* g++.dg/debug/dwarf2/ref-3.C: XFAIL AIX.
[official-gcc.git] / gcc / cse.c
blob4b3cd2e1df4a47db580a640e0473c6e3aaeac32e
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2016 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "insn-config.h"
32 #include "regs.h"
33 #include "emit-rtl.h"
34 #include "recog.h"
35 #include "cfgrtl.h"
36 #include "cfganal.h"
37 #include "cfgcleanup.h"
38 #include "alias.h"
39 #include "toplev.h"
40 #include "params.h"
41 #include "rtlhooks-def.h"
42 #include "tree-pass.h"
43 #include "dbgcnt.h"
44 #include "rtl-iter.h"
46 #ifndef LOAD_EXTEND_OP
47 #define LOAD_EXTEND_OP(M) UNKNOWN
48 #endif
50 /* The basic idea of common subexpression elimination is to go
51 through the code, keeping a record of expressions that would
52 have the same value at the current scan point, and replacing
53 expressions encountered with the cheapest equivalent expression.
55 It is too complicated to keep track of the different possibilities
56 when control paths merge in this code; so, at each label, we forget all
57 that is known and start fresh. This can be described as processing each
58 extended basic block separately. We have a separate pass to perform
59 global CSE.
61 Note CSE can turn a conditional or computed jump into a nop or
62 an unconditional jump. When this occurs we arrange to run the jump
63 optimizer after CSE to delete the unreachable code.
65 We use two data structures to record the equivalent expressions:
66 a hash table for most expressions, and a vector of "quantity
67 numbers" to record equivalent (pseudo) registers.
69 The use of the special data structure for registers is desirable
70 because it is faster. It is possible because registers references
71 contain a fairly small number, the register number, taken from
72 a contiguously allocated series, and two register references are
73 identical if they have the same number. General expressions
74 do not have any such thing, so the only way to retrieve the
75 information recorded on an expression other than a register
76 is to keep it in a hash table.
78 Registers and "quantity numbers":
80 At the start of each basic block, all of the (hardware and pseudo)
81 registers used in the function are given distinct quantity
82 numbers to indicate their contents. During scan, when the code
83 copies one register into another, we copy the quantity number.
84 When a register is loaded in any other way, we allocate a new
85 quantity number to describe the value generated by this operation.
86 `REG_QTY (N)' records what quantity register N is currently thought
87 of as containing.
89 All real quantity numbers are greater than or equal to zero.
90 If register N has not been assigned a quantity, `REG_QTY (N)' will
91 equal -N - 1, which is always negative.
93 Quantity numbers below zero do not exist and none of the `qty_table'
94 entries should be referenced with a negative index.
96 We also maintain a bidirectional chain of registers for each
97 quantity number. The `qty_table` members `first_reg' and `last_reg',
98 and `reg_eqv_table' members `next' and `prev' hold these chains.
100 The first register in a chain is the one whose lifespan is least local.
101 Among equals, it is the one that was seen first.
102 We replace any equivalent register with that one.
104 If two registers have the same quantity number, it must be true that
105 REG expressions with qty_table `mode' must be in the hash table for both
106 registers and must be in the same class.
108 The converse is not true. Since hard registers may be referenced in
109 any mode, two REG expressions might be equivalent in the hash table
110 but not have the same quantity number if the quantity number of one
111 of the registers is not the same mode as those expressions.
113 Constants and quantity numbers
115 When a quantity has a known constant value, that value is stored
116 in the appropriate qty_table `const_rtx'. This is in addition to
117 putting the constant in the hash table as is usual for non-regs.
119 Whether a reg or a constant is preferred is determined by the configuration
120 macro CONST_COSTS and will often depend on the constant value. In any
121 event, expressions containing constants can be simplified, by fold_rtx.
123 When a quantity has a known nearly constant value (such as an address
124 of a stack slot), that value is stored in the appropriate qty_table
125 `const_rtx'.
127 Integer constants don't have a machine mode. However, cse
128 determines the intended machine mode from the destination
129 of the instruction that moves the constant. The machine mode
130 is recorded in the hash table along with the actual RTL
131 constant expression so that different modes are kept separate.
133 Other expressions:
135 To record known equivalences among expressions in general
136 we use a hash table called `table'. It has a fixed number of buckets
137 that contain chains of `struct table_elt' elements for expressions.
138 These chains connect the elements whose expressions have the same
139 hash codes.
141 Other chains through the same elements connect the elements which
142 currently have equivalent values.
144 Register references in an expression are canonicalized before hashing
145 the expression. This is done using `reg_qty' and qty_table `first_reg'.
146 The hash code of a register reference is computed using the quantity
147 number, not the register number.
149 When the value of an expression changes, it is necessary to remove from the
150 hash table not just that expression but all expressions whose values
151 could be different as a result.
153 1. If the value changing is in memory, except in special cases
154 ANYTHING referring to memory could be changed. That is because
155 nobody knows where a pointer does not point.
156 The function `invalidate_memory' removes what is necessary.
158 The special cases are when the address is constant or is
159 a constant plus a fixed register such as the frame pointer
160 or a static chain pointer. When such addresses are stored in,
161 we can tell exactly which other such addresses must be invalidated
162 due to overlap. `invalidate' does this.
163 All expressions that refer to non-constant
164 memory addresses are also invalidated. `invalidate_memory' does this.
166 2. If the value changing is a register, all expressions
167 containing references to that register, and only those,
168 must be removed.
170 Because searching the entire hash table for expressions that contain
171 a register is very slow, we try to figure out when it isn't necessary.
172 Precisely, this is necessary only when expressions have been
173 entered in the hash table using this register, and then the value has
174 changed, and then another expression wants to be added to refer to
175 the register's new value. This sequence of circumstances is rare
176 within any one basic block.
178 `REG_TICK' and `REG_IN_TABLE', accessors for members of
179 cse_reg_info, are used to detect this case. REG_TICK (i) is
180 incremented whenever a value is stored in register i.
181 REG_IN_TABLE (i) holds -1 if no references to register i have been
182 entered in the table; otherwise, it contains the value REG_TICK (i)
183 had when the references were entered. If we want to enter a
184 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
185 remove old references. Until we want to enter a new entry, the
186 mere fact that the two vectors don't match makes the entries be
187 ignored if anyone tries to match them.
189 Registers themselves are entered in the hash table as well as in
190 the equivalent-register chains. However, `REG_TICK' and
191 `REG_IN_TABLE' do not apply to expressions which are simple
192 register references. These expressions are removed from the table
193 immediately when they become invalid, and this can be done even if
194 we do not immediately search for all the expressions that refer to
195 the register.
197 A CLOBBER rtx in an instruction invalidates its operand for further
198 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
199 invalidates everything that resides in memory.
201 Related expressions:
203 Constant expressions that differ only by an additive integer
204 are called related. When a constant expression is put in
205 the table, the related expression with no constant term
206 is also entered. These are made to point at each other
207 so that it is possible to find out if there exists any
208 register equivalent to an expression related to a given expression. */
210 /* Length of qty_table vector. We know in advance we will not need
211 a quantity number this big. */
213 static int max_qty;
215 /* Next quantity number to be allocated.
216 This is 1 + the largest number needed so far. */
218 static int next_qty;
220 /* Per-qty information tracking.
222 `first_reg' and `last_reg' track the head and tail of the
223 chain of registers which currently contain this quantity.
225 `mode' contains the machine mode of this quantity.
227 `const_rtx' holds the rtx of the constant value of this
228 quantity, if known. A summations of the frame/arg pointer
229 and a constant can also be entered here. When this holds
230 a known value, `const_insn' is the insn which stored the
231 constant value.
233 `comparison_{code,const,qty}' are used to track when a
234 comparison between a quantity and some constant or register has
235 been passed. In such a case, we know the results of the comparison
236 in case we see it again. These members record a comparison that
237 is known to be true. `comparison_code' holds the rtx code of such
238 a comparison, else it is set to UNKNOWN and the other two
239 comparison members are undefined. `comparison_const' holds
240 the constant being compared against, or zero if the comparison
241 is not against a constant. `comparison_qty' holds the quantity
242 being compared against when the result is known. If the comparison
243 is not with a register, `comparison_qty' is -1. */
245 struct qty_table_elem
247 rtx const_rtx;
248 rtx_insn *const_insn;
249 rtx comparison_const;
250 int comparison_qty;
251 unsigned int first_reg, last_reg;
252 /* The sizes of these fields should match the sizes of the
253 code and mode fields of struct rtx_def (see rtl.h). */
254 ENUM_BITFIELD(rtx_code) comparison_code : 16;
255 ENUM_BITFIELD(machine_mode) mode : 8;
258 /* The table of all qtys, indexed by qty number. */
259 static struct qty_table_elem *qty_table;
261 /* For machines that have a CC0, we do not record its value in the hash
262 table since its use is guaranteed to be the insn immediately following
263 its definition and any other insn is presumed to invalidate it.
265 Instead, we store below the current and last value assigned to CC0.
266 If it should happen to be a constant, it is stored in preference
267 to the actual assigned value. In case it is a constant, we store
268 the mode in which the constant should be interpreted. */
270 static rtx this_insn_cc0, prev_insn_cc0;
271 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
273 /* Insn being scanned. */
275 static rtx_insn *this_insn;
276 static bool optimize_this_for_speed_p;
278 /* Index by register number, gives the number of the next (or
279 previous) register in the chain of registers sharing the same
280 value.
282 Or -1 if this register is at the end of the chain.
284 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
286 /* Per-register equivalence chain. */
287 struct reg_eqv_elem
289 int next, prev;
292 /* The table of all register equivalence chains. */
293 static struct reg_eqv_elem *reg_eqv_table;
295 struct cse_reg_info
297 /* The timestamp at which this register is initialized. */
298 unsigned int timestamp;
300 /* The quantity number of the register's current contents. */
301 int reg_qty;
303 /* The number of times the register has been altered in the current
304 basic block. */
305 int reg_tick;
307 /* The REG_TICK value at which rtx's containing this register are
308 valid in the hash table. If this does not equal the current
309 reg_tick value, such expressions existing in the hash table are
310 invalid. */
311 int reg_in_table;
313 /* The SUBREG that was set when REG_TICK was last incremented. Set
314 to -1 if the last store was to the whole register, not a subreg. */
315 unsigned int subreg_ticked;
318 /* A table of cse_reg_info indexed by register numbers. */
319 static struct cse_reg_info *cse_reg_info_table;
321 /* The size of the above table. */
322 static unsigned int cse_reg_info_table_size;
324 /* The index of the first entry that has not been initialized. */
325 static unsigned int cse_reg_info_table_first_uninitialized;
327 /* The timestamp at the beginning of the current run of
328 cse_extended_basic_block. We increment this variable at the beginning of
329 the current run of cse_extended_basic_block. The timestamp field of a
330 cse_reg_info entry matches the value of this variable if and only
331 if the entry has been initialized during the current run of
332 cse_extended_basic_block. */
333 static unsigned int cse_reg_info_timestamp;
335 /* A HARD_REG_SET containing all the hard registers for which there is
336 currently a REG expression in the hash table. Note the difference
337 from the above variables, which indicate if the REG is mentioned in some
338 expression in the table. */
340 static HARD_REG_SET hard_regs_in_table;
342 /* True if CSE has altered the CFG. */
343 static bool cse_cfg_altered;
345 /* True if CSE has altered conditional jump insns in such a way
346 that jump optimization should be redone. */
347 static bool cse_jumps_altered;
349 /* True if we put a LABEL_REF into the hash table for an INSN
350 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
351 to put in the note. */
352 static bool recorded_label_ref;
354 /* canon_hash stores 1 in do_not_record
355 if it notices a reference to CC0, PC, or some other volatile
356 subexpression. */
358 static int do_not_record;
360 /* canon_hash stores 1 in hash_arg_in_memory
361 if it notices a reference to memory within the expression being hashed. */
363 static int hash_arg_in_memory;
365 /* The hash table contains buckets which are chains of `struct table_elt's,
366 each recording one expression's information.
367 That expression is in the `exp' field.
369 The canon_exp field contains a canonical (from the point of view of
370 alias analysis) version of the `exp' field.
372 Those elements with the same hash code are chained in both directions
373 through the `next_same_hash' and `prev_same_hash' fields.
375 Each set of expressions with equivalent values
376 are on a two-way chain through the `next_same_value'
377 and `prev_same_value' fields, and all point with
378 the `first_same_value' field at the first element in
379 that chain. The chain is in order of increasing cost.
380 Each element's cost value is in its `cost' field.
382 The `in_memory' field is nonzero for elements that
383 involve any reference to memory. These elements are removed
384 whenever a write is done to an unidentified location in memory.
385 To be safe, we assume that a memory address is unidentified unless
386 the address is either a symbol constant or a constant plus
387 the frame pointer or argument pointer.
389 The `related_value' field is used to connect related expressions
390 (that differ by adding an integer).
391 The related expressions are chained in a circular fashion.
392 `related_value' is zero for expressions for which this
393 chain is not useful.
395 The `cost' field stores the cost of this element's expression.
396 The `regcost' field stores the value returned by approx_reg_cost for
397 this element's expression.
399 The `is_const' flag is set if the element is a constant (including
400 a fixed address).
402 The `flag' field is used as a temporary during some search routines.
404 The `mode' field is usually the same as GET_MODE (`exp'), but
405 if `exp' is a CONST_INT and has no machine mode then the `mode'
406 field is the mode it was being used as. Each constant is
407 recorded separately for each mode it is used with. */
409 struct table_elt
411 rtx exp;
412 rtx canon_exp;
413 struct table_elt *next_same_hash;
414 struct table_elt *prev_same_hash;
415 struct table_elt *next_same_value;
416 struct table_elt *prev_same_value;
417 struct table_elt *first_same_value;
418 struct table_elt *related_value;
419 int cost;
420 int regcost;
421 /* The size of this field should match the size
422 of the mode field of struct rtx_def (see rtl.h). */
423 ENUM_BITFIELD(machine_mode) mode : 8;
424 char in_memory;
425 char is_const;
426 char flag;
429 /* We don't want a lot of buckets, because we rarely have very many
430 things stored in the hash table, and a lot of buckets slows
431 down a lot of loops that happen frequently. */
432 #define HASH_SHIFT 5
433 #define HASH_SIZE (1 << HASH_SHIFT)
434 #define HASH_MASK (HASH_SIZE - 1)
436 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
437 register (hard registers may require `do_not_record' to be set). */
439 #define HASH(X, M) \
440 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
441 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
442 : canon_hash (X, M)) & HASH_MASK)
444 /* Like HASH, but without side-effects. */
445 #define SAFE_HASH(X, M) \
446 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
447 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
448 : safe_hash (X, M)) & HASH_MASK)
450 /* Determine whether register number N is considered a fixed register for the
451 purpose of approximating register costs.
452 It is desirable to replace other regs with fixed regs, to reduce need for
453 non-fixed hard regs.
454 A reg wins if it is either the frame pointer or designated as fixed. */
455 #define FIXED_REGNO_P(N) \
456 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
457 || fixed_regs[N] || global_regs[N])
459 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
460 hard registers and pointers into the frame are the cheapest with a cost
461 of 0. Next come pseudos with a cost of one and other hard registers with
462 a cost of 2. Aside from these special cases, call `rtx_cost'. */
464 #define CHEAP_REGNO(N) \
465 (REGNO_PTR_FRAME_P (N) \
466 || (HARD_REGISTER_NUM_P (N) \
467 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
469 #define COST(X, MODE) \
470 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
471 #define COST_IN(X, MODE, OUTER, OPNO) \
472 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
474 /* Get the number of times this register has been updated in this
475 basic block. */
477 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
479 /* Get the point at which REG was recorded in the table. */
481 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
483 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
484 SUBREG). */
486 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
488 /* Get the quantity number for REG. */
490 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
492 /* Determine if the quantity number for register X represents a valid index
493 into the qty_table. */
495 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
497 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
499 #define CHEAPER(X, Y) \
500 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
502 static struct table_elt *table[HASH_SIZE];
504 /* Chain of `struct table_elt's made so far for this function
505 but currently removed from the table. */
507 static struct table_elt *free_element_chain;
509 /* Set to the cost of a constant pool reference if one was found for a
510 symbolic constant. If this was found, it means we should try to
511 convert constants into constant pool entries if they don't fit in
512 the insn. */
514 static int constant_pool_entries_cost;
515 static int constant_pool_entries_regcost;
517 /* Trace a patch through the CFG. */
519 struct branch_path
521 /* The basic block for this path entry. */
522 basic_block bb;
525 /* This data describes a block that will be processed by
526 cse_extended_basic_block. */
528 struct cse_basic_block_data
530 /* Total number of SETs in block. */
531 int nsets;
532 /* Size of current branch path, if any. */
533 int path_size;
534 /* Current path, indicating which basic_blocks will be processed. */
535 struct branch_path *path;
539 /* Pointers to the live in/live out bitmaps for the boundaries of the
540 current EBB. */
541 static bitmap cse_ebb_live_in, cse_ebb_live_out;
543 /* A simple bitmap to track which basic blocks have been visited
544 already as part of an already processed extended basic block. */
545 static sbitmap cse_visited_basic_blocks;
547 static bool fixed_base_plus_p (rtx x);
548 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
549 static int preferable (int, int, int, int);
550 static void new_basic_block (void);
551 static void make_new_qty (unsigned int, machine_mode);
552 static void make_regs_eqv (unsigned int, unsigned int);
553 static void delete_reg_equiv (unsigned int);
554 static int mention_regs (rtx);
555 static int insert_regs (rtx, struct table_elt *, int);
556 static void remove_from_table (struct table_elt *, unsigned);
557 static void remove_pseudo_from_table (rtx, unsigned);
558 static struct table_elt *lookup (rtx, unsigned, machine_mode);
559 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
560 static rtx lookup_as_function (rtx, enum rtx_code);
561 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
562 machine_mode, int, int);
563 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
564 machine_mode);
565 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
566 static void invalidate (rtx, machine_mode);
567 static void remove_invalid_refs (unsigned int);
568 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
569 machine_mode);
570 static void rehash_using_reg (rtx);
571 static void invalidate_memory (void);
572 static void invalidate_for_call (void);
573 static rtx use_related_value (rtx, struct table_elt *);
575 static inline unsigned canon_hash (rtx, machine_mode);
576 static inline unsigned safe_hash (rtx, machine_mode);
577 static inline unsigned hash_rtx_string (const char *);
579 static rtx canon_reg (rtx, rtx_insn *);
580 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
581 machine_mode *,
582 machine_mode *);
583 static rtx fold_rtx (rtx, rtx_insn *);
584 static rtx equiv_constant (rtx);
585 static void record_jump_equiv (rtx_insn *, bool);
586 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
587 int);
588 static void cse_insn (rtx_insn *);
589 static void cse_prescan_path (struct cse_basic_block_data *);
590 static void invalidate_from_clobbers (rtx_insn *);
591 static void invalidate_from_sets_and_clobbers (rtx_insn *);
592 static rtx cse_process_notes (rtx, rtx, bool *);
593 static void cse_extended_basic_block (struct cse_basic_block_data *);
594 extern void dump_class (struct table_elt*);
595 static void get_cse_reg_info_1 (unsigned int regno);
596 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
598 static void flush_hash_table (void);
599 static bool insn_live_p (rtx_insn *, int *);
600 static bool set_live_p (rtx, rtx_insn *, int *);
601 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
602 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
603 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
604 bool);
607 #undef RTL_HOOKS_GEN_LOWPART
608 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
610 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
612 /* Nonzero if X has the form (PLUS frame-pointer integer). */
614 static bool
615 fixed_base_plus_p (rtx x)
617 switch (GET_CODE (x))
619 case REG:
620 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
621 return true;
622 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
623 return true;
624 return false;
626 case PLUS:
627 if (!CONST_INT_P (XEXP (x, 1)))
628 return false;
629 return fixed_base_plus_p (XEXP (x, 0));
631 default:
632 return false;
636 /* Dump the expressions in the equivalence class indicated by CLASSP.
637 This function is used only for debugging. */
638 DEBUG_FUNCTION void
639 dump_class (struct table_elt *classp)
641 struct table_elt *elt;
643 fprintf (stderr, "Equivalence chain for ");
644 print_rtl (stderr, classp->exp);
645 fprintf (stderr, ": \n");
647 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
649 print_rtl (stderr, elt->exp);
650 fprintf (stderr, "\n");
654 /* Return an estimate of the cost of the registers used in an rtx.
655 This is mostly the number of different REG expressions in the rtx;
656 however for some exceptions like fixed registers we use a cost of
657 0. If any other hard register reference occurs, return MAX_COST. */
659 static int
660 approx_reg_cost (const_rtx x)
662 int cost = 0;
663 subrtx_iterator::array_type array;
664 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
666 const_rtx x = *iter;
667 if (REG_P (x))
669 unsigned int regno = REGNO (x);
670 if (!CHEAP_REGNO (regno))
672 if (regno < FIRST_PSEUDO_REGISTER)
674 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
675 return MAX_COST;
676 cost += 2;
678 else
679 cost += 1;
683 return cost;
686 /* Return a negative value if an rtx A, whose costs are given by COST_A
687 and REGCOST_A, is more desirable than an rtx B.
688 Return a positive value if A is less desirable, or 0 if the two are
689 equally good. */
690 static int
691 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
693 /* First, get rid of cases involving expressions that are entirely
694 unwanted. */
695 if (cost_a != cost_b)
697 if (cost_a == MAX_COST)
698 return 1;
699 if (cost_b == MAX_COST)
700 return -1;
703 /* Avoid extending lifetimes of hardregs. */
704 if (regcost_a != regcost_b)
706 if (regcost_a == MAX_COST)
707 return 1;
708 if (regcost_b == MAX_COST)
709 return -1;
712 /* Normal operation costs take precedence. */
713 if (cost_a != cost_b)
714 return cost_a - cost_b;
715 /* Only if these are identical consider effects on register pressure. */
716 if (regcost_a != regcost_b)
717 return regcost_a - regcost_b;
718 return 0;
721 /* Internal function, to compute cost when X is not a register; called
722 from COST macro to keep it simple. */
724 static int
725 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
727 return ((GET_CODE (x) == SUBREG
728 && REG_P (SUBREG_REG (x))
729 && GET_MODE_CLASS (mode) == MODE_INT
730 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
731 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
732 && subreg_lowpart_p (x)
733 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x))))
735 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
739 /* Initialize CSE_REG_INFO_TABLE. */
741 static void
742 init_cse_reg_info (unsigned int nregs)
744 /* Do we need to grow the table? */
745 if (nregs > cse_reg_info_table_size)
747 unsigned int new_size;
749 if (cse_reg_info_table_size < 2048)
751 /* Compute a new size that is a power of 2 and no smaller
752 than the large of NREGS and 64. */
753 new_size = (cse_reg_info_table_size
754 ? cse_reg_info_table_size : 64);
756 while (new_size < nregs)
757 new_size *= 2;
759 else
761 /* If we need a big table, allocate just enough to hold
762 NREGS registers. */
763 new_size = nregs;
766 /* Reallocate the table with NEW_SIZE entries. */
767 free (cse_reg_info_table);
768 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
769 cse_reg_info_table_size = new_size;
770 cse_reg_info_table_first_uninitialized = 0;
773 /* Do we have all of the first NREGS entries initialized? */
774 if (cse_reg_info_table_first_uninitialized < nregs)
776 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
777 unsigned int i;
779 /* Put the old timestamp on newly allocated entries so that they
780 will all be considered out of date. We do not touch those
781 entries beyond the first NREGS entries to be nice to the
782 virtual memory. */
783 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
784 cse_reg_info_table[i].timestamp = old_timestamp;
786 cse_reg_info_table_first_uninitialized = nregs;
790 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
792 static void
793 get_cse_reg_info_1 (unsigned int regno)
795 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
796 entry will be considered to have been initialized. */
797 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
799 /* Initialize the rest of the entry. */
800 cse_reg_info_table[regno].reg_tick = 1;
801 cse_reg_info_table[regno].reg_in_table = -1;
802 cse_reg_info_table[regno].subreg_ticked = -1;
803 cse_reg_info_table[regno].reg_qty = -regno - 1;
806 /* Find a cse_reg_info entry for REGNO. */
808 static inline struct cse_reg_info *
809 get_cse_reg_info (unsigned int regno)
811 struct cse_reg_info *p = &cse_reg_info_table[regno];
813 /* If this entry has not been initialized, go ahead and initialize
814 it. */
815 if (p->timestamp != cse_reg_info_timestamp)
816 get_cse_reg_info_1 (regno);
818 return p;
821 /* Clear the hash table and initialize each register with its own quantity,
822 for a new basic block. */
824 static void
825 new_basic_block (void)
827 int i;
829 next_qty = 0;
831 /* Invalidate cse_reg_info_table. */
832 cse_reg_info_timestamp++;
834 /* Clear out hash table state for this pass. */
835 CLEAR_HARD_REG_SET (hard_regs_in_table);
837 /* The per-quantity values used to be initialized here, but it is
838 much faster to initialize each as it is made in `make_new_qty'. */
840 for (i = 0; i < HASH_SIZE; i++)
842 struct table_elt *first;
844 first = table[i];
845 if (first != NULL)
847 struct table_elt *last = first;
849 table[i] = NULL;
851 while (last->next_same_hash != NULL)
852 last = last->next_same_hash;
854 /* Now relink this hash entire chain into
855 the free element list. */
857 last->next_same_hash = free_element_chain;
858 free_element_chain = first;
862 prev_insn_cc0 = 0;
865 /* Say that register REG contains a quantity in mode MODE not in any
866 register before and initialize that quantity. */
868 static void
869 make_new_qty (unsigned int reg, machine_mode mode)
871 int q;
872 struct qty_table_elem *ent;
873 struct reg_eqv_elem *eqv;
875 gcc_assert (next_qty < max_qty);
877 q = REG_QTY (reg) = next_qty++;
878 ent = &qty_table[q];
879 ent->first_reg = reg;
880 ent->last_reg = reg;
881 ent->mode = mode;
882 ent->const_rtx = ent->const_insn = NULL;
883 ent->comparison_code = UNKNOWN;
885 eqv = &reg_eqv_table[reg];
886 eqv->next = eqv->prev = -1;
889 /* Make reg NEW equivalent to reg OLD.
890 OLD is not changing; NEW is. */
892 static void
893 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
895 unsigned int lastr, firstr;
896 int q = REG_QTY (old_reg);
897 struct qty_table_elem *ent;
899 ent = &qty_table[q];
901 /* Nothing should become eqv until it has a "non-invalid" qty number. */
902 gcc_assert (REGNO_QTY_VALID_P (old_reg));
904 REG_QTY (new_reg) = q;
905 firstr = ent->first_reg;
906 lastr = ent->last_reg;
908 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
909 hard regs. Among pseudos, if NEW will live longer than any other reg
910 of the same qty, and that is beyond the current basic block,
911 make it the new canonical replacement for this qty. */
912 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
913 /* Certain fixed registers might be of the class NO_REGS. This means
914 that not only can they not be allocated by the compiler, but
915 they cannot be used in substitutions or canonicalizations
916 either. */
917 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
918 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
919 || (new_reg >= FIRST_PSEUDO_REGISTER
920 && (firstr < FIRST_PSEUDO_REGISTER
921 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
922 && !bitmap_bit_p (cse_ebb_live_out, firstr))
923 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
924 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
926 reg_eqv_table[firstr].prev = new_reg;
927 reg_eqv_table[new_reg].next = firstr;
928 reg_eqv_table[new_reg].prev = -1;
929 ent->first_reg = new_reg;
931 else
933 /* If NEW is a hard reg (known to be non-fixed), insert at end.
