchor = false;
4390 struct table_elt *src_const_elt = 0;
4391 int src_cost = MAX_COST;
4392 int src_eqv_cost = MAX_COST;
4393 int src_folded_cost = MAX_COST;
4394 int src_related_cost = MAX_COST;
4395 int src_elt_cost = MAX_COST;
4396 int src_regcost = MAX_COST;
4397 int src_eqv_regcost = MAX_COST;
4398 int src_folded_regcost = MAX_COST;
4399 int src_related_regcost = MAX_COST;
4400 int src_elt_regcost = MAX_COST;
4401 /* Set nonzero if we need to call force_const_mem on with the
4402 contents of src_folded before using it. */
4403 int src_folded_force_flag = 0;
4405 dest = SET_DEST (sets[i].rtl);
4406 src = SET_SRC (sets[i].rtl);
4408 /* If SRC is a constant that has no machine mode,
4409 hash it with the destination's machine mode.
4410 This way we can keep different modes separate. */
4412 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4413 sets[i].mode = mode;
4415 if (src_eqv)
4417 enum machine_mode eqvmode = mode;
4418 if (GET_CODE (dest) == STRICT_LOW_PART)
4419 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4420 do_not_record = 0;
4421 hash_arg_in_memory = 0;
4422 src_eqv_hash = HASH (src_eqv, eqvmode);
4424 /* Find the equivalence class for the equivalent expression. */
4426 if (!do_not_record)
4427 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4429 src_eqv_volatile = do_not_record;
4430 src_eqv_in_memory = hash_arg_in_memory;
4433 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4434 value of the INNER register, not the destination. So it is not
4435 a valid substitution for the source. But save it for later. */
4436 if (GET_CODE (dest) == STRICT_LOW_PART)
4437 src_eqv_here = 0;
4438 else
4439 src_eqv_here = src_eqv;
4441 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4442 simplified result, which may not necessarily be valid. */
4443 src_folded = fold_rtx (src, insn);
4445 #if 0
4446 /* ??? This caused bad code to be generated for the m68k port with -O2.
4447 Suppose src is (CONST_INT -1), and that after truncation src_folded
4448 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4449 At the end we will add src and src_const to the same equivalence
4450 class. We now have 3 and -1 on the same equivalence class. This
4451 causes later instructions to be mis-optimized. */
4452 /* If storing a constant in a bitfield, pre-truncate the constant
4453 so we will be able to record it later. */
4454 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4456 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4458 if (CONST_INT_P (src)
4459 && CONST_INT_P (width)
4460 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4461 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4462 src_folded
4463 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4464 << INTVAL (width)) - 1));
4466 #endif
4468 /* Compute SRC's hash code, and also notice if it
4469 should not be recorded at all. In that case,
4470 prevent any further processing of this assignment. */
4471 do_not_record = 0;
4472 hash_arg_in_memory = 0;
4474 sets[i].src = src;
4475 sets[i].src_hash = HASH (src, mode);
4476 sets[i].src_volatile = do_not_record;
4477 sets[i].src_in_memory = hash_arg_in_memory;
4479 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4480 a pseudo, do not record SRC. Using SRC as a replacement for
4481 anything else will be incorrect in that situation. Note that
4482 this usually occurs only for stack slots, in which case all the
4483 RTL would be referring to SRC, so we don't lose any optimization
4484 opportunities by not having SRC in the hash table. */
4486 if (MEM_P (src)
4487 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4488 && REG_P (dest)
4489 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4490 sets[i].src_volatile = 1;
4492 #if 0
4493 /* It is no longer clear why we used to do this, but it doesn't
4494 appear to still be needed. So let's try without it since this
4495 code hurts cse'ing widened ops. */
4496 /* If source is a paradoxical subreg (such as QI treated as an SI),
4497 treat it as volatile. It may do the work of an SI in one context
4498 where the extra bits are not being used, but cannot replace an SI
4499 in general. */
4500 if (paradoxical_subreg_p (src))
4501 sets[i].src_volatile = 1;
4502 #endif
4504 /* Locate all possible equivalent forms for SRC. Try to replace
4505 SRC in the insn with each cheaper equivalent.
4507 We have the following types of equivalents: SRC itself, a folded
4508 version, a value given in a REG_EQUAL note, or a value related
4509 to a constant.
4511 Each of these equivalents may be part of an additional class
4512 of equivalents (if more than one is in the table, they must be in
4513 the same class; we check for this).
4515 If the source is volatile, we don't do any table lookups.
4517 We note any constant equivalent for possible later use in a
4518 REG_NOTE. */
4520 if (!sets[i].src_volatile)
4521 elt = lookup (src, sets[i].src_hash, mode);
4523 sets[i].src_elt = elt;
4525 if (elt && src_eqv_here && src_eqv_elt)
4527 if (elt->first_same_value != src_eqv_elt->first_same_value)
4529 /* The REG_EQUAL is indicating that two formerly distinct
4530 classes are now equivalent. So merge them. */
4531 merge_equiv_classes (elt, src_eqv_elt);
4532 src_eqv_hash = HASH (src_eqv, elt->mode);
4533 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4536 src_eqv_here = 0;
4539 else if (src_eqv_elt)
4540 elt = src_eqv_elt;
4542 /* Try to find a constant somewhere and record it in `src_const'.
4543 Record its table element, if any, in `src_const_elt'. Look in
4544 any known equivalences first. (If the constant is not in the
4545 table, also set `sets[i].src_const_hash'). */
4546 if (elt)
4547 for (p = elt->first_same_value; p; p = p->next_same_value)
4548 if (p->is_const)
4550 src_const = p->exp;
4551 src_const_elt = elt;
4552 break;
4555 if (src_const == 0
4556 && (CONSTANT_P (src_folded)
4557 /* Consider (minus (label_ref L1) (label_ref L2)) as
4558 "constant" here so we will record it. This allows us
4559 to fold switch statements when an ADDR_DIFF_VEC is used. */
4560 || (GET_CODE (src_folded) == MINUS
4561 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4562 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4563 src_const = src_folded, src_const_elt = elt;
4564 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4565 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4567 /* If we don't know if the constant is in the table, get its
4568 hash code and look it up. */
4569 if (src_const && src_const_elt == 0)
4571 sets[i].src_const_hash = HASH (src_const, mode);
4572 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4575 sets[i].src_const = src_const;
4576 sets[i].src_const_elt = src_const_elt;
4578 /* If the constant and our source are both in the table, mark them as
4579 equivalent. Otherwise, if a constant is in the table but the source
4580 isn't, set ELT to it. */
4581 if (src_const_elt && elt
4582 && src_const_elt->first_same_value != elt->first_same_value)
4583 merge_equiv_classes (elt, src_const_elt);
4584 else if (src_const_elt && elt == 0)
4585 elt = src_const_elt;
4587 /* See if there is a register linearly related to a constant
4588 equivalent of SRC. */
4589 if (src_const
4590 && (GET_CODE (src_const) == CONST
4591 || (src_const_elt && src_const_elt->related_value != 0)))
4593 src_related = use_related_value (src_const, src_const_elt);
4594 if (src_related)
4596 struct table_elt *src_related_elt
4597 = lookup (src_related, HASH (src_related, mode), mode);
4598 if (src_related_elt && elt)
4600 if (elt->first_same_value
4601 != src_related_elt->first_same_value)
4602 /* This can occur when we previously saw a CONST
4603 involving a SYMBOL_REF and then see the SYMBOL_REF
4604 twice. Merge the involved classes. */
4605 merge_equiv_classes (elt, src_related_elt);
4607 src_related = 0;
4608 src_related_elt = 0;
4610 else if (src_related_elt && elt == 0)
4611 elt = src_related_elt;
4615 /* See if we have a CONST_INT that is already in a register in a
4616 wider mode. */
4618 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4619 && GET_MODE_CLASS (mode) == MODE_INT
4620 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4622 enum machine_mode wider_mode;
4624 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4625 wider_mode != VOIDmode
4626 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4627 && src_related == 0;
4628 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4630 struct table_elt *const_elt
4631 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4633 if (const_elt == 0)
4634 continue;
4636 for (const_elt = const_elt->first_same_value;
4637 const_elt; const_elt = const_elt->next_same_value)
4638 if (REG_P (const_elt->exp))
4640 src_related = gen_lowpart (mode, const_elt->exp);
4641 break;
4646 /* Another possibility is that we have an AND with a constant in
4647 a mode narrower than a word. If so, it might have been generated
4648 as part of an "if" which would narrow the AND. If we already
4649 have done the AND in a wider mode, we can use a SUBREG of that
4650 value. */
4652 if (flag_expensive_optimizations && ! src_related
4653 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4654 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4656 enum machine_mode tmode;
4657 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4659 for (tmode = GET_MODE_WIDER_MODE (mode);
4660 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4661 tmode = GET_MODE_WIDER_MODE (tmode))
4663 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4664 struct table_elt *larger_elt;
4666 if (inner)
4668 PUT_MODE (new_and, tmode);
4669 XEXP (new_and, 0) = inner;
4670 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4671 if (larger_elt == 0)
4672 continue;
4674 for (larger_elt = larger_elt->first_same_value;
4675 larger_elt; larger_elt = larger_elt->next_same_value)
4676 if (REG_P (larger_elt->exp))
4678 src_related
4679 = gen_lowpart (mode, larger_elt->exp);
4680 break;
4683 if (src_related)
4684 break;
4689 #ifdef LOAD_EXTEND_OP
4690 /* See if a MEM has already been loaded with a widening operation;
4691 if it has, we can use a subreg of that. Many CISC machines
4692 also have such operations, but this is only likely to be
4693 beneficial on these machines. */
4695 if (flag_expensive_optimizations && src_related == 0
4696 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4697 && GET_MODE_CLASS (mode) == MODE_INT
4698 && MEM_P (src) && ! do_not_record
4699 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4701 struct rtx_def memory_extend_buf;
4702 rtx memory_extend_rtx = &memory_extend_buf;
4703 enum machine_mode tmode;
4705 /* Set what we are trying to extend and the operation it might
4706 have been extended with. */
4707 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4708 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4709 XEXP (memory_extend_rtx, 0) = src;
4711 for (tmode = GET_MODE_WIDER_MODE (mode);
4712 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4713 tmode = GET_MODE_WIDER_MODE (tmode))
4715 struct table_elt *larger_elt;
4717 PUT_MODE (memory_extend_rtx, tmode);
4718 larger_elt = lookup (memory_extend_rtx,
4719 HASH (memory_extend_rtx, tmode), tmode);
4720 if (larger_elt == 0)
4721 continue;
4723 for (larger_elt = larger_elt->first_same_value;
4724 larger_elt; larger_elt = larger_elt->next_same_value)
4725 if (REG_P (larger_elt->exp))
4727 src_related = gen_lowpart (mode, larger_elt->exp);
4728 break;
4731 if (src_related)
4732 break;
4735 #endif /* LOAD_EXTEND_OP */
4737 /* Try to express the constant using a register+offset expression
4738 derived from a constant anchor. */
4740 if (targetm.const_anchor
4741 && !src_related
4742 && src_const
4743 && GET_CODE (src_const) == CONST_INT)
4745 src_related = try_const_anchors (src_const, mode);
4746 src_related_is_const_anchor = src_related != NULL_RTX;
4750 if (src == src_folded)
4751 src_folded = 0;
4753 /* At this point, ELT, if nonzero, points to a class of expressions
4754 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4755 and SRC_RELATED, if nonzero, each contain additional equivalent
4756 expressions. Prune these latter expressions by deleting expressions
4757 already in the equivalence class.
