Daily bump.
[official-gcc.git] / gcc / cse.c
blob72f1c4f6a9e378a71429b335c7a2e5b112ac76c0
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2016 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "insn-config.h"
32 #include "regs.h"
33 #include "emit-rtl.h"
34 #include "recog.h"
35 #include "cfgrtl.h"
36 #include "cfganal.h"
37 #include "cfgcleanup.h"
38 #include "alias.h"
39 #include "toplev.h"
40 #include "params.h"
41 #include "rtlhooks-def.h"
42 #include "tree-pass.h"
43 #include "dbgcnt.h"
44 #include "rtl-iter.h"
46 /* The basic idea of common subexpression elimination is to go
47 through the code, keeping a record of expressions that would
48 have the same value at the current scan point, and replacing
49 expressions encountered with the cheapest equivalent expression.
51 It is too complicated to keep track of the different possibilities
52 when control paths merge in this code; so, at each label, we forget all
53 that is known and start fresh. This can be described as processing each
54 extended basic block separately. We have a separate pass to perform
55 global CSE.
57 Note CSE can turn a conditional or computed jump into a nop or
58 an unconditional jump. When this occurs we arrange to run the jump
59 optimizer after CSE to delete the unreachable code.
61 We use two data structures to record the equivalent expressions:
62 a hash table for most expressions, and a vector of "quantity
63 numbers" to record equivalent (pseudo) registers.
65 The use of the special data structure for registers is desirable
66 because it is faster. It is possible because registers references
67 contain a fairly small number, the register number, taken from
68 a contiguously allocated series, and two register references are
69 identical if they have the same number. General expressions
70 do not have any such thing, so the only way to retrieve the
71 information recorded on an expression other than a register
72 is to keep it in a hash table.
74 Registers and "quantity numbers":
76 At the start of each basic block, all of the (hardware and pseudo)
77 registers used in the function are given distinct quantity
78 numbers to indicate their contents. During scan, when the code
79 copies one register into another, we copy the quantity number.
80 When a register is loaded in any other way, we allocate a new
81 quantity number to describe the value generated by this operation.
82 `REG_QTY (N)' records what quantity register N is currently thought
83 of as containing.
85 All real quantity numbers are greater than or equal to zero.
86 If register N has not been assigned a quantity, `REG_QTY (N)' will
87 equal -N - 1, which is always negative.
89 Quantity numbers below zero do not exist and none of the `qty_table'
90 entries should be referenced with a negative index.
92 We also maintain a bidirectional chain of registers for each
93 quantity number. The `qty_table` members `first_reg' and `last_reg',
94 and `reg_eqv_table' members `next' and `prev' hold these chains.
96 The first register in a chain is the one whose lifespan is least local.
97 Among equals, it is the one that was seen first.
98 We replace any equivalent register with that one.
100 If two registers have the same quantity number, it must be true that
101 REG expressions with qty_table `mode' must be in the hash table for both
102 registers and must be in the same class.
104 The converse is not true. Since hard registers may be referenced in
105 any mode, two REG expressions might be equivalent in the hash table
106 but not have the same quantity number if the quantity number of one
107 of the registers is not the same mode as those expressions.
109 Constants and quantity numbers
111 When a quantity has a known constant value, that value is stored
112 in the appropriate qty_table `const_rtx'. This is in addition to
113 putting the constant in the hash table as is usual for non-regs.
115 Whether a reg or a constant is preferred is determined by the configuration
116 macro CONST_COSTS and will often depend on the constant value. In any
117 event, expressions containing constants can be simplified, by fold_rtx.
119 When a quantity has a known nearly constant value (such as an address
120 of a stack slot), that value is stored in the appropriate qty_table
121 `const_rtx'.
123 Integer constants don't have a machine mode. However, cse
124 determines the intended machine mode from the destination
125 of the instruction that moves the constant. The machine mode
126 is recorded in the hash table along with the actual RTL
127 constant expression so that different modes are kept separate.
129 Other expressions:
131 To record known equivalences among expressions in general
132 we use a hash table called `table'. It has a fixed number of buckets
133 that contain chains of `struct table_elt' elements for expressions.
134 These chains connect the elements whose expressions have the same
135 hash codes.
137 Other chains through the same elements connect the elements which
138 currently have equivalent values.
140 Register references in an expression are canonicalized before hashing
141 the expression. This is done using `reg_qty' and qty_table `first_reg'.
142 The hash code of a register reference is computed using the quantity
143 number, not the register number.
145 When the value of an expression changes, it is necessary to remove from the
146 hash table not just that expression but all expressions whose values
147 could be different as a result.
149 1. If the value changing is in memory, except in special cases
150 ANYTHING referring to memory could be changed. That is because
151 nobody knows where a pointer does not point.
152 The function `invalidate_memory' removes what is necessary.
154 The special cases are when the address is constant or is
155 a constant plus a fixed register such as the frame pointer
156 or a static chain pointer. When such addresses are stored in,
157 we can tell exactly which other such addresses must be invalidated
158 due to overlap. `invalidate' does this.
159 All expressions that refer to non-constant
160 memory addresses are also invalidated. `invalidate_memory' does this.
162 2. If the value changing is a register, all expressions
163 containing references to that register, and only those,
164 must be removed.
166 Because searching the entire hash table for expressions that contain
167 a register is very slow, we try to figure out when it isn't necessary.
168 Precisely, this is necessary only when expressions have been
169 entered in the hash table using this register, and then the value has
170 changed, and then another expression wants to be added to refer to
171 the register's new value. This sequence of circumstances is rare
172 within any one basic block.
174 `REG_TICK' and `REG_IN_TABLE', accessors for members of
175 cse_reg_info, are used to detect this case. REG_TICK (i) is
176 incremented whenever a value is stored in register i.
177 REG_IN_TABLE (i) holds -1 if no references to register i have been
178 entered in the table; otherwise, it contains the value REG_TICK (i)
179 had when the references were entered. If we want to enter a
180 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
181 remove old references. Until we want to enter a new entry, the
182 mere fact that the two vectors don't match makes the entries be
183 ignored if anyone tries to match them.
185 Registers themselves are entered in the hash table as well as in
186 the equivalent-register chains. However, `REG_TICK' and
187 `REG_IN_TABLE' do not apply to expressions which are simple
188 register references. These expressions are removed from the table
189 immediately when they become invalid, and this can be done even if
190 we do not immediately search for all the expressions that refer to
191 the register.
193 A CLOBBER rtx in an instruction invalidates its operand for further
194 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
195 invalidates everything that resides in memory.
197 Related expressions:
199 Constant expressions that differ only by an additive integer
200 are called related. When a constant expression is put in
201 the table, the related expression with no constant term
202 is also entered. These are made to point at each other
203 so that it is possible to find out if there exists any
204 register equivalent to an expression related to a given expression. */
206 /* Length of qty_table vector. We know in advance we will not need
207 a quantity number this big. */
209 static int max_qty;
211 /* Next quantity number to be allocated.
212 This is 1 + the largest number needed so far. */
214 static int next_qty;
216 /* Per-qty information tracking.
218 `first_reg' and `last_reg' track the head and tail of the
219 chain of registers which currently contain this quantity.
221 `mode' contains the machine mode of this quantity.
223 `const_rtx' holds the rtx of the constant value of this
224 quantity, if known. A summations of the frame/arg pointer
225 and a constant can also be entered here. When this holds
226 a known value, `const_insn' is the insn which stored the
227 constant value.
229 `comparison_{code,const,qty}' are used to track when a
230 comparison between a quantity and some constant or register has
231 been passed. In such a case, we know the results of the comparison
232 in case we see it again. These members record a comparison that
233 is known to be true. `comparison_code' holds the rtx code of such
234 a comparison, else it is set to UNKNOWN and the other two
235 comparison members are undefined. `comparison_const' holds
236 the constant being compared against, or zero if the comparison
237 is not against a constant. `comparison_qty' holds the quantity
238 being compared against when the result is known. If the comparison
239 is not with a register, `comparison_qty' is -1. */
241 struct qty_table_elem
243 rtx const_rtx;
244 rtx_insn *const_insn;
245 rtx comparison_const;
246 int comparison_qty;
247 unsigned int first_reg, last_reg;
248 /* The sizes of these fields should match the sizes of the
249 code and mode fields of struct rtx_def (see rtl.h). */
250 ENUM_BITFIELD(rtx_code) comparison_code : 16;
251 ENUM_BITFIELD(machine_mode) mode : 8;
254 /* The table of all qtys, indexed by qty number. */
255 static struct qty_table_elem *qty_table;
257 /* For machines that have a CC0, we do not record its value in the hash
258 table since its use is guaranteed to be the insn immediately following
259 its definition and any other insn is presumed to invalidate it.
261 Instead, we store below the current and last value assigned to CC0.
262 If it should happen to be a constant, it is stored in preference
263 to the actual assigned value. In case it is a constant, we store
264 the mode in which the constant should be interpreted. */
266 static rtx this_insn_cc0, prev_insn_cc0;
267 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
269 /* Insn being scanned. */
271 static rtx_insn *this_insn;
272 static bool optimize_this_for_speed_p;
274 /* Index by register number, gives the number of the next (or
275 previous) register in the chain of registers sharing the same
276 value.
278 Or -1 if this register is at the end of the chain.
280 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
282 /* Per-register equivalence chain. */
283 struct reg_eqv_elem
285 int next, prev;
288 /* The table of all register equivalence chains. */
289 static struct reg_eqv_elem *reg_eqv_table;
291 struct cse_reg_info
293 /* The timestamp at which this register is initialized. */
294 unsigned int timestamp;
296 /* The quantity number of the register's current contents. */
297 int reg_qty;
299 /* The number of times the register has been altered in the current
300 basic block. */
301 int reg_tick;
303 /* The REG_TICK value at which rtx's containing this register are
304 valid in the hash table. If this does not equal the current
305 reg_tick value, such expressions existing in the hash table are
306 invalid. */
307 int reg_in_table;
309 /* The SUBREG that was set when REG_TICK was last incremented. Set
310 to -1 if the last store was to the whole register, not a subreg. */
311 unsigned int subreg_ticked;
314 /* A table of cse_reg_info indexed by register numbers. */
315 static struct cse_reg_info *cse_reg_info_table;
317 /* The size of the above table. */
318 static unsigned int cse_reg_info_table_size;
320 /* The index of the first entry that has not been initialized. */
321 static unsigned int cse_reg_info_table_first_uninitialized;
323 /* The timestamp at the beginning of the current run of
324 cse_extended_basic_block. We increment this variable at the beginning of
325 the current run of cse_extended_basic_block. The timestamp field of a
326 cse_reg_info entry matches the value of this variable if and only
327 if the entry has been initialized during the current run of
328 cse_extended_basic_block. */
329 static unsigned int cse_reg_info_timestamp;
331 /* A HARD_REG_SET containing all the hard registers for which there is
332 currently a REG expression in the hash table. Note the difference
333 from the above variables, which indicate if the REG is mentioned in some
334 expression in the table. */
336 static HARD_REG_SET hard_regs_in_table;
338 /* True if CSE has altered the CFG. */
339 static bool cse_cfg_altered;
341 /* True if CSE has altered conditional jump insns in such a way
342 that jump optimization should be redone. */
343 static bool cse_jumps_altered;
345 /* True if we put a LABEL_REF into the hash table for an INSN
346 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
347 to put in the note. */
348 static bool recorded_label_ref;
350 /* canon_hash stores 1 in do_not_record
351 if it notices a reference to CC0, PC, or some other volatile
352 subexpression. */
354 static int do_not_record;
356 /* canon_hash stores 1 in hash_arg_in_memory
357 if it notices a reference to memory within the expression being hashed. */
359 static int hash_arg_in_memory;
361 /* The hash table contains buckets which are chains of `struct table_elt's,
362 each recording one expression's information.
363 That expression is in the `exp' field.
365 The canon_exp field contains a canonical (from the point of view of
366 alias analysis) version of the `exp' field.
368 Those elements with the same hash code are chained in both directions
369 through the `next_same_hash' and `prev_same_hash' fields.
371 Each set of expressions with equivalent values
372 are on a two-way chain through the `next_same_value'
373 and `prev_same_value' fields, and all point with
374 the `first_same_value' field at the first element in
375 that chain. The chain is in order of increasing cost.
376 Each element's cost value is in its `cost' field.
378 The `in_memory' field is nonzero for elements that
379 involve any reference to memory. These elements are removed
380 whenever a write is done to an unidentified location in memory.
381 To be safe, we assume that a memory address is unidentified unless
382 the address is either a symbol constant or a constant plus
383 the frame pointer or argument pointer.
385 The `related_value' field is used to connect related expressions
386 (that differ by adding an integer).
387 The related expressions are chained in a circular fashion.
388 `related_value' is zero for expressions for which this
389 chain is not useful.
391 The `cost' field stores the cost of this element's expression.
392 The `regcost' field stores the value returned by approx_reg_cost for
393 this element's expression.
395 The `is_const' flag is set if the element is a constant (including
396 a fixed address).
398 The `flag' field is used as a temporary during some search routines.
400 The `mode' field is usually the same as GET_MODE (`exp'), but
401 if `exp' is a CONST_INT and has no machine mode then the `mode'
402 field is the mode it was being used as. Each constant is
403 recorded separately for each mode it is used with. */
405 struct table_elt
407 rtx exp;
408 rtx canon_exp;
409 struct table_elt *next_same_hash;
410 struct table_elt *prev_same_hash;
411 struct table_elt *next_same_value;
412 struct table_elt *prev_same_value;
413 struct table_elt *first_same_value;
414 struct table_elt *related_value;
415 int cost;
416 int regcost;
417 /* The size of this field should match the size
418 of the mode field of struct rtx_def (see rtl.h). */
419 ENUM_BITFIELD(machine_mode) mode : 8;
420 char in_memory;
421 char is_const;
422 char flag;
425 /* We don't want a lot of buckets, because we rarely have very many
426 things stored in the hash table, and a lot of buckets slows
427 down a lot of loops that happen frequently. */
428 #define HASH_SHIFT 5
429 #define HASH_SIZE (1 << HASH_SHIFT)
430 #define HASH_MASK (HASH_SIZE - 1)
432 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
433 register (hard registers may require `do_not_record' to be set). */
435 #define HASH(X, M) \
436 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
437 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
438 : canon_hash (X, M)) & HASH_MASK)
440 /* Like HASH, but without side-effects. */
441 #define SAFE_HASH(X, M) \
442 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
443 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
444 : safe_hash (X, M)) & HASH_MASK)
446 /* Determine whether register number N is considered a fixed register for the
447 purpose of approximating register costs.
448 It is desirable to replace other regs with fixed regs, to reduce need for
449 non-fixed hard regs.
450 A reg wins if it is either the frame pointer or designated as fixed. */
451 #define FIXED_REGNO_P(N) \
452 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
453 || fixed_regs[N] || global_regs[N])
455 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
456 hard registers and pointers into the frame are the cheapest with a cost
457 of 0. Next come pseudos with a cost of one and other hard registers with
458 a cost of 2. Aside from these special cases, call `rtx_cost'. */
460 #define CHEAP_REGNO(N) \
461 (REGNO_PTR_FRAME_P (N) \
462 || (HARD_REGISTER_NUM_P (N) \
463 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
465 #define COST(X, MODE) \
466 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
467 #define COST_IN(X, MODE, OUTER, OPNO) \
468 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
470 /* Get the number of times this register has been updated in this
471 basic block. */
473 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
475 /* Get the point at which REG was recorded in the table. */
477 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
479 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
480 SUBREG). */
482 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
484 /* Get the quantity number for REG. */
486 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
488 /* Determine if the quantity number for register X represents a valid index
489 into the qty_table. */
491 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
493 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
495 #define CHEAPER(X, Y) \
496 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
498 static struct table_elt *table[HASH_SIZE];
500 /* Chain of `struct table_elt's made so far for this function
501 but currently removed from the table. */
503 static struct table_elt *free_element_chain;
505 /* Set to the cost of a constant pool reference if one was found for a
506 symbolic constant. If this was found, it means we should try to
507 convert constants into constant pool entries if they don't fit in
508 the insn. */
510 static int constant_pool_entries_cost;
511 static int constant_pool_entries_regcost;
513 /* Trace a patch through the CFG. */
515 struct branch_path
517 /* The basic block for this path entry. */
518 basic_block bb;
521 /* This data describes a block that will be processed by
522 cse_extended_basic_block. */
524 struct cse_basic_block_data
526 /* Total number of SETs in block. */
527 int nsets;
528 /* Size of current branch path, if any. */
529 int path_size;
530 /* Current path, indicating which basic_blocks will be processed. */
531 struct branch_path *path;
535 /* Pointers to the live in/live out bitmaps for the boundaries of the
536 current EBB. */
537 static bitmap cse_ebb_live_in, cse_ebb_live_out;
539 /* A simple bitmap to track which basic blocks have been visited
540 already as part of an already processed extended basic block. */
541 static sbitmap cse_visited_basic_blocks;
543 static bool fixed_base_plus_p (rtx x);
544 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
545 static int preferable (int, int, int, int);
546 static void new_basic_block (void);
547 static void make_new_qty (unsigned int, machine_mode);
548 static void make_regs_eqv (unsigned int, unsigned int);
549 static void delete_reg_equiv (unsigned int);
550 static int mention_regs (rtx);
551 static int insert_regs (rtx, struct table_elt *, int);
552 static void remove_from_table (struct table_elt *, unsigned);
553 static void remove_pseudo_from_table (rtx, unsigned);
554 static struct table_elt *lookup (rtx, unsigned, machine_mode);
555 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
556 static rtx lookup_as_function (rtx, enum rtx_code);
557 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
558 machine_mode, int, int);
559 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
560 machine_mode);
561 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
562 static void invalidate (rtx, machine_mode);
563 static void remove_invalid_refs (unsigned int);
564 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
565 machine_mode);
566 static void rehash_using_reg (rtx);
567 static void invalidate_memory (void);
568 static void invalidate_for_call (void);
569 static rtx use_related_value (rtx, struct table_elt *);
571 static inline unsigned canon_hash (rtx, machine_mode);
572 static inline unsigned safe_hash (rtx, machine_mode);
573 static inline unsigned hash_rtx_string (const char *);
575 static rtx canon_reg (rtx, rtx_insn *);
576 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
577 machine_mode *,
578 machine_mode *);
579 static rtx fold_rtx (rtx, rtx_insn *);
580 static rtx equiv_constant (rtx);
581 static void record_jump_equiv (rtx_insn *, bool);
582 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
583 int);
584 static void cse_insn (rtx_insn *);
585 static void cse_prescan_path (struct cse_basic_block_data *);
586 static void invalidate_from_clobbers (rtx_insn *);
587 static void invalidate_from_sets_and_clobbers (rtx_insn *);
588 static rtx cse_process_notes (rtx, rtx, bool *);
589 static void cse_extended_basic_block (struct cse_basic_block_data *);
590 extern void dump_class (struct table_elt*);
591 static void get_cse_reg_info_1 (unsigned int regno);
592 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
594 static void flush_hash_table (void);
595 static bool insn_live_p (rtx_insn *, int *);
596 static bool set_live_p (rtx, rtx_insn *, int *);
597 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
598 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
599 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
600 bool);
603 #undef RTL_HOOKS_GEN_LOWPART
604 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
606 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
608 /* Nonzero if X has the form (PLUS frame-pointer integer). */
610 static bool
611 fixed_base_plus_p (rtx x)
613 switch (GET_CODE (x))
615 case REG:
616 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
617 return true;
618 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
619 return true;
620 return false;
622 case PLUS:
623 if (!CONST_INT_P (XEXP (x, 1)))
624 return false;
625 return fixed_base_plus_p (XEXP (x, 0));
627 default:
628 return false;
632 /* Dump the expressions in the equivalence class indicated by CLASSP.
633 This function is used only for debugging. */
634 DEBUG_FUNCTION void
635 dump_class (struct table_elt *classp)
637 struct table_elt *elt;
639 fprintf (stderr, "Equivalence chain for ");
640 print_rtl (stderr, classp->exp);
641 fprintf (stderr, ": \n");
643 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
645 print_rtl (stderr, elt->exp);
646 fprintf (stderr, "\n");
650 /* Return an estimate of the cost of the registers used in an rtx.
651 This is mostly the number of different REG expressions in the rtx;
652 however for some exceptions like fixed registers we use a cost of
653 0. If any other hard register reference occurs, return MAX_COST. */
655 static int
656 approx_reg_cost (const_rtx x)
658 int cost = 0;
659 subrtx_iterator::array_type array;
660 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
662 const_rtx x = *iter;
663 if (REG_P (x))
665 unsigned int regno = REGNO (x);
666 if (!CHEAP_REGNO (regno))
668 if (regno < FIRST_PSEUDO_REGISTER)
670 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
671 return MAX_COST;
672 cost += 2;
674 else
675 cost += 1;
679 return cost;
682 /* Return a negative value if an rtx A, whose costs are given by COST_A
683 and REGCOST_A, is more desirable than an rtx B.
684 Return a positive value if A is less desirable, or 0 if the two are
685 equally good. */
686 static int
687 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
689 /* First, get rid of cases involving expressions that are entirely
690 unwanted. */
691 if (cost_a != cost_b)
693 if (cost_a == MAX_COST)
694 return 1;
695 if (cost_b == MAX_COST)
696 return -1;
699 /* Avoid extending lifetimes of hardregs. */
700 if (regcost_a != regcost_b)
702 if (regcost_a == MAX_COST)
703 return 1;
704 if (regcost_b == MAX_COST)
705 return -1;
708 /* Normal operation costs take precedence. */
709 if (cost_a != cost_b)
710 return cost_a - cost_b;
711 /* Only if these are identical consider effects on register pressure. */
712 if (regcost_a != regcost_b)
713 return regcost_a - regcost_b;
714 return 0;
717 /* Internal function, to compute cost when X is not a register; called
718 from COST macro to keep it simple. */
720 static int
721 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
723 return ((GET_CODE (x) == SUBREG
724 && REG_P (SUBREG_REG (x))
725 && GET_MODE_CLASS (mode) == MODE_INT
726 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
727 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
728 && subreg_lowpart_p (x)
729 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x))))
731 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
735 /* Initialize CSE_REG_INFO_TABLE. */
737 static void
738 init_cse_reg_info (unsigned int nregs)
740 /* Do we need to grow the table? */
741 if (nregs > cse_reg_info_table_size)
743 unsigned int new_size;
745 if (cse_reg_info_table_size < 2048)
747 /* Compute a new size that is a power of 2 and no smaller
748 than the large of NREGS and 64. */
749 new_size = (cse_reg_info_table_size
750 ? cse_reg_info_table_size : 64);
752 while (new_size < nregs)
753 new_size *= 2;
755 else
757 /* If we need a big table, allocate just enough to hold
758 NREGS registers. */
759 new_size = nregs;
762 /* Reallocate the table with NEW_SIZE entries. */
763 free (cse_reg_info_table);
764 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
765 cse_reg_info_table_size = new_size;
766 cse_reg_info_table_first_uninitialized = 0;
769 /* Do we have all of the first NREGS entries initialized? */
770 if (cse_reg_info_table_first_uninitialized < nregs)
772 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
773 unsigned int i;
775 /* Put the old timestamp on newly allocated entries so that they
776 will all be considered out of date. We do not touch those
777 entries beyond the first NREGS entries to be nice to the
778 virtual memory. */
779 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
780 cse_reg_info_table[i].timestamp = old_timestamp;
782 cse_reg_info_table_first_uninitialized = nregs;
786 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
788 static void
789 get_cse_reg_info_1 (unsigned int regno)
791 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
792 entry will be considered to have been initialized. */
793 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
795 /* Initialize the rest of the entry. */
796 cse_reg_info_table[regno].reg_tick = 1;
797 cse_reg_info_table[regno].reg_in_table = -1;
798 cse_reg_info_table[regno].subreg_ticked = -1;
799 cse_reg_info_table[regno].reg_qty = -regno - 1;
802 /* Find a cse_reg_info entry for REGNO. */
804 static inline struct cse_reg_info *
805 get_cse_reg_info (unsigned int regno)
807 struct cse_reg_info *p = &cse_reg_info_table[regno];
809 /* If this entry has not been initialized, go ahead and initialize
810 it. */
811 if (p->timestamp != cse_reg_info_timestamp)
812 get_cse_reg_info_1 (regno);
814 return p;
817 /* Clear the hash table and initialize each register with its own quantity,
818 for a new basic block. */
820 static void
821 new_basic_block (void)
823 int i;
825 next_qty = 0;
827 /* Invalidate cse_reg_info_table. */
828 cse_reg_info_timestamp++;
830 /* Clear out hash table state for this pass. */
831 CLEAR_HARD_REG_SET (hard_regs_in_table);
833 /* The per-quantity values used to be initialized here, but it is
834 much faster to initialize each as it is made in `make_new_qty'. */
836 for (i = 0; i < HASH_SIZE; i++)
838 struct table_elt *first;
840 first = table[i];
841 if (first != NULL)
843 struct table_elt *last = first;
845 table[i] = NULL;
847 while (last->next_same_hash != NULL)
848 last = last->next_same_hash;
850 /* Now relink this hash entire chain into
851 the free element list. */
853 last->next_same_hash = free_element_chain;
854 free_element_chain = first;
858 prev_insn_cc0 = 0;
861 /* Say that register REG contains a quantity in mode MODE not in any
862 register before and initialize that quantity. */
864 static void
865 make_new_qty (unsigned int reg, machine_mode mode)
867 int q;
868 struct qty_table_elem *ent;
869 struct reg_eqv_elem *eqv;
871 gcc_assert (next_qty < max_qty);
873 q = REG_QTY (reg) = next_qty++;
874 ent = &qty_table[q];
875 ent->first_reg = reg;
876 ent->last_reg = reg;
877 ent->mode = mode;
878 ent->const_rtx = ent->const_insn = NULL;
879 ent->comparison_code = UNKNOWN;
881 eqv = &reg_eqv_table[reg];
882 eqv->next = eqv->prev = -1;
885 /* Make reg NEW equivalent to reg OLD.
886 OLD is not changing; NEW is. */
888 static void
889 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
891 unsigned int lastr, firstr;
892 int q = REG_QTY (old_reg);
893 struct qty_table_elem *ent;
895 ent = &qty_table[q];
897 /* Nothing should become eqv until it has a "non-invalid" qty number. */
898 gcc_assert (REGNO_QTY_VALID_P (old_reg));
900 REG_QTY (new_reg) = q;
901 firstr = ent->first_reg;
902 lastr = ent->last_reg;
904 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
905 hard regs. Among pseudos, if NEW will live longer than any other reg
906 of the same qty, and that is beyond the current basic block,
907 make it the new canonical replacement for this qty. */
908 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
909 /* Certain fixed registers might be of the class NO_REGS. This means
910 that not only can they not be allocated by the compiler, but
911 they cannot be used in substitutions or canonicalizations
912 either. */
913 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
914 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
915 || (new_reg >= FIRST_PSEUDO_REGISTER
916 && (firstr < FIRST_PSEUDO_REGISTER
917 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
918 && !bitmap_bit_p (cse_ebb_live_out, firstr))
919 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
920 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
922 reg_eqv_table[firstr].prev = new_reg;
923 reg_eqv_table[new_reg].next = firstr;
924 reg_eqv_table[new_reg].prev = -1;
925 ent->first_reg = new_reg;
927 else
929 /* If NEW is a hard reg (known to be non-fixed), insert at end.