934 Otherwise, insert before any non-fixed hard regs that are at the
935 end. Registers of class NO_REGS cannot be used as an
936 equivalent for anything. */
937 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
938 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
939 && new_reg >= FIRST_PSEUDO_REGISTER)
940 lastr = reg_eqv_table[lastr].prev;
941 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
942 if (reg_eqv_table[lastr].next >= 0)
943 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
944 else
945 qty_table[q].last_reg = new_reg;
946 reg_eqv_table[lastr].next = new_reg;
947 reg_eqv_table[new_reg].prev = lastr;
951 /* Remove REG from its equivalence class. */
953 static void
954 delete_reg_equiv (unsigned int reg)
956 struct qty_table_elem *ent;
957 int q = REG_QTY (reg);
958 int p, n;
960 /* If invalid, do nothing. */
961 if (! REGNO_QTY_VALID_P (reg))
962 return;
964 ent = &qty_table[q];
966 p = reg_eqv_table[reg].prev;
967 n = reg_eqv_table[reg].next;
969 if (n != -1)
970 reg_eqv_table[n].prev = p;
971 else
972 ent->last_reg = p;
973 if (p != -1)
974 reg_eqv_table[p].next = n;
975 else
976 ent->first_reg = n;
978 REG_QTY (reg) = -reg - 1;
981 /* Remove any invalid expressions from the hash table
982 that refer to any of the registers contained in expression X.
984 Make sure that newly inserted references to those registers
985 as subexpressions will be considered valid.
987 mention_regs is not called when a register itself
988 is being stored in the table.
990 Return 1 if we have done something that may have changed the hash code
991 of X. */
993 static int
994 mention_regs (rtx x)
996 enum rtx_code code;
997 int i, j;
998 const char *fmt;
999 int changed = 0;
1001 if (x == 0)
1002 return 0;
1004 code = GET_CODE (x);
1005 if (code == REG)
1007 unsigned int regno = REGNO (x);
1008 unsigned int endregno = END_REGNO (x);
1009 unsigned int i;
1011 for (i = regno; i < endregno; i++)
1013 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1014 remove_invalid_refs (i);
1016 REG_IN_TABLE (i) = REG_TICK (i);
1017 SUBREG_TICKED (i) = -1;
1020 return 0;
1023 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1024 pseudo if they don't use overlapping words. We handle only pseudos
1025 here for simplicity. */
1026 if (code == SUBREG && REG_P (SUBREG_REG (x))
1027 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1029 unsigned int i = REGNO (SUBREG_REG (x));
1031 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1033 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1034 the last store to this register really stored into this
1035 subreg, then remove the memory of this subreg.
1036 Otherwise, remove any memory of the entire register and
1037 all its subregs from the table. */
1038 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1039 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1040 remove_invalid_refs (i);
1041 else
1042 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1045 REG_IN_TABLE (i) = REG_TICK (i);
1046 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1047 return 0;
1050 /* If X is a comparison or a COMPARE and either operand is a register
1051 that does not have a quantity, give it one. This is so that a later
1052 call to record_jump_equiv won't cause X to be assigned a different
1053 hash code and not found in the table after that call.
1055 It is not necessary to do this here, since rehash_using_reg can
1056 fix up the table later, but doing this here eliminates the need to
1057 call that expensive function in the most common case where the only
1058 use of the register is in the comparison. */
1060 if (code == COMPARE || COMPARISON_P (x))
1062 if (REG_P (XEXP (x, 0))
1063 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1064 if (insert_regs (XEXP (x, 0), NULL, 0))
1066 rehash_using_reg (XEXP (x, 0));
1067 changed = 1;
1070 if (REG_P (XEXP (x, 1))
1071 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1072 if (insert_regs (XEXP (x, 1), NULL, 0))
1074 rehash_using_reg (XEXP (x, 1));
1075 changed = 1;
1079 fmt = GET_RTX_FORMAT (code);
1080 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1081 if (fmt[i] == 'e')
1082 changed |= mention_regs (XEXP (x, i));
1083 else if (fmt[i] == 'E')
1084 for (j = 0; j < XVECLEN (x, i); j++)
1085 changed |= mention_regs (XVECEXP (x, i, j));
1087 return changed;
1090 /* Update the register quantities for inserting X into the hash table
1091 with a value equivalent to CLASSP.
1092 (If the class does not contain a REG, it is irrelevant.)
1093 If MODIFIED is nonzero, X is a destination; it is being modified.
1094 Note that delete_reg_equiv should be called on a register
1095 before insert_regs is done on that register with MODIFIED != 0.
1097 Nonzero value means that elements of reg_qty have changed
1098 so X's hash code may be different. */
1100 static int
1101 insert_regs (rtx x, struct table_elt *classp, int modified)
1103 if (REG_P (x))
1105 unsigned int regno = REGNO (x);
1106 int qty_valid;
1108 /* If REGNO is in the equivalence table already but is of the
1109 wrong mode for that equivalence, don't do anything here. */
1111 qty_valid = REGNO_QTY_VALID_P (regno);
1112 if (qty_valid)
1114 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1116 if (ent->mode != GET_MODE (x))
1117 return 0;
1120 if (modified || ! qty_valid)
1122 if (classp)
1123 for (classp = classp->first_same_value;
1124 classp != 0;
1125 classp = classp->next_same_value)
1126 if (REG_P (classp->exp)
1127 && GET_MODE (classp->exp) == GET_MODE (x))
1129 unsigned c_regno = REGNO (classp->exp);
1131 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1133 /* Suppose that 5 is hard reg and 100 and 101 are
1134 pseudos. Consider
1136 (set (reg:si 100) (reg:si 5))
1137 (set (reg:si 5) (reg:si 100))
1138 (set (reg:di 101) (reg:di 5))
1140 We would now set REG_QTY (101) = REG_QTY (5), but the
1141 entry for 5 is in SImode. When we use this later in
1142 copy propagation, we get the register in wrong mode. */
1143 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1144 continue;
1146 make_regs_eqv (regno, c_regno);
1147 return 1;
1150 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1151 than REG_IN_TABLE to find out if there was only a single preceding
1152 invalidation - for the SUBREG - or another one, which would be
1153 for the full register. However, if we find here that REG_TICK
1154 indicates that the register is invalid, it means that it has
1155 been invalidated in a separate operation. The SUBREG might be used
1156 now (then this is a recursive call), or we might use the full REG
1157 now and a SUBREG of it later. So bump up REG_TICK so that
1158 mention_regs will do the right thing. */
1159 if (! modified
1160 && REG_IN_TABLE (regno) >= 0
1161 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1162 REG_TICK (regno)++;
1163 make_new_qty (regno, GET_MODE (x));
1164 return 1;
1167 return 0;
1170 /* If X is a SUBREG, we will likely be inserting the inner register in the
1171 table. If that register doesn't have an assigned quantity number at
1172 this point but does later, the insertion that we will be doing now will
1173 not be accessible because its hash code will have changed. So assign
1174 a quantity number now. */
1176 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1177 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1179 insert_regs (SUBREG_REG (x), NULL, 0);
1180 mention_regs (x);
1181 return 1;
1183 else
1184 return mention_regs (x);
1188 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1189 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1190 CST is equal to an anchor. */
1192 static bool
1193 compute_const_anchors (rtx cst,
1194 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1195 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1197 HOST_WIDE_INT n = INTVAL (cst);
1199 *lower_base = n & ~(targetm.const_anchor - 1);
1200 if (*lower_base == n)
1201 return false;
1203 *upper_base =
1204 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1205 *upper_offs = n - *upper_base;
1206 *lower_offs = n - *lower_base;
1207 return true;
1210 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1212 static void
1213 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1214 machine_mode mode)
1216 struct table_elt *elt;
1217 unsigned hash;
1218 rtx anchor_exp;
1219 rtx exp;
1221 anchor_exp = GEN_INT (anchor);
1222 hash = HASH (anchor_exp, mode);
1223 elt = lookup (anchor_exp, hash, mode);
1224 if (!elt)
1225 elt = insert (anchor_exp, NULL, hash, mode);
1227 exp = plus_constant (mode, reg, offs);
1228 /* REG has just been inserted and the hash codes recomputed. */
1229 mention_regs (exp);
1230 hash = HASH (exp, mode);
1232 /* Use the cost of the register rather than the whole expression. When
1233 looking up constant anchors we will further offset the corresponding
1234 expression therefore it does not make sense to prefer REGs over
1235 reg-immediate additions. Prefer instead the oldest expression. Also
1236 don't prefer pseudos over hard regs so that we derive constants in
1237 argument registers from other argument registers rather than from the
1238 original pseudo that was used to synthesize the constant. */
1239 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1242 /* The constant CST is equivalent to the register REG. Create
1243 equivalences between the two anchors of CST and the corresponding
1244 register-offset expressions using REG. */
1246 static void
1247 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1249 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1251 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1252 &upper_base, &upper_offs))
1253 return;
1255 /* Ignore anchors of value 0. Constants accessible from zero are
1256 simple. */
1257 if (lower_base != 0)
1258 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1260 if (upper_base != 0)
1261 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1264 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1265 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1266 valid expression. Return the cheapest and oldest of such expressions. In
1267 *OLD, return how old the resulting expression is compared to the other
1268 equivalent expressions. */
1270 static rtx
1271 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1272 unsigned *old)
1274 struct table_elt *elt;
1275 unsigned idx;
1276 struct table_elt *match_elt;
1277 rtx match;
1279 /* Find the cheapest and *oldest* expression to maximize the chance of
1280 reusing the same pseudo. */
1282 match_elt = NULL;
1283 match = NULL_RTX;
1284 for (elt = anchor_elt->first_same_value, idx = 0;
1285 elt;
1286 elt = elt->next_same_value, idx++)
1288 if (match_elt && CHEAPER (match_elt, elt))
1289 return match;
1291 if (REG_P (elt->exp)
1292 || (GET_CODE (elt->exp) == PLUS
1293 && REG_P (XEXP (elt->exp, 0))
1294 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1296 rtx x;
1298 /* Ignore expressions that are no longer valid. */
1299 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1300 continue;
1302 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1303 if (REG_P (x)
1304 || (GET_CODE (x) == PLUS
1305 && IN_RANGE (INTVAL (XEXP (x, 1)),
1306 -targetm.const_anchor,
1307 targetm.const_anchor - 1)))
1309 match = x;
1310 match_elt = elt;
1311 *old = idx;
1316 return match;
1319 /* Try to express the constant SRC_CONST using a register+offset expression
1320 derived from a constant anchor. Return it if successful or NULL_RTX,
1321 otherwise. */
1323 static rtx
1324 try_const_anchors (rtx src_const, machine_mode mode)
1326 struct table_elt *lower_elt, *upper_elt;
1327 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1328 rtx lower_anchor_rtx, upper_anchor_rtx;
1329 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1330 unsigned lower_old, upper_old;
1332 /* CONST_INT is used for CC modes, but we should leave those alone. */
1333 if (GET_MODE_CLASS (mode) == MODE_CC)
1334 return NULL_RTX;
1336 gcc_assert (SCALAR_INT_MODE_P (mode));
1337 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1338 &upper_base, &upper_offs))
1339 return NULL_RTX;
1341 lower_anchor_rtx = GEN_INT (lower_base);
1342 upper_anchor_rtx = GEN_INT (upper_base);
1343 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1344 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1346 if (lower_elt)
1347 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1348 if (upper_elt)
1349 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1351 if (!lower_exp)
1352 return upper_exp;
1353 if (!upper_exp)
1354 return lower_exp;
1356 /* Return the older expression. */
1357 return (upper_old > lower_old ? upper_exp : lower_exp);
1360 /* Look in or update the hash table. */
1362 /* Remove table element ELT from use in the table.
1363 HASH is its hash code, made using the HASH macro.
1364 It's an argument because often that is known in advance
1365 and we save much time not recomputing it. */
1367 static void
1368 remove_from_table (struct table_elt *elt, unsigned int hash)
1370 if (elt == 0)
1371 return;
1373 /* Mark this element as removed. See cse_insn. */
1374 elt->first_same_value = 0;
1376 /* Remove the table element from its equivalence class. */
1379 struct table_elt *prev = elt->prev_same_value;
1380 struct table_elt *next = elt->next_same_value;
1382 if (next)
1383 next->prev_same_value = prev;
1385 if (prev)
1386 prev->next_same_value = next;
1387 else
1389 struct table_elt *newfirst = next;
1390 while (next)
1392 next->first_same_value = newfirst;
1393 next = next->next_same_value;
1398 /* Remove the table element from its hash bucket. */
1401 struct table_elt *prev = elt->prev_same_hash;
1402 struct table_elt *next = elt->next_same_hash;
1404 if (next)
1405 next->prev_same_hash = prev;
1407 if (prev)
1408 prev->next_same_hash = next;
1409 else if (table[hash] == elt)
1410 table[hash] = next;
1411 else
1413 /* This entry is not in the proper hash bucket. This can happen
1414 when two classes were merged by `merge_equiv_classes'. Search
1415 for the hash bucket that it heads. This happens only very
1416 rarely, so the cost is acceptable. */
1417 for (hash = 0; hash < HASH_SIZE; hash++)
1418 if (table[hash] == elt)
1419 table[hash] = next;
1423 /* Remove the table element from its related-value circular chain. */
1425 if (elt->related_value != 0 && elt->related_value != elt)
1427 struct table_elt *p = elt->related_value;
1429 while (p->related_value != elt)
1430 p = p->related_value;
1431 p->related_value = elt->related_value;
1432 if (p->related_value == p)
1433 p->related_value = 0;
1436 /* Now add it to the free element chain. */
1437 elt->next_same_hash = free_element_chain;
1438 free_element_chain = elt;
1441 /* Same as above, but X is a pseudo-register. */
1443 static void
1444 remove_pseudo_from_table (rtx x, unsigned int hash)
1446 struct table_elt *elt;
1448 /* Because a pseudo-register can be referenced in more than one
1449 mode, we might have to remove more than one table entry. */
1450 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1451 remove_from_table (elt, hash);
1454 /* Look up X in the hash table and return its table element,
1455 or 0 if X is not in the table.
1457 MODE is the machine-mode of X, or if X is an integer constant
1458 with VOIDmode then MODE is the mode with which X will be used.
1460 Here we are satisfied to find an expression whose tree structure
1461 looks like X. */
1463 static struct table_elt *
1464 lookup (rtx x, unsigned int hash, machine_mode mode)
1466 struct table_elt *p;
1468 for (p = table[hash]; p; p = p->next_same_hash)
1469 if (mode == p->mode && ((x == p->exp && REG_P (x))
1470 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1471 return p;
1473 return 0;
1476 /* Like `lookup' but don't care whether the table element uses invalid regs.
1477 Also ignore discrepancies in the machine mode of a register. */
1479 static struct table_elt *
1480 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1482 struct table_elt *p;
1484 if (REG_P (x))
1486 unsigned int regno = REGNO (x);
1488 /* Don't check the machine mode when comparing registers;
1489 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1490 for (p = table[hash]; p; p = p->next_same_hash)
1491 if (REG_P (p->exp)
1492 && REGNO (p->exp) == regno)
1493 return p;
1495 else
1497 for (p = table[hash]; p; p = p->next_same_hash)
1498 if (mode == p->mode
1499 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1500 return p;
1503 return 0;
1506 /* Look for an expression equivalent to X and with code CODE.
1507 If one is found, return that expression. */
1509 static rtx
1510 lookup_as_function (rtx x, enum rtx_code code)
1512 struct table_elt *p
1513 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1515 if (p == 0)
1516 return 0;
1518 for (p = p->first_same_value; p; p = p->next_same_value)
1519 if (GET_CODE (p->exp) == code
1520 /* Make sure this is a valid entry in the table. */
1521 && exp_equiv_p (p->exp, p->exp, 1, false))
1522 return p->exp;
1524 return 0;
1527 /* Insert X in the hash table, assuming HASH is its hash code and
1528 CLASSP is an element of the class it should go in (or 0 if a new
1529 class should be made). COST is the code of X and reg_cost is the
1530 cost of registers in X. It is inserted at the proper position to
1531 keep the class in the order cheapest first.
1533 MODE is the machine-mode of X, or if X is an integer constant
1534 with VOIDmode then MODE is the mode with which X will be used.
1536 For elements of equal cheapness, the most recent one
1537 goes in front, except that the first element in the list
1538 remains first unless a cheaper element is added. The order of
1539 pseudo-registers does not matter, as canon_reg will be called to
1540 find the cheapest when a register is retrieved from the table.
1542 The in_memory field in the hash table element is set to 0.
1543 The caller must set it nonzero if appropriate.
1545 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1546 and if insert_regs returns a nonzero value
1547 you must then recompute its hash code before calling here.
1549 If necessary, update table showing constant values of quantities. */
1551 static struct table_elt *
1552 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1553 machine_mode mode, int cost, int reg_cost)
1555 struct table_elt *elt;
1557 /* If X is a register and we haven't made a quantity for it,
1558 something is wrong. */
1559 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1561 /* If X is a hard register, show it is being put in the table. */
1562 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1563 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1565 /* Put an element for X into the right hash bucket. */
1567 elt = free_element_chain;
1568 if (elt)
1569 free_element_chain = elt->next_same_hash;
1570 else
1571 elt = XNEW (struct table_elt);
1573 elt->exp = x;
1574 elt->canon_exp = NULL_RTX;
1575 elt->cost = cost;
1576 elt->regcost = reg_cost;
1577 elt->next_same_value = 0;
1578 elt->prev_same_value = 0;
1579 elt->next_same_hash = table[hash];
1580 elt->prev_same_hash = 0;
1581 elt->related_value = 0;
1582 elt->in_memory = 0;
1583 elt->mode = mode;
1584 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1586 if (table[hash])
1587 table[hash]->prev_same_hash = elt;
1588 table[hash] = elt;
1590 /* Put it into the proper value-class. */
1591 if (classp)
1593 classp = classp->first_same_value;
1594 if (CHEAPER (elt, classp))
1595 /* Insert at the head of the class. */
1597 struct table_elt *p;
1598 elt->next_same_value = classp;
1599 classp->prev_same_value = elt;
1600 elt->first_same_value = elt;
1602 for (p = classp; p; p = p->next_same_value)
1603 p->first_same_value = elt;
1605 else
1607 /* Insert not at head of the class. */
1608 /* Put it after the last element cheaper than X. */
1609 struct table_elt *p, *next;
1611 for (p = classp;
1612 (next = p->next_same_value) && CHEAPER (next, elt);
1613 p = next)
1616 /* Put it after P and before NEXT. */
1617 elt->next_same_value = next;
1618 if (next)
1619 next->prev_same_value = elt;
1621 elt->prev_same_value = p;
1622 p->next_same_value = elt;
1623 elt->first_same_value = classp;
1626 else
1627 elt->first_same_value = elt;
1629 /* If this is a constant being set equivalent to a register or a register
1630 being set equivalent to a constant, note the constant equivalence.
1632 If this is a constant, it cannot be equivalent to a different constant,
1633 and a constant is the only thing that can be cheaper than a register. So
1634 we know the register is the head of the class (before the constant was
1635 inserted).
1637 If this is a register that is not already known equivalent to a
1638 constant, we must check the entire class.
1640 If this is a register that is already known equivalent to an insn,
1641 update the qtys `const_insn' to show that `this_insn' is the latest
1642 insn making that quantity equivalent to the constant. */
1644 if (elt->is_const && classp && REG_P (classp->exp)
1645 && !REG_P (x))
1647 int exp_q = REG_QTY (REGNO (classp->exp));
1648 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1650 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1651 exp_ent->const_insn = this_insn;
1654 else if (REG_P (x)
1655 && classp
1656 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1657 && ! elt->is_const)
1659 struct table_elt *p;
1661 for (p = classp; p != 0; p = p->next_same_value)
1663 if (p->is_const && !REG_P (p->exp))
1665 int x_q = REG_QTY (REGNO (x));
1666 struct qty_table_elem *x_ent = &qty_table[x_q];
1668 x_ent->const_rtx
1669 = gen_lowpart (GET_MODE (x), p->exp);
1670 x_ent->const_insn = this_insn;
1671 break;
1676 else if (REG_P (x)
1677 && qty_table[REG_QTY (REGNO (x))].const_rtx
1678 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1679 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1681 /* If this is a constant with symbolic value,
1682 and it has a term with an explicit integer value,
1683 link it up with related expressions. */
1684 if (GET_CODE (x) == CONST)
1686 rtx subexp = get_related_value (x);
1687 unsigned subhash;
1688 struct table_elt *subelt, *subelt_prev;
1690 if (subexp != 0)
1692 /* Get the integer-free subexpression in the hash table. */
1693 subhash = SAFE_HASH (subexp, mode);
1694 subelt = lookup (subexp, subhash, mode);
1695 if (subelt == 0)
1696 subelt = insert (subexp, NULL, subhash, mode);
1697 /* Initialize SUBELT's circular chain if it has none. */
1698 if (subelt->related_value == 0)
1699 subelt->related_value = subelt;
1700 /* Find the element in the circular chain that precedes SUBELT. */
1701 subelt_prev = subelt;
1702 while (subelt_prev->related_value != subelt)
1703 subelt_prev = subelt_prev->related_value;
1704 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1705 This way the element that follows SUBELT is the oldest one. */
1706 elt->related_value = subelt_prev->related_value;
1707 subelt_prev->related_value = elt;
1711 return elt;
1714 /* Wrap insert_with_costs by passing the default costs. */
1716 static struct table_elt *
1717 insert (rtx x, struct table_elt *classp, unsigned int hash,
1718 machine_mode mode)
1720 return insert_with_costs (x, classp, hash, mode,
1721 COST (x, mode), approx_reg_cost (x));
1725 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1726 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1727 the two classes equivalent.
1729 CLASS1 will be the surviving class; CLASS2 should not be used after this
1730 call.
1732 Any invalid entries in CLASS2 will not be copied. */
1734 static void
1735 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1737 struct table_elt *elt, *next, *new_elt;
1739 /* Ensure we start with the head of the classes. */
1740 class1 = class1->first_same_value;
1741 class2 = class2->first_same_value;
1743 /* If they were already equal, forget it. */
1744 if (class1 == class2)
1745 return;
1747 for (elt = class2; elt; elt = next)
1749 unsigned int hash;
1750 rtx exp = elt->exp;
1751 machine_mode mode = elt->mode;
1753 next = elt->next_same_value;
1755 /* Remove old entry, make a new one in CLASS1's class.
1756 Don't do this for invalid entries as we cannot find their
1757 hash code (it also isn't necessary). */
1758 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1760 bool need_rehash = false;
1762 hash_arg_in_memory = 0;
1763 hash = HASH (exp, mode);
1765 if (REG_P (exp))
1767 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1768 delete_reg_equiv (REGNO (exp));
1771 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1772 remove_pseudo_from_table (exp, hash);
1773 else
1774 remove_from_table (elt, hash);
1776 if (insert_regs (exp, class1, 0) || need_rehash)
1778 rehash_using_reg (exp);
1779 hash = HASH (exp, mode);
1781 new_elt = insert (exp, class1, hash, mode);
1782 new_elt->in_memory = hash_arg_in_memory;
1783 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1784 new_elt->cost = MAX_COST;
1789 /* Flush the entire hash table. */
1791 static void
1792 flush_hash_table (void)
1794 int i;
1795 struct table_elt *p;
1797 for (i = 0; i < HASH_SIZE; i++)
1798 for (p = table[i]; p; p = table[i])
1800 /* Note that invalidate can remove elements
1801 after P in the current hash chain. */
1802 if (REG_P (p->exp))
1803 invalidate (p->exp, VOIDmode);
1804 else
1805 remove_from_table (p, i);
1809 /* Check whether an anti dependence exists between X and EXP. MODE and
1810 ADDR are as for canon_anti_dependence. */
1812 static bool
1813 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1815 subrtx_iterator::array_type array;
1816 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1818 const_rtx x = *iter;
1819 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1820 return true;
1822 return false;
1825 /* Remove from the hash table, or mark as invalid, all expressions whose
1826 values could be altered by storing in X. X is a register, a subreg, or
1827 a memory reference with nonvarying address (because, when a memory
1828 reference with a varying address is stored in, all memory references are
1829 removed by invalidate_memory so specific invalidation is superfluous).
1830 FULL_MODE, if not VOIDmode, indicates that this much should be
1831 invalidated instead of just the amount indicated by the mode of X. This
1832 is only used for bitfield stores into memory.
1834 A nonvarying address may be just a register or just a symbol reference,
1835 or it may be either of those plus a numeric offset. */
1837 static void
1838 invalidate (rtx x, machine_mode full_mode)
1840 int i;
1841 struct table_elt *p;
1842 rtx addr;
1844 switch (GET_CODE (x))
1846 case REG:
1848 /* If X is a register, dependencies on its contents are recorded
1849 through the qty number mechanism. Just change the qty number of
1850 the register, mark it as invalid for expressions that refer to it,
1851 and remove it itself. */
1852 unsigned int regno = REGNO (x);
1853 unsigned int hash = HASH (x, GET_MODE (x));
1855 /* Remove REGNO from any quantity list it might be on and indicate
1856 that its value might have changed. If it is a pseudo, remove its
1857 entry from the hash table.