4759 Check for an equivalent identical to the destination. If found,
4760 this is the preferred equivalent since it will likely lead to
4761 elimination of the insn. Indicate this by placing it in
4762 `src_related'. */
4764 if (elt)
4765 elt = elt->first_same_value;
4766 for (p = elt; p; p = p->next_same_value)
4768 enum rtx_code code = GET_CODE (p->exp);
4770 /* If the expression is not valid, ignore it. Then we do not
4771 have to check for validity below. In most cases, we can use
4772 `rtx_equal_p', since canonicalization has already been done. */
4773 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4774 continue;
4776 /* Also skip paradoxical subregs, unless that's what we're
4777 looking for. */
4778 if (paradoxical_subreg_p (p->exp)
4779 && ! (src != 0
4780 && GET_CODE (src) == SUBREG
4781 && GET_MODE (src) == GET_MODE (p->exp)
4782 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4783 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4784 continue;
4786 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4787 src = 0;
4788 else if (src_folded && GET_CODE (src_folded) == code
4789 && rtx_equal_p (src_folded, p->exp))
4790 src_folded = 0;
4791 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4792 && rtx_equal_p (src_eqv_here, p->exp))
4793 src_eqv_here = 0;
4794 else if (src_related && GET_CODE (src_related) == code
4795 && rtx_equal_p (src_related, p->exp))
4796 src_related = 0;
4798 /* This is the same as the destination of the insns, we want
4799 to prefer it. Copy it to src_related. The code below will
4800 then give it a negative cost. */
4801 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4802 src_related = dest;
4805 /* Find the cheapest valid equivalent, trying all the available
4806 possibilities. Prefer items not in the hash table to ones
4807 that are when they are equal cost. Note that we can never
4808 worsen an insn as the current contents will also succeed.
4809 If we find an equivalent identical to the destination, use it as best,
4810 since this insn will probably be eliminated in that case. */
4811 if (src)
4813 if (rtx_equal_p (src, dest))
4814 src_cost = src_regcost = -1;
4815 else
4817 src_cost = COST (src);
4818 src_regcost = approx_reg_cost (src);
4822 if (src_eqv_here)
4824 if (rtx_equal_p (src_eqv_here, dest))
4825 src_eqv_cost = src_eqv_regcost = -1;
4826 else
4828 src_eqv_cost = COST (src_eqv_here);
4829 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4833 if (src_folded)
4835 if (rtx_equal_p (src_folded, dest))
4836 src_folded_cost = src_folded_regcost = -1;
4837 else
4839 src_folded_cost = COST (src_folded);
4840 src_folded_regcost = approx_reg_cost (src_folded);
4844 if (src_related)
4846 if (rtx_equal_p (src_related, dest))
4847 src_related_cost = src_related_regcost = -1;
4848 else
4850 src_related_cost = COST (src_related);
4851 src_related_regcost = approx_reg_cost (src_related);
4853 /* If a const-anchor is used to synthesize a constant that
4854 normally requires multiple instructions then slightly prefer
4855 it over the original sequence. These instructions are likely
4856 to become redundant now. We can't compare against the cost
4857 of src_eqv_here because, on MIPS for example, multi-insn
4858 constants have zero cost; they are assumed to be hoisted from
4859 loops. */
4860 if (src_related_is_const_anchor
4861 && src_related_cost == src_cost
4862 && src_eqv_here)
4863 src_related_cost--;
4867 /* If this was an indirect jump insn, a known label will really be
4868 cheaper even though it looks more expensive. */
4869 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
4870 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
4872 /* Terminate loop when replacement made. This must terminate since
4873 the current contents will be tested and will always be valid. */
4874 while (1)
4876 rtx trial;
4878 /* Skip invalid entries. */
4879 while (elt && !REG_P (elt->exp)
4880 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
4881 elt = elt->next_same_value;
4883 /* A paradoxical subreg would be bad here: it'll be the right
4884 size, but later may be adjusted so that the upper bits aren't
4885 what we want. So reject it. */
4886 if (elt != 0
4887 && paradoxical_subreg_p (elt->exp)
4888 /* It is okay, though, if the rtx we're trying to match
4889 will ignore any of the bits we can't predict. */
4890 && ! (src != 0
4891 && GET_CODE (src) == SUBREG
4892 && GET_MODE (src) == GET_MODE (elt->exp)
4893 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4894 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
4896 elt = elt->next_same_value;
4897 continue;
4900 if (elt)
4902 src_elt_cost = elt->cost;
4903 src_elt_regcost = elt->regcost;
4906 /* Find cheapest and skip it for the next time. For items
4907 of equal cost, use this order:
4908 src_folded, src, src_eqv, src_related and hash table entry. */
4909 if (src_folded
4910 && preferable (src_folded_cost, src_folded_regcost,
4911 src_cost, src_regcost) <= 0
4912 && preferable (src_folded_cost, src_folded_regcost,
4913 src_eqv_cost, src_eqv_regcost) <= 0
4914 && preferable (src_folded_cost, src_folded_regcost,
4915 src_related_cost, src_related_regcost) <= 0
4916 && preferable (src_folded_cost, src_folded_regcost,
4917 src_elt_cost, src_elt_regcost) <= 0)
4919 trial = src_folded, src_folded_cost = MAX_COST;
4920 if (src_folded_force_flag)
4922 rtx forced = force_const_mem (mode, trial);
4923 if (forced)
4924 trial = forced;
4927 else if (src
4928 && preferable (src_cost, src_regcost,
4929 src_eqv_cost, src_eqv_regcost) <= 0
4930 && preferable (src_cost, src_regcost,
4931 src_related_cost, src_related_regcost) <= 0
4932 && preferable (src_cost, src_regcost,
4933 src_elt_cost, src_elt_regcost) <= 0)
4934 trial = src, src_cost = MAX_COST;
4935 else if (src_eqv_here
4936 && preferable (src_eqv_cost, src_eqv_regcost,
4937 src_related_cost, src_related_regcost) <= 0
4938 && preferable (src_eqv_cost, src_eqv_regcost,
4939 src_elt_cost, src_elt_regcost) <= 0)
4940 trial = src_eqv_here, src_eqv_cost = MAX_COST;
4941 else if (src_related
4942 && preferable (src_related_cost, src_related_regcost,
4943 src_elt_cost, src_elt_regcost) <= 0)
4944 trial = src_related, src_related_cost = MAX_COST;
4945 else
4947 trial = elt->exp;
4948 elt = elt->next_same_value;
4949 src_elt_cost = MAX_COST;
4952 /* Avoid creation of overlapping memory moves. */
4953 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
4955 rtx src, dest;
4957 /* BLKmode moves are not handled by cse anyway. */
4958 if (GET_MODE (trial) == BLKmode)
4959 break;
4961 src = canon_rtx (trial);
4962 dest = canon_rtx (SET_DEST (sets[i].rtl));
4964 if (!MEM_P (src) || !MEM_P (dest)
4965 || !nonoverlapping_memrefs_p (src, dest, false))
4966 break;
4969 /* Try to optimize
4970 (set (reg:M N) (const_int A))
4971 (set (reg:M2 O) (const_int B))
4972 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
4973 (reg:M2 O)). */
4974 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
4975 && CONST_INT_P (trial)
4976 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
4977 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
4978 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
4979 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
4980 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
4981 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
4982 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
4983 <= HOST_BITS_PER_WIDE_INT))
4985 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
4986 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4987 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
4988 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
4989 struct table_elt *dest_elt
4990 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
4991 rtx dest_cst = NULL;
4993 if (dest_elt)
4994 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
4995 if (p->is_const && CONST_INT_P (p->exp))
4997 dest_cst = p->exp;
4998 break;
5000 if (dest_cst)
5002 HOST_WIDE_INT val = INTVAL (dest_cst);
5003 HOST_WIDE_INT mask;
5004 unsigned int shift;
5005 if (BITS_BIG_ENDIAN)
5006 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5007 - INTVAL (pos) - INTVAL (width);
5008 else
5009 shift = INTVAL (pos);
5010 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5011 mask = ~(HOST_WIDE_INT) 0;
5012 else
5013 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5014 val &= ~(mask << shift);
5015 val |= (INTVAL (trial) & mask) << shift;
5016 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5017 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5018 dest_reg, 1);
5019 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5020 GEN_INT (val), 1);
5021 if (apply_change_group ())
5023 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5024 if (note)
5026 remove_note (insn, note);
5027 df_notes_rescan (insn);
5029 src_eqv = NULL_RTX;
5030 src_eqv_elt = NULL;
5031 src_eqv_volatile = 0;
5032 src_eqv_in_memory = 0;
5033 src_eqv_hash = 0;
5034 repeat = true;
5035 break;
5040 /* We don't normally have an insn matching (set (pc) (pc)), so
5041 check for this separately here. We will delete such an
5042 insn below.