930 Otherwise, insert before any non-fixed hard regs that are at the
931 end. Registers of class NO_REGS cannot be used as an
932 equivalent for anything. */
933 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
934 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
935 && new_reg >= FIRST_PSEUDO_REGISTER)
936 lastr = reg_eqv_table[lastr].prev;
937 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
938 if (reg_eqv_table[lastr].next >= 0)
939 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
940 else
941 qty_table[q].last_reg = new_reg;
942 reg_eqv_table[lastr].next = new_reg;
943 reg_eqv_table[new_reg].prev = lastr;
947 /* Remove REG from its equivalence class. */
949 static void
950 delete_reg_equiv (unsigned int reg)
952 struct qty_table_elem *ent;
953 int q = REG_QTY (reg);
954 int p, n;
956 /* If invalid, do nothing. */
957 if (! REGNO_QTY_VALID_P (reg))
958 return;
960 ent = &qty_table[q];
962 p = reg_eqv_table[reg].prev;
963 n = reg_eqv_table[reg].next;
965 if (n != -1)
966 reg_eqv_table[n].prev = p;
967 else
968 ent->last_reg = p;
969 if (p != -1)
970 reg_eqv_table[p].next = n;
971 else
972 ent->first_reg = n;
974 REG_QTY (reg) = -reg - 1;
977 /* Remove any invalid expressions from the hash table
978 that refer to any of the registers contained in expression X.
980 Make sure that newly inserted references to those registers
981 as subexpressions will be considered valid.
983 mention_regs is not called when a register itself
984 is being stored in the table.
986 Return 1 if we have done something that may have changed the hash code
987 of X. */
989 static int
990 mention_regs (rtx x)
992 enum rtx_code code;
993 int i, j;
994 const char *fmt;
995 int changed = 0;
997 if (x == 0)
998 return 0;
1000 code = GET_CODE (x);
1001 if (code == REG)
1003 unsigned int regno = REGNO (x);
1004 unsigned int endregno = END_REGNO (x);
1005 unsigned int i;
1007 for (i = regno; i < endregno; i++)
1009 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1010 remove_invalid_refs (i);
1012 REG_IN_TABLE (i) = REG_TICK (i);
1013 SUBREG_TICKED (i) = -1;
1016 return 0;
1019 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1020 pseudo if they don't use overlapping words. We handle only pseudos
1021 here for simplicity. */
1022 if (code == SUBREG && REG_P (SUBREG_REG (x))
1023 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1025 unsigned int i = REGNO (SUBREG_REG (x));
1027 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1029 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1030 the last store to this register really stored into this
1031 subreg, then remove the memory of this subreg.
1032 Otherwise, remove any memory of the entire register and
1033 all its subregs from the table. */
1034 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1035 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1036 remove_invalid_refs (i);
1037 else
1038 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1041 REG_IN_TABLE (i) = REG_TICK (i);
1042 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1043 return 0;
1046 /* If X is a comparison or a COMPARE and either operand is a register
1047 that does not have a quantity, give it one. This is so that a later
1048 call to record_jump_equiv won't cause X to be assigned a different
1049 hash code and not found in the table after that call.
1051 It is not necessary to do this here, since rehash_using_reg can
1052 fix up the table later, but doing this here eliminates the need to
1053 call that expensive function in the most common case where the only
1054 use of the register is in the comparison. */
1056 if (code == COMPARE || COMPARISON_P (x))
1058 if (REG_P (XEXP (x, 0))
1059 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1060 if (insert_regs (XEXP (x, 0), NULL, 0))
1062 rehash_using_reg (XEXP (x, 0));
1063 changed = 1;
1066 if (REG_P (XEXP (x, 1))
1067 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1068 if (insert_regs (XEXP (x, 1), NULL, 0))
1070 rehash_using_reg (XEXP (x, 1));
1071 changed = 1;
1075 fmt = GET_RTX_FORMAT (code);
1076 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1077 if (fmt[i] == 'e')
1078 changed |= mention_regs (XEXP (x, i));
1079 else if (fmt[i] == 'E')
1080 for (j = 0; j < XVECLEN (x, i); j++)
1081 changed |= mention_regs (XVECEXP (x, i, j));
1083 return changed;
1086 /* Update the register quantities for inserting X into the hash table
1087 with a value equivalent to CLASSP.
1088 (If the class does not contain a REG, it is irrelevant.)
1089 If MODIFIED is nonzero, X is a destination; it is being modified.
1090 Note that delete_reg_equiv should be called on a register
1091 before insert_regs is done on that register with MODIFIED != 0.
1093 Nonzero value means that elements of reg_qty have changed
1094 so X's hash code may be different. */
1096 static int
1097 insert_regs (rtx x, struct table_elt *classp, int modified)
1099 if (REG_P (x))
1101 unsigned int regno = REGNO (x);
1102 int qty_valid;
1104 /* If REGNO is in the equivalence table already but is of the
1105 wrong mode for that equivalence, don't do anything here. */
1107 qty_valid = REGNO_QTY_VALID_P (regno);
1108 if (qty_valid)
1110 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1112 if (ent->mode != GET_MODE (x))
1113 return 0;
1116 if (modified || ! qty_valid)
1118 if (classp)
1119 for (classp = classp->first_same_value;
1120 classp != 0;
1121 classp = classp->next_same_value)
1122 if (REG_P (classp->exp)
1123 && GET_MODE (classp->exp) == GET_MODE (x))
1125 unsigned c_regno = REGNO (classp->exp);
1127 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1129 /* Suppose that 5 is hard reg and 100 and 101 are
1130 pseudos. Consider
1132 (set (reg:si 100) (reg:si 5))
1133 (set (reg:si 5) (reg:si 100))
1134 (set (reg:di 101) (reg:di 5))
1136 We would now set REG_QTY (101) = REG_QTY (5), but the
1137 entry for 5 is in SImode. When we use this later in
1138 copy propagation, we get the register in wrong mode. */
1139 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1140 continue;
1142 make_regs_eqv (regno, c_regno);
1143 return 1;
1146 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1147 than REG_IN_TABLE to find out if there was only a single preceding
1148 invalidation - for the SUBREG - or another one, which would be
1149 for the full register. However, if we find here that REG_TICK
1150 indicates that the register is invalid, it means that it has
1151 been invalidated in a separate operation. The SUBREG might be used
1152 now (then this is a recursive call), or we might use the full REG
1153 now and a SUBREG of it later. So bump up REG_TICK so that
1154 mention_regs will do the right thing. */
1155 if (! modified
1156 && REG_IN_TABLE (regno) >= 0
1157 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1158 REG_TICK (regno)++;
1159 make_new_qty (regno, GET_MODE (x));
1160 return 1;
1163 return 0;
1166 /* If X is a SUBREG, we will likely be inserting the inner register in the
1167 table. If that register doesn't have an assigned quantity number at
1168 this point but does later, the insertion that we will be doing now will
1169 not be accessible because its hash code will have changed. So assign
1170 a quantity number now. */
1172 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1173 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1175 insert_regs (SUBREG_REG (x), NULL, 0);
1176 mention_regs (x);
1177 return 1;
1179 else
1180 return mention_regs (x);
1184 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1185 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1186 CST is equal to an anchor. */
1188 static bool
1189 compute_const_anchors (rtx cst,
1190 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1191 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1193 HOST_WIDE_INT n = INTVAL (cst);
1195 *lower_base = n & ~(targetm.const_anchor - 1);
1196 if (*lower_base == n)
1197 return false;
1199 *upper_base =
1200 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1201 *upper_offs = n - *upper_base;
1202 *lower_offs = n - *lower_base;
1203 return true;
1206 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1208 static void
1209 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1210 machine_mode mode)
1212 struct table_elt *elt;
1213 unsigned hash;
1214 rtx anchor_exp;
1215 rtx exp;
1217 anchor_exp = GEN_INT (anchor);
1218 hash = HASH (anchor_exp, mode);
1219 elt = lookup (anchor_exp, hash, mode);
1220 if (!elt)
1221 elt = insert (anchor_exp, NULL, hash, mode);
1223 exp = plus_constant (mode, reg, offs);
1224 /* REG has just been inserted and the hash codes recomputed. */
1225 mention_regs (exp);
1226 hash = HASH (exp, mode);
1228 /* Use the cost of the register rather than the whole expression. When
1229 looking up constant anchors we will further offset the corresponding
1230 expression therefore it does not make sense to prefer REGs over
1231 reg-immediate additions. Prefer instead the oldest expression. Also
1232 don't prefer pseudos over hard regs so that we derive constants in
1233 argument registers from other argument registers rather than from the
1234 original pseudo that was used to synthesize the constant. */
1235 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1238 /* The constant CST is equivalent to the register REG. Create
1239 equivalences between the two anchors of CST and the corresponding
1240 register-offset expressions using REG. */
1242 static void
1243 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1245 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1247 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1248 &upper_base, &upper_offs))
1249 return;
1251 /* Ignore anchors of value 0. Constants accessible from zero are
1252 simple. */
1253 if (lower_base != 0)
1254 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1256 if (upper_base != 0)
1257 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1260 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1261 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1262 valid expression. Return the cheapest and oldest of such expressions. In
1263 *OLD, return how old the resulting expression is compared to the other
1264 equivalent expressions. */
1266 static rtx
1267 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1268 unsigned *old)
1270 struct table_elt *elt;
1271 unsigned idx;
1272 struct table_elt *match_elt;
1273 rtx match;
1275 /* Find the cheapest and *oldest* expression to maximize the chance of
1276 reusing the same pseudo. */
1278 match_elt = NULL;
1279 match = NULL_RTX;
1280 for (elt = anchor_elt->first_same_value, idx = 0;
1281 elt;
1282 elt = elt->next_same_value, idx++)
1284 if (match_elt && CHEAPER (match_elt, elt))
1285 return match;
1287 if (REG_P (elt->exp)
1288 || (GET_CODE (elt->exp) == PLUS
1289 && REG_P (XEXP (elt->exp, 0))
1290 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1292 rtx x;
1294 /* Ignore expressions that are no longer valid. */
1295 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1296 continue;
1298 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1299 if (REG_P (x)
1300 || (GET_CODE (x) == PLUS
1301 && IN_RANGE (INTVAL (XEXP (x, 1)),
1302 -targetm.const_anchor,
1303 targetm.const_anchor - 1)))
1305 match = x;
1306 match_elt = elt;
1307 *old = idx;
1312 return match;
1315 /* Try to express the constant SRC_CONST using a register+offset expression
1316 derived from a constant anchor. Return it if successful or NULL_RTX,
1317 otherwise. */
1319 static rtx
1320 try_const_anchors (rtx src_const, machine_mode mode)
1322 struct table_elt *lower_elt, *upper_elt;
1323 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1324 rtx lower_anchor_rtx, upper_anchor_rtx;
1325 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1326 unsigned lower_old, upper_old;
1328 /* CONST_INT is used for CC modes, but we should leave those alone. */
1329 if (GET_MODE_CLASS (mode) == MODE_CC)
1330 return NULL_RTX;
1332 gcc_assert (SCALAR_INT_MODE_P (mode));
1333 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1334 &upper_base, &upper_offs))
1335 return NULL_RTX;
1337 lower_anchor_rtx = GEN_INT (lower_base);
1338 upper_anchor_rtx = GEN_INT (upper_base);
1339 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1340 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1342 if (lower_elt)
1343 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1344 if (upper_elt)
1345 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1347 if (!lower_exp)
1348 return upper_exp;
1349 if (!upper_exp)
1350 return lower_exp;
1352 /* Return the older expression. */
1353 return (upper_old > lower_old ? upper_exp : lower_exp);
1356 /* Look in or update the hash table. */
1358 /* Remove table element ELT from use in the table.
1359 HASH is its hash code, made using the HASH macro.
1360 It's an argument because often that is known in advance
1361 and we save much time not recomputing it. */
1363 static void
1364 remove_from_table (struct table_elt *elt, unsigned int hash)
1366 if (elt == 0)
1367 return;
1369 /* Mark this element as removed. See cse_insn. */
1370 elt->first_same_value = 0;
1372 /* Remove the table element from its equivalence class. */
1375 struct table_elt *prev = elt->prev_same_value;
1376 struct table_elt *next = elt->next_same_value;
1378 if (next)
1379 next->prev_same_value = prev;
1381 if (prev)
1382 prev->next_same_value = next;
1383 else
1385 struct table_elt *newfirst = next;
1386 while (next)
1388 next->first_same_value = newfirst;
1389 next = next->next_same_value;
1394 /* Remove the table element from its hash bucket. */
1397 struct table_elt *prev = elt->prev_same_hash;
1398 struct table_elt *next = elt->next_same_hash;
1400 if (next)
1401 next->prev_same_hash = prev;
1403 if (prev)
1404 prev->next_same_hash = next;
1405 else if (table[hash] == elt)
1406 table[hash] = next;
1407 else
1409 /* This entry is not in the proper hash bucket. This can happen
1410 when two classes were merged by `merge_equiv_classes'. Search
1411 for the hash bucket that it heads. This happens only very
1412 rarely, so the cost is acceptable. */
1413 for (hash = 0; hash < HASH_SIZE; hash++)
1414 if (table[hash] == elt)
1415 table[hash] = next;
1419 /* Remove the table element from its related-value circular chain. */
1421 if (elt->related_value != 0 && elt->related_value != elt)
1423 struct table_elt *p = elt->related_value;
1425 while (p->related_value != elt)
1426 p = p->related_value;
1427 p->related_value = elt->related_value;
1428 if (p->related_value == p)
1429 p->related_value = 0;
1432 /* Now add it to the free element chain. */
1433 elt->next_same_hash = free_element_chain;
1434 free_element_chain = elt;
1437 /* Same as above, but X is a pseudo-register. */
1439 static void
1440 remove_pseudo_from_table (rtx x, unsigned int hash)
1442 struct table_elt *elt;
1444 /* Because a pseudo-register can be referenced in more than one
1445 mode, we might have to remove more than one table entry. */
1446 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1447 remove_from_table (elt, hash);
1450 /* Look up X in the hash table and return its table element,
1451 or 0 if X is not in the table.
1453 MODE is the machine-mode of X, or if X is an integer constant
1454 with VOIDmode then MODE is the mode with which X will be used.
1456 Here we are satisfied to find an expression whose tree structure
1457 looks like X. */
1459 static struct table_elt *
1460 lookup (rtx x, unsigned int hash, machine_mode mode)
1462 struct table_elt *p;
1464 for (p = table[hash]; p; p = p->next_same_hash)
1465 if (mode == p->mode && ((x == p->exp && REG_P (x))
1466 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1467 return p;
1469 return 0;
1472 /* Like `lookup' but don't care whether the table element uses invalid regs.
1473 Also ignore discrepancies in the machine mode of a register. */
1475 static struct table_elt *
1476 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1478 struct table_elt *p;
1480 if (REG_P (x))
1482 unsigned int regno = REGNO (x);
1484 /* Don't check the machine mode when comparing registers;
1485 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1486 for (p = table[hash]; p; p = p->next_same_hash)
1487 if (REG_P (p->exp)
1488 && REGNO (p->exp) == regno)
1489 return p;
1491 else
1493 for (p = table[hash]; p; p = p->next_same_hash)
1494 if (mode == p->mode
1495 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1496 return p;
1499 return 0;
1502 /* Look for an expression equivalent to X and with code CODE.
1503 If one is found, return that expression. */
1505 static rtx
1506 lookup_as_function (rtx x, enum rtx_code code)
1508 struct table_elt *p
1509 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1511 if (p == 0)
1512 return 0;
1514 for (p = p->first_same_value; p; p = p->next_same_value)
1515 if (GET_CODE (p->exp) == code
1516 /* Make sure this is a valid entry in the table. */
1517 && exp_equiv_p (p->exp, p->exp, 1, false))
1518 return p->exp;
1520 return 0;
1523 /* Insert X in the hash table, assuming HASH is its hash code and
1524 CLASSP is an element of the class it should go in (or 0 if a new
1525 class should be made). COST is the code of X and reg_cost is the
1526 cost of registers in X. It is inserted at the proper position to
1527 keep the class in the order cheapest first.
1529 MODE is the machine-mode of X, or if X is an integer constant
1530 with VOIDmode then MODE is the mode with which X will be used.
1532 For elements of equal cheapness, the most recent one
1533 goes in front, except that the first element in the list
1534 remains first unless a cheaper element is added. The order of
1535 pseudo-registers does not matter, as canon_reg will be called to
1536 find the cheapest when a register is retrieved from the table.
1538 The in_memory field in the hash table element is set to 0.
1539 The caller must set it nonzero if appropriate.
1541 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1542 and if insert_regs returns a nonzero value
1543 you must then recompute its hash code before calling here.
1545 If necessary, update table showing constant values of quantities. */
1547 static struct table_elt *
1548 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1549 machine_mode mode, int cost, int reg_cost)
1551 struct table_elt *elt;
1553 /* If X is a register and we haven't made a quantity for it,
1554 something is wrong. */
1555 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1557 /* If X is a hard register, show it is being put in the table. */
1558 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1559 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1561 /* Put an element for X into the right hash bucket. */
1563 elt = free_element_chain;
1564 if (elt)
1565 free_element_chain = elt->next_same_hash;
1566 else
1567 elt = XNEW (struct table_elt);
1569 elt->exp = x;
1570 elt->canon_exp = NULL_RTX;
1571 elt->cost = cost;
1572 elt->regcost = reg_cost;
1573 elt->next_same_value = 0;
1574 elt->prev_same_value = 0;
1575 elt->next_same_hash = table[hash];
1576 elt->prev_same_hash = 0;
1577 elt->related_value = 0;
1578 elt->in_memory = 0;
1579 elt->mode = mode;
1580 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1582 if (table[hash])
1583 table[hash]->prev_same_hash = elt;
1584 table[hash] = elt;
1586 /* Put it into the proper value-class. */
1587 if (classp)
1589 classp = classp->first_same_value;
1590 if (CHEAPER (elt, classp))
1591 /* Insert at the head of the class. */
1593 struct table_elt *p;
1594 elt->next_same_value = classp;
1595 classp->prev_same_value = elt;
1596 elt->first_same_value = elt;
1598 for (p = classp; p; p = p->next_same_value)
1599 p->first_same_value = elt;
1601 else
1603 /* Insert not at head of the class. */
1604 /* Put it after the last element cheaper than X. */
1605 struct table_elt *p, *next;
1607 for (p = classp;
1608 (next = p->next_same_value) && CHEAPER (next, elt);
1609 p = next)
1612 /* Put it after P and before NEXT. */
1613 elt->next_same_value = next;
1614 if (next)
1615 next->prev_same_value = elt;
1617 elt->prev_same_value = p;
1618 p->next_same_value = elt;
1619 elt->first_same_value = classp;
1622 else
1623 elt->first_same_value = elt;
1625 /* If this is a constant being set equivalent to a register or a register
1626 being set equivalent to a constant, note the constant equivalence.
1628 If this is a constant, it cannot be equivalent to a different constant,
1629 and a constant is the only thing that can be cheaper than a register. So
1630 we know the register is the head of the class (before the constant was
1631 inserted).
1633 If this is a register that is not already known equivalent to a
1634 constant, we must check the entire class.
1636 If this is a register that is already known equivalent to an insn,
1637 update the qtys `const_insn' to show that `this_insn' is the latest
1638 insn making that quantity equivalent to the constant. */
1640 if (elt->is_const && classp && REG_P (classp->exp)
1641 && !REG_P (x))
1643 int exp_q = REG_QTY (REGNO (classp->exp));
1644 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1646 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1647 exp_ent->const_insn = this_insn;
1650 else if (REG_P (x)
1651 && classp
1652 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1653 && ! elt->is_const)
1655 struct table_elt *p;
1657 for (p = classp; p != 0; p = p->next_same_value)
1659 if (p->is_const && !REG_P (p->exp))
1661 int x_q = REG_QTY (REGNO (x));
1662 struct qty_table_elem *x_ent = &qty_table[x_q];
1664 x_ent->const_rtx
1665 = gen_lowpart (GET_MODE (x), p->exp);
1666 x_ent->const_insn = this_insn;
1667 break;
1672 else if (REG_P (x)
1673 && qty_table[REG_QTY (REGNO (x))].const_rtx
1674 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1675 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1677 /* If this is a constant with symbolic value,
1678 and it has a term with an explicit integer value,
1679 link it up with related expressions. */
1680 if (GET_CODE (x) == CONST)
1682 rtx subexp = get_related_value (x);
1683 unsigned subhash;
1684 struct table_elt *subelt, *subelt_prev;
1686 if (subexp != 0)
1688 /* Get the integer-free subexpression in the hash table. */
1689 subhash = SAFE_HASH (subexp, mode);
1690 subelt = lookup (subexp, subhash, mode);
1691 if (subelt == 0)
1692 subelt = insert (subexp, NULL, subhash, mode);
1693 /* Initialize SUBELT's circular chain if it has none. */
1694 if (subelt->related_value == 0)
1695 subelt->related_value = subelt;
1696 /* Find the element in the circular chain that precedes SUBELT. */
1697 subelt_prev = subelt;
1698 while (subelt_prev->related_value != subelt)
1699 subelt_prev = subelt_prev->related_value;
1700 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1701 This way the element that follows SUBELT is the oldest one. */
1702 elt->related_value = subelt_prev->related_value;
1703 subelt_prev->related_value = elt;
1707 return elt;
1710 /* Wrap insert_with_costs by passing the default costs. */
1712 static struct table_elt *
1713 insert (rtx x, struct table_elt *classp, unsigned int hash,
1714 machine_mode mode)
1716 return insert_with_costs (x, classp, hash, mode,
1717 COST (x, mode), approx_reg_cost (x));
1721 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1722 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1723 the two classes equivalent.
1725 CLASS1 will be the surviving class; CLASS2 should not be used after this
1726 call.
1728 Any invalid entries in CLASS2 will not be copied. */
1730 static void
1731 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1733 struct table_elt *elt, *next, *new_elt;
1735 /* Ensure we start with the head of the classes. */
1736 class1 = class1->first_same_value;
1737 class2 = class2->first_same_value;
1739 /* If they were already equal, forget it. */
1740 if (class1 == class2)
1741 return;
1743 for (elt = class2; elt; elt = next)
1745 unsigned int hash;
1746 rtx exp = elt->exp;
1747 machine_mode mode = elt->mode;
1749 next = elt->next_same_value;
1751 /* Remove old entry, make a new one in CLASS1's class.
1752 Don't do this for invalid entries as we cannot find their
1753 hash code (it also isn't necessary). */
1754 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1756 bool need_rehash = false;
1758 hash_arg_in_memory = 0;
1759 hash = HASH (exp, mode);
1761 if (REG_P (exp))
1763 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1764 delete_reg_equiv (REGNO (exp));
1767 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1768 remove_pseudo_from_table (exp, hash);
1769 else
1770 remove_from_table (elt, hash);
1772 if (insert_regs (exp, class1, 0) || need_rehash)
1774 rehash_using_reg (exp);
1775 hash = HASH (exp, mode);
1777 new_elt = insert (exp, class1, hash, mode);
1778 new_elt->in_memory = hash_arg_in_memory;
1779 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1780 new_elt->cost = MAX_COST;
1785 /* Flush the entire hash table. */
1787 static void
1788 flush_hash_table (void)
1790 int i;
1791 struct table_elt *p;
1793 for (i = 0; i < HASH_SIZE; i++)
1794 for (p = table[i]; p; p = table[i])
1796 /* Note that invalidate can remove elements
1797 after P in the current hash chain. */
1798 if (REG_P (p->exp))
1799 invalidate (p->exp, VOIDmode);
1800 else
1801 remove_from_table (p, i);
1805 /* Check whether an anti dependence exists between X and EXP. MODE and
1806 ADDR are as for canon_anti_dependence. */
1808 static bool
1809 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1811 subrtx_iterator::array_type array;
1812 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1814 const_rtx x = *iter;
1815 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1816 return true;
1818 return false;
1821 /* Remove from the hash table, or mark as invalid, all expressions whose
1822 values could be altered by storing in X. X is a register, a subreg, or
1823 a memory reference with nonvarying address (because, when a memory
1824 reference with a varying address is stored in, all memory references are
1825 removed by invalidate_memory so specific invalidation is superfluous).
1826 FULL_MODE, if not VOIDmode, indicates that this much should be
1827 invalidated instead of just the amount indicated by the mode of X. This
1828 is only used for bitfield stores into memory.
1830 A nonvarying address may be just a register or just a symbol reference,
1831 or it may be either of those plus a numeric offset. */
1833 static void
1834 invalidate (rtx x, machine_mode full_mode)
1836 int i;
1837 struct table_elt *p;
1838 rtx addr;
1840 switch (GET_CODE (x))
1842 case REG:
1844 /* If X is a register, dependencies on its contents are recorded
1845 through the qty number mechanism. Just change the qty number of
1846 the register, mark it as invalid for expressions that refer to it,
1847 and remove it itself. */
1848 unsigned int regno = REGNO (x);
1849 unsigned int hash = HASH (x, GET_MODE (x));
1851 /* Remove REGNO from any quantity list it might be on and indicate
1852 that its value might have changed. If it is a pseudo, remove its
1853 entry from the hash table.
