Remove assert in get_def_bb_for_const
[official-gcc.git] / gcc / cse.c
blobbce4fb0262e602bc14f8ad77e52224ed30872f98
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 "tm_p.h"
30 #include "insn-config.h"
31 #include "regs.h"
32 #include "emit-rtl.h"
33 #include "recog.h"
34 #include "cfgrtl.h"
35 #include "cfganal.h"
36 #include "cfgcleanup.h"
37 #include "alias.h"
38 #include "toplev.h"
39 #include "params.h"
40 #include "rtlhooks-def.h"
41 #include "tree-pass.h"
42 #include "dbgcnt.h"
43 #include "rtl-iter.h"
45 #ifndef LOAD_EXTEND_OP
46 #define LOAD_EXTEND_OP(M) UNKNOWN
47 #endif
49 /* The basic idea of common subexpression elimination is to go
50 through the code, keeping a record of expressions that would
51 have the same value at the current scan point, and replacing
52 expressions encountered with the cheapest equivalent expression.
54 It is too complicated to keep track of the different possibilities
55 when control paths merge in this code; so, at each label, we forget all
56 that is known and start fresh. This can be described as processing each
57 extended basic block separately. We have a separate pass to perform
58 global CSE.
60 Note CSE can turn a conditional or computed jump into a nop or
61 an unconditional jump. When this occurs we arrange to run the jump
62 optimizer after CSE to delete the unreachable code.
64 We use two data structures to record the equivalent expressions:
65 a hash table for most expressions, and a vector of "quantity
66 numbers" to record equivalent (pseudo) registers.
68 The use of the special data structure for registers is desirable
69 because it is faster. It is possible because registers references
70 contain a fairly small number, the register number, taken from
71 a contiguously allocated series, and two register references are
72 identical if they have the same number. General expressions
73 do not have any such thing, so the only way to retrieve the
74 information recorded on an expression other than a register
75 is to keep it in a hash table.
77 Registers and "quantity numbers":
79 At the start of each basic block, all of the (hardware and pseudo)
80 registers used in the function are given distinct quantity
81 numbers to indicate their contents. During scan, when the code
82 copies one register into another, we copy the quantity number.
83 When a register is loaded in any other way, we allocate a new
84 quantity number to describe the value generated by this operation.
85 `REG_QTY (N)' records what quantity register N is currently thought
86 of as containing.
88 All real quantity numbers are greater than or equal to zero.
89 If register N has not been assigned a quantity, `REG_QTY (N)' will
90 equal -N - 1, which is always negative.
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
95 We also maintain a bidirectional chain of registers for each
96 quantity number. The `qty_table` members `first_reg' and `last_reg',
97 and `reg_eqv_table' members `next' and `prev' hold these chains.
99 The first register in a chain is the one whose lifespan is least local.
100 Among equals, it is the one that was seen first.
101 We replace any equivalent register with that one.
103 If two registers have the same quantity number, it must be true that
104 REG expressions with qty_table `mode' must be in the hash table for both
105 registers and must be in the same class.
107 The converse is not true. Since hard registers may be referenced in
108 any mode, two REG expressions might be equivalent in the hash table
109 but not have the same quantity number if the quantity number of one
110 of the registers is not the same mode as those expressions.
112 Constants and quantity numbers
114 When a quantity has a known constant value, that value is stored
115 in the appropriate qty_table `const_rtx'. This is in addition to
116 putting the constant in the hash table as is usual for non-regs.
118 Whether a reg or a constant is preferred is determined by the configuration
119 macro CONST_COSTS and will often depend on the constant value. In any
120 event, expressions containing constants can be simplified, by fold_rtx.
122 When a quantity has a known nearly constant value (such as an address
123 of a stack slot), that value is stored in the appropriate qty_table
124 `const_rtx'.
126 Integer constants don't have a machine mode. However, cse
127 determines the intended machine mode from the destination
128 of the instruction that moves the constant. The machine mode
129 is recorded in the hash table along with the actual RTL
130 constant expression so that different modes are kept separate.
132 Other expressions:
134 To record known equivalences among expressions in general
135 we use a hash table called `table'. It has a fixed number of buckets
136 that contain chains of `struct table_elt' elements for expressions.
137 These chains connect the elements whose expressions have the same
138 hash codes.
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
143 Register references in an expression are canonicalized before hashing
144 the expression. This is done using `reg_qty' and qty_table `first_reg'.
145 The hash code of a register reference is computed using the quantity
146 number, not the register number.
148 When the value of an expression changes, it is necessary to remove from the
149 hash table not just that expression but all expressions whose values
150 could be different as a result.
152 1. If the value changing is in memory, except in special cases
153 ANYTHING referring to memory could be changed. That is because
154 nobody knows where a pointer does not point.
155 The function `invalidate_memory' removes what is necessary.
157 The special cases are when the address is constant or is
158 a constant plus a fixed register such as the frame pointer
159 or a static chain pointer. When such addresses are stored in,
160 we can tell exactly which other such addresses must be invalidated
161 due to overlap. `invalidate' does this.
162 All expressions that refer to non-constant
163 memory addresses are also invalidated. `invalidate_memory' does this.
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
169 Because searching the entire hash table for expressions that contain
170 a register is very slow, we try to figure out when it isn't necessary.
171 Precisely, this is necessary only when expressions have been
172 entered in the hash table using this register, and then the value has
173 changed, and then another expression wants to be added to refer to
174 the register's new value. This sequence of circumstances is rare
175 within any one basic block.
177 `REG_TICK' and `REG_IN_TABLE', accessors for members of
178 cse_reg_info, are used to detect this case. REG_TICK (i) is
179 incremented whenever a value is stored in register i.
180 REG_IN_TABLE (i) holds -1 if no references to register i have been
181 entered in the table; otherwise, it contains the value REG_TICK (i)
182 had when the references were entered. If we want to enter a
183 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
184 remove old references. Until we want to enter a new entry, the
185 mere fact that the two vectors don't match makes the entries be
186 ignored if anyone tries to match them.
188 Registers themselves are entered in the hash table as well as in
189 the equivalent-register chains. However, `REG_TICK' and
190 `REG_IN_TABLE' do not apply to expressions which are simple
191 register references. These expressions are removed from the table
192 immediately when they become invalid, and this can be done even if
193 we do not immediately search for all the expressions that refer to
194 the register.
196 A CLOBBER rtx in an instruction invalidates its operand for further
197 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
198 invalidates everything that resides in memory.
200 Related expressions:
202 Constant expressions that differ only by an additive integer
203 are called related. When a constant expression is put in
204 the table, the related expression with no constant term
205 is also entered. These are made to point at each other
206 so that it is possible to find out if there exists any
207 register equivalent to an expression related to a given expression. */
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
212 static int max_qty;
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
217 static int next_qty;
219 /* Per-qty information tracking.
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
224 `mode' contains the machine mode of this quantity.
226 `const_rtx' holds the rtx of the constant value of this
227 quantity, if known. A summations of the frame/arg pointer
228 and a constant can also be entered here. When this holds
229 a known value, `const_insn' is the insn which stored the
230 constant value.
232 `comparison_{code,const,qty}' are used to track when a
233 comparison between a quantity and some constant or register has
234 been passed. In such a case, we know the results of the comparison
235 in case we see it again. These members record a comparison that
236 is known to be true. `comparison_code' holds the rtx code of such
237 a comparison, else it is set to UNKNOWN and the other two
238 comparison members are undefined. `comparison_const' holds
239 the constant being compared against, or zero if the comparison
240 is not against a constant. `comparison_qty' holds the quantity
241 being compared against when the result is known. If the comparison
242 is not with a register, `comparison_qty' is -1. */
244 struct qty_table_elem
246 rtx const_rtx;
247 rtx_insn *const_insn;
248 rtx comparison_const;
249 int comparison_qty;
250 unsigned int first_reg, last_reg;
251 /* The sizes of these fields should match the sizes of the
252 code and mode fields of struct rtx_def (see rtl.h). */
253 ENUM_BITFIELD(rtx_code) comparison_code : 16;
254 ENUM_BITFIELD(machine_mode) mode : 8;
257 /* The table of all qtys, indexed by qty number. */
258 static struct qty_table_elem *qty_table;
260 /* For machines that have a CC0, we do not record its value in the hash
261 table since its use is guaranteed to be the insn immediately following
262 its definition and any other insn is presumed to invalidate it.
264 Instead, we store below the current and last value assigned to CC0.
265 If it should happen to be a constant, it is stored in preference
266 to the actual assigned value. In case it is a constant, we store
267 the mode in which the constant should be interpreted. */
269 static rtx this_insn_cc0, prev_insn_cc0;
270 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
272 /* Insn being scanned. */
274 static rtx_insn *this_insn;
275 static bool optimize_this_for_speed_p;
277 /* Index by register number, gives the number of the next (or
278 previous) register in the chain of registers sharing the same
279 value.
281 Or -1 if this register is at the end of the chain.
283 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
285 /* Per-register equivalence chain. */
286 struct reg_eqv_elem
288 int next, prev;
291 /* The table of all register equivalence chains. */
292 static struct reg_eqv_elem *reg_eqv_table;
294 struct cse_reg_info
296 /* The timestamp at which this register is initialized. */
297 unsigned int timestamp;
299 /* The quantity number of the register's current contents. */
300 int reg_qty;
302 /* The number of times the register has been altered in the current
303 basic block. */
304 int reg_tick;
306 /* The REG_TICK value at which rtx's containing this register are
307 valid in the hash table. If this does not equal the current
308 reg_tick value, such expressions existing in the hash table are
309 invalid. */
310 int reg_in_table;
312 /* The SUBREG that was set when REG_TICK was last incremented. Set
313 to -1 if the last store was to the whole register, not a subreg. */
314 unsigned int subreg_ticked;
317 /* A table of cse_reg_info indexed by register numbers. */
318 static struct cse_reg_info *cse_reg_info_table;
320 /* The size of the above table. */
321 static unsigned int cse_reg_info_table_size;
323 /* The index of the first entry that has not been initialized. */
324 static unsigned int cse_reg_info_table_first_uninitialized;
326 /* The timestamp at the beginning of the current run of
327 cse_extended_basic_block. We increment this variable at the beginning of
328 the current run of cse_extended_basic_block. The timestamp field of a
329 cse_reg_info entry matches the value of this variable if and only
330 if the entry has been initialized during the current run of
331 cse_extended_basic_block. */
332 static unsigned int cse_reg_info_timestamp;
334 /* A HARD_REG_SET containing all the hard registers for which there is
335 currently a REG expression in the hash table. Note the difference
336 from the above variables, which indicate if the REG is mentioned in some
337 expression in the table. */
339 static HARD_REG_SET hard_regs_in_table;
341 /* True if CSE has altered the CFG. */
342 static bool cse_cfg_altered;
344 /* True if CSE has altered conditional jump insns in such a way
345 that jump optimization should be redone. */
346 static bool cse_jumps_altered;
348 /* True if we put a LABEL_REF into the hash table for an INSN
349 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
350 to put in the note. */
351 static bool recorded_label_ref;
353 /* canon_hash stores 1 in do_not_record
354 if it notices a reference to CC0, PC, or some other volatile
355 subexpression. */
357 static int do_not_record;
359 /* canon_hash stores 1 in hash_arg_in_memory
360 if it notices a reference to memory within the expression being hashed. */
362 static int hash_arg_in_memory;
364 /* The hash table contains buckets which are chains of `struct table_elt's,
365 each recording one expression's information.
366 That expression is in the `exp' field.
368 The canon_exp field contains a canonical (from the point of view of
369 alias analysis) version of the `exp' field.
371 Those elements with the same hash code are chained in both directions
372 through the `next_same_hash' and `prev_same_hash' fields.
374 Each set of expressions with equivalent values
375 are on a two-way chain through the `next_same_value'
376 and `prev_same_value' fields, and all point with
377 the `first_same_value' field at the first element in
378 that chain. The chain is in order of increasing cost.
379 Each element's cost value is in its `cost' field.
381 The `in_memory' field is nonzero for elements that
382 involve any reference to memory. These elements are removed
383 whenever a write is done to an unidentified location in memory.
384 To be safe, we assume that a memory address is unidentified unless
385 the address is either a symbol constant or a constant plus
386 the frame pointer or argument pointer.
388 The `related_value' field is used to connect related expressions
389 (that differ by adding an integer).
390 The related expressions are chained in a circular fashion.
391 `related_value' is zero for expressions for which this
392 chain is not useful.
394 The `cost' field stores the cost of this element's expression.
395 The `regcost' field stores the value returned by approx_reg_cost for
396 this element's expression.
398 The `is_const' flag is set if the element is a constant (including
399 a fixed address).
401 The `flag' field is used as a temporary during some search routines.
403 The `mode' field is usually the same as GET_MODE (`exp'), but
404 if `exp' is a CONST_INT and has no machine mode then the `mode'
405 field is the mode it was being used as. Each constant is
406 recorded separately for each mode it is used with. */
408 struct table_elt
410 rtx exp;
411 rtx canon_exp;
412 struct table_elt *next_same_hash;
413 struct table_elt *prev_same_hash;
414 struct table_elt *next_same_value;
415 struct table_elt *prev_same_value;
416 struct table_elt *first_same_value;
417 struct table_elt *related_value;
418 int cost;
419 int regcost;
420 /* The size of this field should match the size
421 of the mode field of struct rtx_def (see rtl.h). */
422 ENUM_BITFIELD(machine_mode) mode : 8;
423 char in_memory;
424 char is_const;
425 char flag;
428 /* We don't want a lot of buckets, because we rarely have very many
429 things stored in the hash table, and a lot of buckets slows
430 down a lot of loops that happen frequently. */
431 #define HASH_SHIFT 5
432 #define HASH_SIZE (1 << HASH_SHIFT)
433 #define HASH_MASK (HASH_SIZE - 1)
435 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
436 register (hard registers may require `do_not_record' to be set). */
438 #define HASH(X, M) \
439 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
440 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
441 : canon_hash (X, M)) & HASH_MASK)
443 /* Like HASH, but without side-effects. */
444 #define SAFE_HASH(X, M) \
445 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
446 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
447 : safe_hash (X, M)) & HASH_MASK)
449 /* Determine whether register number N is considered a fixed register for the
450 purpose of approximating register costs.
451 It is desirable to replace other regs with fixed regs, to reduce need for
452 non-fixed hard regs.
453 A reg wins if it is either the frame pointer or designated as fixed. */
454 #define FIXED_REGNO_P(N) \
455 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
456 || fixed_regs[N] || global_regs[N])
458 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
459 hard registers and pointers into the frame are the cheapest with a cost
460 of 0. Next come pseudos with a cost of one and other hard registers with
461 a cost of 2. Aside from these special cases, call `rtx_cost'. */
463 #define CHEAP_REGNO(N) \
464 (REGNO_PTR_FRAME_P (N) \
465 || (HARD_REGISTER_NUM_P (N) \
466 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
468 #define COST(X, MODE) \
469 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
470 #define COST_IN(X, MODE, OUTER, OPNO) \
471 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
473 /* Get the number of times this register has been updated in this
474 basic block. */
476 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
478 /* Get the point at which REG was recorded in the table. */
480 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
482 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
483 SUBREG). */
485 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
487 /* Get the quantity number for REG. */
489 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
491 /* Determine if the quantity number for register X represents a valid index
492 into the qty_table. */
494 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
496 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
498 #define CHEAPER(X, Y) \
499 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
501 static struct table_elt *table[HASH_SIZE];
503 /* Chain of `struct table_elt's made so far for this function
504 but currently removed from the table. */
506 static struct table_elt *free_element_chain;
508 /* Set to the cost of a constant pool reference if one was found for a
509 symbolic constant. If this was found, it means we should try to
510 convert constants into constant pool entries if they don't fit in
511 the insn. */
513 static int constant_pool_entries_cost;
514 static int constant_pool_entries_regcost;
516 /* Trace a patch through the CFG. */
518 struct branch_path
520 /* The basic block for this path entry. */
521 basic_block bb;
524 /* This data describes a block that will be processed by
525 cse_extended_basic_block. */
527 struct cse_basic_block_data
529 /* Total number of SETs in block. */
530 int nsets;
531 /* Size of current branch path, if any. */
532 int path_size;
533 /* Current path, indicating which basic_blocks will be processed. */
534 struct branch_path *path;
538 /* Pointers to the live in/live out bitmaps for the boundaries of the
539 current EBB. */
540 static bitmap cse_ebb_live_in, cse_ebb_live_out;
542 /* A simple bitmap to track which basic blocks have been visited
543 already as part of an already processed extended basic block. */
544 static sbitmap cse_visited_basic_blocks;
546 static bool fixed_base_plus_p (rtx x);
547 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
548 static int preferable (int, int, int, int);
549 static void new_basic_block (void);
550 static void make_new_qty (unsigned int, machine_mode);
551 static void make_regs_eqv (unsigned int, unsigned int);
552 static void delete_reg_equiv (unsigned int);
553 static int mention_regs (rtx);
554 static int insert_regs (rtx, struct table_elt *, int);
555 static void remove_from_table (struct table_elt *, unsigned);
556 static void remove_pseudo_from_table (rtx, unsigned);
557 static struct table_elt *lookup (rtx, unsigned, machine_mode);
558 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
559 static rtx lookup_as_function (rtx, enum rtx_code);
560 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
561 machine_mode, int, int);
562 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
563 machine_mode);
564 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
565 static void invalidate (rtx, machine_mode);
566 static void remove_invalid_refs (unsigned int);
567 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
568 machine_mode);
569 static void rehash_using_reg (rtx);
570 static void invalidate_memory (void);
571 static void invalidate_for_call (void);
572 static rtx use_related_value (rtx, struct table_elt *);
574 static inline unsigned canon_hash (rtx, machine_mode);
575 static inline unsigned safe_hash (rtx, machine_mode);
576 static inline unsigned hash_rtx_string (const char *);
578 static rtx canon_reg (rtx, rtx_insn *);
579 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
580 machine_mode *,
581 machine_mode *);
582 static rtx fold_rtx (rtx, rtx_insn *);
583 static rtx equiv_constant (rtx);
584 static void record_jump_equiv (rtx_insn *, bool);
585 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
586 int);
587 static void cse_insn (rtx_insn *);
588 static void cse_prescan_path (struct cse_basic_block_data *);
589 static void invalidate_from_clobbers (rtx_insn *);
590 static void invalidate_from_sets_and_clobbers (rtx_insn *);
591 static rtx cse_process_notes (rtx, rtx, bool *);
592 static void cse_extended_basic_block (struct cse_basic_block_data *);
593 extern void dump_class (struct table_elt*);
594 static void get_cse_reg_info_1 (unsigned int regno);
595 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
597 static void flush_hash_table (void);
598 static bool insn_live_p (rtx_insn *, int *);
599 static bool set_live_p (rtx, rtx_insn *, int *);
600 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
601 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
602 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
603 bool);
606 #undef RTL_HOOKS_GEN_LOWPART
607 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
609 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
611 /* Nonzero if X has the form (PLUS frame-pointer integer). */
613 static bool
614 fixed_base_plus_p (rtx x)
616 switch (GET_CODE (x))
618 case REG:
619 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
620 return true;
621 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
622 return true;
623 return false;
625 case PLUS:
626 if (!CONST_INT_P (XEXP (x, 1)))
627 return false;
628 return fixed_base_plus_p (XEXP (x, 0));
630 default:
631 return false;
635 /* Dump the expressions in the equivalence class indicated by CLASSP.
636 This function is used only for debugging. */
637 DEBUG_FUNCTION void
638 dump_class (struct table_elt *classp)
640 struct table_elt *elt;
642 fprintf (stderr, "Equivalence chain for ");
643 print_rtl (stderr, classp->exp);
644 fprintf (stderr, ": \n");
646 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
648 print_rtl (stderr, elt->exp);
649 fprintf (stderr, "\n");
653 /* Return an estimate of the cost of the registers used in an rtx.
654 This is mostly the number of different REG expressions in the rtx;
655 however for some exceptions like fixed registers we use a cost of
656 0. If any other hard register reference occurs, return MAX_COST. */
658 static int
659 approx_reg_cost (const_rtx x)
661 int cost = 0;
662 subrtx_iterator::array_type array;
663 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
665 const_rtx x = *iter;
666 if (REG_P (x))
668 unsigned int regno = REGNO (x);
669 if (!CHEAP_REGNO (regno))
671 if (regno < FIRST_PSEUDO_REGISTER)
673 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
674 return MAX_COST;
675 cost += 2;
677 else
678 cost += 1;
682 return cost;
685 /* Return a negative value if an rtx A, whose costs are given by COST_A
686 and REGCOST_A, is more desirable than an rtx B.
687 Return a positive value if A is less desirable, or 0 if the two are
688 equally good. */
689 static int
690 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
692 /* First, get rid of cases involving expressions that are entirely
693 unwanted. */
694 if (cost_a != cost_b)
696 if (cost_a == MAX_COST)
697 return 1;
698 if (cost_b == MAX_COST)
699 return -1;
702 /* Avoid extending lifetimes of hardregs. */
703 if (regcost_a != regcost_b)
705 if (regcost_a == MAX_COST)
706 return 1;
707 if (regcost_b == MAX_COST)
708 return -1;
711 /* Normal operation costs take precedence. */
712 if (cost_a != cost_b)
713 return cost_a - cost_b;
714 /* Only if these are identical consider effects on register pressure. */
715 if (regcost_a != regcost_b)
716 return regcost_a - regcost_b;
717 return 0;
720 /* Internal function, to compute cost when X is not a register; called
721 from COST macro to keep it simple. */
723 static int
724 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
726 return ((GET_CODE (x) == SUBREG
727 && REG_P (SUBREG_REG (x))
728 && GET_MODE_CLASS (mode) == MODE_INT
729 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
730 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
731 && subreg_lowpart_p (x)
732 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x))))
734 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
738 /* Initialize CSE_REG_INFO_TABLE. */
740 static void
741 init_cse_reg_info (unsigned int nregs)
743 /* Do we need to grow the table? */
744 if (nregs > cse_reg_info_table_size)
746 unsigned int new_size;
748 if (cse_reg_info_table_size < 2048)
750 /* Compute a new size that is a power of 2 and no smaller
751 than the large of NREGS and 64. */
752 new_size = (cse_reg_info_table_size
753 ? cse_reg_info_table_size : 64);
755 while (new_size < nregs)
756 new_size *= 2;
758 else
760 /* If we need a big table, allocate just enough to hold
761 NREGS registers. */
762 new_size = nregs;
765 /* Reallocate the table with NEW_SIZE entries. */
766 free (cse_reg_info_table);
767 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
768 cse_reg_info_table_size = new_size;
769 cse_reg_info_table_first_uninitialized = 0;
772 /* Do we have all of the first NREGS entries initialized? */
773 if (cse_reg_info_table_first_uninitialized < nregs)
775 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
776 unsigned int i;
778 /* Put the old timestamp on newly allocated entries so that they
779 will all be considered out of date. We do not touch those
780 entries beyond the first NREGS entries to be nice to the
781 virtual memory. */
782 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
783 cse_reg_info_table[i].timestamp = old_timestamp;
785 cse_reg_info_table_first_uninitialized = nregs;
789 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
791 static void
792 get_cse_reg_info_1 (unsigned int regno)
794 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
795 entry will be considered to have been initialized. */
796 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
798 /* Initialize the rest of the entry. */
799 cse_reg_info_table[regno].reg_tick = 1;
800 cse_reg_info_table[regno].reg_in_table = -1;
801 cse_reg_info_table[regno].subreg_ticked = -1;
802 cse_reg_info_table[regno].reg_qty = -regno - 1;
805 /* Find a cse_reg_info entry for REGNO. */
807 static inline struct cse_reg_info *
808 get_cse_reg_info (unsigned int regno)
810 struct cse_reg_info *p = &cse_reg_info_table[regno];
812 /* If this entry has not been initialized, go ahead and initialize
813 it. */
814 if (p->timestamp != cse_reg_info_timestamp)
815 get_cse_reg_info_1 (regno);
817 return p;
820 /* Clear the hash table and initialize each register with its own quantity,
821 for a new basic block. */
823 static void
824 new_basic_block (void)
826 int i;
828 next_qty = 0;
830 /* Invalidate cse_reg_info_table. */
831 cse_reg_info_timestamp++;
833 /* Clear out hash table state for this pass. */
834 CLEAR_HARD_REG_SET (hard_regs_in_table);
836 /* The per-quantity values used to be initialized here, but it is
837 much faster to initialize each as it is made in `make_new_qty'. */
839 for (i = 0; i < HASH_SIZE; i++)
841 struct table_elt *first;
843 first = table[i];
844 if (first != NULL)
846 struct table_elt *last = first;
848 table[i] = NULL;
850 while (last->next_same_hash != NULL)
851 last = last->next_same_hash;
853 /* Now relink this hash entire chain into
854 the free element list. */
856 last->next_same_hash = free_element_chain;
857 free_element_chain = first;
861 prev_insn_cc0 = 0;
864 /* Say that register REG contains a quantity in mode MODE not in any
865 register before and initialize that quantity. */
867 static void
868 make_new_qty (unsigned int reg, machine_mode mode)
870 int q;
871 struct qty_table_elem *ent;
872 struct reg_eqv_elem *eqv;
874 gcc_assert (next_qty < max_qty);
876 q = REG_QTY (reg) = next_qty++;
877 ent = &qty_table[q];
878 ent->first_reg = reg;
879 ent->last_reg = reg;
880 ent->mode = mode;
881 ent->const_rtx = ent->const_insn = NULL;
882 ent->comparison_code = UNKNOWN;
884 eqv = &reg_eqv_table[reg];
885 eqv->next = eqv->prev = -1;
888 /* Make reg NEW equivalent to reg OLD.
889 OLD is not changing; NEW is. */
891 static void
892 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
894 unsigned int lastr, firstr;
895 int q = REG_QTY (old_reg);
896 struct qty_table_elem *ent;
898 ent = &qty_table[q];
900 /* Nothing should become eqv until it has a "non-invalid" qty number. */
901 gcc_assert (REGNO_QTY_VALID_P (old_reg));
903 REG_QTY (new_reg) = q;
904 firstr = ent->first_reg;
905 lastr = ent->last_reg;
907 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
908 hard regs. Among pseudos, if NEW will live longer than any other reg
909 of the same qty, and that is beyond the current basic block,
910 make it the new canonical replacement for this qty. */
911 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
912 /* Certain fixed registers might be of the class NO_REGS. This means
913 that not only can they not be allocated by the compiler, but
914 they cannot be used in substitutions or canonicalizations
915 either. */
916 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
917 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
918 || (new_reg >= FIRST_PSEUDO_REGISTER
919 && (firstr < FIRST_PSEUDO_REGISTER
920 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
921 && !bitmap_bit_p (cse_ebb_live_out, firstr))
922 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
923 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
925 reg_eqv_table[firstr].prev = new_reg;
926 reg_eqv_table[new_reg].next = firstr;
927 reg_eqv_table[new_reg].prev = -1;
928 ent->first_reg = new_reg;
930 else
932 /* If NEW is a hard reg (known to be non-fixed), insert at end.