1859 For a hard register, we do the first two actions above for any
1860 additional hard registers corresponding to X. Then, if any of these
1861 registers are in the table, we must remove any REG entries that
1862 overlap these registers. */
1864 delete_reg_equiv (regno);
1865 REG_TICK (regno)++;
1866 SUBREG_TICKED (regno) = -1;
1868 if (regno >= FIRST_PSEUDO_REGISTER)
1869 remove_pseudo_from_table (x, hash);
1870 else
1872 HOST_WIDE_INT in_table
1873 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1874 unsigned int endregno = END_REGNO (x);
1875 unsigned int tregno, tendregno, rn;
1876 struct table_elt *p, *next;
1878 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1880 for (rn = regno + 1; rn < endregno; rn++)
1882 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1883 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1884 delete_reg_equiv (rn);
1885 REG_TICK (rn)++;
1886 SUBREG_TICKED (rn) = -1;
1889 if (in_table)
1890 for (hash = 0; hash < HASH_SIZE; hash++)
1891 for (p = table[hash]; p; p = next)
1893 next = p->next_same_hash;
1895 if (!REG_P (p->exp)
1896 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1897 continue;
1899 tregno = REGNO (p->exp);
1900 tendregno = END_REGNO (p->exp);
1901 if (tendregno > regno && tregno < endregno)
1902 remove_from_table (p, hash);
1906 return;
1908 case SUBREG:
1909 invalidate (SUBREG_REG (x), VOIDmode);
1910 return;
1912 case PARALLEL:
1913 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1914 invalidate (XVECEXP (x, 0, i), VOIDmode);
1915 return;
1917 case EXPR_LIST:
1918 /* This is part of a disjoint return value; extract the location in
1919 question ignoring the offset. */
1920 invalidate (XEXP (x, 0), VOIDmode);
1921 return;
1923 case MEM:
1924 addr = canon_rtx (get_addr (XEXP (x, 0)));
1925 /* Calculate the canonical version of X here so that
1926 true_dependence doesn't generate new RTL for X on each call. */
1927 x = canon_rtx (x);
1929 /* Remove all hash table elements that refer to overlapping pieces of
1930 memory. */
1931 if (full_mode == VOIDmode)
1932 full_mode = GET_MODE (x);
1934 for (i = 0; i < HASH_SIZE; i++)
1936 struct table_elt *next;
1938 for (p = table[i]; p; p = next)
1940 next = p->next_same_hash;
1941 if (p->in_memory)
1943 /* Just canonicalize the expression once;
1944 otherwise each time we call invalidate
1945 true_dependence will canonicalize the
1946 expression again. */
1947 if (!p->canon_exp)
1948 p->canon_exp = canon_rtx (p->exp);
1949 if (check_dependence (p->canon_exp, x, full_mode, addr))
1950 remove_from_table (p, i);
1954 return;
1956 default:
1957 gcc_unreachable ();
1961 /* Invalidate DEST. Used when DEST is not going to be added
1962 into the hash table for some reason, e.g. do_not_record
1963 flagged on it. */
1965 static void
1966 invalidate_dest (rtx dest)
1968 if (REG_P (dest)
1969 || GET_CODE (dest) == SUBREG
1970 || MEM_P (dest))
1971 invalidate (dest, VOIDmode);
1972 else if (GET_CODE (dest) == STRICT_LOW_PART
1973 || GET_CODE (dest) == ZERO_EXTRACT)
1974 invalidate (XEXP (dest, 0), GET_MODE (dest));
1977 /* Remove all expressions that refer to register REGNO,
1978 since they are already invalid, and we are about to
1979 mark that register valid again and don't want the old
1980 expressions to reappear as valid. */
1982 static void
1983 remove_invalid_refs (unsigned int regno)
1985 unsigned int i;
1986 struct table_elt *p, *next;
1988 for (i = 0; i < HASH_SIZE; i++)
1989 for (p = table[i]; p; p = next)
1991 next = p->next_same_hash;
1992 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1993 remove_from_table (p, i);
1997 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1998 and mode MODE. */
1999 static void
2000 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2001 machine_mode mode)
2003 unsigned int i;
2004 struct table_elt *p, *next;
2005 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2007 for (i = 0; i < HASH_SIZE; i++)
2008 for (p = table[i]; p; p = next)
2010 rtx exp = p->exp;
2011 next = p->next_same_hash;
2013 if (!REG_P (exp)
2014 && (GET_CODE (exp) != SUBREG
2015 || !REG_P (SUBREG_REG (exp))
2016 || REGNO (SUBREG_REG (exp)) != regno
2017 || (((SUBREG_BYTE (exp)
2018 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2019 && SUBREG_BYTE (exp) <= end))
2020 && refers_to_regno_p (regno, p->exp))
2021 remove_from_table (p, i);
2025 /* Recompute the hash codes of any valid entries in the hash table that
2026 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2028 This is called when we make a jump equivalence. */
2030 static void
2031 rehash_using_reg (rtx x)
2033 unsigned int i;
2034 struct table_elt *p, *next;
2035 unsigned hash;
2037 if (GET_CODE (x) == SUBREG)
2038 x = SUBREG_REG (x);
2040 /* If X is not a register or if the register is known not to be in any
2041 valid entries in the table, we have no work to do. */
2043 if (!REG_P (x)
2044 || REG_IN_TABLE (REGNO (x)) < 0
2045 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2046 return;
2048 /* Scan all hash chains looking for valid entries that mention X.
2049 If we find one and it is in the wrong hash chain, move it. */
2051 for (i = 0; i < HASH_SIZE; i++)
2052 for (p = table[i]; p; p = next)
2054 next = p->next_same_hash;
2055 if (reg_mentioned_p (x, p->exp)
2056 && exp_equiv_p (p->exp, p->exp, 1, false)
2057 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2059 if (p->next_same_hash)
2060 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2062 if (p->prev_same_hash)
2063 p->prev_same_hash->next_same_hash = p->next_same_hash;
2064 else
2065 table[i] = p->next_same_hash;
2067 p->next_same_hash = table[hash];
2068 p->prev_same_hash = 0;
2069 if (table[hash])
2070 table[hash]->prev_same_hash = p;
2071 table[hash] = p;
2076 /* Remove from the hash table any expression that is a call-clobbered
2077 register. Also update their TICK values. */
2079 static void
2080 invalidate_for_call (void)
2082 unsigned int regno, endregno;
2083 unsigned int i;
2084 unsigned hash;
2085 struct table_elt *p, *next;
2086 int in_table = 0;
2087 hard_reg_set_iterator hrsi;
2089 /* Go through all the hard registers. For each that is clobbered in
2090 a CALL_INSN, remove the register from quantity chains and update
2091 reg_tick if defined. Also see if any of these registers is currently
2092 in the table. */
2093 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2095 delete_reg_equiv (regno);
2096 if (REG_TICK (regno) >= 0)
2098 REG_TICK (regno)++;
2099 SUBREG_TICKED (regno) = -1;
2101 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2104 /* In the case where we have no call-clobbered hard registers in the
2105 table, we are done. Otherwise, scan the table and remove any
2106 entry that overlaps a call-clobbered register. */
2108 if (in_table)
2109 for (hash = 0; hash < HASH_SIZE; hash++)
2110 for (p = table[hash]; p; p = next)
2112 next = p->next_same_hash;
2114 if (!REG_P (p->exp)
2115 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2116 continue;
2118 regno = REGNO (p->exp);
2119 endregno = END_REGNO (p->exp);
2121 for (i = regno; i < endregno; i++)
2122 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2124 remove_from_table (p, hash);
2125 break;
2130 /* Given an expression X of type CONST,
2131 and ELT which is its table entry (or 0 if it
2132 is not in the hash table),
2133 return an alternate expression for X as a register plus integer.
2134 If none can be found, return 0. */
2136 static rtx
2137 use_related_value (rtx x, struct table_elt *elt)
2139 struct table_elt *relt = 0;
2140 struct table_elt *p, *q;
2141 HOST_WIDE_INT offset;
2143 /* First, is there anything related known?
2144 If we have a table element, we can tell from that.
2145 Otherwise, must look it up. */
2147 if (elt != 0 && elt->related_value != 0)
2148 relt = elt;
2149 else if (elt == 0 && GET_CODE (x) == CONST)
2151 rtx subexp = get_related_value (x);
2152 if (subexp != 0)
2153 relt = lookup (subexp,
2154 SAFE_HASH (subexp, GET_MODE (subexp)),
2155 GET_MODE (subexp));
2158 if (relt == 0)
2159 return 0;
2161 /* Search all related table entries for one that has an
2162 equivalent register. */
2164 p = relt;
2165 while (1)
2167 /* This loop is strange in that it is executed in two different cases.
2168 The first is when X is already in the table. Then it is searching
2169 the RELATED_VALUE list of X's class (RELT). The second case is when
2170 X is not in the table. Then RELT points to a class for the related
2171 value.
2173 Ensure that, whatever case we are in, that we ignore classes that have
2174 the same value as X. */
2176 if (rtx_equal_p (x, p->exp))
2177 q = 0;
2178 else
2179 for (q = p->first_same_value; q; q = q->next_same_value)
2180 if (REG_P (q->exp))
2181 break;
2183 if (q)
2184 break;
2186 p = p->related_value;
2188 /* We went all the way around, so there is nothing to be found.
2189 Alternatively, perhaps RELT was in the table for some other reason
2190 and it has no related values recorded. */
2191 if (p == relt || p == 0)
2192 break;
2195 if (q == 0)
2196 return 0;
2198 offset = (get_integer_term (x) - get_integer_term (p->exp));
2199 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2200 return plus_constant (q->mode, q->exp, offset);
2204 /* Hash a string. Just add its bytes up. */
2205 static inline unsigned
2206 hash_rtx_string (const char *ps)
2208 unsigned hash = 0;
2209 const unsigned char *p = (const unsigned char *) ps;
2211 if (p)
2212 while (*p)
2213 hash += *p++;
2215 return hash;
2218 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2219 When the callback returns true, we continue with the new rtx. */
2221 unsigned
2222 hash_rtx_cb (const_rtx x, machine_mode mode,
2223 int *do_not_record_p, int *hash_arg_in_memory_p,
2224 bool have_reg_qty, hash_rtx_callback_function cb)
2226 int i, j;
2227 unsigned hash = 0;
2228 enum rtx_code code;
2229 const char *fmt;
2230 machine_mode newmode;
2231 rtx newx;
2233 /* Used to turn recursion into iteration. We can't rely on GCC's
2234 tail-recursion elimination since we need to keep accumulating values
2235 in HASH. */
2236 repeat:
2237 if (x == 0)
2238 return hash;
2240 /* Invoke the callback first. */
2241 if (cb != NULL
2242 && ((*cb) (x, mode, &newx, &newmode)))
2244 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2245 hash_arg_in_memory_p, have_reg_qty, cb);
2246 return hash;
2249 code = GET_CODE (x);
2250 switch (code)
2252 case REG:
2254 unsigned int regno = REGNO (x);
2256 if (do_not_record_p && !reload_completed)
2258 /* On some machines, we can't record any non-fixed hard register,
2259 because extending its life will cause reload problems. We
2260 consider ap, fp, sp, gp to be fixed for this purpose.
2262 We also consider CCmode registers to be fixed for this purpose;
2263 failure to do so leads to failure to simplify 0<100 type of
2264 conditionals.
2266 On all machines, we can't record any global registers.
2267 Nor should we record any register that is in a small
2268 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2269 bool record;
2271 if (regno >= FIRST_PSEUDO_REGISTER)
2272 record = true;
2273 else if (x == frame_pointer_rtx
2274 || x == hard_frame_pointer_rtx
2275 || x == arg_pointer_rtx
2276 || x == stack_pointer_rtx
2277 || x == pic_offset_table_rtx)
2278 record = true;
2279 else if (global_regs[regno])
2280 record = false;
2281 else if (fixed_regs[regno])
2282 record = true;
2283 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2284 record = true;
2285 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2286 record = false;
2287 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2288 record = false;
2289 else
2290 record = true;
2292 if (!record)
2294 *do_not_record_p = 1;
2295 return 0;
2299 hash += ((unsigned int) REG << 7);
2300 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2301 return hash;
2304 /* We handle SUBREG of a REG specially because the underlying
2305 reg changes its hash value with every value change; we don't
2306 want to have to forget unrelated subregs when one subreg changes. */
2307 case SUBREG:
2309 if (REG_P (SUBREG_REG (x)))
2311 hash += (((unsigned int) SUBREG << 7)
2312 + REGNO (SUBREG_REG (x))
2313 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2314 return hash;
2316 break;
2319 case CONST_INT:
2320 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2321 + (unsigned int) INTVAL (x));
2322 return hash;
2324 case CONST_WIDE_INT:
2325 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2326 hash += CONST_WIDE_INT_ELT (x, i);
2327 return hash;
2329 case CONST_DOUBLE:
2330 /* This is like the general case, except that it only counts
2331 the integers representing the constant. */
2332 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2333 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2334 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2335 + (unsigned int) CONST_DOUBLE_HIGH (x));
2336 else
2337 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2338 return hash;
2340 case CONST_FIXED:
2341 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2342 hash += fixed_hash (CONST_FIXED_VALUE (x));
2343 return hash;
2345 case CONST_VECTOR:
2347 int units;
2348 rtx elt;
2350 units = CONST_VECTOR_NUNITS (x);
2352 for (i = 0; i < units; ++i)
2354 elt = CONST_VECTOR_ELT (x, i);
2355 hash += hash_rtx_cb (elt, GET_MODE (elt),
2356 do_not_record_p, hash_arg_in_memory_p,
2357 have_reg_qty, cb);
2360 return hash;
2363 /* Assume there is only one rtx object for any given label. */
2364 case LABEL_REF:
2365 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2366 differences and differences between each stage's debugging dumps. */
2367 hash += (((unsigned int) LABEL_REF << 7)
2368 + CODE_LABEL_NUMBER (label_ref_label (x)));
2369 return hash;
2371 case SYMBOL_REF:
2373 /* Don't hash on the symbol's address to avoid bootstrap differences.
2374 Different hash values may cause expressions to be recorded in
2375 different orders and thus different registers to be used in the
2376 final assembler. This also avoids differences in the dump files
2377 between various stages. */
2378 unsigned int h = 0;
2379 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2381 while (*p)
2382 h += (h << 7) + *p++; /* ??? revisit */
2384 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2385 return hash;
2388 case MEM:
2389 /* We don't record if marked volatile or if BLKmode since we don't
2390 know the size of the move. */
2391 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2393 *do_not_record_p = 1;
2394 return 0;
2396 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2397 *hash_arg_in_memory_p = 1;
2399 /* Now that we have already found this special case,
2400 might as well speed it up as much as possible. */
2401 hash += (unsigned) MEM;
2402 x = XEXP (x, 0);
2403 goto repeat;
2405 case USE:
2406 /* A USE that mentions non-volatile memory needs special
2407 handling since the MEM may be BLKmode which normally
2408 prevents an entry from being made. Pure calls are
2409 marked by a USE which mentions BLKmode memory.
2410 See calls.c:emit_call_1. */
2411 if (MEM_P (XEXP (x, 0))
2412 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2414 hash += (unsigned) USE;
2415 x = XEXP (x, 0);
2417 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2418 *hash_arg_in_memory_p = 1;
2420 /* Now that we have already found this special case,
2421 might as well speed it up as much as possible. */
2422 hash += (unsigned) MEM;
2423 x = XEXP (x, 0);
2424 goto repeat;
2426 break;
2428 case PRE_DEC:
2429 case PRE_INC:
2430 case POST_DEC:
2431 case POST_INC:
2432 case PRE_MODIFY:
2433 case POST_MODIFY:
2434 case PC:
2435 case CC0:
2436 case CALL:
2437 case UNSPEC_VOLATILE:
2438 if (do_not_record_p) {
2439 *do_not_record_p = 1;
2440 return 0;
2442 else
2443 return hash;
2444 break;
2446 case ASM_OPERANDS:
2447 if (do_not_record_p && MEM_VOLATILE_P (x))
2449 *do_not_record_p = 1;
2450 return 0;
2452 else
2454 /* We don't want to take the filename and line into account. */
2455 hash += (unsigned) code + (unsigned) GET_MODE (x)
2456 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2457 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2458 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2460 if (ASM_OPERANDS_INPUT_LENGTH (x))
2462 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2464 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2465 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2466 do_not_record_p, hash_arg_in_memory_p,
2467 have_reg_qty, cb)
2468 + hash_rtx_string
2469 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2472 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2473 x = ASM_OPERANDS_INPUT (x, 0);
2474 mode = GET_MODE (x);
2475 goto repeat;
2478 return hash;
2480 break;
2482 default:
2483 break;
2486 i = GET_RTX_LENGTH (code) - 1;
2487 hash += (unsigned) code + (unsigned) GET_MODE (x);
2488 fmt = GET_RTX_FORMAT (code);
2489 for (; i >= 0; i--)
2491 switch (fmt[i])
2493 case 'e':
2494 /* If we are about to do the last recursive call
2495 needed at this level, change it into iteration.
2496 This function is called enough to be worth it. */
2497 if (i == 0)
2499 x = XEXP (x, i);
2500 goto repeat;
2503 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2504 hash_arg_in_memory_p,
2505 have_reg_qty, cb);
2506 break;
2508 case 'E':
2509 for (j = 0; j < XVECLEN (x, i); j++)
2510 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2511 hash_arg_in_memory_p,
2512 have_reg_qty, cb);
2513 break;
2515 case 's':
2516 hash += hash_rtx_string (XSTR (x, i));
2517 break;
2519 case 'i':
2520 hash += (unsigned int) XINT (x, i);
2521 break;
2523 case '0': case 't':
2524 /* Unused. */
2525 break;
2527 default:
2528 gcc_unreachable ();
2532 return hash;
2535 /* Hash an rtx. We are careful to make sure the value is never negative.
2536 Equivalent registers hash identically.
2537 MODE is used in hashing for CONST_INTs only;
2538 otherwise the mode of X is used.
2540 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2542 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2543 a MEM rtx which does not have the MEM_READONLY_P flag set.
2545 Note that cse_insn knows that the hash code of a MEM expression
2546 is just (int) MEM plus the hash code of the address. */
2548 unsigned
2549 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2550 int *hash_arg_in_memory_p, bool have_reg_qty)
2552 return hash_rtx_cb (x, mode, do_not_record_p,
2553 hash_arg_in_memory_p, have_reg_qty, NULL);
2556 /* Hash an rtx X for cse via hash_rtx.
2557 Stores 1 in do_not_record if any subexpression is volatile.
2558 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2559 does not have the MEM_READONLY_P flag set. */
2561 static inline unsigned
2562 canon_hash (rtx x, machine_mode mode)
2564 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2567 /* Like canon_hash but with no side effects, i.e. do_not_record
2568 and hash_arg_in_memory are not changed. */
2570 static inline unsigned
2571 safe_hash (rtx x, machine_mode mode)
2573 int dummy_do_not_record;
2574 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2577 /* Return 1 iff X and Y would canonicalize into the same thing,
2578 without actually constructing the canonicalization of either one.
2579 If VALIDATE is nonzero,
2580 we assume X is an expression being processed from the rtl
2581 and Y was found in the hash table. We check register refs
2582 in Y for being marked as valid.
2584 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2587 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2589 int i, j;
2590 enum rtx_code code;
2591 const char *fmt;
2593 /* Note: it is incorrect to assume an expression is equivalent to itself
2594 if VALIDATE is nonzero. */
2595 if (x == y && !validate)
2596 return 1;
2598 if (x == 0 || y == 0)
2599 return x == y;
2601 code = GET_CODE (x);
2602 if (code != GET_CODE (y))
2603 return 0;
2605 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2606 if (GET_MODE (x) != GET_MODE (y))
2607 return 0;
2609 /* MEMs referring to different address space are not equivalent. */
2610 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2611 return 0;
2613 switch (code)
2615 case PC:
2616 case CC0:
2617 CASE_CONST_UNIQUE:
2618 return x == y;
2620 case LABEL_REF:
2621 return label_ref_label (x) == label_ref_label (y);
2623 case SYMBOL_REF:
2624 return XSTR (x, 0) == XSTR (y, 0);
2626 case REG:
2627 if (for_gcse)
2628 return REGNO (x) == REGNO (y);
2629 else
2631 unsigned int regno = REGNO (y);
2632 unsigned int i;
2633 unsigned int endregno = END_REGNO (y);
2635 /* If the quantities are not the same, the expressions are not
2636 equivalent. If there are and we are not to validate, they
2637 are equivalent. Otherwise, ensure all regs are up-to-date. */
2639 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2640 return 0;
2642 if (! validate)
2643 return 1;
2645 for (i = regno; i < endregno; i++)
2646 if (REG_IN_TABLE (i) != REG_TICK (i))
2647 return 0;
2649 return 1;
2652 case MEM:
2653 if (for_gcse)
2655 /* A volatile mem should not be considered equivalent to any
2656 other. */
2657 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2658 return 0;
2660 /* Can't merge two expressions in different alias sets, since we
2661 can decide that the expression is transparent in a block when
2662 it isn't, due to it being set with the different alias set.
2664 Also, can't merge two expressions with different MEM_ATTRS.
2665 They could e.g. be two different entities allocated into the
2666 same space on the stack (see e.g. PR25130). In that case, the
2667 MEM addresses can be the same, even though the two MEMs are
2668 absolutely not equivalent.
2670 But because really all MEM attributes should be the same for
2671 equivalent MEMs, we just use the invariant that MEMs that have
2672 the same attributes share the same mem_attrs data structure. */
2673 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2674 return 0;
2676 /* If we are handling exceptions, we cannot consider two expressions
2677 with different trapping status as equivalent, because simple_mem
2678 might accept one and reject the other. */
2679 if (cfun->can_throw_non_call_exceptions
2680 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2681 return 0;
2683 break;
2685 /* For commutative operations, check both orders. */
2686 case PLUS:
2687 case MULT:
2688 case AND:
2689 case IOR:
2690 case XOR:
2691 case NE:
2692 case EQ:
2693 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2694 validate, for_gcse)
2695 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2696 validate, for_gcse))
2697 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2698 validate, for_gcse)
2699 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2700 validate, for_gcse)));
2702 case ASM_OPERANDS:
2703 /* We don't use the generic code below because we want to
2704 disregard filename and line numbers. */
2706 /* A volatile asm isn't equivalent to any other. */
2707 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2708 return 0;
2710 if (GET_MODE (x) != GET_MODE (y)
2711 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2712 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2713 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2714 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2715 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2716 return 0;
2718 if (ASM_OPERANDS_INPUT_LENGTH (x))
2720 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2721 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2722 ASM_OPERANDS_INPUT (y, i),
2723 validate, for_gcse)
2724 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2725 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2726 return 0;
2729 return 1;
2731 default:
2732 break;
2735 /* Compare the elements. If any pair of corresponding elements
2736 fail to match, return 0 for the whole thing. */
2738 fmt = GET_RTX_FORMAT (code);
2739 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2741 switch (fmt[i])
2743 case 'e':
2744 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2745 validate, for_gcse))
2746 return 0;
2747 break;
2749 case 'E':
2750 if (XVECLEN (x, i) != XVECLEN (y, i))
2751 return 0;
2752 for (j = 0; j < XVECLEN (x, i); j++)
2753 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2754 validate, for_gcse))
2755 return 0;
2756 break;
2758 case 's':
2759 if (strcmp (XSTR (x, i), XSTR (y, i)))
2760 return 0;
2761 break;
2763 case 'i':
2764 if (XINT (x, i) != XINT (y, i))
2765 return 0;
2766 break;
2768 case 'w':
2769 if (XWINT (x, i) != XWINT (y, i))
2770 return 0;
2771 break;
2773 case '0':
2774 case 't':
2775 break;
2777 default:
2778 gcc_unreachable ();
2782 return 1;
2785 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2786 the result if necessary. INSN is as for canon_reg. */
2788 static void
2789 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2791 if (*xloc)
2793 rtx new_rtx = canon_reg (*xloc, insn);
2795 /* If replacing pseudo with hard reg or vice versa, ensure the
2796 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2797 gcc_assert (insn && new_rtx);
2798 validate_change (insn, xloc, new_rtx, 1);
2802 /* Canonicalize an expression:
2803 replace each register reference inside it
2804 with the "oldest" equivalent register.
2806 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2807 after we make our substitution. The calls are made with IN_GROUP nonzero
2808 so apply_change_group must be called upon the outermost return from this
2809 function (unless INSN is zero). The result of apply_change_group can
2810 generally be discarded since the changes we are making are optional. */
2812 static rtx
2813 canon_reg (rtx x, rtx_insn *insn)
2815 int i;
2816 enum rtx_code code;
2817 const char *fmt;
2819 if (x == 0)
2820 return x;
2822 code = GET_CODE (x);
2823 switch (code)
2825 case PC:
2826 case CC0:
2827 case CONST:
2828 CASE_CONST_ANY:
2829 case SYMBOL_REF:
2830 case LABEL_REF:
2831 case ADDR_VEC:
2832 case ADDR_DIFF_VEC:
2833 return x;
2835 case REG:
2837 int first;
2838 int q;
2839 struct qty_table_elem *ent;
2841 /* Never replace a hard reg, because hard regs can appear
2842 in more than one machine mode, and we must preserve the mode
2843 of each occurrence. Also, some hard regs appear in
2844 MEMs that are shared and mustn't be altered. Don't try to
2845 replace any reg that maps to a reg of class NO_REGS. */
2846 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2847 || ! REGNO_QTY_VALID_P (REGNO (x)))
2848 return x;
2850 q = REG_QTY (REGNO (x));
2851 ent = &qty_table[q];
2852 first = ent->first_reg;
2853 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2854 : REGNO_REG_CLASS (first) == NO_REGS ? x
2855 : gen_rtx_REG (ent->mode, first));
2858 default:
2859 break;
2862 fmt = GET_RTX_FORMAT (code);
2863 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2865 int j;
2867 if (fmt[i] == 'e')
2868 validate_canon_reg (&XEXP (x, i), insn);
2869 else if (fmt[i] == 'E')
2870 for (j = 0; j < XVECLEN (x, i); j++)
2871 validate_canon_reg (&XVECEXP (x, i, j), insn);
2874 return x;
2877 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2878 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2879 what values are being compared.
2881 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2882 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2883 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2884 compared to produce cc0.
2886 The return value is the comparison operator and is either the code of
2887 A or the code corresponding to the inverse of the comparison. */
2889 static enum rtx_code
2890 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2891 machine_mode *pmode1, machine_mode *pmode2)
2893 rtx arg1, arg2;
2894 hash_set<rtx> *visited = NULL;
2895 /* Set nonzero when we find something of interest. */
2896 rtx x = NULL;
2898 arg1 = *parg1, arg2 = *parg2;
2900 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2902 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2904 int reverse_code = 0;
2905 struct table_elt *p = 0;
2907 /* Remember state from previous iteration. */
2908 if (x)
2910 if (!visited)
2911 visited = new hash_set<rtx>;
2912 visited->add (x);
2913 x = 0;
2916 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2917 On machines with CC0, this is the only case that can occur, since
2918 fold_rtx will return the COMPARE or item being compared with zero
2919 when given CC0. */
2921 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2922 x = arg1;
2924 /* If ARG1 is a comparison operator and CODE is testing for
2925 STORE_FLAG_VALUE, get the inner arguments. */
2927 else if (COMPARISON_P (arg1))
2929 #ifdef FLOAT_STORE_FLAG_VALUE
2930 REAL_VALUE_TYPE fsfv;
2931 #endif
2933 if (code == NE
2934 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2935 && code == LT && STORE_FLAG_VALUE == -1)
2936 #ifdef FLOAT_STORE_FLAG_VALUE
2937 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2938 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2939 REAL_VALUE_NEGATIVE (fsfv)))
2940 #endif
2942 x = arg1;
2943 else if (code == EQ
2944 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2945 && code == GE && STORE_FLAG_VALUE == -1)
2946 #ifdef FLOAT_STORE_FLAG_VALUE
2947 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2948 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2949 REAL_VALUE_NEGATIVE (fsfv)))
2950 #endif
2952 x = arg1, reverse_code = 1;
2955 /* ??? We could also check for
2957 (ne (and (eq (...) (const_int 1))) (const_int 0))
2959 and related forms, but let's wait until we see them occurring. */
2961 if (x == 0)
2962 /* Look up ARG1 in the hash table and see if it has an equivalence
2963 that lets us see what is being compared. */
2964 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2965 if (p)
2967 p = p->first_same_value;
2969 /* If what we compare is already known to be constant, that is as
2970 good as it gets.