1855 For a hard register, we do the first two actions above for any
1856 additional hard registers corresponding to X. Then, if any of these
1857 registers are in the table, we must remove any REG entries that
1858 overlap these registers. */
1860 delete_reg_equiv (regno);
1861 REG_TICK (regno)++;
1862 SUBREG_TICKED (regno) = -1;
1864 if (regno >= FIRST_PSEUDO_REGISTER)
1865 remove_pseudo_from_table (x, hash);
1866 else
1868 HOST_WIDE_INT in_table
1869 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1870 unsigned int endregno = END_REGNO (x);
1871 unsigned int tregno, tendregno, rn;
1872 struct table_elt *p, *next;
1874 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1876 for (rn = regno + 1; rn < endregno; rn++)
1878 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1879 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1880 delete_reg_equiv (rn);
1881 REG_TICK (rn)++;
1882 SUBREG_TICKED (rn) = -1;
1885 if (in_table)
1886 for (hash = 0; hash < HASH_SIZE; hash++)
1887 for (p = table[hash]; p; p = next)
1889 next = p->next_same_hash;
1891 if (!REG_P (p->exp)
1892 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1893 continue;
1895 tregno = REGNO (p->exp);
1896 tendregno = END_REGNO (p->exp);
1897 if (tendregno > regno && tregno < endregno)
1898 remove_from_table (p, hash);
1902 return;
1904 case SUBREG:
1905 invalidate (SUBREG_REG (x), VOIDmode);
1906 return;
1908 case PARALLEL:
1909 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1910 invalidate (XVECEXP (x, 0, i), VOIDmode);
1911 return;
1913 case EXPR_LIST:
1914 /* This is part of a disjoint return value; extract the location in
1915 question ignoring the offset. */
1916 invalidate (XEXP (x, 0), VOIDmode);
1917 return;
1919 case MEM:
1920 addr = canon_rtx (get_addr (XEXP (x, 0)));
1921 /* Calculate the canonical version of X here so that
1922 true_dependence doesn't generate new RTL for X on each call. */
1923 x = canon_rtx (x);
1925 /* Remove all hash table elements that refer to overlapping pieces of
1926 memory. */
1927 if (full_mode == VOIDmode)
1928 full_mode = GET_MODE (x);
1930 for (i = 0; i < HASH_SIZE; i++)
1932 struct table_elt *next;
1934 for (p = table[i]; p; p = next)
1936 next = p->next_same_hash;
1937 if (p->in_memory)
1939 /* Just canonicalize the expression once;
1940 otherwise each time we call invalidate
1941 true_dependence will canonicalize the
1942 expression again. */
1943 if (!p->canon_exp)
1944 p->canon_exp = canon_rtx (p->exp);
1945 if (check_dependence (p->canon_exp, x, full_mode, addr))
1946 remove_from_table (p, i);
1950 return;
1952 default:
1953 gcc_unreachable ();
1957 /* Invalidate DEST. Used when DEST is not going to be added
1958 into the hash table for some reason, e.g. do_not_record
1959 flagged on it. */
1961 static void
1962 invalidate_dest (rtx dest)
1964 if (REG_P (dest)
1965 || GET_CODE (dest) == SUBREG
1966 || MEM_P (dest))
1967 invalidate (dest, VOIDmode);
1968 else if (GET_CODE (dest) == STRICT_LOW_PART
1969 || GET_CODE (dest) == ZERO_EXTRACT)
1970 invalidate (XEXP (dest, 0), GET_MODE (dest));
1973 /* Remove all expressions that refer to register REGNO,
1974 since they are already invalid, and we are about to
1975 mark that register valid again and don't want the old
1976 expressions to reappear as valid. */
1978 static void
1979 remove_invalid_refs (unsigned int regno)
1981 unsigned int i;
1982 struct table_elt *p, *next;
1984 for (i = 0; i < HASH_SIZE; i++)
1985 for (p = table[i]; p; p = next)
1987 next = p->next_same_hash;
1988 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1989 remove_from_table (p, i);
1993 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1994 and mode MODE. */
1995 static void
1996 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1997 machine_mode mode)
1999 unsigned int i;
2000 struct table_elt *p, *next;
2001 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2003 for (i = 0; i < HASH_SIZE; i++)
2004 for (p = table[i]; p; p = next)
2006 rtx exp = p->exp;
2007 next = p->next_same_hash;
2009 if (!REG_P (exp)
2010 && (GET_CODE (exp) != SUBREG
2011 || !REG_P (SUBREG_REG (exp))
2012 || REGNO (SUBREG_REG (exp)) != regno
2013 || (((SUBREG_BYTE (exp)
2014 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2015 && SUBREG_BYTE (exp) <= end))
2016 && refers_to_regno_p (regno, p->exp))
2017 remove_from_table (p, i);
2021 /* Recompute the hash codes of any valid entries in the hash table that
2022 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2024 This is called when we make a jump equivalence. */
2026 static void
2027 rehash_using_reg (rtx x)
2029 unsigned int i;
2030 struct table_elt *p, *next;
2031 unsigned hash;
2033 if (GET_CODE (x) == SUBREG)
2034 x = SUBREG_REG (x);
2036 /* If X is not a register or if the register is known not to be in any
2037 valid entries in the table, we have no work to do. */
2039 if (!REG_P (x)
2040 || REG_IN_TABLE (REGNO (x)) < 0
2041 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2042 return;
2044 /* Scan all hash chains looking for valid entries that mention X.
2045 If we find one and it is in the wrong hash chain, move it. */
2047 for (i = 0; i < HASH_SIZE; i++)
2048 for (p = table[i]; p; p = next)
2050 next = p->next_same_hash;
2051 if (reg_mentioned_p (x, p->exp)
2052 && exp_equiv_p (p->exp, p->exp, 1, false)
2053 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2055 if (p->next_same_hash)
2056 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2058 if (p->prev_same_hash)
2059 p->prev_same_hash->next_same_hash = p->next_same_hash;
2060 else
2061 table[i] = p->next_same_hash;
2063 p->next_same_hash = table[hash];
2064 p->prev_same_hash = 0;
2065 if (table[hash])
2066 table[hash]->prev_same_hash = p;
2067 table[hash] = p;
2072 /* Remove from the hash table any expression that is a call-clobbered
2073 register. Also update their TICK values. */
2075 static void
2076 invalidate_for_call (void)
2078 unsigned int regno, endregno;
2079 unsigned int i;
2080 unsigned hash;
2081 struct table_elt *p, *next;
2082 int in_table = 0;
2083 hard_reg_set_iterator hrsi;
2085 /* Go through all the hard registers. For each that is clobbered in
2086 a CALL_INSN, remove the register from quantity chains and update
2087 reg_tick if defined. Also see if any of these registers is currently
2088 in the table. */
2089 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2091 delete_reg_equiv (regno);
2092 if (REG_TICK (regno) >= 0)
2094 REG_TICK (regno)++;
2095 SUBREG_TICKED (regno) = -1;
2097 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2100 /* In the case where we have no call-clobbered hard registers in the
2101 table, we are done. Otherwise, scan the table and remove any
2102 entry that overlaps a call-clobbered register. */
2104 if (in_table)
2105 for (hash = 0; hash < HASH_SIZE; hash++)
2106 for (p = table[hash]; p; p = next)
2108 next = p->next_same_hash;
2110 if (!REG_P (p->exp)
2111 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2112 continue;
2114 regno = REGNO (p->exp);
2115 endregno = END_REGNO (p->exp);
2117 for (i = regno; i < endregno; i++)
2118 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2120 remove_from_table (p, hash);
2121 break;
2126 /* Given an expression X of type CONST,
2127 and ELT which is its table entry (or 0 if it
2128 is not in the hash table),
2129 return an alternate expression for X as a register plus integer.
2130 If none can be found, return 0. */
2132 static rtx
2133 use_related_value (rtx x, struct table_elt *elt)
2135 struct table_elt *relt = 0;
2136 struct table_elt *p, *q;
2137 HOST_WIDE_INT offset;
2139 /* First, is there anything related known?
2140 If we have a table element, we can tell from that.
2141 Otherwise, must look it up. */
2143 if (elt != 0 && elt->related_value != 0)
2144 relt = elt;
2145 else if (elt == 0 && GET_CODE (x) == CONST)
2147 rtx subexp = get_related_value (x);
2148 if (subexp != 0)
2149 relt = lookup (subexp,
2150 SAFE_HASH (subexp, GET_MODE (subexp)),
2151 GET_MODE (subexp));
2154 if (relt == 0)
2155 return 0;
2157 /* Search all related table entries for one that has an
2158 equivalent register. */
2160 p = relt;
2161 while (1)
2163 /* This loop is strange in that it is executed in two different cases.
2164 The first is when X is already in the table. Then it is searching
2165 the RELATED_VALUE list of X's class (RELT). The second case is when
2166 X is not in the table. Then RELT points to a class for the related
2167 value.
2169 Ensure that, whatever case we are in, that we ignore classes that have
2170 the same value as X. */
2172 if (rtx_equal_p (x, p->exp))
2173 q = 0;
2174 else
2175 for (q = p->first_same_value; q; q = q->next_same_value)
2176 if (REG_P (q->exp))
2177 break;
2179 if (q)
2180 break;
2182 p = p->related_value;
2184 /* We went all the way around, so there is nothing to be found.
2185 Alternatively, perhaps RELT was in the table for some other reason
2186 and it has no related values recorded. */
2187 if (p == relt || p == 0)
2188 break;
2191 if (q == 0)
2192 return 0;
2194 offset = (get_integer_term (x) - get_integer_term (p->exp));
2195 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2196 return plus_constant (q->mode, q->exp, offset);
2200 /* Hash a string. Just add its bytes up. */
2201 static inline unsigned
2202 hash_rtx_string (const char *ps)
2204 unsigned hash = 0;
2205 const unsigned char *p = (const unsigned char *) ps;
2207 if (p)
2208 while (*p)
2209 hash += *p++;
2211 return hash;
2214 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2215 When the callback returns true, we continue with the new rtx. */
2217 unsigned
2218 hash_rtx_cb (const_rtx x, machine_mode mode,
2219 int *do_not_record_p, int *hash_arg_in_memory_p,
2220 bool have_reg_qty, hash_rtx_callback_function cb)
2222 int i, j;
2223 unsigned hash = 0;
2224 enum rtx_code code;
2225 const char *fmt;
2226 machine_mode newmode;
2227 rtx newx;
2229 /* Used to turn recursion into iteration. We can't rely on GCC's
2230 tail-recursion elimination since we need to keep accumulating values
2231 in HASH. */
2232 repeat:
2233 if (x == 0)
2234 return hash;
2236 /* Invoke the callback first. */
2237 if (cb != NULL
2238 && ((*cb) (x, mode, &newx, &newmode)))
2240 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2241 hash_arg_in_memory_p, have_reg_qty, cb);
2242 return hash;
2245 code = GET_CODE (x);
2246 switch (code)
2248 case REG:
2250 unsigned int regno = REGNO (x);
2252 if (do_not_record_p && !reload_completed)
2254 /* On some machines, we can't record any non-fixed hard register,
2255 because extending its life will cause reload problems. We
2256 consider ap, fp, sp, gp to be fixed for this purpose.
2258 We also consider CCmode registers to be fixed for this purpose;
2259 failure to do so leads to failure to simplify 0<100 type of
2260 conditionals.
2262 On all machines, we can't record any global registers.
2263 Nor should we record any register that is in a small
2264 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2265 bool record;
2267 if (regno >= FIRST_PSEUDO_REGISTER)
2268 record = true;
2269 else if (x == frame_pointer_rtx
2270 || x == hard_frame_pointer_rtx
2271 || x == arg_pointer_rtx
2272 || x == stack_pointer_rtx
2273 || x == pic_offset_table_rtx)
2274 record = true;
2275 else if (global_regs[regno])
2276 record = false;
2277 else if (fixed_regs[regno])
2278 record = true;
2279 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2280 record = true;
2281 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2282 record = false;
2283 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2284 record = false;
2285 else
2286 record = true;
2288 if (!record)
2290 *do_not_record_p = 1;
2291 return 0;
2295 hash += ((unsigned int) REG << 7);
2296 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2297 return hash;
2300 /* We handle SUBREG of a REG specially because the underlying
2301 reg changes its hash value with every value change; we don't
2302 want to have to forget unrelated subregs when one subreg changes. */
2303 case SUBREG:
2305 if (REG_P (SUBREG_REG (x)))
2307 hash += (((unsigned int) SUBREG << 7)
2308 + REGNO (SUBREG_REG (x))
2309 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2310 return hash;
2312 break;
2315 case CONST_INT:
2316 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2317 + (unsigned int) INTVAL (x));
2318 return hash;
2320 case CONST_WIDE_INT:
2321 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2322 hash += CONST_WIDE_INT_ELT (x, i);
2323 return hash;
2325 case CONST_DOUBLE:
2326 /* This is like the general case, except that it only counts
2327 the integers representing the constant. */
2328 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2329 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2330 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2331 + (unsigned int) CONST_DOUBLE_HIGH (x));
2332 else
2333 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2334 return hash;
2336 case CONST_FIXED:
2337 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2338 hash += fixed_hash (CONST_FIXED_VALUE (x));
2339 return hash;
2341 case CONST_VECTOR:
2343 int units;
2344 rtx elt;
2346 units = CONST_VECTOR_NUNITS (x);
2348 for (i = 0; i < units; ++i)
2350 elt = CONST_VECTOR_ELT (x, i);
2351 hash += hash_rtx_cb (elt, GET_MODE (elt),
2352 do_not_record_p, hash_arg_in_memory_p,
2353 have_reg_qty, cb);
2356 return hash;
2359 /* Assume there is only one rtx object for any given label. */
2360 case LABEL_REF:
2361 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2362 differences and differences between each stage's debugging dumps. */
2363 hash += (((unsigned int) LABEL_REF << 7)
2364 + CODE_LABEL_NUMBER (label_ref_label (x)));
2365 return hash;
2367 case SYMBOL_REF:
2369 /* Don't hash on the symbol's address to avoid bootstrap differences.
2370 Different hash values may cause expressions to be recorded in
2371 different orders and thus different registers to be used in the
2372 final assembler. This also avoids differences in the dump files
2373 between various stages. */
2374 unsigned int h = 0;
2375 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2377 while (*p)
2378 h += (h << 7) + *p++; /* ??? revisit */
2380 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2381 return hash;
2384 case MEM:
2385 /* We don't record if marked volatile or if BLKmode since we don't
2386 know the size of the move. */
2387 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2389 *do_not_record_p = 1;
2390 return 0;
2392 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2393 *hash_arg_in_memory_p = 1;
2395 /* Now that we have already found this special case,
2396 might as well speed it up as much as possible. */
2397 hash += (unsigned) MEM;
2398 x = XEXP (x, 0);
2399 goto repeat;
2401 case USE:
2402 /* A USE that mentions non-volatile memory needs special
2403 handling since the MEM may be BLKmode which normally
2404 prevents an entry from being made. Pure calls are
2405 marked by a USE which mentions BLKmode memory.
2406 See calls.c:emit_call_1. */
2407 if (MEM_P (XEXP (x, 0))
2408 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2410 hash += (unsigned) USE;
2411 x = XEXP (x, 0);
2413 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2414 *hash_arg_in_memory_p = 1;
2416 /* Now that we have already found this special case,
2417 might as well speed it up as much as possible. */
2418 hash += (unsigned) MEM;
2419 x = XEXP (x, 0);
2420 goto repeat;
2422 break;
2424 case PRE_DEC:
2425 case PRE_INC:
2426 case POST_DEC:
2427 case POST_INC:
2428 case PRE_MODIFY:
2429 case POST_MODIFY:
2430 case PC:
2431 case CC0:
2432 case CALL:
2433 case UNSPEC_VOLATILE:
2434 if (do_not_record_p) {
2435 *do_not_record_p = 1;
2436 return 0;
2438 else
2439 return hash;
2440 break;
2442 case ASM_OPERANDS:
2443 if (do_not_record_p && MEM_VOLATILE_P (x))
2445 *do_not_record_p = 1;
2446 return 0;
2448 else
2450 /* We don't want to take the filename and line into account. */
2451 hash += (unsigned) code + (unsigned) GET_MODE (x)
2452 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2453 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2454 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2456 if (ASM_OPERANDS_INPUT_LENGTH (x))
2458 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2460 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2461 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2462 do_not_record_p, hash_arg_in_memory_p,
2463 have_reg_qty, cb)
2464 + hash_rtx_string
2465 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2468 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2469 x = ASM_OPERANDS_INPUT (x, 0);
2470 mode = GET_MODE (x);
2471 goto repeat;
2474 return hash;
2476 break;
2478 default:
2479 break;
2482 i = GET_RTX_LENGTH (code) - 1;
2483 hash += (unsigned) code + (unsigned) GET_MODE (x);
2484 fmt = GET_RTX_FORMAT (code);
2485 for (; i >= 0; i--)
2487 switch (fmt[i])
2489 case 'e':
2490 /* If we are about to do the last recursive call
2491 needed at this level, change it into iteration.
2492 This function is called enough to be worth it. */
2493 if (i == 0)
2495 x = XEXP (x, i);
2496 goto repeat;
2499 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2500 hash_arg_in_memory_p,
2501 have_reg_qty, cb);
2502 break;
2504 case 'E':
2505 for (j = 0; j < XVECLEN (x, i); j++)
2506 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2507 hash_arg_in_memory_p,
2508 have_reg_qty, cb);
2509 break;
2511 case 's':
2512 hash += hash_rtx_string (XSTR (x, i));
2513 break;
2515 case 'i':
2516 hash += (unsigned int) XINT (x, i);
2517 break;
2519 case '0': case 't':
2520 /* Unused. */
2521 break;
2523 default:
2524 gcc_unreachable ();
2528 return hash;
2531 /* Hash an rtx. We are careful to make sure the value is never negative.
2532 Equivalent registers hash identically.
2533 MODE is used in hashing for CONST_INTs only;
2534 otherwise the mode of X is used.
2536 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2538 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2539 a MEM rtx which does not have the MEM_READONLY_P flag set.
2541 Note that cse_insn knows that the hash code of a MEM expression
2542 is just (int) MEM plus the hash code of the address. */
2544 unsigned
2545 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2546 int *hash_arg_in_memory_p, bool have_reg_qty)
2548 return hash_rtx_cb (x, mode, do_not_record_p,
2549 hash_arg_in_memory_p, have_reg_qty, NULL);
2552 /* Hash an rtx X for cse via hash_rtx.
2553 Stores 1 in do_not_record if any subexpression is volatile.
2554 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2555 does not have the MEM_READONLY_P flag set. */
2557 static inline unsigned
2558 canon_hash (rtx x, machine_mode mode)
2560 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2563 /* Like canon_hash but with no side effects, i.e. do_not_record
2564 and hash_arg_in_memory are not changed. */
2566 static inline unsigned
2567 safe_hash (rtx x, machine_mode mode)
2569 int dummy_do_not_record;
2570 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2573 /* Return 1 iff X and Y would canonicalize into the same thing,
2574 without actually constructing the canonicalization of either one.
2575 If VALIDATE is nonzero,
2576 we assume X is an expression being processed from the rtl
2577 and Y was found in the hash table. We check register refs
2578 in Y for being marked as valid.
2580 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2583 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2585 int i, j;
2586 enum rtx_code code;
2587 const char *fmt;
2589 /* Note: it is incorrect to assume an expression is equivalent to itself
2590 if VALIDATE is nonzero. */
2591 if (x == y && !validate)
2592 return 1;
2594 if (x == 0 || y == 0)
2595 return x == y;
2597 code = GET_CODE (x);
2598 if (code != GET_CODE (y))
2599 return 0;
2601 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2602 if (GET_MODE (x) != GET_MODE (y))
2603 return 0;
2605 /* MEMs referring to different address space are not equivalent. */
2606 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2607 return 0;
2609 switch (code)
2611 case PC:
2612 case CC0:
2613 CASE_CONST_UNIQUE:
2614 return x == y;
2616 case LABEL_REF:
2617 return label_ref_label (x) == label_ref_label (y);
2619 case SYMBOL_REF:
2620 return XSTR (x, 0) == XSTR (y, 0);
2622 case REG:
2623 if (for_gcse)
2624 return REGNO (x) == REGNO (y);
2625 else
2627 unsigned int regno = REGNO (y);
2628 unsigned int i;
2629 unsigned int endregno = END_REGNO (y);
2631 /* If the quantities are not the same, the expressions are not
2632 equivalent. If there are and we are not to validate, they
2633 are equivalent. Otherwise, ensure all regs are up-to-date. */
2635 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2636 return 0;
2638 if (! validate)
2639 return 1;
2641 for (i = regno; i < endregno; i++)
2642 if (REG_IN_TABLE (i) != REG_TICK (i))
2643 return 0;
2645 return 1;
2648 case MEM:
2649 if (for_gcse)
2651 /* A volatile mem should not be considered equivalent to any
2652 other. */
2653 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2654 return 0;
2656 /* Can't merge two expressions in different alias sets, since we
2657 can decide that the expression is transparent in a block when
2658 it isn't, due to it being set with the different alias set.
2660 Also, can't merge two expressions with different MEM_ATTRS.
2661 They could e.g. be two different entities allocated into the
2662 same space on the stack (see e.g. PR25130). In that case, the
2663 MEM addresses can be the same, even though the two MEMs are
2664 absolutely not equivalent.
2666 But because really all MEM attributes should be the same for
2667 equivalent MEMs, we just use the invariant that MEMs that have
2668 the same attributes share the same mem_attrs data structure. */
2669 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2670 return 0;
2672 /* If we are handling exceptions, we cannot consider two expressions
2673 with different trapping status as equivalent, because simple_mem
2674 might accept one and reject the other. */
2675 if (cfun->can_throw_non_call_exceptions
2676 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2677 return 0;
2679 break;
2681 /* For commutative operations, check both orders. */
2682 case PLUS:
2683 case MULT:
2684 case AND:
2685 case IOR:
2686 case XOR:
2687 case NE:
2688 case EQ:
2689 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2690 validate, for_gcse)
2691 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2692 validate, for_gcse))
2693 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2694 validate, for_gcse)
2695 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2696 validate, for_gcse)));
2698 case ASM_OPERANDS:
2699 /* We don't use the generic code below because we want to
2700 disregard filename and line numbers. */
2702 /* A volatile asm isn't equivalent to any other. */
2703 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2704 return 0;
2706 if (GET_MODE (x) != GET_MODE (y)
2707 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2708 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2709 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2710 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2711 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2712 return 0;
2714 if (ASM_OPERANDS_INPUT_LENGTH (x))
2716 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2717 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2718 ASM_OPERANDS_INPUT (y, i),
2719 validate, for_gcse)
2720 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2721 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2722 return 0;
2725 return 1;
2727 default:
2728 break;
2731 /* Compare the elements. If any pair of corresponding elements
2732 fail to match, return 0 for the whole thing. */
2734 fmt = GET_RTX_FORMAT (code);
2735 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2737 switch (fmt[i])
2739 case 'e':
2740 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2741 validate, for_gcse))
2742 return 0;
2743 break;
2745 case 'E':
2746 if (XVECLEN (x, i) != XVECLEN (y, i))
2747 return 0;
2748 for (j = 0; j < XVECLEN (x, i); j++)
2749 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2750 validate, for_gcse))
2751 return 0;
2752 break;
2754 case 's':
2755 if (strcmp (XSTR (x, i), XSTR (y, i)))
2756 return 0;
2757 break;
2759 case 'i':
2760 if (XINT (x, i) != XINT (y, i))
2761 return 0;
2762 break;
2764 case 'w':
2765 if (XWINT (x, i) != XWINT (y, i))
2766 return 0;
2767 break;
2769 case '0':
2770 case 't':
2771 break;
2773 default:
2774 gcc_unreachable ();
2778 return 1;
2781 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2782 the result if necessary. INSN is as for canon_reg. */
2784 static void
2785 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2787 if (*xloc)
2789 rtx new_rtx = canon_reg (*xloc, insn);
2791 /* If replacing pseudo with hard reg or vice versa, ensure the
2792 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2793 gcc_assert (insn && new_rtx);
2794 validate_change (insn, xloc, new_rtx, 1);
2798 /* Canonicalize an expression:
2799 replace each register reference inside it
2800 with the "oldest" equivalent register.
2802 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2803 after we make our substitution. The calls are made with IN_GROUP nonzero
2804 so apply_change_group must be called upon the outermost return from this
2805 function (unless INSN is zero). The result of apply_change_group can
2806 generally be discarded since the changes we are making are optional. */
2808 static rtx
2809 canon_reg (rtx x, rtx_insn *insn)
2811 int i;
2812 enum rtx_code code;
2813 const char *fmt;
2815 if (x == 0)
2816 return x;
2818 code = GET_CODE (x);
2819 switch (code)
2821 case PC:
2822 case CC0:
2823 case CONST:
2824 CASE_CONST_ANY:
2825 case SYMBOL_REF:
2826 case LABEL_REF:
2827 case ADDR_VEC:
2828 case ADDR_DIFF_VEC:
2829 return x;
2831 case REG:
2833 int first;
2834 int q;
2835 struct qty_table_elem *ent;
2837 /* Never replace a hard reg, because hard regs can appear
2838 in more than one machine mode, and we must preserve the mode
2839 of each occurrence. Also, some hard regs appear in
2840 MEMs that are shared and mustn't be altered. Don't try to
2841 replace any reg that maps to a reg of class NO_REGS. */
2842 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2843 || ! REGNO_QTY_VALID_P (REGNO (x)))
2844 return x;
2846 q = REG_QTY (REGNO (x));
2847 ent = &qty_table[q];
2848 first = ent->first_reg;
2849 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2850 : REGNO_REG_CLASS (first) == NO_REGS ? x
2851 : gen_rtx_REG (ent->mode, first));
2854 default:
2855 break;
2858 fmt = GET_RTX_FORMAT (code);
2859 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2861 int j;
2863 if (fmt[i] == 'e')
2864 validate_canon_reg (&XEXP (x, i), insn);
2865 else if (fmt[i] == 'E')
2866 for (j = 0; j < XVECLEN (x, i); j++)
2867 validate_canon_reg (&XVECEXP (x, i, j), insn);
2870 return x;
2873 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2874 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2875 what values are being compared.
2877 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2878 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2879 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2880 compared to produce cc0.
2882 The return value is the comparison operator and is either the code of
2883 A or the code corresponding to the inverse of the comparison. */
2885 static enum rtx_code
2886 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2887 machine_mode *pmode1, machine_mode *pmode2)
2889 rtx arg1, arg2;
2890 hash_set<rtx> *visited = NULL;
2891 /* Set nonzero when we find something of interest. */
2892 rtx x = NULL;
2894 arg1 = *parg1, arg2 = *parg2;
2896 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2898 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2900 int reverse_code = 0;
2901 struct table_elt *p = 0;
2903 /* Remember state from previous iteration. */
2904 if (x)
2906 if (!visited)
2907 visited = new hash_set<rtx>;
2908 visited->add (x);
2909 x = 0;
2912 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2913 On machines with CC0, this is the only case that can occur, since
2914 fold_rtx will return the COMPARE or item being compared with zero
2915 when given CC0. */
2917 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2918 x = arg1;
2920 /* If ARG1 is a comparison operator and CODE is testing for
2921 STORE_FLAG_VALUE, get the inner arguments. */
2923 else if (COMPARISON_P (arg1))
2925 #ifdef FLOAT_STORE_FLAG_VALUE
2926 REAL_VALUE_TYPE fsfv;
2927 #endif
2929 if (code == NE
2930 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2931 && code == LT && STORE_FLAG_VALUE == -1)
2932 #ifdef FLOAT_STORE_FLAG_VALUE
2933 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2934 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2935 REAL_VALUE_NEGATIVE (fsfv)))
2936 #endif
2938 x = arg1;
2939 else if (code == EQ
2940 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2941 && code == GE && STORE_FLAG_VALUE == -1)
2942 #ifdef FLOAT_STORE_FLAG_VALUE
2943 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2944 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2945 REAL_VALUE_NEGATIVE (fsfv)))
2946 #endif
2948 x = arg1, reverse_code = 1;
2951 /* ??? We could also check for
2953 (ne (and (eq (...) (const_int 1))) (const_int 0))
2955 and related forms, but let's wait until we see them occurring. */
2957 if (x == 0)
2958 /* Look up ARG1 in the hash table and see if it has an equivalence
2959 that lets us see what is being compared. */
2960 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2961 if (p)
2963 p = p->first_same_value;
2965 /* If what we compare is already known to be constant, that is as
2966 good as it gets.