933 Otherwise, insert before any non-fixed hard regs that are at the
934 end. Registers of class NO_REGS cannot be used as an
935 equivalent for anything. */
936 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
937 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
938 && new_reg >= FIRST_PSEUDO_REGISTER)
939 lastr = reg_eqv_table[lastr].prev;
940 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
941 if (reg_eqv_table[lastr].next >= 0)
942 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
943 else
944 qty_table[q].last_reg = new_reg;
945 reg_eqv_table[lastr].next = new_reg;
946 reg_eqv_table[new_reg].prev = lastr;
950 /* Remove REG from its equivalence class. */
952 static void
953 delete_reg_equiv (unsigned int reg)
955 struct qty_table_elem *ent;
956 int q = REG_QTY (reg);
957 int p, n;
959 /* If invalid, do nothing. */
960 if (! REGNO_QTY_VALID_P (reg))
961 return;
963 ent = &qty_table[q];
965 p = reg_eqv_table[reg].prev;
966 n = reg_eqv_table[reg].next;
968 if (n != -1)
969 reg_eqv_table[n].prev = p;
970 else
971 ent->last_reg = p;
972 if (p != -1)
973 reg_eqv_table[p].next = n;
974 else
975 ent->first_reg = n;
977 REG_QTY (reg) = -reg - 1;
980 /* Remove any invalid expressions from the hash table
981 that refer to any of the registers contained in expression X.
983 Make sure that newly inserted references to those registers
984 as subexpressions will be considered valid.
986 mention_regs is not called when a register itself
987 is being stored in the table.
989 Return 1 if we have done something that may have changed the hash code
990 of X. */
992 static int
993 mention_regs (rtx x)
995 enum rtx_code code;
996 int i, j;
997 const char *fmt;
998 int changed = 0;
1000 if (x == 0)
1001 return 0;
1003 code = GET_CODE (x);
1004 if (code == REG)
1006 unsigned int regno = REGNO (x);
1007 unsigned int endregno = END_REGNO (x);
1008 unsigned int i;
1010 for (i = regno; i < endregno; i++)
1012 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1013 remove_invalid_refs (i);
1015 REG_IN_TABLE (i) = REG_TICK (i);
1016 SUBREG_TICKED (i) = -1;
1019 return 0;
1022 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1023 pseudo if they don't use overlapping words. We handle only pseudos
1024 here for simplicity. */
1025 if (code == SUBREG && REG_P (SUBREG_REG (x))
1026 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1028 unsigned int i = REGNO (SUBREG_REG (x));
1030 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1032 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1033 the last store to this register really stored into this
1034 subreg, then remove the memory of this subreg.
1035 Otherwise, remove any memory of the entire register and
1036 all its subregs from the table. */
1037 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1038 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1039 remove_invalid_refs (i);
1040 else
1041 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1044 REG_IN_TABLE (i) = REG_TICK (i);
1045 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1046 return 0;
1049 /* If X is a comparison or a COMPARE and either operand is a register
1050 that does not have a quantity, give it one. This is so that a later
1051 call to record_jump_equiv won't cause X to be assigned a different
1052 hash code and not found in the table after that call.
1054 It is not necessary to do this here, since rehash_using_reg can
1055 fix up the table later, but doing this here eliminates the need to
1056 call that expensive function in the most common case where the only
1057 use of the register is in the comparison. */
1059 if (code == COMPARE || COMPARISON_P (x))
1061 if (REG_P (XEXP (x, 0))
1062 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1063 if (insert_regs (XEXP (x, 0), NULL, 0))
1065 rehash_using_reg (XEXP (x, 0));
1066 changed = 1;
1069 if (REG_P (XEXP (x, 1))
1070 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1071 if (insert_regs (XEXP (x, 1), NULL, 0))
1073 rehash_using_reg (XEXP (x, 1));
1074 changed = 1;
1078 fmt = GET_RTX_FORMAT (code);
1079 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1080 if (fmt[i] == 'e')
1081 changed |= mention_regs (XEXP (x, i));
1082 else if (fmt[i] == 'E')
1083 for (j = 0; j < XVECLEN (x, i); j++)
1084 changed |= mention_regs (XVECEXP (x, i, j));
1086 return changed;
1089 /* Update the register quantities for inserting X into the hash table
1090 with a value equivalent to CLASSP.
1091 (If the class does not contain a REG, it is irrelevant.)
1092 If MODIFIED is nonzero, X is a destination; it is being modified.
1093 Note that delete_reg_equiv should be called on a register
1094 before insert_regs is done on that register with MODIFIED != 0.
1096 Nonzero value means that elements of reg_qty have changed
1097 so X's hash code may be different. */
1099 static int
1100 insert_regs (rtx x, struct table_elt *classp, int modified)
1102 if (REG_P (x))
1104 unsigned int regno = REGNO (x);
1105 int qty_valid;
1107 /* If REGNO is in the equivalence table already but is of the
1108 wrong mode for that equivalence, don't do anything here. */
1110 qty_valid = REGNO_QTY_VALID_P (regno);
1111 if (qty_valid)
1113 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1115 if (ent->mode != GET_MODE (x))
1116 return 0;
1119 if (modified || ! qty_valid)
1121 if (classp)
1122 for (classp = classp->first_same_value;
1123 classp != 0;
1124 classp = classp->next_same_value)
1125 if (REG_P (classp->exp)
1126 && GET_MODE (classp->exp) == GET_MODE (x))
1128 unsigned c_regno = REGNO (classp->exp);
1130 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1132 /* Suppose that 5 is hard reg and 100 and 101 are
1133 pseudos. Consider
1135 (set (reg:si 100) (reg:si 5))
1136 (set (reg:si 5) (reg:si 100))
1137 (set (reg:di 101) (reg:di 5))
1139 We would now set REG_QTY (101) = REG_QTY (5), but the
1140 entry for 5 is in SImode. When we use this later in
1141 copy propagation, we get the register in wrong mode. */
1142 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1143 continue;
1145 make_regs_eqv (regno, c_regno);
1146 return 1;
1149 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1150 than REG_IN_TABLE to find out if there was only a single preceding
1151 invalidation - for the SUBREG - or another one, which would be
1152 for the full register. However, if we find here that REG_TICK
1153 indicates that the register is invalid, it means that it has
1154 been invalidated in a separate operation. The SUBREG might be used
1155 now (then this is a recursive call), or we might use the full REG
1156 now and a SUBREG of it later. So bump up REG_TICK so that
1157 mention_regs will do the right thing. */
1158 if (! modified
1159 && REG_IN_TABLE (regno) >= 0
1160 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1161 REG_TICK (regno)++;
1162 make_new_qty (regno, GET_MODE (x));
1163 return 1;
1166 return 0;
1169 /* If X is a SUBREG, we will likely be inserting the inner register in the
1170 table. If that register doesn't have an assigned quantity number at
1171 this point but does later, the insertion that we will be doing now will
1172 not be accessible because its hash code will have changed. So assign
1173 a quantity number now. */
1175 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1176 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1178 insert_regs (SUBREG_REG (x), NULL, 0);
1179 mention_regs (x);
1180 return 1;
1182 else
1183 return mention_regs (x);
1187 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1188 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1189 CST is equal to an anchor. */
1191 static bool
1192 compute_const_anchors (rtx cst,
1193 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1194 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1196 HOST_WIDE_INT n = INTVAL (cst);
1198 *lower_base = n & ~(targetm.const_anchor - 1);
1199 if (*lower_base == n)
1200 return false;
1202 *upper_base =
1203 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1204 *upper_offs = n - *upper_base;
1205 *lower_offs = n - *lower_base;
1206 return true;
1209 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1211 static void
1212 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1213 machine_mode mode)
1215 struct table_elt *elt;
1216 unsigned hash;
1217 rtx anchor_exp;
1218 rtx exp;
1220 anchor_exp = GEN_INT (anchor);
1221 hash = HASH (anchor_exp, mode);
1222 elt = lookup (anchor_exp, hash, mode);
1223 if (!elt)
1224 elt = insert (anchor_exp, NULL, hash, mode);
1226 exp = plus_constant (mode, reg, offs);
1227 /* REG has just been inserted and the hash codes recomputed. */
1228 mention_regs (exp);
1229 hash = HASH (exp, mode);
1231 /* Use the cost of the register rather than the whole expression. When
1232 looking up constant anchors we will further offset the corresponding
1233 expression therefore it does not make sense to prefer REGs over
1234 reg-immediate additions. Prefer instead the oldest expression. Also
1235 don't prefer pseudos over hard regs so that we derive constants in
1236 argument registers from other argument registers rather than from the
1237 original pseudo that was used to synthesize the constant. */
1238 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1241 /* The constant CST is equivalent to the register REG. Create
1242 equivalences between the two anchors of CST and the corresponding
1243 register-offset expressions using REG. */
1245 static void
1246 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1248 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1250 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1251 &upper_base, &upper_offs))
1252 return;
1254 /* Ignore anchors of value 0. Constants accessible from zero are
1255 simple. */
1256 if (lower_base != 0)
1257 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1259 if (upper_base != 0)
1260 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1263 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1264 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1265 valid expression. Return the cheapest and oldest of such expressions. In
1266 *OLD, return how old the resulting expression is compared to the other
1267 equivalent expressions. */
1269 static rtx
1270 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1271 unsigned *old)
1273 struct table_elt *elt;
1274 unsigned idx;
1275 struct table_elt *match_elt;
1276 rtx match;
1278 /* Find the cheapest and *oldest* expression to maximize the chance of
1279 reusing the same pseudo. */
1281 match_elt = NULL;
1282 match = NULL_RTX;
1283 for (elt = anchor_elt->first_same_value, idx = 0;
1284 elt;
1285 elt = elt->next_same_value, idx++)
1287 if (match_elt && CHEAPER (match_elt, elt))
1288 return match;
1290 if (REG_P (elt->exp)
1291 || (GET_CODE (elt->exp) == PLUS
1292 && REG_P (XEXP (elt->exp, 0))
1293 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1295 rtx x;
1297 /* Ignore expressions that are no longer valid. */
1298 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1299 continue;
1301 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1302 if (REG_P (x)
1303 || (GET_CODE (x) == PLUS
1304 && IN_RANGE (INTVAL (XEXP (x, 1)),
1305 -targetm.const_anchor,
1306 targetm.const_anchor - 1)))
1308 match = x;
1309 match_elt = elt;
1310 *old = idx;
1315 return match;
1318 /* Try to express the constant SRC_CONST using a register+offset expression
1319 derived from a constant anchor. Return it if successful or NULL_RTX,
1320 otherwise. */
1322 static rtx
1323 try_const_anchors (rtx src_const, machine_mode mode)
1325 struct table_elt *lower_elt, *upper_elt;
1326 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1327 rtx lower_anchor_rtx, upper_anchor_rtx;
1328 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1329 unsigned lower_old, upper_old;
1331 /* CONST_INT is used for CC modes, but we should leave those alone. */
1332 if (GET_MODE_CLASS (mode) == MODE_CC)
1333 return NULL_RTX;
1335 gcc_assert (SCALAR_INT_MODE_P (mode));
1336 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1337 &upper_base, &upper_offs))
1338 return NULL_RTX;
1340 lower_anchor_rtx = GEN_INT (lower_base);
1341 upper_anchor_rtx = GEN_INT (upper_base);
1342 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1343 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1345 if (lower_elt)
1346 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1347 if (upper_elt)
1348 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1350 if (!lower_exp)
1351 return upper_exp;
1352 if (!upper_exp)
1353 return lower_exp;
1355 /* Return the older expression. */
1356 return (upper_old > lower_old ? upper_exp : lower_exp);
1359 /* Look in or update the hash table. */
1361 /* Remove table element ELT from use in the table.
1362 HASH is its hash code, made using the HASH macro.
1363 It's an argument because often that is known in advance
1364 and we save much time not recomputing it. */
1366 static void
1367 remove_from_table (struct table_elt *elt, unsigned int hash)
1369 if (elt == 0)
1370 return;
1372 /* Mark this element as removed. See cse_insn. */
1373 elt->first_same_value = 0;
1375 /* Remove the table element from its equivalence class. */
1378 struct table_elt *prev = elt->prev_same_value;
1379 struct table_elt *next = elt->next_same_value;
1381 if (next)
1382 next->prev_same_value = prev;
1384 if (prev)
1385 prev->next_same_value = next;
1386 else
1388 struct table_elt *newfirst = next;
1389 while (next)
1391 next->first_same_value = newfirst;
1392 next = next->next_same_value;
1397 /* Remove the table element from its hash bucket. */
1400 struct table_elt *prev = elt->prev_same_hash;
1401 struct table_elt *next = elt->next_same_hash;
1403 if (next)
1404 next->prev_same_hash = prev;
1406 if (prev)
1407 prev->next_same_hash = next;
1408 else if (table[hash] == elt)
1409 table[hash] = next;
1410 else
1412 /* This entry is not in the proper hash bucket. This can happen
1413 when two classes were merged by `merge_equiv_classes'. Search
1414 for the hash bucket that it heads. This happens only very
1415 rarely, so the cost is acceptable. */
1416 for (hash = 0; hash < HASH_SIZE; hash++)
1417 if (table[hash] == elt)
1418 table[hash] = next;
1422 /* Remove the table element from its related-value circular chain. */
1424 if (elt->related_value != 0 && elt->related_value != elt)
1426 struct table_elt *p = elt->related_value;
1428 while (p->related_value != elt)
1429 p = p->related_value;
1430 p->related_value = elt->related_value;
1431 if (p->related_value == p)
1432 p->related_value = 0;
1435 /* Now add it to the free element chain. */
1436 elt->next_same_hash = free_element_chain;
1437 free_element_chain = elt;
1440 /* Same as above, but X is a pseudo-register. */
1442 static void
1443 remove_pseudo_from_table (rtx x, unsigned int hash)
1445 struct table_elt *elt;
1447 /* Because a pseudo-register can be referenced in more than one
1448 mode, we might have to remove more than one table entry. */
1449 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1450 remove_from_table (elt, hash);
1453 /* Look up X in the hash table and return its table element,
1454 or 0 if X is not in the table.
1456 MODE is the machine-mode of X, or if X is an integer constant
1457 with VOIDmode then MODE is the mode with which X will be used.
1459 Here we are satisfied to find an expression whose tree structure
1460 looks like X. */
1462 static struct table_elt *
1463 lookup (rtx x, unsigned int hash, machine_mode mode)
1465 struct table_elt *p;
1467 for (p = table[hash]; p; p = p->next_same_hash)
1468 if (mode == p->mode && ((x == p->exp && REG_P (x))
1469 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1470 return p;
1472 return 0;
1475 /* Like `lookup' but don't care whether the table element uses invalid regs.
1476 Also ignore discrepancies in the machine mode of a register. */
1478 static struct table_elt *
1479 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1481 struct table_elt *p;
1483 if (REG_P (x))
1485 unsigned int regno = REGNO (x);
1487 /* Don't check the machine mode when comparing registers;
1488 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1489 for (p = table[hash]; p; p = p->next_same_hash)
1490 if (REG_P (p->exp)
1491 && REGNO (p->exp) == regno)
1492 return p;
1494 else
1496 for (p = table[hash]; p; p = p->next_same_hash)
1497 if (mode == p->mode
1498 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1499 return p;
1502 return 0;
1505 /* Look for an expression equivalent to X and with code CODE.
1506 If one is found, return that expression. */
1508 static rtx
1509 lookup_as_function (rtx x, enum rtx_code code)
1511 struct table_elt *p
1512 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1514 if (p == 0)
1515 return 0;
1517 for (p = p->first_same_value; p; p = p->next_same_value)
1518 if (GET_CODE (p->exp) == code
1519 /* Make sure this is a valid entry in the table. */
1520 && exp_equiv_p (p->exp, p->exp, 1, false))
1521 return p->exp;
1523 return 0;
1526 /* Insert X in the hash table, assuming HASH is its hash code and
1527 CLASSP is an element of the class it should go in (or 0 if a new
1528 class should be made). COST is the code of X and reg_cost is the
1529 cost of registers in X. It is inserted at the proper position to
1530 keep the class in the order cheapest first.
1532 MODE is the machine-mode of X, or if X is an integer constant
1533 with VOIDmode then MODE is the mode with which X will be used.
1535 For elements of equal cheapness, the most recent one
1536 goes in front, except that the first element in the list
1537 remains first unless a cheaper element is added. The order of
1538 pseudo-registers does not matter, as canon_reg will be called to
1539 find the cheapest when a register is retrieved from the table.
1541 The in_memory field in the hash table element is set to 0.
1542 The caller must set it nonzero if appropriate.
1544 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1545 and if insert_regs returns a nonzero value
1546 you must then recompute its hash code before calling here.
1548 If necessary, update table showing constant values of quantities. */
1550 static struct table_elt *
1551 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1552 machine_mode mode, int cost, int reg_cost)
1554 struct table_elt *elt;
1556 /* If X is a register and we haven't made a quantity for it,
1557 something is wrong. */
1558 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1560 /* If X is a hard register, show it is being put in the table. */
1561 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1562 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1564 /* Put an element for X into the right hash bucket. */
1566 elt = free_element_chain;
1567 if (elt)
1568 free_element_chain = elt->next_same_hash;
1569 else
1570 elt = XNEW (struct table_elt);
1572 elt->exp = x;
1573 elt->canon_exp = NULL_RTX;
1574 elt->cost = cost;
1575 elt->regcost = reg_cost;
1576 elt->next_same_value = 0;
1577 elt->prev_same_value = 0;
1578 elt->next_same_hash = table[hash];
1579 elt->prev_same_hash = 0;
1580 elt->related_value = 0;
1581 elt->in_memory = 0;
1582 elt->mode = mode;
1583 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1585 if (table[hash])
1586 table[hash]->prev_same_hash = elt;
1587 table[hash] = elt;
1589 /* Put it into the proper value-class. */
1590 if (classp)
1592 classp = classp->first_same_value;
1593 if (CHEAPER (elt, classp))
1594 /* Insert at the head of the class. */
1596 struct table_elt *p;
1597 elt->next_same_value = classp;
1598 classp->prev_same_value = elt;
1599 elt->first_same_value = elt;
1601 for (p = classp; p; p = p->next_same_value)
1602 p->first_same_value = elt;
1604 else
1606 /* Insert not at head of the class. */
1607 /* Put it after the last element cheaper than X. */
1608 struct table_elt *p, *next;
1610 for (p = classp;
1611 (next = p->next_same_value) && CHEAPER (next, elt);
1612 p = next)
1615 /* Put it after P and before NEXT. */
1616 elt->next_same_value = next;
1617 if (next)
1618 next->prev_same_value = elt;
1620 elt->prev_same_value = p;
1621 p->next_same_value = elt;
1622 elt->first_same_value = classp;
1625 else
1626 elt->first_same_value = elt;
1628 /* If this is a constant being set equivalent to a register or a register
1629 being set equivalent to a constant, note the constant equivalence.
1631 If this is a constant, it cannot be equivalent to a different constant,
1632 and a constant is the only thing that can be cheaper than a register. So
1633 we know the register is the head of the class (before the constant was
1634 inserted).
1636 If this is a register that is not already known equivalent to a
1637 constant, we must check the entire class.
1639 If this is a register that is already known equivalent to an insn,
1640 update the qtys `const_insn' to show that `this_insn' is the latest
1641 insn making that quantity equivalent to the constant. */
1643 if (elt->is_const && classp && REG_P (classp->exp)
1644 && !REG_P (x))
1646 int exp_q = REG_QTY (REGNO (classp->exp));
1647 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1649 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1650 exp_ent->const_insn = this_insn;
1653 else if (REG_P (x)
1654 && classp
1655 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1656 && ! elt->is_const)
1658 struct table_elt *p;
1660 for (p = classp; p != 0; p = p->next_same_value)
1662 if (p->is_const && !REG_P (p->exp))
1664 int x_q = REG_QTY (REGNO (x));
1665 struct qty_table_elem *x_ent = &qty_table[x_q];
1667 x_ent->const_rtx
1668 = gen_lowpart (GET_MODE (x), p->exp);
1669 x_ent->const_insn = this_insn;
1670 break;
1675 else if (REG_P (x)
1676 && qty_table[REG_QTY (REGNO (x))].const_rtx
1677 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1678 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1680 /* If this is a constant with symbolic value,
1681 and it has a term with an explicit integer value,
1682 link it up with related expressions. */
1683 if (GET_CODE (x) == CONST)
1685 rtx subexp = get_related_value (x);
1686 unsigned subhash;
1687 struct table_elt *subelt, *subelt_prev;
1689 if (subexp != 0)
1691 /* Get the integer-free subexpression in the hash table. */
1692 subhash = SAFE_HASH (subexp, mode);
1693 subelt = lookup (subexp, subhash, mode);
1694 if (subelt == 0)
1695 subelt = insert (subexp, NULL, subhash, mode);
1696 /* Initialize SUBELT's circular chain if it has none. */
1697 if (subelt->related_value == 0)
1698 subelt->related_value = subelt;
1699 /* Find the element in the circular chain that precedes SUBELT. */
1700 subelt_prev = subelt;
1701 while (subelt_prev->related_value != subelt)
1702 subelt_prev = subelt_prev->related_value;
1703 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1704 This way the element that follows SUBELT is the oldest one. */
1705 elt->related_value = subelt_prev->related_value;
1706 subelt_prev->related_value = elt;
1710 return elt;
1713 /* Wrap insert_with_costs by passing the default costs. */
1715 static struct table_elt *
1716 insert (rtx x, struct table_elt *classp, unsigned int hash,
1717 machine_mode mode)
1719 return insert_with_costs (x, classp, hash, mode,
1720 COST (x, mode), approx_reg_cost (x));
1724 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1725 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1726 the two classes equivalent.
1728 CLASS1 will be the surviving class; CLASS2 should not be used after this
1729 call.
1731 Any invalid entries in CLASS2 will not be copied. */
1733 static void
1734 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1736 struct table_elt *elt, *next, *new_elt;
1738 /* Ensure we start with the head of the classes. */
1739 class1 = class1->first_same_value;
1740 class2 = class2->first_same_value;
1742 /* If they were already equal, forget it. */
1743 if (class1 == class2)
1744 return;
1746 for (elt = class2; elt; elt = next)
1748 unsigned int hash;
1749 rtx exp = elt->exp;
1750 machine_mode mode = elt->mode;
1752 next = elt->next_same_value;
1754 /* Remove old entry, make a new one in CLASS1's class.
1755 Don't do this for invalid entries as we cannot find their
1756 hash code (it also isn't necessary). */
1757 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1759 bool need_rehash = false;
1761 hash_arg_in_memory = 0;
1762 hash = HASH (exp, mode);
1764 if (REG_P (exp))
1766 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1767 delete_reg_equiv (REGNO (exp));
1770 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1771 remove_pseudo_from_table (exp, hash);
1772 else
1773 remove_from_table (elt, hash);
1775 if (insert_regs (exp, class1, 0) || need_rehash)
1777 rehash_using_reg (exp);
1778 hash = HASH (exp, mode);
1780 new_elt = insert (exp, class1, hash, mode);
1781 new_elt->in_memory = hash_arg_in_memory;
1782 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1783 new_elt->cost = MAX_COST;
1788 /* Flush the entire hash table. */
1790 static void
1791 flush_hash_table (void)
1793 int i;
1794 struct table_elt *p;
1796 for (i = 0; i < HASH_SIZE; i++)
1797 for (p = table[i]; p; p = table[i])
1799 /* Note that invalidate can remove elements
1800 after P in the current hash chain. */
1801 if (REG_P (p->exp))
1802 invalidate (p->exp, VOIDmode);
1803 else
1804 remove_from_table (p, i);
1808 /* Check whether an anti dependence exists between X and EXP. MODE and
1809 ADDR are as for canon_anti_dependence. */
1811 static bool
1812 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1814 subrtx_iterator::array_type array;
1815 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1817 const_rtx x = *iter;
1818 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1819 return true;
1821 return false;
1824 /* Remove from the hash table, or mark as invalid, all expressions whose
1825 values could be altered by storing in X. X is a register, a subreg, or
1826 a memory reference with nonvarying address (because, when a memory
1827 reference with a varying address is stored in, all memory references are
1828 removed by invalidate_memory so specific invalidation is superfluous).
1829 FULL_MODE, if not VOIDmode, indicates that this much should be
1830 invalidated instead of just the amount indicated by the mode of X. This
1831 is only used for bitfield stores into memory.
1833 A nonvarying address may be just a register or just a symbol reference,
1834 or it may be either of those plus a numeric offset. */
1836 static void
1837 invalidate (rtx x, machine_mode full_mode)
1839 int i;
1840 struct table_elt *p;
1841 rtx addr;
1843 switch (GET_CODE (x))
1845 case REG:
1847 /* If X is a register, dependencies on its contents are recorded
1848 through the qty number mechanism. Just change the qty number of
1849 the register, mark it as invalid for expressions that refer to it,
1850 and remove it itself. */
1851 unsigned int regno = REGNO (x);
1852 unsigned int hash = HASH (x, GET_MODE (x));
1854 /* Remove REGNO from any quantity list it might be on and indicate
1855 that its value might have changed. If it is a pseudo, remove its
1856 entry from the hash table.