2971 We need to break the loop in this case, because otherwise we
2972 can have an infinite loop when looking at a reg that is known
2973 to be a constant which is the same as a comparison of a reg
2974 against zero which appears later in the insn stream, which in
2975 turn is constant and the same as the comparison of the first reg
2976 against zero... */
2977 if (p->is_const)
2978 break;
2981 for (; p; p = p->next_same_value)
2983 machine_mode inner_mode = GET_MODE (p->exp);
2984 #ifdef FLOAT_STORE_FLAG_VALUE
2985 REAL_VALUE_TYPE fsfv;
2986 #endif
2988 /* If the entry isn't valid, skip it. */
2989 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2990 continue;
2992 /* If it's a comparison we've used before, skip it. */
2993 if (visited && visited->contains (p->exp))
2994 continue;
2996 if (GET_CODE (p->exp) == COMPARE
2997 /* Another possibility is that this machine has a compare insn
2998 that includes the comparison code. In that case, ARG1 would
2999 be equivalent to a comparison operation that would set ARG1 to
3000 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3001 ORIG_CODE is the actual comparison being done; if it is an EQ,
3002 we must reverse ORIG_CODE. On machine with a negative value
3003 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3004 || ((code == NE
3005 || (code == LT
3006 && val_signbit_known_set_p (inner_mode,
3007 STORE_FLAG_VALUE))
3008 #ifdef FLOAT_STORE_FLAG_VALUE
3009 || (code == LT
3010 && SCALAR_FLOAT_MODE_P (inner_mode)
3011 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3012 REAL_VALUE_NEGATIVE (fsfv)))
3013 #endif
3015 && COMPARISON_P (p->exp)))
3017 x = p->exp;
3018 break;
3020 else if ((code == EQ
3021 || (code == GE
3022 && val_signbit_known_set_p (inner_mode,
3023 STORE_FLAG_VALUE))
3024 #ifdef FLOAT_STORE_FLAG_VALUE
3025 || (code == GE
3026 && SCALAR_FLOAT_MODE_P (inner_mode)
3027 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3028 REAL_VALUE_NEGATIVE (fsfv)))
3029 #endif
3031 && COMPARISON_P (p->exp))
3033 reverse_code = 1;
3034 x = p->exp;
3035 break;
3038 /* If this non-trapping address, e.g. fp + constant, the
3039 equivalent is a better operand since it may let us predict
3040 the value of the comparison. */
3041 else if (!rtx_addr_can_trap_p (p->exp))
3043 arg1 = p->exp;
3044 continue;
3048 /* If we didn't find a useful equivalence for ARG1, we are done.
3049 Otherwise, set up for the next iteration. */
3050 if (x == 0)
3051 break;
3053 /* If we need to reverse the comparison, make sure that is
3054 possible -- we can't necessarily infer the value of GE from LT
3055 with floating-point operands. */
3056 if (reverse_code)
3058 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3059 if (reversed == UNKNOWN)
3060 break;
3061 else
3062 code = reversed;
3064 else if (COMPARISON_P (x))
3065 code = GET_CODE (x);
3066 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3069 /* Return our results. Return the modes from before fold_rtx
3070 because fold_rtx might produce const_int, and then it's too late. */
3071 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3072 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3074 if (visited)
3075 delete visited;
3076 return code;
3079 /* If X is a nontrivial arithmetic operation on an argument for which
3080 a constant value can be determined, return the result of operating
3081 on that value, as a constant. Otherwise, return X, possibly with
3082 one or more operands changed to a forward-propagated constant.
3084 If X is a register whose contents are known, we do NOT return
3085 those contents here; equiv_constant is called to perform that task.
3086 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3088 INSN is the insn that we may be modifying. If it is 0, make a copy
3089 of X before modifying it. */
3091 static rtx
3092 fold_rtx (rtx x, rtx_insn *insn)
3094 enum rtx_code code;
3095 machine_mode mode;
3096 const char *fmt;
3097 int i;
3098 rtx new_rtx = 0;
3099 int changed = 0;
3101 /* Operands of X. */
3102 /* Workaround -Wmaybe-uninitialized false positive during
3103 profiledbootstrap by initializing them. */
3104 rtx folded_arg0 = NULL_RTX;
3105 rtx folded_arg1 = NULL_RTX;
3107 /* Constant equivalents of first three operands of X;
3108 0 when no such equivalent is known. */
3109 rtx const_arg0;
3110 rtx const_arg1;
3111 rtx const_arg2;
3113 /* The mode of the first operand of X. We need this for sign and zero
3114 extends. */
3115 machine_mode mode_arg0;
3117 if (x == 0)
3118 return x;
3120 /* Try to perform some initial simplifications on X. */
3121 code = GET_CODE (x);
3122 switch (code)
3124 case MEM:
3125 case SUBREG:
3126 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3127 than it would in other contexts. Basically its mode does not
3128 signify the size of the object read. That information is carried
3129 by size operand. If we happen to have a MEM of the appropriate
3130 mode in our tables with a constant value we could simplify the
3131 extraction incorrectly if we allowed substitution of that value
3132 for the MEM. */
3133 case ZERO_EXTRACT:
3134 case SIGN_EXTRACT:
3135 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3136 return new_rtx;
3137 return x;
3139 case CONST:
3140 CASE_CONST_ANY:
3141 case SYMBOL_REF:
3142 case LABEL_REF:
3143 case REG:
3144 case PC:
3145 /* No use simplifying an EXPR_LIST
3146 since they are used only for lists of args
3147 in a function call's REG_EQUAL note. */
3148 case EXPR_LIST:
3149 return x;
3151 case CC0:
3152 return prev_insn_cc0;
3154 case ASM_OPERANDS:
3155 if (insn)
3157 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3158 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3159 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3161 return x;
3163 case CALL:
3164 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3165 return x;
3166 break;
3168 /* Anything else goes through the loop below. */
3169 default:
3170 break;
3173 mode = GET_MODE (x);
3174 const_arg0 = 0;
3175 const_arg1 = 0;
3176 const_arg2 = 0;
3177 mode_arg0 = VOIDmode;
3179 /* Try folding our operands.
3180 Then see which ones have constant values known. */
3182 fmt = GET_RTX_FORMAT (code);
3183 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3184 if (fmt[i] == 'e')
3186 rtx folded_arg = XEXP (x, i), const_arg;
3187 machine_mode mode_arg = GET_MODE (folded_arg);
3189 switch (GET_CODE (folded_arg))
3191 case MEM:
3192 case REG:
3193 case SUBREG:
3194 const_arg = equiv_constant (folded_arg);
3195 break;
3197 case CONST:
3198 CASE_CONST_ANY:
3199 case SYMBOL_REF:
3200 case LABEL_REF:
3201 const_arg = folded_arg;
3202 break;
3204 case CC0:
3205 /* The cc0-user and cc0-setter may be in different blocks if
3206 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3207 will have been cleared as we exited the block with the
3208 setter.
3210 While we could potentially track cc0 in this case, it just
3211 doesn't seem to be worth it given that cc0 targets are not
3212 terribly common or important these days and trapping math
3213 is rarely used. The combination of those two conditions
3214 necessary to trip this situation is exceedingly rare in the
3215 real world. */
3216 if (!prev_insn_cc0)
3218 const_arg = NULL_RTX;
3220 else
3222 folded_arg = prev_insn_cc0;
3223 mode_arg = prev_insn_cc0_mode;
3224 const_arg = equiv_constant (folded_arg);
3226 break;
3228 default:
3229 folded_arg = fold_rtx (folded_arg, insn);
3230 const_arg = equiv_constant (folded_arg);
3231 break;
3234 /* For the first three operands, see if the operand
3235 is constant or equivalent to a constant. */
3236 switch (i)
3238 case 0:
3239 folded_arg0 = folded_arg;
3240 const_arg0 = const_arg;
3241 mode_arg0 = mode_arg;
3242 break;
3243 case 1:
3244 folded_arg1 = folded_arg;
3245 const_arg1 = const_arg;
3246 break;
3247 case 2:
3248 const_arg2 = const_arg;
3249 break;
3252 /* Pick the least expensive of the argument and an equivalent constant
3253 argument. */
3254 if (const_arg != 0
3255 && const_arg != folded_arg
3256 && (COST_IN (const_arg, mode_arg, code, i)
3257 <= COST_IN (folded_arg, mode_arg, code, i))
3259 /* It's not safe to substitute the operand of a conversion
3260 operator with a constant, as the conversion's identity
3261 depends upon the mode of its operand. This optimization
3262 is handled by the call to simplify_unary_operation. */
3263 && (GET_RTX_CLASS (code) != RTX_UNARY
3264 || GET_MODE (const_arg) == mode_arg0
3265 || (code != ZERO_EXTEND
3266 && code != SIGN_EXTEND
3267 && code != TRUNCATE
3268 && code != FLOAT_TRUNCATE
3269 && code != FLOAT_EXTEND
3270 && code != FLOAT
3271 && code != FIX
3272 && code != UNSIGNED_FLOAT
3273 && code != UNSIGNED_FIX)))
3274 folded_arg = const_arg;
3276 if (folded_arg == XEXP (x, i))
3277 continue;
3279 if (insn == NULL_RTX && !changed)
3280 x = copy_rtx (x);
3281 changed = 1;
3282 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3285 if (changed)
3287 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3288 consistent with the order in X. */
3289 if (canonicalize_change_group (insn, x))
3291 std::swap (const_arg0, const_arg1);
3292 std::swap (folded_arg0, folded_arg1);
3295 apply_change_group ();
3298 /* If X is an arithmetic operation, see if we can simplify it. */
3300 switch (GET_RTX_CLASS (code))
3302 case RTX_UNARY:
3304 /* We can't simplify extension ops unless we know the
3305 original mode. */
3306 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3307 && mode_arg0 == VOIDmode)
3308 break;
3310 new_rtx = simplify_unary_operation (code, mode,
3311 const_arg0 ? const_arg0 : folded_arg0,
3312 mode_arg0);
3314 break;
3316 case RTX_COMPARE:
3317 case RTX_COMM_COMPARE:
3318 /* See what items are actually being compared and set FOLDED_ARG[01]
3319 to those values and CODE to the actual comparison code. If any are
3320 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3321 do anything if both operands are already known to be constant. */
3323 /* ??? Vector mode comparisons are not supported yet. */
3324 if (VECTOR_MODE_P (mode))
3325 break;
3327 if (const_arg0 == 0 || const_arg1 == 0)
3329 struct table_elt *p0, *p1;
3330 rtx true_rtx, false_rtx;
3331 machine_mode mode_arg1;
3333 if (SCALAR_FLOAT_MODE_P (mode))
3335 #ifdef FLOAT_STORE_FLAG_VALUE
3336 true_rtx = (const_double_from_real_value
3337 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3338 #else
3339 true_rtx = NULL_RTX;
3340 #endif
3341 false_rtx = CONST0_RTX (mode);
3343 else
3345 true_rtx = const_true_rtx;
3346 false_rtx = const0_rtx;
3349 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3350 &mode_arg0, &mode_arg1);
3352 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3353 what kinds of things are being compared, so we can't do
3354 anything with this comparison. */
3356 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3357 break;
3359 const_arg0 = equiv_constant (folded_arg0);
3360 const_arg1 = equiv_constant (folded_arg1);
3362 /* If we do not now have two constants being compared, see
3363 if we can nevertheless deduce some things about the
3364 comparison. */
3365 if (const_arg0 == 0 || const_arg1 == 0)
3367 if (const_arg1 != NULL)
3369 rtx cheapest_simplification;
3370 int cheapest_cost;
3371 rtx simp_result;
3372 struct table_elt *p;
3374 /* See if we can find an equivalent of folded_arg0
3375 that gets us a cheaper expression, possibly a
3376 constant through simplifications. */
3377 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3378 mode_arg0);
3380 if (p != NULL)
3382 cheapest_simplification = x;
3383 cheapest_cost = COST (x, mode);
3385 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3387 int cost;
3389 /* If the entry isn't valid, skip it. */
3390 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3391 continue;
3393 /* Try to simplify using this equivalence. */
3394 simp_result
3395 = simplify_relational_operation (code, mode,
3396 mode_arg0,
3397 p->exp,
3398 const_arg1);
3400 if (simp_result == NULL)
3401 continue;
3403 cost = COST (simp_result, mode);
3404 if (cost < cheapest_cost)
3406 cheapest_cost = cost;
3407 cheapest_simplification = simp_result;
3411 /* If we have a cheaper expression now, use that
3412 and try folding it further, from the top. */
3413 if (cheapest_simplification != x)
3414 return fold_rtx (copy_rtx (cheapest_simplification),
3415 insn);
3419 /* See if the two operands are the same. */
3421 if ((REG_P (folded_arg0)
3422 && REG_P (folded_arg1)
3423 && (REG_QTY (REGNO (folded_arg0))
3424 == REG_QTY (REGNO (folded_arg1))))
3425 || ((p0 = lookup (folded_arg0,
3426 SAFE_HASH (folded_arg0, mode_arg0),
3427 mode_arg0))
3428 && (p1 = lookup (folded_arg1,
3429 SAFE_HASH (folded_arg1, mode_arg0),
3430 mode_arg0))
3431 && p0->first_same_value == p1->first_same_value))
3432 folded_arg1 = folded_arg0;
3434 /* If FOLDED_ARG0 is a register, see if the comparison we are
3435 doing now is either the same as we did before or the reverse
3436 (we only check the reverse if not floating-point). */
3437 else if (REG_P (folded_arg0))
3439 int qty = REG_QTY (REGNO (folded_arg0));
3441 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3443 struct qty_table_elem *ent = &qty_table[qty];
3445 if ((comparison_dominates_p (ent->comparison_code, code)
3446 || (! FLOAT_MODE_P (mode_arg0)
3447 && comparison_dominates_p (ent->comparison_code,
3448 reverse_condition (code))))
3449 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3450 || (const_arg1
3451 && rtx_equal_p (ent->comparison_const,
3452 const_arg1))
3453 || (REG_P (folded_arg1)
3454 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3456 if (comparison_dominates_p (ent->comparison_code, code))
3458 if (true_rtx)
3459 return true_rtx;
3460 else
3461 break;
3463 else
3464 return false_rtx;
3471 /* If we are comparing against zero, see if the first operand is
3472 equivalent to an IOR with a constant. If so, we may be able to
3473 determine the result of this comparison. */
3474 if (const_arg1 == const0_rtx && !const_arg0)
3476 rtx y = lookup_as_function (folded_arg0, IOR);
3477 rtx inner_const;
3479 if (y != 0
3480 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3481 && CONST_INT_P (inner_const)
3482 && INTVAL (inner_const) != 0)
3483 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3487 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3488 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3489 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3490 op0, op1);
3492 break;
3494 case RTX_BIN_ARITH:
3495 case RTX_COMM_ARITH:
3496 switch (code)
3498 case PLUS:
3499 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3500 with that LABEL_REF as its second operand. If so, the result is
3501 the first operand of that MINUS. This handles switches with an
3502 ADDR_DIFF_VEC table. */
3503 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3505 rtx y
3506 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3507 : lookup_as_function (folded_arg0, MINUS);
3509 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3510 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3511 return XEXP (y, 0);
3513 /* Now try for a CONST of a MINUS like the above. */
3514 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3515 : lookup_as_function (folded_arg0, CONST))) != 0
3516 && GET_CODE (XEXP (y, 0)) == MINUS
3517 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3518 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3519 return XEXP (XEXP (y, 0), 0);
3522 /* Likewise if the operands are in the other order. */
3523 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3525 rtx y
3526 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3527 : lookup_as_function (folded_arg1, MINUS);
3529 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3530 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3531 return XEXP (y, 0);
3533 /* Now try for a CONST of a MINUS like the above. */
3534 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3535 : lookup_as_function (folded_arg1, CONST))) != 0
3536 && GET_CODE (XEXP (y, 0)) == MINUS
3537 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3538 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3539 return XEXP (XEXP (y, 0), 0);
3542 /* If second operand is a register equivalent to a negative
3543 CONST_INT, see if we can find a register equivalent to the
3544 positive constant. Make a MINUS if so. Don't do this for
3545 a non-negative constant since we might then alternate between
3546 choosing positive and negative constants. Having the positive
3547 constant previously-used is the more common case. Be sure
3548 the resulting constant is non-negative; if const_arg1 were
3549 the smallest negative number this would overflow: depending
3550 on the mode, this would either just be the same value (and
3551 hence not save anything) or be incorrect. */
3552 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3553 && INTVAL (const_arg1) < 0
3554 /* This used to test
3556 -INTVAL (const_arg1) >= 0
3558 But The Sun V5.0 compilers mis-compiled that test. So
3559 instead we test for the problematic value in a more direct
3560 manner and hope the Sun compilers get it correct. */
3561 && INTVAL (const_arg1) !=
3562 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3563 && REG_P (folded_arg1))
3565 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3566 struct table_elt *p
3567 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3569 if (p)
3570 for (p = p->first_same_value; p; p = p->next_same_value)
3571 if (REG_P (p->exp))
3572 return simplify_gen_binary (MINUS, mode, folded_arg0,
3573 canon_reg (p->exp, NULL));
3575 goto from_plus;
3577 case MINUS:
3578 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3579 If so, produce (PLUS Z C2-C). */
3580 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3582 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3583 if (y && CONST_INT_P (XEXP (y, 1)))
3584 return fold_rtx (plus_constant (mode, copy_rtx (y),
3585 -INTVAL (const_arg1)),
3586 NULL);
3589 /* Fall through. */
3591 from_plus:
3592 case SMIN: case SMAX: case UMIN: case UMAX:
3593 case IOR: case AND: case XOR:
3594 case MULT:
3595 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3596 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3597 is known to be of similar form, we may be able to replace the
3598 operation with a combined operation. This may eliminate the
3599 intermediate operation if every use is simplified in this way.
3600 Note that the similar optimization done by combine.c only works
3601 if the intermediate operation's result has only one reference. */
3603 if (REG_P (folded_arg0)
3604 && const_arg1 && CONST_INT_P (const_arg1))
3606 int is_shift
3607 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3608 rtx y, inner_const, new_const;
3609 rtx canon_const_arg1 = const_arg1;
3610 enum rtx_code associate_code;
3612 if (is_shift
3613 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3614 || INTVAL (const_arg1) < 0))
3616 if (SHIFT_COUNT_TRUNCATED)
3617 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3618 & (GET_MODE_BITSIZE (mode)
3619 - 1));
3620 else
3621 break;
3624 y = lookup_as_function (folded_arg0, code);
3625 if (y == 0)
3626 break;
3628 /* If we have compiled a statement like
3629 "if (x == (x & mask1))", and now are looking at
3630 "x & mask2", we will have a case where the first operand
3631 of Y is the same as our first operand. Unless we detect
3632 this case, an infinite loop will result. */
3633 if (XEXP (y, 0) == folded_arg0)
3634 break;
3636 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3637 if (!inner_const || !CONST_INT_P (inner_const))
3638 break;
3640 /* Don't associate these operations if they are a PLUS with the
3641 same constant and it is a power of two. These might be doable
3642 with a pre- or post-increment. Similarly for two subtracts of
3643 identical powers of two with post decrement. */
3645 if (code == PLUS && const_arg1 == inner_const
3646 && ((HAVE_PRE_INCREMENT
3647 && pow2p_hwi (INTVAL (const_arg1)))
3648 || (HAVE_POST_INCREMENT
3649 && pow2p_hwi (INTVAL (const_arg1)))
3650 || (HAVE_PRE_DECREMENT
3651 && pow2p_hwi (- INTVAL (const_arg1)))
3652 || (HAVE_POST_DECREMENT
3653 && pow2p_hwi (- INTVAL (const_arg1)))))
3654 break;
3656 /* ??? Vector mode shifts by scalar
3657 shift operand are not supported yet. */
3658 if (is_shift && VECTOR_MODE_P (mode))
3659 break;
3661 if (is_shift
3662 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3663 || INTVAL (inner_const) < 0))
3665 if (SHIFT_COUNT_TRUNCATED)
3666 inner_const = GEN_INT (INTVAL (inner_const)
3667 & (GET_MODE_BITSIZE (mode) - 1));
3668 else
3669 break;
3672 /* Compute the code used to compose the constants. For example,
3673 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3675 associate_code = (is_shift || code == MINUS ? PLUS : code);
3677 new_const = simplify_binary_operation (associate_code, mode,
3678 canon_const_arg1,
3679 inner_const);
3681 if (new_const == 0)
3682 break;
3684 /* If we are associating shift operations, don't let this
3685 produce a shift of the size of the object or larger.
3686 This could occur when we follow a sign-extend by a right
3687 shift on a machine that does a sign-extend as a pair
3688 of shifts. */
3690 if (is_shift
3691 && CONST_INT_P (new_const)
3692 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3694 /* As an exception, we can turn an ASHIFTRT of this
3695 form into a shift of the number of bits - 1. */
3696 if (code == ASHIFTRT)
3697 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3698 else if (!side_effects_p (XEXP (y, 0)))
3699 return CONST0_RTX (mode);
3700 else
3701 break;
3704 y = copy_rtx (XEXP (y, 0));
3706 /* If Y contains our first operand (the most common way this
3707 can happen is if Y is a MEM), we would do into an infinite
3708 loop if we tried to fold it. So don't in that case. */
3710 if (! reg_mentioned_p (folded_arg0, y))
3711 y = fold_rtx (y, insn);
3713 return simplify_gen_binary (code, mode, y, new_const);
3715 break;
3717 case DIV: case UDIV:
3718 /* ??? The associative optimization performed immediately above is
3719 also possible for DIV and UDIV using associate_code of MULT.
3720 However, we would need extra code to verify that the
3721 multiplication does not overflow, that is, there is no overflow
3722 in the calculation of new_const. */
3723 break;
3725 default:
3726 break;
3729 new_rtx = simplify_binary_operation (code, mode,
3730 const_arg0 ? const_arg0 : folded_arg0,
3731 const_arg1 ? const_arg1 : folded_arg1);
3732 break;
3734 case RTX_OBJ:
3735 /* (lo_sum (high X) X) is simply X. */
3736 if (code == LO_SUM && const_arg0 != 0
3737 && GET_CODE (const_arg0) == HIGH
3738 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3739 return const_arg1;
3740 break;
3742 case RTX_TERNARY:
3743 case RTX_BITFIELD_OPS:
3744 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3745 const_arg0 ? const_arg0 : folded_arg0,
3746 const_arg1 ? const_arg1 : folded_arg1,
3747 const_arg2 ? const_arg2 : XEXP (x, 2));
3748 break;
3750 default:
3751 break;
3754 return new_rtx ? new_rtx : x;
3757 /* Return a constant value currently equivalent to X.
3758 Return 0 if we don't know one. */
3760 static rtx
3761 equiv_constant (rtx x)
3763 if (REG_P (x)
3764 && REGNO_QTY_VALID_P (REGNO (x)))
3766 int x_q = REG_QTY (REGNO (x));
3767 struct qty_table_elem *x_ent = &qty_table[x_q];
3769 if (x_ent->const_rtx)
3770 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3773 if (x == 0 || CONSTANT_P (x))
3774 return x;
3776 if (GET_CODE (x) == SUBREG)
3778 machine_mode mode = GET_MODE (x);
3779 machine_mode imode = GET_MODE (SUBREG_REG (x));
3780 rtx new_rtx;
3782 /* See if we previously assigned a constant value to this SUBREG. */
3783 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3784 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3785 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3786 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3787 return new_rtx;
3789 /* If we didn't and if doing so makes sense, see if we previously
3790 assigned a constant value to the enclosing word mode SUBREG. */
3791 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3792 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3794 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3795 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3797 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3798 new_rtx = lookup_as_function (y, CONST_INT);
3799 if (new_rtx)
3800 return gen_lowpart (mode, new_rtx);
3804 /* Otherwise see if we already have a constant for the inner REG,
3805 and if that is enough to calculate an equivalent constant for
3806 the subreg. Note that the upper bits of paradoxical subregs
3807 are undefined, so they cannot be said to equal anything. */
3808 if (REG_P (SUBREG_REG (x))
3809 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3810 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3811 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3813 return 0;
3816 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3817 the hash table in case its value was seen before. */
3819 if (MEM_P (x))
3821 struct table_elt *elt;
3823 x = avoid_constant_pool_reference (x);
3824 if (CONSTANT_P (x))
3825 return x;
3827 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3828 if (elt == 0)
3829 return 0;
3831 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3832 if (elt->is_const && CONSTANT_P (elt->exp))
3833 return elt->exp;
3836 return 0;
3839 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3840 "taken" branch.
3842 In certain cases, this can cause us to add an equivalence. For example,
3843 if we are following the taken case of
3844 if (i == 2)
3845 we can add the fact that `i' and '2' are now equivalent.
3847 In any case, we can record that this comparison was passed. If the same
3848 comparison is seen later, we will know its value. */
3850 static void
3851 record_jump_equiv (rtx_insn *insn, bool taken)
3853 int cond_known_true;
3854 rtx op0, op1;
3855 rtx set;
3856 machine_mode mode, mode0, mode1;
3857 int reversed_nonequality = 0;
3858 enum rtx_code code;
3860 /* Ensure this is the right kind of insn. */
3861 gcc_assert (any_condjump_p (insn));
3863 set = pc_set (insn);
3865 /* See if this jump condition is known true or false. */
3866 if (taken)
3867 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3868 else
3869 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3871 /* Get the type of comparison being done and the operands being compared.