5044 For other cases such as a table jump or conditional jump
5045 where we know the ultimate target, go ahead and replace the
5046 operand. While that may not make a valid insn, we will
5047 reemit the jump below (and also insert any necessary
5048 barriers). */
5049 if (n_sets == 1 && dest == pc_rtx
5050 && (trial == pc_rtx
5051 || (GET_CODE (trial) == LABEL_REF
5052 && ! condjump_p (insn))))
5054 /* Don't substitute non-local labels, this confuses CFG. */
5055 if (GET_CODE (trial) == LABEL_REF
5056 && LABEL_REF_NONLOCAL_P (trial))
5057 continue;
5059 SET_SRC (sets[i].rtl) = trial;
5060 cse_jumps_altered = true;
5061 break;
5064 /* Reject certain invalid forms of CONST that we create. */
5065 else if (CONSTANT_P (trial)
5066 && GET_CODE (trial) == CONST
5067 /* Reject cases that will cause decode_rtx_const to
5068 die. On the alpha when simplifying a switch, we
5069 get (const (truncate (minus (label_ref)
5070 (label_ref)))). */
5071 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5072 /* Likewise on IA-64, except without the
5073 truncate. */
5074 || (GET_CODE (XEXP (trial, 0)) == MINUS
5075 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5076 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5077 /* Do nothing for this case. */
5080 /* Look for a substitution that makes a valid insn. */
5081 else if (validate_unshare_change
5082 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5084 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5086 /* The result of apply_change_group can be ignored; see
5087 canon_reg. */
5089 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5090 apply_change_group ();
5092 break;
5095 /* If we previously found constant pool entries for
5096 constants and this is a constant, try making a
5097 pool entry. Put it in src_folded unless we already have done
5098 this since that is where it likely came from. */
5100 else if (constant_pool_entries_cost
5101 && CONSTANT_P (trial)
5102 && (src_folded == 0
5103 || (!MEM_P (src_folded)
5104 && ! src_folded_force_flag))
5105 && GET_MODE_CLASS (mode) != MODE_CC
5106 && mode != VOIDmode)
5108 src_folded_force_flag = 1;
5109 src_folded = trial;
5110 src_folded_cost = constant_pool_entries_cost;
5111 src_folded_regcost = constant_pool_entries_regcost;
5115 /* If we changed the insn too much, handle this set from scratch. */
5116 if (repeat)
5118 i--;
5119 continue;
5122 src = SET_SRC (sets[i].rtl);
5124 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5125 However, there is an important exception: If both are registers
5126 that are not the head of their equivalence class, replace SET_SRC
5127 with the head of the class. If we do not do this, we will have
5128 both registers live over a portion of the basic block. This way,
5129 their lifetimes will likely abut instead of overlapping. */
5130 if (REG_P (dest)
5131 && REGNO_QTY_VALID_P (REGNO (dest)))
5133 int dest_q = REG_QTY (REGNO (dest));
5134 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5136 if (dest_ent->mode == GET_MODE (dest)
5137 && dest_ent->first_reg != REGNO (dest)
5138 && REG_P (src) && REGNO (src) == REGNO (dest)
5139 /* Don't do this if the original insn had a hard reg as
5140 SET_SRC or SET_DEST. */
5141 && (!REG_P (sets[i].src)
5142 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5143 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5144 /* We can't call canon_reg here because it won't do anything if
5145 SRC is a hard register. */
5147 int src_q = REG_QTY (REGNO (src));
5148 struct qty_table_elem *src_ent = &qty_table[src_q];
5149 int first = src_ent->first_reg;
5150 rtx new_src
5151 = (first >= FIRST_PSEUDO_REGISTER
5152 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5154 /* We must use validate-change even for this, because this
5155 might be a special no-op instruction, suitable only to
5156 tag notes onto. */
5157 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5159 src = new_src;
5160 /* If we had a constant that is cheaper than what we are now
5161 setting SRC to, use that constant. We ignored it when we
5162 thought we could make this into a no-op. */
5163 if (src_const && COST (src_const) < COST (src)
5164 && validate_change (insn, &SET_SRC (sets[i].rtl),
5165 src_const, 0))
5166 src = src_const;
5171 /* If we made a change, recompute SRC values. */
5172 if (src != sets[i].src)
5174 do_not_record = 0;
5175 hash_arg_in_memory = 0;
5176 sets[i].src = src;
5177 sets[i].src_hash = HASH (src, mode);
5178 sets[i].src_volatile = do_not_record;
5179 sets[i].src_in_memory = hash_arg_in_memory;
5180 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5183 /* If this is a single SET, we are setting a register, and we have an
5184 equivalent constant, we want to add a REG_NOTE. We don't want
5185 to write a REG_EQUAL note for a constant pseudo since verifying that
5186 that pseudo hasn't been eliminated is a pain. Such a note also
5187 won't help anything.
5189 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5190 which can be created for a reference to a compile time computable
5191 entry in a jump table. */
5193 if (n_sets == 1 && src_const && REG_P (dest)
5194 && !REG_P (src_const)
5195 && ! (GET_CODE (src_const) == CONST
5196 && GET_CODE (XEXP (src_const, 0)) == MINUS
5197 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5198 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5200 /* We only want a REG_EQUAL note if src_const != src. */
5201 if (! rtx_equal_p (src, src_const))
5203 /* Make sure that the rtx is not shared. */
5204 src_const = copy_rtx (src_const);
5206 /* Record the actual constant value in a REG_EQUAL note,
5207 making a new one if one does not already exist. */
5208 set_unique_reg_note (insn, REG_EQUAL, src_const);
5209 df_notes_rescan (insn);
5213 /* Now deal with the destination. */
5214 do_not_record = 0;
5216 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5217 while (GET_CODE (dest) == SUBREG
5218 || GET_CODE (dest) == ZERO_EXTRACT
5219 || GET_CODE (dest) == STRICT_LOW_PART)
5220 dest = XEXP (dest, 0);
5222 sets[i].inner_dest = dest;
5224 if (MEM_P (dest))
5226 #ifdef PUSH_ROUNDING
5227 /* Stack pushes invalidate the stack pointer. */
5228 rtx addr = XEXP (dest, 0);
5229 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5230 && XEXP (addr, 0) == stack_pointer_rtx)
5231 invalidate (stack_pointer_rtx, VOIDmode);
5232 #endif
5233 dest = fold_rtx (dest, insn);
5236 /* Compute the hash code of the destination now,
5237 before the effects of this instruction are recorded,
5238 since the register values used in the address computation
5239 are those before this instruction. */
5240 sets[i].dest_hash = HASH (dest, mode);
5242 /* Don't enter a bit-field in the hash table
5243 because the value in it after the store
5244 may not equal what was stored, due to truncation. */
5246 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5248 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5250 if (src_const != 0 && CONST_INT_P (src_const)
5251 && CONST_INT_P (width)
5252 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5253 && ! (INTVAL (src_const)
5254 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5255 /* Exception: if the value is constant,
5256 and it won't be truncated, record it. */
5258 else
5260 /* This is chosen so that the destination will be invalidated
5261 but no new value will be recorded.
5262 We must invalidate because sometimes constant
5263 values can be recorded for bitfields. */
5264 sets[i].src_elt = 0;
5265 sets[i].src_volatile = 1;
5266 src_eqv = 0;
5267 src_eqv_elt = 0;
5271 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5272 the insn. */
5273 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5275 /* One less use of the label this insn used to jump to. */
5276 delete_insn_and_edges (insn);
5277 cse_jumps_altered = true;
5278 /* No more processing for this set. */
5279 sets[i].rtl = 0;
5282 /* If this SET is now setting PC to a label, we know it used to
5283 be a conditional or computed branch. */
5284 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5285 && !LABEL_REF_NONLOCAL_P (src))
5287 /* We reemit the jump in as many cases as possible just in
5288 case the form of an unconditional jump is significantly
5289 different than a computed jump or conditional jump.
5291 If this insn has multiple sets, then reemitting the
5292 jump is nontrivial. So instead we just force rerecognition
5293 and hope for the best. */
5294 if (n_sets == 1)
5296 rtx new_rtx, note;
5298 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5299 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5300 LABEL_NUSES (XEXP (src, 0))++;
5302 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5303 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5304 if (note)
5306 XEXP (note, 1) = NULL_RTX;
5307 REG_NOTES (new_rtx) = note;
5310 delete_insn_and_edges (insn);
5311 insn = new_rtx;
5313 else
5314 INSN_CODE (insn) = -1;
5316 /* Do not bother deleting any unreachable code, let jump do it. */
5317 cse_jumps_altered = true;
5318 sets[i].rtl = 0;
5321 /* If destination is volatile, invalidate it and then do no further
5322 processing for this assignment. */
5324 else if (do_not_record)
5326 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5327 invalidate (dest, VOIDmode);
5328 else if (MEM_P (dest))
5329 invalidate (dest, VOIDmode);
5330 else if (GET_CODE (dest) == STRICT_LOW_PART
5331 || GET_CODE (dest) == ZERO_EXTRACT)
5332 invalidate (XEXP (dest, 0), GET_MODE (dest));
5333 sets[i].rtl = 0;
5336 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5337 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5339 #ifdef HAVE_cc0
5340 /* If setting CC0, record what it was set to, or a constant, if it
5341 is equivalent to a constant. If it is being set to a floating-point
5342 value, make a COMPARE with the appropriate constant of 0. If we
5343 don't do this, later code can interpret this as a test against
5344 const0_rtx, which can cause problems if we try to put it into an
5345 insn as a floating-point operand. */
5346 if (dest == cc0_rtx)
5348 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5349 this_insn_cc0_mode = mode;
5350 if (FLOAT_MODE_P (mode))
5351 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5352 CONST0_RTX (mode));
5354 #endif
5357 /* Now enter all non-volatile source expressions in the hash table
5358 if they are not already present.
5359 Record their equivalence classes in src_elt.
5360 This way we can insert the corresponding destinations into
5361 the same classes even if the actual sources are no longer in them
5362 (having been invalidated). */
5364 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5365 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5367 struct table_elt *elt;
5368 struct table_elt *classp = sets[0].src_elt;
5369 rtx dest = SET_DEST (sets[0].rtl);
5370 enum machine_mode eqvmode = GET_MODE (dest);
5372 if (GET_CODE (dest) == STRICT_LOW_PART)
5374 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5375 classp = 0;
5377 if (insert_regs (src_eqv, classp, 0))
5379 rehash_using_reg (src_eqv);
5380 src_eqv_hash = HASH (src_eqv, eqvmode);
5382 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5383 elt->in_memory = src_eqv_in_memory;
5384 src_eqv_elt = elt;
5386 /* Check to see if src_eqv_elt is the same as a set source which
5387 does not yet have an elt, and if so set the elt of the set source
5388 to src_eqv_elt. */
5389 for (i = 0; i < n_sets; i++)
5390 if (sets[i].rtl && sets[i].src_elt == 0
5391 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5392 sets[i].src_elt = src_eqv_elt;
5395 for (i = 0; i < n_sets; i++)
5396 if (sets[i].rtl && ! sets[i].src_volatile
5397 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5399 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5401 /* REG_EQUAL in setting a STRICT_LOW_PART
5402 gives an equivalent for the entire destination register,
5403 not just for the subreg being stored in now.
5404 This is a more interesting equivalence, so we arrange later
5405 to treat the entire reg as the destination. */
5406 sets[i].src_elt = src_eqv_elt;
5407 sets[i].src_hash = src_eqv_hash;
5409 else
5411 /* Insert source and constant equivalent into hash table, if not
5412 already present. */
5413 struct table_elt *classp = src_eqv_elt;
5414 rtx src = sets[i].src;
5415 rtx dest = SET_DEST (sets[i].rtl);
5416 enum machine_mode mode
5417 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5419 /* It's possible that we have a source value known to be
5420 constant but don't have a REG_EQUAL note on the insn.
5421 Lack of a note will mean src_eqv_elt will be NULL. This
5422 can happen where we've generated a SUBREG to access a
5423 CONST_INT that is already in a register in a wider mode.
5424 Ensure that the source expression is put in the proper
5425 constant class. */
5426 if (!classp)
5427 classp = sets[i].src_const_elt;
5429 if (sets[i].src_elt == 0)
5431 struct table_elt *elt;
5433 /* Note that these insert_regs calls cannot remove
5434 any of the src_elt's, because they would have failed to
5435 match if not still valid. */
5436 if (insert_regs (src, classp, 0))
5438 rehash_using_reg (src);
5439 sets[i].src_hash = HASH (src, mode);
5441 elt = insert (src, classp, sets[i].src_hash, mode);
5442 elt->in_memory = sets[i].src_in_memory;
5443 sets[i].src_elt = classp = elt;
5445 if (sets[i].src_const && sets[i].src_const_elt == 0
5446 && src != sets[i].src_const
5447 && ! rtx_equal_p (sets[i].src_const, src))
5448 sets[i].src_elt = insert (sets[i].src_const, classp,
5449 sets[i].src_const_hash, mode);
5452 else if (sets[i].src_elt == 0)
5453 /* If we did not insert the source into the hash table (e.g., it was
5454 volatile), note the equivalence class for the REG_EQUAL value, if any,
5455 so that the destination goes into that class. */
5456 sets[i].src_elt = src_eqv_elt;
5458 /* Record destination addresses in the hash table. This allows us to
5459 check if they are invalidated by other sets. */
5460 for (i = 0; i < n_sets; i++)
5462 if (sets[i].rtl)
5464 rtx x = sets[i].inner_dest;
5465 struct table_elt *elt;
5466 enum machine_mode mode;
5467 unsigned hash;
5469 if (MEM_P (x))
5471 x = XEXP (x, 0);
5472 mode = GET_MODE (x);
5473 hash = HASH (x, mode);
5474 elt = lookup (x, hash, mode);
5475 if (!elt)
5477 if (insert_regs (x, NULL, 0))
5479 rtx dest = SET_DEST (sets[i].rtl);
5481 rehash_using_reg (x);
5482 hash = HASH (x, mode);
5483 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5485 elt = insert (x, NULL, hash, mode);
5488 sets[i].dest_addr_elt = elt;
5490 else
5491 sets[i].dest_addr_elt = NULL;
5495 invalidate_from_clobbers (x);
5497 /* Some registers are invalidated by subroutine calls. Memory is
5498 invalidated by non-constant calls. */
5500 if (CALL_P (insn))
5502 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5503 invalidate_memory ();
5504 invalidate_for_call ();
5507 /* Now invalidate everything set by this instruction.