1858 For a hard register, we do the first two actions above for any
1859 additional hard registers corresponding to X. Then, if any of these
1860 registers are in the table, we must remove any REG entries that
1861 overlap these registers. */
1863 delete_reg_equiv (regno);
1864 REG_TICK (regno)++;
1865 SUBREG_TICKED (regno) = -1;
1867 if (regno >= FIRST_PSEUDO_REGISTER)
1868 remove_pseudo_from_table (x, hash);
1869 else
1871 HOST_WIDE_INT in_table
1872 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1873 unsigned int endregno = END_REGNO (x);
1874 unsigned int tregno, tendregno, rn;
1875 struct table_elt *p, *next;
1877 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1879 for (rn = regno + 1; rn < endregno; rn++)
1881 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1882 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1883 delete_reg_equiv (rn);
1884 REG_TICK (rn)++;
1885 SUBREG_TICKED (rn) = -1;
1888 if (in_table)
1889 for (hash = 0; hash < HASH_SIZE; hash++)
1890 for (p = table[hash]; p; p = next)
1892 next = p->next_same_hash;
1894 if (!REG_P (p->exp)
1895 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1896 continue;
1898 tregno = REGNO (p->exp);
1899 tendregno = END_REGNO (p->exp);
1900 if (tendregno > regno && tregno < endregno)
1901 remove_from_table (p, hash);
1905 return;
1907 case SUBREG:
1908 invalidate (SUBREG_REG (x), VOIDmode);
1909 return;
1911 case PARALLEL:
1912 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1913 invalidate (XVECEXP (x, 0, i), VOIDmode);
1914 return;
1916 case EXPR_LIST:
1917 /* This is part of a disjoint return value; extract the location in
1918 question ignoring the offset. */
1919 invalidate (XEXP (x, 0), VOIDmode);
1920 return;
1922 case MEM:
1923 addr = canon_rtx (get_addr (XEXP (x, 0)));
1924 /* Calculate the canonical version of X here so that
1925 true_dependence doesn't generate new RTL for X on each call. */
1926 x = canon_rtx (x);
1928 /* Remove all hash table elements that refer to overlapping pieces of
1929 memory. */
1930 if (full_mode == VOIDmode)
1931 full_mode = GET_MODE (x);
1933 for (i = 0; i < HASH_SIZE; i++)
1935 struct table_elt *next;
1937 for (p = table[i]; p; p = next)
1939 next = p->next_same_hash;
1940 if (p->in_memory)
1942 /* Just canonicalize the expression once;
1943 otherwise each time we call invalidate
1944 true_dependence will canonicalize the
1945 expression again. */
1946 if (!p->canon_exp)
1947 p->canon_exp = canon_rtx (p->exp);
1948 if (check_dependence (p->canon_exp, x, full_mode, addr))
1949 remove_from_table (p, i);
1953 return;
1955 default:
1956 gcc_unreachable ();
1960 /* Invalidate DEST. Used when DEST is not going to be added
1961 into the hash table for some reason, e.g. do_not_record
1962 flagged on it. */
1964 static void
1965 invalidate_dest (rtx dest)
1967 if (REG_P (dest)
1968 || GET_CODE (dest) == SUBREG
1969 || MEM_P (dest))
1970 invalidate (dest, VOIDmode);
1971 else if (GET_CODE (dest) == STRICT_LOW_PART
1972 || GET_CODE (dest) == ZERO_EXTRACT)
1973 invalidate (XEXP (dest, 0), GET_MODE (dest));
1976 /* Remove all expressions that refer to register REGNO,
1977 since they are already invalid, and we are about to
1978 mark that register valid again and don't want the old
1979 expressions to reappear as valid. */
1981 static void
1982 remove_invalid_refs (unsigned int regno)
1984 unsigned int i;
1985 struct table_elt *p, *next;
1987 for (i = 0; i < HASH_SIZE; i++)
1988 for (p = table[i]; p; p = next)
1990 next = p->next_same_hash;
1991 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1992 remove_from_table (p, i);
1996 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1997 and mode MODE. */
1998 static void
1999 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2000 machine_mode mode)
2002 unsigned int i;
2003 struct table_elt *p, *next;
2004 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2006 for (i = 0; i < HASH_SIZE; i++)
2007 for (p = table[i]; p; p = next)
2009 rtx exp = p->exp;
2010 next = p->next_same_hash;
2012 if (!REG_P (exp)
2013 && (GET_CODE (exp) != SUBREG
2014 || !REG_P (SUBREG_REG (exp))
2015 || REGNO (SUBREG_REG (exp)) != regno
2016 || (((SUBREG_BYTE (exp)
2017 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2018 && SUBREG_BYTE (exp) <= end))
2019 && refers_to_regno_p (regno, p->exp))
2020 remove_from_table (p, i);
2024 /* Recompute the hash codes of any valid entries in the hash table that
2025 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2027 This is called when we make a jump equivalence. */
2029 static void
2030 rehash_using_reg (rtx x)
2032 unsigned int i;
2033 struct table_elt *p, *next;
2034 unsigned hash;
2036 if (GET_CODE (x) == SUBREG)
2037 x = SUBREG_REG (x);
2039 /* If X is not a register or if the register is known not to be in any
2040 valid entries in the table, we have no work to do. */
2042 if (!REG_P (x)
2043 || REG_IN_TABLE (REGNO (x)) < 0
2044 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2045 return;
2047 /* Scan all hash chains looking for valid entries that mention X.
2048 If we find one and it is in the wrong hash chain, move it. */
2050 for (i = 0; i < HASH_SIZE; i++)
2051 for (p = table[i]; p; p = next)
2053 next = p->next_same_hash;
2054 if (reg_mentioned_p (x, p->exp)
2055 && exp_equiv_p (p->exp, p->exp, 1, false)
2056 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2058 if (p->next_same_hash)
2059 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2061 if (p->prev_same_hash)
2062 p->prev_same_hash->next_same_hash = p->next_same_hash;
2063 else
2064 table[i] = p->next_same_hash;
2066 p->next_same_hash = table[hash];
2067 p->prev_same_hash = 0;
2068 if (table[hash])
2069 table[hash]->prev_same_hash = p;
2070 table[hash] = p;
2075 /* Remove from the hash table any expression that is a call-clobbered
2076 register. Also update their TICK values. */
2078 static void
2079 invalidate_for_call (void)
2081 unsigned int regno, endregno;
2082 unsigned int i;
2083 unsigned hash;
2084 struct table_elt *p, *next;
2085 int in_table = 0;
2086 hard_reg_set_iterator hrsi;
2088 /* Go through all the hard registers. For each that is clobbered in
2089 a CALL_INSN, remove the register from quantity chains and update
2090 reg_tick if defined. Also see if any of these registers is currently
2091 in the table. */
2092 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2094 delete_reg_equiv (regno);
2095 if (REG_TICK (regno) >= 0)
2097 REG_TICK (regno)++;
2098 SUBREG_TICKED (regno) = -1;
2100 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2103 /* In the case where we have no call-clobbered hard registers in the
2104 table, we are done. Otherwise, scan the table and remove any
2105 entry that overlaps a call-clobbered register. */
2107 if (in_table)
2108 for (hash = 0; hash < HASH_SIZE; hash++)
2109 for (p = table[hash]; p; p = next)
2111 next = p->next_same_hash;
2113 if (!REG_P (p->exp)
2114 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2115 continue;
2117 regno = REGNO (p->exp);
2118 endregno = END_REGNO (p->exp);
2120 for (i = regno; i < endregno; i++)
2121 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2123 remove_from_table (p, hash);
2124 break;
2129 /* Given an expression X of type CONST,
2130 and ELT which is its table entry (or 0 if it
2131 is not in the hash table),
2132 return an alternate expression for X as a register plus integer.
2133 If none can be found, return 0. */
2135 static rtx
2136 use_related_value (rtx x, struct table_elt *elt)
2138 struct table_elt *relt = 0;
2139 struct table_elt *p, *q;
2140 HOST_WIDE_INT offset;
2142 /* First, is there anything related known?
2143 If we have a table element, we can tell from that.
2144 Otherwise, must look it up. */
2146 if (elt != 0 && elt->related_value != 0)
2147 relt = elt;
2148 else if (elt == 0 && GET_CODE (x) == CONST)
2150 rtx subexp = get_related_value (x);
2151 if (subexp != 0)
2152 relt = lookup (subexp,
2153 SAFE_HASH (subexp, GET_MODE (subexp)),
2154 GET_MODE (subexp));
2157 if (relt == 0)
2158 return 0;
2160 /* Search all related table entries for one that has an
2161 equivalent register. */
2163 p = relt;
2164 while (1)
2166 /* This loop is strange in that it is executed in two different cases.
2167 The first is when X is already in the table. Then it is searching
2168 the RELATED_VALUE list of X's class (RELT). The second case is when
2169 X is not in the table. Then RELT points to a class for the related
2170 value.
2172 Ensure that, whatever case we are in, that we ignore classes that have
2173 the same value as X. */
2175 if (rtx_equal_p (x, p->exp))
2176 q = 0;
2177 else
2178 for (q = p->first_same_value; q; q = q->next_same_value)
2179 if (REG_P (q->exp))
2180 break;
2182 if (q)
2183 break;
2185 p = p->related_value;
2187 /* We went all the way around, so there is nothing to be found.
2188 Alternatively, perhaps RELT was in the table for some other reason
2189 and it has no related values recorded. */
2190 if (p == relt || p == 0)
2191 break;
2194 if (q == 0)
2195 return 0;
2197 offset = (get_integer_term (x) - get_integer_term (p->exp));
2198 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2199 return plus_constant (q->mode, q->exp, offset);
2203 /* Hash a string. Just add its bytes up. */
2204 static inline unsigned
2205 hash_rtx_string (const char *ps)
2207 unsigned hash = 0;
2208 const unsigned char *p = (const unsigned char *) ps;
2210 if (p)
2211 while (*p)
2212 hash += *p++;
2214 return hash;
2217 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2218 When the callback returns true, we continue with the new rtx. */
2220 unsigned
2221 hash_rtx_cb (const_rtx x, machine_mode mode,
2222 int *do_not_record_p, int *hash_arg_in_memory_p,
2223 bool have_reg_qty, hash_rtx_callback_function cb)
2225 int i, j;
2226 unsigned hash = 0;
2227 enum rtx_code code;
2228 const char *fmt;
2229 machine_mode newmode;
2230 rtx newx;
2232 /* Used to turn recursion into iteration. We can't rely on GCC's
2233 tail-recursion elimination since we need to keep accumulating values
2234 in HASH. */
2235 repeat:
2236 if (x == 0)
2237 return hash;
2239 /* Invoke the callback first. */
2240 if (cb != NULL
2241 && ((*cb) (x, mode, &newx, &newmode)))
2243 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2244 hash_arg_in_memory_p, have_reg_qty, cb);
2245 return hash;
2248 code = GET_CODE (x);
2249 switch (code)
2251 case REG:
2253 unsigned int regno = REGNO (x);
2255 if (do_not_record_p && !reload_completed)
2257 /* On some machines, we can't record any non-fixed hard register,
2258 because extending its life will cause reload problems. We
2259 consider ap, fp, sp, gp to be fixed for this purpose.
2261 We also consider CCmode registers to be fixed for this purpose;
2262 failure to do so leads to failure to simplify 0<100 type of
2263 conditionals.
2265 On all machines, we can't record any global registers.
2266 Nor should we record any register that is in a small
2267 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2268 bool record;
2270 if (regno >= FIRST_PSEUDO_REGISTER)
2271 record = true;
2272 else if (x == frame_pointer_rtx
2273 || x == hard_frame_pointer_rtx
2274 || x == arg_pointer_rtx
2275 || x == stack_pointer_rtx
2276 || x == pic_offset_table_rtx)
2277 record = true;
2278 else if (global_regs[regno])
2279 record = false;
2280 else if (fixed_regs[regno])
2281 record = true;
2282 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2283 record = true;
2284 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2285 record = false;
2286 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2287 record = false;
2288 else
2289 record = true;
2291 if (!record)
2293 *do_not_record_p = 1;
2294 return 0;
2298 hash += ((unsigned int) REG << 7);
2299 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2300 return hash;
2303 /* We handle SUBREG of a REG specially because the underlying
2304 reg changes its hash value with every value change; we don't
2305 want to have to forget unrelated subregs when one subreg changes. */
2306 case SUBREG:
2308 if (REG_P (SUBREG_REG (x)))
2310 hash += (((unsigned int) SUBREG << 7)
2311 + REGNO (SUBREG_REG (x))
2312 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2313 return hash;
2315 break;
2318 case CONST_INT:
2319 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2320 + (unsigned int) INTVAL (x));
2321 return hash;
2323 case CONST_WIDE_INT:
2324 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2325 hash += CONST_WIDE_INT_ELT (x, i);
2326 return hash;
2328 case CONST_DOUBLE:
2329 /* This is like the general case, except that it only counts
2330 the integers representing the constant. */
2331 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2332 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2333 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2334 + (unsigned int) CONST_DOUBLE_HIGH (x));
2335 else
2336 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2337 return hash;
2339 case CONST_FIXED:
2340 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2341 hash += fixed_hash (CONST_FIXED_VALUE (x));
2342 return hash;
2344 case CONST_VECTOR:
2346 int units;
2347 rtx elt;
2349 units = CONST_VECTOR_NUNITS (x);
2351 for (i = 0; i < units; ++i)
2353 elt = CONST_VECTOR_ELT (x, i);
2354 hash += hash_rtx_cb (elt, GET_MODE (elt),
2355 do_not_record_p, hash_arg_in_memory_p,
2356 have_reg_qty, cb);
2359 return hash;
2362 /* Assume there is only one rtx object for any given label. */
2363 case LABEL_REF:
2364 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2365 differences and differences between each stage's debugging dumps. */
2366 hash += (((unsigned int) LABEL_REF << 7)
2367 + CODE_LABEL_NUMBER (LABEL_REF_LABEL (x)));
2368 return hash;
2370 case SYMBOL_REF:
2372 /* Don't hash on the symbol's address to avoid bootstrap differences.
2373 Different hash values may cause expressions to be recorded in
2374 different orders and thus different registers to be used in the
2375 final assembler. This also avoids differences in the dump files
2376 between various stages. */
2377 unsigned int h = 0;
2378 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2380 while (*p)
2381 h += (h << 7) + *p++; /* ??? revisit */
2383 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2384 return hash;
2387 case MEM:
2388 /* We don't record if marked volatile or if BLKmode since we don't
2389 know the size of the move. */
2390 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2392 *do_not_record_p = 1;
2393 return 0;
2395 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2396 *hash_arg_in_memory_p = 1;
2398 /* Now that we have already found this special case,
2399 might as well speed it up as much as possible. */
2400 hash += (unsigned) MEM;
2401 x = XEXP (x, 0);
2402 goto repeat;
2404 case USE:
2405 /* A USE that mentions non-volatile memory needs special
2406 handling since the MEM may be BLKmode which normally
2407 prevents an entry from being made. Pure calls are
2408 marked by a USE which mentions BLKmode memory.
2409 See calls.c:emit_call_1. */
2410 if (MEM_P (XEXP (x, 0))
2411 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2413 hash += (unsigned) USE;
2414 x = XEXP (x, 0);
2416 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2417 *hash_arg_in_memory_p = 1;
2419 /* Now that we have already found this special case,
2420 might as well speed it up as much as possible. */
2421 hash += (unsigned) MEM;
2422 x = XEXP (x, 0);
2423 goto repeat;
2425 break;
2427 case PRE_DEC:
2428 case PRE_INC:
2429 case POST_DEC:
2430 case POST_INC:
2431 case PRE_MODIFY:
2432 case POST_MODIFY:
2433 case PC:
2434 case CC0:
2435 case CALL:
2436 case UNSPEC_VOLATILE:
2437 if (do_not_record_p) {
2438 *do_not_record_p = 1;
2439 return 0;
2441 else
2442 return hash;
2443 break;
2445 case ASM_OPERANDS:
2446 if (do_not_record_p && MEM_VOLATILE_P (x))
2448 *do_not_record_p = 1;
2449 return 0;
2451 else
2453 /* We don't want to take the filename and line into account. */
2454 hash += (unsigned) code + (unsigned) GET_MODE (x)
2455 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2456 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2457 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2459 if (ASM_OPERANDS_INPUT_LENGTH (x))
2461 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2463 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2464 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2465 do_not_record_p, hash_arg_in_memory_p,
2466 have_reg_qty, cb)
2467 + hash_rtx_string
2468 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2471 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2472 x = ASM_OPERANDS_INPUT (x, 0);
2473 mode = GET_MODE (x);
2474 goto repeat;
2477 return hash;
2479 break;
2481 default:
2482 break;
2485 i = GET_RTX_LENGTH (code) - 1;
2486 hash += (unsigned) code + (unsigned) GET_MODE (x);
2487 fmt = GET_RTX_FORMAT (code);
2488 for (; i >= 0; i--)
2490 switch (fmt[i])
2492 case 'e':
2493 /* If we are about to do the last recursive call
2494 needed at this level, change it into iteration.
2495 This function is called enough to be worth it. */
2496 if (i == 0)
2498 x = XEXP (x, i);
2499 goto repeat;
2502 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2503 hash_arg_in_memory_p,
2504 have_reg_qty, cb);
2505 break;
2507 case 'E':
2508 for (j = 0; j < XVECLEN (x, i); j++)
2509 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2510 hash_arg_in_memory_p,
2511 have_reg_qty, cb);
2512 break;
2514 case 's':
2515 hash += hash_rtx_string (XSTR (x, i));
2516 break;
2518 case 'i':
2519 hash += (unsigned int) XINT (x, i);
2520 break;
2522 case '0': case 't':
2523 /* Unused. */
2524 break;
2526 default:
2527 gcc_unreachable ();
2531 return hash;
2534 /* Hash an rtx. We are careful to make sure the value is never negative.
2535 Equivalent registers hash identically.
2536 MODE is used in hashing for CONST_INTs only;
2537 otherwise the mode of X is used.
2539 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2541 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2542 a MEM rtx which does not have the MEM_READONLY_P flag set.
2544 Note that cse_insn knows that the hash code of a MEM expression
2545 is just (int) MEM plus the hash code of the address. */
2547 unsigned
2548 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2549 int *hash_arg_in_memory_p, bool have_reg_qty)
2551 return hash_rtx_cb (x, mode, do_not_record_p,
2552 hash_arg_in_memory_p, have_reg_qty, NULL);
2555 /* Hash an rtx X for cse via hash_rtx.
2556 Stores 1 in do_not_record if any subexpression is volatile.
2557 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2558 does not have the MEM_READONLY_P flag set. */
2560 static inline unsigned
2561 canon_hash (rtx x, machine_mode mode)
2563 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2566 /* Like canon_hash but with no side effects, i.e. do_not_record
2567 and hash_arg_in_memory are not changed. */
2569 static inline unsigned
2570 safe_hash (rtx x, machine_mode mode)
2572 int dummy_do_not_record;
2573 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2576 /* Return 1 iff X and Y would canonicalize into the same thing,
2577 without actually constructing the canonicalization of either one.
2578 If VALIDATE is nonzero,
2579 we assume X is an expression being processed from the rtl
2580 and Y was found in the hash table. We check register refs
2581 in Y for being marked as valid.
2583 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2586 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2588 int i, j;
2589 enum rtx_code code;
2590 const char *fmt;
2592 /* Note: it is incorrect to assume an expression is equivalent to itself
2593 if VALIDATE is nonzero. */
2594 if (x == y && !validate)
2595 return 1;
2597 if (x == 0 || y == 0)
2598 return x == y;
2600 code = GET_CODE (x);
2601 if (code != GET_CODE (y))
2602 return 0;
2604 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2605 if (GET_MODE (x) != GET_MODE (y))
2606 return 0;
2608 /* MEMs referring to different address space are not equivalent. */
2609 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2610 return 0;
2612 switch (code)
2614 case PC:
2615 case CC0:
2616 CASE_CONST_UNIQUE:
2617 return x == y;
2619 case LABEL_REF:
2620 return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y);
2622 case SYMBOL_REF:
2623 return XSTR (x, 0) == XSTR (y, 0);
2625 case REG:
2626 if (for_gcse)
2627 return REGNO (x) == REGNO (y);
2628 else
2630 unsigned int regno = REGNO (y);
2631 unsigned int i;
2632 unsigned int endregno = END_REGNO (y);
2634 /* If the quantities are not the same, the expressions are not
2635 equivalent. If there are and we are not to validate, they
2636 are equivalent. Otherwise, ensure all regs are up-to-date. */
2638 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2639 return 0;
2641 if (! validate)
2642 return 1;
2644 for (i = regno; i < endregno; i++)
2645 if (REG_IN_TABLE (i) != REG_TICK (i))
2646 return 0;
2648 return 1;
2651 case MEM:
2652 if (for_gcse)
2654 /* A volatile mem should not be considered equivalent to any
2655 other. */
2656 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2657 return 0;
2659 /* Can't merge two expressions in different alias sets, since we
2660 can decide that the expression is transparent in a block when
2661 it isn't, due to it being set with the different alias set.
2663 Also, can't merge two expressions with different MEM_ATTRS.
2664 They could e.g. be two different entities allocated into the
2665 same space on the stack (see e.g. PR25130). In that case, the
2666 MEM addresses can be the same, even though the two MEMs are
2667 absolutely not equivalent.
2669 But because really all MEM attributes should be the same for
2670 equivalent MEMs, we just use the invariant that MEMs that have
2671 the same attributes share the same mem_attrs data structure. */
2672 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2673 return 0;
2675 /* If we are handling exceptions, we cannot consider two expressions
2676 with different trapping status as equivalent, because simple_mem
2677 might accept one and reject the other. */
2678 if (cfun->can_throw_non_call_exceptions
2679 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2680 return 0;
2682 break;
2684 /* For commutative operations, check both orders. */
2685 case PLUS:
2686 case MULT:
2687 case AND:
2688 case IOR:
2689 case XOR:
2690 case NE:
2691 case EQ:
2692 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2693 validate, for_gcse)
2694 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2695 validate, for_gcse))
2696 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2697 validate, for_gcse)
2698 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2699 validate, for_gcse)));
2701 case ASM_OPERANDS:
2702 /* We don't use the generic code below because we want to
2703 disregard filename and line numbers. */
2705 /* A volatile asm isn't equivalent to any other. */
2706 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2707 return 0;
2709 if (GET_MODE (x) != GET_MODE (y)
2710 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2711 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2712 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2713 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2714 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2715 return 0;
2717 if (ASM_OPERANDS_INPUT_LENGTH (x))
2719 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2720 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2721 ASM_OPERANDS_INPUT (y, i),
2722 validate, for_gcse)
2723 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2724 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2725 return 0;
2728 return 1;
2730 default:
2731 break;
2734 /* Compare the elements. If any pair of corresponding elements
2735 fail to match, return 0 for the whole thing. */
2737 fmt = GET_RTX_FORMAT (code);
2738 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2740 switch (fmt[i])
2742 case 'e':
2743 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2744 validate, for_gcse))
2745 return 0;
2746 break;
2748 case 'E':
2749 if (XVECLEN (x, i) != XVECLEN (y, i))
2750 return 0;
2751 for (j = 0; j < XVECLEN (x, i); j++)
2752 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2753 validate, for_gcse))
2754 return 0;
2755 break;
2757 case 's':
2758 if (strcmp (XSTR (x, i), XSTR (y, i)))
2759 return 0;
2760 break;
2762 case 'i':
2763 if (XINT (x, i) != XINT (y, i))
2764 return 0;
2765 break;
2767 case 'w':
2768 if (XWINT (x, i) != XWINT (y, i))
2769 return 0;
2770 break;
2772 case '0':
2773 case 't':
2774 break;
2776 default:
2777 gcc_unreachable ();
2781 return 1;
2784 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2785 the result if necessary. INSN is as for canon_reg. */
2787 static void
2788 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2790 if (*xloc)
2792 rtx new_rtx = canon_reg (*xloc, insn);
2794 /* If replacing pseudo with hard reg or vice versa, ensure the
2795 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2796 gcc_assert (insn && new_rtx);
2797 validate_change (insn, xloc, new_rtx, 1);
2801 /* Canonicalize an expression:
2802 replace each register reference inside it
2803 with the "oldest" equivalent register.
2805 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2806 after we make our substitution. The calls are made with IN_GROUP nonzero
2807 so apply_change_group must be called upon the outermost return from this
2808 function (unless INSN is zero). The result of apply_change_group can
2809 generally be discarded since the changes we are making are optional. */
2811 static rtx
2812 canon_reg (rtx x, rtx_insn *insn)
2814 int i;
2815 enum rtx_code code;
2816 const char *fmt;
2818 if (x == 0)
2819 return x;
2821 code = GET_CODE (x);
2822 switch (code)
2824 case PC:
2825 case CC0:
2826 case CONST:
2827 CASE_CONST_ANY:
2828 case SYMBOL_REF:
2829 case LABEL_REF:
2830 case ADDR_VEC:
2831 case ADDR_DIFF_VEC:
2832 return x;
2834 case REG:
2836 int first;
2837 int q;
2838 struct qty_table_elem *ent;
2840 /* Never replace a hard reg, because hard regs can appear
2841 in more than one machine mode, and we must preserve the mode
2842 of each occurrence. Also, some hard regs appear in
2843 MEMs that are shared and mustn't be altered. Don't try to
2844 replace any reg that maps to a reg of class NO_REGS. */
2845 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2846 || ! REGNO_QTY_VALID_P (REGNO (x)))
2847 return x;
2849 q = REG_QTY (REGNO (x));
2850 ent = &qty_table[q];
2851 first = ent->first_reg;
2852 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2853 : REGNO_REG_CLASS (first) == NO_REGS ? x
2854 : gen_rtx_REG (ent->mode, first));
2857 default:
2858 break;
2861 fmt = GET_RTX_FORMAT (code);
2862 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2864 int j;
2866 if (fmt[i] == 'e')
2867 validate_canon_reg (&XEXP (x, i), insn);
2868 else if (fmt[i] == 'E')
2869 for (j = 0; j < XVECLEN (x, i); j++)
2870 validate_canon_reg (&XVECEXP (x, i, j), insn);
2873 return x;
2876 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2877 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2878 what values are being compared.
2880 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2881 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2882 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2883 compared to produce cc0.
2885 The return value is the comparison operator and is either the code of
2886 A or the code corresponding to the inverse of the comparison. */
2888 static enum rtx_code
2889 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2890 machine_mode *pmode1, machine_mode *pmode2)
2892 rtx arg1, arg2;
2893 hash_set<rtx> *visited = NULL;
2894 /* Set nonzero when we find something of interest. */
2895 rtx x = NULL;
2897 arg1 = *parg1, arg2 = *parg2;
2899 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2901 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2903 int reverse_code = 0;
2904 struct table_elt *p = 0;
2906 /* Remember state from previous iteration. */
2907 if (x)
2909 if (!visited)
2910 visited = new hash_set<rtx>;
2911 visited->add (x);
2912 x = 0;
2915 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2916 On machines with CC0, this is the only case that can occur, since
2917 fold_rtx will return the COMPARE or item being compared with zero
2918 when given CC0. */
2920 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2921 x = arg1;
2923 /* If ARG1 is a comparison operator and CODE is testing for
2924 STORE_FLAG_VALUE, get the inner arguments. */
2926 else if (COMPARISON_P (arg1))
2928 #ifdef FLOAT_STORE_FLAG_VALUE
2929 REAL_VALUE_TYPE fsfv;
2930 #endif
2932 if (code == NE
2933 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2934 && code == LT && STORE_FLAG_VALUE == -1)
2935 #ifdef FLOAT_STORE_FLAG_VALUE
2936 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2937 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2938 REAL_VALUE_NEGATIVE (fsfv)))
2939 #endif
2941 x = arg1;
2942 else if (code == EQ
2943 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2944 && code == GE && STORE_FLAG_VALUE == -1)
2945 #ifdef FLOAT_STORE_FLAG_VALUE
2946 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2947 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2948 REAL_VALUE_NEGATIVE (fsfv)))
2949 #endif
2951 x = arg1, reverse_code = 1;
2954 /* ??? We could also check for
2956 (ne (and (eq (...) (const_int 1))) (const_int 0))
2958 and related forms, but let's wait until we see them occurring. */
2960 if (x == 0)
2961 /* Look up ARG1 in the hash table and see if it has an equivalence
2962 that lets us see what is being compared. */
2963 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2964 if (p)
2966 p = p->first_same_value;
2968 /* If what we compare is already known to be constant, that is as
2969 good as it gets.
