PR rtl-optimization/82913
[official-gcc.git] / gcc / cse.c
blob65cc9ae110c1468fa194f587cae11f3fa424a442
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2017 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 scalar_int_mode int_mode, inner_mode;
724 return ((GET_CODE (x) == SUBREG
725 && REG_P (SUBREG_REG (x))
726 && is_int_mode (mode, &int_mode)
727 && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
728 && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
729 && subreg_lowpart_p (x)
730 && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
732 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
736 /* Initialize CSE_REG_INFO_TABLE. */
738 static void
739 init_cse_reg_info (unsigned int nregs)
741 /* Do we need to grow the table? */
742 if (nregs > cse_reg_info_table_size)
744 unsigned int new_size;
746 if (cse_reg_info_table_size < 2048)
748 /* Compute a new size that is a power of 2 and no smaller
749 than the large of NREGS and 64. */
750 new_size = (cse_reg_info_table_size
751 ? cse_reg_info_table_size : 64);
753 while (new_size < nregs)
754 new_size *= 2;
756 else
758 /* If we need a big table, allocate just enough to hold
759 NREGS registers. */
760 new_size = nregs;
763 /* Reallocate the table with NEW_SIZE entries. */
764 free (cse_reg_info_table);
765 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
766 cse_reg_info_table_size = new_size;
767 cse_reg_info_table_first_uninitialized = 0;
770 /* Do we have all of the first NREGS entries initialized? */
771 if (cse_reg_info_table_first_uninitialized < nregs)
773 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
774 unsigned int i;
776 /* Put the old timestamp on newly allocated entries so that they
777 will all be considered out of date. We do not touch those
778 entries beyond the first NREGS entries to be nice to the
779 virtual memory. */
780 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
781 cse_reg_info_table[i].timestamp = old_timestamp;
783 cse_reg_info_table_first_uninitialized = nregs;
787 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
789 static void
790 get_cse_reg_info_1 (unsigned int regno)
792 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
793 entry will be considered to have been initialized. */
794 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
796 /* Initialize the rest of the entry. */
797 cse_reg_info_table[regno].reg_tick = 1;
798 cse_reg_info_table[regno].reg_in_table = -1;
799 cse_reg_info_table[regno].subreg_ticked = -1;
800 cse_reg_info_table[regno].reg_qty = -regno - 1;
803 /* Find a cse_reg_info entry for REGNO. */
805 static inline struct cse_reg_info *
806 get_cse_reg_info (unsigned int regno)
808 struct cse_reg_info *p = &cse_reg_info_table[regno];
810 /* If this entry has not been initialized, go ahead and initialize
811 it. */
812 if (p->timestamp != cse_reg_info_timestamp)
813 get_cse_reg_info_1 (regno);
815 return p;
818 /* Clear the hash table and initialize each register with its own quantity,
819 for a new basic block. */
821 static void
822 new_basic_block (void)
824 int i;
826 next_qty = 0;
828 /* Invalidate cse_reg_info_table. */
829 cse_reg_info_timestamp++;
831 /* Clear out hash table state for this pass. */
832 CLEAR_HARD_REG_SET (hard_regs_in_table);
834 /* The per-quantity values used to be initialized here, but it is
835 much faster to initialize each as it is made in `make_new_qty'. */
837 for (i = 0; i < HASH_SIZE; i++)
839 struct table_elt *first;
841 first = table[i];
842 if (first != NULL)
844 struct table_elt *last = first;
846 table[i] = NULL;
848 while (last->next_same_hash != NULL)
849 last = last->next_same_hash;
851 /* Now relink this hash entire chain into
852 the free element list. */
854 last->next_same_hash = free_element_chain;
855 free_element_chain = first;
859 prev_insn_cc0 = 0;
862 /* Say that register REG contains a quantity in mode MODE not in any
863 register before and initialize that quantity. */
865 static void
866 make_new_qty (unsigned int reg, machine_mode mode)
868 int q;
869 struct qty_table_elem *ent;
870 struct reg_eqv_elem *eqv;
872 gcc_assert (next_qty < max_qty);
874 q = REG_QTY (reg) = next_qty++;
875 ent = &qty_table[q];
876 ent->first_reg = reg;
877 ent->last_reg = reg;
878 ent->mode = mode;
879 ent->const_rtx = ent->const_insn = NULL;
880 ent->comparison_code = UNKNOWN;
882 eqv = &reg_eqv_table[reg];
883 eqv->next = eqv->prev = -1;
886 /* Make reg NEW equivalent to reg OLD.
887 OLD is not changing; NEW is. */
889 static void
890 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
892 unsigned int lastr, firstr;
893 int q = REG_QTY (old_reg);
894 struct qty_table_elem *ent;
896 ent = &qty_table[q];
898 /* Nothing should become eqv until it has a "non-invalid" qty number. */
899 gcc_assert (REGNO_QTY_VALID_P (old_reg));
901 REG_QTY (new_reg) = q;
902 firstr = ent->first_reg;
903 lastr = ent->last_reg;
905 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
906 hard regs. Among pseudos, if NEW will live longer than any other reg
907 of the same qty, and that is beyond the current basic block,
908 make it the new canonical replacement for this qty. */
909 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
910 /* Certain fixed registers might be of the class NO_REGS. This means
911 that not only can they not be allocated by the compiler, but
912 they cannot be used in substitutions or canonicalizations
913 either. */
914 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
915 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
916 || (new_reg >= FIRST_PSEUDO_REGISTER
917 && (firstr < FIRST_PSEUDO_REGISTER
918 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
919 && !bitmap_bit_p (cse_ebb_live_out, firstr))
920 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
921 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
923 reg_eqv_table[firstr].prev = new_reg;
924 reg_eqv_table[new_reg].next = firstr;
925 reg_eqv_table[new_reg].prev = -1;
926 ent->first_reg = new_reg;
928 else
930 /* If NEW is a hard reg (known to be non-fixed), insert at end.
931 Otherwise, insert before any non-fixed hard regs that are at the
932 end. Registers of class NO_REGS cannot be used as an
933 equivalent for anything. */
934 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
935 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
936 && new_reg >= FIRST_PSEUDO_REGISTER)
937 lastr = reg_eqv_table[lastr].prev;
938 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
939 if (reg_eqv_table[lastr].next >= 0)
940 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
941 else
942 qty_table[q].last_reg = new_reg;
943 reg_eqv_table[lastr].next = new_reg;
944 reg_eqv_table[new_reg].prev = lastr;
948 /* Remove REG from its equivalence class. */
950 static void
951 delete_reg_equiv (unsigned int reg)
953 struct qty_table_elem *ent;
954 int q = REG_QTY (reg);
955 int p, n;
957 /* If invalid, do nothing. */
958 if (! REGNO_QTY_VALID_P (reg))
959 return;
961 ent = &qty_table[q];
963 p = reg_eqv_table[reg].prev;
964 n = reg_eqv_table[reg].next;
966 if (n != -1)
967 reg_eqv_table[n].prev = p;
968 else
969 ent->last_reg = p;
970 if (p != -1)
971 reg_eqv_table[p].next = n;
972 else
973 ent->first_reg = n;
975 REG_QTY (reg) = -reg - 1;
978 /* Remove any invalid expressions from the hash table
979 that refer to any of the registers contained in expression X.
981 Make sure that newly inserted references to those registers
982 as subexpressions will be considered valid.
984 mention_regs is not called when a register itself
985 is being stored in the table.
987 Return 1 if we have done something that may have changed the hash code
988 of X. */
990 static int
991 mention_regs (rtx x)
993 enum rtx_code code;
994 int i, j;
995 const char *fmt;
996 int changed = 0;
998 if (x == 0)
999 return 0;
1001 code = GET_CODE (x);
1002 if (code == REG)
1004 unsigned int regno = REGNO (x);
1005 unsigned int endregno = END_REGNO (x);
1006 unsigned int i;
1008 for (i = regno; i < endregno; i++)
1010 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1011 remove_invalid_refs (i);
1013 REG_IN_TABLE (i) = REG_TICK (i);
1014 SUBREG_TICKED (i) = -1;
1017 return 0;
1020 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1021 pseudo if they don't use overlapping words. We handle only pseudos
1022 here for simplicity. */
1023 if (code == SUBREG && REG_P (SUBREG_REG (x))
1024 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1026 unsigned int i = REGNO (SUBREG_REG (x));
1028 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1030 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1031 the last store to this register really stored into this
1032 subreg, then remove the memory of this subreg.
1033 Otherwise, remove any memory of the entire register and
1034 all its subregs from the table. */
1035 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1036 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1037 remove_invalid_refs (i);
1038 else
1039 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1042 REG_IN_TABLE (i) = REG_TICK (i);
1043 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1044 return 0;
1047 /* If X is a comparison or a COMPARE and either operand is a register
1048 that does not have a quantity, give it one. This is so that a later
1049 call to record_jump_equiv won't cause X to be assigned a different
1050 hash code and not found in the table after that call.
1052 It is not necessary to do this here, since rehash_using_reg can
1053 fix up the table later, but doing this here eliminates the need to
1054 call that expensive function in the most common case where the only
1055 use of the register is in the comparison. */
1057 if (code == COMPARE || COMPARISON_P (x))
1059 if (REG_P (XEXP (x, 0))
1060 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1061 if (insert_regs (XEXP (x, 0), NULL, 0))
1063 rehash_using_reg (XEXP (x, 0));
1064 changed = 1;
1067 if (REG_P (XEXP (x, 1))
1068 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1069 if (insert_regs (XEXP (x, 1), NULL, 0))
1071 rehash_using_reg (XEXP (x, 1));
1072 changed = 1;
1076 fmt = GET_RTX_FORMAT (code);
1077 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1078 if (fmt[i] == 'e')
1079 changed |= mention_regs (XEXP (x, i));
1080 else if (fmt[i] == 'E')
1081 for (j = 0; j < XVECLEN (x, i); j++)
1082 changed |= mention_regs (XVECEXP (x, i, j));
1084 return changed;
1087 /* Update the register quantities for inserting X into the hash table
1088 with a value equivalent to CLASSP.
1089 (If the class does not contain a REG, it is irrelevant.)
1090 If MODIFIED is nonzero, X is a destination; it is being modified.
1091 Note that delete_reg_equiv should be called on a register
1092 before insert_regs is done on that register with MODIFIED != 0.