2967 We need to break the loop in this case, because otherwise we
2968 can have an infinite loop when looking at a reg that is known
2969 to be a constant which is the same as a comparison of a reg
2970 against zero which appears later in the insn stream, which in
2971 turn is constant and the same as the comparison of the first reg
2972 against zero... */
2973 if (p->is_const)
2974 break;
2977 for (; p; p = p->next_same_value)
2979 machine_mode inner_mode = GET_MODE (p->exp);
2980 #ifdef FLOAT_STORE_FLAG_VALUE
2981 REAL_VALUE_TYPE fsfv;
2982 #endif
2984 /* If the entry isn't valid, skip it. */
2985 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2986 continue;
2988 /* If it's a comparison we've used before, skip it. */
2989 if (visited && visited->contains (p->exp))
2990 continue;
2992 if (GET_CODE (p->exp) == COMPARE
2993 /* Another possibility is that this machine has a compare insn
2994 that includes the comparison code. In that case, ARG1 would
2995 be equivalent to a comparison operation that would set ARG1 to
2996 either STORE_FLAG_VALUE or zero. If this is an NE operation,
2997 ORIG_CODE is the actual comparison being done; if it is an EQ,
2998 we must reverse ORIG_CODE. On machine with a negative value
2999 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3000 || ((code == NE
3001 || (code == LT
3002 && val_signbit_known_set_p (inner_mode,
3003 STORE_FLAG_VALUE))
3004 #ifdef FLOAT_STORE_FLAG_VALUE
3005 || (code == LT
3006 && SCALAR_FLOAT_MODE_P (inner_mode)
3007 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3008 REAL_VALUE_NEGATIVE (fsfv)))
3009 #endif
3011 && COMPARISON_P (p->exp)))
3013 x = p->exp;
3014 break;
3016 else if ((code == EQ
3017 || (code == GE
3018 && val_signbit_known_set_p (inner_mode,
3019 STORE_FLAG_VALUE))
3020 #ifdef FLOAT_STORE_FLAG_VALUE
3021 || (code == GE
3022 && SCALAR_FLOAT_MODE_P (inner_mode)
3023 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3024 REAL_VALUE_NEGATIVE (fsfv)))
3025 #endif
3027 && COMPARISON_P (p->exp))
3029 reverse_code = 1;
3030 x = p->exp;
3031 break;
3034 /* If this non-trapping address, e.g. fp + constant, the
3035 equivalent is a better operand since it may let us predict
3036 the value of the comparison. */
3037 else if (!rtx_addr_can_trap_p (p->exp))
3039 arg1 = p->exp;
3040 continue;
3044 /* If we didn't find a useful equivalence for ARG1, we are done.
3045 Otherwise, set up for the next iteration. */
3046 if (x == 0)
3047 break;
3049 /* If we need to reverse the comparison, make sure that is
3050 possible -- we can't necessarily infer the value of GE from LT
3051 with floating-point operands. */
3052 if (reverse_code)
3054 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3055 if (reversed == UNKNOWN)
3056 break;
3057 else
3058 code = reversed;
3060 else if (COMPARISON_P (x))
3061 code = GET_CODE (x);
3062 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3065 /* Return our results. Return the modes from before fold_rtx
3066 because fold_rtx might produce const_int, and then it's too late. */
3067 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3068 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3070 if (visited)
3071 delete visited;
3072 return code;
3075 /* If X is a nontrivial arithmetic operation on an argument for which
3076 a constant value can be determined, return the result of operating
3077 on that value, as a constant. Otherwise, return X, possibly with
3078 one or more operands changed to a forward-propagated constant.
3080 If X is a register whose contents are known, we do NOT return
3081 those contents here; equiv_constant is called to perform that task.
3082 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3084 INSN is the insn that we may be modifying. If it is 0, make a copy
3085 of X before modifying it. */
3087 static rtx
3088 fold_rtx (rtx x, rtx_insn *insn)
3090 enum rtx_code code;
3091 machine_mode mode;
3092 const char *fmt;
3093 int i;
3094 rtx new_rtx = 0;
3095 int changed = 0;
3097 /* Operands of X. */
3098 /* Workaround -Wmaybe-uninitialized false positive during
3099 profiledbootstrap by initializing them. */
3100 rtx folded_arg0 = NULL_RTX;
3101 rtx folded_arg1 = NULL_RTX;
3103 /* Constant equivalents of first three operands of X;
3104 0 when no such equivalent is known. */
3105 rtx const_arg0;
3106 rtx const_arg1;
3107 rtx const_arg2;
3109 /* The mode of the first operand of X. We need this for sign and zero
3110 extends. */
3111 machine_mode mode_arg0;
3113 if (x == 0)
3114 return x;
3116 /* Try to perform some initial simplifications on X. */
3117 code = GET_CODE (x);
3118 switch (code)
3120 case MEM:
3121 case SUBREG:
3122 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3123 than it would in other contexts. Basically its mode does not
3124 signify the size of the object read. That information is carried
3125 by size operand. If we happen to have a MEM of the appropriate
3126 mode in our tables with a constant value we could simplify the
3127 extraction incorrectly if we allowed substitution of that value
3128 for the MEM. */
3129 case ZERO_EXTRACT:
3130 case SIGN_EXTRACT:
3131 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3132 return new_rtx;
3133 return x;
3135 case CONST:
3136 CASE_CONST_ANY:
3137 case SYMBOL_REF:
3138 case LABEL_REF:
3139 case REG:
3140 case PC:
3141 /* No use simplifying an EXPR_LIST
3142 since they are used only for lists of args
3143 in a function call's REG_EQUAL note. */
3144 case EXPR_LIST:
3145 return x;
3147 case CC0:
3148 return prev_insn_cc0;
3150 case ASM_OPERANDS:
3151 if (insn)
3153 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3154 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3155 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3157 return x;
3159 case CALL:
3160 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3161 return x;
3162 break;
3164 /* Anything else goes through the loop below. */
3165 default:
3166 break;
3169 mode = GET_MODE (x);
3170 const_arg0 = 0;
3171 const_arg1 = 0;
3172 const_arg2 = 0;
3173 mode_arg0 = VOIDmode;
3175 /* Try folding our operands.
3176 Then see which ones have constant values known. */
3178 fmt = GET_RTX_FORMAT (code);
3179 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3180 if (fmt[i] == 'e')
3182 rtx folded_arg = XEXP (x, i), const_arg;
3183 machine_mode mode_arg = GET_MODE (folded_arg);
3185 switch (GET_CODE (folded_arg))
3187 case MEM:
3188 case REG:
3189 case SUBREG:
3190 const_arg = equiv_constant (folded_arg);
3191 break;
3193 case CONST:
3194 CASE_CONST_ANY:
3195 case SYMBOL_REF:
3196 case LABEL_REF:
3197 const_arg = folded_arg;
3198 break;
3200 case CC0:
3201 /* The cc0-user and cc0-setter may be in different blocks if
3202 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3203 will have been cleared as we exited the block with the
3204 setter.
3206 While we could potentially track cc0 in this case, it just
3207 doesn't seem to be worth it given that cc0 targets are not
3208 terribly common or important these days and trapping math
3209 is rarely used. The combination of those two conditions
3210 necessary to trip this situation is exceedingly rare in the
3211 real world. */
3212 if (!prev_insn_cc0)
3214 const_arg = NULL_RTX;
3216 else
3218 folded_arg = prev_insn_cc0;
3219 mode_arg = prev_insn_cc0_mode;
3220 const_arg = equiv_constant (folded_arg);
3222 break;
3224 default:
3225 folded_arg = fold_rtx (folded_arg, insn);
3226 const_arg = equiv_constant (folded_arg);
3227 break;
3230 /* For the first three operands, see if the operand
3231 is constant or equivalent to a constant. */
3232 switch (i)
3234 case 0:
3235 folded_arg0 = folded_arg;
3236 const_arg0 = const_arg;
3237 mode_arg0 = mode_arg;
3238 break;
3239 case 1:
3240 folded_arg1 = folded_arg;
3241 const_arg1 = const_arg;
3242 break;
3243 case 2:
3244 const_arg2 = const_arg;
3245 break;
3248 /* Pick the least expensive of the argument and an equivalent constant
3249 argument. */
3250 if (const_arg != 0
3251 && const_arg != folded_arg
3252 && (COST_IN (const_arg, mode_arg, code, i)
3253 <= COST_IN (folded_arg, mode_arg, code, i))
3255 /* It's not safe to substitute the operand of a conversion
3256 operator with a constant, as the conversion's identity
3257 depends upon the mode of its operand. This optimization
3258 is handled by the call to simplify_unary_operation. */
3259 && (GET_RTX_CLASS (code) != RTX_UNARY
3260 || GET_MODE (const_arg) == mode_arg0
3261 || (code != ZERO_EXTEND
3262 && code != SIGN_EXTEND
3263 && code != TRUNCATE
3264 && code != FLOAT_TRUNCATE
3265 && code != FLOAT_EXTEND
3266 && code != FLOAT
3267 && code != FIX
3268 && code != UNSIGNED_FLOAT
3269 && code != UNSIGNED_FIX)))
3270 folded_arg = const_arg;
3272 if (folded_arg == XEXP (x, i))
3273 continue;
3275 if (insn == NULL_RTX && !changed)
3276 x = copy_rtx (x);
3277 changed = 1;
3278 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3281 if (changed)
3283 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3284 consistent with the order in X. */
3285 if (canonicalize_change_group (insn, x))
3287 std::swap (const_arg0, const_arg1);
3288 std::swap (folded_arg0, folded_arg1);
3291 apply_change_group ();
3294 /* If X is an arithmetic operation, see if we can simplify it. */
3296 switch (GET_RTX_CLASS (code))
3298 case RTX_UNARY:
3300 /* We can't simplify extension ops unless we know the
3301 original mode. */
3302 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3303 && mode_arg0 == VOIDmode)
3304 break;
3306 new_rtx = simplify_unary_operation (code, mode,
3307 const_arg0 ? const_arg0 : folded_arg0,
3308 mode_arg0);
3310 break;
3312 case RTX_COMPARE:
3313 case RTX_COMM_COMPARE:
3314 /* See what items are actually being compared and set FOLDED_ARG[01]
3315 to those values and CODE to the actual comparison code. If any are
3316 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3317 do anything if both operands are already known to be constant. */
3319 /* ??? Vector mode comparisons are not supported yet. */
3320 if (VECTOR_MODE_P (mode))
3321 break;
3323 if (const_arg0 == 0 || const_arg1 == 0)
3325 struct table_elt *p0, *p1;
3326 rtx true_rtx, false_rtx;
3327 machine_mode mode_arg1;
3329 if (SCALAR_FLOAT_MODE_P (mode))
3331 #ifdef FLOAT_STORE_FLAG_VALUE
3332 true_rtx = (const_double_from_real_value
3333 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3334 #else
3335 true_rtx = NULL_RTX;
3336 #endif
3337 false_rtx = CONST0_RTX (mode);
3339 else
3341 true_rtx = const_true_rtx;
3342 false_rtx = const0_rtx;
3345 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3346 &mode_arg0, &mode_arg1);
3348 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3349 what kinds of things are being compared, so we can't do
3350 anything with this comparison. */
3352 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3353 break;
3355 const_arg0 = equiv_constant (folded_arg0);
3356 const_arg1 = equiv_constant (folded_arg1);
3358 /* If we do not now have two constants being compared, see
3359 if we can nevertheless deduce some things about the
3360 comparison. */
3361 if (const_arg0 == 0 || const_arg1 == 0)
3363 if (const_arg1 != NULL)
3365 rtx cheapest_simplification;
3366 int cheapest_cost;
3367 rtx simp_result;
3368 struct table_elt *p;
3370 /* See if we can find an equivalent of folded_arg0
3371 that gets us a cheaper expression, possibly a
3372 constant through simplifications. */
3373 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3374 mode_arg0);
3376 if (p != NULL)
3378 cheapest_simplification = x;
3379 cheapest_cost = COST (x, mode);
3381 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3383 int cost;
3385 /* If the entry isn't valid, skip it. */
3386 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3387 continue;
3389 /* Try to simplify using this equivalence. */
3390 simp_result
3391 = simplify_relational_operation (code, mode,
3392 mode_arg0,
3393 p->exp,
3394 const_arg1);
3396 if (simp_result == NULL)
3397 continue;
3399 cost = COST (simp_result, mode);
3400 if (cost < cheapest_cost)
3402 cheapest_cost = cost;
3403 cheapest_simplification = simp_result;
3407 /* If we have a cheaper expression now, use that
3408 and try folding it further, from the top. */
3409 if (cheapest_simplification != x)
3410 return fold_rtx (copy_rtx (cheapest_simplification),
3411 insn);
3415 /* See if the two operands are the same. */
3417 if ((REG_P (folded_arg0)
3418 && REG_P (folded_arg1)
3419 && (REG_QTY (REGNO (folded_arg0))
3420 == REG_QTY (REGNO (folded_arg1))))
3421 || ((p0 = lookup (folded_arg0,
3422 SAFE_HASH (folded_arg0, mode_arg0),
3423 mode_arg0))
3424 && (p1 = lookup (folded_arg1,
3425 SAFE_HASH (folded_arg1, mode_arg0),
3426 mode_arg0))
3427 && p0->first_same_value == p1->first_same_value))
3428 folded_arg1 = folded_arg0;
3430 /* If FOLDED_ARG0 is a register, see if the comparison we are
3431 doing now is either the same as we did before or the reverse
3432 (we only check the reverse if not floating-point). */
3433 else if (REG_P (folded_arg0))
3435 int qty = REG_QTY (REGNO (folded_arg0));
3437 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3439 struct qty_table_elem *ent = &qty_table[qty];
3441 if ((comparison_dominates_p (ent->comparison_code, code)
3442 || (! FLOAT_MODE_P (mode_arg0)
3443 && comparison_dominates_p (ent->comparison_code,
3444 reverse_condition (code))))
3445 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3446 || (const_arg1
3447 && rtx_equal_p (ent->comparison_const,
3448 const_arg1))
3449 || (REG_P (folded_arg1)
3450 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3452 if (comparison_dominates_p (ent->comparison_code, code))
3454 if (true_rtx)
3455 return true_rtx;
3456 else
3457 break;
3459 else
3460 return false_rtx;
3467 /* If we are comparing against zero, see if the first operand is
3468 equivalent to an IOR with a constant. If so, we may be able to
3469 determine the result of this comparison. */
3470 if (const_arg1 == const0_rtx && !const_arg0)
3472 rtx y = lookup_as_function (folded_arg0, IOR);
3473 rtx inner_const;
3475 if (y != 0
3476 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3477 && CONST_INT_P (inner_const)
3478 && INTVAL (inner_const) != 0)
3479 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3483 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3484 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3485 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3486 op0, op1);
3488 break;
3490 case RTX_BIN_ARITH:
3491 case RTX_COMM_ARITH:
3492 switch (code)
3494 case PLUS:
3495 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3496 with that LABEL_REF as its second operand. If so, the result is
3497 the first operand of that MINUS. This handles switches with an
3498 ADDR_DIFF_VEC table. */
3499 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3501 rtx y
3502 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3503 : lookup_as_function (folded_arg0, MINUS);
3505 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3506 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3507 return XEXP (y, 0);
3509 /* Now try for a CONST of a MINUS like the above. */
3510 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3511 : lookup_as_function (folded_arg0, CONST))) != 0
3512 && GET_CODE (XEXP (y, 0)) == MINUS
3513 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3514 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3515 return XEXP (XEXP (y, 0), 0);
3518 /* Likewise if the operands are in the other order. */
3519 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3521 rtx y
3522 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3523 : lookup_as_function (folded_arg1, MINUS);
3525 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3526 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3527 return XEXP (y, 0);
3529 /* Now try for a CONST of a MINUS like the above. */
3530 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3531 : lookup_as_function (folded_arg1, CONST))) != 0
3532 && GET_CODE (XEXP (y, 0)) == MINUS
3533 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3534 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3535 return XEXP (XEXP (y, 0), 0);
3538 /* If second operand is a register equivalent to a negative
3539 CONST_INT, see if we can find a register equivalent to the
3540 positive constant. Make a MINUS if so. Don't do this for
3541 a non-negative constant since we might then alternate between
3542 choosing positive and negative constants. Having the positive
3543 constant previously-used is the more common case. Be sure
3544 the resulting constant is non-negative; if const_arg1 were
3545 the smallest negative number this would overflow: depending
3546 on the mode, this would either just be the same value (and
3547 hence not save anything) or be incorrect. */
3548 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3549 && INTVAL (const_arg1) < 0
3550 /* This used to test
3552 -INTVAL (const_arg1) >= 0
3554 But The Sun V5.0 compilers mis-compiled that test. So
3555 instead we test for the problematic value in a more direct
3556 manner and hope the Sun compilers get it correct. */
3557 && INTVAL (const_arg1) !=
3558 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3559 && REG_P (folded_arg1))
3561 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3562 struct table_elt *p
3563 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3565 if (p)
3566 for (p = p->first_same_value; p; p = p->next_same_value)
3567 if (REG_P (p->exp))
3568 return simplify_gen_binary (MINUS, mode, folded_arg0,
3569 canon_reg (p->exp, NULL));
3571 goto from_plus;
3573 case MINUS:
3574 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3575 If so, produce (PLUS Z C2-C). */
3576 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3578 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3579 if (y && CONST_INT_P (XEXP (y, 1)))
3580 return fold_rtx (plus_constant (mode, copy_rtx (y),
3581 -INTVAL (const_arg1)),
3582 NULL);
3585 /* Fall through. */
3587 from_plus:
3588 case SMIN: case SMAX: case UMIN: case UMAX:
3589 case IOR: case AND: case XOR:
3590 case MULT:
3591 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3592 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3593 is known to be of similar form, we may be able to replace the
3594 operation with a combined operation. This may eliminate the
3595 intermediate operation if every use is simplified in this way.
3596 Note that the similar optimization done by combine.c only works
3597 if the intermediate operation's result has only one reference. */
3599 if (REG_P (folded_arg0)
3600 && const_arg1 && CONST_INT_P (const_arg1))
3602 int is_shift
3603 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3604 rtx y, inner_const, new_const;
3605 rtx canon_const_arg1 = const_arg1;
3606 enum rtx_code associate_code;
3608 if (is_shift
3609 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3610 || INTVAL (const_arg1) < 0))
3612 if (SHIFT_COUNT_TRUNCATED)
3613 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3614 & (GET_MODE_BITSIZE (mode)
3615 - 1));
3616 else
3617 break;
3620 y = lookup_as_function (folded_arg0, code);
3621 if (y == 0)
3622 break;
3624 /* If we have compiled a statement like
3625 "if (x == (x & mask1))", and now are looking at
3626 "x & mask2", we will have a case where the first operand
3627 of Y is the same as our first operand. Unless we detect
3628 this case, an infinite loop will result. */
3629 if (XEXP (y, 0) == folded_arg0)
3630 break;
3632 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3633 if (!inner_const || !CONST_INT_P (inner_const))
3634 break;
3636 /* Don't associate these operations if they are a PLUS with the
3637 same constant and it is a power of two. These might be doable
3638 with a pre- or post-increment. Similarly for two subtracts of
3639 identical powers of two with post decrement. */
3641 if (code == PLUS && const_arg1 == inner_const
3642 && ((HAVE_PRE_INCREMENT
3643 && pow2p_hwi (INTVAL (const_arg1)))
3644 || (HAVE_POST_INCREMENT
3645 && pow2p_hwi (INTVAL (const_arg1)))
3646 || (HAVE_PRE_DECREMENT
3647 && pow2p_hwi (- INTVAL (const_arg1)))
3648 || (HAVE_POST_DECREMENT
3649 && pow2p_hwi (- INTVAL (const_arg1)))))
3650 break;
3652 /* ??? Vector mode shifts by scalar
3653 shift operand are not supported yet. */
3654 if (is_shift && VECTOR_MODE_P (mode))
3655 break;
3657 if (is_shift
3658 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3659 || INTVAL (inner_const) < 0))
3661 if (SHIFT_COUNT_TRUNCATED)
3662 inner_const = GEN_INT (INTVAL (inner_const)
3663 & (GET_MODE_BITSIZE (mode) - 1));
3664 else
3665 break;
3668 /* Compute the code used to compose the constants. For example,
3669 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3671 associate_code = (is_shift || code == MINUS ? PLUS : code);
3673 new_const = simplify_binary_operation (associate_code, mode,
3674 canon_const_arg1,
3675 inner_const);
3677 if (new_const == 0)
3678 break;
3680 /* If we are associating shift operations, don't let this
3681 produce a shift of the size of the object or larger.
3682 This could occur when we follow a sign-extend by a right
3683 shift on a machine that does a sign-extend as a pair
3684 of shifts. */
3686 if (is_shift
3687 && CONST_INT_P (new_const)
3688 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3690 /* As an exception, we can turn an ASHIFTRT of this
3691 form into a shift of the number of bits - 1. */
3692 if (code == ASHIFTRT)
3693 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3694 else if (!side_effects_p (XEXP (y, 0)))
3695 return CONST0_RTX (mode);
3696 else
3697 break;
3700 y = copy_rtx (XEXP (y, 0));
3702 /* If Y contains our first operand (the most common way this
3703 can happen is if Y is a MEM), we would do into an infinite
3704 loop if we tried to fold it. So don't in that case. */
3706 if (! reg_mentioned_p (folded_arg0, y))
3707 y = fold_rtx (y, insn);
3709 return simplify_gen_binary (code, mode, y, new_const);
3711 break;
3713 case DIV: case UDIV:
3714 /* ??? The associative optimization performed immediately above is
3715 also possible for DIV and UDIV using associate_code of MULT.
3716 However, we would need extra code to verify that the
3717 multiplication does not overflow, that is, there is no overflow
3718 in the calculation of new_const. */
3719 break;
3721 default:
3722 break;
3725 new_rtx = simplify_binary_operation (code, mode,
3726 const_arg0 ? const_arg0 : folded_arg0,
3727 const_arg1 ? const_arg1 : folded_arg1);
3728 break;
3730 case RTX_OBJ:
3731 /* (lo_sum (high X) X) is simply X. */
3732 if (code == LO_SUM && const_arg0 != 0
3733 && GET_CODE (const_arg0) == HIGH
3734 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3735 return const_arg1;
3736 break;
3738 case RTX_TERNARY:
3739 case RTX_BITFIELD_OPS:
3740 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3741 const_arg0 ? const_arg0 : folded_arg0,
3742 const_arg1 ? const_arg1 : folded_arg1,
3743 const_arg2 ? const_arg2 : XEXP (x, 2));
3744 break;
3746 default:
3747 break;
3750 return new_rtx ? new_rtx : x;
3753 /* Return a constant value currently equivalent to X.
3754 Return 0 if we don't know one. */
3756 static rtx
3757 equiv_constant (rtx x)
3759 if (REG_P (x)
3760 && REGNO_QTY_VALID_P (REGNO (x)))
3762 int x_q = REG_QTY (REGNO (x));
3763 struct qty_table_elem *x_ent = &qty_table[x_q];
3765 if (x_ent->const_rtx)
3766 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3769 if (x == 0 || CONSTANT_P (x))
3770 return x;
3772 if (GET_CODE (x) == SUBREG)
3774 machine_mode mode = GET_MODE (x);
3775 machine_mode imode = GET_MODE (SUBREG_REG (x));
3776 rtx new_rtx;
3778 /* See if we previously assigned a constant value to this SUBREG. */
3779 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3780 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3781 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3782 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3783 return new_rtx;
3785 /* If we didn't and if doing so makes sense, see if we previously
3786 assigned a constant value to the enclosing word mode SUBREG. */
3787 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3788 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3790 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3791 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3793 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3794 new_rtx = lookup_as_function (y, CONST_INT);
3795 if (new_rtx)
3796 return gen_lowpart (mode, new_rtx);
3800 /* Otherwise see if we already have a constant for the inner REG,
3801 and if that is enough to calculate an equivalent constant for
3802 the subreg. Note that the upper bits of paradoxical subregs
3803 are undefined, so they cannot be said to equal anything. */
3804 if (REG_P (SUBREG_REG (x))
3805 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3806 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3807 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3809 return 0;
3812 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3813 the hash table in case its value was seen before. */
3815 if (MEM_P (x))
3817 struct table_elt *elt;
3819 x = avoid_constant_pool_reference (x);
3820 if (CONSTANT_P (x))
3821 return x;
3823 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3824 if (elt == 0)
3825 return 0;
3827 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3828 if (elt->is_const && CONSTANT_P (elt->exp))
3829 return elt->exp;
3832 return 0;
3835 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3836 "taken" branch.
3838 In certain cases, this can cause us to add an equivalence. For example,
3839 if we are following the taken case of
3840 if (i == 2)
3841 we can add the fact that `i' and '2' are now equivalent.
3843 In any case, we can record that this comparison was passed. If the same
3844 comparison is seen later, we will know its value. */
3846 static void
3847 record_jump_equiv (rtx_insn *insn, bool taken)
3849 int cond_known_true;
3850 rtx op0, op1;
3851 rtx set;
3852 machine_mode mode, mode0, mode1;
3853 int reversed_nonequality = 0;
3854 enum rtx_code code;
3856 /* Ensure this is the right kind of insn. */
3857 gcc_assert (any_condjump_p (insn));
3859 set = pc_set (insn);
3861 /* See if this jump condition is known true or false. */
3862 if (taken)
3863 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3864 else
3865 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3867 /* Get the type of comparison being done and the operands being compared.