2970 We need to break the loop in this case, because otherwise we
2971 can have an infinite loop when looking at a reg that is known
2972 to be a constant which is the same as a comparison of a reg
2973 against zero which appears later in the insn stream, which in
2974 turn is constant and the same as the comparison of the first reg
2975 against zero... */
2976 if (p->is_const)
2977 break;
2980 for (; p; p = p->next_same_value)
2982 machine_mode inner_mode = GET_MODE (p->exp);
2983 #ifdef FLOAT_STORE_FLAG_VALUE
2984 REAL_VALUE_TYPE fsfv;
2985 #endif
2987 /* If the entry isn't valid, skip it. */
2988 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2989 continue;
2991 /* If it's a comparison we've used before, skip it. */
2992 if (visited && visited->contains (p->exp))
2993 continue;
2995 if (GET_CODE (p->exp) == COMPARE
2996 /* Another possibility is that this machine has a compare insn
2997 that includes the comparison code. In that case, ARG1 would
2998 be equivalent to a comparison operation that would set ARG1 to
2999 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3000 ORIG_CODE is the actual comparison being done; if it is an EQ,
3001 we must reverse ORIG_CODE. On machine with a negative value
3002 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3003 || ((code == NE
3004 || (code == LT
3005 && val_signbit_known_set_p (inner_mode,
3006 STORE_FLAG_VALUE))
3007 #ifdef FLOAT_STORE_FLAG_VALUE
3008 || (code == LT
3009 && SCALAR_FLOAT_MODE_P (inner_mode)
3010 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3011 REAL_VALUE_NEGATIVE (fsfv)))
3012 #endif
3014 && COMPARISON_P (p->exp)))
3016 x = p->exp;
3017 break;
3019 else if ((code == EQ
3020 || (code == GE
3021 && val_signbit_known_set_p (inner_mode,
3022 STORE_FLAG_VALUE))
3023 #ifdef FLOAT_STORE_FLAG_VALUE
3024 || (code == GE
3025 && SCALAR_FLOAT_MODE_P (inner_mode)
3026 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3027 REAL_VALUE_NEGATIVE (fsfv)))
3028 #endif
3030 && COMPARISON_P (p->exp))
3032 reverse_code = 1;
3033 x = p->exp;
3034 break;
3037 /* If this non-trapping address, e.g. fp + constant, the
3038 equivalent is a better operand since it may let us predict
3039 the value of the comparison. */
3040 else if (!rtx_addr_can_trap_p (p->exp))
3042 arg1 = p->exp;
3043 continue;
3047 /* If we didn't find a useful equivalence for ARG1, we are done.
3048 Otherwise, set up for the next iteration. */
3049 if (x == 0)
3050 break;
3052 /* If we need to reverse the comparison, make sure that is
3053 possible -- we can't necessarily infer the value of GE from LT
3054 with floating-point operands. */
3055 if (reverse_code)
3057 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3058 if (reversed == UNKNOWN)
3059 break;
3060 else
3061 code = reversed;
3063 else if (COMPARISON_P (x))
3064 code = GET_CODE (x);
3065 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3068 /* Return our results. Return the modes from before fold_rtx
3069 because fold_rtx might produce const_int, and then it's too late. */
3070 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3071 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3073 if (visited)
3074 delete visited;
3075 return code;
3078 /* If X is a nontrivial arithmetic operation on an argument for which
3079 a constant value can be determined, return the result of operating
3080 on that value, as a constant. Otherwise, return X, possibly with
3081 one or more operands changed to a forward-propagated constant.
3083 If X is a register whose contents are known, we do NOT return
3084 those contents here; equiv_constant is called to perform that task.
3085 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3087 INSN is the insn that we may be modifying. If it is 0, make a copy
3088 of X before modifying it. */
3090 static rtx
3091 fold_rtx (rtx x, rtx_insn *insn)
3093 enum rtx_code code;
3094 machine_mode mode;
3095 const char *fmt;
3096 int i;
3097 rtx new_rtx = 0;
3098 int changed = 0;
3100 /* Operands of X. */
3101 /* Workaround -Wmaybe-uninitialized false positive during
3102 profiledbootstrap by initializing them. */
3103 rtx folded_arg0 = NULL_RTX;
3104 rtx folded_arg1 = NULL_RTX;
3106 /* Constant equivalents of first three operands of X;
3107 0 when no such equivalent is known. */
3108 rtx const_arg0;
3109 rtx const_arg1;
3110 rtx const_arg2;
3112 /* The mode of the first operand of X. We need this for sign and zero
3113 extends. */
3114 machine_mode mode_arg0;
3116 if (x == 0)
3117 return x;
3119 /* Try to perform some initial simplifications on X. */
3120 code = GET_CODE (x);
3121 switch (code)
3123 case MEM:
3124 case SUBREG:
3125 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3126 than it would in other contexts. Basically its mode does not
3127 signify the size of the object read. That information is carried
3128 by size operand. If we happen to have a MEM of the appropriate
3129 mode in our tables with a constant value we could simplify the
3130 extraction incorrectly if we allowed substitution of that value
3131 for the MEM. */
3132 case ZERO_EXTRACT:
3133 case SIGN_EXTRACT:
3134 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3135 return new_rtx;
3136 return x;
3138 case CONST:
3139 CASE_CONST_ANY:
3140 case SYMBOL_REF:
3141 case LABEL_REF:
3142 case REG:
3143 case PC:
3144 /* No use simplifying an EXPR_LIST
3145 since they are used only for lists of args
3146 in a function call's REG_EQUAL note. */
3147 case EXPR_LIST:
3148 return x;
3150 case CC0:
3151 return prev_insn_cc0;
3153 case ASM_OPERANDS:
3154 if (insn)
3156 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3157 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3158 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3160 return x;
3162 case CALL:
3163 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3164 return x;
3165 break;
3167 /* Anything else goes through the loop below. */
3168 default:
3169 break;
3172 mode = GET_MODE (x);
3173 const_arg0 = 0;
3174 const_arg1 = 0;
3175 const_arg2 = 0;
3176 mode_arg0 = VOIDmode;
3178 /* Try folding our operands.
3179 Then see which ones have constant values known. */
3181 fmt = GET_RTX_FORMAT (code);
3182 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3183 if (fmt[i] == 'e')
3185 rtx folded_arg = XEXP (x, i), const_arg;
3186 machine_mode mode_arg = GET_MODE (folded_arg);
3188 switch (GET_CODE (folded_arg))
3190 case MEM:
3191 case REG:
3192 case SUBREG:
3193 const_arg = equiv_constant (folded_arg);
3194 break;
3196 case CONST:
3197 CASE_CONST_ANY:
3198 case SYMBOL_REF:
3199 case LABEL_REF:
3200 const_arg = folded_arg;
3201 break;
3203 case CC0:
3204 /* The cc0-user and cc0-setter may be in different blocks if
3205 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3206 will have been cleared as we exited the block with the
3207 setter.
3209 While we could potentially track cc0 in this case, it just
3210 doesn't seem to be worth it given that cc0 targets are not
3211 terribly common or important these days and trapping math
3212 is rarely used. The combination of those two conditions
3213 necessary to trip this situation is exceedingly rare in the
3214 real world. */
3215 if (!prev_insn_cc0)
3217 const_arg = NULL_RTX;
3219 else
3221 folded_arg = prev_insn_cc0;
3222 mode_arg = prev_insn_cc0_mode;
3223 const_arg = equiv_constant (folded_arg);
3225 break;
3227 default:
3228 folded_arg = fold_rtx (folded_arg, insn);
3229 const_arg = equiv_constant (folded_arg);
3230 break;
3233 /* For the first three operands, see if the operand
3234 is constant or equivalent to a constant. */
3235 switch (i)
3237 case 0:
3238 folded_arg0 = folded_arg;
3239 const_arg0 = const_arg;
3240 mode_arg0 = mode_arg;
3241 break;
3242 case 1:
3243 folded_arg1 = folded_arg;
3244 const_arg1 = const_arg;
3245 break;
3246 case 2:
3247 const_arg2 = const_arg;
3248 break;
3251 /* Pick the least expensive of the argument and an equivalent constant
3252 argument. */
3253 if (const_arg != 0
3254 && const_arg != folded_arg
3255 && (COST_IN (const_arg, mode_arg, code, i)
3256 <= COST_IN (folded_arg, mode_arg, code, i))
3258 /* It's not safe to substitute the operand of a conversion
3259 operator with a constant, as the conversion's identity
3260 depends upon the mode of its operand. This optimization
3261 is handled by the call to simplify_unary_operation. */
3262 && (GET_RTX_CLASS (code) != RTX_UNARY
3263 || GET_MODE (const_arg) == mode_arg0
3264 || (code != ZERO_EXTEND
3265 && code != SIGN_EXTEND
3266 && code != TRUNCATE
3267 && code != FLOAT_TRUNCATE
3268 && code != FLOAT_EXTEND
3269 && code != FLOAT
3270 && code != FIX
3271 && code != UNSIGNED_FLOAT
3272 && code != UNSIGNED_FIX)))
3273 folded_arg = const_arg;
3275 if (folded_arg == XEXP (x, i))
3276 continue;
3278 if (insn == NULL_RTX && !changed)
3279 x = copy_rtx (x);
3280 changed = 1;
3281 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3284 if (changed)
3286 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3287 consistent with the order in X. */
3288 if (canonicalize_change_group (insn, x))
3290 std::swap (const_arg0, const_arg1);
3291 std::swap (folded_arg0, folded_arg1);
3294 apply_change_group ();
3297 /* If X is an arithmetic operation, see if we can simplify it. */
3299 switch (GET_RTX_CLASS (code))
3301 case RTX_UNARY:
3303 /* We can't simplify extension ops unless we know the
3304 original mode. */
3305 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3306 && mode_arg0 == VOIDmode)
3307 break;
3309 new_rtx = simplify_unary_operation (code, mode,
3310 const_arg0 ? const_arg0 : folded_arg0,
3311 mode_arg0);
3313 break;
3315 case RTX_COMPARE:
3316 case RTX_COMM_COMPARE:
3317 /* See what items are actually being compared and set FOLDED_ARG[01]
3318 to those values and CODE to the actual comparison code. If any are
3319 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3320 do anything if both operands are already known to be constant. */
3322 /* ??? Vector mode comparisons are not supported yet. */
3323 if (VECTOR_MODE_P (mode))
3324 break;
3326 if (const_arg0 == 0 || const_arg1 == 0)
3328 struct table_elt *p0, *p1;
3329 rtx true_rtx, false_rtx;
3330 machine_mode mode_arg1;
3332 if (SCALAR_FLOAT_MODE_P (mode))
3334 #ifdef FLOAT_STORE_FLAG_VALUE
3335 true_rtx = (const_double_from_real_value
3336 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3337 #else
3338 true_rtx = NULL_RTX;
3339 #endif
3340 false_rtx = CONST0_RTX (mode);
3342 else
3344 true_rtx = const_true_rtx;
3345 false_rtx = const0_rtx;
3348 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3349 &mode_arg0, &mode_arg1);
3351 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3352 what kinds of things are being compared, so we can't do
3353 anything with this comparison. */
3355 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3356 break;
3358 const_arg0 = equiv_constant (folded_arg0);
3359 const_arg1 = equiv_constant (folded_arg1);
3361 /* If we do not now have two constants being compared, see
3362 if we can nevertheless deduce some things about the
3363 comparison. */
3364 if (const_arg0 == 0 || const_arg1 == 0)
3366 if (const_arg1 != NULL)
3368 rtx cheapest_simplification;
3369 int cheapest_cost;
3370 rtx simp_result;
3371 struct table_elt *p;
3373 /* See if we can find an equivalent of folded_arg0
3374 that gets us a cheaper expression, possibly a
3375 constant through simplifications. */
3376 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3377 mode_arg0);
3379 if (p != NULL)
3381 cheapest_simplification = x;
3382 cheapest_cost = COST (x, mode);
3384 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3386 int cost;
3388 /* If the entry isn't valid, skip it. */
3389 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3390 continue;
3392 /* Try to simplify using this equivalence. */
3393 simp_result
3394 = simplify_relational_operation (code, mode,
3395 mode_arg0,
3396 p->exp,
3397 const_arg1);
3399 if (simp_result == NULL)
3400 continue;
3402 cost = COST (simp_result, mode);
3403 if (cost < cheapest_cost)
3405 cheapest_cost = cost;
3406 cheapest_simplification = simp_result;
3410 /* If we have a cheaper expression now, use that
3411 and try folding it further, from the top. */
3412 if (cheapest_simplification != x)
3413 return fold_rtx (copy_rtx (cheapest_simplification),
3414 insn);
3418 /* See if the two operands are the same. */
3420 if ((REG_P (folded_arg0)
3421 && REG_P (folded_arg1)
3422 && (REG_QTY (REGNO (folded_arg0))
3423 == REG_QTY (REGNO (folded_arg1))))
3424 || ((p0 = lookup (folded_arg0,
3425 SAFE_HASH (folded_arg0, mode_arg0),
3426 mode_arg0))
3427 && (p1 = lookup (folded_arg1,
3428 SAFE_HASH (folded_arg1, mode_arg0),
3429 mode_arg0))
3430 && p0->first_same_value == p1->first_same_value))
3431 folded_arg1 = folded_arg0;
3433 /* If FOLDED_ARG0 is a register, see if the comparison we are
3434 doing now is either the same as we did before or the reverse
3435 (we only check the reverse if not floating-point). */
3436 else if (REG_P (folded_arg0))
3438 int qty = REG_QTY (REGNO (folded_arg0));
3440 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3442 struct qty_table_elem *ent = &qty_table[qty];
3444 if ((comparison_dominates_p (ent->comparison_code, code)
3445 || (! FLOAT_MODE_P (mode_arg0)
3446 && comparison_dominates_p (ent->comparison_code,
3447 reverse_condition (code))))
3448 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3449 || (const_arg1
3450 && rtx_equal_p (ent->comparison_const,
3451 const_arg1))
3452 || (REG_P (folded_arg1)
3453 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3455 if (comparison_dominates_p (ent->comparison_code, code))
3457 if (true_rtx)
3458 return true_rtx;
3459 else
3460 break;
3462 else
3463 return false_rtx;
3470 /* If we are comparing against zero, see if the first operand is
3471 equivalent to an IOR with a constant. If so, we may be able to
3472 determine the result of this comparison. */
3473 if (const_arg1 == const0_rtx && !const_arg0)
3475 rtx y = lookup_as_function (folded_arg0, IOR);
3476 rtx inner_const;
3478 if (y != 0
3479 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3480 && CONST_INT_P (inner_const)
3481 && INTVAL (inner_const) != 0)
3482 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3486 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3487 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3488 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3489 op0, op1);
3491 break;
3493 case RTX_BIN_ARITH:
3494 case RTX_COMM_ARITH:
3495 switch (code)
3497 case PLUS:
3498 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3499 with that LABEL_REF as its second operand. If so, the result is
3500 the first operand of that MINUS. This handles switches with an
3501 ADDR_DIFF_VEC table. */
3502 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3504 rtx y
3505 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3506 : lookup_as_function (folded_arg0, MINUS);
3508 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3509 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg1))
3510 return XEXP (y, 0);
3512 /* Now try for a CONST of a MINUS like the above. */
3513 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3514 : lookup_as_function (folded_arg0, CONST))) != 0
3515 && GET_CODE (XEXP (y, 0)) == MINUS
3516 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3517 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg1))
3518 return XEXP (XEXP (y, 0), 0);
3521 /* Likewise if the operands are in the other order. */
3522 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3524 rtx y
3525 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3526 : lookup_as_function (folded_arg1, MINUS);
3528 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3529 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg0))
3530 return XEXP (y, 0);
3532 /* Now try for a CONST of a MINUS like the above. */
3533 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3534 : lookup_as_function (folded_arg1, CONST))) != 0
3535 && GET_CODE (XEXP (y, 0)) == MINUS
3536 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3537 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg0))
3538 return XEXP (XEXP (y, 0), 0);
3541 /* If second operand is a register equivalent to a negative
3542 CONST_INT, see if we can find a register equivalent to the
3543 positive constant. Make a MINUS if so. Don't do this for
3544 a non-negative constant since we might then alternate between
3545 choosing positive and negative constants. Having the positive
3546 constant previously-used is the more common case. Be sure
3547 the resulting constant is non-negative; if const_arg1 were
3548 the smallest negative number this would overflow: depending
3549 on the mode, this would either just be the same value (and
3550 hence not save anything) or be incorrect. */
3551 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3552 && INTVAL (const_arg1) < 0
3553 /* This used to test
3555 -INTVAL (const_arg1) >= 0
3557 But The Sun V5.0 compilers mis-compiled that test. So
3558 instead we test for the problematic value in a more direct
3559 manner and hope the Sun compilers get it correct. */
3560 && INTVAL (const_arg1) !=
3561 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3562 && REG_P (folded_arg1))
3564 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3565 struct table_elt *p
3566 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3568 if (p)
3569 for (p = p->first_same_value; p; p = p->next_same_value)
3570 if (REG_P (p->exp))
3571 return simplify_gen_binary (MINUS, mode, folded_arg0,
3572 canon_reg (p->exp, NULL));
3574 goto from_plus;
3576 case MINUS:
3577 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3578 If so, produce (PLUS Z C2-C). */
3579 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3581 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3582 if (y && CONST_INT_P (XEXP (y, 1)))
3583 return fold_rtx (plus_constant (mode, copy_rtx (y),
3584 -INTVAL (const_arg1)),
3585 NULL);
3588 /* Fall through. */
3590 from_plus:
3591 case SMIN: case SMAX: case UMIN: case UMAX:
3592 case IOR: case AND: case XOR:
3593 case MULT:
3594 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3595 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3596 is known to be of similar form, we may be able to replace the
3597 operation with a combined operation. This may eliminate the
3598 intermediate operation if every use is simplified in this way.
3599 Note that the similar optimization done by combine.c only works
3600 if the intermediate operation's result has only one reference. */
3602 if (REG_P (folded_arg0)
3603 && const_arg1 && CONST_INT_P (const_arg1))
3605 int is_shift
3606 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3607 rtx y, inner_const, new_const;
3608 rtx canon_const_arg1 = const_arg1;
3609 enum rtx_code associate_code;
3611 if (is_shift
3612 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3613 || INTVAL (const_arg1) < 0))
3615 if (SHIFT_COUNT_TRUNCATED)
3616 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3617 & (GET_MODE_BITSIZE (mode)
3618 - 1));
3619 else
3620 break;
3623 y = lookup_as_function (folded_arg0, code);
3624 if (y == 0)
3625 break;
3627 /* If we have compiled a statement like
3628 "if (x == (x & mask1))", and now are looking at
3629 "x & mask2", we will have a case where the first operand
3630 of Y is the same as our first operand. Unless we detect
3631 this case, an infinite loop will result. */
3632 if (XEXP (y, 0) == folded_arg0)
3633 break;
3635 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3636 if (!inner_const || !CONST_INT_P (inner_const))
3637 break;
3639 /* Don't associate these operations if they are a PLUS with the
3640 same constant and it is a power of two. These might be doable
3641 with a pre- or post-increment. Similarly for two subtracts of
3642 identical powers of two with post decrement. */
3644 if (code == PLUS && const_arg1 == inner_const
3645 && ((HAVE_PRE_INCREMENT
3646 && exact_log2 (INTVAL (const_arg1)) >= 0)
3647 || (HAVE_POST_INCREMENT
3648 && exact_log2 (INTVAL (const_arg1)) >= 0)
3649 || (HAVE_PRE_DECREMENT
3650 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3651 || (HAVE_POST_DECREMENT
3652 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3653 break;
3655 /* ??? Vector mode shifts by scalar
3656 shift operand are not supported yet. */
3657 if (is_shift && VECTOR_MODE_P (mode))
3658 break;
3660 if (is_shift
3661 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3662 || INTVAL (inner_const) < 0))
3664 if (SHIFT_COUNT_TRUNCATED)
3665 inner_const = GEN_INT (INTVAL (inner_const)
3666 & (GET_MODE_BITSIZE (mode) - 1));
3667 else
3668 break;
3671 /* Compute the code used to compose the constants. For example,
3672 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3674 associate_code = (is_shift || code == MINUS ? PLUS : code);
3676 new_const = simplify_binary_operation (associate_code, mode,
3677 canon_const_arg1,
3678 inner_const);
3680 if (new_const == 0)
3681 break;
3683 /* If we are associating shift operations, don't let this
3684 produce a shift of the size of the object or larger.
3685 This could occur when we follow a sign-extend by a right
3686 shift on a machine that does a sign-extend as a pair
3687 of shifts. */
3689 if (is_shift
3690 && CONST_INT_P (new_const)
3691 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3693 /* As an exception, we can turn an ASHIFTRT of this
3694 form into a shift of the number of bits - 1. */
3695 if (code == ASHIFTRT)
3696 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3697 else if (!side_effects_p (XEXP (y, 0)))
3698 return CONST0_RTX (mode);
3699 else
3700 break;
3703 y = copy_rtx (XEXP (y, 0));
3705 /* If Y contains our first operand (the most common way this
3706 can happen is if Y is a MEM), we would do into an infinite
3707 loop if we tried to fold it. So don't in that case. */
3709 if (! reg_mentioned_p (folded_arg0, y))
3710 y = fold_rtx (y, insn);
3712 return simplify_gen_binary (code, mode, y, new_const);
3714 break;
3716 case DIV: case UDIV:
3717 /* ??? The associative optimization performed immediately above is
3718 also possible for DIV and UDIV using associate_code of MULT.
3719 However, we would need extra code to verify that the
3720 multiplication does not overflow, that is, there is no overflow
3721 in the calculation of new_const. */
3722 break;
3724 default:
3725 break;
3728 new_rtx = simplify_binary_operation (code, mode,
3729 const_arg0 ? const_arg0 : folded_arg0,
3730 const_arg1 ? const_arg1 : folded_arg1);
3731 break;
3733 case RTX_OBJ:
3734 /* (lo_sum (high X) X) is simply X. */
3735 if (code == LO_SUM && const_arg0 != 0
3736 && GET_CODE (const_arg0) == HIGH
3737 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3738 return const_arg1;
3739 break;
3741 case RTX_TERNARY:
3742 case RTX_BITFIELD_OPS:
3743 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3744 const_arg0 ? const_arg0 : folded_arg0,
3745 const_arg1 ? const_arg1 : folded_arg1,
3746 const_arg2 ? const_arg2 : XEXP (x, 2));
3747 break;
3749 default:
3750 break;
3753 return new_rtx ? new_rtx : x;
3756 /* Return a constant value currently equivalent to X.
3757 Return 0 if we don't know one. */
3759 static rtx
3760 equiv_constant (rtx x)
3762 if (REG_P (x)
3763 && REGNO_QTY_VALID_P (REGNO (x)))
3765 int x_q = REG_QTY (REGNO (x));
3766 struct qty_table_elem *x_ent = &qty_table[x_q];
3768 if (x_ent->const_rtx)
3769 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3772 if (x == 0 || CONSTANT_P (x))
3773 return x;
3775 if (GET_CODE (x) == SUBREG)
3777 machine_mode mode = GET_MODE (x);
3778 machine_mode imode = GET_MODE (SUBREG_REG (x));
3779 rtx new_rtx;
3781 /* See if we previously assigned a constant value to this SUBREG. */
3782 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3783 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3784 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3785 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3786 return new_rtx;
3788 /* If we didn't and if doing so makes sense, see if we previously
3789 assigned a constant value to the enclosing word mode SUBREG. */
3790 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3791 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3793 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3794 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3796 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3797 new_rtx = lookup_as_function (y, CONST_INT);
3798 if (new_rtx)
3799 return gen_lowpart (mode, new_rtx);
3803 /* Otherwise see if we already have a constant for the inner REG,
3804 and if that is enough to calculate an equivalent constant for
3805 the subreg. Note that the upper bits of paradoxical subregs
3806 are undefined, so they cannot be said to equal anything. */
3807 if (REG_P (SUBREG_REG (x))
3808 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3809 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3810 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3812 return 0;
3815 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3816 the hash table in case its value was seen before. */
3818 if (MEM_P (x))
3820 struct table_elt *elt;
3822 x = avoid_constant_pool_reference (x);
3823 if (CONSTANT_P (x))
3824 return x;
3826 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3827 if (elt == 0)
3828 return 0;
3830 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3831 if (elt->is_const && CONSTANT_P (elt->exp))
3832 return elt->exp;
3835 return 0;
3838 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3839 "taken" branch.
3841 In certain cases, this can cause us to add an equivalence. For example,
3842 if we are following the taken case of
3843 if (i == 2)
3844 we can add the fact that `i' and '2' are now equivalent.
3846 In any case, we can record that this comparison was passed. If the same
3847 comparison is seen later, we will know its value. */
3849 static void
3850 record_jump_equiv (rtx_insn *insn, bool taken)
3852 int cond_known_true;
3853 rtx op0, op1;
3854 rtx set;
3855 machine_mode mode, mode0, mode1;
3856 int reversed_nonequality = 0;
3857 enum rtx_code code;
3859 /* Ensure this is the right kind of insn. */
3860 gcc_assert (any_condjump_p (insn));
3862 set = pc_set (insn);
3864 /* See if this jump condition is known true or false. */
3865 if (taken)
3866 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3867 else
3868 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3870 /* Get the type of comparison being done and the operands being compared.