1094 Nonzero value means that elements of reg_qty have changed
1095 so X's hash code may be different. */
1097 static int
1098 insert_regs (rtx x, struct table_elt *classp, int modified)
1100 if (REG_P (x))
1102 unsigned int regno = REGNO (x);
1103 int qty_valid;
1105 /* If REGNO is in the equivalence table already but is of the
1106 wrong mode for that equivalence, don't do anything here. */
1108 qty_valid = REGNO_QTY_VALID_P (regno);
1109 if (qty_valid)
1111 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1113 if (ent->mode != GET_MODE (x))
1114 return 0;
1117 if (modified || ! qty_valid)
1119 if (classp)
1120 for (classp = classp->first_same_value;
1121 classp != 0;
1122 classp = classp->next_same_value)
1123 if (REG_P (classp->exp)
1124 && GET_MODE (classp->exp) == GET_MODE (x))
1126 unsigned c_regno = REGNO (classp->exp);
1128 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1130 /* Suppose that 5 is hard reg and 100 and 101 are
1131 pseudos. Consider
1133 (set (reg:si 100) (reg:si 5))
1134 (set (reg:si 5) (reg:si 100))
1135 (set (reg:di 101) (reg:di 5))
1137 We would now set REG_QTY (101) = REG_QTY (5), but the
1138 entry for 5 is in SImode. When we use this later in
1139 copy propagation, we get the register in wrong mode. */
1140 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1141 continue;
1143 make_regs_eqv (regno, c_regno);
1144 return 1;
1147 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1148 than REG_IN_TABLE to find out if there was only a single preceding
1149 invalidation - for the SUBREG - or another one, which would be
1150 for the full register. However, if we find here that REG_TICK
1151 indicates that the register is invalid, it means that it has
1152 been invalidated in a separate operation. The SUBREG might be used
1153 now (then this is a recursive call), or we might use the full REG
1154 now and a SUBREG of it later. So bump up REG_TICK so that
1155 mention_regs will do the right thing. */
1156 if (! modified
1157 && REG_IN_TABLE (regno) >= 0
1158 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1159 REG_TICK (regno)++;
1160 make_new_qty (regno, GET_MODE (x));
1161 return 1;
1164 return 0;
1167 /* If X is a SUBREG, we will likely be inserting the inner register in the
1168 table. If that register doesn't have an assigned quantity number at
1169 this point but does later, the insertion that we will be doing now will
1170 not be accessible because its hash code will have changed. So assign
1171 a quantity number now. */
1173 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1174 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1176 insert_regs (SUBREG_REG (x), NULL, 0);
1177 mention_regs (x);
1178 return 1;
1180 else
1181 return mention_regs (x);
1185 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1186 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1187 CST is equal to an anchor. */
1189 static bool
1190 compute_const_anchors (rtx cst,
1191 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1192 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1194 HOST_WIDE_INT n = INTVAL (cst);
1196 *lower_base = n & ~(targetm.const_anchor - 1);
1197 if (*lower_base == n)
1198 return false;
1200 *upper_base =
1201 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1202 *upper_offs = n - *upper_base;
1203 *lower_offs = n - *lower_base;
1204 return true;
1207 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1209 static void
1210 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1211 machine_mode mode)
1213 struct table_elt *elt;
1214 unsigned hash;
1215 rtx anchor_exp;
1216 rtx exp;
1218 anchor_exp = GEN_INT (anchor);
1219 hash = HASH (anchor_exp, mode);
1220 elt = lookup (anchor_exp, hash, mode);
1221 if (!elt)
1222 elt = insert (anchor_exp, NULL, hash, mode);
1224 exp = plus_constant (mode, reg, offs);
1225 /* REG has just been inserted and the hash codes recomputed. */
1226 mention_regs (exp);
1227 hash = HASH (exp, mode);
1229 /* Use the cost of the register rather than the whole expression. When
1230 looking up constant anchors we will further offset the corresponding
1231 expression therefore it does not make sense to prefer REGs over
1232 reg-immediate additions. Prefer instead the oldest expression. Also
1233 don't prefer pseudos over hard regs so that we derive constants in
1234 argument registers from other argument registers rather than from the
1235 original pseudo that was used to synthesize the constant. */
1236 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1239 /* The constant CST is equivalent to the register REG. Create
1240 equivalences between the two anchors of CST and the corresponding
1241 register-offset expressions using REG. */
1243 static void
1244 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1246 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1248 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1249 &upper_base, &upper_offs))
1250 return;
1252 /* Ignore anchors of value 0. Constants accessible from zero are
1253 simple. */
1254 if (lower_base != 0)
1255 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1257 if (upper_base != 0)
1258 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1261 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1262 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1263 valid expression. Return the cheapest and oldest of such expressions. In
1264 *OLD, return how old the resulting expression is compared to the other
1265 equivalent expressions. */
1267 static rtx
1268 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1269 unsigned *old)
1271 struct table_elt *elt;
1272 unsigned idx;
1273 struct table_elt *match_elt;
1274 rtx match;
1276 /* Find the cheapest and *oldest* expression to maximize the chance of
1277 reusing the same pseudo. */
1279 match_elt = NULL;
1280 match = NULL_RTX;
1281 for (elt = anchor_elt->first_same_value, idx = 0;
1282 elt;
1283 elt = elt->next_same_value, idx++)
1285 if (match_elt && CHEAPER (match_elt, elt))
1286 return match;
1288 if (REG_P (elt->exp)
1289 || (GET_CODE (elt->exp) == PLUS
1290 && REG_P (XEXP (elt->exp, 0))
1291 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1293 rtx x;
1295 /* Ignore expressions that are no longer valid. */
1296 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1297 continue;
1299 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1300 if (REG_P (x)
1301 || (GET_CODE (x) == PLUS
1302 && IN_RANGE (INTVAL (XEXP (x, 1)),
1303 -targetm.const_anchor,
1304 targetm.const_anchor - 1)))
1306 match = x;
1307 match_elt = elt;
1308 *old = idx;
1313 return match;
1316 /* Try to express the constant SRC_CONST using a register+offset expression
1317 derived from a constant anchor. Return it if successful or NULL_RTX,
1318 otherwise. */
1320 static rtx
1321 try_const_anchors (rtx src_const, machine_mode mode)
1323 struct table_elt *lower_elt, *upper_elt;
1324 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1325 rtx lower_anchor_rtx, upper_anchor_rtx;
1326 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1327 unsigned lower_old, upper_old;
1329 /* CONST_INT is used for CC modes, but we should leave those alone. */
1330 if (GET_MODE_CLASS (mode) == MODE_CC)
1331 return NULL_RTX;
1333 gcc_assert (SCALAR_INT_MODE_P (mode));
1334 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1335 &upper_base, &upper_offs))
1336 return NULL_RTX;
1338 lower_anchor_rtx = GEN_INT (lower_base);
1339 upper_anchor_rtx = GEN_INT (upper_base);
1340 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1341 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1343 if (lower_elt)
1344 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1345 if (upper_elt)
1346 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1348 if (!lower_exp)
1349 return upper_exp;
1350 if (!upper_exp)
1351 return lower_exp;
1353 /* Return the older expression. */
1354 return (upper_old > lower_old ? upper_exp : lower_exp);
1357 /* Look in or update the hash table. */
1359 /* Remove table element ELT from use in the table.
1360 HASH is its hash code, made using the HASH macro.
1361 It's an argument because often that is known in advance
1362 and we save much time not recomputing it. */
1364 static void
1365 remove_from_table (struct table_elt *elt, unsigned int hash)
1367 if (elt == 0)
1368 return;
1370 /* Mark this element as removed. See cse_insn. */
1371 elt->first_same_value = 0;
1373 /* Remove the table element from its equivalence class. */
1376 struct table_elt *prev = elt->prev_same_value;
1377 struct table_elt *next = elt->next_same_value;
1379 if (next)
1380 next->prev_same_value = prev;
1382 if (prev)
1383 prev->next_same_value = next;
1384 else
1386 struct table_elt *newfirst = next;
1387 while (next)
1389 next->first_same_value = newfirst;
1390 next = next->next_same_value;
1395 /* Remove the table element from its hash bucket. */
1398 struct table_elt *prev = elt->prev_same_hash;
1399 struct table_elt *next = elt->next_same_hash;
1401 if (next)
1402 next->prev_same_hash = prev;
1404 if (prev)
1405 prev->next_same_hash = next;
1406 else if (table[hash] == elt)
1407 table[hash] = next;
1408 else
1410 /* This entry is not in the proper hash bucket. This can happen
1411 when two classes were merged by `merge_equiv_classes'. Search
1412 for the hash bucket that it heads. This happens only very
1413 rarely, so the cost is acceptable. */
1414 for (hash = 0; hash < HASH_SIZE; hash++)
1415 if (table[hash] == elt)
1416 table[hash] = next;
1420 /* Remove the table element from its related-value circular chain. */
1422 if (elt->related_value != 0 && elt->related_value != elt)
1424 struct table_elt *p = elt->related_value;
1426 while (p->related_value != elt)
1427 p = p->related_value;
1428 p->related_value = elt->related_value;
1429 if (p->related_value == p)
1430 p->related_value = 0;
1433 /* Now add it to the free element chain. */
1434 elt->next_same_hash = free_element_chain;
1435 free_element_chain = elt;
1438 /* Same as above, but X is a pseudo-register. */
1440 static void
1441 remove_pseudo_from_table (rtx x, unsigned int hash)
1443 struct table_elt *elt;
1445 /* Because a pseudo-register can be referenced in more than one
1446 mode, we might have to remove more than one table entry. */
1447 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1448 remove_from_table (elt, hash);
1451 /* Look up X in the hash table and return its table element,
1452 or 0 if X is not in the table.
1454 MODE is the machine-mode of X, or if X is an integer constant
1455 with VOIDmode then MODE is the mode with which X will be used.
1457 Here we are satisfied to find an expression whose tree structure
1458 looks like X. */
1460 static struct table_elt *
1461 lookup (rtx x, unsigned int hash, machine_mode mode)
1463 struct table_elt *p;
1465 for (p = table[hash]; p; p = p->next_same_hash)
1466 if (mode == p->mode && ((x == p->exp && REG_P (x))
1467 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1468 return p;
1470 return 0;
1473 /* Like `lookup' but don't care whether the table element uses invalid regs.
1474 Also ignore discrepancies in the machine mode of a register. */
1476 static struct table_elt *
1477 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1479 struct table_elt *p;
1481 if (REG_P (x))
1483 unsigned int regno = REGNO (x);
1485 /* Don't check the machine mode when comparing registers;
1486 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1487 for (p = table[hash]; p; p = p->next_same_hash)
1488 if (REG_P (p->exp)
1489 && REGNO (p->exp) == regno)
1490 return p;
1492 else
1494 for (p = table[hash]; p; p = p->next_same_hash)
1495 if (mode == p->mode
1496 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1497 return p;
1500 return 0;
1503 /* Look for an expression equivalent to X and with code CODE.
1504 If one is found, return that expression. */
1506 static rtx
1507 lookup_as_function (rtx x, enum rtx_code code)
1509 struct table_elt *p
1510 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1512 if (p == 0)
1513 return 0;
1515 for (p = p->first_same_value; p; p = p->next_same_value)
1516 if (GET_CODE (p->exp) == code
1517 /* Make sure this is a valid entry in the table. */
1518 && exp_equiv_p (p->exp, p->exp, 1, false))
1519 return p->exp;
1521 return 0;
1524 /* Insert X in the hash table, assuming HASH is its hash code and
1525 CLASSP is an element of the class it should go in (or 0 if a new
1526 class should be made). COST is the code of X and reg_cost is the
1527 cost of registers in X. It is inserted at the proper position to
1528 keep the class in the order cheapest first.
1530 MODE is the machine-mode of X, or if X is an integer constant
1531 with VOIDmode then MODE is the mode with which X will be used.
1533 For elements of equal cheapness, the most recent one
1534 goes in front, except that the first element in the list
1535 remains first unless a cheaper element is added. The order of
1536 pseudo-registers does not matter, as canon_reg will be called to
1537 find the cheapest when a register is retrieved from the table.
1539 The in_memory field in the hash table element is set to 0.
1540 The caller must set it nonzero if appropriate.
1542 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1543 and if insert_regs returns a nonzero value
1544 you must then recompute its hash code before calling here.
1546 If necessary, update table showing constant values of quantities. */
1548 static struct table_elt *
1549 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1550 machine_mode mode, int cost, int reg_cost)
1552 struct table_elt *elt;
1554 /* If X is a register and we haven't made a quantity for it,
1555 something is wrong. */
1556 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1558 /* If X is a hard register, show it is being put in the table. */
1559 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1560 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1562 /* Put an element for X into the right hash bucket. */
1564 elt = free_element_chain;
1565 if (elt)
1566 free_element_chain = elt->next_same_hash;
1567 else
1568 elt = XNEW (struct table_elt);
1570 elt->exp = x;
1571 elt->canon_exp = NULL_RTX;
1572 elt->cost = cost;
1573 elt->regcost = reg_cost;
1574 elt->next_same_value = 0;
1575 elt->prev_same_value = 0;
1576 elt->next_same_hash = table[hash];
1577 elt->prev_same_hash = 0;
1578 elt->related_value = 0;
1579 elt->in_memory = 0;
1580 elt->mode = mode;
1581 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1583 if (table[hash])
1584 table[hash]->prev_same_hash = elt;
1585 table[hash] = elt;
1587 /* Put it into the proper value-class. */
1588 if (classp)
1590 classp = classp->first_same_value;
1591 if (CHEAPER (elt, classp))
1592 /* Insert at the head of the class. */
1594 struct table_elt *p;
1595 elt->next_same_value = classp;
1596 classp->prev_same_value = elt;
1597 elt->first_same_value = elt;
1599 for (p = classp; p; p = p->next_same_value)
1600 p->first_same_value = elt;
1602 else
1604 /* Insert not at head of the class. */
1605 /* Put it after the last element cheaper than X. */
1606 struct table_elt *p, *next;
1608 for (p = classp;
1609 (next = p->next_same_value) && CHEAPER (next, elt);
1610 p = next)
1613 /* Put it after P and before NEXT. */
1614 elt->next_same_value = next;
1615 if (next)
1616 next->prev_same_value = elt;
1618 elt->prev_same_value = p;
1619 p->next_same_value = elt;
1620 elt->first_same_value = classp;
1623 else
1624 elt->first_same_value = elt;
1626 /* If this is a constant being set equivalent to a register or a register
1627 being set equivalent to a constant, note the constant equivalence.
1629 If this is a constant, it cannot be equivalent to a different constant,
1630 and a constant is the only thing that can be cheaper than a register. So
1631 we know the register is the head of the class (before the constant was
1632 inserted).
1634 If this is a register that is not already known equivalent to a
1635 constant, we must check the entire class.