3872 If we had to reverse a non-equality condition, record that fact so we
3873 know that it isn't valid for floating-point. */
3874 code = GET_CODE (XEXP (SET_SRC (set), 0));
3875 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3876 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3878 /* On a cc0 target the cc0-setter and cc0-user may end up in different
3879 blocks. When that happens the tracking of the cc0-setter via
3880 PREV_INSN_CC0 is spoiled. That means that fold_rtx may return
3881 NULL_RTX. In those cases, there's nothing to record. */
3882 if (op0 == NULL_RTX || op1 == NULL_RTX)
3883 return;
3885 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3886 if (! cond_known_true)
3888 code = reversed_comparison_code_parts (code, op0, op1, insn);
3890 /* Don't remember if we can't find the inverse. */
3891 if (code == UNKNOWN)
3892 return;
3895 /* The mode is the mode of the non-constant. */
3896 mode = mode0;
3897 if (mode1 != VOIDmode)
3898 mode = mode1;
3900 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3903 /* Yet another form of subreg creation. In this case, we want something in
3904 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3906 static rtx
3907 record_jump_cond_subreg (machine_mode mode, rtx op)
3909 machine_mode op_mode = GET_MODE (op);
3910 if (op_mode == mode || op_mode == VOIDmode)
3911 return op;
3912 return lowpart_subreg (mode, op, op_mode);
3915 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3916 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3917 Make any useful entries we can with that information. Called from
3918 above function and called recursively. */
3920 static void
3921 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3922 rtx op1, int reversed_nonequality)
3924 unsigned op0_hash, op1_hash;
3925 int op0_in_memory, op1_in_memory;
3926 struct table_elt *op0_elt, *op1_elt;
3928 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3929 we know that they are also equal in the smaller mode (this is also
3930 true for all smaller modes whether or not there is a SUBREG, but
3931 is not worth testing for with no SUBREG). */
3933 /* Note that GET_MODE (op0) may not equal MODE. */
3934 if (code == EQ && paradoxical_subreg_p (op0))
3936 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3937 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3938 if (tem)
3939 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3940 reversed_nonequality);
3943 if (code == EQ && paradoxical_subreg_p (op1))
3945 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3946 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3947 if (tem)
3948 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3949 reversed_nonequality);
3952 /* Similarly, if this is an NE comparison, and either is a SUBREG
3953 making a smaller mode, we know the whole thing is also NE. */
3955 /* Note that GET_MODE (op0) may not equal MODE;
3956 if we test MODE instead, we can get an infinite recursion
3957 alternating between two modes each wider than MODE. */
3959 if (code == NE && GET_CODE (op0) == SUBREG
3960 && subreg_lowpart_p (op0)
3961 && (GET_MODE_SIZE (GET_MODE (op0))
3962 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3964 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3965 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3966 if (tem)
3967 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3968 reversed_nonequality);
3971 if (code == NE && GET_CODE (op1) == SUBREG
3972 && subreg_lowpart_p (op1)
3973 && (GET_MODE_SIZE (GET_MODE (op1))
3974 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3976 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3977 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3978 if (tem)
3979 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3980 reversed_nonequality);
3983 /* Hash both operands. */
3985 do_not_record = 0;
3986 hash_arg_in_memory = 0;
3987 op0_hash = HASH (op0, mode);
3988 op0_in_memory = hash_arg_in_memory;
3990 if (do_not_record)
3991 return;
3993 do_not_record = 0;
3994 hash_arg_in_memory = 0;
3995 op1_hash = HASH (op1, mode);
3996 op1_in_memory = hash_arg_in_memory;
3998 if (do_not_record)
3999 return;
4001 /* Look up both operands. */
4002 op0_elt = lookup (op0, op0_hash, mode);
4003 op1_elt = lookup (op1, op1_hash, mode);
4005 /* If both operands are already equivalent or if they are not in the
4006 table but are identical, do nothing. */
4007 if ((op0_elt != 0 && op1_elt != 0
4008 && op0_elt->first_same_value == op1_elt->first_same_value)
4009 || op0 == op1 || rtx_equal_p (op0, op1))
4010 return;
4012 /* If we aren't setting two things equal all we can do is save this
4013 comparison. Similarly if this is floating-point. In the latter
4014 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4015 If we record the equality, we might inadvertently delete code
4016 whose intent was to change -0 to +0. */
4018 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4020 struct qty_table_elem *ent;
4021 int qty;
4023 /* If we reversed a floating-point comparison, if OP0 is not a
4024 register, or if OP1 is neither a register or constant, we can't
4025 do anything. */
4027 if (!REG_P (op1))
4028 op1 = equiv_constant (op1);
4030 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4031 || !REG_P (op0) || op1 == 0)
4032 return;
4034 /* Put OP0 in the hash table if it isn't already. This gives it a
4035 new quantity number. */
4036 if (op0_elt == 0)
4038 if (insert_regs (op0, NULL, 0))
4040 rehash_using_reg (op0);
4041 op0_hash = HASH (op0, mode);
4043 /* If OP0 is contained in OP1, this changes its hash code
4044 as well. Faster to rehash than to check, except
4045 for the simple case of a constant. */
4046 if (! CONSTANT_P (op1))
4047 op1_hash = HASH (op1,mode);
4050 op0_elt = insert (op0, NULL, op0_hash, mode);
4051 op0_elt->in_memory = op0_in_memory;
4054 qty = REG_QTY (REGNO (op0));
4055 ent = &qty_table[qty];
4057 ent->comparison_code = code;
4058 if (REG_P (op1))
4060 /* Look it up again--in case op0 and op1 are the same. */
4061 op1_elt = lookup (op1, op1_hash, mode);
4063 /* Put OP1 in the hash table so it gets a new quantity number. */
4064 if (op1_elt == 0)
4066 if (insert_regs (op1, NULL, 0))
4068 rehash_using_reg (op1);
4069 op1_hash = HASH (op1, mode);
4072 op1_elt = insert (op1, NULL, op1_hash, mode);
4073 op1_elt->in_memory = op1_in_memory;
4076 ent->comparison_const = NULL_RTX;
4077 ent->comparison_qty = REG_QTY (REGNO (op1));
4079 else
4081 ent->comparison_const = op1;
4082 ent->comparison_qty = -1;
4085 return;
4088 /* If either side is still missing an equivalence, make it now,
4089 then merge the equivalences. */
4091 if (op0_elt == 0)
4093 if (insert_regs (op0, NULL, 0))
4095 rehash_using_reg (op0);
4096 op0_hash = HASH (op0, mode);
4099 op0_elt = insert (op0, NULL, op0_hash, mode);
4100 op0_elt->in_memory = op0_in_memory;
4103 if (op1_elt == 0)
4105 if (insert_regs (op1, NULL, 0))
4107 rehash_using_reg (op1);
4108 op1_hash = HASH (op1, mode);
4111 op1_elt = insert (op1, NULL, op1_hash, mode);
4112 op1_elt->in_memory = op1_in_memory;
4115 merge_equiv_classes (op0_elt, op1_elt);
4118 /* CSE processing for one instruction.
4120 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4121 but the few that "leak through" are cleaned up by cse_insn, and complex
4122 addressing modes are often formed here.
4124 The main function is cse_insn, and between here and that function
4125 a couple of helper functions is defined to keep the size of cse_insn
4126 within reasonable proportions.
4128 Data is shared between the main and helper functions via STRUCT SET,
4129 that contains all data related for every set in the instruction that
4130 is being processed.
4132 Note that cse_main processes all sets in the instruction. Most
4133 passes in GCC only process simple SET insns or single_set insns, but
4134 CSE processes insns with multiple sets as well. */
4136 /* Data on one SET contained in the instruction. */
4138 struct set
4140 /* The SET rtx itself. */
4141 rtx rtl;
4142 /* The SET_SRC of the rtx (the original value, if it is changing). */
4143 rtx src;
4144 /* The hash-table element for the SET_SRC of the SET. */
4145 struct table_elt *src_elt;
4146 /* Hash value for the SET_SRC. */
4147 unsigned src_hash;
4148 /* Hash value for the SET_DEST. */
4149 unsigned dest_hash;
4150 /* The SET_DEST, with SUBREG, etc., stripped. */
4151 rtx inner_dest;
4152 /* Nonzero if the SET_SRC is in memory. */
4153 char src_in_memory;
4154 /* Nonzero if the SET_SRC contains something
4155 whose value cannot be predicted and understood. */
4156 char src_volatile;
4157 /* Original machine mode, in case it becomes a CONST_INT.
4158 The size of this field should match the size of the mode
4159 field of struct rtx_def (see rtl.h). */
4160 ENUM_BITFIELD(machine_mode) mode : 8;
4161 /* A constant equivalent for SET_SRC, if any. */
4162 rtx src_const;
4163 /* Hash value of constant equivalent for SET_SRC. */
4164 unsigned src_const_hash;
4165 /* Table entry for constant equivalent for SET_SRC, if any. */
4166 struct table_elt *src_const_elt;
4167 /* Table entry for the destination address. */
4168 struct table_elt *dest_addr_elt;
4171 /* Special handling for (set REG0 REG1) where REG0 is the
4172 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4173 be used in the sequel, so (if easily done) change this insn to
4174 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4175 that computed their value. Then REG1 will become a dead store
4176 and won't cloud the situation for later optimizations.
4178 Do not make this change if REG1 is a hard register, because it will
4179 then be used in the sequel and we may be changing a two-operand insn
4180 into a three-operand insn.
4182 This is the last transformation that cse_insn will try to do. */
4184 static void
4185 try_back_substitute_reg (rtx set, rtx_insn *insn)
4187 rtx dest = SET_DEST (set);
4188 rtx src = SET_SRC (set);
4190 if (REG_P (dest)
4191 && REG_P (src) && ! HARD_REGISTER_P (src)
4192 && REGNO_QTY_VALID_P (REGNO (src)))
4194 int src_q = REG_QTY (REGNO (src));
4195 struct qty_table_elem *src_ent = &qty_table[src_q];
4197 if (src_ent->first_reg == REGNO (dest))
4199 /* Scan for the previous nonnote insn, but stop at a basic
4200 block boundary. */
4201 rtx_insn *prev = insn;
4202 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4205 prev = PREV_INSN (prev);
4207 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4209 /* Do not swap the registers around if the previous instruction
4210 attaches a REG_EQUIV note to REG1.
4212 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4213 from the pseudo that originally shadowed an incoming argument
4214 to another register. Some uses of REG_EQUIV might rely on it
4215 being attached to REG1 rather than REG2.
4217 This section previously turned the REG_EQUIV into a REG_EQUAL
4218 note. We cannot do that because REG_EQUIV may provide an
4219 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4220 if (NONJUMP_INSN_P (prev)
4221 && GET_CODE (PATTERN (prev)) == SET
4222 && SET_DEST (PATTERN (prev)) == src
4223 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4225 rtx note;
4227 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4228 validate_change (insn, &SET_DEST (set), src, 1);
4229 validate_change (insn, &SET_SRC (set), dest, 1);
4230 apply_change_group ();
4232 /* If INSN has a REG_EQUAL note, and this note mentions
4233 REG0, then we must delete it, because the value in
4234 REG0 has changed. If the note's value is REG1, we must
4235 also delete it because that is now this insn's dest. */
4236 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4237 if (note != 0
4238 && (reg_mentioned_p (dest, XEXP (note, 0))
4239 || rtx_equal_p (src, XEXP (note, 0))))
4240 remove_note (insn, note);
4246 /* Record all the SETs in this instruction into SETS_PTR,
4247 and return the number of recorded sets. */
4248 static int
4249 find_sets_in_insn (rtx_insn *insn, struct set **psets)
4251 struct set *sets = *psets;
4252 int n_sets = 0;
4253 rtx x = PATTERN (insn);
4255 if (GET_CODE (x) == SET)
4257 /* Ignore SETs that are unconditional jumps.
4258 They never need cse processing, so this does not hurt.
4259 The reason is not efficiency but rather
4260 so that we can test at the end for instructions
4261 that have been simplified to unconditional jumps
4262 and not be misled by unchanged instructions
4263 that were unconditional jumps to begin with. */
4264 if (SET_DEST (x) == pc_rtx
4265 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4267 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4268 The hard function value register is used only once, to copy to
4269 someplace else, so it isn't worth cse'ing. */
4270 else if (GET_CODE (SET_SRC (x)) == CALL)
4272 else
4273 sets[n_sets++].rtl = x;
4275 else if (GET_CODE (x) == PARALLEL)
4277 int i, lim = XVECLEN (x, 0);
4279 /* Go over the expressions of the PARALLEL in forward order, to
4280 put them in the same order in the SETS array. */
4281 for (i = 0; i < lim; i++)
4283 rtx y = XVECEXP (x, 0, i);
4284 if (GET_CODE (y) == SET)
4286 /* As above, we ignore unconditional jumps and call-insns and
4287 ignore the result of apply_change_group. */
4288 if (SET_DEST (y) == pc_rtx
4289 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4291 else if (GET_CODE (SET_SRC (y)) == CALL)
4293 else
4294 sets[n_sets++].rtl = y;
4299 return n_sets;
4302 /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */
4304 static void
4305 canon_asm_operands (rtx x, rtx_insn *insn)
4307 for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4309 rtx input = ASM_OPERANDS_INPUT (x, i);
4310 if (!(REG_P (input) && HARD_REGISTER_P (input)))
4312 input = canon_reg (input, insn);
4313 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4318 /* Where possible, substitute every register reference in the N_SETS
4319 number of SETS in INSN with the canonical register.
4321 Register canonicalization propagatest the earliest register (i.e.
4322 one that is set before INSN) with the same value. This is a very
4323 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4324 to RTL. For instance, a CONST for an address is usually expanded
4325 multiple times to loads into different registers, thus creating many
4326 subexpressions of the form:
4328 (set (reg1) (some_const))
4329 (set (mem (... reg1 ...) (thing)))
4330 (set (reg2) (some_const))
4331 (set (mem (... reg2 ...) (thing)))
4333 After canonicalizing, the code takes the following form:
4335 (set (reg1) (some_const))
4336 (set (mem (... reg1 ...) (thing)))
4337 (set (reg2) (some_const))
4338 (set (mem (... reg1 ...) (thing)))
4340 The set to reg2 is now trivially dead, and the memory reference (or
4341 address, or whatever) may be a candidate for further CSEing.
4343 In this function, the result of apply_change_group can be ignored;
4344 see canon_reg. */
4346 static void
4347 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4349 struct set *sets = *psets;
4350 rtx tem;
4351 rtx x = PATTERN (insn);
4352 int i;
4354 if (CALL_P (insn))
4356 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4357 if (GET_CODE (XEXP (tem, 0)) != SET)
4358 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4361 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4363 canon_reg (SET_SRC (x), insn);
4364 apply_change_group ();
4365 fold_rtx (SET_SRC (x), insn);
4367 else if (GET_CODE (x) == CLOBBER)
4369 /* If we clobber memory, canon the address.
4370 This does nothing when a register is clobbered
4371 because we have already invalidated the reg. */
4372 if (MEM_P (XEXP (x, 0)))
4373 canon_reg (XEXP (x, 0), insn);
4375 else if (GET_CODE (x) == USE
4376 && ! (REG_P (XEXP (x, 0))
4377 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4378 /* Canonicalize a USE of a pseudo register or memory location. */
4379 canon_reg (x, insn);
4380 else if (GET_CODE (x) == ASM_OPERANDS)
4381 canon_asm_operands (x, insn);
4382 else if (GET_CODE (x) == CALL)
4384 canon_reg (x, insn);
4385 apply_change_group ();
4386 fold_rtx (x, insn);
4388 else if (DEBUG_INSN_P (insn))
4389 canon_reg (PATTERN (insn), insn);
4390 else if (GET_CODE (x) == PARALLEL)
4392 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4394 rtx y = XVECEXP (x, 0, i);
4395 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4397 canon_reg (SET_SRC (y), insn);
4398 apply_change_group ();
4399 fold_rtx (SET_SRC (y), insn);
4401 else if (GET_CODE (y) == CLOBBER)
4403 if (MEM_P (XEXP (y, 0)))
4404 canon_reg (XEXP (y, 0), insn);
4406 else if (GET_CODE (y) == USE
4407 && ! (REG_P (XEXP (y, 0))
4408 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4409 canon_reg (y, insn);
4410 else if (GET_CODE (y) == ASM_OPERANDS)
4411 canon_asm_operands (y, insn);
4412 else if (GET_CODE (y) == CALL)
4414 canon_reg (y, insn);
4415 apply_change_group ();
4416 fold_rtx (y, insn);
4421 if (n_sets == 1 && REG_NOTES (insn) != 0
4422 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4424 /* We potentially will process this insn many times. Therefore,
4425 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4426 unique set in INSN.
4428 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4429 because cse_insn handles those specially. */
4430 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4431 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4432 remove_note (insn, tem);
4433 else
4435 canon_reg (XEXP (tem, 0), insn);
4436 apply_change_group ();
4437 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4438 df_notes_rescan (insn);
4442 /* Canonicalize sources and addresses of destinations.
4443 We do this in a separate pass to avoid problems when a MATCH_DUP is
4444 present in the insn pattern. In that case, we want to ensure that
4445 we don't break the duplicate nature of the pattern. So we will replace
4446 both operands at the same time. Otherwise, we would fail to find an
4447 equivalent substitution in the loop calling validate_change below.
4449 We used to suppress canonicalization of DEST if it appears in SRC,
4450 but we don't do this any more. */
4452 for (i = 0; i < n_sets; i++)
4454 rtx dest = SET_DEST (sets[i].rtl);
4455 rtx src = SET_SRC (sets[i].rtl);
4456 rtx new_rtx = canon_reg (src, insn);
4458 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4460 if (GET_CODE (dest) == ZERO_EXTRACT)
4462 validate_change (insn, &XEXP (dest, 1),
4463 canon_reg (XEXP (dest, 1), insn), 1);
4464 validate_change (insn, &XEXP (dest, 2),
4465 canon_reg (XEXP (dest, 2), insn), 1);
4468 while (GET_CODE (dest) == SUBREG
4469 || GET_CODE (dest) == ZERO_EXTRACT
4470 || GET_CODE (dest) == STRICT_LOW_PART)
4471 dest = XEXP (dest, 0);
4473 if (MEM_P (dest))
4474 canon_reg (dest, insn);
4477 /* Now that we have done all the replacements, we can apply the change
4478 group and see if they all work. Note that this will cause some
4479 canonicalizations that would have worked individually not to be applied
4480 because some other canonicalization didn't work, but this should not
4481 occur often.
4483 The result of apply_change_group can be ignored; see canon_reg. */
4485 apply_change_group ();
4488 /* Main function of CSE.
4489 First simplify sources and addresses of all assignments
4490 in the instruction, using previously-computed equivalents values.
4491 Then install the new sources and destinations in the table
4492 of available values. */
4494 static void
4495 cse_insn (rtx_insn *insn)
4497 rtx x = PATTERN (insn);
4498 int i;
4499 rtx tem;
4500 int n_sets = 0;
4502 rtx src_eqv = 0;
4503 struct table_elt *src_eqv_elt = 0;
4504 int src_eqv_volatile = 0;
4505 int src_eqv_in_memory = 0;
4506 unsigned src_eqv_hash = 0;
4508 struct set *sets = (struct set *) 0;
4510 if (GET_CODE (x) == SET)
4511 sets = XALLOCA (struct set);
4512 else if (GET_CODE (x) == PARALLEL)
4513 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4515 this_insn = insn;
4516 /* Records what this insn does to set CC0. */
4517 this_insn_cc0 = 0;
4518 this_insn_cc0_mode = VOIDmode;
4520 /* Find all regs explicitly clobbered in this insn,
4521 to ensure they are not replaced with any other regs
4522 elsewhere in this insn. */
4523 invalidate_from_sets_and_clobbers (insn);
4525 /* Record all the SETs in this instruction. */
4526 n_sets = find_sets_in_insn (insn, &sets);
4528 /* Substitute the canonical register where possible. */
4529 canonicalize_insn (insn, &sets, n_sets);
4531 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4532 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4533 latter condition is necessary because SRC_EQV is handled specially for
4534 this case, and if it isn't set, then there will be no equivalence
4535 for the destination. */
4536 if (n_sets == 1 && REG_NOTES (insn) != 0
4537 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4540 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4541 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4542 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4543 src_eqv = copy_rtx (XEXP (tem, 0));
4544 /* If DEST is of the form ZERO_EXTACT, as in:
4545 (set (zero_extract:SI (reg:SI 119)
4546 (const_int 16 [0x10])
4547 (const_int 16 [0x10]))
4548 (const_int 51154 [0xc7d2]))
4549 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4550 point. Note that this is different from SRC_EQV. We can however
4551 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4552 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4553 && CONST_INT_P (XEXP (tem, 0))
4554 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4555 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4557 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4558 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4559 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4560 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4561 HOST_WIDE_INT mask;
4562 unsigned int shift;
4563 if (BITS_BIG_ENDIAN)
4564 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
4565 - INTVAL (pos) - INTVAL (width);
4566 else
4567 shift = INTVAL (pos);
4568 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4569 mask = HOST_WIDE_INT_M1;
4570 else
4571 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4572 val = (val >> shift) & mask;
4573 src_eqv = GEN_INT (val);
4577 /* Set sets[i].src_elt to the class each source belongs to.
4578 Detect assignments from or to volatile things
4579 and set set[i] to zero so they will be ignored
4580 in the rest of this function.
4582 Nothing in this loop changes the hash table or the register chains. */
4584 for (i = 0; i < n_sets; i++)
4586 bool repeat = false;
4587 bool mem_noop_insn = false;
4588 rtx src, dest;
4589 rtx src_folded;
4590 struct table_elt *elt = 0, *p;
4591 machine_mode mode;
4592 rtx src_eqv_here;
4593 rtx src_const = 0;
4594 rtx src_related = 0;
4595 bool src_related_is_const_anchor = false;
4596 struct table_elt *src_const_elt = 0;
4597 int src_cost = MAX_COST;
4598 int src_eqv_cost = MAX_COST;
4599 int src_folded_cost = MAX_COST;
4600 int src_related_cost = MAX_COST;
4601 int src_elt_cost = MAX_COST;
4602 int src_regcost = MAX_COST;
4603 int src_eqv_regcost = MAX_COST;
4604 int src_folded_regcost = MAX_COST;
4605 int src_related_regcost = MAX_COST;
4606 int src_elt_regcost = MAX_COST;
4607 /* Set nonzero if we need to call force_const_mem on with the
4608 contents of src_folded before using it. */
4609 int src_folded_force_flag = 0;
4611 dest = SET_DEST (sets[i].rtl);
4612 src = SET_SRC (sets[i].rtl);
4614 /* If SRC is a constant that has no machine mode,
4615 hash it with the destination's machine mode.
4616 This way we can keep different modes separate. */
4618 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4619 sets[i].mode = mode;
4621 if (src_eqv)
4623 machine_mode eqvmode = mode;
4624 if (GET_CODE (dest) == STRICT_LOW_PART)
4625 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4626 do_not_record = 0;
4627 hash_arg_in_memory = 0;
4628 src_eqv_hash = HASH (src_eqv, eqvmode);
4630 /* Find the equivalence class for the equivalent expression. */
4632 if (!do_not_record)
4633 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4635 src_eqv_volatile = do_not_record;
4636 src_eqv_in_memory = hash_arg_in_memory;
4639 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4640 value of the INNER register, not the destination. So it is not
4641 a valid substitution for the source. But save it for later. */
4642 if (GET_CODE (dest) == STRICT_LOW_PART)
4643 src_eqv_here = 0;
4644 else
4645 src_eqv_here = src_eqv;
4647 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4648 simplified result, which may not necessarily be valid. */
4649 src_folded = fold_rtx (src, NULL);
4651 #if 0
4652 /* ??? This caused bad code to be generated for the m68k port with -O2.
4653 Suppose src is (CONST_INT -1), and that after truncation src_folded
4654 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4655 At the end we will add src and src_const to the same equivalence
4656 class. We now have 3 and -1 on the same equivalence class. This
4657 causes later instructions to be mis-optimized. */
4658 /* If storing a constant in a bitfield, pre-truncate the constant
4659 so we will be able to record it later. */
4660 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4662 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4664 if (CONST_INT_P (src)
4665 && CONST_INT_P (width)
4666 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4667 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4668 src_folded
4669 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4670 << INTVAL (width)) - 1));
4672 #endif
4674 /* Compute SRC's hash code, and also notice if it
4675 should not be recorded at all. In that case,
4676 prevent any further processing of this assignment. */
4677 do_not_record = 0;
4678 hash_arg_in_memory = 0;
4680 sets[i].src = src;
4681 sets[i].src_hash = HASH (src, mode);
4682 sets[i].src_volatile = do_not_record;
4683 sets[i].src_in_memory = hash_arg_in_memory;
4685 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4686 a pseudo, do not record SRC. Using SRC as a replacement for
4687 anything else will be incorrect in that situation. Note that
4688 this usually occurs only for stack slots, in which case all the
4689 RTL would be referring to SRC, so we don't lose any optimization
4690 opportunities by not having SRC in the hash table. */
4692 if (MEM_P (src)
4693 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4694 && REG_P (dest)
4695 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4696 sets[i].src_volatile = 1;
4698 else if (GET_CODE (src) == ASM_OPERANDS
4699 && GET_CODE (x) == PARALLEL)
4701 /* Do not record result of a non-volatile inline asm with
4702 more than one result. */
4703 if (n_sets > 1)
4704 sets[i].src_volatile = 1;
4706 int j, lim = XVECLEN (x, 0);
4707 for (j = 0; j < lim; j++)
4709 rtx y = XVECEXP (x, 0, j);
4710 /* And do not record result of a non-volatile inline asm
4711 with "memory" clobber. */
4712 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4714 sets[i].src_volatile = 1;
4715 break;
4720 #if 0
4721 /* It is no longer clear why we used to do this, but it doesn't
4722 appear to still be needed. So let's try without it since this
4723 code hurts cse'ing widened ops. */
4724 /* If source is a paradoxical subreg (such as QI treated as an SI),
4725 treat it as volatile. It may do the work of an SI in one context
4726 where the extra bits are not being used, but cannot replace an SI
4727 in general. */
4728 if (paradoxical_subreg_p (src))
4729 sets[i].src_volatile = 1;
4730 #endif
4732 /* Locate all possible equivalent forms for SRC. Try to replace
4733 SRC in the insn with each cheaper equivalent.
4735 We have the following types of equivalents: SRC itself, a folded
4736 version, a value given in a REG_EQUAL note, or a value related
4737 to a constant.
4739 Each of these equivalents may be part of an additional class
4740 of equivalents (if more than one is in the table, they must be in
4741 the same class; we check for this).
4743 If the source is volatile, we don't do any table lookups.
4745 We note any constant equivalent for possible later use in a
4746 REG_NOTE. */
4748 if (!sets[i].src_volatile)
4749 elt = lookup (src, sets[i].src_hash, mode);
4751 sets[i].src_elt = elt;
4753 if (elt && src_eqv_here && src_eqv_elt)
4755 if (elt->first_same_value != src_eqv_elt->first_same_value)
4757 /* The REG_EQUAL is indicating that two formerly distinct
4758 classes are now equivalent. So merge them. */
4759 merge_equiv_classes (elt, src_eqv_elt);
4760 src_eqv_hash = HASH (src_eqv, elt->mode);
4761 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4764 src_eqv_here = 0;
4767 else if (src_eqv_elt)
4768 elt = src_eqv_elt;
4770 /* Try to find a constant somewhere and record it in `src_const'.