5508 If a SUBREG or other funny destination is being set,
5509 sets[i].rtl is still nonzero, so here we invalidate the reg
5510 a part of which is being set. */
5512 for (i = 0; i < n_sets; i++)
5513 if (sets[i].rtl)
5515 /* We can't use the inner dest, because the mode associated with
5516 a ZERO_EXTRACT is significant. */
5517 rtx dest = SET_DEST (sets[i].rtl);
5519 /* Needed for registers to remove the register from its
5520 previous quantity's chain.
5521 Needed for memory if this is a nonvarying address, unless
5522 we have just done an invalidate_memory that covers even those. */
5523 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5524 invalidate (dest, VOIDmode);
5525 else if (MEM_P (dest))
5526 invalidate (dest, VOIDmode);
5527 else if (GET_CODE (dest) == STRICT_LOW_PART
5528 || GET_CODE (dest) == ZERO_EXTRACT)
5529 invalidate (XEXP (dest, 0), GET_MODE (dest));
5532 /* A volatile ASM invalidates everything. */
5533 if (NONJUMP_INSN_P (insn)
5534 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5535 && MEM_VOLATILE_P (PATTERN (insn)))
5536 flush_hash_table ();
5538 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5539 the regs restored by the longjmp come from a later time
5540 than the setjmp. */
5541 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5543 flush_hash_table ();
5544 goto done;
5547 /* Make sure registers mentioned in destinations
5548 are safe for use in an expression to be inserted.
5549 This removes from the hash table
5550 any invalid entry that refers to one of these registers.
5552 We don't care about the return value from mention_regs because
5553 we are going to hash the SET_DEST values unconditionally. */
5555 for (i = 0; i < n_sets; i++)
5557 if (sets[i].rtl)
5559 rtx x = SET_DEST (sets[i].rtl);
5561 if (!REG_P (x))
5562 mention_regs (x);
5563 else
5565 /* We used to rely on all references to a register becoming
5566 inaccessible when a register changes to a new quantity,
5567 since that changes the hash code. However, that is not
5568 safe, since after HASH_SIZE new quantities we get a
5569 hash 'collision' of a register with its own invalid
5570 entries. And since SUBREGs have been changed not to
5571 change their hash code with the hash code of the register,
5572 it wouldn't work any longer at all. So we have to check
5573 for any invalid references lying around now.
5574 This code is similar to the REG case in mention_regs,
5575 but it knows that reg_tick has been incremented, and
5576 it leaves reg_in_table as -1 . */
5577 unsigned int regno = REGNO (x);
5578 unsigned int endregno = END_REGNO (x);
5579 unsigned int i;
5581 for (i = regno; i < endregno; i++)
5583 if (REG_IN_TABLE (i) >= 0)
5585 remove_invalid_refs (i);
5586 REG_IN_TABLE (i) = -1;
5593 /* We may have just removed some of the src_elt's from the hash table.
5594 So replace each one with the current head of the same class.
5595 Also check if destination addresses have been removed. */
5597 for (i = 0; i < n_sets; i++)
5598 if (sets[i].rtl)
5600 if (sets[i].dest_addr_elt
5601 && sets[i].dest_addr_elt->first_same_value == 0)
5603 /* The elt was removed, which means this destination is not
5604 valid after this instruction. */
5605 sets[i].rtl = NULL_RTX;
5607 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5608 /* If elt was removed, find current head of same class,
5609 or 0 if nothing remains of that class. */
5611 struct table_elt *elt = sets[i].src_elt;
5613 while (elt && elt->prev_same_value)
5614 elt = elt->prev_same_value;
5616 while (elt && elt->first_same_value == 0)
5617 elt = elt->next_same_value;
5618 sets[i].src_elt = elt ? elt->first_same_value : 0;
5622 /* Now insert the destinations into their equivalence classes. */
5624 for (i = 0; i < n_sets; i++)
5625 if (sets[i].rtl)
5627 rtx dest = SET_DEST (sets[i].rtl);
5628 struct table_elt *elt;
5630 /* Don't record value if we are not supposed to risk allocating
5631 floating-point values in registers that might be wider than
5632 memory. */
5633 if ((flag_float_store
5634 && MEM_P (dest)
5635 && FLOAT_MODE_P (GET_MODE (dest)))
5636 /* Don't record BLKmode values, because we don't know the
5637 size of it, and can't be sure that other BLKmode values
5638 have the same or smaller size. */
5639 || GET_MODE (dest) == BLKmode
5640 /* If we didn't put a REG_EQUAL value or a source into the hash
5641 table, there is no point is recording DEST. */
5642 || sets[i].src_elt == 0
5643 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5644 or SIGN_EXTEND, don't record DEST since it can cause
5645 some tracking to be wrong.
5647 ??? Think about this more later. */
5648 || (paradoxical_subreg_p (dest)
5649 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5650 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5651 continue;
5653 /* STRICT_LOW_PART isn't part of the value BEING set,
5654 and neither is the SUBREG inside it.
5655 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5656 if (GET_CODE (dest) == STRICT_LOW_PART)
5657 dest = SUBREG_REG (XEXP (dest, 0));
5659 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5660 /* Registers must also be inserted into chains for quantities. */
5661 if (insert_regs (dest, sets[i].src_elt, 1))
5663 /* If `insert_regs' changes something, the hash code must be
5664 recalculated. */
5665 rehash_using_reg (dest);
5666 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5669 elt = insert (dest, sets[i].src_elt,
5670 sets[i].dest_hash, GET_MODE (dest));
5672 /* If this is a constant, insert the constant anchors with the
5673 equivalent register-offset expressions using register DEST. */
5674 if (targetm.const_anchor
5675 && REG_P (dest)
5676 && SCALAR_INT_MODE_P (GET_MODE (dest))
5677 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5678 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5680 elt->in_memory = (MEM_P (sets[i].inner_dest)
5681 && !MEM_READONLY_P (sets[i].inner_dest));
5683 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5684 narrower than M2, and both M1 and M2 are the same number of words,
5685 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5686 make that equivalence as well.
5688 However, BAR may have equivalences for which gen_lowpart
5689 will produce a simpler value than gen_lowpart applied to
5690 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5691 BAR's equivalences. If we don't get a simplified form, make
5692 the SUBREG. It will not be used in an equivalence, but will
5693 cause two similar assignments to be detected.
5695 Note the loop below will find SUBREG_REG (DEST) since we have
5696 already entered SRC and DEST of the SET in the table. */
5698 if (GET_CODE (dest) == SUBREG
5699 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5700 / UNITS_PER_WORD)
5701 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5702 && (GET_MODE_SIZE (GET_MODE (dest))
5703 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5704 && sets[i].src_elt != 0)
5706 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5707 struct table_elt *elt, *classp = 0;
5709 for (elt = sets[i].src_elt->first_same_value; elt;
5710 elt = elt->next_same_value)
5712 rtx new_src = 0;
5713 unsigned src_hash;
5714 struct table_elt *src_elt;
5715 int byte = 0;
5717 /* Ignore invalid entries. */
5718 if (!REG_P (elt->exp)
5719 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5720 continue;
5722 /* We may have already been playing subreg games. If the
5723 mode is already correct for the destination, use it. */
5724 if (GET_MODE (elt->exp) == new_mode)
5725 new_src = elt->exp;
5726 else
5728 /* Calculate big endian correction for the SUBREG_BYTE.
5729 We have already checked that M1 (GET_MODE (dest))
5730 is not narrower than M2 (new_mode). */
5731 if (BYTES_BIG_ENDIAN)
5732 byte = (GET_MODE_SIZE (GET_MODE (dest))
5733 - GET_MODE_SIZE (new_mode));
5735 new_src = simplify_gen_subreg (new_mode, elt->exp,
5736 GET_MODE (dest), byte);
5739 /* The call to simplify_gen_subreg fails if the value
5740 is VOIDmode, yet we can't do any simplification, e.g.
5741 for EXPR_LISTs denoting function call results.
5742 It is invalid to construct a SUBREG with a VOIDmode
5743 SUBREG_REG, hence a zero new_src means we can't do
5744 this substitution. */
5745 if (! new_src)
5746 continue;
5748 src_hash = HASH (new_src, new_mode);
5749 src_elt = lookup (new_src, src_hash, new_mode);
5751 /* Put the new source in the hash table is if isn't
5752 already. */
5753 if (src_elt == 0)
5755 if (insert_regs (new_src, classp, 0))
5757 rehash_using_reg (new_src);
5758 src_hash = HASH (new_src, new_mode);
5760 src_elt = insert (new_src, classp, src_hash, new_mode);
5761 src_elt->in_memory = elt->in_memory;
5763 else if (classp && classp != src_elt->first_same_value)
5764 /* Show that two things that we've seen before are
5765 actually the same. */
5766 merge_equiv_classes (src_elt, classp);
5768 classp = src_elt->first_same_value;
5769 /* Ignore invalid entries. */
5770 while (classp
5771 && !REG_P (classp->exp)
5772 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5773 classp = classp->next_same_value;
5778 /* Special handling for (set REG0 REG1) where REG0 is the
5779 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5780 be used in the sequel, so (if easily done) change this insn to
5781 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5782 that computed their value. Then REG1 will become a dead store
5783 and won't cloud the situation for later optimizations.
5785 Do not make this change if REG1 is a hard register, because it will
5786 then be used in the sequel and we may be changing a two-operand insn
5787 into a three-operand insn.
5789 Also do not do this if we are operating on a copy of INSN. */
5791 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
5792 && NEXT_INSN (PREV_INSN (insn)) == insn
5793 && REG_P (SET_SRC (sets[0].rtl))
5794 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
5795 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
5797 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
5798 struct qty_table_elem *src_ent = &qty_table[src_q];
5800 if (src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
5802 /* Scan for the previous nonnote insn, but stop at a basic
5803 block boundary. */
5804 rtx prev = insn;
5805 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
5808 prev = PREV_INSN (prev);
5810 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
5812 /* Do not swap the registers around if the previous instruction
5813 attaches a REG_EQUIV note to REG1.
5815 ??? It's not entirely clear whether we can transfer a REG_EQUIV
5816 from the pseudo that originally shadowed an incoming argument
5817 to another register. Some uses of REG_EQUIV might rely on it
5818 being attached to REG1 rather than REG2.