1637 If this is a register that is already known equivalent to an insn,
1638 update the qtys `const_insn' to show that `this_insn' is the latest
1639 insn making that quantity equivalent to the constant. */
1641 if (elt->is_const && classp && REG_P (classp->exp)
1642 && !REG_P (x))
1644 int exp_q = REG_QTY (REGNO (classp->exp));
1645 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1647 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1648 exp_ent->const_insn = this_insn;
1651 else if (REG_P (x)
1652 && classp
1653 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1654 && ! elt->is_const)
1656 struct table_elt *p;
1658 for (p = classp; p != 0; p = p->next_same_value)
1660 if (p->is_const && !REG_P (p->exp))
1662 int x_q = REG_QTY (REGNO (x));
1663 struct qty_table_elem *x_ent = &qty_table[x_q];
1665 x_ent->const_rtx
1666 = gen_lowpart (GET_MODE (x), p->exp);
1667 x_ent->const_insn = this_insn;
1668 break;
1673 else if (REG_P (x)
1674 && qty_table[REG_QTY (REGNO (x))].const_rtx
1675 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1676 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1678 /* If this is a constant with symbolic value,
1679 and it has a term with an explicit integer value,
1680 link it up with related expressions. */
1681 if (GET_CODE (x) == CONST)
1683 rtx subexp = get_related_value (x);
1684 unsigned subhash;
1685 struct table_elt *subelt, *subelt_prev;
1687 if (subexp != 0)
1689 /* Get the integer-free subexpression in the hash table. */
1690 subhash = SAFE_HASH (subexp, mode);
1691 subelt = lookup (subexp, subhash, mode);
1692 if (subelt == 0)
1693 subelt = insert (subexp, NULL, subhash, mode);
1694 /* Initialize SUBELT's circular chain if it has none. */
1695 if (subelt->related_value == 0)
1696 subelt->related_value = subelt;
1697 /* Find the element in the circular chain that precedes SUBELT. */
1698 subelt_prev = subelt;
1699 while (subelt_prev->related_value != subelt)
1700 subelt_prev = subelt_prev->related_value;
1701 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1702 This way the element that follows SUBELT is the oldest one. */
1703 elt->related_value = subelt_prev->related_value;
1704 subelt_prev->related_value = elt;
1708 return elt;
1711 /* Wrap insert_with_costs by passing the default costs. */
1713 static struct table_elt *
1714 insert (rtx x, struct table_elt *classp, unsigned int hash,
1715 machine_mode mode)
1717 return insert_with_costs (x, classp, hash, mode,
1718 COST (x, mode), approx_reg_cost (x));
1722 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1723 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1724 the two classes equivalent.
1726 CLASS1 will be the surviving class; CLASS2 should not be used after this
1727 call.
1729 Any invalid entries in CLASS2 will not be copied. */
1731 static void
1732 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1734 struct table_elt *elt, *next, *new_elt;
1736 /* Ensure we start with the head of the classes. */
1737 class1 = class1->first_same_value;
1738 class2 = class2->first_same_value;
1740 /* If they were already equal, forget it. */
1741 if (class1 == class2)
1742 return;
1744 for (elt = class2; elt; elt = next)
1746 unsigned int hash;
1747 rtx exp = elt->exp;
1748 machine_mode mode = elt->mode;
1750 next = elt->next_same_value;
1752 /* Remove old entry, make a new one in CLASS1's class.
1753 Don't do this for invalid entries as we cannot find their
1754 hash code (it also isn't necessary). */
1755 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1757 bool need_rehash = false;
1759 hash_arg_in_memory = 0;
1760 hash = HASH (exp, mode);
1762 if (REG_P (exp))
1764 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1765 delete_reg_equiv (REGNO (exp));
1768 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1769 remove_pseudo_from_table (exp, hash);
1770 else
1771 remove_from_table (elt, hash);
1773 if (insert_regs (exp, class1, 0) || need_rehash)
1775 rehash_using_reg (exp);
1776 hash = HASH (exp, mode);
1778 new_elt = insert (exp, class1, hash, mode);
1779 new_elt->in_memory = hash_arg_in_memory;
1780 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1781 new_elt->cost = MAX_COST;
1786 /* Flush the entire hash table. */
1788 static void
1789 flush_hash_table (void)
1791 int i;
1792 struct table_elt *p;
1794 for (i = 0; i < HASH_SIZE; i++)
1795 for (p = table[i]; p; p = table[i])
1797 /* Note that invalidate can remove elements
1798 after P in the current hash chain. */
1799 if (REG_P (p->exp))
1800 invalidate (p->exp, VOIDmode);
1801 else
1802 remove_from_table (p, i);
1806 /* Check whether an anti dependence exists between X and EXP. MODE and
1807 ADDR are as for canon_anti_dependence. */
1809 static bool
1810 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1812 subrtx_iterator::array_type array;
1813 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1815 const_rtx x = *iter;
1816 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1817 return true;
1819 return false;
1822 /* Remove from the hash table, or mark as invalid, all expressions whose
1823 values could be altered by storing in X. X is a register, a subreg, or
1824 a memory reference with nonvarying address (because, when a memory
1825 reference with a varying address is stored in, all memory references are
1826 removed by invalidate_memory so specific invalidation is superfluous).
1827 FULL_MODE, if not VOIDmode, indicates that this much should be
1828 invalidated instead of just the amount indicated by the mode of X. This
1829 is only used for bitfield stores into memory.
1831 A nonvarying address may be just a register or just a symbol reference,
1832 or it may be either of those plus a numeric offset. */
1834 static void
1835 invalidate (rtx x, machine_mode full_mode)
1837 int i;
1838 struct table_elt *p;
1839 rtx addr;
1841 switch (GET_CODE (x))
1843 case REG:
1845 /* If X is a register, dependencies on its contents are recorded
1846 through the qty number mechanism. Just change the qty number of
1847 the register, mark it as invalid for expressions that refer to it,
1848 and remove it itself. */
1849 unsigned int regno = REGNO (x);
1850 unsigned int hash = HASH (x, GET_MODE (x));
1852 /* Remove REGNO from any quantity list it might be on and indicate
1853 that its value might have changed. If it is a pseudo, remove its
1854 entry from the hash table.
1856 For a hard register, we do the first two actions above for any
1857 additional hard registers corresponding to X. Then, if any of these
1858 registers are in the table, we must remove any REG entries that
1859 overlap these registers. */
1861 delete_reg_equiv (regno);
1862 REG_TICK (regno)++;
1863 SUBREG_TICKED (regno) = -1;
1865 if (regno >= FIRST_PSEUDO_REGISTER)
1866 remove_pseudo_from_table (x, hash);
1867 else
1869 HOST_WIDE_INT in_table
1870 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1871 unsigned int endregno = END_REGNO (x);
1872 unsigned int tregno, tendregno, rn;
1873 struct table_elt *p, *next;
1875 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1877 for (rn = regno + 1; rn < endregno; rn++)
1879 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1880 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1881 delete_reg_equiv (rn);
1882 REG_TICK (rn)++;
1883 SUBREG_TICKED (rn) = -1;
1886 if (in_table)
1887 for (hash = 0; hash < HASH_SIZE; hash++)
1888 for (p = table[hash]; p; p = next)
1890 next = p->next_same_hash;
1892 if (!REG_P (p->exp)
1893 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1894 continue;
1896 tregno = REGNO (p->exp);
1897 tendregno = END_REGNO (p->exp);
1898 if (tendregno > regno && tregno < endregno)
1899 remove_from_table (p, hash);
1903 return;
1905 case SUBREG:
1906 invalidate (SUBREG_REG (x), VOIDmode);
1907 return;
1909 case PARALLEL:
1910 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1911 invalidate (XVECEXP (x, 0, i), VOIDmode);
1912 return;
1914 case EXPR_LIST:
1915 /* This is part of a disjoint return value; extract the location in
1916 question ignoring the offset. */
1917 invalidate (XEXP (x, 0), VOIDmode);
1918 return;
1920 case MEM:
1921 addr = canon_rtx (get_addr (XEXP (x, 0)));
1922 /* Calculate the canonical version of X here so that
1923 true_dependence doesn't generate new RTL for X on each call. */
1924 x = canon_rtx (x);
1926 /* Remove all hash table elements that refer to overlapping pieces of
1927 memory. */
1928 if (full_mode == VOIDmode)
1929 full_mode = GET_MODE (x);
1931 for (i = 0; i < HASH_SIZE; i++)
1933 struct table_elt *next;
1935 for (p = table[i]; p; p = next)
1937 next = p->next_same_hash;
1938 if (p->in_memory)
1940 /* Just canonicalize the expression once;
1941 otherwise each time we call invalidate
1942 true_dependence will canonicalize the
1943 expression again. */
1944 if (!p->canon_exp)
1945 p->canon_exp = canon_rtx (p->exp);
1946 if (check_dependence (p->canon_exp, x, full_mode, addr))
1947 remove_from_table (p, i);
1951 return;
1953 default:
1954 gcc_unreachable ();
1958 /* Invalidate DEST. Used when DEST is not going to be added
1959 into the hash table for some reason, e.g. do_not_record
1960 flagged on it. */
1962 static void
1963 invalidate_dest (rtx dest)
1965 if (REG_P (dest)
1966 || GET_CODE (dest) == SUBREG
1967 || MEM_P (dest))
1968 invalidate (dest, VOIDmode);
1969 else if (GET_CODE (dest) == STRICT_LOW_PART
1970 || GET_CODE (dest) == ZERO_EXTRACT)
1971 invalidate (XEXP (dest, 0), GET_MODE (dest));
1974 /* Remove all expressions that refer to register REGNO,
1975 since they are already invalid, and we are about to
1976 mark that register valid again and don't want the old
1977 expressions to reappear as valid. */
1979 static void
1980 remove_invalid_refs (unsigned int regno)
1982 unsigned int i;
1983 struct table_elt *p, *next;
1985 for (i = 0; i < HASH_SIZE; i++)
1986 for (p = table[i]; p; p = next)
1988 next = p->next_same_hash;
1989 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1990 remove_from_table (p, i);
1994 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1995 and mode MODE. */
1996 static void
1997 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
1998 machine_mode mode)
2000 unsigned int i;
2001 struct table_elt *p, *next;
2002 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2004 for (i = 0; i < HASH_SIZE; i++)
2005 for (p = table[i]; p; p = next)
2007 rtx exp = p->exp;
2008 next = p->next_same_hash;
2010 if (!REG_P (exp)
2011 && (GET_CODE (exp) != SUBREG
2012 || !REG_P (SUBREG_REG (exp))
2013 || REGNO (SUBREG_REG (exp)) != regno
2014 || (((SUBREG_BYTE (exp)
2015 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2016 && SUBREG_BYTE (exp) <= end))
2017 && refers_to_regno_p (regno, p->exp))
2018 remove_from_table (p, i);
2022 /* Recompute the hash codes of any valid entries in the hash table that
2023 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2025 This is called when we make a jump equivalence. */
2027 static void
2028 rehash_using_reg (rtx x)
2030 unsigned int i;
2031 struct table_elt *p, *next;
2032 unsigned hash;
2034 if (GET_CODE (x) == SUBREG)
2035 x = SUBREG_REG (x);
2037 /* If X is not a register or if the register is known not to be in any
2038 valid entries in the table, we have no work to do. */
2040 if (!REG_P (x)
2041 || REG_IN_TABLE (REGNO (x)) < 0
2042 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2043 return;
2045 /* Scan all hash chains looking for valid entries that mention X.
2046 If we find one and it is in the wrong hash chain, move it. */
2048 for (i = 0; i < HASH_SIZE; i++)
2049 for (p = table[i]; p; p = next)
2051 next = p->next_same_hash;
2052 if (reg_mentioned_p (x, p->exp)
2053 && exp_equiv_p (p->exp, p->exp, 1, false)
2054 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2056 if (p->next_same_hash)
2057 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2059 if (p->prev_same_hash)
2060 p->prev_same_hash->next_same_hash = p->next_same_hash;
2061 else
2062 table[i] = p->next_same_hash;
2064 p->next_same_hash = table[hash];
2065 p->prev_same_hash = 0;
2066 if (table[hash])
2067 table[hash]->prev_same_hash = p;
2068 table[hash] = p;
2073 /* Remove from the hash table any expression that is a call-clobbered
2074 register. Also update their TICK values. */
2076 static void
2077 invalidate_for_call (void)
2079 unsigned int regno, endregno;
2080 unsigned int i;
2081 unsigned hash;
2082 struct table_elt *p, *next;
2083 int in_table = 0;
2084 hard_reg_set_iterator hrsi;
2086 /* Go through all the hard registers. For each that is clobbered in
2087 a CALL_INSN, remove the register from quantity chains and update
2088 reg_tick if defined. Also see if any of these registers is currently
2089 in the table. */
2090 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2092 delete_reg_equiv (regno);
2093 if (REG_TICK (regno) >= 0)
2095 REG_TICK (regno)++;
2096 SUBREG_TICKED (regno) = -1;
2098 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2101 /* In the case where we have no call-clobbered hard registers in the
2102 table, we are done. Otherwise, scan the table and remove any
2103 entry that overlaps a call-clobbered register. */
2105 if (in_table)
2106 for (hash = 0; hash < HASH_SIZE; hash++)
2107 for (p = table[hash]; p; p = next)
2109 next = p->next_same_hash;
2111 if (!REG_P (p->exp)
2112 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2113 continue;
2115 regno = REGNO (p->exp);
2116 endregno = END_REGNO (p->exp);
2118 for (i = regno; i < endregno; i++)
2119 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2121 remove_from_table (p, hash);
2122 break;
2127 /* Given an expression X of type CONST,
2128 and ELT which is its table entry (or 0 if it
2129 is not in the hash table),
2130 return an alternate expression for X as a register plus integer.