3871 If we had to reverse a non-equality condition, record that fact so we
3872 know that it isn't valid for floating-point. */
3873 code = GET_CODE (XEXP (SET_SRC (set), 0));
3874 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3875 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3877 /* On a cc0 target the cc0-setter and cc0-user may end up in different
3878 blocks. When that happens the tracking of the cc0-setter via
3879 PREV_INSN_CC0 is spoiled. That means that fold_rtx may return
3880 NULL_RTX. In those cases, there's nothing to record. */
3881 if (op0 == NULL_RTX || op1 == NULL_RTX)
3882 return;
3884 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3885 if (! cond_known_true)
3887 code = reversed_comparison_code_parts (code, op0, op1, insn);
3889 /* Don't remember if we can't find the inverse. */
3890 if (code == UNKNOWN)
3891 return;
3894 /* The mode is the mode of the non-constant. */
3895 mode = mode0;
3896 if (mode1 != VOIDmode)
3897 mode = mode1;
3899 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3902 /* Yet another form of subreg creation. In this case, we want something in
3903 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3905 static rtx
3906 record_jump_cond_subreg (machine_mode mode, rtx op)
3908 machine_mode op_mode = GET_MODE (op);
3909 if (op_mode == mode || op_mode == VOIDmode)
3910 return op;
3911 return lowpart_subreg (mode, op, op_mode);
3914 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3915 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3916 Make any useful entries we can with that information. Called from
3917 above function and called recursively. */
3919 static void
3920 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3921 rtx op1, int reversed_nonequality)
3923 unsigned op0_hash, op1_hash;
3924 int op0_in_memory, op1_in_memory;
3925 struct table_elt *op0_elt, *op1_elt;
3927 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3928 we know that they are also equal in the smaller mode (this is also
3929 true for all smaller modes whether or not there is a SUBREG, but
3930 is not worth testing for with no SUBREG). */
3932 /* Note that GET_MODE (op0) may not equal MODE. */
3933 if (code == EQ && paradoxical_subreg_p (op0))
3935 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3936 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3937 if (tem)
3938 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3939 reversed_nonequality);
3942 if (code == EQ && paradoxical_subreg_p (op1))
3944 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3945 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3946 if (tem)
3947 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3948 reversed_nonequality);
3951 /* Similarly, if this is an NE comparison, and either is a SUBREG
3952 making a smaller mode, we know the whole thing is also NE. */
3954 /* Note that GET_MODE (op0) may not equal MODE;
3955 if we test MODE instead, we can get an infinite recursion
3956 alternating between two modes each wider than MODE. */
3958 if (code == NE && GET_CODE (op0) == SUBREG
3959 && subreg_lowpart_p (op0)
3960 && (GET_MODE_SIZE (GET_MODE (op0))
3961 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3963 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3964 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3965 if (tem)
3966 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3967 reversed_nonequality);
3970 if (code == NE && GET_CODE (op1) == SUBREG
3971 && subreg_lowpart_p (op1)
3972 && (GET_MODE_SIZE (GET_MODE (op1))
3973 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3975 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3976 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3977 if (tem)
3978 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3979 reversed_nonequality);
3982 /* Hash both operands. */
3984 do_not_record = 0;
3985 hash_arg_in_memory = 0;
3986 op0_hash = HASH (op0, mode);
3987 op0_in_memory = hash_arg_in_memory;
3989 if (do_not_record)
3990 return;
3992 do_not_record = 0;
3993 hash_arg_in_memory = 0;
3994 op1_hash = HASH (op1, mode);
3995 op1_in_memory = hash_arg_in_memory;
3997 if (do_not_record)
3998 return;
4000 /* Look up both operands. */
4001 op0_elt = lookup (op0, op0_hash, mode);
4002 op1_elt = lookup (op1, op1_hash, mode);
4004 /* If both operands are already equivalent or if they are not in the
4005 table but are identical, do nothing. */
4006 if ((op0_elt != 0 && op1_elt != 0
4007 && op0_elt->first_same_value == op1_elt->first_same_value)
4008 || op0 == op1 || rtx_equal_p (op0, op1))
4009 return;
4011 /* If we aren't setting two things equal all we can do is save this
4012 comparison. Similarly if this is floating-point. In the latter
4013 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4014 If we record the equality, we might inadvertently delete code
4015 whose intent was to change -0 to +0. */
4017 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4019 struct qty_table_elem *ent;
4020 int qty;
4022 /* If we reversed a floating-point comparison, if OP0 is not a
4023 register, or if OP1 is neither a register or constant, we can't
4024 do anything. */
4026 if (!REG_P (op1))
4027 op1 = equiv_constant (op1);
4029 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4030 || !REG_P (op0) || op1 == 0)
4031 return;
4033 /* Put OP0 in the hash table if it isn't already. This gives it a
4034 new quantity number. */
4035 if (op0_elt == 0)
4037 if (insert_regs (op0, NULL, 0))
4039 rehash_using_reg (op0);
4040 op0_hash = HASH (op0, mode);
4042 /* If OP0 is contained in OP1, this changes its hash code
4043 as well. Faster to rehash than to check, except
4044 for the simple case of a constant. */
4045 if (! CONSTANT_P (op1))
4046 op1_hash = HASH (op1,mode);
4049 op0_elt = insert (op0, NULL, op0_hash, mode);
4050 op0_elt->in_memory = op0_in_memory;
4053 qty = REG_QTY (REGNO (op0));
4054 ent = &qty_table[qty];
4056 ent->comparison_code = code;
4057 if (REG_P (op1))
4059 /* Look it up again--in case op0 and op1 are the same. */
4060 op1_elt = lookup (op1, op1_hash, mode);
4062 /* Put OP1 in the hash table so it gets a new quantity number. */
4063 if (op1_elt == 0)
4065 if (insert_regs (op1, NULL, 0))
4067 rehash_using_reg (op1);
4068 op1_hash = HASH (op1, mode);
4071 op1_elt = insert (op1, NULL, op1_hash, mode);
4072 op1_elt->in_memory = op1_in_memory;
4075 ent->comparison_const = NULL_RTX;
4076 ent->comparison_qty = REG_QTY (REGNO (op1));
4078 else
4080 ent->comparison_const = op1;
4081 ent->comparison_qty = -1;
4084 return;
4087 /* If either side is still missing an equivalence, make it now,
4088 then merge the equivalences. */
4090 if (op0_elt == 0)
4092 if (insert_regs (op0, NULL, 0))
4094 rehash_using_reg (op0);
4095 op0_hash = HASH (op0, mode);
4098 op0_elt = insert (op0, NULL, op0_hash, mode);
4099 op0_elt->in_memory = op0_in_memory;
4102 if (op1_elt == 0)
4104 if (insert_regs (op1, NULL, 0))
4106 rehash_using_reg (op1);
4107 op1_hash = HASH (op1, mode);
4110 op1_elt = insert (op1, NULL, op1_hash, mode);
4111 op1_elt->in_memory = op1_in_memory;
4114 merge_equiv_classes (op0_elt, op1_elt);
4117 /* CSE processing for one instruction.
4119 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4120 but the few that "leak through" are cleaned up by cse_insn, and complex
4121 addressing modes are often formed here.
4123 The main function is cse_insn, and between here and that function
4124 a couple of helper functions is defined to keep the size of cse_insn
4125 within reasonable proportions.
4127 Data is shared between the main and helper functions via STRUCT SET,
4128 that contains all data related for every set in the instruction that
4129 is being processed.
4131 Note that cse_main processes all sets in the instruction. Most
4132 passes in GCC only process simple SET insns or single_set insns, but
4133 CSE processes insns with multiple sets as well. */
4135 /* Data on one SET contained in the instruction. */
4137 struct set
4139 /* The SET rtx itself. */
4140 rtx rtl;
4141 /* The SET_SRC of the rtx (the original value, if it is changing). */
4142 rtx src;
4143 /* The hash-table element for the SET_SRC of the SET. */
4144 struct table_elt *src_elt;
4145 /* Hash value for the SET_SRC. */
4146 unsigned src_hash;
4147 /* Hash value for the SET_DEST. */
4148 unsigned dest_hash;
4149 /* The SET_DEST, with SUBREG, etc., stripped. */
4150 rtx inner_dest;
4151 /* Nonzero if the SET_SRC is in memory. */
4152 char src_in_memory;
4153 /* Nonzero if the SET_SRC contains something
4154 whose value cannot be predicted and understood. */
4155 char src_volatile;
4156 /* Original machine mode, in case it becomes a CONST_INT.
4157 The size of this field should match the size of the mode
4158 field of struct rtx_def (see rtl.h). */
4159 ENUM_BITFIELD(machine_mode) mode : 8;
4160 /* A constant equivalent for SET_SRC, if any. */
4161 rtx src_const;
4162 /* Hash value of constant equivalent for SET_SRC. */
4163 unsigned src_const_hash;
4164 /* Table entry for constant equivalent for SET_SRC, if any. */
4165 struct table_elt *src_const_elt;
4166 /* Table entry for the destination address. */
4167 struct table_elt *dest_addr_elt;
4170 /* Special handling for (set REG0 REG1) where REG0 is the
4171 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4172 be used in the sequel, so (if easily done) change this insn to
4173 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4174 that computed their value. Then REG1 will become a dead store
4175 and won't cloud the situation for later optimizations.
4177 Do not make this change if REG1 is a hard register, because it will
4178 then be used in the sequel and we may be changing a two-operand insn
4179 into a three-operand insn.
4181 This is the last transformation that cse_insn will try to do. */
4183 static void
4184 try_back_substitute_reg (rtx set, rtx_insn *insn)
4186 rtx dest = SET_DEST (set);
4187 rtx src = SET_SRC (set);
4189 if (REG_P (dest)
4190 && REG_P (src) && ! HARD_REGISTER_P (src)
4191 && REGNO_QTY_VALID_P (REGNO (src)))
4193 int src_q = REG_QTY (REGNO (src));
4194 struct qty_table_elem *src_ent = &qty_table[src_q];
4196 if (src_ent->first_reg == REGNO (dest))
4198 /* Scan for the previous nonnote insn, but stop at a basic
4199 block boundary. */
4200 rtx_insn *prev = insn;
4201 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4204 prev = PREV_INSN (prev);
4206 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4208 /* Do not swap the registers around if the previous instruction
4209 attaches a REG_EQUIV note to REG1.
4211 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4212 from the pseudo that originally shadowed an incoming argument
4213 to another register. Some uses of REG_EQUIV might rely on it
4214 being attached to REG1 rather than REG2.
4216 This section previously turned the REG_EQUIV into a REG_EQUAL
4217 note. We cannot do that because REG_EQUIV may provide an
4218 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4219 if (NONJUMP_INSN_P (prev)
4220 && GET_CODE (PATTERN (prev)) == SET
4221 && SET_DEST (PATTERN (prev)) == src
4222 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4224 rtx note;
4226 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4227 validate_change (insn, &SET_DEST (set), src, 1);
4228 validate_change (insn, &SET_SRC (set), dest, 1);
4229 apply_change_group ();
4231 /* If INSN has a REG_EQUAL note, and this note mentions
4232 REG0, then we must delete it, because the value in
4233 REG0 has changed. If the note's value is REG1, we must
4234 also delete it because that is now this insn's dest. */
4235 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4236 if (note != 0
4237 && (reg_mentioned_p (dest, XEXP (note, 0))
4238 || rtx_equal_p (src, XEXP (note, 0))))
4239 remove_note (insn, note);
4245 /* Record all the SETs in this instruction into SETS_PTR,
4246 and return the number of recorded sets. */
4247 static int
4248 find_sets_in_insn (rtx_insn *insn, struct set **psets)
4250 struct set *sets = *psets;
4251 int n_sets = 0;
4252 rtx x = PATTERN (insn);
4254 if (GET_CODE (x) == SET)
4256 /* Ignore SETs that are unconditional jumps.
4257 They never need cse processing, so this does not hurt.
4258 The reason is not efficiency but rather
4259 so that we can test at the end for instructions
4260 that have been simplified to unconditional jumps
4261 and not be misled by unchanged instructions
4262 that were unconditional jumps to begin with. */
4263 if (SET_DEST (x) == pc_rtx
4264 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4266 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4267 The hard function value register is used only once, to copy to
4268 someplace else, so it isn't worth cse'ing. */
4269 else if (GET_CODE (SET_SRC (x)) == CALL)
4271 else
4272 sets[n_sets++].rtl = x;
4274 else if (GET_CODE (x) == PARALLEL)
4276 int i, lim = XVECLEN (x, 0);
4278 /* Go over the expressions of the PARALLEL in forward order, to
4279 put them in the same order in the SETS array. */
4280 for (i = 0; i < lim; i++)
4282 rtx y = XVECEXP (x, 0, i);
4283 if (GET_CODE (y) == SET)
4285 /* As above, we ignore unconditional jumps and call-insns and
4286 ignore the result of apply_change_group. */
4287 if (SET_DEST (y) == pc_rtx
4288 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4290 else if (GET_CODE (SET_SRC (y)) == CALL)
4292 else
4293 sets[n_sets++].rtl = y;
4298 return n_sets;
4301 /* Where possible, substitute every register reference in the N_SETS
4302 number of SETS in INSN with the canonical register.
4304 Register canonicalization propagatest the earliest register (i.e.
4305 one that is set before INSN) with the same value. This is a very
4306 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4307 to RTL. For instance, a CONST for an address is usually expanded
4308 multiple times to loads into different registers, thus creating many
4309 subexpressions of the form:
4311 (set (reg1) (some_const))
4312 (set (mem (... reg1 ...) (thing)))
4313 (set (reg2) (some_const))
4314 (set (mem (... reg2 ...) (thing)))
4316 After canonicalizing, the code takes the following form:
4318 (set (reg1) (some_const))
4319 (set (mem (... reg1 ...) (thing)))
4320 (set (reg2) (some_const))
4321 (set (mem (... reg1 ...) (thing)))
4323 The set to reg2 is now trivially dead, and the memory reference (or
4324 address, or whatever) may be a candidate for further CSEing.
4326 In this function, the result of apply_change_group can be ignored;
4327 see canon_reg. */
4329 static void
4330 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4332 struct set *sets = *psets;
4333 rtx tem;
4334 rtx x = PATTERN (insn);
4335 int i;
4337 if (CALL_P (insn))
4339 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4340 if (GET_CODE (XEXP (tem, 0)) != SET)
4341 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4344 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4346 canon_reg (SET_SRC (x), insn);
4347 apply_change_group ();
4348 fold_rtx (SET_SRC (x), insn);
4350 else if (GET_CODE (x) == CLOBBER)
4352 /* If we clobber memory, canon the address.
4353 This does nothing when a register is clobbered
4354 because we have already invalidated the reg. */
4355 if (MEM_P (XEXP (x, 0)))
4356 canon_reg (XEXP (x, 0), insn);
4358 else if (GET_CODE (x) == USE
4359 && ! (REG_P (XEXP (x, 0))
4360 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4361 /* Canonicalize a USE of a pseudo register or memory location. */
4362 canon_reg (x, insn);
4363 else if (GET_CODE (x) == ASM_OPERANDS)
4365 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4367 rtx input = ASM_OPERANDS_INPUT (x, i);
4368 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4370 input = canon_reg (input, insn);
4371 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4375 else if (GET_CODE (x) == CALL)
4377 canon_reg (x, insn);
4378 apply_change_group ();
4379 fold_rtx (x, insn);
4381 else if (DEBUG_INSN_P (insn))
4382 canon_reg (PATTERN (insn), insn);
4383 else if (GET_CODE (x) == PARALLEL)
4385 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4387 rtx y = XVECEXP (x, 0, i);
4388 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4390 canon_reg (SET_SRC (y), insn);
4391 apply_change_group ();
4392 fold_rtx (SET_SRC (y), insn);
4394 else if (GET_CODE (y) == CLOBBER)
4396 if (MEM_P (XEXP (y, 0)))
4397 canon_reg (XEXP (y, 0), insn);
4399 else if (GET_CODE (y) == USE
4400 && ! (REG_P (XEXP (y, 0))
4401 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4402 canon_reg (y, insn);
4403 else if (GET_CODE (y) == CALL)
4405 canon_reg (y, insn);
4406 apply_change_group ();
4407 fold_rtx (y, insn);
4412 if (n_sets == 1 && REG_NOTES (insn) != 0
4413 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4415 /* We potentially will process this insn many times. Therefore,
4416 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4417 unique set in INSN.
4419 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4420 because cse_insn handles those specially. */
4421 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4422 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4423 remove_note (insn, tem);
4424 else
4426 canon_reg (XEXP (tem, 0), insn);
4427 apply_change_group ();
4428 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4429 df_notes_rescan (insn);
4433 /* Canonicalize sources and addresses of destinations.
4434 We do this in a separate pass to avoid problems when a MATCH_DUP is
4435 present in the insn pattern. In that case, we want to ensure that
4436 we don't break the duplicate nature of the pattern. So we will replace
4437 both operands at the same time. Otherwise, we would fail to find an
4438 equivalent substitution in the loop calling validate_change below.
4440 We used to suppress canonicalization of DEST if it appears in SRC,
4441 but we don't do this any more. */
4443 for (i = 0; i < n_sets; i++)
4445 rtx dest = SET_DEST (sets[i].rtl);
4446 rtx src = SET_SRC (sets[i].rtl);
4447 rtx new_rtx = canon_reg (src, insn);
4449 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4451 if (GET_CODE (dest) == ZERO_EXTRACT)
4453 validate_change (insn, &XEXP (dest, 1),
4454 canon_reg (XEXP (dest, 1), insn), 1);
4455 validate_change (insn, &XEXP (dest, 2),
4456 canon_reg (XEXP (dest, 2), insn), 1);
4459 while (GET_CODE (dest) == SUBREG
4460 || GET_CODE (dest) == ZERO_EXTRACT
4461 || GET_CODE (dest) == STRICT_LOW_PART)
4462 dest = XEXP (dest, 0);
4464 if (MEM_P (dest))
4465 canon_reg (dest, insn);
4468 /* Now that we have done all the replacements, we can apply the change
4469 group and see if they all work. Note that this will cause some
4470 canonicalizations that would have worked individually not to be applied
4471 because some other canonicalization didn't work, but this should not
4472 occur often.
4474 The result of apply_change_group can be ignored; see canon_reg. */
4476 apply_change_group ();
4479 /* Main function of CSE.
4480 First simplify sources and addresses of all assignments
4481 in the instruction, using previously-computed equivalents values.
4482 Then install the new sources and destinations in the table
4483 of available values. */
4485 static void
4486 cse_insn (rtx_insn *insn)
4488 rtx x = PATTERN (insn);
4489 int i;
4490 rtx tem;
4491 int n_sets = 0;
4493 rtx src_eqv = 0;
4494 struct table_elt *src_eqv_elt = 0;
4495 int src_eqv_volatile = 0;
4496 int src_eqv_in_memory = 0;
4497 unsigned src_eqv_hash = 0;
4499 struct set *sets = (struct set *) 0;
4501 if (GET_CODE (x) == SET)
4502 sets = XALLOCA (struct set);
4503 else if (GET_CODE (x) == PARALLEL)
4504 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4506 this_insn = insn;
4507 /* Records what this insn does to set CC0. */
4508 this_insn_cc0 = 0;
4509 this_insn_cc0_mode = VOIDmode;
4511 /* Find all regs explicitly clobbered in this insn,
4512 to ensure they are not replaced with any other regs
4513 elsewhere in this insn. */
4514 invalidate_from_sets_and_clobbers (insn);
4516 /* Record all the SETs in this instruction. */
4517 n_sets = find_sets_in_insn (insn, &sets);
4519 /* Substitute the canonical register where possible. */
4520 canonicalize_insn (insn, &sets, n_sets);
4522 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4523 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4524 latter condition is necessary because SRC_EQV is handled specially for
4525 this case, and if it isn't set, then there will be no equivalence
4526 for the destination. */
4527 if (n_sets == 1 && REG_NOTES (insn) != 0
4528 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4531 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4532 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4533 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4534 src_eqv = copy_rtx (XEXP (tem, 0));
4535 /* If DEST is of the form ZERO_EXTACT, as in:
4536 (set (zero_extract:SI (reg:SI 119)
4537 (const_int 16 [0x10])
4538 (const_int 16 [0x10]))
4539 (const_int 51154 [0xc7d2]))
4540 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4541 point. Note that this is different from SRC_EQV. We can however
4542 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4543 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4544 && CONST_INT_P (XEXP (tem, 0))
4545 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4546 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4548 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4549 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4550 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4551 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4552 HOST_WIDE_INT mask;
4553 unsigned int shift;
4554 if (BITS_BIG_ENDIAN)
4555 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
4556 - INTVAL (pos) - INTVAL (width);
4557 else
4558 shift = INTVAL (pos);
4559 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4560 mask = ~(HOST_WIDE_INT) 0;
4561 else
4562 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
4563 val = (val >> shift) & mask;
4564 src_eqv = GEN_INT (val);
4568 /* Set sets[i].src_elt to the class each source belongs to.
4569 Detect assignments from or to volatile things
4570 and set set[i] to zero so they will be ignored
4571 in the rest of this function.
4573 Nothing in this loop changes the hash table or the register chains. */
4575 for (i = 0; i < n_sets; i++)
4577 bool repeat = false;
4578 bool mem_noop_insn = false;
4579 rtx src, dest;
4580 rtx src_folded;
4581 struct table_elt *elt = 0, *p;
4582 machine_mode mode;
4583 rtx src_eqv_here;
4584 rtx src_const = 0;
4585 rtx src_related = 0;
4586 bool src_related_is_const_anchor = false;
4587 struct table_elt *src_const_elt = 0;
4588 int src_cost = MAX_COST;
4589 int src_eqv_cost = MAX_COST;
4590 int src_folded_cost = MAX_COST;
4591 int src_related_cost = MAX_COST;
4592 int src_elt_cost = MAX_COST;
4593 int src_regcost = MAX_COST;
4594 int src_eqv_regcost = MAX_COST;
4595 int src_folded_regcost = MAX_COST;
4596 int src_related_regcost = MAX_COST;
4597 int src_elt_regcost = MAX_COST;
4598 /* Set nonzero if we need to call force_const_mem on with the
4599 contents of src_folded before using it. */
4600 int src_folded_force_flag = 0;
4602 dest = SET_DEST (sets[i].rtl);
4603 src = SET_SRC (sets[i].rtl);
4605 /* If SRC is a constant that has no machine mode,
4606 hash it with the destination's machine mode.
4607 This way we can keep different modes separate. */
4609 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4610 sets[i].mode = mode;
4612 if (src_eqv)
4614 machine_mode eqvmode = mode;
4615 if (GET_CODE (dest) == STRICT_LOW_PART)
4616 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4617 do_not_record = 0;
4618 hash_arg_in_memory = 0;
4619 src_eqv_hash = HASH (src_eqv, eqvmode);
4621 /* Find the equivalence class for the equivalent expression. */
4623 if (!do_not_record)
4624 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4626 src_eqv_volatile = do_not_record;
4627 src_eqv_in_memory = hash_arg_in_memory;
4630 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4631 value of the INNER register, not the destination. So it is not
4632 a valid substitution for the source. But save it for later. */
4633 if (GET_CODE (dest) == STRICT_LOW_PART)
4634 src_eqv_here = 0;
4635 else
4636 src_eqv_here = src_eqv;
4638 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4639 simplified result, which may not necessarily be valid. */
4640 src_folded = fold_rtx (src, NULL);
4642 #if 0
4643 /* ??? This caused bad code to be generated for the m68k port with -O2.
4644 Suppose src is (CONST_INT -1), and that after truncation src_folded
4645 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4646 At the end we will add src and src_const to the same equivalence
4647 class. We now have 3 and -1 on the same equivalence class. This
4648 causes later instructions to be mis-optimized. */
4649 /* If storing a constant in a bitfield, pre-truncate the constant
4650 so we will be able to record it later. */
4651 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4653 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4655 if (CONST_INT_P (src)
4656 && CONST_INT_P (width)
4657 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4658 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4659 src_folded
4660 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4661 << INTVAL (width)) - 1));
4663 #endif
4665 /* Compute SRC's hash code, and also notice if it
4666 should not be recorded at all. In that case,
4667 prevent any further processing of this assignment. */
4668 do_not_record = 0;
4669 hash_arg_in_memory = 0;
4671 sets[i].src = src;
4672 sets[i].src_hash = HASH (src, mode);
4673 sets[i].src_volatile = do_not_record;
4674 sets[i].src_in_memory = hash_arg_in_memory;
4676 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4677 a pseudo, do not record SRC. Using SRC as a replacement for
4678 anything else will be incorrect in that situation. Note that
4679 this usually occurs only for stack slots, in which case all the
4680 RTL would be referring to SRC, so we don't lose any optimization
4681 opportunities by not having SRC in the hash table. */
4683 if (MEM_P (src)
4684 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4685 && REG_P (dest)
4686 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4687 sets[i].src_volatile = 1;
4689 else if (GET_CODE (src) == ASM_OPERANDS
4690 && GET_CODE (x) == PARALLEL)
4692 /* Do not record result of a non-volatile inline asm with
4693 more than one result. */
4694 if (n_sets > 1)
4695 sets[i].src_volatile = 1;
4697 int j, lim = XVECLEN (x, 0);
4698 for (j = 0; j < lim; j++)
4700 rtx y = XVECEXP (x, 0, j);
4701 /* And do not record result of a non-volatile inline asm
4702 with "memory" clobber. */
4703 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4705 sets[i].src_volatile = 1;
4706 break;
4711 #if 0
4712 /* It is no longer clear why we used to do this, but it doesn't
4713 appear to still be needed. So let's try without it since this
4714 code hurts cse'ing widened ops. */
4715 /* If source is a paradoxical subreg (such as QI treated as an SI),
4716 treat it as volatile. It may do the work of an SI in one context
4717 where the extra bits are not being used, but cannot replace an SI
4718 in general. */
4719 if (paradoxical_subreg_p (src))
4720 sets[i].src_volatile = 1;
4721 #endif
4723 /* Locate all possible equivalent forms for SRC. Try to replace
4724 SRC in the insn with each cheaper equivalent.
4726 We have the following types of equivalents: SRC itself, a folded
4727 version, a value given in a REG_EQUAL note, or a value related
4728 to a constant.
4730 Each of these equivalents may be part of an additional class
4731 of equivalents (if more than one is in the table, they must be in
4732 the same class; we check for this).
4734 If the source is volatile, we don't do any table lookups.
4736 We note any constant equivalent for possible later use in a
4737 REG_NOTE. */
4739 if (!sets[i].src_volatile)
4740 elt = lookup (src, sets[i].src_hash, mode);
4742 sets[i].src_elt = elt;
4744 if (elt && src_eqv_here && src_eqv_elt)
4746 if (elt->first_same_value != src_eqv_elt->first_same_value)
4748 /* The REG_EQUAL is indicating that two formerly distinct
4749 classes are now equivalent. So merge them. */
4750 merge_equiv_classes (elt, src_eqv_elt);
4751 src_eqv_hash = HASH (src_eqv, elt->mode);
4752 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4755 src_eqv_here = 0;
4758 else if (src_eqv_elt)
4759 elt = src_eqv_elt;
4761 /* Try to find a constant somewhere and record it in `src_const'.