4771 Record its table element, if any, in `src_const_elt'. Look in
4772 any known equivalences first. (If the constant is not in the
4773 table, also set `sets[i].src_const_hash'). */
4774 if (elt)
4775 for (p = elt->first_same_value; p; p = p->next_same_value)
4776 if (p->is_const)
4778 src_const = p->exp;
4779 src_const_elt = elt;
4780 break;
4783 if (src_const == 0
4784 && (CONSTANT_P (src_folded)
4785 /* Consider (minus (label_ref L1) (label_ref L2)) as
4786 "constant" here so we will record it. This allows us
4787 to fold switch statements when an ADDR_DIFF_VEC is used. */
4788 || (GET_CODE (src_folded) == MINUS
4789 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4790 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4791 src_const = src_folded, src_const_elt = elt;
4792 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4793 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4795 /* If we don't know if the constant is in the table, get its
4796 hash code and look it up. */
4797 if (src_const && src_const_elt == 0)
4799 sets[i].src_const_hash = HASH (src_const, mode);
4800 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4803 sets[i].src_const = src_const;
4804 sets[i].src_const_elt = src_const_elt;
4806 /* If the constant and our source are both in the table, mark them as
4807 equivalent. Otherwise, if a constant is in the table but the source
4808 isn't, set ELT to it. */
4809 if (src_const_elt && elt
4810 && src_const_elt->first_same_value != elt->first_same_value)
4811 merge_equiv_classes (elt, src_const_elt);
4812 else if (src_const_elt && elt == 0)
4813 elt = src_const_elt;
4815 /* See if there is a register linearly related to a constant
4816 equivalent of SRC. */
4817 if (src_const
4818 && (GET_CODE (src_const) == CONST
4819 || (src_const_elt && src_const_elt->related_value != 0)))
4821 src_related = use_related_value (src_const, src_const_elt);
4822 if (src_related)
4824 struct table_elt *src_related_elt
4825 = lookup (src_related, HASH (src_related, mode), mode);
4826 if (src_related_elt && elt)
4828 if (elt->first_same_value
4829 != src_related_elt->first_same_value)
4830 /* This can occur when we previously saw a CONST
4831 involving a SYMBOL_REF and then see the SYMBOL_REF
4832 twice. Merge the involved classes. */
4833 merge_equiv_classes (elt, src_related_elt);
4835 src_related = 0;
4836 src_related_elt = 0;
4838 else if (src_related_elt && elt == 0)
4839 elt = src_related_elt;
4843 /* See if we have a CONST_INT that is already in a register in a
4844 wider mode. */
4846 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4847 && GET_MODE_CLASS (mode) == MODE_INT
4848 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4850 machine_mode wider_mode;
4852 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4853 wider_mode != VOIDmode
4854 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4855 && src_related == 0;
4856 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4858 struct table_elt *const_elt
4859 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4861 if (const_elt == 0)
4862 continue;
4864 for (const_elt = const_elt->first_same_value;
4865 const_elt; const_elt = const_elt->next_same_value)
4866 if (REG_P (const_elt->exp))
4868 src_related = gen_lowpart (mode, const_elt->exp);
4869 break;
4874 /* Another possibility is that we have an AND with a constant in
4875 a mode narrower than a word. If so, it might have been generated
4876 as part of an "if" which would narrow the AND. If we already
4877 have done the AND in a wider mode, we can use a SUBREG of that
4878 value. */
4880 if (flag_expensive_optimizations && ! src_related
4881 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4882 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4884 machine_mode tmode;
4885 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4887 for (tmode = GET_MODE_WIDER_MODE (mode);
4888 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4889 tmode = GET_MODE_WIDER_MODE (tmode))
4891 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4892 struct table_elt *larger_elt;
4894 if (inner)
4896 PUT_MODE (new_and, tmode);
4897 XEXP (new_and, 0) = inner;
4898 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4899 if (larger_elt == 0)
4900 continue;
4902 for (larger_elt = larger_elt->first_same_value;
4903 larger_elt; larger_elt = larger_elt->next_same_value)
4904 if (REG_P (larger_elt->exp))
4906 src_related
4907 = gen_lowpart (mode, larger_elt->exp);
4908 break;
4911 if (src_related)
4912 break;
4917 /* See if a MEM has already been loaded with a widening operation;
4918 if it has, we can use a subreg of that. Many CISC machines
4919 also have such operations, but this is only likely to be
4920 beneficial on these machines. */
4922 if (flag_expensive_optimizations && src_related == 0
4923 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4924 && GET_MODE_CLASS (mode) == MODE_INT
4925 && MEM_P (src) && ! do_not_record
4926 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4928 struct rtx_def memory_extend_buf;
4929 rtx memory_extend_rtx = &memory_extend_buf;
4930 machine_mode tmode;
4932 /* Set what we are trying to extend and the operation it might
4933 have been extended with. */
4934 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4935 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4936 XEXP (memory_extend_rtx, 0) = src;
4938 for (tmode = GET_MODE_WIDER_MODE (mode);
4939 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4940 tmode = GET_MODE_WIDER_MODE (tmode))
4942 struct table_elt *larger_elt;
4944 PUT_MODE (memory_extend_rtx, tmode);
4945 larger_elt = lookup (memory_extend_rtx,
4946 HASH (memory_extend_rtx, tmode), tmode);
4947 if (larger_elt == 0)
4948 continue;
4950 for (larger_elt = larger_elt->first_same_value;
4951 larger_elt; larger_elt = larger_elt->next_same_value)
4952 if (REG_P (larger_elt->exp))
4954 src_related = gen_lowpart (mode, larger_elt->exp);
4955 break;
4958 if (src_related)
4959 break;
4963 /* Try to express the constant using a register+offset expression
4964 derived from a constant anchor. */
4966 if (targetm.const_anchor
4967 && !src_related
4968 && src_const
4969 && GET_CODE (src_const) == CONST_INT)
4971 src_related = try_const_anchors (src_const, mode);
4972 src_related_is_const_anchor = src_related != NULL_RTX;
4976 if (src == src_folded)
4977 src_folded = 0;
4979 /* At this point, ELT, if nonzero, points to a class of expressions
4980 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4981 and SRC_RELATED, if nonzero, each contain additional equivalent
4982 expressions. Prune these latter expressions by deleting expressions
4983 already in the equivalence class.
4985 Check for an equivalent identical to the destination. If found,
4986 this is the preferred equivalent since it will likely lead to
4987 elimination of the insn. Indicate this by placing it in
4988 `src_related'. */
4990 if (elt)
4991 elt = elt->first_same_value;
4992 for (p = elt; p; p = p->next_same_value)
4994 enum rtx_code code = GET_CODE (p->exp);
4996 /* If the expression is not valid, ignore it. Then we do not
4997 have to check for validity below. In most cases, we can use
4998 `rtx_equal_p', since canonicalization has already been done. */
4999 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5000 continue;
5002 /* Also skip paradoxical subregs, unless that's what we're
5003 looking for. */
5004 if (paradoxical_subreg_p (p->exp)
5005 && ! (src != 0
5006 && GET_CODE (src) == SUBREG
5007 && GET_MODE (src) == GET_MODE (p->exp)
5008 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5009 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5010 continue;
5012 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5013 src = 0;
5014 else if (src_folded && GET_CODE (src_folded) == code
5015 && rtx_equal_p (src_folded, p->exp))
5016 src_folded = 0;
5017 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5018 && rtx_equal_p (src_eqv_here, p->exp))
5019 src_eqv_here = 0;
5020 else if (src_related && GET_CODE (src_related) == code
5021 && rtx_equal_p (src_related, p->exp))
5022 src_related = 0;
5024 /* This is the same as the destination of the insns, we want
5025 to prefer it. Copy it to src_related. The code below will
5026 then give it a negative cost. */
5027 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5028 src_related = dest;
5031 /* Find the cheapest valid equivalent, trying all the available
5032 possibilities. Prefer items not in the hash table to ones
5033 that are when they are equal cost. Note that we can never
5034 worsen an insn as the current contents will also succeed.
5035 If we find an equivalent identical to the destination, use it as best,
5036 since this insn will probably be eliminated in that case. */
5037 if (src)
5039 if (rtx_equal_p (src, dest))
5040 src_cost = src_regcost = -1;
5041 else
5043 src_cost = COST (src, mode);
5044 src_regcost = approx_reg_cost (src);
5048 if (src_eqv_here)
5050 if (rtx_equal_p (src_eqv_here, dest))
5051 src_eqv_cost = src_eqv_regcost = -1;
5052 else
5054 src_eqv_cost = COST (src_eqv_here, mode);
5055 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5059 if (src_folded)
5061 if (rtx_equal_p (src_folded, dest))
5062 src_folded_cost = src_folded_regcost = -1;
5063 else
5065 src_folded_cost = COST (src_folded, mode);
5066 src_folded_regcost = approx_reg_cost (src_folded);
5070 if (src_related)
5072 if (rtx_equal_p (src_related, dest))
5073 src_related_cost = src_related_regcost = -1;
5074 else
5076 src_related_cost = COST (src_related, mode);
5077 src_related_regcost = approx_reg_cost (src_related);
5079 /* If a const-anchor is used to synthesize a constant that
5080 normally requires multiple instructions then slightly prefer
5081 it over the original sequence. These instructions are likely
5082 to become redundant now. We can't compare against the cost
5083 of src_eqv_here because, on MIPS for example, multi-insn
5084 constants have zero cost; they are assumed to be hoisted from
5085 loops. */
5086 if (src_related_is_const_anchor
5087 && src_related_cost == src_cost
5088 && src_eqv_here)
5089 src_related_cost--;
5093 /* If this was an indirect jump insn, a known label will really be
5094 cheaper even though it looks more expensive. */
5095 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5096 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5098 /* Terminate loop when replacement made. This must terminate since
5099 the current contents will be tested and will always be valid. */
5100 while (1)
5102 rtx trial;
5104 /* Skip invalid entries. */
5105 while (elt && !REG_P (elt->exp)
5106 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5107 elt = elt->next_same_value;
5109 /* A paradoxical subreg would be bad here: it'll be the right
5110 size, but later may be adjusted so that the upper bits aren't
5111 what we want. So reject it. */
5112 if (elt != 0
5113 && paradoxical_subreg_p (elt->exp)
5114 /* It is okay, though, if the rtx we're trying to match
5115 will ignore any of the bits we can't predict. */
5116 && ! (src != 0
5117 && GET_CODE (src) == SUBREG
5118 && GET_MODE (src) == GET_MODE (elt->exp)
5119 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5120 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5122 elt = elt->next_same_value;
5123 continue;
5126 if (elt)
5128 src_elt_cost = elt->cost;
5129 src_elt_regcost = elt->regcost;
5132 /* Find cheapest and skip it for the next time. For items
5133 of equal cost, use this order:
5134 src_folded, src, src_eqv, src_related and hash table entry. */
5135 if (src_folded
5136 && preferable (src_folded_cost, src_folded_regcost,
5137 src_cost, src_regcost) <= 0
5138 && preferable (src_folded_cost, src_folded_regcost,
5139 src_eqv_cost, src_eqv_regcost) <= 0
5140 && preferable (src_folded_cost, src_folded_regcost,
5141 src_related_cost, src_related_regcost) <= 0
5142 && preferable (src_folded_cost, src_folded_regcost,
5143 src_elt_cost, src_elt_regcost) <= 0)
5145 trial = src_folded, src_folded_cost = MAX_COST;
5146 if (src_folded_force_flag)
5148 rtx forced = force_const_mem (mode, trial);
5149 if (forced)
5150 trial = forced;
5153 else if (src
5154 && preferable (src_cost, src_regcost,
5155 src_eqv_cost, src_eqv_regcost) <= 0
5156 && preferable (src_cost, src_regcost,
5157 src_related_cost, src_related_regcost) <= 0
5158 && preferable (src_cost, src_regcost,
5159 src_elt_cost, src_elt_regcost) <= 0)
5160 trial = src, src_cost = MAX_COST;
5161 else if (src_eqv_here
5162 && preferable (src_eqv_cost, src_eqv_regcost,
5163 src_related_cost, src_related_regcost) <= 0
5164 && preferable (src_eqv_cost, src_eqv_regcost,
5165 src_elt_cost, src_elt_regcost) <= 0)
5166 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5167 else if (src_related
5168 && preferable (src_related_cost, src_related_regcost,
5169 src_elt_cost, src_elt_regcost) <= 0)
5170 trial = src_related, src_related_cost = MAX_COST;
5171 else
5173 trial = elt->exp;
5174 elt = elt->next_same_value;
5175 src_elt_cost = MAX_COST;
5178 /* Avoid creation of overlapping memory moves. */
5179 if (MEM_P (trial) && MEM_P (dest) && !rtx_equal_p (trial, dest))
5181 rtx src, dest;
5183 /* BLKmode moves are not handled by cse anyway. */
5184 if (GET_MODE (trial) == BLKmode)
5185 break;
5187 src = canon_rtx (trial);
5188 dest = canon_rtx (SET_DEST (sets[i].rtl));
5190 if (!MEM_P (src) || !MEM_P (dest)
5191 || !nonoverlapping_memrefs_p (src, dest, false))
5192 break;
5195 /* Try to optimize
5196 (set (reg:M N) (const_int A))
5197 (set (reg:M2 O) (const_int B))
5198 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5199 (reg:M2 O)). */
5200 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5201 && CONST_INT_P (trial)
5202 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5203 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5204 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5205 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5206 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5207 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5208 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5209 <= HOST_BITS_PER_WIDE_INT))
5211 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5212 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5213 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5214 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5215 struct table_elt *dest_elt
5216 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5217 rtx dest_cst = NULL;
5219 if (dest_elt)
5220 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5221 if (p->is_const && CONST_INT_P (p->exp))
5223 dest_cst = p->exp;
5224 break;
5226 if (dest_cst)
5228 HOST_WIDE_INT val = INTVAL (dest_cst);
5229 HOST_WIDE_INT mask;
5230 unsigned int shift;
5231 if (BITS_BIG_ENDIAN)
5232 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5233 - INTVAL (pos) - INTVAL (width);
5234 else
5235 shift = INTVAL (pos);
5236 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5237 mask = HOST_WIDE_INT_M1;
5238 else
5239 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5240 val &= ~(mask << shift);
5241 val |= (INTVAL (trial) & mask) << shift;
5242 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5243 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5244 dest_reg, 1);
5245 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5246 GEN_INT (val), 1);
5247 if (apply_change_group ())
5249 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5250 if (note)
5252 remove_note (insn, note);
5253 df_notes_rescan (insn);
5255 src_eqv = NULL_RTX;
5256 src_eqv_elt = NULL;
5257 src_eqv_volatile = 0;
5258 src_eqv_in_memory = 0;
5259 src_eqv_hash = 0;
5260 repeat = true;
5261 break;
5266 /* We don't normally have an insn matching (set (pc) (pc)), so
5267 check for this separately here. We will delete such an
5268 insn below.
5270 For other cases such as a table jump or conditional jump
5271 where we know the ultimate target, go ahead and replace the
5272 operand. While that may not make a valid insn, we will
5273 reemit the jump below (and also insert any necessary
5274 barriers). */
5275 if (n_sets == 1 && dest == pc_rtx
5276 && (trial == pc_rtx
5277 || (GET_CODE (trial) == LABEL_REF
5278 && ! condjump_p (insn))))
5280 /* Don't substitute non-local labels, this confuses CFG. */
5281 if (GET_CODE (trial) == LABEL_REF
5282 && LABEL_REF_NONLOCAL_P (trial))
5283 continue;
5285 SET_SRC (sets[i].rtl) = trial;
5286 cse_jumps_altered = true;
5287 break;
5290 /* Similarly, lots of targets don't allow no-op
5291 (set (mem x) (mem x)) moves. */
5292 else if (n_sets == 1
5293 && MEM_P (trial)
5294 && MEM_P (dest)
5295 && rtx_equal_p (trial, dest)
5296 && !side_effects_p (dest)
5297 && (cfun->can_delete_dead_exceptions
5298 || insn_nothrow_p (insn)))
5300 SET_SRC (sets[i].rtl) = trial;
5301 mem_noop_insn = true;
5302 break;
5305 /* Reject certain invalid forms of CONST that we create. */
5306 else if (CONSTANT_P (trial)
5307 && GET_CODE (trial) == CONST
5308 /* Reject cases that will cause decode_rtx_const to
5309 die. On the alpha when simplifying a switch, we
5310 get (const (truncate (minus (label_ref)
5311 (label_ref)))). */
5312 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5313 /* Likewise on IA-64, except without the
5314 truncate. */
5315 || (GET_CODE (XEXP (trial, 0)) == MINUS
5316 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5317 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5318 /* Do nothing for this case. */
5321 /* Look for a substitution that makes a valid insn. */
5322 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5323 trial, 0))
5325 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5327 /* The result of apply_change_group can be ignored; see
5328 canon_reg. */
5330 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5331 apply_change_group ();
5333 break;
5336 /* If we previously found constant pool entries for
5337 constants and this is a constant, try making a
5338 pool entry. Put it in src_folded unless we already have done
5339 this since that is where it likely came from. */
5341 else if (constant_pool_entries_cost
5342 && CONSTANT_P (trial)
5343 && (src_folded == 0
5344 || (!MEM_P (src_folded)
5345 && ! src_folded_force_flag))
5346 && GET_MODE_CLASS (mode) != MODE_CC
5347 && mode != VOIDmode)
5349 src_folded_force_flag = 1;
5350 src_folded = trial;
5351 src_folded_cost = constant_pool_entries_cost;
5352 src_folded_regcost = constant_pool_entries_regcost;
5356 /* If we changed the insn too much, handle this set from scratch. */
5357 if (repeat)
5359 i--;
5360 continue;
5363 src = SET_SRC (sets[i].rtl);
5365 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5366 However, there is an important exception: If both are registers
5367 that are not the head of their equivalence class, replace SET_SRC
5368 with the head of the class. If we do not do this, we will have
5369 both registers live over a portion of the basic block. This way,
5370 their lifetimes will likely abut instead of overlapping. */
5371 if (REG_P (dest)
5372 && REGNO_QTY_VALID_P (REGNO (dest)))
5374 int dest_q = REG_QTY (REGNO (dest));
5375 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5377 if (dest_ent->mode == GET_MODE (dest)
5378 && dest_ent->first_reg != REGNO (dest)
5379 && REG_P (src) && REGNO (src) == REGNO (dest)
5380 /* Don't do this if the original insn had a hard reg as
5381 SET_SRC or SET_DEST. */
5382 && (!REG_P (sets[i].src)
5383 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5384 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5385 /* We can't call canon_reg here because it won't do anything if
5386 SRC is a hard register. */
5388 int src_q = REG_QTY (REGNO (src));
5389 struct qty_table_elem *src_ent = &qty_table[src_q];
5390 int first = src_ent->first_reg;
5391 rtx new_src
5392 = (first >= FIRST_PSEUDO_REGISTER
5393 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5395 /* We must use validate-change even for this, because this
5396 might be a special no-op instruction, suitable only to
5397 tag notes onto. */
5398 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5400 src = new_src;
5401 /* If we had a constant that is cheaper than what we are now
5402 setting SRC to, use that constant. We ignored it when we
5403 thought we could make this into a no-op. */
5404 if (src_const && COST (src_const, mode) < COST (src, mode)
5405 && validate_change (insn, &SET_SRC (sets[i].rtl),
5406 src_const, 0))
5407 src = src_const;
5412 /* If we made a change, recompute SRC values. */
5413 if (src != sets[i].src)
5415 do_not_record = 0;
5416 hash_arg_in_memory = 0;
5417 sets[i].src = src;
5418 sets[i].src_hash = HASH (src, mode);
5419 sets[i].src_volatile = do_not_record;
5420 sets[i].src_in_memory = hash_arg_in_memory;
5421 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5424 /* If this is a single SET, we are setting a register, and we have an
5425 equivalent constant, we want to add a REG_EQUAL note if the constant
5426 is different from the source. We don't want to do it for a constant
5427 pseudo since verifying that this pseudo hasn't been eliminated is a
5428 pain; moreover such a note won't help anything.
5430 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5431 which can be created for a reference to a compile time computable
5432 entry in a jump table. */
5433 if (n_sets == 1
5434 && REG_P (dest)
5435 && src_const
5436 && !REG_P (src_const)
5437 && !(GET_CODE (src_const) == SUBREG
5438 && REG_P (SUBREG_REG (src_const)))
5439 && !(GET_CODE (src_const) == CONST
5440 && GET_CODE (XEXP (src_const, 0)) == MINUS
5441 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5442 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5443 && !rtx_equal_p (src, src_const))
5445 /* Make sure that the rtx is not shared. */
5446 src_const = copy_rtx (src_const);
5448 /* Record the actual constant value in a REG_EQUAL note,
5449 making a new one if one does not already exist. */
5450 set_unique_reg_note (insn, REG_EQUAL, src_const);
5451 df_notes_rescan (insn);
5454 /* Now deal with the destination. */
5455 do_not_record = 0;
5457 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5458 while (GET_CODE (dest) == SUBREG
5459 || GET_CODE (dest) == ZERO_EXTRACT
5460 || GET_CODE (dest) == STRICT_LOW_PART)
5461 dest = XEXP (dest, 0);
5463 sets[i].inner_dest = dest;
5465 if (MEM_P (dest))
5467 #ifdef PUSH_ROUNDING
5468 /* Stack pushes invalidate the stack pointer. */
5469 rtx addr = XEXP (dest, 0);
5470 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5471 && XEXP (addr, 0) == stack_pointer_rtx)
5472 invalidate (stack_pointer_rtx, VOIDmode);
5473 #endif
5474 dest = fold_rtx (dest, insn);
5477 /* Compute the hash code of the destination now,
5478 before the effects of this instruction are recorded,
5479 since the register values used in the address computation
5480 are those before this instruction. */
5481 sets[i].dest_hash = HASH (dest, mode);
5483 /* Don't enter a bit-field in the hash table
5484 because the value in it after the store
5485 may not equal what was stored, due to truncation. */
5487 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5489 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5491 if (src_const != 0 && CONST_INT_P (src_const)
5492 && CONST_INT_P (width)
5493 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5494 && ! (INTVAL (src_const)
5495 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5496 /* Exception: if the value is constant,
5497 and it won't be truncated, record it. */
5499 else
5501 /* This is chosen so that the destination will be invalidated
5502 but no new value will be recorded.
5503 We must invalidate because sometimes constant
5504 values can be recorded for bitfields. */
5505 sets[i].src_elt = 0;
5506 sets[i].src_volatile = 1;
5507 src_eqv = 0;
5508 src_eqv_elt = 0;
5512 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5513 the insn. */
5514 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5516 /* One less use of the label this insn used to jump to. */
5517 cse_cfg_altered |= delete_insn_and_edges (insn);
5518 cse_jumps_altered = true;
5519 /* No more processing for this set. */
5520 sets[i].rtl = 0;
5523 /* Similarly for no-op MEM moves. */
5524 else if (mem_noop_insn)
5526 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5527 cse_cfg_altered = true;
5528 cse_cfg_altered |= delete_insn_and_edges (insn);
5529 /* No more processing for this set. */
5530 sets[i].rtl = 0;
5533 /* If this SET is now setting PC to a label, we know it used to
5534 be a conditional or computed branch. */
5535 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5536 && !LABEL_REF_NONLOCAL_P (src))
5538 /* We reemit the jump in as many cases as possible just in
5539 case the form of an unconditional jump is significantly
5540 different than a computed jump or conditional jump.
5542 If this insn has multiple sets, then reemitting the
5543 jump is nontrivial. So instead we just force rerecognition
5544 and hope for the best. */
5545 if (n_sets == 1)
5547 rtx_jump_insn *new_rtx;
5548 rtx note;
5550 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5551 new_rtx = emit_jump_insn_before (seq, insn);
5552 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5553 LABEL_NUSES (XEXP (src, 0))++;
5555 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5556 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5557 if (note)
5559 XEXP (note, 1) = NULL_RTX;
5560 REG_NOTES (new_rtx) = note;
5563 cse_cfg_altered |= delete_insn_and_edges (insn);
5564 insn = new_rtx;
5566 else
5567 INSN_CODE (insn) = -1;
5569 /* Do not bother deleting any unreachable code, let jump do it. */
5570 cse_jumps_altered = true;
5571 sets[i].rtl = 0;
5574 /* If destination is volatile, invalidate it and then do no further
5575 processing for this assignment. */
5577 else if (do_not_record)
5579 invalidate_dest (dest);
5580 sets[i].rtl = 0;
5583 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5585 do_not_record = 0;
5586 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5587 if (do_not_record)
5589 invalidate_dest (SET_DEST (sets[i].rtl));
5590 sets[i].rtl = 0;
5594 /* If setting CC0, record what it was set to, or a constant, if it
5595 is equivalent to a constant. If it is being set to a floating-point
5596 value, make a COMPARE with the appropriate constant of 0. If we
5597 don't do this, later code can interpret this as a test against
5598 const0_rtx, which can cause problems if we try to put it into an
5599 insn as a floating-point operand. */
5600 if (dest == cc0_rtx)
5602 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5603 this_insn_cc0_mode = mode;
5604 if (FLOAT_MODE_P (mode))
5605 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5606 CONST0_RTX (mode));
5610 /* Now enter all non-volatile source expressions in the hash table
5611 if they are not already present.
5612 Record their equivalence classes in src_elt.
5613 This way we can insert the corresponding destinations into
5614 the same classes even if the actual sources are no longer in them
5615 (having been invalidated). */
5617 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5618 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5620 struct table_elt *elt;
5621 struct table_elt *classp = sets[0].src_elt;
5622 rtx dest = SET_DEST (sets[0].rtl);
5623 machine_mode eqvmode = GET_MODE (dest);
5625 if (GET_CODE (dest) == STRICT_LOW_PART)
5627 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5628 classp = 0;
5630 if (insert_regs (src_eqv, classp, 0))
5632 rehash_using_reg (src_eqv);
5633 src_eqv_hash = HASH (src_eqv, eqvmode);
5635 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5636 elt->in_memory = src_eqv_in_memory;
5637 src_eqv_elt = elt;
5639 /* Check to see if src_eqv_elt is the same as a set source which
5640 does not yet have an elt, and if so set the elt of the set source
5641 to src_eqv_elt. */
5642 for (i = 0; i < n_sets; i++)
5643 if (sets[i].rtl && sets[i].src_elt == 0
5644 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5645 sets[i].src_elt = src_eqv_elt;
5648 for (i = 0; i < n_sets; i++)
5649 if (sets[i].rtl && ! sets[i].src_volatile
5650 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5652 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5654 /* REG_EQUAL in setting a STRICT_LOW_PART
5655 gives an equivalent for the entire destination register,
5656 not just for the subreg being stored in now.
5657 This is a more interesting equivalence, so we arrange later
5658 to treat the entire reg as the destination. */
5659 sets[i].src_elt = src_eqv_elt;
5660 sets[i].src_hash = src_eqv_hash;
5662 else
5664 /* Insert source and constant equivalent into hash table, if not
5665 already present. */
5666 struct table_elt *classp = src_eqv_elt;
5667 rtx src = sets[i].src;
5668 rtx dest = SET_DEST (sets[i].rtl);
5669 machine_mode mode
5670 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5672 /* It's possible that we have a source value known to be
5673 constant but don't have a REG_EQUAL note on the insn.
5674 Lack of a note will mean src_eqv_elt will be NULL. This
5675 can happen where we've generated a SUBREG to access a
5676 CONST_INT that is already in a register in a wider mode.
5677 Ensure that the source expression is put in the proper
5678 constant class. */
5679 if (!classp)
5680 classp = sets[i].src_const_elt;
5682 if (sets[i].src_elt == 0)
5684 struct table_elt *elt;
5686 /* Note that these insert_regs calls cannot remove
5687 any of the src_elt's, because they would have failed to
5688 match if not still valid. */
5689 if (insert_regs (src, classp, 0))
5691 rehash_using_reg (src);
5692 sets[i].src_hash = HASH (src, mode);
5694 elt = insert (src, classp, sets[i].src_hash, mode);
5695 elt->in_memory = sets[i].src_in_memory;
5696 /* If inline asm has any clobbers, ensure we only reuse
5697 existing inline asms and never try to put the ASM_OPERANDS
5698 into an insn that isn't inline asm. */
5699 if (GET_CODE (src) == ASM_OPERANDS
5700 && GET_CODE (x) == PARALLEL)
5701 elt->cost = MAX_COST;
5702 sets[i].src_elt = classp = elt;
5704 if (sets[i].src_const && sets[i].src_const_elt == 0
5705 && src != sets[i].src_const
5706 && ! rtx_equal_p (sets[i].src_const, src))
5707 sets[i].src_elt = insert (sets[i].src_const, classp,
5708 sets[i].src_const_hash, mode);
5711 else if (sets[i].src_elt == 0)
5712 /* If we did not insert the source into the hash table (e.g., it was
5713 volatile), note the equivalence class for the REG_EQUAL value, if any,
5714 so that the destination goes into that class. */
5715 sets[i].src_elt = src_eqv_elt;
5717 /* Record destination addresses in the hash table. This allows us to
5718 check if they are invalidated by other sets. */
5719 for (i = 0; i < n_sets; i++)
5721 if (sets[i].rtl)
5723 rtx x = sets[i].inner_dest;
5724 struct table_elt *elt;
5725 machine_mode mode;
5726 unsigned hash;
5728 if (MEM_P (x))
5730 x = XEXP (x, 0);
5731 mode = GET_MODE (x);
5732 hash = HASH (x, mode);
5733 elt = lookup (x, hash, mode);
5734 if (!elt)
5736 if (insert_regs (x, NULL, 0))
5738 rtx dest = SET_DEST (sets[i].rtl);
5740 rehash_using_reg (x);
5741 hash = HASH (x, mode);
5742 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5744 elt = insert (x, NULL, hash, mode);
5747 sets[i].dest_addr_elt = elt;
5749 else
5750 sets[i].dest_addr_elt = NULL;
5754 invalidate_from_clobbers (insn);
5756 /* Some registers are invalidated by subroutine calls. Memory is
5757 invalidated by non-constant calls. */
5759 if (CALL_P (insn))
5761 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5762 invalidate_memory ();
5763 else
5764 /* For const/pure calls, invalidate any argument slots, because
5765 those are owned by the callee. */
5766 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5767 if (GET_CODE (XEXP (tem, 0)) == USE
5768 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5769 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5770 invalidate_for_call ();
5773 /* Now invalidate everything set by this instruction.