5820 This section previously turned the REG_EQUIV into a REG_EQUAL
5821 note. We cannot do that because REG_EQUIV may provide an
5822 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
5823 if (NONJUMP_INSN_P (prev)
5824 && GET_CODE (PATTERN (prev)) == SET
5825 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
5826 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
5828 rtx dest = SET_DEST (sets[0].rtl);
5829 rtx src = SET_SRC (sets[0].rtl);
5830 rtx note;
5832 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
5833 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
5834 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
5835 apply_change_group ();
5837 /* If INSN has a REG_EQUAL note, and this note mentions
5838 REG0, then we must delete it, because the value in
5839 REG0 has changed. If the note's value is REG1, we must
5840 also delete it because that is now this insn's dest. */
5841 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5842 if (note != 0
5843 && (reg_mentioned_p (dest, XEXP (note, 0))
5844 || rtx_equal_p (src, XEXP (note, 0))))
5845 remove_note (insn, note);
5850 done:;
5853 /* Remove from the hash table all expressions that reference memory. */
5855 static void
5856 invalidate_memory (void)
5858 int i;
5859 struct table_elt *p, *next;
5861 for (i = 0; i < HASH_SIZE; i++)
5862 for (p = table[i]; p; p = next)
5864 next = p->next_same_hash;
5865 if (p->in_memory)
5866 remove_from_table (p, i);
5870 /* Perform invalidation on the basis of everything about an insn
5871 except for invalidating the actual places that are SET in it.
5872 This includes the places CLOBBERed, and anything that might
5873 alias with something that is SET or CLOBBERed.
5875 X is the pattern of the insn. */
5877 static void
5878 invalidate_from_clobbers (rtx x)
5880 if (GET_CODE (x) == CLOBBER)
5882 rtx ref = XEXP (x, 0);
5883 if (ref)
5885 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5886 || MEM_P (ref))
5887 invalidate (ref, VOIDmode);
5888 else if (GET_CODE (ref) == STRICT_LOW_PART
5889 || GET_CODE (ref) == ZERO_EXTRACT)
5890 invalidate (XEXP (ref, 0), GET_MODE (ref));
5893 else if (GET_CODE (x) == PARALLEL)
5895 int i;
5896 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5898 rtx y = XVECEXP (x, 0, i);
5899 if (GET_CODE (y) == CLOBBER)
5901 rtx ref = XEXP (y, 0);
5902 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5903 || MEM_P (ref))
5904 invalidate (ref, VOIDmode);
5905 else if (GET_CODE (ref) == STRICT_LOW_PART
5906 || GET_CODE (ref) == ZERO_EXTRACT)
5907 invalidate (XEXP (ref, 0), GET_MODE (ref));
5913 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
5914 and replace any registers in them with either an equivalent constant
5915 or the canonical form of the register. If we are inside an address,
5916 only do this if the address remains valid.
5918 OBJECT is 0 except when within a MEM in which case it is the MEM.
5920 Return the replacement for X. */
5922 static rtx
5923 cse_process_notes_1 (rtx x, rtx object, bool *changed)
5925 enum rtx_code code = GET_CODE (x);
5926 const char *fmt = GET_RTX_FORMAT (code);
5927 int i;
5929 switch (code)
5931 case CONST_INT:
5932 case CONST:
5933 case SYMBOL_REF:
5934 case LABEL_REF:
5935 case CONST_DOUBLE:
5936 case CONST_FIXED:
5937 case CONST_VECTOR:
5938 case PC:
5939 case CC0:
5940 case LO_SUM:
5941 return x;
5943 case MEM:
5944 validate_change (x, &XEXP (x, 0),
5945 cse_process_notes (XEXP (x, 0), x, changed), 0);
5946 return x;
5948 case EXPR_LIST:
5949 case INSN_LIST:
5950 if (REG_NOTE_KIND (x) == REG_EQUAL)
5951 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
5952 if (XEXP (x, 1))
5953 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
5954 return x;
5956 case SIGN_EXTEND:
5957 case ZERO_EXTEND:
5958 case SUBREG:
5960 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
5961 /* We don't substitute VOIDmode constants into these rtx,
5962 since they would impede folding. */
5963 if (GET_MODE (new_rtx) != VOIDmode)
5964 validate_change (object, &XEXP (x, 0), new_rtx, 0);
5965 return x;
5968 case REG:
5969 i = REG_QTY (REGNO (x));
5971 /* Return a constant or a constant register. */
5972 if (REGNO_QTY_VALID_P (REGNO (x)))
5974 struct qty_table_elem *ent = &qty_table[i];
5976 if (ent->const_rtx != NULL_RTX
5977 && (CONSTANT_P (ent->const_rtx)
5978 || REG_P (ent->const_rtx)))
5980 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
5981 if (new_rtx)
5982 return copy_rtx (new_rtx);
5986 /* Otherwise, canonicalize this register. */
5987 return canon_reg (x, NULL_RTX);
5989 default:
5990 break;
5993 for (i = 0; i < GET_RTX_LENGTH (code); i++)
5994 if (fmt[i] == 'e')
5995 validate_change (object, &XEXP (x, i),
5996 cse_process_notes (XEXP (x, i), object, changed), 0);
5998 return x;
6001 static rtx
6002 cse_process_notes (rtx x, rtx object, bool *changed)
6004 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6005 if (new_rtx != x)
6006 *changed = true;
6007 return new_rtx;
6011 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6013 DATA is a pointer to a struct cse_basic_block_data, that is used to
6014 describe the path.
6015 It is filled with a queue of basic blocks, starting with FIRST_BB
6016 and following a trace through the CFG.
6018 If all paths starting at FIRST_BB have been followed, or no new path
6019 starting at FIRST_BB can be constructed, this function returns FALSE.
6020 Otherwise, DATA->path is filled and the function returns TRUE indicating
6021 that a path to follow was found.
6023 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6024 block in the path will be FIRST_BB. */
6026 static bool
6027 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6028 int follow_jumps)
6030 basic_block bb;
6031 edge e;
6032 int path_size;
6034 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6036 /* See if there is a previous path. */
6037 path_size = data->path_size;
6039 /* There is a previous path. Make sure it started with FIRST_BB. */
6040 if (path_size)
6041 gcc_assert (data->path[0].bb == first_bb);
6043 /* There was only one basic block in the last path. Clear the path and
6044 return, so that paths starting at another basic block can be tried. */
6045 if (path_size == 1)
6047 path_size = 0;
6048 goto done;
6051 /* If the path was empty from the beginning, construct a new path. */
6052 if (path_size == 0)
6053 data->path[path_size++].bb = first_bb;
6054 else
6056 /* Otherwise, path_size must be equal to or greater than 2, because
6057 a previous path exists that is at least two basic blocks long.
6059 Update the previous branch path, if any. If the last branch was
6060 previously along the branch edge, take the fallthrough edge now. */
6061 while (path_size >= 2)
6063 basic_block last_bb_in_path, previous_bb_in_path;
6064 edge e;
6066 --path_size;
6067 last_bb_in_path = data->path[path_size].bb;
6068 previous_bb_in_path = data->path[path_size - 1].bb;
6070 /* If we previously followed a path along the branch edge, try
6071 the fallthru edge now. */
6072 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6073 && any_condjump_p (BB_END (previous_bb_in_path))
6074 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6075 && e == BRANCH_EDGE (previous_bb_in_path))
6077 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6078 if (bb != EXIT_BLOCK_PTR
6079 && single_pred_p (bb)
6080 /* We used to assert here that we would only see blocks
6081 that we have not visited yet. But we may end up
6082 visiting basic blocks twice if the CFG has changed
6083 in this run of cse_main, because when the CFG changes
6084 the topological sort of the CFG also changes. A basic
6085 blocks that previously had more than two predecessors
6086 may now have a single predecessor, and become part of
6087 a path that starts at another basic block.
6089 We still want to visit each basic block only once, so
6090 halt the path here if we have already visited BB. */
6091 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6093 SET_BIT (cse_visited_basic_blocks, bb->index);
6094 data->path[path_size++].bb = bb;
6095 break;
6099 data->path[path_size].bb = NULL;
6102 /* If only one block remains in the path, bail. */
6103 if (path_size == 1)
6105 path_size = 0;
6106 goto done;
6110 /* Extend the path if possible. */
6111 if (follow_jumps)
6113 bb = data->path[path_size - 1].bb;
6114 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6116 if (single_succ_p (bb))
6117 e = single_succ_edge (bb);
6118 else if (EDGE_COUNT (bb->succs) == 2
6119 && any_condjump_p (BB_END (bb)))
6121 /* First try to follow the branch. If that doesn't lead
6122 to a useful path, follow the fallthru edge. */
6123 e = BRANCH_EDGE (bb);
6124 if (!single_pred_p (e->dest))
6125 e = FALLTHRU_EDGE (bb);
6127 else
6128 e = NULL;
6130 if (e
6131 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6132 && e->dest != EXIT_BLOCK_PTR
6133 && single_pred_p (e->dest)
6134 /* Avoid visiting basic blocks twice. The large comment
6135 above explains why this can happen. */
6136 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6138 basic_block bb2 = e->dest;
6139 SET_BIT (cse_visited_basic_blocks, bb2->index);
6140 data->path[path_size++].bb = bb2;
6141 bb = bb2;
6143 else
6144 bb = NULL;
6148 done:
6149 data->path_size = path_size;
6150 return path_size != 0;
6153 /* Dump the path in DATA to file F. NSETS is the number of sets
6154 in the path. */
6156 static void
6157 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6159 int path_entry;
6161 fprintf (f, ";; Following path with %d sets: ", nsets);
6162 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6163 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6164 fputc ('\n', dump_file);
6165 fflush (f);
6169 /* Return true if BB has exception handling successor edges. */
6171 static bool
6172 have_eh_succ_edges (basic_block bb)
6174 edge e;
6175 edge_iterator ei;
6177 FOR_EACH_EDGE (e, ei, bb->succs)
6178 if (e->flags & EDGE_EH)
6179 return true;
6181 return false;
6185 /* Scan to the end of the path described by DATA. Return an estimate of
6186 the total number of SETs of all insns in the path. */
6188 static void
6189 cse_prescan_path (struct cse_basic_block_data *data)
6191 int nsets = 0;
6192 int path_size = data->path_size;
6193 int path_entry;
6195 /* Scan to end of each basic block in the path. */
6196 for (path_entry = 0; path_entry < path_size; path_entry++)
6198 basic_block bb;
6199 rtx insn;
6201 bb = data->path[path_entry].bb;
6203 FOR_BB_INSNS (bb, insn)
6205 if (!INSN_P (insn))
6206 continue;
6208 /* A PARALLEL can have lots of SETs in it,
6209 especially if it is really an ASM_OPERANDS. */
6210 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6211 nsets += XVECLEN (PATTERN (insn), 0);
6212 else
6213 nsets += 1;
6217 data->nsets = nsets;
6220 /* Process a single extended basic block described by EBB_DATA. */
6222 static void
6223 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6225 int path_size = ebb_data->path_size;
6226 int path_entry;
6227 int num_insns = 0;
6229 /* Allocate the space needed by qty_table. */
6230 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6232 new_basic_block ();
6233 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6234 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6235 for (path_entry = 0; path_entry < path_size; path_entry++)
6237 basic_block bb;
6238 rtx insn;
6240 bb = ebb_data->path[path_entry].bb;
6242 /* Invalidate recorded information for eh regs if there is an EH
6243 edge pointing to that bb. */
6244 if (bb_has_eh_pred (bb))
6246 df_ref *def_rec;
6248 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6250 df_ref def = *def_rec;
6251 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6252 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6256 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6257 FOR_BB_INSNS (bb, insn)
6259 /* If we have processed 1,000 insns, flush the hash table to
6260 avoid extreme quadratic behavior. We must not include NOTEs
6261 in the count since there may be more of them when generating
6262 debugging information. If we clear the table at different
6263 times, code generated with -g -O might be different than code
6264 generated with -O but not -g.