2131 If none can be found, return 0. */
2133 static rtx
2134 use_related_value (rtx x, struct table_elt *elt)
2136 struct table_elt *relt = 0;
2137 struct table_elt *p, *q;
2138 HOST_WIDE_INT offset;
2140 /* First, is there anything related known?
2141 If we have a table element, we can tell from that.
2142 Otherwise, must look it up. */
2144 if (elt != 0 && elt->related_value != 0)
2145 relt = elt;
2146 else if (elt == 0 && GET_CODE (x) == CONST)
2148 rtx subexp = get_related_value (x);
2149 if (subexp != 0)
2150 relt = lookup (subexp,
2151 SAFE_HASH (subexp, GET_MODE (subexp)),
2152 GET_MODE (subexp));
2155 if (relt == 0)
2156 return 0;
2158 /* Search all related table entries for one that has an
2159 equivalent register. */
2161 p = relt;
2162 while (1)
2164 /* This loop is strange in that it is executed in two different cases.
2165 The first is when X is already in the table. Then it is searching
2166 the RELATED_VALUE list of X's class (RELT). The second case is when
2167 X is not in the table. Then RELT points to a class for the related
2168 value.
2170 Ensure that, whatever case we are in, that we ignore classes that have
2171 the same value as X. */
2173 if (rtx_equal_p (x, p->exp))
2174 q = 0;
2175 else
2176 for (q = p->first_same_value; q; q = q->next_same_value)
2177 if (REG_P (q->exp))
2178 break;
2180 if (q)
2181 break;
2183 p = p->related_value;
2185 /* We went all the way around, so there is nothing to be found.
2186 Alternatively, perhaps RELT was in the table for some other reason
2187 and it has no related values recorded. */
2188 if (p == relt || p == 0)
2189 break;
2192 if (q == 0)
2193 return 0;
2195 offset = (get_integer_term (x) - get_integer_term (p->exp));
2196 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2197 return plus_constant (q->mode, q->exp, offset);
2201 /* Hash a string. Just add its bytes up. */
2202 static inline unsigned
2203 hash_rtx_string (const char *ps)
2205 unsigned hash = 0;
2206 const unsigned char *p = (const unsigned char *) ps;
2208 if (p)
2209 while (*p)
2210 hash += *p++;
2212 return hash;
2215 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2216 When the callback returns true, we continue with the new rtx. */
2218 unsigned
2219 hash_rtx_cb (const_rtx x, machine_mode mode,
2220 int *do_not_record_p, int *hash_arg_in_memory_p,
2221 bool have_reg_qty, hash_rtx_callback_function cb)
2223 int i, j;
2224 unsigned hash = 0;
2225 enum rtx_code code;
2226 const char *fmt;
2227 machine_mode newmode;
2228 rtx newx;
2230 /* Used to turn recursion into iteration. We can't rely on GCC's
2231 tail-recursion elimination since we need to keep accumulating values
2232 in HASH. */
2233 repeat:
2234 if (x == 0)
2235 return hash;
2237 /* Invoke the callback first. */
2238 if (cb != NULL
2239 && ((*cb) (x, mode, &newx, &newmode)))
2241 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2242 hash_arg_in_memory_p, have_reg_qty, cb);
2243 return hash;
2246 code = GET_CODE (x);
2247 switch (code)
2249 case REG:
2251 unsigned int regno = REGNO (x);
2253 if (do_not_record_p && !reload_completed)
2255 /* On some machines, we can't record any non-fixed hard register,
2256 because extending its life will cause reload problems. We
2257 consider ap, fp, sp, gp to be fixed for this purpose.
2259 We also consider CCmode registers to be fixed for this purpose;
2260 failure to do so leads to failure to simplify 0<100 type of
2261 conditionals.
2263 On all machines, we can't record any global registers.
2264 Nor should we record any register that is in a small
2265 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2266 bool record;
2268 if (regno >= FIRST_PSEUDO_REGISTER)
2269 record = true;
2270 else if (x == frame_pointer_rtx
2271 || x == hard_frame_pointer_rtx
2272 || x == arg_pointer_rtx
2273 || x == stack_pointer_rtx
2274 || x == pic_offset_table_rtx)
2275 record = true;
2276 else if (global_regs[regno])
2277 record = false;
2278 else if (fixed_regs[regno])
2279 record = true;
2280 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2281 record = true;
2282 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2283 record = false;
2284 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2285 record = false;
2286 else
2287 record = true;
2289 if (!record)
2291 *do_not_record_p = 1;
2292 return 0;
2296 hash += ((unsigned int) REG << 7);
2297 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2298 return hash;
2301 /* We handle SUBREG of a REG specially because the underlying
2302 reg changes its hash value with every value change; we don't
2303 want to have to forget unrelated subregs when one subreg changes. */
2304 case SUBREG:
2306 if (REG_P (SUBREG_REG (x)))
2308 hash += (((unsigned int) SUBREG << 7)
2309 + REGNO (SUBREG_REG (x))
2310 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2311 return hash;
2313 break;
2316 case CONST_INT:
2317 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2318 + (unsigned int) INTVAL (x));
2319 return hash;
2321 case CONST_WIDE_INT:
2322 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2323 hash += CONST_WIDE_INT_ELT (x, i);
2324 return hash;
2326 case CONST_DOUBLE:
2327 /* This is like the general case, except that it only counts
2328 the integers representing the constant. */
2329 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2330 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2331 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2332 + (unsigned int) CONST_DOUBLE_HIGH (x));
2333 else
2334 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2335 return hash;
2337 case CONST_FIXED:
2338 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2339 hash += fixed_hash (CONST_FIXED_VALUE (x));
2340 return hash;
2342 case CONST_VECTOR:
2344 int units;
2345 rtx elt;
2347 units = CONST_VECTOR_NUNITS (x);
2349 for (i = 0; i < units; ++i)
2351 elt = CONST_VECTOR_ELT (x, i);
2352 hash += hash_rtx_cb (elt, GET_MODE (elt),
2353 do_not_record_p, hash_arg_in_memory_p,
2354 have_reg_qty, cb);
2357 return hash;
2360 /* Assume there is only one rtx object for any given label. */
2361 case LABEL_REF:
2362 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2363 differences and differences between each stage's debugging dumps. */
2364 hash += (((unsigned int) LABEL_REF << 7)
2365 + CODE_LABEL_NUMBER (label_ref_label (x)));
2366 return hash;
2368 case SYMBOL_REF:
2370 /* Don't hash on the symbol's address to avoid bootstrap differences.
2371 Different hash values may cause expressions to be recorded in
2372 different orders and thus different registers to be used in the
2373 final assembler. This also avoids differences in the dump files
2374 between various stages. */
2375 unsigned int h = 0;
2376 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2378 while (*p)
2379 h += (h << 7) + *p++; /* ??? revisit */
2381 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2382 return hash;
2385 case MEM:
2386 /* We don't record if marked volatile or if BLKmode since we don't
2387 know the size of the move. */
2388 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2390 *do_not_record_p = 1;
2391 return 0;
2393 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2394 *hash_arg_in_memory_p = 1;
2396 /* Now that we have already found this special case,
2397 might as well speed it up as much as possible. */
2398 hash += (unsigned) MEM;
2399 x = XEXP (x, 0);
2400 goto repeat;
2402 case USE:
2403 /* A USE that mentions non-volatile memory needs special
2404 handling since the MEM may be BLKmode which normally
2405 prevents an entry from being made. Pure calls are
2406 marked by a USE which mentions BLKmode memory.
2407 See calls.c:emit_call_1. */
2408 if (MEM_P (XEXP (x, 0))
2409 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2411 hash += (unsigned) USE;
2412 x = XEXP (x, 0);
2414 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2415 *hash_arg_in_memory_p = 1;
2417 /* Now that we have already found this special case,
2418 might as well speed it up as much as possible. */
2419 hash += (unsigned) MEM;
2420 x = XEXP (x, 0);
2421 goto repeat;
2423 break;
2425 case PRE_DEC:
2426 case PRE_INC:
2427 case POST_DEC:
2428 case POST_INC:
2429 case PRE_MODIFY:
2430 case POST_MODIFY:
2431 case PC:
2432 case CC0:
2433 case CALL:
2434 case UNSPEC_VOLATILE:
2435 if (do_not_record_p) {
2436 *do_not_record_p = 1;
2437 return 0;
2439 else
2440 return hash;
2441 break;
2443 case ASM_OPERANDS:
2444 if (do_not_record_p && MEM_VOLATILE_P (x))
2446 *do_not_record_p = 1;
2447 return 0;
2449 else
2451 /* We don't want to take the filename and line into account. */
2452 hash += (unsigned) code + (unsigned) GET_MODE (x)
2453 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2454 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2455 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2457 if (ASM_OPERANDS_INPUT_LENGTH (x))
2459 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2461 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2462 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2463 do_not_record_p, hash_arg_in_memory_p,
2464 have_reg_qty, cb)
2465 + hash_rtx_string
2466 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2469 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2470 x = ASM_OPERANDS_INPUT (x, 0);
2471 mode = GET_MODE (x);
2472 goto repeat;
2475 return hash;
2477 break;
2479 default:
2480 break;
2483 i = GET_RTX_LENGTH (code) - 1;
2484 hash += (unsigned) code + (unsigned) GET_MODE (x);
2485 fmt = GET_RTX_FORMAT (code);
2486 for (; i >= 0; i--)
2488 switch (fmt[i])
2490 case 'e':
2491 /* If we are about to do the last recursive call
2492 needed at this level, change it into iteration.
2493 This function is called enough to be worth it. */
2494 if (i == 0)
2496 x = XEXP (x, i);
2497 goto repeat;
2500 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2501 hash_arg_in_memory_p,
2502 have_reg_qty, cb);
2503 break;
2505 case 'E':
2506 for (j = 0; j < XVECLEN (x, i); j++)
2507 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2508 hash_arg_in_memory_p,
2509 have_reg_qty, cb);
2510 break;
2512 case 's':
2513 hash += hash_rtx_string (XSTR (x, i));
2514 break;
2516 case 'i':
2517 hash += (unsigned int) XINT (x, i);
2518 break;
2520 case '0': case 't':
2521 /* Unused. */
2522 break;
2524 default:
2525 gcc_unreachable ();
2529 return hash;
2532 /* Hash an rtx. We are careful to make sure the value is never negative.
2533 Equivalent registers hash identically.
2534 MODE is used in hashing for CONST_INTs only;
2535 otherwise the mode of X is used.
2537 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2539 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2540 a MEM rtx which does not have the MEM_READONLY_P flag set.
2542 Note that cse_insn knows that the hash code of a MEM expression
2543 is just (int) MEM plus the hash code of the address. */
2545 unsigned
2546 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2547 int *hash_arg_in_memory_p, bool have_reg_qty)
2549 return hash_rtx_cb (x, mode, do_not_record_p,
2550 hash_arg_in_memory_p, have_reg_qty, NULL);
2553 /* Hash an rtx X for cse via hash_rtx.
2554 Stores 1 in do_not_record if any subexpression is volatile.
2555 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2556 does not have the MEM_READONLY_P flag set. */
2558 static inline unsigned
2559 canon_hash (rtx x, machine_mode mode)
2561 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2564 /* Like canon_hash but with no side effects, i.e. do_not_record
2565 and hash_arg_in_memory are not changed. */
2567 static inline unsigned
2568 safe_hash (rtx x, machine_mode mode)
2570 int dummy_do_not_record;
2571 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2574 /* Return 1 iff X and Y would canonicalize into the same thing,
2575 without actually constructing the canonicalization of either one.
2576 If VALIDATE is nonzero,
2577 we assume X is an expression being processed from the rtl
2578 and Y was found in the hash table. We check register refs
2579 in Y for being marked as valid.
2581 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2584 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2586 int i, j;
2587 enum rtx_code code;
2588 const char *fmt;
2590 /* Note: it is incorrect to assume an expression is equivalent to itself
2591 if VALIDATE is nonzero. */
2592 if (x == y && !validate)
2593 return 1;
2595 if (x == 0 || y == 0)
2596 return x == y;
2598 code = GET_CODE (x);
2599 if (code != GET_CODE (y))
2600 return 0;
2602 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2603 if (GET_MODE (x) != GET_MODE (y))
2604 return 0;
2606 /* MEMs referring to different address space are not equivalent. */
2607 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2608 return 0;
2610 switch (code)
2612 case PC:
2613 case CC0:
2614 CASE_CONST_UNIQUE:
2615 return x == y;
2617 case LABEL_REF:
2618 return label_ref_label (x) == label_ref_label (y);
2620 case SYMBOL_REF:
2621 return XSTR (x, 0) == XSTR (y, 0);
2623 case REG:
2624 if (for_gcse)
2625 return REGNO (x) == REGNO (y);
2626 else
2628 unsigned int regno = REGNO (y);
2629 unsigned int i;
2630 unsigned int endregno = END_REGNO (y);
2632 /* If the quantities are not the same, the expressions are not
2633 equivalent. If there are and we are not to validate, they
2634 are equivalent. Otherwise, ensure all regs are up-to-date. */
2636 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2637 return 0;
2639 if (! validate)
2640 return 1;
2642 for (i = regno; i < endregno; i++)
2643 if (REG_IN_TABLE (i) != REG_TICK (i))
2644 return 0;
2646 return 1;
2649 case MEM:
2650 if (for_gcse)
2652 /* A volatile mem should not be considered equivalent to any
2653 other. */
2654 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2655 return 0;
2657 /* Can't merge two expressions in different alias sets, since we
2658 can decide that the expression is transparent in a block when
2659 it isn't, due to it being set with the different alias set.