4762 Record its table element, if any, in `src_const_elt'. Look in
4763 any known equivalences first. (If the constant is not in the
4764 table, also set `sets[i].src_const_hash'). */
4765 if (elt)
4766 for (p = elt->first_same_value; p; p = p->next_same_value)
4767 if (p->is_const)
4769 src_const = p->exp;
4770 src_const_elt = elt;
4771 break;
4774 if (src_const == 0
4775 && (CONSTANT_P (src_folded)
4776 /* Consider (minus (label_ref L1) (label_ref L2)) as
4777 "constant" here so we will record it. This allows us
4778 to fold switch statements when an ADDR_DIFF_VEC is used. */
4779 || (GET_CODE (src_folded) == MINUS
4780 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4781 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4782 src_const = src_folded, src_const_elt = elt;
4783 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4784 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4786 /* If we don't know if the constant is in the table, get its
4787 hash code and look it up. */
4788 if (src_const && src_const_elt == 0)
4790 sets[i].src_const_hash = HASH (src_const, mode);
4791 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4794 sets[i].src_const = src_const;
4795 sets[i].src_const_elt = src_const_elt;
4797 /* If the constant and our source are both in the table, mark them as
4798 equivalent. Otherwise, if a constant is in the table but the source
4799 isn't, set ELT to it. */
4800 if (src_const_elt && elt
4801 && src_const_elt->first_same_value != elt->first_same_value)
4802 merge_equiv_classes (elt, src_const_elt);
4803 else if (src_const_elt && elt == 0)
4804 elt = src_const_elt;
4806 /* See if there is a register linearly related to a constant
4807 equivalent of SRC. */
4808 if (src_const
4809 && (GET_CODE (src_const) == CONST
4810 || (src_const_elt && src_const_elt->related_value != 0)))
4812 src_related = use_related_value (src_const, src_const_elt);
4813 if (src_related)
4815 struct table_elt *src_related_elt
4816 = lookup (src_related, HASH (src_related, mode), mode);
4817 if (src_related_elt && elt)
4819 if (elt->first_same_value
4820 != src_related_elt->first_same_value)
4821 /* This can occur when we previously saw a CONST
4822 involving a SYMBOL_REF and then see the SYMBOL_REF
4823 twice. Merge the involved classes. */
4824 merge_equiv_classes (elt, src_related_elt);
4826 src_related = 0;
4827 src_related_elt = 0;
4829 else if (src_related_elt && elt == 0)
4830 elt = src_related_elt;
4834 /* See if we have a CONST_INT that is already in a register in a
4835 wider mode. */
4837 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4838 && GET_MODE_CLASS (mode) == MODE_INT
4839 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4841 machine_mode wider_mode;
4843 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4844 wider_mode != VOIDmode
4845 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4846 && src_related == 0;
4847 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4849 struct table_elt *const_elt
4850 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4852 if (const_elt == 0)
4853 continue;
4855 for (const_elt = const_elt->first_same_value;
4856 const_elt; const_elt = const_elt->next_same_value)
4857 if (REG_P (const_elt->exp))
4859 src_related = gen_lowpart (mode, const_elt->exp);
4860 break;
4865 /* Another possibility is that we have an AND with a constant in
4866 a mode narrower than a word. If so, it might have been generated
4867 as part of an "if" which would narrow the AND. If we already
4868 have done the AND in a wider mode, we can use a SUBREG of that
4869 value. */
4871 if (flag_expensive_optimizations && ! src_related
4872 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4873 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4875 machine_mode tmode;
4876 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4878 for (tmode = GET_MODE_WIDER_MODE (mode);
4879 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4880 tmode = GET_MODE_WIDER_MODE (tmode))
4882 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4883 struct table_elt *larger_elt;
4885 if (inner)
4887 PUT_MODE (new_and, tmode);
4888 XEXP (new_and, 0) = inner;
4889 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4890 if (larger_elt == 0)
4891 continue;
4893 for (larger_elt = larger_elt->first_same_value;
4894 larger_elt; larger_elt = larger_elt->next_same_value)
4895 if (REG_P (larger_elt->exp))
4897 src_related
4898 = gen_lowpart (mode, larger_elt->exp);
4899 break;
4902 if (src_related)
4903 break;
4908 /* See if a MEM has already been loaded with a widening operation;
4909 if it has, we can use a subreg of that. Many CISC machines
4910 also have such operations, but this is only likely to be
4911 beneficial on these machines. */
4913 if (flag_expensive_optimizations && src_related == 0
4914 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4915 && GET_MODE_CLASS (mode) == MODE_INT
4916 && MEM_P (src) && ! do_not_record
4917 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4919 struct rtx_def memory_extend_buf;
4920 rtx memory_extend_rtx = &memory_extend_buf;
4921 machine_mode tmode;
4923 /* Set what we are trying to extend and the operation it might
4924 have been extended with. */
4925 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4926 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4927 XEXP (memory_extend_rtx, 0) = src;
4929 for (tmode = GET_MODE_WIDER_MODE (mode);
4930 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4931 tmode = GET_MODE_WIDER_MODE (tmode))
4933 struct table_elt *larger_elt;
4935 PUT_MODE (memory_extend_rtx, tmode);
4936 larger_elt = lookup (memory_extend_rtx,
4937 HASH (memory_extend_rtx, tmode), tmode);
4938 if (larger_elt == 0)
4939 continue;
4941 for (larger_elt = larger_elt->first_same_value;
4942 larger_elt; larger_elt = larger_elt->next_same_value)
4943 if (REG_P (larger_elt->exp))
4945 src_related = gen_lowpart (mode, larger_elt->exp);
4946 break;
4949 if (src_related)
4950 break;
4954 /* Try to express the constant using a register+offset expression
4955 derived from a constant anchor. */
4957 if (targetm.const_anchor
4958 && !src_related
4959 && src_const
4960 && GET_CODE (src_const) == CONST_INT)
4962 src_related = try_const_anchors (src_const, mode);
4963 src_related_is_const_anchor = src_related != NULL_RTX;
4967 if (src == src_folded)
4968 src_folded = 0;
4970 /* At this point, ELT, if nonzero, points to a class of expressions
4971 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4972 and SRC_RELATED, if nonzero, each contain additional equivalent
4973 expressions. Prune these latter expressions by deleting expressions
4974 already in the equivalence class.
4976 Check for an equivalent identical to the destination. If found,
4977 this is the preferred equivalent since it will likely lead to
4978 elimination of the insn. Indicate this by placing it in
4979 `src_related'. */
4981 if (elt)
4982 elt = elt->first_same_value;
4983 for (p = elt; p; p = p->next_same_value)
4985 enum rtx_code code = GET_CODE (p->exp);
4987 /* If the expression is not valid, ignore it. Then we do not
4988 have to check for validity below. In most cases, we can use
4989 `rtx_equal_p', since canonicalization has already been done. */
4990 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4991 continue;
4993 /* Also skip paradoxical subregs, unless that's what we're
4994 looking for. */
4995 if (paradoxical_subreg_p (p->exp)
4996 && ! (src != 0
4997 && GET_CODE (src) == SUBREG
4998 && GET_MODE (src) == GET_MODE (p->exp)
4999 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5000 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5001 continue;
5003 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5004 src = 0;
5005 else if (src_folded && GET_CODE (src_folded) == code
5006 && rtx_equal_p (src_folded, p->exp))
5007 src_folded = 0;
5008 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5009 && rtx_equal_p (src_eqv_here, p->exp))
5010 src_eqv_here = 0;
5011 else if (src_related && GET_CODE (src_related) == code
5012 && rtx_equal_p (src_related, p->exp))
5013 src_related = 0;
5015 /* This is the same as the destination of the insns, we want
5016 to prefer it. Copy it to src_related. The code below will
5017 then give it a negative cost. */
5018 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5019 src_related = dest;
5022 /* Find the cheapest valid equivalent, trying all the available
5023 possibilities. Prefer items not in the hash table to ones
5024 that are when they are equal cost. Note that we can never
5025 worsen an insn as the current contents will also succeed.
5026 If we find an equivalent identical to the destination, use it as best,
5027 since this insn will probably be eliminated in that case. */
5028 if (src)
5030 if (rtx_equal_p (src, dest))
5031 src_cost = src_regcost = -1;
5032 else
5034 src_cost = COST (src, mode);
5035 src_regcost = approx_reg_cost (src);
5039 if (src_eqv_here)
5041 if (rtx_equal_p (src_eqv_here, dest))
5042 src_eqv_cost = src_eqv_regcost = -1;
5043 else
5045 src_eqv_cost = COST (src_eqv_here, mode);
5046 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5050 if (src_folded)
5052 if (rtx_equal_p (src_folded, dest))
5053 src_folded_cost = src_folded_regcost = -1;
5054 else
5056 src_folded_cost = COST (src_folded, mode);
5057 src_folded_regcost = approx_reg_cost (src_folded);
5061 if (src_related)
5063 if (rtx_equal_p (src_related, dest))
5064 src_related_cost = src_related_regcost = -1;
5065 else
5067 src_related_cost = COST (src_related, mode);
5068 src_related_regcost = approx_reg_cost (src_related);
5070 /* If a const-anchor is used to synthesize a constant that
5071 normally requires multiple instructions then slightly prefer
5072 it over the original sequence. These instructions are likely
5073 to become redundant now. We can't compare against the cost
5074 of src_eqv_here because, on MIPS for example, multi-insn
5075 constants have zero cost; they are assumed to be hoisted from
5076 loops. */
5077 if (src_related_is_const_anchor
5078 && src_related_cost == src_cost
5079 && src_eqv_here)
5080 src_related_cost--;
5084 /* If this was an indirect jump insn, a known label will really be
5085 cheaper even though it looks more expensive. */
5086 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5087 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5089 /* Terminate loop when replacement made. This must terminate since
5090 the current contents will be tested and will always be valid. */
5091 while (1)
5093 rtx trial;
5095 /* Skip invalid entries. */
5096 while (elt && !REG_P (elt->exp)
5097 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5098 elt = elt->next_same_value;
5100 /* A paradoxical subreg would be bad here: it'll be the right
5101 size, but later may be adjusted so that the upper bits aren't
5102 what we want. So reject it. */
5103 if (elt != 0
5104 && paradoxical_subreg_p (elt->exp)
5105 /* It is okay, though, if the rtx we're trying to match
5106 will ignore any of the bits we can't predict. */
5107 && ! (src != 0
5108 && GET_CODE (src) == SUBREG
5109 && GET_MODE (src) == GET_MODE (elt->exp)
5110 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5111 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5113 elt = elt->next_same_value;
5114 continue;
5117 if (elt)
5119 src_elt_cost = elt->cost;
5120 src_elt_regcost = elt->regcost;
5123 /* Find cheapest and skip it for the next time. For items
5124 of equal cost, use this order:
5125 src_folded, src, src_eqv, src_related and hash table entry. */
5126 if (src_folded
5127 && preferable (src_folded_cost, src_folded_regcost,
5128 src_cost, src_regcost) <= 0
5129 && preferable (src_folded_cost, src_folded_regcost,
5130 src_eqv_cost, src_eqv_regcost) <= 0
5131 && preferable (src_folded_cost, src_folded_regcost,
5132 src_related_cost, src_related_regcost) <= 0
5133 && preferable (src_folded_cost, src_folded_regcost,
5134 src_elt_cost, src_elt_regcost) <= 0)
5136 trial = src_folded, src_folded_cost = MAX_COST;
5137 if (src_folded_force_flag)
5139 rtx forced = force_const_mem (mode, trial);
5140 if (forced)
5141 trial = forced;
5144 else if (src
5145 && preferable (src_cost, src_regcost,
5146 src_eqv_cost, src_eqv_regcost) <= 0
5147 && preferable (src_cost, src_regcost,
5148 src_related_cost, src_related_regcost) <= 0
5149 && preferable (src_cost, src_regcost,
5150 src_elt_cost, src_elt_regcost) <= 0)
5151 trial = src, src_cost = MAX_COST;
5152 else if (src_eqv_here
5153 && preferable (src_eqv_cost, src_eqv_regcost,
5154 src_related_cost, src_related_regcost) <= 0
5155 && preferable (src_eqv_cost, src_eqv_regcost,
5156 src_elt_cost, src_elt_regcost) <= 0)
5157 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5158 else if (src_related
5159 && preferable (src_related_cost, src_related_regcost,
5160 src_elt_cost, src_elt_regcost) <= 0)
5161 trial = src_related, src_related_cost = MAX_COST;
5162 else
5164 trial = elt->exp;
5165 elt = elt->next_same_value;
5166 src_elt_cost = MAX_COST;
5169 /* Avoid creation of overlapping memory moves. */
5170 if (MEM_P (trial) && MEM_P (dest) && !rtx_equal_p (trial, dest))
5172 rtx src, dest;
5174 /* BLKmode moves are not handled by cse anyway. */
5175 if (GET_MODE (trial) == BLKmode)
5176 break;
5178 src = canon_rtx (trial);
5179 dest = canon_rtx (SET_DEST (sets[i].rtl));
5181 if (!MEM_P (src) || !MEM_P (dest)
5182 || !nonoverlapping_memrefs_p (src, dest, false))
5183 break;
5186 /* Try to optimize
5187 (set (reg:M N) (const_int A))
5188 (set (reg:M2 O) (const_int B))
5189 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5190 (reg:M2 O)). */
5191 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5192 && CONST_INT_P (trial)
5193 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5194 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5195 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5196 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5197 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5198 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5199 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5200 <= HOST_BITS_PER_WIDE_INT))
5202 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5203 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5204 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5205 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5206 struct table_elt *dest_elt
5207 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5208 rtx dest_cst = NULL;
5210 if (dest_elt)
5211 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5212 if (p->is_const && CONST_INT_P (p->exp))
5214 dest_cst = p->exp;
5215 break;
5217 if (dest_cst)
5219 HOST_WIDE_INT val = INTVAL (dest_cst);
5220 HOST_WIDE_INT mask;
5221 unsigned int shift;
5222 if (BITS_BIG_ENDIAN)
5223 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5224 - INTVAL (pos) - INTVAL (width);
5225 else
5226 shift = INTVAL (pos);
5227 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5228 mask = ~(HOST_WIDE_INT) 0;
5229 else
5230 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5231 val &= ~(mask << shift);
5232 val |= (INTVAL (trial) & mask) << shift;
5233 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5234 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5235 dest_reg, 1);
5236 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5237 GEN_INT (val), 1);
5238 if (apply_change_group ())
5240 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5241 if (note)
5243 remove_note (insn, note);
5244 df_notes_rescan (insn);
5246 src_eqv = NULL_RTX;
5247 src_eqv_elt = NULL;
5248 src_eqv_volatile = 0;
5249 src_eqv_in_memory = 0;
5250 src_eqv_hash = 0;
5251 repeat = true;
5252 break;
5257 /* We don't normally have an insn matching (set (pc) (pc)), so
5258 check for this separately here. We will delete such an
5259 insn below.
5261 For other cases such as a table jump or conditional jump
5262 where we know the ultimate target, go ahead and replace the
5263 operand. While that may not make a valid insn, we will
5264 reemit the jump below (and also insert any necessary
5265 barriers). */
5266 if (n_sets == 1 && dest == pc_rtx
5267 && (trial == pc_rtx
5268 || (GET_CODE (trial) == LABEL_REF
5269 && ! condjump_p (insn))))
5271 /* Don't substitute non-local labels, this confuses CFG. */
5272 if (GET_CODE (trial) == LABEL_REF
5273 && LABEL_REF_NONLOCAL_P (trial))
5274 continue;
5276 SET_SRC (sets[i].rtl) = trial;
5277 cse_jumps_altered = true;
5278 break;
5281 /* Similarly, lots of targets don't allow no-op
5282 (set (mem x) (mem x)) moves. */
5283 else if (n_sets == 1
5284 && MEM_P (trial)
5285 && MEM_P (dest)
5286 && rtx_equal_p (trial, dest)
5287 && !side_effects_p (dest)
5288 && (cfun->can_delete_dead_exceptions
5289 || insn_nothrow_p (insn)))
5291 SET_SRC (sets[i].rtl) = trial;
5292 mem_noop_insn = true;
5293 break;
5296 /* Reject certain invalid forms of CONST that we create. */
5297 else if (CONSTANT_P (trial)
5298 && GET_CODE (trial) == CONST
5299 /* Reject cases that will cause decode_rtx_const to
5300 die. On the alpha when simplifying a switch, we
5301 get (const (truncate (minus (label_ref)
5302 (label_ref)))). */
5303 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5304 /* Likewise on IA-64, except without the
5305 truncate. */
5306 || (GET_CODE (XEXP (trial, 0)) == MINUS
5307 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5308 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5309 /* Do nothing for this case. */
5312 /* Look for a substitution that makes a valid insn. */
5313 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5314 trial, 0))
5316 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5318 /* The result of apply_change_group can be ignored; see
5319 canon_reg. */
5321 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5322 apply_change_group ();
5324 break;
5327 /* If we previously found constant pool entries for
5328 constants and this is a constant, try making a
5329 pool entry. Put it in src_folded unless we already have done
5330 this since that is where it likely came from. */
5332 else if (constant_pool_entries_cost
5333 && CONSTANT_P (trial)
5334 && (src_folded == 0
5335 || (!MEM_P (src_folded)
5336 && ! src_folded_force_flag))
5337 && GET_MODE_CLASS (mode) != MODE_CC
5338 && mode != VOIDmode)
5340 src_folded_force_flag = 1;
5341 src_folded = trial;
5342 src_folded_cost = constant_pool_entries_cost;
5343 src_folded_regcost = constant_pool_entries_regcost;
5347 /* If we changed the insn too much, handle this set from scratch. */
5348 if (repeat)
5350 i--;
5351 continue;
5354 src = SET_SRC (sets[i].rtl);
5356 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5357 However, there is an important exception: If both are registers
5358 that are not the head of their equivalence class, replace SET_SRC
5359 with the head of the class. If we do not do this, we will have
5360 both registers live over a portion of the basic block. This way,
5361 their lifetimes will likely abut instead of overlapping. */
5362 if (REG_P (dest)
5363 && REGNO_QTY_VALID_P (REGNO (dest)))
5365 int dest_q = REG_QTY (REGNO (dest));
5366 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5368 if (dest_ent->mode == GET_MODE (dest)
5369 && dest_ent->first_reg != REGNO (dest)
5370 && REG_P (src) && REGNO (src) == REGNO (dest)
5371 /* Don't do this if the original insn had a hard reg as
5372 SET_SRC or SET_DEST. */
5373 && (!REG_P (sets[i].src)
5374 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5375 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5376 /* We can't call canon_reg here because it won't do anything if
5377 SRC is a hard register. */
5379 int src_q = REG_QTY (REGNO (src));
5380 struct qty_table_elem *src_ent = &qty_table[src_q];
5381 int first = src_ent->first_reg;
5382 rtx new_src
5383 = (first >= FIRST_PSEUDO_REGISTER
5384 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5386 /* We must use validate-change even for this, because this
5387 might be a special no-op instruction, suitable only to
5388 tag notes onto. */
5389 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5391 src = new_src;
5392 /* If we had a constant that is cheaper than what we are now
5393 setting SRC to, use that constant. We ignored it when we
5394 thought we could make this into a no-op. */
5395 if (src_const && COST (src_const, mode) < COST (src, mode)
5396 && validate_change (insn, &SET_SRC (sets[i].rtl),
5397 src_const, 0))
5398 src = src_const;
5403 /* If we made a change, recompute SRC values. */
5404 if (src != sets[i].src)
5406 do_not_record = 0;
5407 hash_arg_in_memory = 0;
5408 sets[i].src = src;
5409 sets[i].src_hash = HASH (src, mode);
5410 sets[i].src_volatile = do_not_record;
5411 sets[i].src_in_memory = hash_arg_in_memory;
5412 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5415 /* If this is a single SET, we are setting a register, and we have an
5416 equivalent constant, we want to add a REG_EQUAL note if the constant
5417 is different from the source. We don't want to do it for a constant
5418 pseudo since verifying that this pseudo hasn't been eliminated is a
5419 pain; moreover such a note won't help anything.
5421 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5422 which can be created for a reference to a compile time computable
5423 entry in a jump table. */
5424 if (n_sets == 1
5425 && REG_P (dest)
5426 && src_const
5427 && !REG_P (src_const)
5428 && !(GET_CODE (src_const) == SUBREG
5429 && REG_P (SUBREG_REG (src_const)))
5430 && !(GET_CODE (src_const) == CONST
5431 && GET_CODE (XEXP (src_const, 0)) == MINUS
5432 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5433 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5434 && !rtx_equal_p (src, src_const))
5436 /* Make sure that the rtx is not shared. */
5437 src_const = copy_rtx (src_const);
5439 /* Record the actual constant value in a REG_EQUAL note,
5440 making a new one if one does not already exist. */
5441 set_unique_reg_note (insn, REG_EQUAL, src_const);
5442 df_notes_rescan (insn);
5445 /* Now deal with the destination. */
5446 do_not_record = 0;
5448 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5449 while (GET_CODE (dest) == SUBREG
5450 || GET_CODE (dest) == ZERO_EXTRACT
5451 || GET_CODE (dest) == STRICT_LOW_PART)
5452 dest = XEXP (dest, 0);
5454 sets[i].inner_dest = dest;
5456 if (MEM_P (dest))
5458 #ifdef PUSH_ROUNDING
5459 /* Stack pushes invalidate the stack pointer. */
5460 rtx addr = XEXP (dest, 0);
5461 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5462 && XEXP (addr, 0) == stack_pointer_rtx)
5463 invalidate (stack_pointer_rtx, VOIDmode);
5464 #endif
5465 dest = fold_rtx (dest, insn);
5468 /* Compute the hash code of the destination now,
5469 before the effects of this instruction are recorded,
5470 since the register values used in the address computation
5471 are those before this instruction. */
5472 sets[i].dest_hash = HASH (dest, mode);
5474 /* Don't enter a bit-field in the hash table
5475 because the value in it after the store
5476 may not equal what was stored, due to truncation. */
5478 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5480 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5482 if (src_const != 0 && CONST_INT_P (src_const)
5483 && CONST_INT_P (width)
5484 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5485 && ! (INTVAL (src_const)
5486 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5487 /* Exception: if the value is constant,
5488 and it won't be truncated, record it. */
5490 else
5492 /* This is chosen so that the destination will be invalidated
5493 but no new value will be recorded.
5494 We must invalidate because sometimes constant
5495 values can be recorded for bitfields. */
5496 sets[i].src_elt = 0;
5497 sets[i].src_volatile = 1;
5498 src_eqv = 0;
5499 src_eqv_elt = 0;
5503 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5504 the insn. */
5505 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5507 /* One less use of the label this insn used to jump to. */
5508 cse_cfg_altered |= delete_insn_and_edges (insn);
5509 cse_jumps_altered = true;
5510 /* No more processing for this set. */
5511 sets[i].rtl = 0;
5514 /* Similarly for no-op MEM moves. */
5515 else if (mem_noop_insn)
5517 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5518 cse_cfg_altered = true;
5519 cse_cfg_altered |= delete_insn_and_edges (insn);
5520 /* No more processing for this set. */
5521 sets[i].rtl = 0;
5524 /* If this SET is now setting PC to a label, we know it used to
5525 be a conditional or computed branch. */
5526 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5527 && !LABEL_REF_NONLOCAL_P (src))
5529 /* We reemit the jump in as many cases as possible just in
5530 case the form of an unconditional jump is significantly
5531 different than a computed jump or conditional jump.
5533 If this insn has multiple sets, then reemitting the
5534 jump is nontrivial. So instead we just force rerecognition
5535 and hope for the best. */
5536 if (n_sets == 1)
5538 rtx_jump_insn *new_rtx;
5539 rtx note;
5541 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5542 new_rtx = emit_jump_insn_before (seq, insn);
5543 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5544 LABEL_NUSES (XEXP (src, 0))++;
5546 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5547 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5548 if (note)
5550 XEXP (note, 1) = NULL_RTX;
5551 REG_NOTES (new_rtx) = note;
5554 cse_cfg_altered |= delete_insn_and_edges (insn);
5555 insn = new_rtx;
5557 else
5558 INSN_CODE (insn) = -1;
5560 /* Do not bother deleting any unreachable code, let jump do it. */
5561 cse_jumps_altered = true;
5562 sets[i].rtl = 0;
5565 /* If destination is volatile, invalidate it and then do no further
5566 processing for this assignment. */
5568 else if (do_not_record)
5570 invalidate_dest (dest);
5571 sets[i].rtl = 0;
5574 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5576 do_not_record = 0;
5577 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5578 if (do_not_record)
5580 invalidate_dest (SET_DEST (sets[i].rtl));
5581 sets[i].rtl = 0;
5585 /* If setting CC0, record what it was set to, or a constant, if it
5586 is equivalent to a constant. If it is being set to a floating-point
5587 value, make a COMPARE with the appropriate constant of 0. If we
5588 don't do this, later code can interpret this as a test against
5589 const0_rtx, which can cause problems if we try to put it into an
5590 insn as a floating-point operand. */
5591 if (dest == cc0_rtx)
5593 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5594 this_insn_cc0_mode = mode;
5595 if (FLOAT_MODE_P (mode))
5596 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5597 CONST0_RTX (mode));
5601 /* Now enter all non-volatile source expressions in the hash table
5602 if they are not already present.
5603 Record their equivalence classes in src_elt.
5604 This way we can insert the corresponding destinations into
5605 the same classes even if the actual sources are no longer in them
5606 (having been invalidated). */
5608 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5609 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5611 struct table_elt *elt;
5612 struct table_elt *classp = sets[0].src_elt;
5613 rtx dest = SET_DEST (sets[0].rtl);
5614 machine_mode eqvmode = GET_MODE (dest);
5616 if (GET_CODE (dest) == STRICT_LOW_PART)
5618 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5619 classp = 0;
5621 if (insert_regs (src_eqv, classp, 0))
5623 rehash_using_reg (src_eqv);
5624 src_eqv_hash = HASH (src_eqv, eqvmode);
5626 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5627 elt->in_memory = src_eqv_in_memory;
5628 src_eqv_elt = elt;
5630 /* Check to see if src_eqv_elt is the same as a set source which
5631 does not yet have an elt, and if so set the elt of the set source
5632 to src_eqv_elt. */
5633 for (i = 0; i < n_sets; i++)
5634 if (sets[i].rtl && sets[i].src_elt == 0
5635 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5636 sets[i].src_elt = src_eqv_elt;
5639 for (i = 0; i < n_sets; i++)
5640 if (sets[i].rtl && ! sets[i].src_volatile
5641 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5643 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5645 /* REG_EQUAL in setting a STRICT_LOW_PART
5646 gives an equivalent for the entire destination register,
5647 not just for the subreg being stored in now.
5648 This is a more interesting equivalence, so we arrange later
5649 to treat the entire reg as the destination. */
5650 sets[i].src_elt = src_eqv_elt;
5651 sets[i].src_hash = src_eqv_hash;
5653 else
5655 /* Insert source and constant equivalent into hash table, if not
5656 already present. */
5657 struct table_elt *classp = src_eqv_elt;
5658 rtx src = sets[i].src;
5659 rtx dest = SET_DEST (sets[i].rtl);
5660 machine_mode mode
5661 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5663 /* It's possible that we have a source value known to be
5664 constant but don't have a REG_EQUAL note on the insn.
5665 Lack of a note will mean src_eqv_elt will be NULL. This
5666 can happen where we've generated a SUBREG to access a
5667 CONST_INT that is already in a register in a wider mode.
5668 Ensure that the source expression is put in the proper
5669 constant class. */
5670 if (!classp)
5671 classp = sets[i].src_const_elt;
5673 if (sets[i].src_elt == 0)
5675 struct table_elt *elt;
5677 /* Note that these insert_regs calls cannot remove
5678 any of the src_elt's, because they would have failed to
5679 match if not still valid. */
5680 if (insert_regs (src, classp, 0))
5682 rehash_using_reg (src);
5683 sets[i].src_hash = HASH (src, mode);
5685 elt = insert (src, classp, sets[i].src_hash, mode);
5686 elt->in_memory = sets[i].src_in_memory;
5687 /* If inline asm has any clobbers, ensure we only reuse
5688 existing inline asms and never try to put the ASM_OPERANDS
5689 into an insn that isn't inline asm. */
5690 if (GET_CODE (src) == ASM_OPERANDS
5691 && GET_CODE (x) == PARALLEL)
5692 elt->cost = MAX_COST;
5693 sets[i].src_elt = classp = elt;
5695 if (sets[i].src_const && sets[i].src_const_elt == 0
5696 && src != sets[i].src_const
5697 && ! rtx_equal_p (sets[i].src_const, src))
5698 sets[i].src_elt = insert (sets[i].src_const, classp,
5699 sets[i].src_const_hash, mode);
5702 else if (sets[i].src_elt == 0)
5703 /* If we did not insert the source into the hash table (e.g., it was
5704 volatile), note the equivalence class for the REG_EQUAL value, if any,
5705 so that the destination goes into that class. */
5706 sets[i].src_elt = src_eqv_elt;
5708 /* Record destination addresses in the hash table. This allows us to
5709 check if they are invalidated by other sets. */
5710 for (i = 0; i < n_sets; i++)
5712 if (sets[i].rtl)
5714 rtx x = sets[i].inner_dest;
5715 struct table_elt *elt;
5716 machine_mode mode;
5717 unsigned hash;
5719 if (MEM_P (x))
5721 x = XEXP (x, 0);
5722 mode = GET_MODE (x);
5723 hash = HASH (x, mode);
5724 elt = lookup (x, hash, mode);
5725 if (!elt)
5727 if (insert_regs (x, NULL, 0))
5729 rtx dest = SET_DEST (sets[i].rtl);
5731 rehash_using_reg (x);
5732 hash = HASH (x, mode);
5733 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5735 elt = insert (x, NULL, hash, mode);
5738 sets[i].dest_addr_elt = elt;
5740 else
5741 sets[i].dest_addr_elt = NULL;
5745 invalidate_from_clobbers (insn);
5747 /* Some registers are invalidated by subroutine calls. Memory is
5748 invalidated by non-constant calls. */
5750 if (CALL_P (insn))
5752 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5753 invalidate_memory ();
5754 invalidate_for_call ();
5757 /* Now invalidate everything set by this instruction.