3868 If we had to reverse a non-equality condition, record that fact so we
3869 know that it isn't valid for floating-point. */
3870 code = GET_CODE (XEXP (SET_SRC (set), 0));
3871 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3872 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3874 /* On a cc0 target the cc0-setter and cc0-user may end up in different
3875 blocks. When that happens the tracking of the cc0-setter via
3876 PREV_INSN_CC0 is spoiled. That means that fold_rtx may return
3877 NULL_RTX. In those cases, there's nothing to record. */
3878 if (op0 == NULL_RTX || op1 == NULL_RTX)
3879 return;
3881 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3882 if (! cond_known_true)
3884 code = reversed_comparison_code_parts (code, op0, op1, insn);
3886 /* Don't remember if we can't find the inverse. */
3887 if (code == UNKNOWN)
3888 return;
3891 /* The mode is the mode of the non-constant. */
3892 mode = mode0;
3893 if (mode1 != VOIDmode)
3894 mode = mode1;
3896 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3899 /* Yet another form of subreg creation. In this case, we want something in
3900 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3902 static rtx
3903 record_jump_cond_subreg (machine_mode mode, rtx op)
3905 machine_mode op_mode = GET_MODE (op);
3906 if (op_mode == mode || op_mode == VOIDmode)
3907 return op;
3908 return lowpart_subreg (mode, op, op_mode);
3911 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3912 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3913 Make any useful entries we can with that information. Called from
3914 above function and called recursively. */
3916 static void
3917 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3918 rtx op1, int reversed_nonequality)
3920 unsigned op0_hash, op1_hash;
3921 int op0_in_memory, op1_in_memory;
3922 struct table_elt *op0_elt, *op1_elt;
3924 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3925 we know that they are also equal in the smaller mode (this is also
3926 true for all smaller modes whether or not there is a SUBREG, but
3927 is not worth testing for with no SUBREG). */
3929 /* Note that GET_MODE (op0) may not equal MODE. */
3930 if (code == EQ && paradoxical_subreg_p (op0))
3932 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3933 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3934 if (tem)
3935 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3936 reversed_nonequality);
3939 if (code == EQ && paradoxical_subreg_p (op1))
3941 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3942 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3943 if (tem)
3944 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3945 reversed_nonequality);
3948 /* Similarly, if this is an NE comparison, and either is a SUBREG
3949 making a smaller mode, we know the whole thing is also NE. */
3951 /* Note that GET_MODE (op0) may not equal MODE;
3952 if we test MODE instead, we can get an infinite recursion
3953 alternating between two modes each wider than MODE. */
3955 if (code == NE && GET_CODE (op0) == SUBREG
3956 && subreg_lowpart_p (op0)
3957 && (GET_MODE_SIZE (GET_MODE (op0))
3958 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3960 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3961 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3962 if (tem)
3963 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3964 reversed_nonequality);
3967 if (code == NE && GET_CODE (op1) == SUBREG
3968 && subreg_lowpart_p (op1)
3969 && (GET_MODE_SIZE (GET_MODE (op1))
3970 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3972 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3973 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3974 if (tem)
3975 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3976 reversed_nonequality);
3979 /* Hash both operands. */
3981 do_not_record = 0;
3982 hash_arg_in_memory = 0;
3983 op0_hash = HASH (op0, mode);
3984 op0_in_memory = hash_arg_in_memory;
3986 if (do_not_record)
3987 return;
3989 do_not_record = 0;
3990 hash_arg_in_memory = 0;
3991 op1_hash = HASH (op1, mode);
3992 op1_in_memory = hash_arg_in_memory;
3994 if (do_not_record)
3995 return;
3997 /* Look up both operands. */
3998 op0_elt = lookup (op0, op0_hash, mode);
3999 op1_elt = lookup (op1, op1_hash, mode);
4001 /* If both operands are already equivalent or if they are not in the
4002 table but are identical, do nothing. */
4003 if ((op0_elt != 0 && op1_elt != 0
4004 && op0_elt->first_same_value == op1_elt->first_same_value)
4005 || op0 == op1 || rtx_equal_p (op0, op1))
4006 return;
4008 /* If we aren't setting two things equal all we can do is save this
4009 comparison. Similarly if this is floating-point. In the latter
4010 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4011 If we record the equality, we might inadvertently delete code
4012 whose intent was to change -0 to +0. */
4014 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4016 struct qty_table_elem *ent;
4017 int qty;
4019 /* If we reversed a floating-point comparison, if OP0 is not a
4020 register, or if OP1 is neither a register or constant, we can't
4021 do anything. */
4023 if (!REG_P (op1))
4024 op1 = equiv_constant (op1);
4026 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4027 || !REG_P (op0) || op1 == 0)
4028 return;
4030 /* Put OP0 in the hash table if it isn't already. This gives it a
4031 new quantity number. */
4032 if (op0_elt == 0)
4034 if (insert_regs (op0, NULL, 0))
4036 rehash_using_reg (op0);
4037 op0_hash = HASH (op0, mode);
4039 /* If OP0 is contained in OP1, this changes its hash code
4040 as well. Faster to rehash than to check, except
4041 for the simple case of a constant. */
4042 if (! CONSTANT_P (op1))
4043 op1_hash = HASH (op1,mode);
4046 op0_elt = insert (op0, NULL, op0_hash, mode);
4047 op0_elt->in_memory = op0_in_memory;
4050 qty = REG_QTY (REGNO (op0));
4051 ent = &qty_table[qty];
4053 ent->comparison_code = code;
4054 if (REG_P (op1))
4056 /* Look it up again--in case op0 and op1 are the same. */
4057 op1_elt = lookup (op1, op1_hash, mode);
4059 /* Put OP1 in the hash table so it gets a new quantity number. */
4060 if (op1_elt == 0)
4062 if (insert_regs (op1, NULL, 0))
4064 rehash_using_reg (op1);
4065 op1_hash = HASH (op1, mode);
4068 op1_elt = insert (op1, NULL, op1_hash, mode);
4069 op1_elt->in_memory = op1_in_memory;
4072 ent->comparison_const = NULL_RTX;
4073 ent->comparison_qty = REG_QTY (REGNO (op1));
4075 else
4077 ent->comparison_const = op1;
4078 ent->comparison_qty = -1;
4081 return;
4084 /* If either side is still missing an equivalence, make it now,
4085 then merge the equivalences. */
4087 if (op0_elt == 0)
4089 if (insert_regs (op0, NULL, 0))
4091 rehash_using_reg (op0);
4092 op0_hash = HASH (op0, mode);
4095 op0_elt = insert (op0, NULL, op0_hash, mode);
4096 op0_elt->in_memory = op0_in_memory;
4099 if (op1_elt == 0)
4101 if (insert_regs (op1, NULL, 0))
4103 rehash_using_reg (op1);
4104 op1_hash = HASH (op1, mode);
4107 op1_elt = insert (op1, NULL, op1_hash, mode);
4108 op1_elt->in_memory = op1_in_memory;
4111 merge_equiv_classes (op0_elt, op1_elt);
4114 /* CSE processing for one instruction.
4116 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4117 but the few that "leak through" are cleaned up by cse_insn, and complex
4118 addressing modes are often formed here.
4120 The main function is cse_insn, and between here and that function
4121 a couple of helper functions is defined to keep the size of cse_insn
4122 within reasonable proportions.
4124 Data is shared between the main and helper functions via STRUCT SET,
4125 that contains all data related for every set in the instruction that
4126 is being processed.
4128 Note that cse_main processes all sets in the instruction. Most
4129 passes in GCC only process simple SET insns or single_set insns, but
4130 CSE processes insns with multiple sets as well. */
4132 /* Data on one SET contained in the instruction. */
4134 struct set
4136 /* The SET rtx itself. */
4137 rtx rtl;
4138 /* The SET_SRC of the rtx (the original value, if it is changing). */
4139 rtx src;
4140 /* The hash-table element for the SET_SRC of the SET. */
4141 struct table_elt *src_elt;
4142 /* Hash value for the SET_SRC. */
4143 unsigned src_hash;
4144 /* Hash value for the SET_DEST. */
4145 unsigned dest_hash;
4146 /* The SET_DEST, with SUBREG, etc., stripped. */
4147 rtx inner_dest;
4148 /* Nonzero if the SET_SRC is in memory. */
4149 char src_in_memory;
4150 /* Nonzero if the SET_SRC contains something
4151 whose value cannot be predicted and understood. */
4152 char src_volatile;
4153 /* Original machine mode, in case it becomes a CONST_INT.
4154 The size of this field should match the size of the mode
4155 field of struct rtx_def (see rtl.h). */
4156 ENUM_BITFIELD(machine_mode) mode : 8;
4157 /* A constant equivalent for SET_SRC, if any. */
4158 rtx src_const;
4159 /* Hash value of constant equivalent for SET_SRC. */
4160 unsigned src_const_hash;
4161 /* Table entry for constant equivalent for SET_SRC, if any. */
4162 struct table_elt *src_const_elt;
4163 /* Table entry for the destination address. */
4164 struct table_elt *dest_addr_elt;
4167 /* Special handling for (set REG0 REG1) where REG0 is the
4168 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4169 be used in the sequel, so (if easily done) change this insn to
4170 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4171 that computed their value. Then REG1 will become a dead store
4172 and won't cloud the situation for later optimizations.
4174 Do not make this change if REG1 is a hard register, because it will
4175 then be used in the sequel and we may be changing a two-operand insn
4176 into a three-operand insn.
4178 This is the last transformation that cse_insn will try to do. */
4180 static void
4181 try_back_substitute_reg (rtx set, rtx_insn *insn)
4183 rtx dest = SET_DEST (set);
4184 rtx src = SET_SRC (set);
4186 if (REG_P (dest)
4187 && REG_P (src) && ! HARD_REGISTER_P (src)
4188 && REGNO_QTY_VALID_P (REGNO (src)))
4190 int src_q = REG_QTY (REGNO (src));
4191 struct qty_table_elem *src_ent = &qty_table[src_q];
4193 if (src_ent->first_reg == REGNO (dest))
4195 /* Scan for the previous nonnote insn, but stop at a basic
4196 block boundary. */
4197 rtx_insn *prev = insn;
4198 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4201 prev = PREV_INSN (prev);
4203 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4205 /* Do not swap the registers around if the previous instruction
4206 attaches a REG_EQUIV note to REG1.
4208 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4209 from the pseudo that originally shadowed an incoming argument
4210 to another register. Some uses of REG_EQUIV might rely on it
4211 being attached to REG1 rather than REG2.
4213 This section previously turned the REG_EQUIV into a REG_EQUAL
4214 note. We cannot do that because REG_EQUIV may provide an
4215 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4216 if (NONJUMP_INSN_P (prev)
4217 && GET_CODE (PATTERN (prev)) == SET
4218 && SET_DEST (PATTERN (prev)) == src
4219 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4221 rtx note;
4223 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4224 validate_change (insn, &SET_DEST (set), src, 1);
4225 validate_change (insn, &SET_SRC (set), dest, 1);
4226 apply_change_group ();
4228 /* If INSN has a REG_EQUAL note, and this note mentions
4229 REG0, then we must delete it, because the value in
4230 REG0 has changed. If the note's value is REG1, we must
4231 also delete it because that is now this insn's dest. */
4232 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4233 if (note != 0
4234 && (reg_mentioned_p (dest, XEXP (note, 0))
4235 || rtx_equal_p (src, XEXP (note, 0))))
4236 remove_note (insn, note);
4242 /* Record all the SETs in this instruction into SETS_PTR,
4243 and return the number of recorded sets. */
4244 static int
4245 find_sets_in_insn (rtx_insn *insn, struct set **psets)
4247 struct set *sets = *psets;
4248 int n_sets = 0;
4249 rtx x = PATTERN (insn);
4251 if (GET_CODE (x) == SET)
4253 /* Ignore SETs that are unconditional jumps.
4254 They never need cse processing, so this does not hurt.
4255 The reason is not efficiency but rather
4256 so that we can test at the end for instructions
4257 that have been simplified to unconditional jumps
4258 and not be misled by unchanged instructions
4259 that were unconditional jumps to begin with. */
4260 if (SET_DEST (x) == pc_rtx
4261 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4263 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4264 The hard function value register is used only once, to copy to
4265 someplace else, so it isn't worth cse'ing. */
4266 else if (GET_CODE (SET_SRC (x)) == CALL)
4268 else
4269 sets[n_sets++].rtl = x;
4271 else if (GET_CODE (x) == PARALLEL)
4273 int i, lim = XVECLEN (x, 0);
4275 /* Go over the expressions of the PARALLEL in forward order, to
4276 put them in the same order in the SETS array. */
4277 for (i = 0; i < lim; i++)
4279 rtx y = XVECEXP (x, 0, i);
4280 if (GET_CODE (y) == SET)
4282 /* As above, we ignore unconditional jumps and call-insns and
4283 ignore the result of apply_change_group. */
4284 if (SET_DEST (y) == pc_rtx
4285 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4287 else if (GET_CODE (SET_SRC (y)) == CALL)
4289 else
4290 sets[n_sets++].rtl = y;
4295 return n_sets;
4298 /* Subroutine of canonicalize_insn. X is an ASM_OPERANDS in INSN. */
4300 static void
4301 canon_asm_operands (rtx x, rtx_insn *insn)
4303 for (int i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4305 rtx input = ASM_OPERANDS_INPUT (x, i);
4306 if (!(REG_P (input) && HARD_REGISTER_P (input)))
4308 input = canon_reg (input, insn);
4309 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4314 /* Where possible, substitute every register reference in the N_SETS
4315 number of SETS in INSN with the canonical register.
4317 Register canonicalization propagatest the earliest register (i.e.
4318 one that is set before INSN) with the same value. This is a very
4319 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4320 to RTL. For instance, a CONST for an address is usually expanded
4321 multiple times to loads into different registers, thus creating many
4322 subexpressions of the form:
4324 (set (reg1) (some_const))
4325 (set (mem (... reg1 ...) (thing)))
4326 (set (reg2) (some_const))
4327 (set (mem (... reg2 ...) (thing)))
4329 After canonicalizing, the code takes the following form:
4331 (set (reg1) (some_const))
4332 (set (mem (... reg1 ...) (thing)))
4333 (set (reg2) (some_const))
4334 (set (mem (... reg1 ...) (thing)))
4336 The set to reg2 is now trivially dead, and the memory reference (or
4337 address, or whatever) may be a candidate for further CSEing.
4339 In this function, the result of apply_change_group can be ignored;
4340 see canon_reg. */
4342 static void
4343 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4345 struct set *sets = *psets;
4346 rtx tem;
4347 rtx x = PATTERN (insn);
4348 int i;
4350 if (CALL_P (insn))
4352 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4353 if (GET_CODE (XEXP (tem, 0)) != SET)
4354 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4357 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4359 canon_reg (SET_SRC (x), insn);
4360 apply_change_group ();
4361 fold_rtx (SET_SRC (x), insn);
4363 else if (GET_CODE (x) == CLOBBER)
4365 /* If we clobber memory, canon the address.
4366 This does nothing when a register is clobbered
4367 because we have already invalidated the reg. */
4368 if (MEM_P (XEXP (x, 0)))
4369 canon_reg (XEXP (x, 0), insn);
4371 else if (GET_CODE (x) == USE
4372 && ! (REG_P (XEXP (x, 0))
4373 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4374 /* Canonicalize a USE of a pseudo register or memory location. */
4375 canon_reg (x, insn);
4376 else if (GET_CODE (x) == ASM_OPERANDS)
4377 canon_asm_operands (x, insn);
4378 else if (GET_CODE (x) == CALL)
4380 canon_reg (x, insn);
4381 apply_change_group ();
4382 fold_rtx (x, insn);
4384 else if (DEBUG_INSN_P (insn))
4385 canon_reg (PATTERN (insn), insn);
4386 else if (GET_CODE (x) == PARALLEL)
4388 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4390 rtx y = XVECEXP (x, 0, i);
4391 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4393 canon_reg (SET_SRC (y), insn);
4394 apply_change_group ();
4395 fold_rtx (SET_SRC (y), insn);
4397 else if (GET_CODE (y) == CLOBBER)
4399 if (MEM_P (XEXP (y, 0)))
4400 canon_reg (XEXP (y, 0), insn);
4402 else if (GET_CODE (y) == USE
4403 && ! (REG_P (XEXP (y, 0))
4404 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4405 canon_reg (y, insn);
4406 else if (GET_CODE (y) == ASM_OPERANDS)
4407 canon_asm_operands (y, insn);
4408 else if (GET_CODE (y) == CALL)
4410 canon_reg (y, insn);
4411 apply_change_group ();
4412 fold_rtx (y, insn);
4417 if (n_sets == 1 && REG_NOTES (insn) != 0
4418 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4420 /* We potentially will process this insn many times. Therefore,
4421 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4422 unique set in INSN.
4424 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4425 because cse_insn handles those specially. */
4426 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4427 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4428 remove_note (insn, tem);
4429 else
4431 canon_reg (XEXP (tem, 0), insn);
4432 apply_change_group ();
4433 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4434 df_notes_rescan (insn);
4438 /* Canonicalize sources and addresses of destinations.
4439 We do this in a separate pass to avoid problems when a MATCH_DUP is
4440 present in the insn pattern. In that case, we want to ensure that
4441 we don't break the duplicate nature of the pattern. So we will replace
4442 both operands at the same time. Otherwise, we would fail to find an
4443 equivalent substitution in the loop calling validate_change below.
4445 We used to suppress canonicalization of DEST if it appears in SRC,
4446 but we don't do this any more. */
4448 for (i = 0; i < n_sets; i++)
4450 rtx dest = SET_DEST (sets[i].rtl);
4451 rtx src = SET_SRC (sets[i].rtl);
4452 rtx new_rtx = canon_reg (src, insn);
4454 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4456 if (GET_CODE (dest) == ZERO_EXTRACT)
4458 validate_change (insn, &XEXP (dest, 1),
4459 canon_reg (XEXP (dest, 1), insn), 1);
4460 validate_change (insn, &XEXP (dest, 2),
4461 canon_reg (XEXP (dest, 2), insn), 1);
4464 while (GET_CODE (dest) == SUBREG
4465 || GET_CODE (dest) == ZERO_EXTRACT
4466 || GET_CODE (dest) == STRICT_LOW_PART)
4467 dest = XEXP (dest, 0);
4469 if (MEM_P (dest))
4470 canon_reg (dest, insn);
4473 /* Now that we have done all the replacements, we can apply the change
4474 group and see if they all work. Note that this will cause some
4475 canonicalizations that would have worked individually not to be applied
4476 because some other canonicalization didn't work, but this should not
4477 occur often.
4479 The result of apply_change_group can be ignored; see canon_reg. */
4481 apply_change_group ();
4484 /* Main function of CSE.
4485 First simplify sources and addresses of all assignments
4486 in the instruction, using previously-computed equivalents values.
4487 Then install the new sources and destinations in the table
4488 of available values. */
4490 static void
4491 cse_insn (rtx_insn *insn)
4493 rtx x = PATTERN (insn);
4494 int i;
4495 rtx tem;
4496 int n_sets = 0;
4498 rtx src_eqv = 0;
4499 struct table_elt *src_eqv_elt = 0;
4500 int src_eqv_volatile = 0;
4501 int src_eqv_in_memory = 0;
4502 unsigned src_eqv_hash = 0;
4504 struct set *sets = (struct set *) 0;
4506 if (GET_CODE (x) == SET)
4507 sets = XALLOCA (struct set);
4508 else if (GET_CODE (x) == PARALLEL)
4509 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4511 this_insn = insn;
4512 /* Records what this insn does to set CC0. */
4513 this_insn_cc0 = 0;
4514 this_insn_cc0_mode = VOIDmode;
4516 /* Find all regs explicitly clobbered in this insn,
4517 to ensure they are not replaced with any other regs
4518 elsewhere in this insn. */
4519 invalidate_from_sets_and_clobbers (insn);
4521 /* Record all the SETs in this instruction. */
4522 n_sets = find_sets_in_insn (insn, &sets);
4524 /* Substitute the canonical register where possible. */
4525 canonicalize_insn (insn, &sets, n_sets);
4527 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4528 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4529 latter condition is necessary because SRC_EQV is handled specially for
4530 this case, and if it isn't set, then there will be no equivalence
4531 for the destination. */
4532 if (n_sets == 1 && REG_NOTES (insn) != 0
4533 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4536 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4537 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4538 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4539 src_eqv = copy_rtx (XEXP (tem, 0));
4540 /* If DEST is of the form ZERO_EXTACT, as in:
4541 (set (zero_extract:SI (reg:SI 119)
4542 (const_int 16 [0x10])
4543 (const_int 16 [0x10]))
4544 (const_int 51154 [0xc7d2]))
4545 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4546 point. Note that this is different from SRC_EQV. We can however
4547 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4548 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4549 && CONST_INT_P (XEXP (tem, 0))
4550 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4551 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4553 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4554 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4555 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4556 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4557 HOST_WIDE_INT mask;
4558 unsigned int shift;
4559 if (BITS_BIG_ENDIAN)
4560 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
4561 - INTVAL (pos) - INTVAL (width);
4562 else
4563 shift = INTVAL (pos);
4564 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4565 mask = HOST_WIDE_INT_M1;
4566 else
4567 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4568 val = (val >> shift) & mask;
4569 src_eqv = GEN_INT (val);
4573 /* Set sets[i].src_elt to the class each source belongs to.
4574 Detect assignments from or to volatile things
4575 and set set[i] to zero so they will be ignored
4576 in the rest of this function.
4578 Nothing in this loop changes the hash table or the register chains. */
4580 for (i = 0; i < n_sets; i++)
4582 bool repeat = false;
4583 bool mem_noop_insn = false;
4584 rtx src, dest;
4585 rtx src_folded;
4586 struct table_elt *elt = 0, *p;
4587 machine_mode mode;
4588 rtx src_eqv_here;
4589 rtx src_const = 0;
4590 rtx src_related = 0;
4591 bool src_related_is_const_anchor = false;
4592 struct table_elt *src_const_elt = 0;
4593 int src_cost = MAX_COST;
4594 int src_eqv_cost = MAX_COST;
4595 int src_folded_cost = MAX_COST;
4596 int src_related_cost = MAX_COST;
4597 int src_elt_cost = MAX_COST;
4598 int src_regcost = MAX_COST;
4599 int src_eqv_regcost = MAX_COST;
4600 int src_folded_regcost = MAX_COST;
4601 int src_related_regcost = MAX_COST;
4602 int src_elt_regcost = MAX_COST;
4603 /* Set nonzero if we need to call force_const_mem on with the
4604 contents of src_folded before using it. */
4605 int src_folded_force_flag = 0;
4607 dest = SET_DEST (sets[i].rtl);
4608 src = SET_SRC (sets[i].rtl);
4610 /* If SRC is a constant that has no machine mode,
4611 hash it with the destination's machine mode.
4612 This way we can keep different modes separate. */
4614 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4615 sets[i].mode = mode;
4617 if (src_eqv)
4619 machine_mode eqvmode = mode;
4620 if (GET_CODE (dest) == STRICT_LOW_PART)
4621 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4622 do_not_record = 0;
4623 hash_arg_in_memory = 0;
4624 src_eqv_hash = HASH (src_eqv, eqvmode);
4626 /* Find the equivalence class for the equivalent expression. */
4628 if (!do_not_record)
4629 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4631 src_eqv_volatile = do_not_record;
4632 src_eqv_in_memory = hash_arg_in_memory;
4635 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4636 value of the INNER register, not the destination. So it is not
4637 a valid substitution for the source. But save it for later. */
4638 if (GET_CODE (dest) == STRICT_LOW_PART)
4639 src_eqv_here = 0;
4640 else
4641 src_eqv_here = src_eqv;
4643 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4644 simplified result, which may not necessarily be valid. */
4645 src_folded = fold_rtx (src, NULL);
4647 #if 0
4648 /* ??? This caused bad code to be generated for the m68k port with -O2.
4649 Suppose src is (CONST_INT -1), and that after truncation src_folded
4650 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4651 At the end we will add src and src_const to the same equivalence
4652 class. We now have 3 and -1 on the same equivalence class. This
4653 causes later instructions to be mis-optimized. */
4654 /* If storing a constant in a bitfield, pre-truncate the constant
4655 so we will be able to record it later. */
4656 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4658 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4660 if (CONST_INT_P (src)
4661 && CONST_INT_P (width)
4662 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4663 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4664 src_folded
4665 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4666 << INTVAL (width)) - 1));
4668 #endif
4670 /* Compute SRC's hash code, and also notice if it
4671 should not be recorded at all. In that case,
4672 prevent any further processing of this assignment. */
4673 do_not_record = 0;
4674 hash_arg_in_memory = 0;
4676 sets[i].src = src;
4677 sets[i].src_hash = HASH (src, mode);
4678 sets[i].src_volatile = do_not_record;
4679 sets[i].src_in_memory = hash_arg_in_memory;
4681 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4682 a pseudo, do not record SRC. Using SRC as a replacement for
4683 anything else will be incorrect in that situation. Note that
4684 this usually occurs only for stack slots, in which case all the
4685 RTL would be referring to SRC, so we don't lose any optimization
4686 opportunities by not having SRC in the hash table. */
4688 if (MEM_P (src)
4689 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4690 && REG_P (dest)
4691 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4692 sets[i].src_volatile = 1;
4694 else if (GET_CODE (src) == ASM_OPERANDS
4695 && GET_CODE (x) == PARALLEL)
4697 /* Do not record result of a non-volatile inline asm with
4698 more than one result. */
4699 if (n_sets > 1)
4700 sets[i].src_volatile = 1;
4702 int j, lim = XVECLEN (x, 0);
4703 for (j = 0; j < lim; j++)
4705 rtx y = XVECEXP (x, 0, j);
4706 /* And do not record result of a non-volatile inline asm
4707 with "memory" clobber. */
4708 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4710 sets[i].src_volatile = 1;
4711 break;
4716 #if 0
4717 /* It is no longer clear why we used to do this, but it doesn't
4718 appear to still be needed. So let's try without it since this
4719 code hurts cse'ing widened ops. */
4720 /* If source is a paradoxical subreg (such as QI treated as an SI),
4721 treat it as volatile. It may do the work of an SI in one context
4722 where the extra bits are not being used, but cannot replace an SI
4723 in general. */
4724 if (paradoxical_subreg_p (src))
4725 sets[i].src_volatile = 1;
4726 #endif
4728 /* Locate all possible equivalent forms for SRC. Try to replace
4729 SRC in the insn with each cheaper equivalent.
4731 We have the following types of equivalents: SRC itself, a folded
4732 version, a value given in a REG_EQUAL note, or a value related
4733 to a constant.
4735 Each of these equivalents may be part of an additional class
4736 of equivalents (if more than one is in the table, they must be in
4737 the same class; we check for this).
4739 If the source is volatile, we don't do any table lookups.
4741 We note any constant equivalent for possible later use in a
4742 REG_NOTE. */
4744 if (!sets[i].src_volatile)
4745 elt = lookup (src, sets[i].src_hash, mode);
4747 sets[i].src_elt = elt;
4749 if (elt && src_eqv_here && src_eqv_elt)
4751 if (elt->first_same_value != src_eqv_elt->first_same_value)
4753 /* The REG_EQUAL is indicating that two formerly distinct
4754 classes are now equivalent. So merge them. */
4755 merge_equiv_classes (elt, src_eqv_elt);
4756 src_eqv_hash = HASH (src_eqv, elt->mode);
4757 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4760 src_eqv_here = 0;
4763 else if (src_eqv_elt)
4764 elt = src_eqv_elt;
4766 /* Try to find a constant somewhere and record it in `src_const'.