2661 Also, can't merge two expressions with different MEM_ATTRS.
2662 They could e.g. be two different entities allocated into the
2663 same space on the stack (see e.g. PR25130). In that case, the
2664 MEM addresses can be the same, even though the two MEMs are
2665 absolutely not equivalent.
2667 But because really all MEM attributes should be the same for
2668 equivalent MEMs, we just use the invariant that MEMs that have
2669 the same attributes share the same mem_attrs data structure. */
2670 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2671 return 0;
2673 /* If we are handling exceptions, we cannot consider two expressions
2674 with different trapping status as equivalent, because simple_mem
2675 might accept one and reject the other. */
2676 if (cfun->can_throw_non_call_exceptions
2677 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2678 return 0;
2680 break;
2682 /* For commutative operations, check both orders. */
2683 case PLUS:
2684 case MULT:
2685 case AND:
2686 case IOR:
2687 case XOR:
2688 case NE:
2689 case EQ:
2690 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2691 validate, for_gcse)
2692 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2693 validate, for_gcse))
2694 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2695 validate, for_gcse)
2696 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2697 validate, for_gcse)));
2699 case ASM_OPERANDS:
2700 /* We don't use the generic code below because we want to
2701 disregard filename and line numbers. */
2703 /* A volatile asm isn't equivalent to any other. */
2704 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2705 return 0;
2707 if (GET_MODE (x) != GET_MODE (y)
2708 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2709 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2710 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2711 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2712 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2713 return 0;
2715 if (ASM_OPERANDS_INPUT_LENGTH (x))
2717 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2718 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2719 ASM_OPERANDS_INPUT (y, i),
2720 validate, for_gcse)
2721 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2722 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2723 return 0;
2726 return 1;
2728 default:
2729 break;
2732 /* Compare the elements. If any pair of corresponding elements
2733 fail to match, return 0 for the whole thing. */
2735 fmt = GET_RTX_FORMAT (code);
2736 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2738 switch (fmt[i])
2740 case 'e':
2741 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2742 validate, for_gcse))
2743 return 0;
2744 break;
2746 case 'E':
2747 if (XVECLEN (x, i) != XVECLEN (y, i))
2748 return 0;
2749 for (j = 0; j < XVECLEN (x, i); j++)
2750 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2751 validate, for_gcse))
2752 return 0;
2753 break;
2755 case 's':
2756 if (strcmp (XSTR (x, i), XSTR (y, i)))
2757 return 0;
2758 break;
2760 case 'i':
2761 if (XINT (x, i) != XINT (y, i))
2762 return 0;
2763 break;
2765 case 'w':
2766 if (XWINT (x, i) != XWINT (y, i))
2767 return 0;
2768 break;
2770 case '0':
2771 case 't':
2772 break;
2774 default:
2775 gcc_unreachable ();
2779 return 1;
2782 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2783 the result if necessary. INSN is as for canon_reg. */
2785 static void
2786 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2788 if (*xloc)
2790 rtx new_rtx = canon_reg (*xloc, insn);
2792 /* If replacing pseudo with hard reg or vice versa, ensure the
2793 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2794 gcc_assert (insn && new_rtx);
2795 validate_change (insn, xloc, new_rtx, 1);
2799 /* Canonicalize an expression:
2800 replace each register reference inside it
2801 with the "oldest" equivalent register.
2803 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2804 after we make our substitution. The calls are made with IN_GROUP nonzero
2805 so apply_change_group must be called upon the outermost return from this
2806 function (unless INSN is zero). The result of apply_change_group can
2807 generally be discarded since the changes we are making are optional. */
2809 static rtx
2810 canon_reg (rtx x, rtx_insn *insn)
2812 int i;
2813 enum rtx_code code;
2814 const char *fmt;
2816 if (x == 0)
2817 return x;
2819 code = GET_CODE (x);
2820 switch (code)
2822 case PC:
2823 case CC0:
2824 case CONST:
2825 CASE_CONST_ANY:
2826 case SYMBOL_REF:
2827 case LABEL_REF:
2828 case ADDR_VEC:
2829 case ADDR_DIFF_VEC:
2830 return x;
2832 case REG:
2834 int first;
2835 int q;
2836 struct qty_table_elem *ent;
2838 /* Never replace a hard reg, because hard regs can appear
2839 in more than one machine mode, and we must preserve the mode
2840 of each occurrence. Also, some hard regs appear in
2841 MEMs that are shared and mustn't be altered. Don't try to
2842 replace any reg that maps to a reg of class NO_REGS. */
2843 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2844 || ! REGNO_QTY_VALID_P (REGNO (x)))
2845 return x;
2847 q = REG_QTY (REGNO (x));
2848 ent = &qty_table[q];
2849 first = ent->first_reg;
2850 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2851 : REGNO_REG_CLASS (first) == NO_REGS ? x
2852 : gen_rtx_REG (ent->mode, first));
2855 default:
2856 break;
2859 fmt = GET_RTX_FORMAT (code);
2860 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2862 int j;
2864 if (fmt[i] == 'e')
2865 validate_canon_reg (&XEXP (x, i), insn);
2866 else if (fmt[i] == 'E')
2867 for (j = 0; j < XVECLEN (x, i); j++)
2868 validate_canon_reg (&XVECEXP (x, i, j), insn);
2871 return x;
2874 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2875 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2876 what values are being compared.
2878 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2879 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2880 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2881 compared to produce cc0.
2883 The return value is the comparison operator and is either the code of
2884 A or the code corresponding to the inverse of the comparison. */
2886 static enum rtx_code
2887 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2888 machine_mode *pmode1, machine_mode *pmode2)
2890 rtx arg1, arg2;
2891 hash_set<rtx> *visited = NULL;
2892 /* Set nonzero when we find something of interest. */
2893 rtx x = NULL;
2895 arg1 = *parg1, arg2 = *parg2;
2897 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2899 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2901 int reverse_code = 0;
2902 struct table_elt *p = 0;
2904 /* Remember state from previous iteration. */
2905 if (x)
2907 if (!visited)
2908 visited = new hash_set<rtx>;
2909 visited->add (x);
2910 x = 0;
2913 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2914 On machines with CC0, this is the only case that can occur, since
2915 fold_rtx will return the COMPARE or item being compared with zero
2916 when given CC0. */
2918 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2919 x = arg1;
2921 /* If ARG1 is a comparison operator and CODE is testing for
2922 STORE_FLAG_VALUE, get the inner arguments. */
2924 else if (COMPARISON_P (arg1))
2926 #ifdef FLOAT_STORE_FLAG_VALUE
2927 REAL_VALUE_TYPE fsfv;
2928 #endif
2930 if (code == NE
2931 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2932 && code == LT && STORE_FLAG_VALUE == -1)
2933 #ifdef FLOAT_STORE_FLAG_VALUE
2934 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2935 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2936 REAL_VALUE_NEGATIVE (fsfv)))
2937 #endif
2939 x = arg1;
2940 else if (code == EQ
2941 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2942 && code == GE && STORE_FLAG_VALUE == -1)
2943 #ifdef FLOAT_STORE_FLAG_VALUE
2944 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2945 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2946 REAL_VALUE_NEGATIVE (fsfv)))
2947 #endif
2949 x = arg1, reverse_code = 1;
2952 /* ??? We could also check for
2954 (ne (and (eq (...) (const_int 1))) (const_int 0))
2956 and related forms, but let's wait until we see them occurring. */
2958 if (x == 0)
2959 /* Look up ARG1 in the hash table and see if it has an equivalence
2960 that lets us see what is being compared. */
2961 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2962 if (p)
2964 p = p->first_same_value;
2966 /* If what we compare is already known to be constant, that is as
2967 good as it gets.
2968 We need to break the loop in this case, because otherwise we
2969 can have an infinite loop when looking at a reg that is known
2970 to be a constant which is the same as a comparison of a reg
2971 against zero which appears later in the insn stream, which in
2972 turn is constant and the same as the comparison of the first reg
2973 against zero... */
2974 if (p->is_const)
2975 break;
2978 for (; p; p = p->next_same_value)
2980 machine_mode inner_mode = GET_MODE (p->exp);
2981 #ifdef FLOAT_STORE_FLAG_VALUE
2982 REAL_VALUE_TYPE fsfv;
2983 #endif
2985 /* If the entry isn't valid, skip it. */
2986 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2987 continue;
2989 /* If it's a comparison we've used before, skip it. */
2990 if (visited && visited->contains (p->exp))
2991 continue;
2993 if (GET_CODE (p->exp) == COMPARE
2994 /* Another possibility is that this machine has a compare insn
2995 that includes the comparison code. In that case, ARG1 would
2996 be equivalent to a comparison operation that would set ARG1 to
2997 either STORE_FLAG_VALUE or zero. If this is an NE operation,
2998 ORIG_CODE is the actual comparison being done; if it is an EQ,
2999 we must reverse ORIG_CODE. On machine with a negative value
3000 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3001 || ((code == NE
3002 || (code == LT
3003 && val_signbit_known_set_p (inner_mode,
3004 STORE_FLAG_VALUE))
3005 #ifdef FLOAT_STORE_FLAG_VALUE
3006 || (code == LT
3007 && SCALAR_FLOAT_MODE_P (inner_mode)
3008 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3009 REAL_VALUE_NEGATIVE (fsfv)))
3010 #endif
3012 && COMPARISON_P (p->exp)))
3014 x = p->exp;
3015 break;
3017 else if ((code == EQ
3018 || (code == GE
3019 && val_signbit_known_set_p (inner_mode,
3020 STORE_FLAG_VALUE))
3021 #ifdef FLOAT_STORE_FLAG_VALUE
3022 || (code == GE
3023 && SCALAR_FLOAT_MODE_P (inner_mode)
3024 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3025 REAL_VALUE_NEGATIVE (fsfv)))
3026 #endif
3028 && COMPARISON_P (p->exp))
3030 reverse_code = 1;
3031 x = p->exp;
3032 break;
3035 /* If this non-trapping address, e.g. fp + constant, the
3036 equivalent is a better operand since it may let us predict
3037 the value of the comparison. */
3038 else if (!rtx_addr_can_trap_p (p->exp))
3040 arg1 = p->exp;
3041 continue;
3045 /* If we didn't find a useful equivalence for ARG1, we are done.
3046 Otherwise, set up for the next iteration. */
3047 if (x == 0)
3048 break;
3050 /* If we need to reverse the comparison, make sure that is
3051 possible -- we can't necessarily infer the value of GE from LT
3052 with floating-point operands. */
3053 if (reverse_code)
3055 enum rtx_code reversed = reversed_comparison_code (x, NULL);
3056 if (reversed == UNKNOWN)
3057 break;
3058 else
3059 code = reversed;
3061 else if (COMPARISON_P (x))
3062 code = GET_CODE (x);
3063 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3066 /* Return our results. Return the modes from before fold_rtx
3067 because fold_rtx might produce const_int, and then it's too late. */
3068 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3069 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3071 if (visited)
3072 delete visited;
3073 return code;
3076 /* If X is a nontrivial arithmetic operation on an argument for which
3077 a constant value can be determined, return the result of operating
3078 on that value, as a constant. Otherwise, return X, possibly with
3079 one or more operands changed to a forward-propagated constant.
3081 If X is a register whose contents are known, we do NOT return
3082 those contents here; equiv_constant is called to perform that task.