5774 If a SUBREG or other funny destination is being set,
5775 sets[i].rtl is still nonzero, so here we invalidate the reg
5776 a part of which is being set. */
5778 for (i = 0; i < n_sets; i++)
5779 if (sets[i].rtl)
5781 /* We can't use the inner dest, because the mode associated with
5782 a ZERO_EXTRACT is significant. */
5783 rtx dest = SET_DEST (sets[i].rtl);
5785 /* Needed for registers to remove the register from its
5786 previous quantity's chain.
5787 Needed for memory if this is a nonvarying address, unless
5788 we have just done an invalidate_memory that covers even those. */
5789 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5790 invalidate (dest, VOIDmode);
5791 else if (MEM_P (dest))
5792 invalidate (dest, VOIDmode);
5793 else if (GET_CODE (dest) == STRICT_LOW_PART
5794 || GET_CODE (dest) == ZERO_EXTRACT)
5795 invalidate (XEXP (dest, 0), GET_MODE (dest));
5798 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5799 the regs restored by the longjmp come from a later time
5800 than the setjmp. */
5801 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5803 flush_hash_table ();
5804 goto done;
5807 /* Make sure registers mentioned in destinations
5808 are safe for use in an expression to be inserted.
5809 This removes from the hash table
5810 any invalid entry that refers to one of these registers.
5812 We don't care about the return value from mention_regs because
5813 we are going to hash the SET_DEST values unconditionally. */
5815 for (i = 0; i < n_sets; i++)
5817 if (sets[i].rtl)
5819 rtx x = SET_DEST (sets[i].rtl);
5821 if (!REG_P (x))
5822 mention_regs (x);
5823 else
5825 /* We used to rely on all references to a register becoming
5826 inaccessible when a register changes to a new quantity,
5827 since that changes the hash code. However, that is not
5828 safe, since after HASH_SIZE new quantities we get a
5829 hash 'collision' of a register with its own invalid
5830 entries. And since SUBREGs have been changed not to
5831 change their hash code with the hash code of the register,
5832 it wouldn't work any longer at all. So we have to check
5833 for any invalid references lying around now.
5834 This code is similar to the REG case in mention_regs,
5835 but it knows that reg_tick has been incremented, and
5836 it leaves reg_in_table as -1 . */
5837 unsigned int regno = REGNO (x);
5838 unsigned int endregno = END_REGNO (x);
5839 unsigned int i;
5841 for (i = regno; i < endregno; i++)
5843 if (REG_IN_TABLE (i) >= 0)
5845 remove_invalid_refs (i);
5846 REG_IN_TABLE (i) = -1;
5853 /* We may have just removed some of the src_elt's from the hash table.
5854 So replace each one with the current head of the same class.
5855 Also check if destination addresses have been removed. */
5857 for (i = 0; i < n_sets; i++)
5858 if (sets[i].rtl)
5860 if (sets[i].dest_addr_elt
5861 && sets[i].dest_addr_elt->first_same_value == 0)
5863 /* The elt was removed, which means this destination is not
5864 valid after this instruction. */
5865 sets[i].rtl = NULL_RTX;
5867 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5868 /* If elt was removed, find current head of same class,
5869 or 0 if nothing remains of that class. */
5871 struct table_elt *elt = sets[i].src_elt;
5873 while (elt && elt->prev_same_value)
5874 elt = elt->prev_same_value;
5876 while (elt && elt->first_same_value == 0)
5877 elt = elt->next_same_value;
5878 sets[i].src_elt = elt ? elt->first_same_value : 0;
5882 /* Now insert the destinations into their equivalence classes. */
5884 for (i = 0; i < n_sets; i++)
5885 if (sets[i].rtl)
5887 rtx dest = SET_DEST (sets[i].rtl);
5888 struct table_elt *elt;
5890 /* Don't record value if we are not supposed to risk allocating
5891 floating-point values in registers that might be wider than
5892 memory. */
5893 if ((flag_float_store
5894 && MEM_P (dest)
5895 && FLOAT_MODE_P (GET_MODE (dest)))
5896 /* Don't record BLKmode values, because we don't know the
5897 size of it, and can't be sure that other BLKmode values
5898 have the same or smaller size. */
5899 || GET_MODE (dest) == BLKmode
5900 /* If we didn't put a REG_EQUAL value or a source into the hash
5901 table, there is no point is recording DEST. */
5902 || sets[i].src_elt == 0)
5903 continue;
5905 /* STRICT_LOW_PART isn't part of the value BEING set,
5906 and neither is the SUBREG inside it.
5907 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5908 if (GET_CODE (dest) == STRICT_LOW_PART)
5909 dest = SUBREG_REG (XEXP (dest, 0));
5911 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5912 /* Registers must also be inserted into chains for quantities. */
5913 if (insert_regs (dest, sets[i].src_elt, 1))
5915 /* If `insert_regs' changes something, the hash code must be
5916 recalculated. */
5917 rehash_using_reg (dest);
5918 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5921 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5922 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5923 if (paradoxical_subreg_p (dest))
5924 continue;
5926 elt = insert (dest, sets[i].src_elt,
5927 sets[i].dest_hash, GET_MODE (dest));
5929 /* If this is a constant, insert the constant anchors with the
5930 equivalent register-offset expressions using register DEST. */
5931 if (targetm.const_anchor
5932 && REG_P (dest)
5933 && SCALAR_INT_MODE_P (GET_MODE (dest))
5934 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5935 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5937 elt->in_memory = (MEM_P (sets[i].inner_dest)
5938 && !MEM_READONLY_P (sets[i].inner_dest));
5940 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5941 narrower than M2, and both M1 and M2 are the same number of words,
5942 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5943 make that equivalence as well.
5945 However, BAR may have equivalences for which gen_lowpart
5946 will produce a simpler value than gen_lowpart applied to
5947 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5948 BAR's equivalences. If we don't get a simplified form, make
5949 the SUBREG. It will not be used in an equivalence, but will
5950 cause two similar assignments to be detected.
5952 Note the loop below will find SUBREG_REG (DEST) since we have
5953 already entered SRC and DEST of the SET in the table. */
5955 if (GET_CODE (dest) == SUBREG
5956 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5957 / UNITS_PER_WORD)
5958 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5959 && (GET_MODE_SIZE (GET_MODE (dest))
5960 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5961 && sets[i].src_elt != 0)
5963 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5964 struct table_elt *elt, *classp = 0;
5966 for (elt = sets[i].src_elt->first_same_value; elt;
5967 elt = elt->next_same_value)
5969 rtx new_src = 0;
5970 unsigned src_hash;
5971 struct table_elt *src_elt;
5972 int byte = 0;
5974 /* Ignore invalid entries. */
5975 if (!REG_P (elt->exp)
5976 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5977 continue;
5979 /* We may have already been playing subreg games. If the
5980 mode is already correct for the destination, use it. */
5981 if (GET_MODE (elt->exp) == new_mode)
5982 new_src = elt->exp;
5983 else
5985 /* Calculate big endian correction for the SUBREG_BYTE.
5986 We have already checked that M1 (GET_MODE (dest))
5987 is not narrower than M2 (new_mode). */
5988 if (BYTES_BIG_ENDIAN)
5989 byte = (GET_MODE_SIZE (GET_MODE (dest))
5990 - GET_MODE_SIZE (new_mode));
5992 new_src = simplify_gen_subreg (new_mode, elt->exp,
5993 GET_MODE (dest), byte);
5996 /* The call to simplify_gen_subreg fails if the value
5997 is VOIDmode, yet we can't do any simplification, e.g.
5998 for EXPR_LISTs denoting function call results.
5999 It is invalid to construct a SUBREG with a VOIDmode
6000 SUBREG_REG, hence a zero new_src means we can't do
6001 this substitution. */
6002 if (! new_src)
6003 continue;
6005 src_hash = HASH (new_src, new_mode);
6006 src_elt = lookup (new_src, src_hash, new_mode);
6008 /* Put the new source in the hash table is if isn't
6009 already. */
6010 if (src_elt == 0)
6012 if (insert_regs (new_src, classp, 0))
6014 rehash_using_reg (new_src);
6015 src_hash = HASH (new_src, new_mode);
6017 src_elt = insert (new_src, classp, src_hash, new_mode);
6018 src_elt->in_memory = elt->in_memory;
6019 if (GET_CODE (new_src) == ASM_OPERANDS
6020 && elt->cost == MAX_COST)
6021 src_elt->cost = MAX_COST;
6023 else if (classp && classp != src_elt->first_same_value)
6024 /* Show that two things that we've seen before are
6025 actually the same. */
6026 merge_equiv_classes (src_elt, classp);
6028 classp = src_elt->first_same_value;
6029 /* Ignore invalid entries. */
6030 while (classp
6031 && !REG_P (classp->exp)
6032 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6033 classp = classp->next_same_value;
6038 /* Special handling for (set REG0 REG1) where REG0 is the
6039 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6040 be used in the sequel, so (if easily done) change this insn to
6041 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6042 that computed their value. Then REG1 will become a dead store
6043 and won't cloud the situation for later optimizations.
6045 Do not make this change if REG1 is a hard register, because it will
6046 then be used in the sequel and we may be changing a two-operand insn
6047 into a three-operand insn.
6049 Also do not do this if we are operating on a copy of INSN. */
6051 if (n_sets == 1 && sets[0].rtl)
6052 try_back_substitute_reg (sets[0].rtl, insn);
6054 done:;
6057 /* Remove from the hash table all expressions that reference memory. */
6059 static void
6060 invalidate_memory (void)
6062 int i;
6063 struct table_elt *p, *next;
6065 for (i = 0; i < HASH_SIZE; i++)
6066 for (p = table[i]; p; p = next)
6068 next = p->next_same_hash;
6069 if (p->in_memory)
6070 remove_from_table (p, i);
6074 /* Perform invalidation on the basis of everything about INSN,
6075 except for invalidating the actual places that are SET in it.
6076 This includes the places CLOBBERed, and anything that might
6077 alias with something that is SET or CLOBBERed. */
6079 static void
6080 invalidate_from_clobbers (rtx_insn *insn)
6082 rtx x = PATTERN (insn);
6084 if (GET_CODE (x) == CLOBBER)
6086 rtx ref = XEXP (x, 0);
6087 if (ref)
6089 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6090 || MEM_P (ref))
6091 invalidate (ref, VOIDmode);
6092 else if (GET_CODE (ref) == STRICT_LOW_PART
6093 || GET_CODE (ref) == ZERO_EXTRACT)
6094 invalidate (XEXP (ref, 0), GET_MODE (ref));
6097 else if (GET_CODE (x) == PARALLEL)
6099 int i;
6100 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6102 rtx y = XVECEXP (x, 0, i);
6103 if (GET_CODE (y) == CLOBBER)
6105 rtx ref = XEXP (y, 0);
6106 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6107 || MEM_P (ref))
6108 invalidate (ref, VOIDmode);
6109 else if (GET_CODE (ref) == STRICT_LOW_PART
6110 || GET_CODE (ref) == ZERO_EXTRACT)
6111 invalidate (XEXP (ref, 0), GET_MODE (ref));
6117 /* Perform invalidation on the basis of everything about INSN.
6118 This includes the places CLOBBERed, and anything that might
6119 alias with something that is SET or CLOBBERed. */
6121 static void
6122 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6124 rtx tem;
6125 rtx x = PATTERN (insn);
6127 if (CALL_P (insn))
6129 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6130 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6131 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6134 /* Ensure we invalidate the destination register of a CALL insn.
6135 This is necessary for machines where this register is a fixed_reg,
6136 because no other code would invalidate it. */
6137 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6138 invalidate (SET_DEST (x), VOIDmode);
6140 else if (GET_CODE (x) == PARALLEL)
6142 int i;
6144 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6146 rtx y = XVECEXP (x, 0, i);
6147 if (GET_CODE (y) == CLOBBER)
6149 rtx clobbered = XEXP (y, 0);
6151 if (REG_P (clobbered)
6152 || GET_CODE (clobbered) == SUBREG)
6153 invalidate (clobbered, VOIDmode);
6154 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6155 || GET_CODE (clobbered) == ZERO_EXTRACT)
6156 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6158 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6159 invalidate (SET_DEST (y), VOIDmode);
6164 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6165 and replace any registers in them with either an equivalent constant
6166 or the canonical form of the register. If we are inside an address,
6167 only do this if the address remains valid.
6169 OBJECT is 0 except when within a MEM in which case it is the MEM.
6171 Return the replacement for X. */
6173 static rtx
6174 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6176 enum rtx_code code = GET_CODE (x);
6177 const char *fmt = GET_RTX_FORMAT (code);
6178 int i;
6180 switch (code)
6182 case CONST:
6183 case SYMBOL_REF:
6184 case LABEL_REF:
6185 CASE_CONST_ANY:
6186 case PC:
6187 case CC0:
6188 case LO_SUM:
6189 return x;
6191 case MEM:
6192 validate_change (x, &XEXP (x, 0),
6193 cse_process_notes (XEXP (x, 0), x, changed), 0);
6194 return x;
6196 case EXPR_LIST:
6197 if (REG_NOTE_KIND (x) == REG_EQUAL)
6198 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6199 /* Fall through. */
6201 case INSN_LIST:
6202 case INT_LIST:
6203 if (XEXP (x, 1))
6204 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6205 return x;
6207 case SIGN_EXTEND:
6208 case ZERO_EXTEND:
6209 case SUBREG:
6211 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6212 /* We don't substitute VOIDmode constants into these rtx,
6213 since they would impede folding. */
6214 if (GET_MODE (new_rtx) != VOIDmode)
6215 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6216 return x;
6219 case UNSIGNED_FLOAT:
6221 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6222 /* We don't substitute negative VOIDmode constants into these rtx,
6223 since they would impede folding. */
6224 if (GET_MODE (new_rtx) != VOIDmode
6225 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6226 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6227 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6228 return x;
6231 case REG:
6232 i = REG_QTY (REGNO (x));
6234 /* Return a constant or a constant register. */
6235 if (REGNO_QTY_VALID_P (REGNO (x)))
6237 struct qty_table_elem *ent = &qty_table[i];
6239 if (ent->const_rtx != NULL_RTX
6240 && (CONSTANT_P (ent->const_rtx)
6241 || REG_P (ent->const_rtx)))
6243 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6244 if (new_rtx)
6245 return copy_rtx (new_rtx);
6249 /* Otherwise, canonicalize this register. */
6250 return canon_reg (x, NULL);
6252 default:
6253 break;
6256 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6257 if (fmt[i] == 'e')
6258 validate_change (object, &XEXP (x, i),
6259 cse_process_notes (XEXP (x, i), object, changed), 0);
6261 return x;
6264 static rtx
6265 cse_process_notes (rtx x, rtx object, bool *changed)
6267 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6268 if (new_rtx != x)
6269 *changed = true;
6270 return new_rtx;
6274 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6276 DATA is a pointer to a struct cse_basic_block_data, that is used to
6277 describe the path.
6278 It is filled with a queue of basic blocks, starting with FIRST_BB
6279 and following a trace through the CFG.
6281 If all paths starting at FIRST_BB have been followed, or no new path
6282 starting at FIRST_BB can be constructed, this function returns FALSE.
6283 Otherwise, DATA->path is filled and the function returns TRUE indicating
6284 that a path to follow was found.
6286 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6287 block in the path will be FIRST_BB. */
6289 static bool
6290 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6291 int follow_jumps)
6293 basic_block bb;
6294 edge e;
6295 int path_size;
6297 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6299 /* See if there is a previous path. */
6300 path_size = data->path_size;
6302 /* There is a previous path. Make sure it started with FIRST_BB. */
6303 if (path_size)
6304 gcc_assert (data->path[0].bb == first_bb);
6306 /* There was only one basic block in the last path. Clear the path and
6307 return, so that paths starting at another basic block can be tried. */
6308 if (path_size == 1)
6310 path_size = 0;
6311 goto done;
6314 /* If the path was empty from the beginning, construct a new path. */
6315 if (path_size == 0)
6316 data->path[path_size++].bb = first_bb;
6317 else
6319 /* Otherwise, path_size must be equal to or greater than 2, because
6320 a previous path exists that is at least two basic blocks long.
6322 Update the previous branch path, if any. If the last branch was
6323 previously along the branch edge, take the fallthrough edge now. */
6324 while (path_size >= 2)
6326 basic_block last_bb_in_path, previous_bb_in_path;
6327 edge e;
6329 --path_size;
6330 last_bb_in_path = data->path[path_size].bb;
6331 previous_bb_in_path = data->path[path_size - 1].bb;
6333 /* If we previously followed a path along the branch edge, try
6334 the fallthru edge now. */
6335 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6336 && any_condjump_p (BB_END (previous_bb_in_path))
6337 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6338 && e == BRANCH_EDGE (previous_bb_in_path))
6340 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6341 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6342 && single_pred_p (bb)
6343 /* We used to assert here that we would only see blocks
6344 that we have not visited yet. But we may end up
6345 visiting basic blocks twice if the CFG has changed
6346 in this run of cse_main, because when the CFG changes
6347 the topological sort of the CFG also changes. A basic
6348 blocks that previously had more than two predecessors
6349 may now have a single predecessor, and become part of
6350 a path that starts at another basic block.
6352 We still want to visit each basic block only once, so
6353 halt the path here if we have already visited BB. */
6354 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6356 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6357 data->path[path_size++].bb = bb;
6358 break;
6362 data->path[path_size].bb = NULL;
6365 /* If only one block remains in the path, bail. */
6366 if (path_size == 1)
6368 path_size = 0;
6369 goto done;
6373 /* Extend the path if possible. */
6374 if (follow_jumps)
6376 bb = data->path[path_size - 1].bb;
6377 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6379 if (single_succ_p (bb))
6380 e = single_succ_edge (bb);
6381 else if (EDGE_COUNT (bb->succs) == 2
6382 && any_condjump_p (BB_END (bb)))
6384 /* First try to follow the branch. If that doesn't lead
6385 to a useful path, follow the fallthru edge. */
6386 e = BRANCH_EDGE (bb);
6387 if (!single_pred_p (e->dest))
6388 e = FALLTHRU_EDGE (bb);
6390 else
6391 e = NULL;
6393 if (e
6394 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6395 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6396 && single_pred_p (e->dest)
6397 /* Avoid visiting basic blocks twice. The large comment
6398 above explains why this can happen. */
6399 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6401 basic_block bb2 = e->dest;
6402 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6403 data->path[path_size++].bb = bb2;
6404 bb = bb2;
6406 else
6407 bb = NULL;
6411 done:
6412 data->path_size = path_size;
6413 return path_size != 0;
6416 /* Dump the path in DATA to file F. NSETS is the number of sets
6417 in the path. */
6419 static void
6420 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6422 int path_entry;
6424 fprintf (f, ";; Following path with %d sets: ", nsets);
6425 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6426 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6427 fputc ('\n', dump_file);
6428 fflush (f);
6432 /* Return true if BB has exception handling successor edges. */
6434 static bool
6435 have_eh_succ_edges (basic_block bb)
6437 edge e;
6438 edge_iterator ei;
6440 FOR_EACH_EDGE (e, ei, bb->succs)
6441 if (e->flags & EDGE_EH)
6442 return true;
6444 return false;
6448 /* Scan to the end of the path described by DATA. Return an estimate of
6449 the total number of SETs of all insns in the path. */
6451 static void
6452 cse_prescan_path (struct cse_basic_block_data *data)
6454 int nsets = 0;
6455 int path_size = data->path_size;
6456 int path_entry;
6458 /* Scan to end of each basic block in the path. */
6459 for (path_entry = 0; path_entry < path_size; path_entry++)
6461 basic_block bb;
6462 rtx_insn *insn;
6464 bb = data->path[path_entry].bb;
6466 FOR_BB_INSNS (bb, insn)
6468 if (!INSN_P (insn))
6469 continue;
6471 /* A PARALLEL can have lots of SETs in it,
6472 especially if it is really an ASM_OPERANDS. */
6473 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6474 nsets += XVECLEN (PATTERN (insn), 0);
6475 else
6476 nsets += 1;
6480 data->nsets = nsets;
6483 /* Return true if the pattern of INSN uses a LABEL_REF for which
6484 there isn't a REG_LABEL_OPERAND note. */
6486 static bool
6487 check_for_label_ref (rtx_insn *insn)
6489 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6490 note for it, we must rerun jump since it needs to place the note. If
6491 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6492 don't do this since no REG_LABEL_OPERAND will be added. */
6493 subrtx_iterator::array_type array;
6494 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6496 const_rtx x = *iter;
6497 if (GET_CODE (x) == LABEL_REF
6498 && !LABEL_REF_NONLOCAL_P (x)
6499 && (!JUMP_P (insn)
6500 || !label_is_jump_target_p (label_ref_label (x), insn))
6501 && LABEL_P (label_ref_label (x))
6502 && INSN_UID (label_ref_label (x)) != 0
6503 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6504 return true;
6506 return false;
6509 /* Process a single extended basic block described by EBB_DATA. */
6511 static void
6512 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6514 int path_size = ebb_data->path_size;
6515 int path_entry;
6516 int num_insns = 0;
6518 /* Allocate the space needed by qty_table. */
6519 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6521 new_basic_block ();
6522 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6523 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6524 for (path_entry = 0; path_entry < path_size; path_entry++)
6526 basic_block bb;
6527 rtx_insn *insn;
6529 bb = ebb_data->path[path_entry].bb;
6531 /* Invalidate recorded information for eh regs if there is an EH
6532 edge pointing to that bb. */
6533 if (bb_has_eh_pred (bb))
6535 df_ref def;
6537 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6538 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6539 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6542 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6543 FOR_BB_INSNS (bb, insn)
6545 /* If we have processed 1,000 insns, flush the hash table to
6546 avoid extreme quadratic behavior. We must not include NOTEs
6547 in the count since there may be more of them when generating
6548 debugging information. If we clear the table at different
6549 times, code generated with -g -O might be different than code
6550 generated with -O but not -g.
6552 FIXME: This is a real kludge and needs to be done some other
6553 way. */
6554 if (NONDEBUG_INSN_P (insn)
6555 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6557 flush_hash_table ();
6558 num_insns = 0;
6561 if (INSN_P (insn))
6563 /* Process notes first so we have all notes in canonical forms
6564 when looking for duplicate operations. */
6565 if (REG_NOTES (insn))
6567 bool changed = false;
6568 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6569 NULL_RTX, &changed);
6570 if (changed)
6571 df_notes_rescan (insn);
6574 cse_insn (insn);
6576 /* If we haven't already found an insn where we added a LABEL_REF,
6577 check this one. */
6578 if (INSN_P (insn) && !recorded_label_ref
6579 && check_for_label_ref (insn))
6580 recorded_label_ref = true;
6582 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6584 /* If the previous insn sets CC0 and this insn no
6585 longer references CC0, delete the previous insn.
6586 Here we use fact that nothing expects CC0 to be
6587 valid over an insn, which is true until the final
6588 pass. */
6589 rtx_insn *prev_insn;
6590 rtx tem;
6592 prev_insn = prev_nonnote_nondebug_insn (insn);
6593 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6594 && (tem = single_set (prev_insn)) != NULL_RTX
6595 && SET_DEST (tem) == cc0_rtx
6596 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6597 delete_insn (prev_insn);
6599 /* If this insn is not the last insn in the basic
6600 block, it will be PREV_INSN(insn) in the next
6601 iteration. If we recorded any CC0-related
6602 information for this insn, remember it. */
6603 if (insn != BB_END (bb))
6605 prev_insn_cc0 = this_insn_cc0;
6606 prev_insn_cc0_mode = this_insn_cc0_mode;
6612 /* With non-call exceptions, we are not always able to update
6613 the CFG properly inside cse_insn. So clean up possibly
6614 redundant EH edges here. */
6615 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6616 cse_cfg_altered |= purge_dead_edges (bb);
6618 /* If we changed a conditional jump, we may have terminated
6619 the path we are following. Check that by verifying that
6620 the edge we would take still exists. If the edge does
6621 not exist anymore, purge the remainder of the path.
6622 Note that this will cause us to return to the caller. */
6623 if (path_entry < path_size - 1)
6625 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6626 if (!find_edge (bb, next_bb))
6630 path_size--;
6632 /* If we truncate the path, we must also reset the
6633 visited bit on the remaining blocks in the path,
6634 or we will never visit them at all. */
6635 bitmap_clear_bit (cse_visited_basic_blocks,
6636 ebb_data->path[path_size].bb->index);
6637 ebb_data->path[path_size].bb = NULL;
6639 while (path_size - 1 != path_entry);
6640 ebb_data->path_size = path_size;
6644 /* If this is a conditional jump insn, record any known
6645 equivalences due to the condition being tested. */
6646 insn = BB_END (bb);
6647 if (path_entry < path_size - 1
6648 && JUMP_P (insn)
6649 && single_set (insn)
6650 && any_condjump_p (insn))
6652 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6653 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6654 record_jump_equiv (insn, taken);
6657 /* Clear the CC0-tracking related insns, they can't provide
6658 useful information across basic block boundaries. */
6659 prev_insn_cc0 = 0;
6662 gcc_assert (next_qty <= max_qty);
6664 free (qty_table);
6668 /* Perform cse on the instructions of a function.
6669 F is the first instruction.
6670 NREGS is one plus the highest pseudo-reg number used in the instruction.
6672 Return 2 if jump optimizations should be redone due to simplifications
6673 in conditional jump instructions.
6674 Return 1 if the CFG should be cleaned up because it has been modified.
6675 Return 0 otherwise. */
6677 static int
6678 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6680 struct cse_basic_block_data ebb_data;
6681 basic_block bb;
6682 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6683 int i, n_blocks;
6685 /* CSE doesn't use dominane info but can invalidate it in different ways.