4767 Record its table element, if any, in `src_const_elt'. Look in
4768 any known equivalences first. (If the constant is not in the
4769 table, also set `sets[i].src_const_hash'). */
4770 if (elt)
4771 for (p = elt->first_same_value; p; p = p->next_same_value)
4772 if (p->is_const)
4774 src_const = p->exp;
4775 src_const_elt = elt;
4776 break;
4779 if (src_const == 0
4780 && (CONSTANT_P (src_folded)
4781 /* Consider (minus (label_ref L1) (label_ref L2)) as
4782 "constant" here so we will record it. This allows us
4783 to fold switch statements when an ADDR_DIFF_VEC is used. */
4784 || (GET_CODE (src_folded) == MINUS
4785 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4786 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4787 src_const = src_folded, src_const_elt = elt;
4788 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4789 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4791 /* If we don't know if the constant is in the table, get its
4792 hash code and look it up. */
4793 if (src_const && src_const_elt == 0)
4795 sets[i].src_const_hash = HASH (src_const, mode);
4796 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4799 sets[i].src_const = src_const;
4800 sets[i].src_const_elt = src_const_elt;
4802 /* If the constant and our source are both in the table, mark them as
4803 equivalent. Otherwise, if a constant is in the table but the source
4804 isn't, set ELT to it. */
4805 if (src_const_elt && elt
4806 && src_const_elt->first_same_value != elt->first_same_value)
4807 merge_equiv_classes (elt, src_const_elt);
4808 else if (src_const_elt && elt == 0)
4809 elt = src_const_elt;
4811 /* See if there is a register linearly related to a constant
4812 equivalent of SRC. */
4813 if (src_const
4814 && (GET_CODE (src_const) == CONST
4815 || (src_const_elt && src_const_elt->related_value != 0)))
4817 src_related = use_related_value (src_const, src_const_elt);
4818 if (src_related)
4820 struct table_elt *src_related_elt
4821 = lookup (src_related, HASH (src_related, mode), mode);
4822 if (src_related_elt && elt)
4824 if (elt->first_same_value
4825 != src_related_elt->first_same_value)
4826 /* This can occur when we previously saw a CONST
4827 involving a SYMBOL_REF and then see the SYMBOL_REF
4828 twice. Merge the involved classes. */
4829 merge_equiv_classes (elt, src_related_elt);
4831 src_related = 0;
4832 src_related_elt = 0;
4834 else if (src_related_elt && elt == 0)
4835 elt = src_related_elt;
4839 /* See if we have a CONST_INT that is already in a register in a
4840 wider mode. */
4842 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4843 && GET_MODE_CLASS (mode) == MODE_INT
4844 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4846 machine_mode wider_mode;
4848 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4849 wider_mode != VOIDmode
4850 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4851 && src_related == 0;
4852 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4854 struct table_elt *const_elt
4855 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4857 if (const_elt == 0)
4858 continue;
4860 for (const_elt = const_elt->first_same_value;
4861 const_elt; const_elt = const_elt->next_same_value)
4862 if (REG_P (const_elt->exp))
4864 src_related = gen_lowpart (mode, const_elt->exp);
4865 break;
4870 /* Another possibility is that we have an AND with a constant in
4871 a mode narrower than a word. If so, it might have been generated
4872 as part of an "if" which would narrow the AND. If we already
4873 have done the AND in a wider mode, we can use a SUBREG of that
4874 value. */
4876 if (flag_expensive_optimizations && ! src_related
4877 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4878 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4880 machine_mode tmode;
4881 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4883 for (tmode = GET_MODE_WIDER_MODE (mode);
4884 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4885 tmode = GET_MODE_WIDER_MODE (tmode))
4887 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4888 struct table_elt *larger_elt;
4890 if (inner)
4892 PUT_MODE (new_and, tmode);
4893 XEXP (new_and, 0) = inner;
4894 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4895 if (larger_elt == 0)
4896 continue;
4898 for (larger_elt = larger_elt->first_same_value;
4899 larger_elt; larger_elt = larger_elt->next_same_value)
4900 if (REG_P (larger_elt->exp))
4902 src_related
4903 = gen_lowpart (mode, larger_elt->exp);
4904 break;
4907 if (src_related)
4908 break;
4913 /* See if a MEM has already been loaded with a widening operation;
4914 if it has, we can use a subreg of that. Many CISC machines
4915 also have such operations, but this is only likely to be
4916 beneficial on these machines. */
4918 rtx_code extend_op;
4919 if (flag_expensive_optimizations && src_related == 0
4920 && MEM_P (src) && ! do_not_record
4921 && (extend_op = load_extend_op (mode)) != UNKNOWN)
4923 struct rtx_def memory_extend_buf;
4924 rtx memory_extend_rtx = &memory_extend_buf;
4925 machine_mode tmode;
4927 /* Set what we are trying to extend and the operation it might
4928 have been extended with. */
4929 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4930 PUT_CODE (memory_extend_rtx, extend_op);
4931 XEXP (memory_extend_rtx, 0) = src;
4933 for (tmode = GET_MODE_WIDER_MODE (mode);
4934 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4935 tmode = GET_MODE_WIDER_MODE (tmode))
4937 struct table_elt *larger_elt;
4939 PUT_MODE (memory_extend_rtx, tmode);
4940 larger_elt = lookup (memory_extend_rtx,
4941 HASH (memory_extend_rtx, tmode), tmode);
4942 if (larger_elt == 0)
4943 continue;
4945 for (larger_elt = larger_elt->first_same_value;
4946 larger_elt; larger_elt = larger_elt->next_same_value)
4947 if (REG_P (larger_elt->exp))
4949 src_related = gen_lowpart (mode, larger_elt->exp);
4950 break;
4953 if (src_related)
4954 break;
4958 /* Try to express the constant using a register+offset expression
4959 derived from a constant anchor. */
4961 if (targetm.const_anchor
4962 && !src_related
4963 && src_const
4964 && GET_CODE (src_const) == CONST_INT)
4966 src_related = try_const_anchors (src_const, mode);
4967 src_related_is_const_anchor = src_related != NULL_RTX;
4971 if (src == src_folded)
4972 src_folded = 0;
4974 /* At this point, ELT, if nonzero, points to a class of expressions
4975 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4976 and SRC_RELATED, if nonzero, each contain additional equivalent
4977 expressions. Prune these latter expressions by deleting expressions
4978 already in the equivalence class.
4980 Check for an equivalent identical to the destination. If found,
4981 this is the preferred equivalent since it will likely lead to
4982 elimination of the insn. Indicate this by placing it in
4983 `src_related'. */
4985 if (elt)
4986 elt = elt->first_same_value;
4987 for (p = elt; p; p = p->next_same_value)
4989 enum rtx_code code = GET_CODE (p->exp);
4991 /* If the expression is not valid, ignore it. Then we do not
4992 have to check for validity below. In most cases, we can use
4993 `rtx_equal_p', since canonicalization has already been done. */
4994 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4995 continue;
4997 /* Also skip paradoxical subregs, unless that's what we're
4998 looking for. */
4999 if (paradoxical_subreg_p (p->exp)
5000 && ! (src != 0
5001 && GET_CODE (src) == SUBREG
5002 && GET_MODE (src) == GET_MODE (p->exp)
5003 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5004 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5005 continue;
5007 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5008 src = 0;
5009 else if (src_folded && GET_CODE (src_folded) == code
5010 && rtx_equal_p (src_folded, p->exp))
5011 src_folded = 0;
5012 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5013 && rtx_equal_p (src_eqv_here, p->exp))
5014 src_eqv_here = 0;
5015 else if (src_related && GET_CODE (src_related) == code
5016 && rtx_equal_p (src_related, p->exp))
5017 src_related = 0;
5019 /* This is the same as the destination of the insns, we want
5020 to prefer it. Copy it to src_related. The code below will
5021 then give it a negative cost. */
5022 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5023 src_related = dest;
5026 /* Find the cheapest valid equivalent, trying all the available
5027 possibilities. Prefer items not in the hash table to ones
5028 that are when they are equal cost. Note that we can never
5029 worsen an insn as the current contents will also succeed.
5030 If we find an equivalent identical to the destination, use it as best,
5031 since this insn will probably be eliminated in that case. */
5032 if (src)
5034 if (rtx_equal_p (src, dest))
5035 src_cost = src_regcost = -1;
5036 else
5038 src_cost = COST (src, mode);
5039 src_regcost = approx_reg_cost (src);
5043 if (src_eqv_here)
5045 if (rtx_equal_p (src_eqv_here, dest))
5046 src_eqv_cost = src_eqv_regcost = -1;
5047 else
5049 src_eqv_cost = COST (src_eqv_here, mode);
5050 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5054 if (src_folded)
5056 if (rtx_equal_p (src_folded, dest))
5057 src_folded_cost = src_folded_regcost = -1;
5058 else
5060 src_folded_cost = COST (src_folded, mode);
5061 src_folded_regcost = approx_reg_cost (src_folded);
5065 if (src_related)
5067 if (rtx_equal_p (src_related, dest))
5068 src_related_cost = src_related_regcost = -1;
5069 else
5071 src_related_cost = COST (src_related, mode);
5072 src_related_regcost = approx_reg_cost (src_related);
5074 /* If a const-anchor is used to synthesize a constant that
5075 normally requires multiple instructions then slightly prefer
5076 it over the original sequence. These instructions are likely
5077 to become redundant now. We can't compare against the cost
5078 of src_eqv_here because, on MIPS for example, multi-insn
5079 constants have zero cost; they are assumed to be hoisted from
5080 loops. */
5081 if (src_related_is_const_anchor
5082 && src_related_cost == src_cost
5083 && src_eqv_here)
5084 src_related_cost--;
5088 /* If this was an indirect jump insn, a known label will really be
5089 cheaper even though it looks more expensive. */
5090 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5091 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5093 /* Terminate loop when replacement made. This must terminate since
5094 the current contents will be tested and will always be valid. */
5095 while (1)
5097 rtx trial;
5099 /* Skip invalid entries. */
5100 while (elt && !REG_P (elt->exp)
5101 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5102 elt = elt->next_same_value;
5104 /* A paradoxical subreg would be bad here: it'll be the right
5105 size, but later may be adjusted so that the upper bits aren't
5106 what we want. So reject it. */
5107 if (elt != 0
5108 && paradoxical_subreg_p (elt->exp)
5109 /* It is okay, though, if the rtx we're trying to match
5110 will ignore any of the bits we can't predict. */
5111 && ! (src != 0
5112 && GET_CODE (src) == SUBREG
5113 && GET_MODE (src) == GET_MODE (elt->exp)
5114 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5115 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5117 elt = elt->next_same_value;
5118 continue;
5121 if (elt)
5123 src_elt_cost = elt->cost;
5124 src_elt_regcost = elt->regcost;
5127 /* Find cheapest and skip it for the next time. For items
5128 of equal cost, use this order:
5129 src_folded, src, src_eqv, src_related and hash table entry. */
5130 if (src_folded
5131 && preferable (src_folded_cost, src_folded_regcost,
5132 src_cost, src_regcost) <= 0
5133 && preferable (src_folded_cost, src_folded_regcost,
5134 src_eqv_cost, src_eqv_regcost) <= 0
5135 && preferable (src_folded_cost, src_folded_regcost,
5136 src_related_cost, src_related_regcost) <= 0
5137 && preferable (src_folded_cost, src_folded_regcost,
5138 src_elt_cost, src_elt_regcost) <= 0)
5140 trial = src_folded, src_folded_cost = MAX_COST;
5141 if (src_folded_force_flag)
5143 rtx forced = force_const_mem (mode, trial);
5144 if (forced)
5145 trial = forced;
5148 else if (src
5149 && preferable (src_cost, src_regcost,
5150 src_eqv_cost, src_eqv_regcost) <= 0
5151 && preferable (src_cost, src_regcost,
5152 src_related_cost, src_related_regcost) <= 0
5153 && preferable (src_cost, src_regcost,
5154 src_elt_cost, src_elt_regcost) <= 0)
5155 trial = src, src_cost = MAX_COST;
5156 else if (src_eqv_here
5157 && preferable (src_eqv_cost, src_eqv_regcost,
5158 src_related_cost, src_related_regcost) <= 0
5159 && preferable (src_eqv_cost, src_eqv_regcost,
5160 src_elt_cost, src_elt_regcost) <= 0)
5161 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5162 else if (src_related
5163 && preferable (src_related_cost, src_related_regcost,
5164 src_elt_cost, src_elt_regcost) <= 0)
5165 trial = src_related, src_related_cost = MAX_COST;
5166 else
5168 trial = elt->exp;
5169 elt = elt->next_same_value;
5170 src_elt_cost = MAX_COST;
5173 /* Avoid creation of overlapping memory moves. */
5174 if (MEM_P (trial) && MEM_P (dest) && !rtx_equal_p (trial, dest))
5176 rtx src, dest;
5178 /* BLKmode moves are not handled by cse anyway. */
5179 if (GET_MODE (trial) == BLKmode)
5180 break;
5182 src = canon_rtx (trial);
5183 dest = canon_rtx (SET_DEST (sets[i].rtl));
5185 if (!MEM_P (src) || !MEM_P (dest)
5186 || !nonoverlapping_memrefs_p (src, dest, false))
5187 break;
5190 /* Try to optimize
5191 (set (reg:M N) (const_int A))
5192 (set (reg:M2 O) (const_int B))
5193 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5194 (reg:M2 O)). */
5195 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5196 && CONST_INT_P (trial)
5197 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5198 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5199 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5200 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5201 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5202 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5203 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5204 <= HOST_BITS_PER_WIDE_INT))
5206 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5207 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5208 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5209 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5210 struct table_elt *dest_elt
5211 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5212 rtx dest_cst = NULL;
5214 if (dest_elt)
5215 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5216 if (p->is_const && CONST_INT_P (p->exp))
5218 dest_cst = p->exp;
5219 break;
5221 if (dest_cst)
5223 HOST_WIDE_INT val = INTVAL (dest_cst);
5224 HOST_WIDE_INT mask;
5225 unsigned int shift;
5226 if (BITS_BIG_ENDIAN)
5227 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5228 - INTVAL (pos) - INTVAL (width);
5229 else
5230 shift = INTVAL (pos);
5231 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5232 mask = HOST_WIDE_INT_M1;
5233 else
5234 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5235 val &= ~(mask << shift);
5236 val |= (INTVAL (trial) & mask) << shift;
5237 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5238 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5239 dest_reg, 1);
5240 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5241 GEN_INT (val), 1);
5242 if (apply_change_group ())
5244 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5245 if (note)
5247 remove_note (insn, note);
5248 df_notes_rescan (insn);
5250 src_eqv = NULL_RTX;
5251 src_eqv_elt = NULL;
5252 src_eqv_volatile = 0;
5253 src_eqv_in_memory = 0;
5254 src_eqv_hash = 0;
5255 repeat = true;
5256 break;
5261 /* We don't normally have an insn matching (set (pc) (pc)), so
5262 check for this separately here. We will delete such an
5263 insn below.
5265 For other cases such as a table jump or conditional jump
5266 where we know the ultimate target, go ahead and replace the
5267 operand. While that may not make a valid insn, we will
5268 reemit the jump below (and also insert any necessary
5269 barriers). */
5270 if (n_sets == 1 && dest == pc_rtx
5271 && (trial == pc_rtx
5272 || (GET_CODE (trial) == LABEL_REF
5273 && ! condjump_p (insn))))
5275 /* Don't substitute non-local labels, this confuses CFG. */
5276 if (GET_CODE (trial) == LABEL_REF
5277 && LABEL_REF_NONLOCAL_P (trial))
5278 continue;
5280 SET_SRC (sets[i].rtl) = trial;
5281 cse_jumps_altered = true;
5282 break;
5285 /* Similarly, lots of targets don't allow no-op
5286 (set (mem x) (mem x)) moves. */
5287 else if (n_sets == 1
5288 && MEM_P (trial)
5289 && MEM_P (dest)
5290 && rtx_equal_p (trial, dest)
5291 && !side_effects_p (dest)
5292 && (cfun->can_delete_dead_exceptions
5293 || insn_nothrow_p (insn)))
5295 SET_SRC (sets[i].rtl) = trial;
5296 mem_noop_insn = true;
5297 break;
5300 /* Reject certain invalid forms of CONST that we create. */
5301 else if (CONSTANT_P (trial)
5302 && GET_CODE (trial) == CONST
5303 /* Reject cases that will cause decode_rtx_const to
5304 die. On the alpha when simplifying a switch, we
5305 get (const (truncate (minus (label_ref)
5306 (label_ref)))). */
5307 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5308 /* Likewise on IA-64, except without the
5309 truncate. */
5310 || (GET_CODE (XEXP (trial, 0)) == MINUS
5311 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5312 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5313 /* Do nothing for this case. */
5316 /* Look for a substitution that makes a valid insn. */
5317 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5318 trial, 0))
5320 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5322 /* The result of apply_change_group can be ignored; see
5323 canon_reg. */
5325 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5326 apply_change_group ();
5328 break;
5331 /* If we previously found constant pool entries for
5332 constants and this is a constant, try making a
5333 pool entry. Put it in src_folded unless we already have done
5334 this since that is where it likely came from. */
5336 else if (constant_pool_entries_cost
5337 && CONSTANT_P (trial)
5338 && (src_folded == 0
5339 || (!MEM_P (src_folded)
5340 && ! src_folded_force_flag))
5341 && GET_MODE_CLASS (mode) != MODE_CC
5342 && mode != VOIDmode)
5344 src_folded_force_flag = 1;
5345 src_folded = trial;
5346 src_folded_cost = constant_pool_entries_cost;
5347 src_folded_regcost = constant_pool_entries_regcost;
5351 /* If we changed the insn too much, handle this set from scratch. */
5352 if (repeat)
5354 i--;
5355 continue;
5358 src = SET_SRC (sets[i].rtl);
5360 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5361 However, there is an important exception: If both are registers
5362 that are not the head of their equivalence class, replace SET_SRC
5363 with the head of the class. If we do not do this, we will have
5364 both registers live over a portion of the basic block. This way,
5365 their lifetimes will likely abut instead of overlapping. */
5366 if (REG_P (dest)
5367 && REGNO_QTY_VALID_P (REGNO (dest)))
5369 int dest_q = REG_QTY (REGNO (dest));
5370 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5372 if (dest_ent->mode == GET_MODE (dest)
5373 && dest_ent->first_reg != REGNO (dest)
5374 && REG_P (src) && REGNO (src) == REGNO (dest)
5375 /* Don't do this if the original insn had a hard reg as
5376 SET_SRC or SET_DEST. */
5377 && (!REG_P (sets[i].src)
5378 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5379 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5380 /* We can't call canon_reg here because it won't do anything if
5381 SRC is a hard register. */
5383 int src_q = REG_QTY (REGNO (src));
5384 struct qty_table_elem *src_ent = &qty_table[src_q];
5385 int first = src_ent->first_reg;
5386 rtx new_src
5387 = (first >= FIRST_PSEUDO_REGISTER
5388 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5390 /* We must use validate-change even for this, because this
5391 might be a special no-op instruction, suitable only to
5392 tag notes onto. */
5393 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5395 src = new_src;
5396 /* If we had a constant that is cheaper than what we are now
5397 setting SRC to, use that constant. We ignored it when we
5398 thought we could make this into a no-op. */
5399 if (src_const && COST (src_const, mode) < COST (src, mode)
5400 && validate_change (insn, &SET_SRC (sets[i].rtl),
5401 src_const, 0))
5402 src = src_const;
5407 /* If we made a change, recompute SRC values. */
5408 if (src != sets[i].src)
5410 do_not_record = 0;
5411 hash_arg_in_memory = 0;
5412 sets[i].src = src;
5413 sets[i].src_hash = HASH (src, mode);
5414 sets[i].src_volatile = do_not_record;
5415 sets[i].src_in_memory = hash_arg_in_memory;
5416 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5419 /* If this is a single SET, we are setting a register, and we have an
5420 equivalent constant, we want to add a REG_EQUAL note if the constant
5421 is different from the source. We don't want to do it for a constant
5422 pseudo since verifying that this pseudo hasn't been eliminated is a
5423 pain; moreover such a note won't help anything.
5425 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5426 which can be created for a reference to a compile time computable
5427 entry in a jump table. */
5428 if (n_sets == 1
5429 && REG_P (dest)
5430 && src_const
5431 && !REG_P (src_const)
5432 && !(GET_CODE (src_const) == SUBREG
5433 && REG_P (SUBREG_REG (src_const)))
5434 && !(GET_CODE (src_const) == CONST
5435 && GET_CODE (XEXP (src_const, 0)) == MINUS
5436 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5437 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5438 && !rtx_equal_p (src, src_const))
5440 /* Make sure that the rtx is not shared. */
5441 src_const = copy_rtx (src_const);
5443 /* Record the actual constant value in a REG_EQUAL note,
5444 making a new one if one does not already exist. */
5445 set_unique_reg_note (insn, REG_EQUAL, src_const);
5446 df_notes_rescan (insn);
5449 /* Now deal with the destination. */
5450 do_not_record = 0;
5452 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5453 while (GET_CODE (dest) == SUBREG
5454 || GET_CODE (dest) == ZERO_EXTRACT
5455 || GET_CODE (dest) == STRICT_LOW_PART)
5456 dest = XEXP (dest, 0);
5458 sets[i].inner_dest = dest;
5460 if (MEM_P (dest))
5462 #ifdef PUSH_ROUNDING
5463 /* Stack pushes invalidate the stack pointer. */
5464 rtx addr = XEXP (dest, 0);
5465 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5466 && XEXP (addr, 0) == stack_pointer_rtx)
5467 invalidate (stack_pointer_rtx, VOIDmode);
5468 #endif
5469 dest = fold_rtx (dest, insn);
5472 /* Compute the hash code of the destination now,
5473 before the effects of this instruction are recorded,
5474 since the register values used in the address computation
5475 are those before this instruction. */
5476 sets[i].dest_hash = HASH (dest, mode);
5478 /* Don't enter a bit-field in the hash table
5479 because the value in it after the store
5480 may not equal what was stored, due to truncation. */
5482 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5484 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5486 if (src_const != 0 && CONST_INT_P (src_const)
5487 && CONST_INT_P (width)
5488 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5489 && ! (INTVAL (src_const)
5490 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5491 /* Exception: if the value is constant,
5492 and it won't be truncated, record it. */
5494 else
5496 /* This is chosen so that the destination will be invalidated
5497 but no new value will be recorded.
5498 We must invalidate because sometimes constant
5499 values can be recorded for bitfields. */
5500 sets[i].src_elt = 0;
5501 sets[i].src_volatile = 1;
5502 src_eqv = 0;
5503 src_eqv_elt = 0;
5507 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5508 the insn. */
5509 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5511 /* One less use of the label this insn used to jump to. */
5512 cse_cfg_altered |= delete_insn_and_edges (insn);
5513 cse_jumps_altered = true;
5514 /* No more processing for this set. */
5515 sets[i].rtl = 0;
5518 /* Similarly for no-op MEM moves. */
5519 else if (mem_noop_insn)
5521 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5522 cse_cfg_altered = true;
5523 cse_cfg_altered |= delete_insn_and_edges (insn);
5524 /* No more processing for this set. */
5525 sets[i].rtl = 0;
5528 /* If this SET is now setting PC to a label, we know it used to
5529 be a conditional or computed branch. */
5530 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5531 && !LABEL_REF_NONLOCAL_P (src))
5533 /* We reemit the jump in as many cases as possible just in
5534 case the form of an unconditional jump is significantly
5535 different than a computed jump or conditional jump.
5537 If this insn has multiple sets, then reemitting the
5538 jump is nontrivial. So instead we just force rerecognition
5539 and hope for the best. */
5540 if (n_sets == 1)
5542 rtx_jump_insn *new_rtx;
5543 rtx note;
5545 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5546 new_rtx = emit_jump_insn_before (seq, insn);
5547 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5548 LABEL_NUSES (XEXP (src, 0))++;
5550 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5551 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5552 if (note)
5554 XEXP (note, 1) = NULL_RTX;
5555 REG_NOTES (new_rtx) = note;
5558 cse_cfg_altered |= delete_insn_and_edges (insn);
5559 insn = new_rtx;
5561 else
5562 INSN_CODE (insn) = -1;
5564 /* Do not bother deleting any unreachable code, let jump do it. */
5565 cse_jumps_altered = true;
5566 sets[i].rtl = 0;
5569 /* If destination is volatile, invalidate it and then do no further
5570 processing for this assignment. */
5572 else if (do_not_record)
5574 invalidate_dest (dest);
5575 sets[i].rtl = 0;
5578 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5580 do_not_record = 0;
5581 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5582 if (do_not_record)
5584 invalidate_dest (SET_DEST (sets[i].rtl));
5585 sets[i].rtl = 0;
5589 /* If setting CC0, record what it was set to, or a constant, if it
5590 is equivalent to a constant. If it is being set to a floating-point
5591 value, make a COMPARE with the appropriate constant of 0. If we
5592 don't do this, later code can interpret this as a test against
5593 const0_rtx, which can cause problems if we try to put it into an
5594 insn as a floating-point operand. */
5595 if (dest == cc0_rtx)
5597 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5598 this_insn_cc0_mode = mode;
5599 if (FLOAT_MODE_P (mode))
5600 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5601 CONST0_RTX (mode));
5605 /* Now enter all non-volatile source expressions in the hash table
5606 if they are not already present.
5607 Record their equivalence classes in src_elt.
5608 This way we can insert the corresponding destinations into
5609 the same classes even if the actual sources are no longer in them
5610 (having been invalidated). */
5612 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5613 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5615 struct table_elt *elt;
5616 struct table_elt *classp = sets[0].src_elt;
5617 rtx dest = SET_DEST (sets[0].rtl);
5618 machine_mode eqvmode = GET_MODE (dest);
5620 if (GET_CODE (dest) == STRICT_LOW_PART)
5622 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5623 classp = 0;
5625 if (insert_regs (src_eqv, classp, 0))
5627 rehash_using_reg (src_eqv);
5628 src_eqv_hash = HASH (src_eqv, eqvmode);
5630 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5631 elt->in_memory = src_eqv_in_memory;
5632 src_eqv_elt = elt;
5634 /* Check to see if src_eqv_elt is the same as a set source which
5635 does not yet have an elt, and if so set the elt of the set source
5636 to src_eqv_elt. */
5637 for (i = 0; i < n_sets; i++)
5638 if (sets[i].rtl && sets[i].src_elt == 0
5639 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5640 sets[i].src_elt = src_eqv_elt;
5643 for (i = 0; i < n_sets; i++)
5644 if (sets[i].rtl && ! sets[i].src_volatile
5645 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5647 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5649 /* REG_EQUAL in setting a STRICT_LOW_PART
5650 gives an equivalent for the entire destination register,
5651 not just for the subreg being stored in now.
5652 This is a more interesting equivalence, so we arrange later
5653 to treat the entire reg as the destination. */
5654 sets[i].src_elt = src_eqv_elt;
5655 sets[i].src_hash = src_eqv_hash;
5657 else
5659 /* Insert source and constant equivalent into hash table, if not
5660 already present. */
5661 struct table_elt *classp = src_eqv_elt;
5662 rtx src = sets[i].src;
5663 rtx dest = SET_DEST (sets[i].rtl);
5664 machine_mode mode
5665 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5667 /* It's possible that we have a source value known to be
5668 constant but don't have a REG_EQUAL note on the insn.
5669 Lack of a note will mean src_eqv_elt will be NULL. This
5670 can happen where we've generated a SUBREG to access a
5671 CONST_INT that is already in a register in a wider mode.
5672 Ensure that the source expression is put in the proper
5673 constant class. */
5674 if (!classp)
5675 classp = sets[i].src_const_elt;
5677 if (sets[i].src_elt == 0)
5679 struct table_elt *elt;
5681 /* Note that these insert_regs calls cannot remove
5682 any of the src_elt's, because they would have failed to
5683 match if not still valid. */
5684 if (insert_regs (src, classp, 0))
5686 rehash_using_reg (src);
5687 sets[i].src_hash = HASH (src, mode);
5689 elt = insert (src, classp, sets[i].src_hash, mode);
5690 elt->in_memory = sets[i].src_in_memory;
5691 /* If inline asm has any clobbers, ensure we only reuse
5692 existing inline asms and never try to put the ASM_OPERANDS
5693 into an insn that isn't inline asm. */
5694 if (GET_CODE (src) == ASM_OPERANDS
5695 && GET_CODE (x) == PARALLEL)
5696 elt->cost = MAX_COST;
5697 sets[i].src_elt = classp = elt;
5699 if (sets[i].src_const && sets[i].src_const_elt == 0
5700 && src != sets[i].src_const
5701 && ! rtx_equal_p (sets[i].src_const, src))
5702 sets[i].src_elt = insert (sets[i].src_const, classp,
5703 sets[i].src_const_hash, mode);
5706 else if (sets[i].src_elt == 0)
5707 /* If we did not insert the source into the hash table (e.g., it was
5708 volatile), note the equivalence class for the REG_EQUAL value, if any,
5709 so that the destination goes into that class. */
5710 sets[i].src_elt = src_eqv_elt;
5712 /* Record destination addresses in the hash table. This allows us to
5713 check if they are invalidated by other sets. */
5714 for (i = 0; i < n_sets; i++)
5716 if (sets[i].rtl)
5718 rtx x = sets[i].inner_dest;
5719 struct table_elt *elt;
5720 machine_mode mode;
5721 unsigned hash;
5723 if (MEM_P (x))
5725 x = XEXP (x, 0);
5726 mode = GET_MODE (x);
5727 hash = HASH (x, mode);
5728 elt = lookup (x, hash, mode);
5729 if (!elt)
5731 if (insert_regs (x, NULL, 0))
5733 rtx dest = SET_DEST (sets[i].rtl);
5735 rehash_using_reg (x);
5736 hash = HASH (x, mode);
5737 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5739 elt = insert (x, NULL, hash, mode);
5742 sets[i].dest_addr_elt = elt;
5744 else
5745 sets[i].dest_addr_elt = NULL;
5749 invalidate_from_clobbers (insn);
5751 /* Some registers are invalidated by subroutine calls. Memory is
5752 invalidated by non-constant calls. */
5754 if (CALL_P (insn))
5756 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5757 invalidate_memory ();
5758 else
5759 /* For const/pure calls, invalidate any argument slots, because
5760 those are owned by the callee. */
5761 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5762 if (GET_CODE (XEXP (tem, 0)) == USE
5763 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5764 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5765 invalidate_for_call ();
5768 /* Now invalidate everything set by this instruction.