3083 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3085 INSN is the insn that we may be modifying. If it is 0, make a copy
3086 of X before modifying it. */
3088 static rtx
3089 fold_rtx (rtx x, rtx_insn *insn)
3091 enum rtx_code code;
3092 machine_mode mode;
3093 const char *fmt;
3094 int i;
3095 rtx new_rtx = 0;
3096 int changed = 0;
3098 /* Operands of X. */
3099 /* Workaround -Wmaybe-uninitialized false positive during
3100 profiledbootstrap by initializing them. */
3101 rtx folded_arg0 = NULL_RTX;
3102 rtx folded_arg1 = NULL_RTX;
3104 /* Constant equivalents of first three operands of X;
3105 0 when no such equivalent is known. */
3106 rtx const_arg0;
3107 rtx const_arg1;
3108 rtx const_arg2;
3110 /* The mode of the first operand of X. We need this for sign and zero
3111 extends. */
3112 machine_mode mode_arg0;
3114 if (x == 0)
3115 return x;
3117 /* Try to perform some initial simplifications on X. */
3118 code = GET_CODE (x);
3119 switch (code)
3121 case MEM:
3122 case SUBREG:
3123 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3124 than it would in other contexts. Basically its mode does not
3125 signify the size of the object read. That information is carried
3126 by size operand. If we happen to have a MEM of the appropriate
3127 mode in our tables with a constant value we could simplify the
3128 extraction incorrectly if we allowed substitution of that value
3129 for the MEM. */
3130 case ZERO_EXTRACT:
3131 case SIGN_EXTRACT:
3132 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3133 return new_rtx;
3134 return x;
3136 case CONST:
3137 CASE_CONST_ANY:
3138 case SYMBOL_REF:
3139 case LABEL_REF:
3140 case REG:
3141 case PC:
3142 /* No use simplifying an EXPR_LIST
3143 since they are used only for lists of args
3144 in a function call's REG_EQUAL note. */
3145 case EXPR_LIST:
3146 return x;
3148 case CC0:
3149 return prev_insn_cc0;
3151 case ASM_OPERANDS:
3152 if (insn)
3154 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3155 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3156 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3158 return x;
3160 case CALL:
3161 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3162 return x;
3163 break;
3165 /* Anything else goes through the loop below. */
3166 default:
3167 break;
3170 mode = GET_MODE (x);
3171 const_arg0 = 0;
3172 const_arg1 = 0;
3173 const_arg2 = 0;
3174 mode_arg0 = VOIDmode;
3176 /* Try folding our operands.
3177 Then see which ones have constant values known. */
3179 fmt = GET_RTX_FORMAT (code);
3180 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3181 if (fmt[i] == 'e')
3183 rtx folded_arg = XEXP (x, i), const_arg;
3184 machine_mode mode_arg = GET_MODE (folded_arg);
3186 switch (GET_CODE (folded_arg))
3188 case MEM:
3189 case REG:
3190 case SUBREG:
3191 const_arg = equiv_constant (folded_arg);
3192 break;
3194 case CONST:
3195 CASE_CONST_ANY:
3196 case SYMBOL_REF:
3197 case LABEL_REF:
3198 const_arg = folded_arg;
3199 break;
3201 case CC0:
3202 /* The cc0-user and cc0-setter may be in different blocks if
3203 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3204 will have been cleared as we exited the block with the
3205 setter.
3207 While we could potentially track cc0 in this case, it just
3208 doesn't seem to be worth it given that cc0 targets are not
3209 terribly common or important these days and trapping math
3210 is rarely used. The combination of those two conditions
3211 necessary to trip this situation is exceedingly rare in the
3212 real world. */
3213 if (!prev_insn_cc0)
3215 const_arg = NULL_RTX;
3217 else
3219 folded_arg = prev_insn_cc0;
3220 mode_arg = prev_insn_cc0_mode;
3221 const_arg = equiv_constant (folded_arg);
3223 break;
3225 default:
3226 folded_arg = fold_rtx (folded_arg, insn);
3227 const_arg = equiv_constant (folded_arg);
3228 break;
3231 /* For the first three operands, see if the operand
3232 is constant or equivalent to a constant. */
3233 switch (i)
3235 case 0:
3236 folded_arg0 = folded_arg;
3237 const_arg0 = const_arg;
3238 mode_arg0 = mode_arg;
3239 break;
3240 case 1:
3241 folded_arg1 = folded_arg;
3242 const_arg1 = const_arg;
3243 break;
3244 case 2:
3245 const_arg2 = const_arg;
3246 break;
3249 /* Pick the least expensive of the argument and an equivalent constant
3250 argument. */
3251 if (const_arg != 0
3252 && const_arg != folded_arg
3253 && (COST_IN (const_arg, mode_arg, code, i)
3254 <= COST_IN (folded_arg, mode_arg, code, i))
3256 /* It's not safe to substitute the operand of a conversion
3257 operator with a constant, as the conversion's identity
3258 depends upon the mode of its operand. This optimization
3259 is handled by the call to simplify_unary_operation. */
3260 && (GET_RTX_CLASS (code) != RTX_UNARY
3261 || GET_MODE (const_arg) == mode_arg0
3262 || (code != ZERO_EXTEND
3263 && code != SIGN_EXTEND
3264 && code != TRUNCATE
3265 && code != FLOAT_TRUNCATE
3266 && code != FLOAT_EXTEND
3267 && code != FLOAT
3268 && code != FIX
3269 && code != UNSIGNED_FLOAT
3270 && code != UNSIGNED_FIX)))
3271 folded_arg = const_arg;
3273 if (folded_arg == XEXP (x, i))
3274 continue;
3276 if (insn == NULL_RTX && !changed)
3277 x = copy_rtx (x);
3278 changed = 1;
3279 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3282 if (changed)
3284 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3285 consistent with the order in X. */
3286 if (canonicalize_change_group (insn, x))
3288 std::swap (const_arg0, const_arg1);
3289 std::swap (folded_arg0, folded_arg1);
3292 apply_change_group ();
3295 /* If X is an arithmetic operation, see if we can simplify it. */
3297 switch (GET_RTX_CLASS (code))
3299 case RTX_UNARY:
3301 /* We can't simplify extension ops unless we know the
3302 original mode. */
3303 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3304 && mode_arg0 == VOIDmode)
3305 break;
3307 new_rtx = simplify_unary_operation (code, mode,
3308 const_arg0 ? const_arg0 : folded_arg0,
3309 mode_arg0);
3311 break;
3313 case RTX_COMPARE:
3314 case RTX_COMM_COMPARE:
3315 /* See what items are actually being compared and set FOLDED_ARG[01]
3316 to those values and CODE to the actual comparison code. If any are
3317 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3318 do anything if both operands are already known to be constant. */
3320 /* ??? Vector mode comparisons are not supported yet. */
3321 if (VECTOR_MODE_P (mode))
3322 break;
3324 if (const_arg0 == 0 || const_arg1 == 0)
3326 struct table_elt *p0, *p1;
3327 rtx true_rtx, false_rtx;
3328 machine_mode mode_arg1;
3330 if (SCALAR_FLOAT_MODE_P (mode))
3332 #ifdef FLOAT_STORE_FLAG_VALUE
3333 true_rtx = (const_double_from_real_value
3334 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3335 #else
3336 true_rtx = NULL_RTX;
3337 #endif
3338 false_rtx = CONST0_RTX (mode);
3340 else
3342 true_rtx = const_true_rtx;
3343 false_rtx = const0_rtx;
3346 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3347 &mode_arg0, &mode_arg1);
3349 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3350 what kinds of things are being compared, so we can't do
3351 anything with this comparison. */
3353 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3354 break;
3356 const_arg0 = equiv_constant (folded_arg0);
3357 const_arg1 = equiv_constant (folded_arg1);
3359 /* If we do not now have two constants being compared, see
3360 if we can nevertheless deduce some things about the
3361 comparison. */
3362 if (const_arg0 == 0 || const_arg1 == 0)
3364 if (const_arg1 != NULL)
3366 rtx cheapest_simplification;
3367 int cheapest_cost;
3368 rtx simp_result;
3369 struct table_elt *p;
3371 /* See if we can find an equivalent of folded_arg0
3372 that gets us a cheaper expression, possibly a
3373 constant through simplifications. */
3374 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3375 mode_arg0);
3377 if (p != NULL)
3379 cheapest_simplification = x;
3380 cheapest_cost = COST (x, mode);
3382 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3384 int cost;
3386 /* If the entry isn't valid, skip it. */
3387 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3388 continue;
3390 /* Try to simplify using this equivalence. */
3391 simp_result
3392 = simplify_relational_operation (code, mode,
3393 mode_arg0,
3394 p->exp,
3395 const_arg1);
3397 if (simp_result == NULL)
3398 continue;
3400 cost = COST (simp_result, mode);
3401 if (cost < cheapest_cost)
3403 cheapest_cost = cost;
3404 cheapest_simplification = simp_result;
3408 /* If we have a cheaper expression now, use that
3409 and try folding it further, from the top. */
3410 if (cheapest_simplification != x)
3411 return fold_rtx (copy_rtx (cheapest_simplification),
3412 insn);
3416 /* See if the two operands are the same. */
3418 if ((REG_P (folded_arg0)
3419 && REG_P (folded_arg1)
3420 && (REG_QTY (REGNO (folded_arg0))
3421 == REG_QTY (REGNO (folded_arg1))))
3422 || ((p0 = lookup (folded_arg0,
3423 SAFE_HASH (folded_arg0, mode_arg0),
3424 mode_arg0))
3425 && (p1 = lookup (folded_arg1,
3426 SAFE_HASH (folded_arg1, mode_arg0),
3427 mode_arg0))
3428 && p0->first_same_value == p1->first_same_value))
3429 folded_arg1 = folded_arg0;
3431 /* If FOLDED_ARG0 is a register, see if the comparison we are
3432 doing now is either the same as we did before or the reverse
3433 (we only check the reverse if not floating-point). */
3434 else if (REG_P (folded_arg0))
3436 int qty = REG_QTY (REGNO (folded_arg0));
3438 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3440 struct qty_table_elem *ent = &qty_table[qty];
3442 if ((comparison_dominates_p (ent->comparison_code, code)
3443 || (! FLOAT_MODE_P (mode_arg0)
3444 && comparison_dominates_p (ent->comparison_code,
3445 reverse_condition (code))))
3446 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3447 || (const_arg1
3448 && rtx_equal_p (ent->comparison_const,
3449 const_arg1))
3450 || (REG_P (folded_arg1)
3451 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3453 if (comparison_dominates_p (ent->comparison_code, code))
3455 if (true_rtx)
3456 return true_rtx;
3457 else
3458 break;
3460 else
3461 return false_rtx;
3468 /* If we are comparing against zero, see if the first operand is
3469 equivalent to an IOR with a constant. If so, we may be able to
3470 determine the result of this comparison. */
3471 if (const_arg1 == const0_rtx && !const_arg0)
3473 rtx y = lookup_as_function (folded_arg0, IOR);
3474 rtx inner_const;
3476 if (y != 0
3477 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3478 && CONST_INT_P (inner_const)
3479 && INTVAL (inner_const) != 0)
3480 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3484 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3485 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3486 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3487 op0, op1);
3489 break;
3491 case RTX_BIN_ARITH:
3492 case RTX_COMM_ARITH:
3493 switch (code)
3495 case PLUS:
3496 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3497 with that LABEL_REF as its second operand. If so, the result is
3498 the first operand of that MINUS. This handles switches with an
3499 ADDR_DIFF_VEC table. */
3500 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3502 rtx y
3503 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3504 : lookup_as_function (folded_arg0, MINUS);
3506 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3507 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg1))
3508 return XEXP (y, 0);
3510 /* Now try for a CONST of a MINUS like the above. */
3511 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3512 : lookup_as_function (folded_arg0, CONST))) != 0
3513 && GET_CODE (XEXP (y, 0)) == MINUS
3514 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3515 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg1))
3516 return XEXP (XEXP (y, 0), 0);
3519 /* Likewise if the operands are in the other order. */
3520 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3522 rtx y
3523 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3524 : lookup_as_function (folded_arg1, MINUS);
3526 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3527 && label_ref_label (XEXP (y, 1)) == label_ref_label (const_arg0))
3528 return XEXP (y, 0);
3530 /* Now try for a CONST of a MINUS like the above. */
3531 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3532 : lookup_as_function (folded_arg1, CONST))) != 0
3533 && GET_CODE (XEXP (y, 0)) == MINUS
3534 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3535 && label_ref_label (XEXP (XEXP (y, 0), 1)) == label_ref_label (const_arg0))
3536 return XEXP (XEXP (y, 0), 0);
3539 /* If second operand is a register equivalent to a negative
3540 CONST_INT, see if we can find a register equivalent to the
3541 positive constant. Make a MINUS if so. Don't do this for
3542 a non-negative constant since we might then alternate between
3543 choosing positive and negative constants. Having the positive
3544 constant previously-used is the more common case. Be sure
3545 the resulting constant is non-negative; if const_arg1 were
3546 the smallest negative number this would overflow: depending
3547 on the mode, this would either just be the same value (and
3548 hence not save anything) or be incorrect. */
3549 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3550 && INTVAL (const_arg1) < 0
3551 /* This used to test
3553 -INTVAL (const_arg1) >= 0
3555 But The Sun V5.0 compilers mis-compiled that test. So
3556 instead we test for the problematic value in a more direct
3557 manner and hope the Sun compilers get it correct. */
3558 && INTVAL (const_arg1) !=
3559 (HOST_WIDE_INT_1 << (HOST_BITS_PER_WIDE_INT - 1))
3560 && REG_P (folded_arg1))
3562 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3563 struct table_elt *p
3564 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3566 if (p)
3567 for (p = p->first_same_value; p; p = p->next_same_value)
3568 if (REG_P (p->exp))
3569 return simplify_gen_binary (MINUS, mode, folded_arg0,
3570 canon_reg (p->exp, NULL));
3572 goto from_plus;
3574 case MINUS:
3575 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3576 If so, produce (PLUS Z C2-C). */
3577 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3579 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3580 if (y && CONST_INT_P (XEXP (y, 1)))
3581 return fold_rtx (plus_constant (mode, copy_rtx (y),
3582 -INTVAL (const_arg1)),
3583 NULL);
3586 /* Fall through. */
3588 from_plus:
3589 case SMIN: case SMAX: case UMIN: case UMAX:
3590 case IOR: case AND: case XOR:
3591 case MULT:
3592 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3593 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3594 is known to be of similar form, we may be able to replace the
3595 operation with a combined operation. This may eliminate the
3596 intermediate operation if every use is simplified in this way.