6266 FIXME: This is a real kludge and needs to be done some other
6267 way. */
6268 if (NONDEBUG_INSN_P (insn)
6269 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6271 flush_hash_table ();
6272 num_insns = 0;
6275 if (INSN_P (insn))
6277 /* Process notes first so we have all notes in canonical forms
6278 when looking for duplicate operations. */
6279 if (REG_NOTES (insn))
6281 bool changed = false;
6282 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6283 NULL_RTX, &changed);
6284 if (changed)
6285 df_notes_rescan (insn);
6288 cse_insn (insn);
6290 /* If we haven't already found an insn where we added a LABEL_REF,
6291 check this one. */
6292 if (INSN_P (insn) && !recorded_label_ref
6293 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6294 (void *) insn))
6295 recorded_label_ref = true;
6297 #ifdef HAVE_cc0
6298 if (NONDEBUG_INSN_P (insn))
6300 /* If the previous insn sets CC0 and this insn no
6301 longer references CC0, delete the previous insn.
6302 Here we use fact that nothing expects CC0 to be
6303 valid over an insn, which is true until the final
6304 pass. */
6305 rtx prev_insn, tem;
6307 prev_insn = prev_nonnote_nondebug_insn (insn);
6308 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6309 && (tem = single_set (prev_insn)) != NULL_RTX
6310 && SET_DEST (tem) == cc0_rtx
6311 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6312 delete_insn (prev_insn);
6314 /* If this insn is not the last insn in the basic
6315 block, it will be PREV_INSN(insn) in the next
6316 iteration. If we recorded any CC0-related
6317 information for this insn, remember it. */
6318 if (insn != BB_END (bb))
6320 prev_insn_cc0 = this_insn_cc0;
6321 prev_insn_cc0_mode = this_insn_cc0_mode;
6324 #endif
6328 /* With non-call exceptions, we are not always able to update
6329 the CFG properly inside cse_insn. So clean up possibly
6330 redundant EH edges here. */
6331 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6332 cse_cfg_altered |= purge_dead_edges (bb);
6334 /* If we changed a conditional jump, we may have terminated
6335 the path we are following. Check that by verifying that
6336 the edge we would take still exists. If the edge does
6337 not exist anymore, purge the remainder of the path.
6338 Note that this will cause us to return to the caller. */
6339 if (path_entry < path_size - 1)
6341 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6342 if (!find_edge (bb, next_bb))
6346 path_size--;
6348 /* If we truncate the path, we must also reset the
6349 visited bit on the remaining blocks in the path,
6350 or we will never visit them at all. */
6351 RESET_BIT (cse_visited_basic_blocks,
6352 ebb_data->path[path_size].bb->index);
6353 ebb_data->path[path_size].bb = NULL;
6355 while (path_size - 1 != path_entry);
6356 ebb_data->path_size = path_size;
6360 /* If this is a conditional jump insn, record any known
6361 equivalences due to the condition being tested. */
6362 insn = BB_END (bb);
6363 if (path_entry < path_size - 1
6364 && JUMP_P (insn)
6365 && single_set (insn)
6366 && any_condjump_p (insn))
6368 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6369 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6370 record_jump_equiv (insn, taken);
6373 #ifdef HAVE_cc0
6374 /* Clear the CC0-tracking related insns, they can't provide
6375 useful information across basic block boundaries. */
6376 prev_insn_cc0 = 0;
6377 #endif
6380 gcc_assert (next_qty <= max_qty);
6382 free (qty_table);
6386 /* Perform cse on the instructions of a function.
6387 F is the first instruction.
6388 NREGS is one plus the highest pseudo-reg number used in the instruction.
6390 Return 2 if jump optimizations should be redone due to simplifications
6391 in conditional jump instructions.
6392 Return 1 if the CFG should be cleaned up because it has been modified.
6393 Return 0 otherwise. */
6396 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6398 struct cse_basic_block_data ebb_data;
6399 basic_block bb;
6400 int *rc_order = XNEWVEC (int, last_basic_block);
6401 int i, n_blocks;
6403 df_set_flags (DF_LR_RUN_DCE);
6404 df_analyze ();
6405 df_set_flags (DF_DEFER_INSN_RESCAN);
6407 reg_scan (get_insns (), max_reg_num ());
6408 init_cse_reg_info (nregs);
6410 ebb_data.path = XNEWVEC (struct branch_path,
6411 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6413 cse_cfg_altered = false;
6414 cse_jumps_altered = false;
6415 recorded_label_ref = false;
6416 constant_pool_entries_cost = 0;
6417 constant_pool_entries_regcost = 0;
6418 ebb_data.path_size = 0;
6419 ebb_data.nsets = 0;
6420 rtl_hooks = cse_rtl_hooks;
6422 init_recog ();
6423 init_alias_analysis ();
6425 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6427 /* Set up the table of already visited basic blocks. */
6428 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6429 sbitmap_zero (cse_visited_basic_blocks);
6431 /* Loop over basic blocks in reverse completion order (RPO),
6432 excluding the ENTRY and EXIT blocks. */
6433 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6434 i = 0;
6435 while (i < n_blocks)
6437 /* Find the first block in the RPO queue that we have not yet
6438 processed before. */
6441 bb = BASIC_BLOCK (rc_order[i++]);
6443 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6444 && i < n_blocks);
6446 /* Find all paths starting with BB, and process them. */
6447 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6449 /* Pre-scan the path. */
6450 cse_prescan_path (&ebb_data);
6452 /* If this basic block has no sets, skip it. */
6453 if (ebb_data.nsets == 0)
6454 continue;
6456 /* Get a reasonable estimate for the maximum number of qty's
6457 needed for this path. For this, we take the number of sets
6458 and multiply that by MAX_RECOG_OPERANDS. */
6459 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6461 /* Dump the path we're about to process. */
6462 if (dump_file)
6463 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6465 cse_extended_basic_block (&ebb_data);
6469 /* Clean up. */
6470 end_alias_analysis ();
6471 free (reg_eqv_table);
6472 free (ebb_data.path);
6473 sbitmap_free (cse_visited_basic_blocks);
6474 free (rc_order);
6475 rtl_hooks = general_rtl_hooks;
6477 if (cse_jumps_altered || recorded_label_ref)
6478 return 2;
6479 else if (cse_cfg_altered)
6480 return 1;
6481 else
6482 return 0;
6485 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6486 which there isn't a REG_LABEL_OPERAND note.
6487 Return one if so. DATA is the insn. */
6489 static int
6490 check_for_label_ref (rtx *rtl, void *data)
6492 rtx insn = (rtx) data;
6494 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6495 note for it, we must rerun jump since it needs to place the note. If
6496 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6497 don't do this since no REG_LABEL_OPERAND will be added. */
6498 return (GET_CODE (*rtl) == LABEL_REF
6499 && ! LABEL_REF_NONLOCAL_P (*rtl)
6500 && (!JUMP_P (insn)
6501 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6502 && LABEL_P (XEXP (*rtl, 0))
6503 && INSN_UID (XEXP (*rtl, 0)) != 0
6504 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6507 /* Count the number of times registers are used (not set) in X.
6508 COUNTS is an array in which we accumulate the count, INCR is how much
6509 we count each register usage.
6511 Don't count a usage of DEST, which is the SET_DEST of a SET which
6512 contains X in its SET_SRC. This is because such a SET does not
6513 modify the liveness of DEST.
6514 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6515 We must then count uses of a SET_DEST regardless, because the insn can't be
6516 deleted here. */
6518 static void
6519 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6521 enum rtx_code code;
6522 rtx note;
6523 const char *fmt;
6524 int i, j;
6526 if (x == 0)
6527 return;
6529 switch (code = GET_CODE (x))
6531 case REG:
6532 if (x != dest)
6533 counts[REGNO (x)] += incr;
6534 return;
6536 case PC:
6537 case CC0:
6538 case CONST:
6539 case CONST_INT:
6540 case CONST_DOUBLE:
6541 case CONST_FIXED:
6542 case CONST_VECTOR:
6543 case SYMBOL_REF:
6544 case LABEL_REF:
6545 return;
6547 case CLOBBER:
6548 /* If we are clobbering a MEM, mark any registers inside the address
6549 as being used. */
6550 if (MEM_P (XEXP (x, 0)))
6551 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6552 return;
6554 case SET:
6555 /* Unless we are setting a REG, count everything in SET_DEST. */
6556 if (!REG_P (SET_DEST (x)))
6557 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6558 count_reg_usage (SET_SRC (x), counts,
6559 dest ? dest : SET_DEST (x),
6560 incr);
6561 return;
6563 case DEBUG_INSN:
6564 return;
6566 case CALL_INSN:
6567 case INSN:
6568 case JUMP_INSN:
6569 /* We expect dest to be NULL_RTX here. If the insn may trap,
6570 or if it cannot be deleted due to side-effects, mark this fact
6571 by setting DEST to pc_rtx. */
6572 if (insn_could_throw_p (x) || side_effects_p (PATTERN (x)))
6573 dest = pc_rtx;
6574 if (code == CALL_INSN)
6575 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6576 count_reg_usage (PATTERN (x), counts, dest, incr);
6578 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6579 use them. */
6581 note = find_reg_equal_equiv_note (x);
6582 if (note)
6584 rtx eqv = XEXP (note, 0);
6586 if (GET_CODE (eqv) == EXPR_LIST)
6587 /* This REG_EQUAL note describes the result of a function call.