5758 If a SUBREG or other funny destination is being set,
5759 sets[i].rtl is still nonzero, so here we invalidate the reg
5760 a part of which is being set. */
5762 for (i = 0; i < n_sets; i++)
5763 if (sets[i].rtl)
5765 /* We can't use the inner dest, because the mode associated with
5766 a ZERO_EXTRACT is significant. */
5767 rtx dest = SET_DEST (sets[i].rtl);
5769 /* Needed for registers to remove the register from its
5770 previous quantity's chain.
5771 Needed for memory if this is a nonvarying address, unless
5772 we have just done an invalidate_memory that covers even those. */
5773 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5774 invalidate (dest, VOIDmode);
5775 else if (MEM_P (dest))
5776 invalidate (dest, VOIDmode);
5777 else if (GET_CODE (dest) == STRICT_LOW_PART
5778 || GET_CODE (dest) == ZERO_EXTRACT)
5779 invalidate (XEXP (dest, 0), GET_MODE (dest));
5782 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5783 the regs restored by the longjmp come from a later time
5784 than the setjmp. */
5785 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5787 flush_hash_table ();
5788 goto done;
5791 /* Make sure registers mentioned in destinations
5792 are safe for use in an expression to be inserted.
5793 This removes from the hash table
5794 any invalid entry that refers to one of these registers.
5796 We don't care about the return value from mention_regs because
5797 we are going to hash the SET_DEST values unconditionally. */
5799 for (i = 0; i < n_sets; i++)
5801 if (sets[i].rtl)
5803 rtx x = SET_DEST (sets[i].rtl);
5805 if (!REG_P (x))
5806 mention_regs (x);
5807 else
5809 /* We used to rely on all references to a register becoming
5810 inaccessible when a register changes to a new quantity,
5811 since that changes the hash code. However, that is not
5812 safe, since after HASH_SIZE new quantities we get a
5813 hash 'collision' of a register with its own invalid
5814 entries. And since SUBREGs have been changed not to
5815 change their hash code with the hash code of the register,
5816 it wouldn't work any longer at all. So we have to check
5817 for any invalid references lying around now.
5818 This code is similar to the REG case in mention_regs,
5819 but it knows that reg_tick has been incremented, and
5820 it leaves reg_in_table as -1 . */
5821 unsigned int regno = REGNO (x);
5822 unsigned int endregno = END_REGNO (x);
5823 unsigned int i;
5825 for (i = regno; i < endregno; i++)
5827 if (REG_IN_TABLE (i) >= 0)
5829 remove_invalid_refs (i);
5830 REG_IN_TABLE (i) = -1;
5837 /* We may have just removed some of the src_elt's from the hash table.
5838 So replace each one with the current head of the same class.
5839 Also check if destination addresses have been removed. */
5841 for (i = 0; i < n_sets; i++)
5842 if (sets[i].rtl)
5844 if (sets[i].dest_addr_elt
5845 && sets[i].dest_addr_elt->first_same_value == 0)
5847 /* The elt was removed, which means this destination is not
5848 valid after this instruction. */
5849 sets[i].rtl = NULL_RTX;
5851 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5852 /* If elt was removed, find current head of same class,
5853 or 0 if nothing remains of that class. */
5855 struct table_elt *elt = sets[i].src_elt;
5857 while (elt && elt->prev_same_value)
5858 elt = elt->prev_same_value;
5860 while (elt && elt->first_same_value == 0)
5861 elt = elt->next_same_value;
5862 sets[i].src_elt = elt ? elt->first_same_value : 0;
5866 /* Now insert the destinations into their equivalence classes. */
5868 for (i = 0; i < n_sets; i++)
5869 if (sets[i].rtl)
5871 rtx dest = SET_DEST (sets[i].rtl);
5872 struct table_elt *elt;
5874 /* Don't record value if we are not supposed to risk allocating
5875 floating-point values in registers that might be wider than
5876 memory. */
5877 if ((flag_float_store
5878 && MEM_P (dest)
5879 && FLOAT_MODE_P (GET_MODE (dest)))
5880 /* Don't record BLKmode values, because we don't know the
5881 size of it, and can't be sure that other BLKmode values
5882 have the same or smaller size. */
5883 || GET_MODE (dest) == BLKmode
5884 /* If we didn't put a REG_EQUAL value or a source into the hash
5885 table, there is no point is recording DEST. */
5886 || sets[i].src_elt == 0
5887 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5888 or SIGN_EXTEND, don't record DEST since it can cause
5889 some tracking to be wrong.
5891 ??? Think about this more later. */
5892 || (paradoxical_subreg_p (dest)
5893 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5894 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5895 continue;
5897 /* STRICT_LOW_PART isn't part of the value BEING set,
5898 and neither is the SUBREG inside it.
5899 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5900 if (GET_CODE (dest) == STRICT_LOW_PART)
5901 dest = SUBREG_REG (XEXP (dest, 0));
5903 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5904 /* Registers must also be inserted into chains for quantities. */
5905 if (insert_regs (dest, sets[i].src_elt, 1))
5907 /* If `insert_regs' changes something, the hash code must be
5908 recalculated. */
5909 rehash_using_reg (dest);
5910 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5913 elt = insert (dest, sets[i].src_elt,
5914 sets[i].dest_hash, GET_MODE (dest));
5916 /* If this is a constant, insert the constant anchors with the
5917 equivalent register-offset expressions using register DEST. */
5918 if (targetm.const_anchor
5919 && REG_P (dest)
5920 && SCALAR_INT_MODE_P (GET_MODE (dest))
5921 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5922 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5924 elt->in_memory = (MEM_P (sets[i].inner_dest)
5925 && !MEM_READONLY_P (sets[i].inner_dest));
5927 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5928 narrower than M2, and both M1 and M2 are the same number of words,
5929 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5930 make that equivalence as well.
5932 However, BAR may have equivalences for which gen_lowpart
5933 will produce a simpler value than gen_lowpart applied to
5934 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5935 BAR's equivalences. If we don't get a simplified form, make
5936 the SUBREG. It will not be used in an equivalence, but will
5937 cause two similar assignments to be detected.
5939 Note the loop below will find SUBREG_REG (DEST) since we have
5940 already entered SRC and DEST of the SET in the table. */
5942 if (GET_CODE (dest) == SUBREG
5943 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5944 / UNITS_PER_WORD)
5945 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5946 && (GET_MODE_SIZE (GET_MODE (dest))
5947 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5948 && sets[i].src_elt != 0)
5950 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5951 struct table_elt *elt, *classp = 0;
5953 for (elt = sets[i].src_elt->first_same_value; elt;
5954 elt = elt->next_same_value)
5956 rtx new_src = 0;
5957 unsigned src_hash;
5958 struct table_elt *src_elt;
5959 int byte = 0;
5961 /* Ignore invalid entries. */
5962 if (!REG_P (elt->exp)
5963 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5964 continue;
5966 /* We may have already been playing subreg games. If the
5967 mode is already correct for the destination, use it. */
5968 if (GET_MODE (elt->exp) == new_mode)
5969 new_src = elt->exp;
5970 else
5972 /* Calculate big endian correction for the SUBREG_BYTE.
5973 We have already checked that M1 (GET_MODE (dest))
5974 is not narrower than M2 (new_mode). */
5975 if (BYTES_BIG_ENDIAN)
5976 byte = (GET_MODE_SIZE (GET_MODE (dest))
5977 - GET_MODE_SIZE (new_mode));
5979 new_src = simplify_gen_subreg (new_mode, elt->exp,
5980 GET_MODE (dest), byte);
5983 /* The call to simplify_gen_subreg fails if the value
5984 is VOIDmode, yet we can't do any simplification, e.g.
5985 for EXPR_LISTs denoting function call results.
5986 It is invalid to construct a SUBREG with a VOIDmode
5987 SUBREG_REG, hence a zero new_src means we can't do
5988 this substitution. */
5989 if (! new_src)
5990 continue;
5992 src_hash = HASH (new_src, new_mode);
5993 src_elt = lookup (new_src, src_hash, new_mode);
5995 /* Put the new source in the hash table is if isn't
5996 already. */
5997 if (src_elt == 0)
5999 if (insert_regs (new_src, classp, 0))
6001 rehash_using_reg (new_src);
6002 src_hash = HASH (new_src, new_mode);
6004 src_elt = insert (new_src, classp, src_hash, new_mode);
6005 src_elt->in_memory = elt->in_memory;
6006 if (GET_CODE (new_src) == ASM_OPERANDS
6007 && elt->cost == MAX_COST)
6008 src_elt->cost = MAX_COST;
6010 else if (classp && classp != src_elt->first_same_value)
6011 /* Show that two things that we've seen before are
6012 actually the same. */
6013 merge_equiv_classes (src_elt, classp);
6015 classp = src_elt->first_same_value;
6016 /* Ignore invalid entries. */
6017 while (classp
6018 && !REG_P (classp->exp)
6019 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6020 classp = classp->next_same_value;
6025 /* Special handling for (set REG0 REG1) where REG0 is the
6026 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6027 be used in the sequel, so (if easily done) change this insn to
6028 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6029 that computed their value. Then REG1 will become a dead store
6030 and won't cloud the situation for later optimizations.
6032 Do not make this change if REG1 is a hard register, because it will
6033 then be used in the sequel and we may be changing a two-operand insn
6034 into a three-operand insn.
6036 Also do not do this if we are operating on a copy of INSN. */
6038 if (n_sets == 1 && sets[0].rtl)
6039 try_back_substitute_reg (sets[0].rtl, insn);
6041 done:;
6044 /* Remove from the hash table all expressions that reference memory. */
6046 static void
6047 invalidate_memory (void)
6049 int i;
6050 struct table_elt *p, *next;
6052 for (i = 0; i < HASH_SIZE; i++)
6053 for (p = table[i]; p; p = next)
6055 next = p->next_same_hash;
6056 if (p->in_memory)
6057 remove_from_table (p, i);
6061 /* Perform invalidation on the basis of everything about INSN,
6062 except for invalidating the actual places that are SET in it.
6063 This includes the places CLOBBERed, and anything that might
6064 alias with something that is SET or CLOBBERed. */
6066 static void
6067 invalidate_from_clobbers (rtx_insn *insn)
6069 rtx x = PATTERN (insn);
6071 if (GET_CODE (x) == CLOBBER)
6073 rtx ref = XEXP (x, 0);
6074 if (ref)
6076 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6077 || MEM_P (ref))
6078 invalidate (ref, VOIDmode);
6079 else if (GET_CODE (ref) == STRICT_LOW_PART
6080 || GET_CODE (ref) == ZERO_EXTRACT)
6081 invalidate (XEXP (ref, 0), GET_MODE (ref));
6084 else if (GET_CODE (x) == PARALLEL)
6086 int i;
6087 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6089 rtx y = XVECEXP (x, 0, i);
6090 if (GET_CODE (y) == CLOBBER)
6092 rtx ref = XEXP (y, 0);
6093 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6094 || MEM_P (ref))
6095 invalidate (ref, VOIDmode);
6096 else if (GET_CODE (ref) == STRICT_LOW_PART
6097 || GET_CODE (ref) == ZERO_EXTRACT)
6098 invalidate (XEXP (ref, 0), GET_MODE (ref));
6104 /* Perform invalidation on the basis of everything about INSN.
6105 This includes the places CLOBBERed, and anything that might
6106 alias with something that is SET or CLOBBERed. */
6108 static void
6109 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6111 rtx tem;
6112 rtx x = PATTERN (insn);
6114 if (CALL_P (insn))
6116 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6117 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6118 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6121 /* Ensure we invalidate the destination register of a CALL insn.
6122 This is necessary for machines where this register is a fixed_reg,
6123 because no other code would invalidate it. */
6124 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6125 invalidate (SET_DEST (x), VOIDmode);
6127 else if (GET_CODE (x) == PARALLEL)
6129 int i;
6131 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6133 rtx y = XVECEXP (x, 0, i);
6134 if (GET_CODE (y) == CLOBBER)
6136 rtx clobbered = XEXP (y, 0);
6138 if (REG_P (clobbered)
6139 || GET_CODE (clobbered) == SUBREG)
6140 invalidate (clobbered, VOIDmode);
6141 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6142 || GET_CODE (clobbered) == ZERO_EXTRACT)
6143 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6145 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6146 invalidate (SET_DEST (y), VOIDmode);
6151 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6152 and replace any registers in them with either an equivalent constant
6153 or the canonical form of the register. If we are inside an address,
6154 only do this if the address remains valid.
6156 OBJECT is 0 except when within a MEM in which case it is the MEM.
6158 Return the replacement for X. */
6160 static rtx
6161 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6163 enum rtx_code code = GET_CODE (x);
6164 const char *fmt = GET_RTX_FORMAT (code);
6165 int i;
6167 switch (code)
6169 case CONST:
6170 case SYMBOL_REF:
6171 case LABEL_REF:
6172 CASE_CONST_ANY:
6173 case PC:
6174 case CC0:
6175 case LO_SUM:
6176 return x;
6178 case MEM:
6179 validate_change (x, &XEXP (x, 0),
6180 cse_process_notes (XEXP (x, 0), x, changed), 0);
6181 return x;
6183 case EXPR_LIST:
6184 if (REG_NOTE_KIND (x) == REG_EQUAL)
6185 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6186 /* Fall through. */
6188 case INSN_LIST:
6189 case INT_LIST:
6190 if (XEXP (x, 1))
6191 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6192 return x;
6194 case SIGN_EXTEND:
6195 case ZERO_EXTEND:
6196 case SUBREG:
6198 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6199 /* We don't substitute VOIDmode constants into these rtx,
6200 since they would impede folding. */
6201 if (GET_MODE (new_rtx) != VOIDmode)
6202 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6203 return x;
6206 case UNSIGNED_FLOAT:
6208 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6209 /* We don't substitute negative VOIDmode constants into these rtx,
6210 since they would impede folding. */
6211 if (GET_MODE (new_rtx) != VOIDmode
6212 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6213 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6214 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6215 return x;
6218 case REG:
6219 i = REG_QTY (REGNO (x));
6221 /* Return a constant or a constant register. */
6222 if (REGNO_QTY_VALID_P (REGNO (x)))
6224 struct qty_table_elem *ent = &qty_table[i];
6226 if (ent->const_rtx != NULL_RTX
6227 && (CONSTANT_P (ent->const_rtx)
6228 || REG_P (ent->const_rtx)))
6230 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6231 if (new_rtx)
6232 return copy_rtx (new_rtx);
6236 /* Otherwise, canonicalize this register. */
6237 return canon_reg (x, NULL);
6239 default:
6240 break;
6243 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6244 if (fmt[i] == 'e')
6245 validate_change (object, &XEXP (x, i),
6246 cse_process_notes (XEXP (x, i), object, changed), 0);
6248 return x;
6251 static rtx
6252 cse_process_notes (rtx x, rtx object, bool *changed)
6254 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6255 if (new_rtx != x)
6256 *changed = true;
6257 return new_rtx;
6261 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6263 DATA is a pointer to a struct cse_basic_block_data, that is used to
6264 describe the path.
6265 It is filled with a queue of basic blocks, starting with FIRST_BB
6266 and following a trace through the CFG.
6268 If all paths starting at FIRST_BB have been followed, or no new path
6269 starting at FIRST_BB can be constructed, this function returns FALSE.
6270 Otherwise, DATA->path is filled and the function returns TRUE indicating
6271 that a path to follow was found.
6273 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6274 block in the path will be FIRST_BB. */
6276 static bool
6277 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6278 int follow_jumps)
6280 basic_block bb;
6281 edge e;
6282 int path_size;
6284 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6286 /* See if there is a previous path. */
6287 path_size = data->path_size;
6289 /* There is a previous path. Make sure it started with FIRST_BB. */
6290 if (path_size)
6291 gcc_assert (data->path[0].bb == first_bb);
6293 /* There was only one basic block in the last path. Clear the path and
6294 return, so that paths starting at another basic block can be tried. */
6295 if (path_size == 1)
6297 path_size = 0;
6298 goto done;
6301 /* If the path was empty from the beginning, construct a new path. */
6302 if (path_size == 0)
6303 data->path[path_size++].bb = first_bb;
6304 else
6306 /* Otherwise, path_size must be equal to or greater than 2, because
6307 a previous path exists that is at least two basic blocks long.
6309 Update the previous branch path, if any. If the last branch was
6310 previously along the branch edge, take the fallthrough edge now. */
6311 while (path_size >= 2)
6313 basic_block last_bb_in_path, previous_bb_in_path;
6314 edge e;
6316 --path_size;
6317 last_bb_in_path = data->path[path_size].bb;
6318 previous_bb_in_path = data->path[path_size - 1].bb;
6320 /* If we previously followed a path along the branch edge, try
6321 the fallthru edge now. */
6322 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6323 && any_condjump_p (BB_END (previous_bb_in_path))
6324 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6325 && e == BRANCH_EDGE (previous_bb_in_path))
6327 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6328 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6329 && single_pred_p (bb)
6330 /* We used to assert here that we would only see blocks
6331 that we have not visited yet. But we may end up
6332 visiting basic blocks twice if the CFG has changed
6333 in this run of cse_main, because when the CFG changes
6334 the topological sort of the CFG also changes. A basic
6335 blocks that previously had more than two predecessors
6336 may now have a single predecessor, and become part of
6337 a path that starts at another basic block.
6339 We still want to visit each basic block only once, so
6340 halt the path here if we have already visited BB. */
6341 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6343 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6344 data->path[path_size++].bb = bb;
6345 break;
6349 data->path[path_size].bb = NULL;
6352 /* If only one block remains in the path, bail. */
6353 if (path_size == 1)
6355 path_size = 0;
6356 goto done;
6360 /* Extend the path if possible. */
6361 if (follow_jumps)
6363 bb = data->path[path_size - 1].bb;
6364 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6366 if (single_succ_p (bb))
6367 e = single_succ_edge (bb);
6368 else if (EDGE_COUNT (bb->succs) == 2
6369 && any_condjump_p (BB_END (bb)))
6371 /* First try to follow the branch. If that doesn't lead
6372 to a useful path, follow the fallthru edge. */
6373 e = BRANCH_EDGE (bb);
6374 if (!single_pred_p (e->dest))
6375 e = FALLTHRU_EDGE (bb);
6377 else
6378 e = NULL;
6380 if (e
6381 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6382 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6383 && single_pred_p (e->dest)
6384 /* Avoid visiting basic blocks twice. The large comment
6385 above explains why this can happen. */
6386 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6388 basic_block bb2 = e->dest;
6389 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6390 data->path[path_size++].bb = bb2;
6391 bb = bb2;
6393 else
6394 bb = NULL;
6398 done:
6399 data->path_size = path_size;
6400 return path_size != 0;
6403 /* Dump the path in DATA to file F. NSETS is the number of sets
6404 in the path. */
6406 static void
6407 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6409 int path_entry;
6411 fprintf (f, ";; Following path with %d sets: ", nsets);
6412 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6413 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6414 fputc ('\n', dump_file);
6415 fflush (f);
6419 /* Return true if BB has exception handling successor edges. */
6421 static bool
6422 have_eh_succ_edges (basic_block bb)
6424 edge e;
6425 edge_iterator ei;
6427 FOR_EACH_EDGE (e, ei, bb->succs)
6428 if (e->flags & EDGE_EH)
6429 return true;
6431 return false;
6435 /* Scan to the end of the path described by DATA. Return an estimate of
6436 the total number of SETs of all insns in the path. */
6438 static void
6439 cse_prescan_path (struct cse_basic_block_data *data)
6441 int nsets = 0;
6442 int path_size = data->path_size;
6443 int path_entry;
6445 /* Scan to end of each basic block in the path. */
6446 for (path_entry = 0; path_entry < path_size; path_entry++)
6448 basic_block bb;
6449 rtx_insn *insn;
6451 bb = data->path[path_entry].bb;
6453 FOR_BB_INSNS (bb, insn)
6455 if (!INSN_P (insn))
6456 continue;
6458 /* A PARALLEL can have lots of SETs in it,
6459 especially if it is really an ASM_OPERANDS. */
6460 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6461 nsets += XVECLEN (PATTERN (insn), 0);
6462 else
6463 nsets += 1;
6467 data->nsets = nsets;
6470 /* Return true if the pattern of INSN uses a LABEL_REF for which
6471 there isn't a REG_LABEL_OPERAND note. */
6473 static bool
6474 check_for_label_ref (rtx_insn *insn)
6476 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6477 note for it, we must rerun jump since it needs to place the note. If
6478 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6479 don't do this since no REG_LABEL_OPERAND will be added. */
6480 subrtx_iterator::array_type array;
6481 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6483 const_rtx x = *iter;
6484 if (GET_CODE (x) == LABEL_REF
6485 && !LABEL_REF_NONLOCAL_P (x)
6486 && (!JUMP_P (insn)
6487 || !label_is_jump_target_p (LABEL_REF_LABEL (x), insn))
6488 && LABEL_P (LABEL_REF_LABEL (x))
6489 && INSN_UID (LABEL_REF_LABEL (x)) != 0
6490 && !find_reg_note (insn, REG_LABEL_OPERAND, LABEL_REF_LABEL (x)))
6491 return true;
6493 return false;
6496 /* Process a single extended basic block described by EBB_DATA. */
6498 static void
6499 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6501 int path_size = ebb_data->path_size;
6502 int path_entry;
6503 int num_insns = 0;
6505 /* Allocate the space needed by qty_table. */
6506 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6508 new_basic_block ();
6509 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6510 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6511 for (path_entry = 0; path_entry < path_size; path_entry++)
6513 basic_block bb;
6514 rtx_insn *insn;
6516 bb = ebb_data->path[path_entry].bb;
6518 /* Invalidate recorded information for eh regs if there is an EH
6519 edge pointing to that bb. */
6520 if (bb_has_eh_pred (bb))
6522 df_ref def;
6524 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6525 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6526 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6529 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6530 FOR_BB_INSNS (bb, insn)
6532 /* If we have processed 1,000 insns, flush the hash table to
6533 avoid extreme quadratic behavior. We must not include NOTEs
6534 in the count since there may be more of them when generating
6535 debugging information. If we clear the table at different
6536 times, code generated with -g -O might be different than code
6537 generated with -O but not -g.
6539 FIXME: This is a real kludge and needs to be done some other
6540 way. */
6541 if (NONDEBUG_INSN_P (insn)
6542 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6544 flush_hash_table ();
6545 num_insns = 0;
6548 if (INSN_P (insn))
6550 /* Process notes first so we have all notes in canonical forms
6551 when looking for duplicate operations. */
6552 if (REG_NOTES (insn))
6554 bool changed = false;
6555 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6556 NULL_RTX, &changed);
6557 if (changed)
6558 df_notes_rescan (insn);
6561 cse_insn (insn);
6563 /* If we haven't already found an insn where we added a LABEL_REF,
6564 check this one. */
6565 if (INSN_P (insn) && !recorded_label_ref
6566 && check_for_label_ref (insn))
6567 recorded_label_ref = true;
6569 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6571 /* If the previous insn sets CC0 and this insn no
6572 longer references CC0, delete the previous insn.
6573 Here we use fact that nothing expects CC0 to be
6574 valid over an insn, which is true until the final
6575 pass. */
6576 rtx_insn *prev_insn;
6577 rtx tem;
6579 prev_insn = prev_nonnote_nondebug_insn (insn);
6580 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6581 && (tem = single_set (prev_insn)) != NULL_RTX
6582 && SET_DEST (tem) == cc0_rtx
6583 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6584 delete_insn (prev_insn);
6586 /* If this insn is not the last insn in the basic
6587 block, it will be PREV_INSN(insn) in the next
6588 iteration. If we recorded any CC0-related
6589 information for this insn, remember it. */
6590 if (insn != BB_END (bb))
6592 prev_insn_cc0 = this_insn_cc0;
6593 prev_insn_cc0_mode = this_insn_cc0_mode;
6599 /* With non-call exceptions, we are not always able to update
6600 the CFG properly inside cse_insn. So clean up possibly
6601 redundant EH edges here. */
6602 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6603 cse_cfg_altered |= purge_dead_edges (bb);
6605 /* If we changed a conditional jump, we may have terminated
6606 the path we are following. Check that by verifying that
6607 the edge we would take still exists. If the edge does
6608 not exist anymore, purge the remainder of the path.
6609 Note that this will cause us to return to the caller. */
6610 if (path_entry < path_size - 1)
6612 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6613 if (!find_edge (bb, next_bb))
6617 path_size--;
6619 /* If we truncate the path, we must also reset the
6620 visited bit on the remaining blocks in the path,
6621 or we will never visit them at all. */
6622 bitmap_clear_bit (cse_visited_basic_blocks,
6623 ebb_data->path[path_size].bb->index);
6624 ebb_data->path[path_size].bb = NULL;
6626 while (path_size - 1 != path_entry);
6627 ebb_data->path_size = path_size;
6631 /* If this is a conditional jump insn, record any known
6632 equivalences due to the condition being tested. */
6633 insn = BB_END (bb);
6634 if (path_entry < path_size - 1
6635 && JUMP_P (insn)
6636 && single_set (insn)
6637 && any_condjump_p (insn))
6639 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6640 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6641 record_jump_equiv (insn, taken);
6644 /* Clear the CC0-tracking related insns, they can't provide
6645 useful information across basic block boundaries. */
6646 prev_insn_cc0 = 0;
6649 gcc_assert (next_qty <= max_qty);
6651 free (qty_table);
6655 /* Perform cse on the instructions of a function.
6656 F is the first instruction.
6657 NREGS is one plus the highest pseudo-reg number used in the instruction.
6659 Return 2 if jump optimizations should be redone due to simplifications
6660 in conditional jump instructions.
6661 Return 1 if the CFG should be cleaned up because it has been modified.
6662 Return 0 otherwise. */
6664 static int
6665 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6667 struct cse_basic_block_data ebb_data;
6668 basic_block bb;
6669 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6670 int i, n_blocks;
6672 /* CSE doesn't use dominane info but can invalidate it in different ways.