3597 Note that the similar optimization done by combine.c only works
3598 if the intermediate operation's result has only one reference. */
3600 if (REG_P (folded_arg0)
3601 && const_arg1 && CONST_INT_P (const_arg1))
3603 int is_shift
3604 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3605 rtx y, inner_const, new_const;
3606 rtx canon_const_arg1 = const_arg1;
3607 enum rtx_code associate_code;
3609 if (is_shift
3610 && (INTVAL (const_arg1) >= GET_MODE_UNIT_PRECISION (mode)
3611 || INTVAL (const_arg1) < 0))
3613 if (SHIFT_COUNT_TRUNCATED)
3614 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3615 & (GET_MODE_UNIT_BITSIZE (mode)
3616 - 1));
3617 else
3618 break;
3621 y = lookup_as_function (folded_arg0, code);
3622 if (y == 0)
3623 break;
3625 /* If we have compiled a statement like
3626 "if (x == (x & mask1))", and now are looking at
3627 "x & mask2", we will have a case where the first operand
3628 of Y is the same as our first operand. Unless we detect
3629 this case, an infinite loop will result. */
3630 if (XEXP (y, 0) == folded_arg0)
3631 break;
3633 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3634 if (!inner_const || !CONST_INT_P (inner_const))
3635 break;
3637 /* Don't associate these operations if they are a PLUS with the
3638 same constant and it is a power of two. These might be doable
3639 with a pre- or post-increment. Similarly for two subtracts of
3640 identical powers of two with post decrement. */
3642 if (code == PLUS && const_arg1 == inner_const
3643 && ((HAVE_PRE_INCREMENT
3644 && pow2p_hwi (INTVAL (const_arg1)))
3645 || (HAVE_POST_INCREMENT
3646 && pow2p_hwi (INTVAL (const_arg1)))
3647 || (HAVE_PRE_DECREMENT
3648 && pow2p_hwi (- INTVAL (const_arg1)))
3649 || (HAVE_POST_DECREMENT
3650 && pow2p_hwi (- INTVAL (const_arg1)))))
3651 break;
3653 /* ??? Vector mode shifts by scalar
3654 shift operand are not supported yet. */
3655 if (is_shift && VECTOR_MODE_P (mode))
3656 break;
3658 if (is_shift
3659 && (INTVAL (inner_const) >= GET_MODE_UNIT_PRECISION (mode)
3660 || INTVAL (inner_const) < 0))
3662 if (SHIFT_COUNT_TRUNCATED)
3663 inner_const = GEN_INT (INTVAL (inner_const)
3664 & (GET_MODE_UNIT_BITSIZE (mode)
3665 - 1));
3666 else
3667 break;
3670 /* Compute the code used to compose the constants. For example,
3671 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3673 associate_code = (is_shift || code == MINUS ? PLUS : code);
3675 new_const = simplify_binary_operation (associate_code, mode,
3676 canon_const_arg1,
3677 inner_const);
3679 if (new_const == 0)
3680 break;
3682 /* If we are associating shift operations, don't let this
3683 produce a shift of the size of the object or larger.
3684 This could occur when we follow a sign-extend by a right
3685 shift on a machine that does a sign-extend as a pair
3686 of shifts. */
3688 if (is_shift
3689 && CONST_INT_P (new_const)
3690 && INTVAL (new_const) >= GET_MODE_UNIT_PRECISION (mode))
3692 /* As an exception, we can turn an ASHIFTRT of this
3693 form into a shift of the number of bits - 1. */
3694 if (code == ASHIFTRT)
3695 new_const = GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1);
3696 else if (!side_effects_p (XEXP (y, 0)))
3697 return CONST0_RTX (mode);
3698 else
3699 break;
3702 y = copy_rtx (XEXP (y, 0));
3704 /* If Y contains our first operand (the most common way this
3705 can happen is if Y is a MEM), we would do into an infinite
3706 loop if we tried to fold it. So don't in that case. */
3708 if (! reg_mentioned_p (folded_arg0, y))
3709 y = fold_rtx (y, insn);
3711 return simplify_gen_binary (code, mode, y, new_const);
3713 break;
3715 case DIV: case UDIV:
3716 /* ??? The associative optimization performed immediately above is
3717 also possible for DIV and UDIV using associate_code of MULT.
3718 However, we would need extra code to verify that the
3719 multiplication does not overflow, that is, there is no overflow
3720 in the calculation of new_const. */
3721 break;
3723 default:
3724 break;
3727 new_rtx = simplify_binary_operation (code, mode,
3728 const_arg0 ? const_arg0 : folded_arg0,
3729 const_arg1 ? const_arg1 : folded_arg1);
3730 break;
3732 case RTX_OBJ:
3733 /* (lo_sum (high X) X) is simply X. */
3734 if (code == LO_SUM && const_arg0 != 0
3735 && GET_CODE (const_arg0) == HIGH
3736 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3737 return const_arg1;
3738 break;
3740 case RTX_TERNARY:
3741 case RTX_BITFIELD_OPS:
3742 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3743 const_arg0 ? const_arg0 : folded_arg0,
3744 const_arg1 ? const_arg1 : folded_arg1,
3745 const_arg2 ? const_arg2 : XEXP (x, 2));
3746 break;
3748 default:
3749 break;
3752 return new_rtx ? new_rtx : x;
3755 /* Return a constant value currently equivalent to X.
3756 Return 0 if we don't know one. */
3758 static rtx
3759 equiv_constant (rtx x)
3761 if (REG_P (x)
3762 && REGNO_QTY_VALID_P (REGNO (x)))
3764 int x_q = REG_QTY (REGNO (x));
3765 struct qty_table_elem *x_ent = &qty_table[x_q];
3767 if (x_ent->const_rtx)
3768 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3771 if (x == 0 || CONSTANT_P (x))
3772 return x;
3774 if (GET_CODE (x) == SUBREG)
3776 machine_mode mode = GET_MODE (x);
3777 machine_mode imode = GET_MODE (SUBREG_REG (x));
3778 rtx new_rtx;
3780 /* See if we previously assigned a constant value to this SUBREG. */
3781 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3782 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3783 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3784 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3785 return new_rtx;
3787 /* If we didn't and if doing so makes sense, see if we previously
3788 assigned a constant value to the enclosing word mode SUBREG. */
3789 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3790 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3792 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3793 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3795 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3796 new_rtx = lookup_as_function (y, CONST_INT);
3797 if (new_rtx)
3798 return gen_lowpart (mode, new_rtx);
3802 /* Otherwise see if we already have a constant for the inner REG,
3803 and if that is enough to calculate an equivalent constant for
3804 the subreg. Note that the upper bits of paradoxical subregs
3805 are undefined, so they cannot be said to equal anything. */
3806 if (REG_P (SUBREG_REG (x))
3807 && !paradoxical_subreg_p (x)
3808 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3809 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3811 return 0;
3814 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3815 the hash table in case its value was seen before. */
3817 if (MEM_P (x))
3819 struct table_elt *elt;
3821 x = avoid_constant_pool_reference (x);
3822 if (CONSTANT_P (x))
3823 return x;
3825 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3826 if (elt == 0)
3827 return 0;
3829 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3830 if (elt->is_const && CONSTANT_P (elt->exp))
3831 return elt->exp;
3834 return 0;
3837 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3838 "taken" branch.
3840 In certain cases, this can cause us to add an equivalence. For example,
3841 if we are following the taken case of
3842 if (i == 2)
3843 we can add the fact that `i' and '2' are now equivalent.
3845 In any case, we can record that this comparison was passed. If the same
3846 comparison is seen later, we will know its value. */
3848 static void
3849 record_jump_equiv (rtx_insn *insn, bool taken)
3851 int cond_known_true;
3852 rtx op0, op1;
3853 rtx set;
3854 machine_mode mode, mode0, mode1;
3855 int reversed_nonequality = 0;
3856 enum rtx_code code;
3858 /* Ensure this is the right kind of insn. */
3859 gcc_assert (any_condjump_p (insn));
3861 set = pc_set (insn);
3863 /* See if this jump condition is known true or false. */
3864 if (taken)
3865 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3866 else
3867 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3869 /* Get the type of comparison being done and the operands being compared.
3870 If we had to reverse a non-equality condition, record that fact so we
3871 know that it isn't valid for floating-point. */
3872 code = GET_CODE (XEXP (SET_SRC (set), 0));
3873 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3874 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3876 /* On a cc0 target the cc0-setter and cc0-user may end up in different
3877 blocks. When that happens the tracking of the cc0-setter via
3878 PREV_INSN_CC0 is spoiled. That means that fold_rtx may return
3879 NULL_RTX. In those cases, there's nothing to record. */
3880 if (op0 == NULL_RTX || op1 == NULL_RTX)
3881 return;
3883 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3884 if (! cond_known_true)
3886 code = reversed_comparison_code_parts (code, op0, op1, insn);
3888 /* Don't remember if we can't find the inverse. */
3889 if (code == UNKNOWN)
3890 return;
3893 /* The mode is the mode of the non-constant. */
3894 mode = mode0;
3895 if (mode1 != VOIDmode)
3896 mode = mode1;
3898 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3901 /* Yet another form of subreg creation. In this case, we want something in
3902 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3904 static rtx
3905 record_jump_cond_subreg (machine_mode mode, rtx op)
3907 machine_mode op_mode = GET_MODE (op);
3908 if (op_mode == mode || op_mode == VOIDmode)
3909 return op;
3910 return lowpart_subreg (mode, op, op_mode);
3913 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3914 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3915 Make any useful entries we can with that information. Called from
3916 above function and called recursively. */
3918 static void
3919 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3920 rtx op1, int reversed_nonequality)
3922 unsigned op0_hash, op1_hash;
3923 int op0_in_memory, op1_in_memory;
3924 struct table_elt *op0_elt, *op1_elt;
3926 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3927 we know that they are also equal in the smaller mode (this is also
3928 true for all smaller modes whether or not there is a SUBREG, but
3929 is not worth testing for with no SUBREG). */
3931 /* Note that GET_MODE (op0) may not equal MODE. */
3932 if (code == EQ && paradoxical_subreg_p (op0))
3934 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3935 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3936 if (tem)
3937 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3938 reversed_nonequality);
3941 if (code == EQ && paradoxical_subreg_p (op1))
3943 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3944 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3945 if (tem)
3946 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3947 reversed_nonequality);
3950 /* Similarly, if this is an NE comparison, and either is a SUBREG
3951 making a smaller mode, we know the whole thing is also NE. */
3953 /* Note that GET_MODE (op0) may not equal MODE;
3954 if we test MODE instead, we can get an infinite recursion
3955 alternating between two modes each wider than MODE. */
3957 if (code == NE
3958 && partial_subreg_p (op0)
3959 && subreg_lowpart_p (op0))
3961 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3962 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3963 if (tem)
3964 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3965 reversed_nonequality);
3968 if (code == NE
3969 && partial_subreg_p (op1)
3970 && subreg_lowpart_p (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 /* Hash value of constant equivalent for SET_SRC. */
4158 unsigned src_const_hash;
4159 /* A constant equivalent for SET_SRC, if any. */
4160 rtx src_const;
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 /* This is the mode of XEXP (tem, 0) as well. */
4555 scalar_int_mode dest_mode
4556 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
4557 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4558 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4559 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4560 HOST_WIDE_INT mask;
4561 unsigned int shift;
4562 if (BITS_BIG_ENDIAN)
4563 shift = (GET_MODE_PRECISION (dest_mode)
4564 - INTVAL (pos) - INTVAL (width));
4565 else
4566 shift = INTVAL (pos);
4567 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4568 mask = HOST_WIDE_INT_M1;
4569 else
4570 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
4571 val = (val >> shift) & mask;
4572 src_eqv = GEN_INT (val);
4576 /* Set sets[i].src_elt to the class each source belongs to.
4577 Detect assignments from or to volatile things
4578 and set set[i] to zero so they will be ignored
4579 in the rest of this function.