5769 If a SUBREG or other funny destination is being set,
5770 sets[i].rtl is still nonzero, so here we invalidate the reg
5771 a part of which is being set. */
5773 for (i = 0; i < n_sets; i++)
5774 if (sets[i].rtl)
5776 /* We can't use the inner dest, because the mode associated with
5777 a ZERO_EXTRACT is significant. */
5778 rtx dest = SET_DEST (sets[i].rtl);
5780 /* Needed for registers to remove the register from its
5781 previous quantity's chain.
5782 Needed for memory if this is a nonvarying address, unless
5783 we have just done an invalidate_memory that covers even those. */
5784 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5785 invalidate (dest, VOIDmode);
5786 else if (MEM_P (dest))
5787 invalidate (dest, VOIDmode);
5788 else if (GET_CODE (dest) == STRICT_LOW_PART
5789 || GET_CODE (dest) == ZERO_EXTRACT)
5790 invalidate (XEXP (dest, 0), GET_MODE (dest));
5793 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5794 the regs restored by the longjmp come from a later time
5795 than the setjmp. */
5796 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5798 flush_hash_table ();
5799 goto done;
5802 /* Make sure registers mentioned in destinations
5803 are safe for use in an expression to be inserted.
5804 This removes from the hash table
5805 any invalid entry that refers to one of these registers.
5807 We don't care about the return value from mention_regs because
5808 we are going to hash the SET_DEST values unconditionally. */
5810 for (i = 0; i < n_sets; i++)
5812 if (sets[i].rtl)
5814 rtx x = SET_DEST (sets[i].rtl);
5816 if (!REG_P (x))
5817 mention_regs (x);
5818 else
5820 /* We used to rely on all references to a register becoming
5821 inaccessible when a register changes to a new quantity,
5822 since that changes the hash code. However, that is not
5823 safe, since after HASH_SIZE new quantities we get a
5824 hash 'collision' of a register with its own invalid
5825 entries. And since SUBREGs have been changed not to
5826 change their hash code with the hash code of the register,
5827 it wouldn't work any longer at all. So we have to check
5828 for any invalid references lying around now.
5829 This code is similar to the REG case in mention_regs,
5830 but it knows that reg_tick has been incremented, and
5831 it leaves reg_in_table as -1 . */
5832 unsigned int regno = REGNO (x);
5833 unsigned int endregno = END_REGNO (x);
5834 unsigned int i;
5836 for (i = regno; i < endregno; i++)
5838 if (REG_IN_TABLE (i) >= 0)
5840 remove_invalid_refs (i);
5841 REG_IN_TABLE (i) = -1;
5848 /* We may have just removed some of the src_elt's from the hash table.
5849 So replace each one with the current head of the same class.
5850 Also check if destination addresses have been removed. */
5852 for (i = 0; i < n_sets; i++)
5853 if (sets[i].rtl)
5855 if (sets[i].dest_addr_elt
5856 && sets[i].dest_addr_elt->first_same_value == 0)
5858 /* The elt was removed, which means this destination is not
5859 valid after this instruction. */
5860 sets[i].rtl = NULL_RTX;
5862 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5863 /* If elt was removed, find current head of same class,
5864 or 0 if nothing remains of that class. */
5866 struct table_elt *elt = sets[i].src_elt;
5868 while (elt && elt->prev_same_value)
5869 elt = elt->prev_same_value;
5871 while (elt && elt->first_same_value == 0)
5872 elt = elt->next_same_value;
5873 sets[i].src_elt = elt ? elt->first_same_value : 0;
5877 /* Now insert the destinations into their equivalence classes. */
5879 for (i = 0; i < n_sets; i++)
5880 if (sets[i].rtl)
5882 rtx dest = SET_DEST (sets[i].rtl);
5883 struct table_elt *elt;
5885 /* Don't record value if we are not supposed to risk allocating
5886 floating-point values in registers that might be wider than
5887 memory. */
5888 if ((flag_float_store
5889 && MEM_P (dest)
5890 && FLOAT_MODE_P (GET_MODE (dest)))
5891 /* Don't record BLKmode values, because we don't know the
5892 size of it, and can't be sure that other BLKmode values
5893 have the same or smaller size. */
5894 || GET_MODE (dest) == BLKmode
5895 /* If we didn't put a REG_EQUAL value or a source into the hash
5896 table, there is no point is recording DEST. */
5897 || sets[i].src_elt == 0)
5898 continue;
5900 /* STRICT_LOW_PART isn't part of the value BEING set,
5901 and neither is the SUBREG inside it.
5902 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5903 if (GET_CODE (dest) == STRICT_LOW_PART)
5904 dest = SUBREG_REG (XEXP (dest, 0));
5906 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5907 /* Registers must also be inserted into chains for quantities. */
5908 if (insert_regs (dest, sets[i].src_elt, 1))
5910 /* If `insert_regs' changes something, the hash code must be
5911 recalculated. */
5912 rehash_using_reg (dest);
5913 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5916 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5917 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5918 if (paradoxical_subreg_p (dest))
5919 continue;
5921 elt = insert (dest, sets[i].src_elt,
5922 sets[i].dest_hash, GET_MODE (dest));
5924 /* If this is a constant, insert the constant anchors with the
5925 equivalent register-offset expressions using register DEST. */
5926 if (targetm.const_anchor
5927 && REG_P (dest)
5928 && SCALAR_INT_MODE_P (GET_MODE (dest))
5929 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5930 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5932 elt->in_memory = (MEM_P (sets[i].inner_dest)
5933 && !MEM_READONLY_P (sets[i].inner_dest));
5935 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5936 narrower than M2, and both M1 and M2 are the same number of words,
5937 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5938 make that equivalence as well.
5940 However, BAR may have equivalences for which gen_lowpart
5941 will produce a simpler value than gen_lowpart applied to
5942 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5943 BAR's equivalences. If we don't get a simplified form, make
5944 the SUBREG. It will not be used in an equivalence, but will
5945 cause two similar assignments to be detected.
5947 Note the loop below will find SUBREG_REG (DEST) since we have
5948 already entered SRC and DEST of the SET in the table. */
5950 if (GET_CODE (dest) == SUBREG
5951 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5952 / UNITS_PER_WORD)
5953 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5954 && (GET_MODE_SIZE (GET_MODE (dest))
5955 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5956 && sets[i].src_elt != 0)
5958 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5959 struct table_elt *elt, *classp = 0;
5961 for (elt = sets[i].src_elt->first_same_value; elt;
5962 elt = elt->next_same_value)
5964 rtx new_src = 0;
5965 unsigned src_hash;
5966 struct table_elt *src_elt;
5967 int byte = 0;
5969 /* Ignore invalid entries. */
5970 if (!REG_P (elt->exp)
5971 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5972 continue;
5974 /* We may have already been playing subreg games. If the
5975 mode is already correct for the destination, use it. */
5976 if (GET_MODE (elt->exp) == new_mode)
5977 new_src = elt->exp;
5978 else
5980 /* Calculate big endian correction for the SUBREG_BYTE.
5981 We have already checked that M1 (GET_MODE (dest))
5982 is not narrower than M2 (new_mode). */
5983 if (BYTES_BIG_ENDIAN)
5984 byte = (GET_MODE_SIZE (GET_MODE (dest))
5985 - GET_MODE_SIZE (new_mode));
5987 new_src = simplify_gen_subreg (new_mode, elt->exp,
5988 GET_MODE (dest), byte);
5991 /* The call to simplify_gen_subreg fails if the value
5992 is VOIDmode, yet we can't do any simplification, e.g.
5993 for EXPR_LISTs denoting function call results.
5994 It is invalid to construct a SUBREG with a VOIDmode
5995 SUBREG_REG, hence a zero new_src means we can't do
5996 this substitution. */
5997 if (! new_src)
5998 continue;
6000 src_hash = HASH (new_src, new_mode);
6001 src_elt = lookup (new_src, src_hash, new_mode);
6003 /* Put the new source in the hash table is if isn't
6004 already. */
6005 if (src_elt == 0)
6007 if (insert_regs (new_src, classp, 0))
6009 rehash_using_reg (new_src);
6010 src_hash = HASH (new_src, new_mode);
6012 src_elt = insert (new_src, classp, src_hash, new_mode);
6013 src_elt->in_memory = elt->in_memory;
6014 if (GET_CODE (new_src) == ASM_OPERANDS
6015 && elt->cost == MAX_COST)
6016 src_elt->cost = MAX_COST;
6018 else if (classp && classp != src_elt->first_same_value)
6019 /* Show that two things that we've seen before are
6020 actually the same. */
6021 merge_equiv_classes (src_elt, classp);
6023 classp = src_elt->first_same_value;
6024 /* Ignore invalid entries. */
6025 while (classp
6026 && !REG_P (classp->exp)
6027 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6028 classp = classp->next_same_value;
6033 /* Special handling for (set REG0 REG1) where REG0 is the
6034 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6035 be used in the sequel, so (if easily done) change this insn to
6036 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6037 that computed their value. Then REG1 will become a dead store
6038 and won't cloud the situation for later optimizations.
6040 Do not make this change if REG1 is a hard register, because it will
6041 then be used in the sequel and we may be changing a two-operand insn
6042 into a three-operand insn.
6044 Also do not do this if we are operating on a copy of INSN. */
6046 if (n_sets == 1 && sets[0].rtl)
6047 try_back_substitute_reg (sets[0].rtl, insn);
6049 done:;
6052 /* Remove from the hash table all expressions that reference memory. */
6054 static void
6055 invalidate_memory (void)
6057 int i;
6058 struct table_elt *p, *next;
6060 for (i = 0; i < HASH_SIZE; i++)
6061 for (p = table[i]; p; p = next)
6063 next = p->next_same_hash;
6064 if (p->in_memory)
6065 remove_from_table (p, i);
6069 /* Perform invalidation on the basis of everything about INSN,
6070 except for invalidating the actual places that are SET in it.
6071 This includes the places CLOBBERed, and anything that might
6072 alias with something that is SET or CLOBBERed. */
6074 static void
6075 invalidate_from_clobbers (rtx_insn *insn)
6077 rtx x = PATTERN (insn);
6079 if (GET_CODE (x) == CLOBBER)
6081 rtx ref = XEXP (x, 0);
6082 if (ref)
6084 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6085 || MEM_P (ref))
6086 invalidate (ref, VOIDmode);
6087 else if (GET_CODE (ref) == STRICT_LOW_PART
6088 || GET_CODE (ref) == ZERO_EXTRACT)
6089 invalidate (XEXP (ref, 0), GET_MODE (ref));
6092 else if (GET_CODE (x) == PARALLEL)
6094 int i;
6095 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6097 rtx y = XVECEXP (x, 0, i);
6098 if (GET_CODE (y) == CLOBBER)
6100 rtx ref = XEXP (y, 0);
6101 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6102 || MEM_P (ref))
6103 invalidate (ref, VOIDmode);
6104 else if (GET_CODE (ref) == STRICT_LOW_PART
6105 || GET_CODE (ref) == ZERO_EXTRACT)
6106 invalidate (XEXP (ref, 0), GET_MODE (ref));
6112 /* Perform invalidation on the basis of everything about INSN.
6113 This includes the places CLOBBERed, and anything that might
6114 alias with something that is SET or CLOBBERed. */
6116 static void
6117 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6119 rtx tem;
6120 rtx x = PATTERN (insn);
6122 if (CALL_P (insn))
6124 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6125 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6126 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6129 /* Ensure we invalidate the destination register of a CALL insn.
6130 This is necessary for machines where this register is a fixed_reg,
6131 because no other code would invalidate it. */
6132 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6133 invalidate (SET_DEST (x), VOIDmode);
6135 else if (GET_CODE (x) == PARALLEL)
6137 int i;
6139 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6141 rtx y = XVECEXP (x, 0, i);
6142 if (GET_CODE (y) == CLOBBER)
6144 rtx clobbered = XEXP (y, 0);
6146 if (REG_P (clobbered)
6147 || GET_CODE (clobbered) == SUBREG)
6148 invalidate (clobbered, VOIDmode);
6149 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6150 || GET_CODE (clobbered) == ZERO_EXTRACT)
6151 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6153 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6154 invalidate (SET_DEST (y), VOIDmode);
6159 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6160 and replace any registers in them with either an equivalent constant
6161 or the canonical form of the register. If we are inside an address,
6162 only do this if the address remains valid.
6164 OBJECT is 0 except when within a MEM in which case it is the MEM.
6166 Return the replacement for X. */
6168 static rtx
6169 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6171 enum rtx_code code = GET_CODE (x);
6172 const char *fmt = GET_RTX_FORMAT (code);
6173 int i;
6175 switch (code)
6177 case CONST:
6178 case SYMBOL_REF:
6179 case LABEL_REF:
6180 CASE_CONST_ANY:
6181 case PC:
6182 case CC0:
6183 case LO_SUM:
6184 return x;
6186 case MEM:
6187 validate_change (x, &XEXP (x, 0),
6188 cse_process_notes (XEXP (x, 0), x, changed), 0);
6189 return x;
6191 case EXPR_LIST:
6192 if (REG_NOTE_KIND (x) == REG_EQUAL)
6193 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6194 /* Fall through. */
6196 case INSN_LIST:
6197 case INT_LIST:
6198 if (XEXP (x, 1))
6199 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6200 return x;
6202 case SIGN_EXTEND:
6203 case ZERO_EXTEND:
6204 case SUBREG:
6206 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6207 /* We don't substitute VOIDmode constants into these rtx,
6208 since they would impede folding. */
6209 if (GET_MODE (new_rtx) != VOIDmode)
6210 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6211 return x;
6214 case UNSIGNED_FLOAT:
6216 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6217 /* We don't substitute negative VOIDmode constants into these rtx,
6218 since they would impede folding. */
6219 if (GET_MODE (new_rtx) != VOIDmode
6220 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6221 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6222 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6223 return x;
6226 case REG:
6227 i = REG_QTY (REGNO (x));
6229 /* Return a constant or a constant register. */
6230 if (REGNO_QTY_VALID_P (REGNO (x)))
6232 struct qty_table_elem *ent = &qty_table[i];
6234 if (ent->const_rtx != NULL_RTX
6235 && (CONSTANT_P (ent->const_rtx)
6236 || REG_P (ent->const_rtx)))
6238 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6239 if (new_rtx)
6240 return copy_rtx (new_rtx);
6244 /* Otherwise, canonicalize this register. */
6245 return canon_reg (x, NULL);
6247 default:
6248 break;
6251 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6252 if (fmt[i] == 'e')
6253 validate_change (object, &XEXP (x, i),
6254 cse_process_notes (XEXP (x, i), object, changed), 0);
6256 return x;
6259 static rtx
6260 cse_process_notes (rtx x, rtx object, bool *changed)
6262 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6263 if (new_rtx != x)
6264 *changed = true;
6265 return new_rtx;
6269 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6271 DATA is a pointer to a struct cse_basic_block_data, that is used to
6272 describe the path.
6273 It is filled with a queue of basic blocks, starting with FIRST_BB
6274 and following a trace through the CFG.
6276 If all paths starting at FIRST_BB have been followed, or no new path
6277 starting at FIRST_BB can be constructed, this function returns FALSE.
6278 Otherwise, DATA->path is filled and the function returns TRUE indicating
6279 that a path to follow was found.
6281 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6282 block in the path will be FIRST_BB. */
6284 static bool
6285 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6286 int follow_jumps)
6288 basic_block bb;
6289 edge e;
6290 int path_size;
6292 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6294 /* See if there is a previous path. */
6295 path_size = data->path_size;
6297 /* There is a previous path. Make sure it started with FIRST_BB. */
6298 if (path_size)
6299 gcc_assert (data->path[0].bb == first_bb);
6301 /* There was only one basic block in the last path. Clear the path and
6302 return, so that paths starting at another basic block can be tried. */
6303 if (path_size == 1)
6305 path_size = 0;
6306 goto done;
6309 /* If the path was empty from the beginning, construct a new path. */
6310 if (path_size == 0)
6311 data->path[path_size++].bb = first_bb;
6312 else
6314 /* Otherwise, path_size must be equal to or greater than 2, because
6315 a previous path exists that is at least two basic blocks long.
6317 Update the previous branch path, if any. If the last branch was
6318 previously along the branch edge, take the fallthrough edge now. */
6319 while (path_size >= 2)
6321 basic_block last_bb_in_path, previous_bb_in_path;
6322 edge e;
6324 --path_size;
6325 last_bb_in_path = data->path[path_size].bb;
6326 previous_bb_in_path = data->path[path_size - 1].bb;
6328 /* If we previously followed a path along the branch edge, try
6329 the fallthru edge now. */
6330 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6331 && any_condjump_p (BB_END (previous_bb_in_path))
6332 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6333 && e == BRANCH_EDGE (previous_bb_in_path))
6335 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6336 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6337 && single_pred_p (bb)
6338 /* We used to assert here that we would only see blocks
6339 that we have not visited yet. But we may end up
6340 visiting basic blocks twice if the CFG has changed
6341 in this run of cse_main, because when the CFG changes
6342 the topological sort of the CFG also changes. A basic
6343 blocks that previously had more than two predecessors
6344 may now have a single predecessor, and become part of
6345 a path that starts at another basic block.
6347 We still want to visit each basic block only once, so
6348 halt the path here if we have already visited BB. */
6349 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6351 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6352 data->path[path_size++].bb = bb;
6353 break;
6357 data->path[path_size].bb = NULL;
6360 /* If only one block remains in the path, bail. */
6361 if (path_size == 1)
6363 path_size = 0;
6364 goto done;
6368 /* Extend the path if possible. */
6369 if (follow_jumps)
6371 bb = data->path[path_size - 1].bb;
6372 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6374 if (single_succ_p (bb))
6375 e = single_succ_edge (bb);
6376 else if (EDGE_COUNT (bb->succs) == 2
6377 && any_condjump_p (BB_END (bb)))
6379 /* First try to follow the branch. If that doesn't lead
6380 to a useful path, follow the fallthru edge. */
6381 e = BRANCH_EDGE (bb);
6382 if (!single_pred_p (e->dest))
6383 e = FALLTHRU_EDGE (bb);
6385 else
6386 e = NULL;
6388 if (e
6389 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6390 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6391 && single_pred_p (e->dest)
6392 /* Avoid visiting basic blocks twice. The large comment
6393 above explains why this can happen. */
6394 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6396 basic_block bb2 = e->dest;
6397 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6398 data->path[path_size++].bb = bb2;
6399 bb = bb2;
6401 else
6402 bb = NULL;
6406 done:
6407 data->path_size = path_size;
6408 return path_size != 0;
6411 /* Dump the path in DATA to file F. NSETS is the number of sets
6412 in the path. */
6414 static void
6415 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6417 int path_entry;
6419 fprintf (f, ";; Following path with %d sets: ", nsets);
6420 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6421 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6422 fputc ('\n', dump_file);
6423 fflush (f);
6427 /* Return true if BB has exception handling successor edges. */
6429 static bool
6430 have_eh_succ_edges (basic_block bb)
6432 edge e;
6433 edge_iterator ei;
6435 FOR_EACH_EDGE (e, ei, bb->succs)
6436 if (e->flags & EDGE_EH)
6437 return true;
6439 return false;
6443 /* Scan to the end of the path described by DATA. Return an estimate of
6444 the total number of SETs of all insns in the path. */
6446 static void
6447 cse_prescan_path (struct cse_basic_block_data *data)
6449 int nsets = 0;
6450 int path_size = data->path_size;
6451 int path_entry;
6453 /* Scan to end of each basic block in the path. */
6454 for (path_entry = 0; path_entry < path_size; path_entry++)
6456 basic_block bb;
6457 rtx_insn *insn;
6459 bb = data->path[path_entry].bb;
6461 FOR_BB_INSNS (bb, insn)
6463 if (!INSN_P (insn))
6464 continue;
6466 /* A PARALLEL can have lots of SETs in it,
6467 especially if it is really an ASM_OPERANDS. */
6468 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6469 nsets += XVECLEN (PATTERN (insn), 0);
6470 else
6471 nsets += 1;
6475 data->nsets = nsets;
6478 /* Return true if the pattern of INSN uses a LABEL_REF for which
6479 there isn't a REG_LABEL_OPERAND note. */
6481 static bool
6482 check_for_label_ref (rtx_insn *insn)
6484 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6485 note for it, we must rerun jump since it needs to place the note. If
6486 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6487 don't do this since no REG_LABEL_OPERAND will be added. */
6488 subrtx_iterator::array_type array;
6489 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6491 const_rtx x = *iter;
6492 if (GET_CODE (x) == LABEL_REF
6493 && !LABEL_REF_NONLOCAL_P (x)
6494 && (!JUMP_P (insn)
6495 || !label_is_jump_target_p (label_ref_label (x), insn))
6496 && LABEL_P (label_ref_label (x))
6497 && INSN_UID (label_ref_label (x)) != 0
6498 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6499 return true;
6501 return false;
6504 /* Process a single extended basic block described by EBB_DATA. */
6506 static void
6507 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6509 int path_size = ebb_data->path_size;
6510 int path_entry;
6511 int num_insns = 0;
6513 /* Allocate the space needed by qty_table. */
6514 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6516 new_basic_block ();
6517 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6518 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6519 for (path_entry = 0; path_entry < path_size; path_entry++)
6521 basic_block bb;
6522 rtx_insn *insn;
6524 bb = ebb_data->path[path_entry].bb;
6526 /* Invalidate recorded information for eh regs if there is an EH
6527 edge pointing to that bb. */
6528 if (bb_has_eh_pred (bb))
6530 df_ref def;
6532 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6533 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6534 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6537 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6538 FOR_BB_INSNS (bb, insn)
6540 /* If we have processed 1,000 insns, flush the hash table to
6541 avoid extreme quadratic behavior. We must not include NOTEs
6542 in the count since there may be more of them when generating
6543 debugging information. If we clear the table at different
6544 times, code generated with -g -O might be different than code
6545 generated with -O but not -g.
6547 FIXME: This is a real kludge and needs to be done some other
6548 way. */
6549 if (NONDEBUG_INSN_P (insn)
6550 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6552 flush_hash_table ();
6553 num_insns = 0;
6556 if (INSN_P (insn))
6558 /* Process notes first so we have all notes in canonical forms
6559 when looking for duplicate operations. */
6560 if (REG_NOTES (insn))
6562 bool changed = false;
6563 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6564 NULL_RTX, &changed);
6565 if (changed)
6566 df_notes_rescan (insn);
6569 cse_insn (insn);
6571 /* If we haven't already found an insn where we added a LABEL_REF,
6572 check this one. */
6573 if (INSN_P (insn) && !recorded_label_ref
6574 && check_for_label_ref (insn))
6575 recorded_label_ref = true;
6577 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6579 /* If the previous insn sets CC0 and this insn no
6580 longer references CC0, delete the previous insn.
6581 Here we use fact that nothing expects CC0 to be
6582 valid over an insn, which is true until the final
6583 pass. */
6584 rtx_insn *prev_insn;
6585 rtx tem;
6587 prev_insn = prev_nonnote_nondebug_insn (insn);
6588 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6589 && (tem = single_set (prev_insn)) != NULL_RTX
6590 && SET_DEST (tem) == cc0_rtx
6591 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6592 delete_insn (prev_insn);
6594 /* If this insn is not the last insn in the basic
6595 block, it will be PREV_INSN(insn) in the next
6596 iteration. If we recorded any CC0-related
6597 information for this insn, remember it. */
6598 if (insn != BB_END (bb))
6600 prev_insn_cc0 = this_insn_cc0;
6601 prev_insn_cc0_mode = this_insn_cc0_mode;
6607 /* With non-call exceptions, we are not always able to update
6608 the CFG properly inside cse_insn. So clean up possibly
6609 redundant EH edges here. */
6610 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6611 cse_cfg_altered |= purge_dead_edges (bb);
6613 /* If we changed a conditional jump, we may have terminated
6614 the path we are following. Check that by verifying that
6615 the edge we would take still exists. If the edge does
6616 not exist anymore, purge the remainder of the path.
6617 Note that this will cause us to return to the caller. */
6618 if (path_entry < path_size - 1)
6620 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6621 if (!find_edge (bb, next_bb))
6625 path_size--;
6627 /* If we truncate the path, we must also reset the
6628 visited bit on the remaining blocks in the path,
6629 or we will never visit them at all. */
6630 bitmap_clear_bit (cse_visited_basic_blocks,
6631 ebb_data->path[path_size].bb->index);
6632 ebb_data->path[path_size].bb = NULL;
6634 while (path_size - 1 != path_entry);
6635 ebb_data->path_size = path_size;
6639 /* If this is a conditional jump insn, record any known
6640 equivalences due to the condition being tested. */
6641 insn = BB_END (bb);
6642 if (path_entry < path_size - 1
6643 && JUMP_P (insn)
6644 && single_set (insn)
6645 && any_condjump_p (insn))
6647 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6648 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6649 record_jump_equiv (insn, taken);
6652 /* Clear the CC0-tracking related insns, they can't provide
6653 useful information across basic block boundaries. */
6654 prev_insn_cc0 = 0;
6657 gcc_assert (next_qty <= max_qty);
6659 free (qty_table);
6663 /* Perform cse on the instructions of a function.
6664 F is the first instruction.
6665 NREGS is one plus the highest pseudo-reg number used in the instruction.
6667 Return 2 if jump optimizations should be redone due to simplifications
6668 in conditional jump instructions.
6669 Return 1 if the CFG should be cleaned up because it has been modified.
6670 Return 0 otherwise. */
6672 static int
6673 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6675 struct cse_basic_block_data ebb_data;
6676 basic_block bb;
6677 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6678 int i, n_blocks;
6680 /* CSE doesn't use dominane info but can invalidate it in different ways.