6588 Process all the arguments. */
6591 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6592 eqv = XEXP (eqv, 1);
6594 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6595 else
6596 count_reg_usage (eqv, counts, dest, incr);
6598 return;
6600 case EXPR_LIST:
6601 if (REG_NOTE_KIND (x) == REG_EQUAL
6602 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6603 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6604 involving registers in the address. */
6605 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6606 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6608 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6609 return;
6611 case ASM_OPERANDS:
6612 /* Iterate over just the inputs, not the constraints as well. */
6613 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6614 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6615 return;
6617 case INSN_LIST:
6618 gcc_unreachable ();
6620 default:
6621 break;
6624 fmt = GET_RTX_FORMAT (code);
6625 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6627 if (fmt[i] == 'e')
6628 count_reg_usage (XEXP (x, i), counts, dest, incr);
6629 else if (fmt[i] == 'E')
6630 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6631 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6635 /* Return true if X is a dead register. */
6637 static inline int
6638 is_dead_reg (rtx x, int *counts)
6640 return (REG_P (x)
6641 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6642 && counts[REGNO (x)] == 0);
6645 /* Return true if set is live. */
6646 static bool
6647 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6648 int *counts)
6650 #ifdef HAVE_cc0
6651 rtx tem;
6652 #endif
6654 if (set_noop_p (set))
6657 #ifdef HAVE_cc0
6658 else if (GET_CODE (SET_DEST (set)) == CC0
6659 && !side_effects_p (SET_SRC (set))
6660 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6661 || !INSN_P (tem)
6662 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6663 return false;
6664 #endif
6665 else if (!is_dead_reg (SET_DEST (set), counts)
6666 || side_effects_p (SET_SRC (set)))
6667 return true;
6668 return false;
6671 /* Return true if insn is live. */
6673 static bool
6674 insn_live_p (rtx insn, int *counts)
6676 int i;
6677 if (insn_could_throw_p (insn))
6678 return true;
6679 else if (GET_CODE (PATTERN (insn)) == SET)
6680 return set_live_p (PATTERN (insn), insn, counts);
6681 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6683 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6685 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6687 if (GET_CODE (elt) == SET)
6689 if (set_live_p (elt, insn, counts))
6690 return true;
6692 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6693 return true;
6695 return false;
6697 else if (DEBUG_INSN_P (insn))
6699 rtx next;
6701 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6702 if (NOTE_P (next))
6703 continue;
6704 else if (!DEBUG_INSN_P (next))
6705 return true;
6706 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6707 return false;
6709 return true;
6711 else
6712 return true;
6715 /* Count the number of stores into pseudo. Callback for note_stores. */
6717 static void
6718 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6720 int *counts = (int *) data;
6721 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6722 counts[REGNO (x)]++;
6725 struct dead_debug_insn_data
6727 int *counts;
6728 rtx *replacements;
6729 bool seen_repl;
6732 /* Return if a DEBUG_INSN needs to be reset because some dead
6733 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6735 static int
6736 is_dead_debug_insn (rtx *loc, void *data)
6738 rtx x = *loc;
6739 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6741 if (is_dead_reg (x, ddid->counts))
6743 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6744 ddid->seen_repl = true;
6745 else
6746 return 1;
6748 return 0;
6751 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6752 Callback for simplify_replace_fn_rtx. */
6754 static rtx
6755 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6757 rtx *replacements = (rtx *) data;
6759 if (REG_P (x)
6760 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6761 && replacements[REGNO (x)] != NULL_RTX)
6763 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6764 return replacements[REGNO (x)];
6765 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6766 GET_MODE (replacements[REGNO (x)]));
6768 return NULL_RTX;
6771 /* Scan all the insns and delete any that are dead; i.e., they store a register
6772 that is never used or they copy a register to itself.
6774 This is used to remove insns made obviously dead by cse, loop or other
6775 optimizations. It improves the heuristics in loop since it won't try to
6776 move dead invariants out of loops or make givs for dead quantities. The
6777 remaining passes of the compilation are also sped up. */
6780 delete_trivially_dead_insns (rtx insns, int nreg)
6782 int *counts;
6783 rtx insn, prev;
6784 rtx *replacements = NULL;
6785 int ndead = 0;
6787 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6788 /* First count the number of times each register is used. */
6789 if (MAY_HAVE_DEBUG_INSNS)
6791 counts = XCNEWVEC (int, nreg * 3);
6792 for (insn = insns; insn; insn = NEXT_INSN (insn))
6793 if (DEBUG_INSN_P (insn))
6794 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6795 NULL_RTX, 1);
6796 else if (INSN_P (insn))
6798 count_reg_usage (insn, counts, NULL_RTX, 1);
6799 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6801 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6802 First one counts how many times each pseudo is used outside
6803 of debug insns, second counts how many times each pseudo is
6804 used in debug insns and third counts how many times a pseudo
6805 is stored. */
6807 else
6809 counts = XCNEWVEC (int, nreg);
6810 for (insn = insns; insn; insn = NEXT_INSN (insn))
6811 if (INSN_P (insn))
6812 count_reg_usage (insn, counts, NULL_RTX, 1);
6813 /* If no debug insns can be present, COUNTS is just an array
6814 which counts how many times each pseudo is used. */
6816 /* Go from the last insn to the first and delete insns that only set unused
6817 registers or copy a register to itself. As we delete an insn, remove
6818 usage counts for registers it uses.
6820 The first jump optimization pass may leave a real insn as the last
6821 insn in the function. We must not skip that insn or we may end
6822 up deleting code that is not really dead.
6824 If some otherwise unused register is only used in DEBUG_INSNs,
6825 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6826 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6827 has been created for the unused register, replace it with
6828 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6829 for (insn = get_last_insn (); insn; insn = prev)
6831 int live_insn = 0;
6833 prev = PREV_INSN (insn);
6834 if (!INSN_P (insn))
6835 continue;
6837 live_insn = insn_live_p (insn, counts);
6839 /* If this is a dead insn, delete it and show registers in it aren't
6840 being used. */
6842 if (! live_insn && dbg_cnt (delete_trivial_dead))
6844 if (DEBUG_INSN_P (insn))
6845 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6846 NULL_RTX, -1);
6847 else
6849 rtx set;
6850 if (MAY_HAVE_DEBUG_INSNS
6851 && (set = single_set (insn)) != NULL_RTX
6852 && is_dead_reg (SET_DEST (set), counts)
6853 /* Used at least once in some DEBUG_INSN. */
6854 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6855 /* And set exactly once. */
6856 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6857 && !side_effects_p (SET_SRC (set))
6858 && asm_noperands (PATTERN (insn)) < 0)
6860 rtx dval, bind;
6862 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6863 dval = make_debug_expr_from_rtl (SET_DEST (set));
6865 /* Emit a debug bind insn before the insn in which
6866 reg dies. */
6867 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
6868 DEBUG_EXPR_TREE_DECL (dval),
6869 SET_SRC (set),
6870 VAR_INIT_STATUS_INITIALIZED);
6871 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
6873 bind = emit_debug_insn_before (bind, insn);
6874 df_insn_rescan (bind);
6876 if (replacements == NULL)
6877 replacements = XCNEWVEC (rtx, nreg);
6878 replacements[REGNO (SET_DEST (set))] = dval;
6881 count_reg_usage (insn, counts, NULL_RTX, -1);
6882 ndead++;
6884 delete_insn_and_edges (insn);
6888 if (MAY_HAVE_DEBUG_INSNS)
6890 struct dead_debug_insn_data ddid;
6891 ddid.counts = counts;
6892 ddid.replacements = replacements;
6893 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
6894 if (DEBUG_INSN_P (insn))
6896 /* If this debug insn references a dead register that wasn't replaced
6897 with an DEBUG_EXPR, reset the DEBUG_INSN. */
6898 ddid.seen_repl = false;
6899 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
6900 is_dead_debug_insn, &ddid))
6902 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
6903 df_insn_rescan (insn);
6905 else if (ddid.seen_repl)
6907 INSN_VAR_LOCATION_LOC (insn)
6908 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
6909 NULL_RTX, replace_dead_reg,
6910 replacements);
6911 df_insn_rescan (insn);
6914 free (replacements);
6917 if (dump_file && ndead)
6918 fprintf (dump_file, "Deleted %i trivially dead insns\n",
6919 ndead);
6920 /* Clean up. */
6921 free (counts);
6922 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
6923 return ndead;
6926 /* This function is called via for_each_rtx. The argument, NEWREG, is
6927 a condition code register with the desired mode. If we are looking
6928 at the same register in a different mode, replace it with
6929 NEWREG. */
6931 static int
6932 cse_change_cc_mode (rtx *loc, void *data)
6934 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
6936 if (*loc
6937 && REG_P (*loc)
6938 && REGNO (*loc) == REGNO (args->newreg)
6939 && GET_MODE (*loc) != GET_MODE (args->newreg))
6941 validate_change (args->insn, loc, args->newreg, 1);
6943 return -1;
6945 return 0;
6948 /* Change the mode of any reference to the register REGNO (NEWREG) to
6949 GET_MODE (NEWREG) in INSN. */
6951 static void
6952 cse_change_cc_mode_insn (rtx insn, rtx newreg)
6954 struct change_cc_mode_args args;
6955 int success;
6957 if (!INSN_P (insn))
6958 return;
6960 args.insn = insn;
6961 args.newreg = newreg;
6963 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
6964 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
6966 /* If the following assertion was triggered, there is most probably
6967 something wrong with the cc_modes_compatible back end function.
6968 CC modes only can be considered compatible if the insn - with the mode
6969 replaced by any of the compatible modes - can still be recognized. */
6970 success = apply_change_group ();
6971 gcc_assert (success);
6974 /* Change the mode of any reference to the register REGNO (NEWREG) to
6975 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
6976 any instruction which modifies NEWREG. */
6978 static void
6979 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
6981 rtx insn;
6983 for (insn = start; insn != end; insn = NEXT_INSN (insn))
6985 if (! INSN_P (insn))
6986 continue;
6988 if (reg_set_p (newreg, insn))
6989 return;
6991 cse_change_cc_mode_insn (insn, newreg);
6995 /* BB is a basic block which finishes with CC_REG as a condition code
6996 register which is set to CC_SRC. Look through the successors of BB
6997 to find blocks which have a single predecessor (i.e., this one),
6998 and look through those blocks for an assignment to CC_REG which is
6999 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7000 permitted to change the mode of CC_SRC to a compatible mode. This
7001 returns VOIDmode if no equivalent assignments were found.
7002 Otherwise it returns the mode which CC_SRC should wind up with.
7003 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7004 but is passed unmodified down to recursive calls in order to prevent
7005 endless recursion.