6686 For simplicity free dominance info here. */
6687 free_dominance_info (CDI_DOMINATORS);
6689 df_set_flags (DF_LR_RUN_DCE);
6690 df_note_add_problem ();
6691 df_analyze ();
6692 df_set_flags (DF_DEFER_INSN_RESCAN);
6694 reg_scan (get_insns (), max_reg_num ());
6695 init_cse_reg_info (nregs);
6697 ebb_data.path = XNEWVEC (struct branch_path,
6698 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6700 cse_cfg_altered = false;
6701 cse_jumps_altered = false;
6702 recorded_label_ref = false;
6703 constant_pool_entries_cost = 0;
6704 constant_pool_entries_regcost = 0;
6705 ebb_data.path_size = 0;
6706 ebb_data.nsets = 0;
6707 rtl_hooks = cse_rtl_hooks;
6709 init_recog ();
6710 init_alias_analysis ();
6712 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6714 /* Set up the table of already visited basic blocks. */
6715 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6716 bitmap_clear (cse_visited_basic_blocks);
6718 /* Loop over basic blocks in reverse completion order (RPO),
6719 excluding the ENTRY and EXIT blocks. */
6720 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6721 i = 0;
6722 while (i < n_blocks)
6724 /* Find the first block in the RPO queue that we have not yet
6725 processed before. */
6728 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6730 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6731 && i < n_blocks);
6733 /* Find all paths starting with BB, and process them. */
6734 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6736 /* Pre-scan the path. */
6737 cse_prescan_path (&ebb_data);
6739 /* If this basic block has no sets, skip it. */
6740 if (ebb_data.nsets == 0)
6741 continue;
6743 /* Get a reasonable estimate for the maximum number of qty's
6744 needed for this path. For this, we take the number of sets
6745 and multiply that by MAX_RECOG_OPERANDS. */
6746 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6748 /* Dump the path we're about to process. */
6749 if (dump_file)
6750 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6752 cse_extended_basic_block (&ebb_data);
6756 /* Clean up. */
6757 end_alias_analysis ();
6758 free (reg_eqv_table);
6759 free (ebb_data.path);
6760 sbitmap_free (cse_visited_basic_blocks);
6761 free (rc_order);
6762 rtl_hooks = general_rtl_hooks;
6764 if (cse_jumps_altered || recorded_label_ref)
6765 return 2;
6766 else if (cse_cfg_altered)
6767 return 1;
6768 else
6769 return 0;
6772 /* Count the number of times registers are used (not set) in X.
6773 COUNTS is an array in which we accumulate the count, INCR is how much
6774 we count each register usage.
6776 Don't count a usage of DEST, which is the SET_DEST of a SET which
6777 contains X in its SET_SRC. This is because such a SET does not
6778 modify the liveness of DEST.
6779 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6780 We must then count uses of a SET_DEST regardless, because the insn can't be
6781 deleted here. */
6783 static void
6784 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6786 enum rtx_code code;
6787 rtx note;
6788 const char *fmt;
6789 int i, j;
6791 if (x == 0)
6792 return;
6794 switch (code = GET_CODE (x))
6796 case REG:
6797 if (x != dest)
6798 counts[REGNO (x)] += incr;
6799 return;
6801 case PC:
6802 case CC0:
6803 case CONST:
6804 CASE_CONST_ANY:
6805 case SYMBOL_REF:
6806 case LABEL_REF:
6807 return;
6809 case CLOBBER:
6810 /* If we are clobbering a MEM, mark any registers inside the address
6811 as being used. */
6812 if (MEM_P (XEXP (x, 0)))
6813 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6814 return;
6816 case SET:
6817 /* Unless we are setting a REG, count everything in SET_DEST. */
6818 if (!REG_P (SET_DEST (x)))
6819 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6820 count_reg_usage (SET_SRC (x), counts,
6821 dest ? dest : SET_DEST (x),
6822 incr);
6823 return;
6825 case DEBUG_INSN:
6826 return;
6828 case CALL_INSN:
6829 case INSN:
6830 case JUMP_INSN:
6831 /* We expect dest to be NULL_RTX here. If the insn may throw,
6832 or if it cannot be deleted due to side-effects, mark this fact
6833 by setting DEST to pc_rtx. */
6834 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6835 || side_effects_p (PATTERN (x)))
6836 dest = pc_rtx;
6837 if (code == CALL_INSN)
6838 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6839 count_reg_usage (PATTERN (x), counts, dest, incr);
6841 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6842 use them. */
6844 note = find_reg_equal_equiv_note (x);
6845 if (note)
6847 rtx eqv = XEXP (note, 0);
6849 if (GET_CODE (eqv) == EXPR_LIST)
6850 /* This REG_EQUAL note describes the result of a function call.
6851 Process all the arguments. */
6854 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6855 eqv = XEXP (eqv, 1);
6857 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6858 else
6859 count_reg_usage (eqv, counts, dest, incr);
6861 return;
6863 case EXPR_LIST:
6864 if (REG_NOTE_KIND (x) == REG_EQUAL
6865 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6866 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6867 involving registers in the address. */
6868 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6869 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6871 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6872 return;
6874 case ASM_OPERANDS:
6875 /* Iterate over just the inputs, not the constraints as well. */
6876 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6877 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6878 return;
6880 case INSN_LIST:
6881 case INT_LIST:
6882 gcc_unreachable ();
6884 default:
6885 break;
6888 fmt = GET_RTX_FORMAT (code);
6889 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6891 if (fmt[i] == 'e')
6892 count_reg_usage (XEXP (x, i), counts, dest, incr);
6893 else if (fmt[i] == 'E')
6894 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6895 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6899 /* Return true if X is a dead register. */
6901 static inline int
6902 is_dead_reg (const_rtx x, int *counts)
6904 return (REG_P (x)
6905 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6906 && counts[REGNO (x)] == 0);
6909 /* Return true if set is live. */
6910 static bool
6911 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6912 int *counts)
6914 rtx_insn *tem;
6916 if (set_noop_p (set))
6919 else if (GET_CODE (SET_DEST (set)) == CC0
6920 && !side_effects_p (SET_SRC (set))
6921 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6922 || !INSN_P (tem)
6923 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6924 return false;
6925 else if (!is_dead_reg (SET_DEST (set), counts)
6926 || side_effects_p (SET_SRC (set)))
6927 return true;
6928 return false;
6931 /* Return true if insn is live. */
6933 static bool
6934 insn_live_p (rtx_insn *insn, int *counts)
6936 int i;
6937 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6938 return true;
6939 else if (GET_CODE (PATTERN (insn)) == SET)
6940 return set_live_p (PATTERN (insn), insn, counts);
6941 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6943 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6945 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6947 if (GET_CODE (elt) == SET)
6949 if (set_live_p (elt, insn, counts))
6950 return true;
6952 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6953 return true;
6955 return false;
6957 else if (DEBUG_INSN_P (insn))
6959 rtx_insn *next;
6961 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6962 if (NOTE_P (next))
6963 continue;
6964 else if (!DEBUG_INSN_P (next))
6965 return true;
6966 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6967 return false;
6969 return true;
6971 else
6972 return true;
6975 /* Count the number of stores into pseudo. Callback for note_stores. */
6977 static void
6978 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6980 int *counts = (int *) data;
6981 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6982 counts[REGNO (x)]++;
6985 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6986 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6987 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6988 Set *SEEN_REPL to true if we see a dead register that does have
6989 a replacement. */
6991 static bool
6992 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6993 bool *seen_repl)
6995 subrtx_iterator::array_type array;
6996 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6998 const_rtx x = *iter;
6999 if (is_dead_reg (x, counts))
7001 if (replacements && replacements[REGNO (x)] != NULL_RTX)
7002 *seen_repl = true;
7003 else
7004 return true;
7007 return false;
7010 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
7011 Callback for simplify_replace_fn_rtx. */
7013 static rtx
7014 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
7016 rtx *replacements = (rtx *) data;
7018 if (REG_P (x)
7019 && REGNO (x) >= FIRST_PSEUDO_REGISTER
7020 && replacements[REGNO (x)] != NULL_RTX)
7022 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
7023 return replacements[REGNO (x)];
7024 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
7025 GET_MODE (replacements[REGNO (x)]));
7027 return NULL_RTX;
7030 /* Scan all the insns and delete any that are dead; i.e., they store a register
7031 that is never used or they copy a register to itself.
7033 This is used to remove insns made obviously dead by cse, loop or other
7034 optimizations. It improves the heuristics in loop since it won't try to
7035 move dead invariants out of loops or make givs for dead quantities. The
7036 remaining passes of the compilation are also sped up. */
7039 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
7041 int *counts;
7042 rtx_insn *insn, *prev;
7043 rtx *replacements = NULL;
7044 int ndead = 0;
7046 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7047 /* First count the number of times each register is used. */
7048 if (MAY_HAVE_DEBUG_INSNS)
7050 counts = XCNEWVEC (int, nreg * 3);
7051 for (insn = insns; insn; insn = NEXT_INSN (insn))
7052 if (DEBUG_INSN_P (insn))
7053 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7054 NULL_RTX, 1);
7055 else if (INSN_P (insn))
7057 count_reg_usage (insn, counts, NULL_RTX, 1);
7058 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7060 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7061 First one counts how many times each pseudo is used outside
7062 of debug insns, second counts how many times each pseudo is
7063 used in debug insns and third counts how many times a pseudo
7064 is stored. */
7066 else
7068 counts = XCNEWVEC (int, nreg);
7069 for (insn = insns; insn; insn = NEXT_INSN (insn))
7070 if (INSN_P (insn))
7071 count_reg_usage (insn, counts, NULL_RTX, 1);
7072 /* If no debug insns can be present, COUNTS is just an array
7073 which counts how many times each pseudo is used. */
7075 /* Pseudo PIC register should be considered as used due to possible
7076 new usages generated. */
7077 if (!reload_completed
7078 && pic_offset_table_rtx
7079 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7080 counts[REGNO (pic_offset_table_rtx)]++;
7081 /* Go from the last insn to the first and delete insns that only set unused
7082 registers or copy a register to itself. As we delete an insn, remove
7083 usage counts for registers it uses.
7085 The first jump optimization pass may leave a real insn as the last
7086 insn in the function. We must not skip that insn or we may end
7087 up deleting code that is not really dead.
7089 If some otherwise unused register is only used in DEBUG_INSNs,
7090 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7091 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7092 has been created for the unused register, replace it with
7093 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7094 for (insn = get_last_insn (); insn; insn = prev)
7096 int live_insn = 0;
7098 prev = PREV_INSN (insn);
7099 if (!INSN_P (insn))
7100 continue;
7102 live_insn = insn_live_p (insn, counts);
7104 /* If this is a dead insn, delete it and show registers in it aren't
7105 being used. */
7107 if (! live_insn && dbg_cnt (delete_trivial_dead))
7109 if (DEBUG_INSN_P (insn))
7110 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7111 NULL_RTX, -1);
7112 else
7114 rtx set;
7115 if (MAY_HAVE_DEBUG_INSNS
7116 && (set = single_set (insn)) != NULL_RTX
7117 && is_dead_reg (SET_DEST (set), counts)
7118 /* Used at least once in some DEBUG_INSN. */
7119 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7120 /* And set exactly once. */
7121 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7122 && !side_effects_p (SET_SRC (set))
7123 && asm_noperands (PATTERN (insn)) < 0)
7125 rtx dval, bind_var_loc;
7126 rtx_insn *bind;
7128 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7129 dval = make_debug_expr_from_rtl (SET_DEST (set));
7131 /* Emit a debug bind insn before the insn in which
7132 reg dies. */
7133 bind_var_loc =
7134 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7135 DEBUG_EXPR_TREE_DECL (dval),
7136 SET_SRC (set),
7137 VAR_INIT_STATUS_INITIALIZED);
7138 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7140 bind = emit_debug_insn_before (bind_var_loc, insn);
7141 df_insn_rescan (bind);
7143 if (replacements == NULL)
7144 replacements = XCNEWVEC (rtx, nreg);
7145 replacements[REGNO (SET_DEST (set))] = dval;
7148 count_reg_usage (insn, counts, NULL_RTX, -1);
7149 ndead++;
7151 cse_cfg_altered |= delete_insn_and_edges (insn);
7155 if (MAY_HAVE_DEBUG_INSNS)
7157 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7158 if (DEBUG_INSN_P (insn))
7160 /* If this debug insn references a dead register that wasn't replaced
7161 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7162 bool seen_repl = false;
7163 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7164 counts, replacements, &seen_repl))
7166 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7167 df_insn_rescan (insn);
7169 else if (seen_repl)
7171 INSN_VAR_LOCATION_LOC (insn)
7172 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7173 NULL_RTX, replace_dead_reg,
7174 replacements);
7175 df_insn_rescan (insn);
7178 free (replacements);
7181 if (dump_file && ndead)
7182 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7183 ndead);
7184 /* Clean up. */
7185 free (counts);
7186 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7187 return ndead;
7190 /* If LOC contains references to NEWREG in a different mode, change them
7191 to use NEWREG instead. */
7193 static void
7194 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7195 rtx *loc, rtx_insn *insn, rtx newreg)
7197 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7199 rtx *loc = *iter;
7200 rtx x = *loc;
7201 if (x
7202 && REG_P (x)
7203 && REGNO (x) == REGNO (newreg)
7204 && GET_MODE (x) != GET_MODE (newreg))
7206 validate_change (insn, loc, newreg, 1);
7207 iter.skip_subrtxes ();
7212 /* Change the mode of any reference to the register REGNO (NEWREG) to
7213 GET_MODE (NEWREG) in INSN. */
7215 static void
7216 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7218 int success;
7220 if (!INSN_P (insn))
7221 return;
7223 subrtx_ptr_iterator::array_type array;
7224 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7225 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7227 /* If the following assertion was triggered, there is most probably
7228 something wrong with the cc_modes_compatible back end function.
7229 CC modes only can be considered compatible if the insn - with the mode
7230 replaced by any of the compatible modes - can still be recognized. */
7231 success = apply_change_group ();
7232 gcc_assert (success);
7235 /* Change the mode of any reference to the register REGNO (NEWREG) to
7236 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7237 any instruction which modifies NEWREG. */
7239 static void
7240 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7242 rtx_insn *insn;
7244 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7246 if (! INSN_P (insn))
7247 continue;
7249 if (reg_set_p (newreg, insn))
7250 return;
7252 cse_change_cc_mode_insn (insn, newreg);
7256 /* BB is a basic block which finishes with CC_REG as a condition code
7257 register which is set to CC_SRC. Look through the successors of BB
7258 to find blocks which have a single predecessor (i.e., this one),
7259 and look through those blocks for an assignment to CC_REG which is
7260 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7261 permitted to change the mode of CC_SRC to a compatible mode. This
7262 returns VOIDmode if no equivalent assignments were found.
7263 Otherwise it returns the mode which CC_SRC should wind up with.
7264 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7265 but is passed unmodified down to recursive calls in order to prevent
7266 endless recursion.
7268 The main complexity in this function is handling the mode issues.
7269 We may have more than one duplicate which we can eliminate, and we
7270 try to find a mode which will work for multiple duplicates. */
7272 static machine_mode
7273 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7274 bool can_change_mode)
7276 bool found_equiv;
7277 machine_mode mode;
7278 unsigned int insn_count;
7279 edge e;
7280 rtx_insn *insns[2];
7281 machine_mode modes[2];
7282 rtx_insn *last_insns[2];
7283 unsigned int i;
7284 rtx newreg;
7285 edge_iterator ei;
7287 /* We expect to have two successors. Look at both before picking
7288 the final mode for the comparison. If we have more successors
7289 (i.e., some sort of table jump, although that seems unlikely),
7290 then we require all beyond the first two to use the same
7291 mode. */
7293 found_equiv = false;
7294 mode = GET_MODE (cc_src);
7295 insn_count = 0;
7296 FOR_EACH_EDGE (e, ei, bb->succs)
7298 rtx_insn *insn;
7299 rtx_insn *end;
7301 if (e->flags & EDGE_COMPLEX)
7302 continue;
7304 if (EDGE_COUNT (e->dest->preds) != 1
7305 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7306 /* Avoid endless recursion on unreachable blocks. */
7307 || e->dest == orig_bb)
7308 continue;
7310 end = NEXT_INSN (BB_END (e->dest));
7311 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7313 rtx set;
7315 if (! INSN_P (insn))
7316 continue;
7318 /* If CC_SRC is modified, we have to stop looking for
7319 something which uses it. */
7320 if (modified_in_p (cc_src, insn))
7321 break;
7323 /* Check whether INSN sets CC_REG to CC_SRC. */
7324 set = single_set (insn);
7325 if (set
7326 && REG_P (SET_DEST (set))
7327 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7329 bool found;
7330 machine_mode set_mode;
7331 machine_mode comp_mode;
7333 found = false;
7334 set_mode = GET_MODE (SET_SRC (set));
7335 comp_mode = set_mode;
7336 if (rtx_equal_p (cc_src, SET_SRC (set)))
7337 found = true;
7338 else if (GET_CODE (cc_src) == COMPARE
7339 && GET_CODE (SET_SRC (set)) == COMPARE
7340 && mode != set_mode
7341 && rtx_equal_p (XEXP (cc_src, 0),
7342 XEXP (SET_SRC (set), 0))
7343 && rtx_equal_p (XEXP (cc_src, 1),
7344 XEXP (SET_SRC (set), 1)))
7347 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7348 if (comp_mode != VOIDmode
7349 && (can_change_mode || comp_mode == mode))
7350 found = true;
7353 if (found)
7355 found_equiv = true;
7356 if (insn_count < ARRAY_SIZE (insns))
7358 insns[insn_count] = insn;
7359 modes[insn_count] = set_mode;
7360 last_insns[insn_count] = end;
7361 ++insn_count;
7363 if (mode != comp_mode)
7365 gcc_assert (can_change_mode);
7366 mode = comp_mode;
7368 /* The modified insn will be re-recognized later. */
7369 PUT_MODE (cc_src, mode);
7372 else
7374 if (set_mode != mode)
7376 /* We found a matching expression in the
7377 wrong mode, but we don't have room to
7378 store it in the array. Punt. This case
7379 should be rare. */
7380 break;
7382 /* INSN sets CC_REG to a value equal to CC_SRC
7383 with the right mode. We can simply delete
7384 it. */
7385 delete_insn (insn);
7388 /* We found an instruction to delete. Keep looking,
7389 in the hopes of finding a three-way jump. */
7390 continue;
7393 /* We found an instruction which sets the condition
7394 code, so don't look any farther. */
7395 break;
7398 /* If INSN sets CC_REG in some other way, don't look any
7399 farther. */
7400 if (reg_set_p (cc_reg, insn))
7401 break;
7404 /* If we fell off the bottom of the block, we can keep looking
7405 through successors. We pass CAN_CHANGE_MODE as false because
7406 we aren't prepared to handle compatibility between the
7407 further blocks and this block. */
7408 if (insn == end)
7410 machine_mode submode;
7412 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7413 if (submode != VOIDmode)
7415 gcc_assert (submode == mode);
7416 found_equiv = true;
7417 can_change_mode = false;
7422 if (! found_equiv)
7423 return VOIDmode;
7425 /* Now INSN_COUNT is the number of instructions we found which set
7426 CC_REG to a value equivalent to CC_SRC. The instructions are in
7427 INSNS. The modes used by those instructions are in MODES. */
7429 newreg = NULL_RTX;
7430 for (i = 0; i < insn_count; ++i)
7432 if (modes[i] != mode)
7434 /* We need to change the mode of CC_REG in INSNS[i] and
7435 subsequent instructions. */
7436 if (! newreg)
7438 if (GET_MODE (cc_reg) == mode)
7439 newreg = cc_reg;
7440 else
7441 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7443 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7444 newreg);
7447 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7450 return mode;
7453 /* If we have a fixed condition code register (or two), walk through
7454 the instructions and try to eliminate duplicate assignments. */
7456 static void
7457 cse_condition_code_reg (void)
7459 unsigned int cc_regno_1;
7460 unsigned int cc_regno_2;
7461 rtx cc_reg_1;
7462 rtx cc_reg_2;
7463 basic_block bb;
7465 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7466 return;
7468 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7469 if (cc_regno_2 != INVALID_REGNUM)
7470 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7471 else
7472 cc_reg_2 = NULL_RTX;
7474 FOR_EACH_BB_FN (bb, cfun)
7476 rtx_insn *last_insn;
7477 rtx cc_reg;
7478 rtx_insn *insn;
7479 rtx_insn *cc_src_insn;
7480 rtx cc_src;
7481 machine_mode mode;
7482 machine_mode orig_mode;
7484 /* Look for blocks which end with a conditional jump based on a
7485 condition code register. Then look for the instruction which
7486 sets the condition code register. Then look through the
7487 successor blocks for instructions which set the condition
7488 code register to the same value. There are other possible
7489 uses of the condition code register, but these are by far the
7490 most common and the ones which we are most likely to be able
7491 to optimize. */
7493 last_insn = BB_END (bb);
7494 if (!JUMP_P (last_insn))
7495 continue;
7497 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7498 cc_reg = cc_reg_1;
7499 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7500 cc_reg = cc_reg_2;
7501 else
7502 continue;
7504 cc_src_insn = NULL;
7505 cc_src = NULL_RTX;
7506 for (insn = PREV_INSN (last_insn);
7507 insn && insn != PREV_INSN (BB_HEAD (bb));
7508 insn = PREV_INSN (insn))
7510 rtx set;
7512 if (! INSN_P (insn))
7513 continue;
7514 set = single_set (insn);
7515 if (set
7516 && REG_P (SET_DEST (set))
7517 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7519 cc_src_insn = insn;
7520 cc_src = SET_SRC (set);
7521 break;
7523 else if (reg_set_p (cc_reg, insn))
7524 break;
7527 if (! cc_src_insn)
7528 continue;
7530 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7531 continue;
7533 /* Now CC_REG is a condition code register used for a
7534 conditional jump at the end of the block, and CC_SRC, in
7535 CC_SRC_INSN, is the value to which that condition code
7536 register is set, and CC_SRC is still meaningful at the end of
7537 the basic block. */
7539 orig_mode = GET_MODE (cc_src);
7540 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7541 if (mode != VOIDmode)
7543 gcc_assert (mode == GET_MODE (cc_src));
7544 if (mode != orig_mode)
7546 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7548 cse_change_cc_mode_insn (cc_src_insn, newreg);
7550 /* Do the same in the following insns that use the
7551 current value of CC_REG within BB. */
7552 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7553 NEXT_INSN (last_insn),
7554 newreg);
7561 /* Perform common subexpression elimination. Nonzero value from
7562 `cse_main' means that jumps were simplified and some code may now
7563 be unreachable, so do jump optimization again. */
7564 static unsigned int
7565 rest_of_handle_cse (void)
7567 int tem;
7569 if (dump_file)
7570 dump_flow_info (dump_file, dump_flags);
7572 tem = cse_main (get_insns (), max_reg_num ());
7574 /* If we are not running more CSE passes, then we are no longer
7575 expecting CSE to be run. But always rerun it in a cheap mode. */
7576 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7578 if (tem == 2)
7580 timevar_push (TV_JUMP);
7581 rebuild_jump_labels (get_insns ());
7582 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7583 timevar_pop (TV_JUMP);
7585 else if (tem == 1 || optimize > 1)
7586 cse_cfg_altered |= cleanup_cfg (0);
7588 return 0;
7591 namespace {
7593 const pass_data pass_data_cse =
7595 RTL_PASS, /* type */
7596 "cse1", /* name */
7597 OPTGROUP_NONE, /* optinfo_flags */
7598 TV_CSE, /* tv_id */
7599 0, /* properties_required */
7600 0, /* properties_provided */
7601 0, /* properties_destroyed */
7602 0, /* todo_flags_start */
7603 TODO_df_finish, /* todo_flags_finish */
7606 class pass_cse : public rtl_opt_pass
7608 public:
7609 pass_cse (gcc::context *ctxt)
7610 : rtl_opt_pass (pass_data_cse, ctxt)
7613 /* opt_pass methods: */
7614 virtual bool gate (function *) { return optimize > 0; }
7615 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7617 }; // class pass_cse
7619 } // anon namespace
7621 rtl_opt_pass *
7622 make_pass_cse (gcc::context *ctxt)
7624 return new pass_cse (ctxt);
7628 /* Run second CSE pass after loop optimizations. */
7629 static unsigned int
7630 rest_of_handle_cse2 (void)
7632 int tem;
7634 if (dump_file)
7635 dump_flow_info (dump_file, dump_flags);
7637 tem = cse_main (get_insns (), max_reg_num ());
7639 /* Run a pass to eliminate duplicated assignments to condition code
7640 registers. We have to run this after bypass_jumps, because it
7641 makes it harder for that pass to determine whether a jump can be
7642 bypassed safely. */
7643 cse_condition_code_reg ();
7645 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7647 if (tem == 2)
7649 timevar_push (TV_JUMP);
7650 rebuild_jump_labels (get_insns ());
7651 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7652 timevar_pop (TV_JUMP);
7654 else if (tem == 1)
7655 cse_cfg_altered |= cleanup_cfg (0);
7657 cse_not_expected = 1;
7658 return 0;
7662 namespace {
7664 const pass_data pass_data_cse2 =
7666 RTL_PASS, /* type */
7667 "cse2", /* name */
7668 OPTGROUP_NONE, /* optinfo_flags */
7669 TV_CSE2, /* tv_id */
7670 0, /* properties_required */
7671 0, /* properties_provided */
7672 0, /* properties_destroyed */
7673 0, /* todo_flags_start */
7674 TODO_df_finish, /* todo_flags_finish */
7677 class pass_cse2 : public rtl_opt_pass
7679 public:
7680 pass_cse2 (gcc::context *ctxt)
7681 : rtl_opt_pass (pass_data_cse2, ctxt)
7684 /* opt_pass methods: */
7685 virtual bool gate (function *)
7687 return optimize > 0 && flag_rerun_cse_after_loop;
7690 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7692 }; // class pass_cse2
7694 } // anon namespace
7696 rtl_opt_pass *
7697 make_pass_cse2 (gcc::context *ctxt)
7699 return new pass_cse2 (ctxt);
7702 /* Run second CSE pass after loop optimizations. */
7703 static unsigned int
7704 rest_of_handle_cse_after_global_opts (void)
7706 int save_cfj;
7707 int tem;
7709 /* We only want to do local CSE, so don't follow jumps. */
7710 save_cfj = flag_cse_follow_jumps;
7711 flag_cse_follow_jumps = 0;
7713 rebuild_jump_labels (get_insns ());
7714 tem = cse_main (get_insns (), max_reg_num ());
7715 cse_cfg_altered |= purge_all_dead_edges ();
7716 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7718 cse_not_expected = !flag_rerun_cse_after_loop;
7720 /* If cse altered any jumps, rerun jump opts to clean things up. */
7721 if (tem == 2)
7723 timevar_push (TV_JUMP);
7724 rebuild_jump_labels (get_insns ());
7725 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7726 timevar_pop (TV_JUMP);
7728 else if (tem == 1)
7729 cse_cfg_altered |= cleanup_cfg (0);
7731 flag_cse_follow_jumps = save_cfj;
7732 return 0;
7735 namespace {
7737 const pass_data pass_data_cse_after_global_opts =
7739 RTL_PASS, /* type */
7740 "cse_local", /* name */
7741 OPTGROUP_NONE, /* optinfo_flags */
7742 TV_CSE, /* tv_id */
7743 0, /* properties_required */
7744 0, /* properties_provided */
7745 0, /* properties_destroyed */
7746 0, /* todo_flags_start */
7747 TODO_df_finish, /* todo_flags_finish */
7750 class pass_cse_after_global_opts : public rtl_opt_pass
7752 public:
7753 pass_cse_after_global_opts (gcc::context *ctxt)
7754 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7757 /* opt_pass methods: */
7758 virtual bool gate (function *)
7760 return optimize > 0 && flag_rerun_cse_after_global_opts;
7763 virtual unsigned int execute (function *)
7765 return rest_of_handle_cse_after_global_opts ();
7768 }; // class pass_cse_after_global_opts
7770 } // anon namespace
7772 rtl_opt_pass *
7773 make_pass_cse_after_global_opts (gcc::context *ctxt)
7775 return new pass_cse_after_global_opts (ctxt);