6681 For simplicity free dominance info here. */
6682 free_dominance_info (CDI_DOMINATORS);
6684 df_set_flags (DF_LR_RUN_DCE);
6685 df_note_add_problem ();
6686 df_analyze ();
6687 df_set_flags (DF_DEFER_INSN_RESCAN);
6689 reg_scan (get_insns (), max_reg_num ());
6690 init_cse_reg_info (nregs);
6692 ebb_data.path = XNEWVEC (struct branch_path,
6693 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6695 cse_cfg_altered = false;
6696 cse_jumps_altered = false;
6697 recorded_label_ref = false;
6698 constant_pool_entries_cost = 0;
6699 constant_pool_entries_regcost = 0;
6700 ebb_data.path_size = 0;
6701 ebb_data.nsets = 0;
6702 rtl_hooks = cse_rtl_hooks;
6704 init_recog ();
6705 init_alias_analysis ();
6707 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6709 /* Set up the table of already visited basic blocks. */
6710 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6711 bitmap_clear (cse_visited_basic_blocks);
6713 /* Loop over basic blocks in reverse completion order (RPO),
6714 excluding the ENTRY and EXIT blocks. */
6715 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6716 i = 0;
6717 while (i < n_blocks)
6719 /* Find the first block in the RPO queue that we have not yet
6720 processed before. */
6723 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6725 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6726 && i < n_blocks);
6728 /* Find all paths starting with BB, and process them. */
6729 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6731 /* Pre-scan the path. */
6732 cse_prescan_path (&ebb_data);
6734 /* If this basic block has no sets, skip it. */
6735 if (ebb_data.nsets == 0)
6736 continue;
6738 /* Get a reasonable estimate for the maximum number of qty's
6739 needed for this path. For this, we take the number of sets
6740 and multiply that by MAX_RECOG_OPERANDS. */
6741 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6743 /* Dump the path we're about to process. */
6744 if (dump_file)
6745 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6747 cse_extended_basic_block (&ebb_data);
6751 /* Clean up. */
6752 end_alias_analysis ();
6753 free (reg_eqv_table);
6754 free (ebb_data.path);
6755 sbitmap_free (cse_visited_basic_blocks);
6756 free (rc_order);
6757 rtl_hooks = general_rtl_hooks;
6759 if (cse_jumps_altered || recorded_label_ref)
6760 return 2;
6761 else if (cse_cfg_altered)
6762 return 1;
6763 else
6764 return 0;
6767 /* Count the number of times registers are used (not set) in X.
6768 COUNTS is an array in which we accumulate the count, INCR is how much
6769 we count each register usage.
6771 Don't count a usage of DEST, which is the SET_DEST of a SET which
6772 contains X in its SET_SRC. This is because such a SET does not
6773 modify the liveness of DEST.
6774 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6775 We must then count uses of a SET_DEST regardless, because the insn can't be
6776 deleted here. */
6778 static void
6779 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6781 enum rtx_code code;
6782 rtx note;
6783 const char *fmt;
6784 int i, j;
6786 if (x == 0)
6787 return;
6789 switch (code = GET_CODE (x))
6791 case REG:
6792 if (x != dest)
6793 counts[REGNO (x)] += incr;
6794 return;
6796 case PC:
6797 case CC0:
6798 case CONST:
6799 CASE_CONST_ANY:
6800 case SYMBOL_REF:
6801 case LABEL_REF:
6802 return;
6804 case CLOBBER:
6805 /* If we are clobbering a MEM, mark any registers inside the address
6806 as being used. */
6807 if (MEM_P (XEXP (x, 0)))
6808 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6809 return;
6811 case SET:
6812 /* Unless we are setting a REG, count everything in SET_DEST. */
6813 if (!REG_P (SET_DEST (x)))
6814 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6815 count_reg_usage (SET_SRC (x), counts,
6816 dest ? dest : SET_DEST (x),
6817 incr);
6818 return;
6820 case DEBUG_INSN:
6821 return;
6823 case CALL_INSN:
6824 case INSN:
6825 case JUMP_INSN:
6826 /* We expect dest to be NULL_RTX here. If the insn may throw,
6827 or if it cannot be deleted due to side-effects, mark this fact
6828 by setting DEST to pc_rtx. */
6829 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6830 || side_effects_p (PATTERN (x)))
6831 dest = pc_rtx;
6832 if (code == CALL_INSN)
6833 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6834 count_reg_usage (PATTERN (x), counts, dest, incr);
6836 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6837 use them. */
6839 note = find_reg_equal_equiv_note (x);
6840 if (note)
6842 rtx eqv = XEXP (note, 0);
6844 if (GET_CODE (eqv) == EXPR_LIST)
6845 /* This REG_EQUAL note describes the result of a function call.
6846 Process all the arguments. */
6849 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6850 eqv = XEXP (eqv, 1);
6852 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6853 else
6854 count_reg_usage (eqv, counts, dest, incr);
6856 return;
6858 case EXPR_LIST:
6859 if (REG_NOTE_KIND (x) == REG_EQUAL
6860 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6861 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6862 involving registers in the address. */
6863 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6864 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6866 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6867 return;
6869 case ASM_OPERANDS:
6870 /* Iterate over just the inputs, not the constraints as well. */
6871 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6872 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6873 return;
6875 case INSN_LIST:
6876 case INT_LIST:
6877 gcc_unreachable ();
6879 default:
6880 break;
6883 fmt = GET_RTX_FORMAT (code);
6884 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6886 if (fmt[i] == 'e')
6887 count_reg_usage (XEXP (x, i), counts, dest, incr);
6888 else if (fmt[i] == 'E')
6889 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6890 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6894 /* Return true if X is a dead register. */
6896 static inline int
6897 is_dead_reg (const_rtx x, int *counts)
6899 return (REG_P (x)
6900 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6901 && counts[REGNO (x)] == 0);
6904 /* Return true if set is live. */
6905 static bool
6906 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6907 int *counts)
6909 rtx_insn *tem;
6911 if (set_noop_p (set))
6914 else if (GET_CODE (SET_DEST (set)) == CC0
6915 && !side_effects_p (SET_SRC (set))
6916 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6917 || !INSN_P (tem)
6918 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6919 return false;
6920 else if (!is_dead_reg (SET_DEST (set), counts)
6921 || side_effects_p (SET_SRC (set)))
6922 return true;
6923 return false;
6926 /* Return true if insn is live. */
6928 static bool
6929 insn_live_p (rtx_insn *insn, int *counts)
6931 int i;
6932 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6933 return true;
6934 else if (GET_CODE (PATTERN (insn)) == SET)
6935 return set_live_p (PATTERN (insn), insn, counts);
6936 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6938 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6940 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6942 if (GET_CODE (elt) == SET)
6944 if (set_live_p (elt, insn, counts))
6945 return true;
6947 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6948 return true;
6950 return false;
6952 else if (DEBUG_INSN_P (insn))
6954 rtx_insn *next;
6956 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6957 if (NOTE_P (next))
6958 continue;
6959 else if (!DEBUG_INSN_P (next))
6960 return true;
6961 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6962 return false;
6964 return true;
6966 else
6967 return true;
6970 /* Count the number of stores into pseudo. Callback for note_stores. */
6972 static void
6973 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6975 int *counts = (int *) data;
6976 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6977 counts[REGNO (x)]++;
6980 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6981 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6982 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6983 Set *SEEN_REPL to true if we see a dead register that does have
6984 a replacement. */
6986 static bool
6987 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6988 bool *seen_repl)
6990 subrtx_iterator::array_type array;
6991 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6993 const_rtx x = *iter;
6994 if (is_dead_reg (x, counts))
6996 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6997 *seen_repl = true;
6998 else
6999 return true;
7002 return false;
7005 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
7006 Callback for simplify_replace_fn_rtx. */
7008 static rtx
7009 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
7011 rtx *replacements = (rtx *) data;
7013 if (REG_P (x)
7014 && REGNO (x) >= FIRST_PSEUDO_REGISTER
7015 && replacements[REGNO (x)] != NULL_RTX)
7017 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
7018 return replacements[REGNO (x)];
7019 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
7020 GET_MODE (replacements[REGNO (x)]));
7022 return NULL_RTX;
7025 /* Scan all the insns and delete any that are dead; i.e., they store a register
7026 that is never used or they copy a register to itself.
7028 This is used to remove insns made obviously dead by cse, loop or other
7029 optimizations. It improves the heuristics in loop since it won't try to
7030 move dead invariants out of loops or make givs for dead quantities. The
7031 remaining passes of the compilation are also sped up. */
7034 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
7036 int *counts;
7037 rtx_insn *insn, *prev;
7038 rtx *replacements = NULL;
7039 int ndead = 0;
7041 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7042 /* First count the number of times each register is used. */
7043 if (MAY_HAVE_DEBUG_INSNS)
7045 counts = XCNEWVEC (int, nreg * 3);
7046 for (insn = insns; insn; insn = NEXT_INSN (insn))
7047 if (DEBUG_INSN_P (insn))
7048 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7049 NULL_RTX, 1);
7050 else if (INSN_P (insn))
7052 count_reg_usage (insn, counts, NULL_RTX, 1);
7053 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7055 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7056 First one counts how many times each pseudo is used outside
7057 of debug insns, second counts how many times each pseudo is
7058 used in debug insns and third counts how many times a pseudo
7059 is stored. */
7061 else
7063 counts = XCNEWVEC (int, nreg);
7064 for (insn = insns; insn; insn = NEXT_INSN (insn))
7065 if (INSN_P (insn))
7066 count_reg_usage (insn, counts, NULL_RTX, 1);
7067 /* If no debug insns can be present, COUNTS is just an array
7068 which counts how many times each pseudo is used. */
7070 /* Pseudo PIC register should be considered as used due to possible
7071 new usages generated. */
7072 if (!reload_completed
7073 && pic_offset_table_rtx
7074 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7075 counts[REGNO (pic_offset_table_rtx)]++;
7076 /* Go from the last insn to the first and delete insns that only set unused
7077 registers or copy a register to itself. As we delete an insn, remove
7078 usage counts for registers it uses.
7080 The first jump optimization pass may leave a real insn as the last
7081 insn in the function. We must not skip that insn or we may end
7082 up deleting code that is not really dead.
7084 If some otherwise unused register is only used in DEBUG_INSNs,
7085 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7086 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7087 has been created for the unused register, replace it with
7088 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7089 for (insn = get_last_insn (); insn; insn = prev)
7091 int live_insn = 0;
7093 prev = PREV_INSN (insn);
7094 if (!INSN_P (insn))
7095 continue;
7097 live_insn = insn_live_p (insn, counts);
7099 /* If this is a dead insn, delete it and show registers in it aren't
7100 being used. */
7102 if (! live_insn && dbg_cnt (delete_trivial_dead))
7104 if (DEBUG_INSN_P (insn))
7105 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7106 NULL_RTX, -1);
7107 else
7109 rtx set;
7110 if (MAY_HAVE_DEBUG_INSNS
7111 && (set = single_set (insn)) != NULL_RTX
7112 && is_dead_reg (SET_DEST (set), counts)
7113 /* Used at least once in some DEBUG_INSN. */
7114 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7115 /* And set exactly once. */
7116 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7117 && !side_effects_p (SET_SRC (set))
7118 && asm_noperands (PATTERN (insn)) < 0)
7120 rtx dval, bind_var_loc;
7121 rtx_insn *bind;
7123 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7124 dval = make_debug_expr_from_rtl (SET_DEST (set));
7126 /* Emit a debug bind insn before the insn in which
7127 reg dies. */
7128 bind_var_loc =
7129 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7130 DEBUG_EXPR_TREE_DECL (dval),
7131 SET_SRC (set),
7132 VAR_INIT_STATUS_INITIALIZED);
7133 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7135 bind = emit_debug_insn_before (bind_var_loc, insn);
7136 df_insn_rescan (bind);
7138 if (replacements == NULL)
7139 replacements = XCNEWVEC (rtx, nreg);
7140 replacements[REGNO (SET_DEST (set))] = dval;
7143 count_reg_usage (insn, counts, NULL_RTX, -1);
7144 ndead++;
7146 cse_cfg_altered |= delete_insn_and_edges (insn);
7150 if (MAY_HAVE_DEBUG_INSNS)
7152 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7153 if (DEBUG_INSN_P (insn))
7155 /* If this debug insn references a dead register that wasn't replaced
7156 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7157 bool seen_repl = false;
7158 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7159 counts, replacements, &seen_repl))
7161 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7162 df_insn_rescan (insn);
7164 else if (seen_repl)
7166 INSN_VAR_LOCATION_LOC (insn)
7167 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7168 NULL_RTX, replace_dead_reg,
7169 replacements);
7170 df_insn_rescan (insn);
7173 free (replacements);
7176 if (dump_file && ndead)
7177 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7178 ndead);
7179 /* Clean up. */
7180 free (counts);
7181 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7182 return ndead;
7185 /* If LOC contains references to NEWREG in a different mode, change them
7186 to use NEWREG instead. */
7188 static void
7189 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7190 rtx *loc, rtx_insn *insn, rtx newreg)
7192 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7194 rtx *loc = *iter;
7195 rtx x = *loc;
7196 if (x
7197 && REG_P (x)
7198 && REGNO (x) == REGNO (newreg)
7199 && GET_MODE (x) != GET_MODE (newreg))
7201 validate_change (insn, loc, newreg, 1);
7202 iter.skip_subrtxes ();
7207 /* Change the mode of any reference to the register REGNO (NEWREG) to
7208 GET_MODE (NEWREG) in INSN. */
7210 static void
7211 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7213 int success;
7215 if (!INSN_P (insn))
7216 return;
7218 subrtx_ptr_iterator::array_type array;
7219 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7220 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7222 /* If the following assertion was triggered, there is most probably
7223 something wrong with the cc_modes_compatible back end function.
7224 CC modes only can be considered compatible if the insn - with the mode
7225 replaced by any of the compatible modes - can still be recognized. */
7226 success = apply_change_group ();
7227 gcc_assert (success);
7230 /* Change the mode of any reference to the register REGNO (NEWREG) to
7231 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7232 any instruction which modifies NEWREG. */
7234 static void
7235 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7237 rtx_insn *insn;
7239 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7241 if (! INSN_P (insn))
7242 continue;
7244 if (reg_set_p (newreg, insn))
7245 return;
7247 cse_change_cc_mode_insn (insn, newreg);
7251 /* BB is a basic block which finishes with CC_REG as a condition code
7252 register which is set to CC_SRC. Look through the successors of BB
7253 to find blocks which have a single predecessor (i.e., this one),
7254 and look through those blocks for an assignment to CC_REG which is
7255 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7256 permitted to change the mode of CC_SRC to a compatible mode. This
7257 returns VOIDmode if no equivalent assignments were found.
7258 Otherwise it returns the mode which CC_SRC should wind up with.
7259 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7260 but is passed unmodified down to recursive calls in order to prevent
7261 endless recursion.
7263 The main complexity in this function is handling the mode issues.
7264 We may have more than one duplicate which we can eliminate, and we
7265 try to find a mode which will work for multiple duplicates. */
7267 static machine_mode
7268 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7269 bool can_change_mode)
7271 bool found_equiv;
7272 machine_mode mode;
7273 unsigned int insn_count;
7274 edge e;
7275 rtx_insn *insns[2];
7276 machine_mode modes[2];
7277 rtx_insn *last_insns[2];
7278 unsigned int i;
7279 rtx newreg;
7280 edge_iterator ei;
7282 /* We expect to have two successors. Look at both before picking
7283 the final mode for the comparison. If we have more successors
7284 (i.e., some sort of table jump, although that seems unlikely),
7285 then we require all beyond the first two to use the same
7286 mode. */
7288 found_equiv = false;
7289 mode = GET_MODE (cc_src);
7290 insn_count = 0;
7291 FOR_EACH_EDGE (e, ei, bb->succs)
7293 rtx_insn *insn;
7294 rtx_insn *end;
7296 if (e->flags & EDGE_COMPLEX)
7297 continue;
7299 if (EDGE_COUNT (e->dest->preds) != 1
7300 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7301 /* Avoid endless recursion on unreachable blocks. */
7302 || e->dest == orig_bb)
7303 continue;
7305 end = NEXT_INSN (BB_END (e->dest));
7306 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7308 rtx set;
7310 if (! INSN_P (insn))
7311 continue;
7313 /* If CC_SRC is modified, we have to stop looking for
7314 something which uses it. */
7315 if (modified_in_p (cc_src, insn))
7316 break;
7318 /* Check whether INSN sets CC_REG to CC_SRC. */
7319 set = single_set (insn);
7320 if (set
7321 && REG_P (SET_DEST (set))
7322 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7324 bool found;
7325 machine_mode set_mode;
7326 machine_mode comp_mode;
7328 found = false;
7329 set_mode = GET_MODE (SET_SRC (set));
7330 comp_mode = set_mode;
7331 if (rtx_equal_p (cc_src, SET_SRC (set)))
7332 found = true;
7333 else if (GET_CODE (cc_src) == COMPARE
7334 && GET_CODE (SET_SRC (set)) == COMPARE
7335 && mode != set_mode
7336 && rtx_equal_p (XEXP (cc_src, 0),
7337 XEXP (SET_SRC (set), 0))
7338 && rtx_equal_p (XEXP (cc_src, 1),
7339 XEXP (SET_SRC (set), 1)))
7342 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7343 if (comp_mode != VOIDmode
7344 && (can_change_mode || comp_mode == mode))
7345 found = true;
7348 if (found)
7350 found_equiv = true;
7351 if (insn_count < ARRAY_SIZE (insns))
7353 insns[insn_count] = insn;
7354 modes[insn_count] = set_mode;
7355 last_insns[insn_count] = end;
7356 ++insn_count;
7358 if (mode != comp_mode)
7360 gcc_assert (can_change_mode);
7361 mode = comp_mode;
7363 /* The modified insn will be re-recognized later. */
7364 PUT_MODE (cc_src, mode);
7367 else
7369 if (set_mode != mode)
7371 /* We found a matching expression in the
7372 wrong mode, but we don't have room to
7373 store it in the array. Punt. This case
7374 should be rare. */
7375 break;
7377 /* INSN sets CC_REG to a value equal to CC_SRC
7378 with the right mode. We can simply delete
7379 it. */
7380 delete_insn (insn);
7383 /* We found an instruction to delete. Keep looking,
7384 in the hopes of finding a three-way jump. */
7385 continue;
7388 /* We found an instruction which sets the condition
7389 code, so don't look any farther. */
7390 break;
7393 /* If INSN sets CC_REG in some other way, don't look any
7394 farther. */
7395 if (reg_set_p (cc_reg, insn))
7396 break;
7399 /* If we fell off the bottom of the block, we can keep looking
7400 through successors. We pass CAN_CHANGE_MODE as false because
7401 we aren't prepared to handle compatibility between the
7402 further blocks and this block. */
7403 if (insn == end)
7405 machine_mode submode;
7407 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7408 if (submode != VOIDmode)
7410 gcc_assert (submode == mode);
7411 found_equiv = true;
7412 can_change_mode = false;
7417 if (! found_equiv)
7418 return VOIDmode;
7420 /* Now INSN_COUNT is the number of instructions we found which set
7421 CC_REG to a value equivalent to CC_SRC. The instructions are in
7422 INSNS. The modes used by those instructions are in MODES. */
7424 newreg = NULL_RTX;
7425 for (i = 0; i < insn_count; ++i)
7427 if (modes[i] != mode)
7429 /* We need to change the mode of CC_REG in INSNS[i] and
7430 subsequent instructions. */
7431 if (! newreg)
7433 if (GET_MODE (cc_reg) == mode)
7434 newreg = cc_reg;
7435 else
7436 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7438 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7439 newreg);
7442 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7445 return mode;
7448 /* If we have a fixed condition code register (or two), walk through
7449 the instructions and try to eliminate duplicate assignments. */
7451 static void
7452 cse_condition_code_reg (void)
7454 unsigned int cc_regno_1;
7455 unsigned int cc_regno_2;
7456 rtx cc_reg_1;
7457 rtx cc_reg_2;
7458 basic_block bb;
7460 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7461 return;
7463 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7464 if (cc_regno_2 != INVALID_REGNUM)
7465 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7466 else
7467 cc_reg_2 = NULL_RTX;
7469 FOR_EACH_BB_FN (bb, cfun)
7471 rtx_insn *last_insn;
7472 rtx cc_reg;
7473 rtx_insn *insn;
7474 rtx_insn *cc_src_insn;
7475 rtx cc_src;
7476 machine_mode mode;
7477 machine_mode orig_mode;
7479 /* Look for blocks which end with a conditional jump based on a
7480 condition code register. Then look for the instruction which
7481 sets the condition code register. Then look through the
7482 successor blocks for instructions which set the condition
7483 code register to the same value. There are other possible
7484 uses of the condition code register, but these are by far the
7485 most common and the ones which we are most likely to be able
7486 to optimize. */
7488 last_insn = BB_END (bb);
7489 if (!JUMP_P (last_insn))
7490 continue;
7492 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7493 cc_reg = cc_reg_1;
7494 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7495 cc_reg = cc_reg_2;
7496 else
7497 continue;
7499 cc_src_insn = NULL;
7500 cc_src = NULL_RTX;
7501 for (insn = PREV_INSN (last_insn);
7502 insn && insn != PREV_INSN (BB_HEAD (bb));
7503 insn = PREV_INSN (insn))
7505 rtx set;
7507 if (! INSN_P (insn))
7508 continue;
7509 set = single_set (insn);
7510 if (set
7511 && REG_P (SET_DEST (set))
7512 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7514 cc_src_insn = insn;
7515 cc_src = SET_SRC (set);
7516 break;
7518 else if (reg_set_p (cc_reg, insn))
7519 break;
7522 if (! cc_src_insn)
7523 continue;
7525 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7526 continue;
7528 /* Now CC_REG is a condition code register used for a
7529 conditional jump at the end of the block, and CC_SRC, in
7530 CC_SRC_INSN, is the value to which that condition code
7531 register is set, and CC_SRC is still meaningful at the end of
7532 the basic block. */
7534 orig_mode = GET_MODE (cc_src);
7535 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7536 if (mode != VOIDmode)
7538 gcc_assert (mode == GET_MODE (cc_src));
7539 if (mode != orig_mode)
7541 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7543 cse_change_cc_mode_insn (cc_src_insn, newreg);
7545 /* Do the same in the following insns that use the
7546 current value of CC_REG within BB. */
7547 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7548 NEXT_INSN (last_insn),
7549 newreg);
7556 /* Perform common subexpression elimination. Nonzero value from
7557 `cse_main' means that jumps were simplified and some code may now
7558 be unreachable, so do jump optimization again. */
7559 static unsigned int
7560 rest_of_handle_cse (void)
7562 int tem;
7564 if (dump_file)
7565 dump_flow_info (dump_file, dump_flags);
7567 tem = cse_main (get_insns (), max_reg_num ());
7569 /* If we are not running more CSE passes, then we are no longer
7570 expecting CSE to be run. But always rerun it in a cheap mode. */
7571 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7573 if (tem == 2)
7575 timevar_push (TV_JUMP);
7576 rebuild_jump_labels (get_insns ());
7577 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7578 timevar_pop (TV_JUMP);
7580 else if (tem == 1 || optimize > 1)
7581 cse_cfg_altered |= cleanup_cfg (0);
7583 return 0;
7586 namespace {
7588 const pass_data pass_data_cse =
7590 RTL_PASS, /* type */
7591 "cse1", /* name */
7592 OPTGROUP_NONE, /* optinfo_flags */
7593 TV_CSE, /* tv_id */
7594 0, /* properties_required */
7595 0, /* properties_provided */
7596 0, /* properties_destroyed */
7597 0, /* todo_flags_start */
7598 TODO_df_finish, /* todo_flags_finish */
7601 class pass_cse : public rtl_opt_pass
7603 public:
7604 pass_cse (gcc::context *ctxt)
7605 : rtl_opt_pass (pass_data_cse, ctxt)
7608 /* opt_pass methods: */
7609 virtual bool gate (function *) { return optimize > 0; }
7610 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7612 }; // class pass_cse
7614 } // anon namespace
7616 rtl_opt_pass *
7617 make_pass_cse (gcc::context *ctxt)
7619 return new pass_cse (ctxt);
7623 /* Run second CSE pass after loop optimizations. */
7624 static unsigned int
7625 rest_of_handle_cse2 (void)
7627 int tem;
7629 if (dump_file)
7630 dump_flow_info (dump_file, dump_flags);
7632 tem = cse_main (get_insns (), max_reg_num ());
7634 /* Run a pass to eliminate duplicated assignments to condition code
7635 registers. We have to run this after bypass_jumps, because it
7636 makes it harder for that pass to determine whether a jump can be
7637 bypassed safely. */
7638 cse_condition_code_reg ();
7640 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7642 if (tem == 2)
7644 timevar_push (TV_JUMP);
7645 rebuild_jump_labels (get_insns ());
7646 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7647 timevar_pop (TV_JUMP);
7649 else if (tem == 1)
7650 cse_cfg_altered |= cleanup_cfg (0);
7652 cse_not_expected = 1;
7653 return 0;
7657 namespace {
7659 const pass_data pass_data_cse2 =
7661 RTL_PASS, /* type */
7662 "cse2", /* name */
7663 OPTGROUP_NONE, /* optinfo_flags */
7664 TV_CSE2, /* tv_id */
7665 0, /* properties_required */
7666 0, /* properties_provided */
7667 0, /* properties_destroyed */
7668 0, /* todo_flags_start */
7669 TODO_df_finish, /* todo_flags_finish */
7672 class pass_cse2 : public rtl_opt_pass
7674 public:
7675 pass_cse2 (gcc::context *ctxt)
7676 : rtl_opt_pass (pass_data_cse2, ctxt)
7679 /* opt_pass methods: */
7680 virtual bool gate (function *)
7682 return optimize > 0 && flag_rerun_cse_after_loop;
7685 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7687 }; // class pass_cse2
7689 } // anon namespace
7691 rtl_opt_pass *
7692 make_pass_cse2 (gcc::context *ctxt)
7694 return new pass_cse2 (ctxt);
7697 /* Run second CSE pass after loop optimizations. */
7698 static unsigned int
7699 rest_of_handle_cse_after_global_opts (void)
7701 int save_cfj;
7702 int tem;
7704 /* We only want to do local CSE, so don't follow jumps. */
7705 save_cfj = flag_cse_follow_jumps;
7706 flag_cse_follow_jumps = 0;
7708 rebuild_jump_labels (get_insns ());
7709 tem = cse_main (get_insns (), max_reg_num ());
7710 cse_cfg_altered |= purge_all_dead_edges ();
7711 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7713 cse_not_expected = !flag_rerun_cse_after_loop;
7715 /* If cse altered any jumps, rerun jump opts to clean things up. */
7716 if (tem == 2)
7718 timevar_push (TV_JUMP);
7719 rebuild_jump_labels (get_insns ());
7720 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7721 timevar_pop (TV_JUMP);
7723 else if (tem == 1)
7724 cse_cfg_altered |= cleanup_cfg (0);
7726 flag_cse_follow_jumps = save_cfj;
7727 return 0;
7730 namespace {
7732 const pass_data pass_data_cse_after_global_opts =
7734 RTL_PASS, /* type */
7735 "cse_local", /* name */
7736 OPTGROUP_NONE, /* optinfo_flags */
7737 TV_CSE, /* tv_id */
7738 0, /* properties_required */
7739 0, /* properties_provided */
7740 0, /* properties_destroyed */
7741 0, /* todo_flags_start */
7742 TODO_df_finish, /* todo_flags_finish */
7745 class pass_cse_after_global_opts : public rtl_opt_pass
7747 public:
7748 pass_cse_after_global_opts (gcc::context *ctxt)
7749 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7752 /* opt_pass methods: */
7753 virtual bool gate (function *)
7755 return optimize > 0 && flag_rerun_cse_after_global_opts;
7758 virtual unsigned int execute (function *)
7760 return rest_of_handle_cse_after_global_opts ();
7763 }; // class pass_cse_after_global_opts
7765 } // anon namespace
7767 rtl_opt_pass *
7768 make_pass_cse_after_global_opts (gcc::context *ctxt)
7770 return new pass_cse_after_global_opts (ctxt);