7007 The main complexity in this function is handling the mode issues.
7008 We may have more than one duplicate which we can eliminate, and we
7009 try to find a mode which will work for multiple duplicates. */
7011 static enum machine_mode
7012 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7013 bool can_change_mode)
7015 bool found_equiv;
7016 enum machine_mode mode;
7017 unsigned int insn_count;
7018 edge e;
7019 rtx insns[2];
7020 enum machine_mode modes[2];
7021 rtx last_insns[2];
7022 unsigned int i;
7023 rtx newreg;
7024 edge_iterator ei;
7026 /* We expect to have two successors. Look at both before picking
7027 the final mode for the comparison. If we have more successors
7028 (i.e., some sort of table jump, although that seems unlikely),
7029 then we require all beyond the first two to use the same
7030 mode. */
7032 found_equiv = false;
7033 mode = GET_MODE (cc_src);
7034 insn_count = 0;
7035 FOR_EACH_EDGE (e, ei, bb->succs)
7037 rtx insn;
7038 rtx end;
7040 if (e->flags & EDGE_COMPLEX)
7041 continue;
7043 if (EDGE_COUNT (e->dest->preds) != 1
7044 || e->dest == EXIT_BLOCK_PTR
7045 /* Avoid endless recursion on unreachable blocks. */
7046 || e->dest == orig_bb)
7047 continue;
7049 end = NEXT_INSN (BB_END (e->dest));
7050 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7052 rtx set;
7054 if (! INSN_P (insn))
7055 continue;
7057 /* If CC_SRC is modified, we have to stop looking for
7058 something which uses it. */
7059 if (modified_in_p (cc_src, insn))
7060 break;
7062 /* Check whether INSN sets CC_REG to CC_SRC. */
7063 set = single_set (insn);
7064 if (set
7065 && REG_P (SET_DEST (set))
7066 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7068 bool found;
7069 enum machine_mode set_mode;
7070 enum machine_mode comp_mode;
7072 found = false;
7073 set_mode = GET_MODE (SET_SRC (set));
7074 comp_mode = set_mode;
7075 if (rtx_equal_p (cc_src, SET_SRC (set)))
7076 found = true;
7077 else if (GET_CODE (cc_src) == COMPARE
7078 && GET_CODE (SET_SRC (set)) == COMPARE
7079 && mode != set_mode
7080 && rtx_equal_p (XEXP (cc_src, 0),
7081 XEXP (SET_SRC (set), 0))
7082 && rtx_equal_p (XEXP (cc_src, 1),
7083 XEXP (SET_SRC (set), 1)))
7086 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7087 if (comp_mode != VOIDmode
7088 && (can_change_mode || comp_mode == mode))
7089 found = true;
7092 if (found)
7094 found_equiv = true;
7095 if (insn_count < ARRAY_SIZE (insns))
7097 insns[insn_count] = insn;
7098 modes[insn_count] = set_mode;
7099 last_insns[insn_count] = end;
7100 ++insn_count;
7102 if (mode != comp_mode)
7104 gcc_assert (can_change_mode);
7105 mode = comp_mode;
7107 /* The modified insn will be re-recognized later. */
7108 PUT_MODE (cc_src, mode);
7111 else
7113 if (set_mode != mode)
7115 /* We found a matching expression in the
7116 wrong mode, but we don't have room to
7117 store it in the array. Punt. This case
7118 should be rare. */
7119 break;
7121 /* INSN sets CC_REG to a value equal to CC_SRC
7122 with the right mode. We can simply delete
7123 it. */
7124 delete_insn (insn);
7127 /* We found an instruction to delete. Keep looking,
7128 in the hopes of finding a three-way jump. */
7129 continue;
7132 /* We found an instruction which sets the condition
7133 code, so don't look any farther. */
7134 break;
7137 /* If INSN sets CC_REG in some other way, don't look any
7138 farther. */
7139 if (reg_set_p (cc_reg, insn))
7140 break;
7143 /* If we fell off the bottom of the block, we can keep looking
7144 through successors. We pass CAN_CHANGE_MODE as false because
7145 we aren't prepared to handle compatibility between the
7146 further blocks and this block. */
7147 if (insn == end)
7149 enum machine_mode submode;
7151 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7152 if (submode != VOIDmode)
7154 gcc_assert (submode == mode);
7155 found_equiv = true;
7156 can_change_mode = false;
7161 if (! found_equiv)
7162 return VOIDmode;
7164 /* Now INSN_COUNT is the number of instructions we found which set
7165 CC_REG to a value equivalent to CC_SRC. The instructions are in
7166 INSNS. The modes used by those instructions are in MODES. */
7168 newreg = NULL_RTX;
7169 for (i = 0; i < insn_count; ++i)
7171 if (modes[i] != mode)
7173 /* We need to change the mode of CC_REG in INSNS[i] and
7174 subsequent instructions. */
7175 if (! newreg)
7177 if (GET_MODE (cc_reg) == mode)
7178 newreg = cc_reg;
7179 else
7180 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7182 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7183 newreg);
7186 delete_insn_and_edges (insns[i]);
7189 return mode;
7192 /* If we have a fixed condition code register (or two), walk through
7193 the instructions and try to eliminate duplicate assignments. */
7195 static void
7196 cse_condition_code_reg (void)
7198 unsigned int cc_regno_1;
7199 unsigned int cc_regno_2;
7200 rtx cc_reg_1;
7201 rtx cc_reg_2;
7202 basic_block bb;
7204 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7205 return;
7207 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7208 if (cc_regno_2 != INVALID_REGNUM)
7209 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7210 else
7211 cc_reg_2 = NULL_RTX;
7213 FOR_EACH_BB (bb)
7215 rtx last_insn;
7216 rtx cc_reg;
7217 rtx insn;
7218 rtx cc_src_insn;
7219 rtx cc_src;
7220 enum machine_mode mode;
7221 enum machine_mode orig_mode;
7223 /* Look for blocks which end with a conditional jump based on a
7224 condition code register. Then look for the instruction which
7225 sets the condition code register. Then look through the
7226 successor blocks for instructions which set the condition
7227 code register to the same value. There are other possible
7228 uses of the condition code register, but these are by far the
7229 most common and the ones which we are most likely to be able
7230 to optimize. */
7232 last_insn = BB_END (bb);
7233 if (!JUMP_P (last_insn))
7234 continue;
7236 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7237 cc_reg = cc_reg_1;
7238 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7239 cc_reg = cc_reg_2;
7240 else
7241 continue;
7243 cc_src_insn = NULL_RTX;
7244 cc_src = NULL_RTX;
7245 for (insn = PREV_INSN (last_insn);
7246 insn && insn != PREV_INSN (BB_HEAD (bb));
7247 insn = PREV_INSN (insn))
7249 rtx set;
7251 if (! INSN_P (insn))
7252 continue;
7253 set = single_set (insn);
7254 if (set
7255 && REG_P (SET_DEST (set))
7256 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7258 cc_src_insn = insn;
7259 cc_src = SET_SRC (set);
7260 break;
7262 else if (reg_set_p (cc_reg, insn))
7263 break;
7266 if (! cc_src_insn)
7267 continue;
7269 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7270 continue;
7272 /* Now CC_REG is a condition code register used for a
7273 conditional jump at the end of the block, and CC_SRC, in
7274 CC_SRC_INSN, is the value to which that condition code
7275 register is set, and CC_SRC is still meaningful at the end of
7276 the basic block. */
7278 orig_mode = GET_MODE (cc_src);
7279 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7280 if (mode != VOIDmode)
7282 gcc_assert (mode == GET_MODE (cc_src));
7283 if (mode != orig_mode)
7285 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7287 cse_change_cc_mode_insn (cc_src_insn, newreg);
7289 /* Do the same in the following insns that use the
7290 current value of CC_REG within BB. */
7291 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7292 NEXT_INSN (last_insn),
7293 newreg);
7300 /* Perform common subexpression elimination. Nonzero value from
7301 `cse_main' means that jumps were simplified and some code may now
7302 be unreachable, so do jump optimization again. */
7303 static bool
7304 gate_handle_cse (void)
7306 return optimize > 0;
7309 static unsigned int
7310 rest_of_handle_cse (void)
7312 int tem;
7314 if (dump_file)
7315 dump_flow_info (dump_file, dump_flags);
7317 tem = cse_main (get_insns (), max_reg_num ());
7319 /* If we are not running more CSE passes, then we are no longer
7320 expecting CSE to be run. But always rerun it in a cheap mode. */
7321 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7323 if (tem == 2)
7325 timevar_push (TV_JUMP);
7326 rebuild_jump_labels (get_insns ());
7327 cleanup_cfg (0);
7328 timevar_pop (TV_JUMP);
7330 else if (tem == 1 || optimize > 1)
7331 cleanup_cfg (0);
7333 return 0;
7336 struct rtl_opt_pass pass_cse =
7339 RTL_PASS,
7340 "cse1", /* name */
7341 gate_handle_cse, /* gate */
7342 rest_of_handle_cse, /* execute */
7343 NULL, /* sub */
7344 NULL, /* next */
7345 0, /* static_pass_number */
7346 TV_CSE, /* tv_id */
7347 0, /* properties_required */
7348 0, /* properties_provided */
7349 0, /* properties_destroyed */
7350 0, /* todo_flags_start */
7351 TODO_df_finish | TODO_verify_rtl_sharing |
7352 TODO_ggc_collect |
7353 TODO_verify_flow, /* todo_flags_finish */
7358 static bool
7359 gate_handle_cse2 (void)
7361 return optimize > 0 && flag_rerun_cse_after_loop;
7364 /* Run second CSE pass after loop optimizations. */
7365 static unsigned int
7366 rest_of_handle_cse2 (void)
7368 int tem;
7370 if (dump_file)
7371 dump_flow_info (dump_file, dump_flags);
7373 tem = cse_main (get_insns (), max_reg_num ());
7375 /* Run a pass to eliminate duplicated assignments to condition code
7376 registers. We have to run this after bypass_jumps, because it
7377 makes it harder for that pass to determine whether a jump can be
7378 bypassed safely. */
7379 cse_condition_code_reg ();
7381 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7383 if (tem == 2)
7385 timevar_push (TV_JUMP);
7386 rebuild_jump_labels (get_insns ());
7387 cleanup_cfg (0);
7388 timevar_pop (TV_JUMP);
7390 else if (tem == 1)
7391 cleanup_cfg (0);
7393 cse_not_expected = 1;
7394 return 0;
7398 struct rtl_opt_pass pass_cse2 =
7401 RTL_PASS,
7402 "cse2", /* name */
7403 gate_handle_cse2, /* gate */
7404 rest_of_handle_cse2, /* execute */
7405 NULL, /* sub */
7406 NULL, /* next */
7407 0, /* static_pass_number */
7408 TV_CSE2, /* tv_id */
7409 0, /* properties_required */
7410 0, /* properties_provided */
7411 0, /* properties_destroyed */
7412 0, /* todo_flags_start */
7413 TODO_df_finish | TODO_verify_rtl_sharing |
7414 TODO_ggc_collect |
7415 TODO_verify_flow /* todo_flags_finish */
7419 static bool
7420 gate_handle_cse_after_global_opts (void)
7422 return optimize > 0 && flag_rerun_cse_after_global_opts;
7425 /* Run second CSE pass after loop optimizations. */
7426 static unsigned int
7427 rest_of_handle_cse_after_global_opts (void)
7429 int save_cfj;
7430 int tem;
7432 /* We only want to do local CSE, so don't follow jumps. */
7433 save_cfj = flag_cse_follow_jumps;
7434 flag_cse_follow_jumps = 0;
7436 rebuild_jump_labels (get_insns ());
7437 tem = cse_main (get_insns (), max_reg_num ());
7438 purge_all_dead_edges ();
7439 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7441 cse_not_expected = !flag_rerun_cse_after_loop;
7443 /* If cse altered any jumps, rerun jump opts to clean things up. */
7444 if (tem == 2)
7446 timevar_push (TV_JUMP);
7447 rebuild_jump_labels (get_insns ());
7448 cleanup_cfg (0);
7449 timevar_pop (TV_JUMP);
7451 else if (tem == 1)
7452 cleanup_cfg (0);
7454 flag_cse_follow_jumps = save_cfj;
7455 return 0;
7458 struct rtl_opt_pass pass_cse_after_global_opts =
7461 RTL_PASS,
7462 "cse_local", /* name */
7463 gate_handle_cse_after_global_opts, /* gate */
7464 rest_of_handle_cse_after_global_opts, /* execute */
7465 NULL, /* sub */
7466 NULL, /* next */
7467 0, /* static_pass_number */
7468 TV_CSE, /* tv_id */
7469 0, /* properties_required */
7470 0, /* properties_provided */
7471 0, /* properties_destroyed */
7472 0, /* todo_flags_start */
7473 TODO_df_finish | TODO_verify_rtl_sharing |
7474 TODO_ggc_collect |
7475 TODO_verify_flow /* todo_flags_finish */