6673 For simplicity free dominance info here. */
6674 free_dominance_info (CDI_DOMINATORS);
6676 df_set_flags (DF_LR_RUN_DCE);
6677 df_note_add_problem ();
6678 df_analyze ();
6679 df_set_flags (DF_DEFER_INSN_RESCAN);
6681 reg_scan (get_insns (), max_reg_num ());
6682 init_cse_reg_info (nregs);
6684 ebb_data.path = XNEWVEC (struct branch_path,
6685 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6687 cse_cfg_altered = false;
6688 cse_jumps_altered = false;
6689 recorded_label_ref = false;
6690 constant_pool_entries_cost = 0;
6691 constant_pool_entries_regcost = 0;
6692 ebb_data.path_size = 0;
6693 ebb_data.nsets = 0;
6694 rtl_hooks = cse_rtl_hooks;
6696 init_recog ();
6697 init_alias_analysis ();
6699 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6701 /* Set up the table of already visited basic blocks. */
6702 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6703 bitmap_clear (cse_visited_basic_blocks);
6705 /* Loop over basic blocks in reverse completion order (RPO),
6706 excluding the ENTRY and EXIT blocks. */
6707 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6708 i = 0;
6709 while (i < n_blocks)
6711 /* Find the first block in the RPO queue that we have not yet
6712 processed before. */
6715 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6717 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6718 && i < n_blocks);
6720 /* Find all paths starting with BB, and process them. */
6721 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6723 /* Pre-scan the path. */
6724 cse_prescan_path (&ebb_data);
6726 /* If this basic block has no sets, skip it. */
6727 if (ebb_data.nsets == 0)
6728 continue;
6730 /* Get a reasonable estimate for the maximum number of qty's
6731 needed for this path. For this, we take the number of sets
6732 and multiply that by MAX_RECOG_OPERANDS. */
6733 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6735 /* Dump the path we're about to process. */
6736 if (dump_file)
6737 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6739 cse_extended_basic_block (&ebb_data);
6743 /* Clean up. */
6744 end_alias_analysis ();
6745 free (reg_eqv_table);
6746 free (ebb_data.path);
6747 sbitmap_free (cse_visited_basic_blocks);
6748 free (rc_order);
6749 rtl_hooks = general_rtl_hooks;
6751 if (cse_jumps_altered || recorded_label_ref)
6752 return 2;
6753 else if (cse_cfg_altered)
6754 return 1;
6755 else
6756 return 0;
6759 /* Count the number of times registers are used (not set) in X.
6760 COUNTS is an array in which we accumulate the count, INCR is how much
6761 we count each register usage.
6763 Don't count a usage of DEST, which is the SET_DEST of a SET which
6764 contains X in its SET_SRC. This is because such a SET does not
6765 modify the liveness of DEST.
6766 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6767 We must then count uses of a SET_DEST regardless, because the insn can't be
6768 deleted here. */
6770 static void
6771 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6773 enum rtx_code code;
6774 rtx note;
6775 const char *fmt;
6776 int i, j;
6778 if (x == 0)
6779 return;
6781 switch (code = GET_CODE (x))
6783 case REG:
6784 if (x != dest)
6785 counts[REGNO (x)] += incr;
6786 return;
6788 case PC:
6789 case CC0:
6790 case CONST:
6791 CASE_CONST_ANY:
6792 case SYMBOL_REF:
6793 case LABEL_REF:
6794 return;
6796 case CLOBBER:
6797 /* If we are clobbering a MEM, mark any registers inside the address
6798 as being used. */
6799 if (MEM_P (XEXP (x, 0)))
6800 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6801 return;
6803 case SET:
6804 /* Unless we are setting a REG, count everything in SET_DEST. */
6805 if (!REG_P (SET_DEST (x)))
6806 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6807 count_reg_usage (SET_SRC (x), counts,
6808 dest ? dest : SET_DEST (x),
6809 incr);
6810 return;
6812 case DEBUG_INSN:
6813 return;
6815 case CALL_INSN:
6816 case INSN:
6817 case JUMP_INSN:
6818 /* We expect dest to be NULL_RTX here. If the insn may throw,
6819 or if it cannot be deleted due to side-effects, mark this fact
6820 by setting DEST to pc_rtx. */
6821 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6822 || side_effects_p (PATTERN (x)))
6823 dest = pc_rtx;
6824 if (code == CALL_INSN)
6825 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6826 count_reg_usage (PATTERN (x), counts, dest, incr);
6828 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6829 use them. */
6831 note = find_reg_equal_equiv_note (x);
6832 if (note)
6834 rtx eqv = XEXP (note, 0);
6836 if (GET_CODE (eqv) == EXPR_LIST)
6837 /* This REG_EQUAL note describes the result of a function call.
6838 Process all the arguments. */
6841 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6842 eqv = XEXP (eqv, 1);
6844 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6845 else
6846 count_reg_usage (eqv, counts, dest, incr);
6848 return;
6850 case EXPR_LIST:
6851 if (REG_NOTE_KIND (x) == REG_EQUAL
6852 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6853 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6854 involving registers in the address. */
6855 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6856 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6858 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6859 return;
6861 case ASM_OPERANDS:
6862 /* Iterate over just the inputs, not the constraints as well. */
6863 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6864 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6865 return;
6867 case INSN_LIST:
6868 case INT_LIST:
6869 gcc_unreachable ();
6871 default:
6872 break;
6875 fmt = GET_RTX_FORMAT (code);
6876 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6878 if (fmt[i] == 'e')
6879 count_reg_usage (XEXP (x, i), counts, dest, incr);
6880 else if (fmt[i] == 'E')
6881 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6882 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6886 /* Return true if X is a dead register. */
6888 static inline int
6889 is_dead_reg (const_rtx x, int *counts)
6891 return (REG_P (x)
6892 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6893 && counts[REGNO (x)] == 0);
6896 /* Return true if set is live. */
6897 static bool
6898 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6899 int *counts)
6901 rtx_insn *tem;
6903 if (set_noop_p (set))
6906 else if (GET_CODE (SET_DEST (set)) == CC0
6907 && !side_effects_p (SET_SRC (set))
6908 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6909 || !INSN_P (tem)
6910 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6911 return false;
6912 else if (!is_dead_reg (SET_DEST (set), counts)
6913 || side_effects_p (SET_SRC (set)))
6914 return true;
6915 return false;
6918 /* Return true if insn is live. */
6920 static bool
6921 insn_live_p (rtx_insn *insn, int *counts)
6923 int i;
6924 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6925 return true;
6926 else if (GET_CODE (PATTERN (insn)) == SET)
6927 return set_live_p (PATTERN (insn), insn, counts);
6928 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6930 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6932 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6934 if (GET_CODE (elt) == SET)
6936 if (set_live_p (elt, insn, counts))
6937 return true;
6939 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6940 return true;
6942 return false;
6944 else if (DEBUG_INSN_P (insn))
6946 rtx_insn *next;
6948 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6949 if (NOTE_P (next))
6950 continue;
6951 else if (!DEBUG_INSN_P (next))
6952 return true;
6953 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6954 return false;
6956 return true;
6958 else
6959 return true;
6962 /* Count the number of stores into pseudo. Callback for note_stores. */
6964 static void
6965 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6967 int *counts = (int *) data;
6968 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6969 counts[REGNO (x)]++;
6972 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6973 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6974 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6975 Set *SEEN_REPL to true if we see a dead register that does have
6976 a replacement. */
6978 static bool
6979 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6980 bool *seen_repl)
6982 subrtx_iterator::array_type array;
6983 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6985 const_rtx x = *iter;
6986 if (is_dead_reg (x, counts))
6988 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6989 *seen_repl = true;
6990 else
6991 return true;
6994 return false;
6997 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6998 Callback for simplify_replace_fn_rtx. */
7000 static rtx
7001 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
7003 rtx *replacements = (rtx *) data;
7005 if (REG_P (x)
7006 && REGNO (x) >= FIRST_PSEUDO_REGISTER
7007 && replacements[REGNO (x)] != NULL_RTX)
7009 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
7010 return replacements[REGNO (x)];
7011 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
7012 GET_MODE (replacements[REGNO (x)]));
7014 return NULL_RTX;
7017 /* Scan all the insns and delete any that are dead; i.e., they store a register
7018 that is never used or they copy a register to itself.
7020 This is used to remove insns made obviously dead by cse, loop or other
7021 optimizations. It improves the heuristics in loop since it won't try to
7022 move dead invariants out of loops or make givs for dead quantities. The
7023 remaining passes of the compilation are also sped up. */
7026 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
7028 int *counts;
7029 rtx_insn *insn, *prev;
7030 rtx *replacements = NULL;
7031 int ndead = 0;
7033 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7034 /* First count the number of times each register is used. */
7035 if (MAY_HAVE_DEBUG_INSNS)
7037 counts = XCNEWVEC (int, nreg * 3);
7038 for (insn = insns; insn; insn = NEXT_INSN (insn))
7039 if (DEBUG_INSN_P (insn))
7040 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7041 NULL_RTX, 1);
7042 else if (INSN_P (insn))
7044 count_reg_usage (insn, counts, NULL_RTX, 1);
7045 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7047 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7048 First one counts how many times each pseudo is used outside
7049 of debug insns, second counts how many times each pseudo is
7050 used in debug insns and third counts how many times a pseudo
7051 is stored. */
7053 else
7055 counts = XCNEWVEC (int, nreg);
7056 for (insn = insns; insn; insn = NEXT_INSN (insn))
7057 if (INSN_P (insn))
7058 count_reg_usage (insn, counts, NULL_RTX, 1);
7059 /* If no debug insns can be present, COUNTS is just an array
7060 which counts how many times each pseudo is used. */
7062 /* Pseudo PIC register should be considered as used due to possible
7063 new usages generated. */
7064 if (!reload_completed
7065 && pic_offset_table_rtx
7066 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7067 counts[REGNO (pic_offset_table_rtx)]++;
7068 /* Go from the last insn to the first and delete insns that only set unused
7069 registers or copy a register to itself. As we delete an insn, remove
7070 usage counts for registers it uses.
7072 The first jump optimization pass may leave a real insn as the last
7073 insn in the function. We must not skip that insn or we may end
7074 up deleting code that is not really dead.
7076 If some otherwise unused register is only used in DEBUG_INSNs,
7077 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7078 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7079 has been created for the unused register, replace it with
7080 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7081 for (insn = get_last_insn (); insn; insn = prev)
7083 int live_insn = 0;
7085 prev = PREV_INSN (insn);
7086 if (!INSN_P (insn))
7087 continue;
7089 live_insn = insn_live_p (insn, counts);
7091 /* If this is a dead insn, delete it and show registers in it aren't
7092 being used. */
7094 if (! live_insn && dbg_cnt (delete_trivial_dead))
7096 if (DEBUG_INSN_P (insn))
7097 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7098 NULL_RTX, -1);
7099 else
7101 rtx set;
7102 if (MAY_HAVE_DEBUG_INSNS
7103 && (set = single_set (insn)) != NULL_RTX
7104 && is_dead_reg (SET_DEST (set), counts)
7105 /* Used at least once in some DEBUG_INSN. */
7106 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7107 /* And set exactly once. */
7108 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7109 && !side_effects_p (SET_SRC (set))
7110 && asm_noperands (PATTERN (insn)) < 0)
7112 rtx dval, bind_var_loc;
7113 rtx_insn *bind;
7115 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7116 dval = make_debug_expr_from_rtl (SET_DEST (set));
7118 /* Emit a debug bind insn before the insn in which
7119 reg dies. */
7120 bind_var_loc =
7121 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7122 DEBUG_EXPR_TREE_DECL (dval),
7123 SET_SRC (set),
7124 VAR_INIT_STATUS_INITIALIZED);
7125 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7127 bind = emit_debug_insn_before (bind_var_loc, insn);
7128 df_insn_rescan (bind);
7130 if (replacements == NULL)
7131 replacements = XCNEWVEC (rtx, nreg);
7132 replacements[REGNO (SET_DEST (set))] = dval;
7135 count_reg_usage (insn, counts, NULL_RTX, -1);
7136 ndead++;
7138 cse_cfg_altered |= delete_insn_and_edges (insn);
7142 if (MAY_HAVE_DEBUG_INSNS)
7144 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7145 if (DEBUG_INSN_P (insn))
7147 /* If this debug insn references a dead register that wasn't replaced
7148 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7149 bool seen_repl = false;
7150 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7151 counts, replacements, &seen_repl))
7153 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7154 df_insn_rescan (insn);
7156 else if (seen_repl)
7158 INSN_VAR_LOCATION_LOC (insn)
7159 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7160 NULL_RTX, replace_dead_reg,
7161 replacements);
7162 df_insn_rescan (insn);
7165 free (replacements);
7168 if (dump_file && ndead)
7169 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7170 ndead);
7171 /* Clean up. */
7172 free (counts);
7173 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7174 return ndead;
7177 /* If LOC contains references to NEWREG in a different mode, change them
7178 to use NEWREG instead. */
7180 static void
7181 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7182 rtx *loc, rtx_insn *insn, rtx newreg)
7184 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7186 rtx *loc = *iter;
7187 rtx x = *loc;
7188 if (x
7189 && REG_P (x)
7190 && REGNO (x) == REGNO (newreg)
7191 && GET_MODE (x) != GET_MODE (newreg))
7193 validate_change (insn, loc, newreg, 1);
7194 iter.skip_subrtxes ();
7199 /* Change the mode of any reference to the register REGNO (NEWREG) to
7200 GET_MODE (NEWREG) in INSN. */
7202 static void
7203 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7205 int success;
7207 if (!INSN_P (insn))
7208 return;
7210 subrtx_ptr_iterator::array_type array;
7211 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7212 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7214 /* If the following assertion was triggered, there is most probably
7215 something wrong with the cc_modes_compatible back end function.
7216 CC modes only can be considered compatible if the insn - with the mode
7217 replaced by any of the compatible modes - can still be recognized. */
7218 success = apply_change_group ();
7219 gcc_assert (success);
7222 /* Change the mode of any reference to the register REGNO (NEWREG) to
7223 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7224 any instruction which modifies NEWREG. */
7226 static void
7227 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7229 rtx_insn *insn;
7231 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7233 if (! INSN_P (insn))
7234 continue;
7236 if (reg_set_p (newreg, insn))
7237 return;
7239 cse_change_cc_mode_insn (insn, newreg);
7243 /* BB is a basic block which finishes with CC_REG as a condition code
7244 register which is set to CC_SRC. Look through the successors of BB
7245 to find blocks which have a single predecessor (i.e., this one),
7246 and look through those blocks for an assignment to CC_REG which is
7247 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7248 permitted to change the mode of CC_SRC to a compatible mode. This
7249 returns VOIDmode if no equivalent assignments were found.
7250 Otherwise it returns the mode which CC_SRC should wind up with.
7251 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7252 but is passed unmodified down to recursive calls in order to prevent
7253 endless recursion.
7255 The main complexity in this function is handling the mode issues.
7256 We may have more than one duplicate which we can eliminate, and we
7257 try to find a mode which will work for multiple duplicates. */
7259 static machine_mode
7260 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7261 bool can_change_mode)
7263 bool found_equiv;
7264 machine_mode mode;
7265 unsigned int insn_count;
7266 edge e;
7267 rtx_insn *insns[2];
7268 machine_mode modes[2];
7269 rtx_insn *last_insns[2];
7270 unsigned int i;
7271 rtx newreg;
7272 edge_iterator ei;
7274 /* We expect to have two successors. Look at both before picking
7275 the final mode for the comparison. If we have more successors
7276 (i.e., some sort of table jump, although that seems unlikely),
7277 then we require all beyond the first two to use the same
7278 mode. */
7280 found_equiv = false;
7281 mode = GET_MODE (cc_src);
7282 insn_count = 0;
7283 FOR_EACH_EDGE (e, ei, bb->succs)
7285 rtx_insn *insn;
7286 rtx_insn *end;
7288 if (e->flags & EDGE_COMPLEX)
7289 continue;
7291 if (EDGE_COUNT (e->dest->preds) != 1
7292 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7293 /* Avoid endless recursion on unreachable blocks. */
7294 || e->dest == orig_bb)
7295 continue;
7297 end = NEXT_INSN (BB_END (e->dest));
7298 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7300 rtx set;
7302 if (! INSN_P (insn))
7303 continue;
7305 /* If CC_SRC is modified, we have to stop looking for
7306 something which uses it. */
7307 if (modified_in_p (cc_src, insn))
7308 break;
7310 /* Check whether INSN sets CC_REG to CC_SRC. */
7311 set = single_set (insn);
7312 if (set
7313 && REG_P (SET_DEST (set))
7314 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7316 bool found;
7317 machine_mode set_mode;
7318 machine_mode comp_mode;
7320 found = false;
7321 set_mode = GET_MODE (SET_SRC (set));
7322 comp_mode = set_mode;
7323 if (rtx_equal_p (cc_src, SET_SRC (set)))
7324 found = true;
7325 else if (GET_CODE (cc_src) == COMPARE
7326 && GET_CODE (SET_SRC (set)) == COMPARE
7327 && mode != set_mode
7328 && rtx_equal_p (XEXP (cc_src, 0),
7329 XEXP (SET_SRC (set), 0))
7330 && rtx_equal_p (XEXP (cc_src, 1),
7331 XEXP (SET_SRC (set), 1)))
7334 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7335 if (comp_mode != VOIDmode
7336 && (can_change_mode || comp_mode == mode))
7337 found = true;
7340 if (found)
7342 found_equiv = true;
7343 if (insn_count < ARRAY_SIZE (insns))
7345 insns[insn_count] = insn;
7346 modes[insn_count] = set_mode;
7347 last_insns[insn_count] = end;
7348 ++insn_count;
7350 if (mode != comp_mode)
7352 gcc_assert (can_change_mode);
7353 mode = comp_mode;
7355 /* The modified insn will be re-recognized later. */
7356 PUT_MODE (cc_src, mode);
7359 else
7361 if (set_mode != mode)
7363 /* We found a matching expression in the
7364 wrong mode, but we don't have room to
7365 store it in the array. Punt. This case
7366 should be rare. */
7367 break;
7369 /* INSN sets CC_REG to a value equal to CC_SRC
7370 with the right mode. We can simply delete
7371 it. */
7372 delete_insn (insn);
7375 /* We found an instruction to delete. Keep looking,
7376 in the hopes of finding a three-way jump. */
7377 continue;
7380 /* We found an instruction which sets the condition
7381 code, so don't look any farther. */
7382 break;
7385 /* If INSN sets CC_REG in some other way, don't look any
7386 farther. */
7387 if (reg_set_p (cc_reg, insn))
7388 break;
7391 /* If we fell off the bottom of the block, we can keep looking
7392 through successors. We pass CAN_CHANGE_MODE as false because
7393 we aren't prepared to handle compatibility between the
7394 further blocks and this block. */
7395 if (insn == end)
7397 machine_mode submode;
7399 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7400 if (submode != VOIDmode)
7402 gcc_assert (submode == mode);
7403 found_equiv = true;
7404 can_change_mode = false;
7409 if (! found_equiv)
7410 return VOIDmode;
7412 /* Now INSN_COUNT is the number of instructions we found which set
7413 CC_REG to a value equivalent to CC_SRC. The instructions are in
7414 INSNS. The modes used by those instructions are in MODES. */
7416 newreg = NULL_RTX;
7417 for (i = 0; i < insn_count; ++i)
7419 if (modes[i] != mode)
7421 /* We need to change the mode of CC_REG in INSNS[i] and
7422 subsequent instructions. */
7423 if (! newreg)
7425 if (GET_MODE (cc_reg) == mode)
7426 newreg = cc_reg;
7427 else
7428 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7430 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7431 newreg);
7434 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7437 return mode;
7440 /* If we have a fixed condition code register (or two), walk through
7441 the instructions and try to eliminate duplicate assignments. */
7443 static void
7444 cse_condition_code_reg (void)
7446 unsigned int cc_regno_1;
7447 unsigned int cc_regno_2;
7448 rtx cc_reg_1;
7449 rtx cc_reg_2;
7450 basic_block bb;
7452 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7453 return;
7455 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7456 if (cc_regno_2 != INVALID_REGNUM)
7457 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7458 else
7459 cc_reg_2 = NULL_RTX;
7461 FOR_EACH_BB_FN (bb, cfun)
7463 rtx_insn *last_insn;
7464 rtx cc_reg;
7465 rtx_insn *insn;
7466 rtx_insn *cc_src_insn;
7467 rtx cc_src;
7468 machine_mode mode;
7469 machine_mode orig_mode;
7471 /* Look for blocks which end with a conditional jump based on a
7472 condition code register. Then look for the instruction which
7473 sets the condition code register. Then look through the
7474 successor blocks for instructions which set the condition
7475 code register to the same value. There are other possible
7476 uses of the condition code register, but these are by far the
7477 most common and the ones which we are most likely to be able
7478 to optimize. */
7480 last_insn = BB_END (bb);
7481 if (!JUMP_P (last_insn))
7482 continue;
7484 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7485 cc_reg = cc_reg_1;
7486 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7487 cc_reg = cc_reg_2;
7488 else
7489 continue;
7491 cc_src_insn = NULL;
7492 cc_src = NULL_RTX;
7493 for (insn = PREV_INSN (last_insn);
7494 insn && insn != PREV_INSN (BB_HEAD (bb));
7495 insn = PREV_INSN (insn))
7497 rtx set;
7499 if (! INSN_P (insn))
7500 continue;
7501 set = single_set (insn);
7502 if (set
7503 && REG_P (SET_DEST (set))
7504 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7506 cc_src_insn = insn;
7507 cc_src = SET_SRC (set);
7508 break;
7510 else if (reg_set_p (cc_reg, insn))
7511 break;
7514 if (! cc_src_insn)
7515 continue;
7517 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7518 continue;
7520 /* Now CC_REG is a condition code register used for a
7521 conditional jump at the end of the block, and CC_SRC, in
7522 CC_SRC_INSN, is the value to which that condition code
7523 register is set, and CC_SRC is still meaningful at the end of
7524 the basic block. */
7526 orig_mode = GET_MODE (cc_src);
7527 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7528 if (mode != VOIDmode)
7530 gcc_assert (mode == GET_MODE (cc_src));
7531 if (mode != orig_mode)
7533 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7535 cse_change_cc_mode_insn (cc_src_insn, newreg);
7537 /* Do the same in the following insns that use the
7538 current value of CC_REG within BB. */
7539 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7540 NEXT_INSN (last_insn),
7541 newreg);
7548 /* Perform common subexpression elimination. Nonzero value from
7549 `cse_main' means that jumps were simplified and some code may now
7550 be unreachable, so do jump optimization again. */
7551 static unsigned int
7552 rest_of_handle_cse (void)
7554 int tem;
7556 if (dump_file)
7557 dump_flow_info (dump_file, dump_flags);
7559 tem = cse_main (get_insns (), max_reg_num ());
7561 /* If we are not running more CSE passes, then we are no longer
7562 expecting CSE to be run. But always rerun it in a cheap mode. */
7563 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7565 if (tem == 2)
7567 timevar_push (TV_JUMP);
7568 rebuild_jump_labels (get_insns ());
7569 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7570 timevar_pop (TV_JUMP);
7572 else if (tem == 1 || optimize > 1)
7573 cse_cfg_altered |= cleanup_cfg (0);
7575 return 0;
7578 namespace {
7580 const pass_data pass_data_cse =
7582 RTL_PASS, /* type */
7583 "cse1", /* name */
7584 OPTGROUP_NONE, /* optinfo_flags */
7585 TV_CSE, /* tv_id */
7586 0, /* properties_required */
7587 0, /* properties_provided */
7588 0, /* properties_destroyed */
7589 0, /* todo_flags_start */
7590 TODO_df_finish, /* todo_flags_finish */
7593 class pass_cse : public rtl_opt_pass
7595 public:
7596 pass_cse (gcc::context *ctxt)
7597 : rtl_opt_pass (pass_data_cse, ctxt)
7600 /* opt_pass methods: */
7601 virtual bool gate (function *) { return optimize > 0; }
7602 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7604 }; // class pass_cse
7606 } // anon namespace
7608 rtl_opt_pass *
7609 make_pass_cse (gcc::context *ctxt)
7611 return new pass_cse (ctxt);
7615 /* Run second CSE pass after loop optimizations. */
7616 static unsigned int
7617 rest_of_handle_cse2 (void)
7619 int tem;
7621 if (dump_file)
7622 dump_flow_info (dump_file, dump_flags);
7624 tem = cse_main (get_insns (), max_reg_num ());
7626 /* Run a pass to eliminate duplicated assignments to condition code
7627 registers. We have to run this after bypass_jumps, because it
7628 makes it harder for that pass to determine whether a jump can be
7629 bypassed safely. */
7630 cse_condition_code_reg ();
7632 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7634 if (tem == 2)
7636 timevar_push (TV_JUMP);
7637 rebuild_jump_labels (get_insns ());
7638 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7639 timevar_pop (TV_JUMP);
7641 else if (tem == 1)
7642 cse_cfg_altered |= cleanup_cfg (0);
7644 cse_not_expected = 1;
7645 return 0;
7649 namespace {
7651 const pass_data pass_data_cse2 =
7653 RTL_PASS, /* type */
7654 "cse2", /* name */
7655 OPTGROUP_NONE, /* optinfo_flags */
7656 TV_CSE2, /* tv_id */
7657 0, /* properties_required */
7658 0, /* properties_provided */
7659 0, /* properties_destroyed */
7660 0, /* todo_flags_start */
7661 TODO_df_finish, /* todo_flags_finish */
7664 class pass_cse2 : public rtl_opt_pass
7666 public:
7667 pass_cse2 (gcc::context *ctxt)
7668 : rtl_opt_pass (pass_data_cse2, ctxt)
7671 /* opt_pass methods: */
7672 virtual bool gate (function *)
7674 return optimize > 0 && flag_rerun_cse_after_loop;
7677 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7679 }; // class pass_cse2
7681 } // anon namespace
7683 rtl_opt_pass *
7684 make_pass_cse2 (gcc::context *ctxt)
7686 return new pass_cse2 (ctxt);
7689 /* Run second CSE pass after loop optimizations. */
7690 static unsigned int
7691 rest_of_handle_cse_after_global_opts (void)
7693 int save_cfj;
7694 int tem;
7696 /* We only want to do local CSE, so don't follow jumps. */
7697 save_cfj = flag_cse_follow_jumps;
7698 flag_cse_follow_jumps = 0;
7700 rebuild_jump_labels (get_insns ());
7701 tem = cse_main (get_insns (), max_reg_num ());
7702 cse_cfg_altered |= purge_all_dead_edges ();
7703 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7705 cse_not_expected = !flag_rerun_cse_after_loop;
7707 /* If cse altered any jumps, rerun jump opts to clean things up. */
7708 if (tem == 2)
7710 timevar_push (TV_JUMP);
7711 rebuild_jump_labels (get_insns ());
7712 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7713 timevar_pop (TV_JUMP);
7715 else if (tem == 1)
7716 cse_cfg_altered |= cleanup_cfg (0);
7718 flag_cse_follow_jumps = save_cfj;
7719 return 0;
7722 namespace {
7724 const pass_data pass_data_cse_after_global_opts =
7726 RTL_PASS, /* type */
7727 "cse_local", /* name */
7728 OPTGROUP_NONE, /* optinfo_flags */
7729 TV_CSE, /* tv_id */
7730 0, /* properties_required */
7731 0, /* properties_provided */
7732 0, /* properties_destroyed */
7733 0, /* todo_flags_start */
7734 TODO_df_finish, /* todo_flags_finish */
7737 class pass_cse_after_global_opts : public rtl_opt_pass
7739 public:
7740 pass_cse_after_global_opts (gcc::context *ctxt)
7741 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7744 /* opt_pass methods: */
7745 virtual bool gate (function *)
7747 return optimize > 0 && flag_rerun_cse_after_global_opts;
7750 virtual unsigned int execute (function *)
7752 return rest_of_handle_cse_after_global_opts ();
7755 }; // class pass_cse_after_global_opts
7757 } // anon namespace
7759 rtl_opt_pass *
7760 make_pass_cse_after_global_opts (gcc::context *ctxt)
7762 return new pass_cse_after_global_opts (ctxt);