4581 Nothing in this loop changes the hash table or the register chains. */
4583 for (i = 0; i < n_sets; i++)
4585 bool repeat = false;
4586 bool mem_noop_insn = false;
4587 rtx src, dest;
4588 rtx src_folded;
4589 struct table_elt *elt = 0, *p;
4590 machine_mode mode;
4591 rtx src_eqv_here;
4592 rtx src_const = 0;
4593 rtx src_related = 0;
4594 bool src_related_is_const_anchor = false;
4595 struct table_elt *src_const_elt = 0;
4596 int src_cost = MAX_COST;
4597 int src_eqv_cost = MAX_COST;
4598 int src_folded_cost = MAX_COST;
4599 int src_related_cost = MAX_COST;
4600 int src_elt_cost = MAX_COST;
4601 int src_regcost = MAX_COST;
4602 int src_eqv_regcost = MAX_COST;
4603 int src_folded_regcost = MAX_COST;
4604 int src_related_regcost = MAX_COST;
4605 int src_elt_regcost = MAX_COST;
4606 /* Set nonzero if we need to call force_const_mem on with the
4607 contents of src_folded before using it. */
4608 int src_folded_force_flag = 0;
4609 scalar_int_mode int_mode;
4611 dest = SET_DEST (sets[i].rtl);
4612 src = SET_SRC (sets[i].rtl);
4614 /* If SRC is a constant that has no machine mode,
4615 hash it with the destination's machine mode.
4616 This way we can keep different modes separate. */
4618 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4619 sets[i].mode = mode;
4621 if (src_eqv)
4623 machine_mode eqvmode = mode;
4624 if (GET_CODE (dest) == STRICT_LOW_PART)
4625 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4626 do_not_record = 0;
4627 hash_arg_in_memory = 0;
4628 src_eqv_hash = HASH (src_eqv, eqvmode);
4630 /* Find the equivalence class for the equivalent expression. */
4632 if (!do_not_record)
4633 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4635 src_eqv_volatile = do_not_record;
4636 src_eqv_in_memory = hash_arg_in_memory;
4639 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4640 value of the INNER register, not the destination. So it is not
4641 a valid substitution for the source. But save it for later. */
4642 if (GET_CODE (dest) == STRICT_LOW_PART)
4643 src_eqv_here = 0;
4644 else
4645 src_eqv_here = src_eqv;
4647 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4648 simplified result, which may not necessarily be valid. */
4649 src_folded = fold_rtx (src, NULL);
4651 #if 0
4652 /* ??? This caused bad code to be generated for the m68k port with -O2.
4653 Suppose src is (CONST_INT -1), and that after truncation src_folded
4654 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4655 At the end we will add src and src_const to the same equivalence
4656 class. We now have 3 and -1 on the same equivalence class. This
4657 causes later instructions to be mis-optimized. */
4658 /* If storing a constant in a bitfield, pre-truncate the constant
4659 so we will be able to record it later. */
4660 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4662 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4664 if (CONST_INT_P (src)
4665 && CONST_INT_P (width)
4666 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4667 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4668 src_folded
4669 = GEN_INT (INTVAL (src) & ((HOST_WIDE_INT_1
4670 << INTVAL (width)) - 1));
4672 #endif
4674 /* Compute SRC's hash code, and also notice if it
4675 should not be recorded at all. In that case,
4676 prevent any further processing of this assignment. */
4677 do_not_record = 0;
4678 hash_arg_in_memory = 0;
4680 sets[i].src = src;
4681 sets[i].src_hash = HASH (src, mode);
4682 sets[i].src_volatile = do_not_record;
4683 sets[i].src_in_memory = hash_arg_in_memory;
4685 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4686 a pseudo, do not record SRC. Using SRC as a replacement for
4687 anything else will be incorrect in that situation. Note that
4688 this usually occurs only for stack slots, in which case all the
4689 RTL would be referring to SRC, so we don't lose any optimization
4690 opportunities by not having SRC in the hash table. */
4692 if (MEM_P (src)
4693 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4694 && REG_P (dest)
4695 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4696 sets[i].src_volatile = 1;
4698 else if (GET_CODE (src) == ASM_OPERANDS
4699 && GET_CODE (x) == PARALLEL)
4701 /* Do not record result of a non-volatile inline asm with
4702 more than one result. */
4703 if (n_sets > 1)
4704 sets[i].src_volatile = 1;
4706 int j, lim = XVECLEN (x, 0);
4707 for (j = 0; j < lim; j++)
4709 rtx y = XVECEXP (x, 0, j);
4710 /* And do not record result of a non-volatile inline asm
4711 with "memory" clobber. */
4712 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4714 sets[i].src_volatile = 1;
4715 break;
4720 #if 0
4721 /* It is no longer clear why we used to do this, but it doesn't
4722 appear to still be needed. So let's try without it since this
4723 code hurts cse'ing widened ops. */
4724 /* If source is a paradoxical subreg (such as QI treated as an SI),
4725 treat it as volatile. It may do the work of an SI in one context
4726 where the extra bits are not being used, but cannot replace an SI
4727 in general. */
4728 if (paradoxical_subreg_p (src))
4729 sets[i].src_volatile = 1;
4730 #endif
4732 /* Locate all possible equivalent forms for SRC. Try to replace
4733 SRC in the insn with each cheaper equivalent.
4735 We have the following types of equivalents: SRC itself, a folded
4736 version, a value given in a REG_EQUAL note, or a value related
4737 to a constant.
4739 Each of these equivalents may be part of an additional class
4740 of equivalents (if more than one is in the table, they must be in
4741 the same class; we check for this).
4743 If the source is volatile, we don't do any table lookups.
4745 We note any constant equivalent for possible later use in a
4746 REG_NOTE. */
4748 if (!sets[i].src_volatile)
4749 elt = lookup (src, sets[i].src_hash, mode);
4751 sets[i].src_elt = elt;
4753 if (elt && src_eqv_here && src_eqv_elt)
4755 if (elt->first_same_value != src_eqv_elt->first_same_value)
4757 /* The REG_EQUAL is indicating that two formerly distinct
4758 classes are now equivalent. So merge them. */
4759 merge_equiv_classes (elt, src_eqv_elt);
4760 src_eqv_hash = HASH (src_eqv, elt->mode);
4761 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4764 src_eqv_here = 0;
4767 else if (src_eqv_elt)
4768 elt = src_eqv_elt;
4770 /* Try to find a constant somewhere and record it in `src_const'.
4771 Record its table element, if any, in `src_const_elt'. Look in
4772 any known equivalences first. (If the constant is not in the
4773 table, also set `sets[i].src_const_hash'). */
4774 if (elt)
4775 for (p = elt->first_same_value; p; p = p->next_same_value)
4776 if (p->is_const)
4778 src_const = p->exp;
4779 src_const_elt = elt;
4780 break;
4783 if (src_const == 0
4784 && (CONSTANT_P (src_folded)
4785 /* Consider (minus (label_ref L1) (label_ref L2)) as
4786 "constant" here so we will record it. This allows us
4787 to fold switch statements when an ADDR_DIFF_VEC is used. */
4788 || (GET_CODE (src_folded) == MINUS
4789 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4790 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4791 src_const = src_folded, src_const_elt = elt;
4792 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4793 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4795 /* If we don't know if the constant is in the table, get its
4796 hash code and look it up. */
4797 if (src_const && src_const_elt == 0)
4799 sets[i].src_const_hash = HASH (src_const, mode);
4800 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4803 sets[i].src_const = src_const;
4804 sets[i].src_const_elt = src_const_elt;
4806 /* If the constant and our source are both in the table, mark them as
4807 equivalent. Otherwise, if a constant is in the table but the source
4808 isn't, set ELT to it. */
4809 if (src_const_elt && elt
4810 && src_const_elt->first_same_value != elt->first_same_value)
4811 merge_equiv_classes (elt, src_const_elt);
4812 else if (src_const_elt && elt == 0)
4813 elt = src_const_elt;
4815 /* See if there is a register linearly related to a constant
4816 equivalent of SRC. */
4817 if (src_const
4818 && (GET_CODE (src_const) == CONST
4819 || (src_const_elt && src_const_elt->related_value != 0)))
4821 src_related = use_related_value (src_const, src_const_elt);
4822 if (src_related)
4824 struct table_elt *src_related_elt
4825 = lookup (src_related, HASH (src_related, mode), mode);
4826 if (src_related_elt && elt)
4828 if (elt->first_same_value
4829 != src_related_elt->first_same_value)
4830 /* This can occur when we previously saw a CONST
4831 involving a SYMBOL_REF and then see the SYMBOL_REF
4832 twice. Merge the involved classes. */
4833 merge_equiv_classes (elt, src_related_elt);
4835 src_related = 0;
4836 src_related_elt = 0;
4838 else if (src_related_elt && elt == 0)
4839 elt = src_related_elt;
4843 /* See if we have a CONST_INT that is already in a register in a
4844 wider mode. */
4846 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4847 && is_int_mode (mode, &int_mode)
4848 && GET_MODE_PRECISION (int_mode) < BITS_PER_WORD)
4850 opt_scalar_int_mode wider_mode_iter;
4851 FOR_EACH_WIDER_MODE (wider_mode_iter, int_mode)
4853 scalar_int_mode wider_mode = wider_mode_iter.require ();
4854 if (GET_MODE_PRECISION (wider_mode) > BITS_PER_WORD)
4855 break;
4857 struct table_elt *const_elt
4858 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4860 if (const_elt == 0)
4861 continue;
4863 for (const_elt = const_elt->first_same_value;
4864 const_elt; const_elt = const_elt->next_same_value)
4865 if (REG_P (const_elt->exp))
4867 src_related = gen_lowpart (int_mode, const_elt->exp);
4868 break;
4871 if (src_related != 0)
4872 break;
4876 /* Another possibility is that we have an AND with a constant in
4877 a mode narrower than a word. If so, it might have been generated
4878 as part of an "if" which would narrow the AND. If we already
4879 have done the AND in a wider mode, we can use a SUBREG of that
4880 value. */
4882 if (flag_expensive_optimizations && ! src_related
4883 && is_a <scalar_int_mode> (mode, &int_mode)
4884 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4885 && GET_MODE_SIZE (int_mode) < UNITS_PER_WORD)
4887 opt_scalar_int_mode tmode_iter;
4888 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4890 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4892 scalar_int_mode tmode = tmode_iter.require ();
4893 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4894 break;
4896 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4897 struct table_elt *larger_elt;
4899 if (inner)
4901 PUT_MODE (new_and, tmode);
4902 XEXP (new_and, 0) = inner;
4903 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4904 if (larger_elt == 0)
4905 continue;
4907 for (larger_elt = larger_elt->first_same_value;
4908 larger_elt; larger_elt = larger_elt->next_same_value)
4909 if (REG_P (larger_elt->exp))
4911 src_related
4912 = gen_lowpart (int_mode, larger_elt->exp);
4913 break;
4916 if (src_related)
4917 break;
4922 /* See if a MEM has already been loaded with a widening operation;
4923 if it has, we can use a subreg of that. Many CISC machines
4924 also have such operations, but this is only likely to be
4925 beneficial on these machines. */
4927 rtx_code extend_op;
4928 if (flag_expensive_optimizations && src_related == 0
4929 && MEM_P (src) && ! do_not_record
4930 && is_a <scalar_int_mode> (mode, &int_mode)
4931 && (extend_op = load_extend_op (int_mode)) != UNKNOWN)
4933 struct rtx_def memory_extend_buf;
4934 rtx memory_extend_rtx = &memory_extend_buf;
4936 /* Set what we are trying to extend and the operation it might
4937 have been extended with. */
4938 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4939 PUT_CODE (memory_extend_rtx, extend_op);
4940 XEXP (memory_extend_rtx, 0) = src;
4942 opt_scalar_int_mode tmode_iter;
4943 FOR_EACH_WIDER_MODE (tmode_iter, int_mode)
4945 struct table_elt *larger_elt;
4947 scalar_int_mode tmode = tmode_iter.require ();
4948 if (GET_MODE_SIZE (tmode) > UNITS_PER_WORD)
4949 break;
4951 PUT_MODE (memory_extend_rtx, tmode);
4952 larger_elt = lookup (memory_extend_rtx,
4953 HASH (memory_extend_rtx, tmode), tmode);
4954 if (larger_elt == 0)
4955 continue;
4957 for (larger_elt = larger_elt->first_same_value;
4958 larger_elt; larger_elt = larger_elt->next_same_value)
4959 if (REG_P (larger_elt->exp))
4961 src_related = gen_lowpart (int_mode, larger_elt->exp);
4962 break;
4965 if (src_related)
4966 break;
4970 /* Try to express the constant using a register+offset expression
4971 derived from a constant anchor. */
4973 if (targetm.const_anchor
4974 && !src_related
4975 && src_const
4976 && GET_CODE (src_const) == CONST_INT)
4978 src_related = try_const_anchors (src_const, mode);
4979 src_related_is_const_anchor = src_related != NULL_RTX;
4983 if (src == src_folded)
4984 src_folded = 0;
4986 /* At this point, ELT, if nonzero, points to a class of expressions
4987 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4988 and SRC_RELATED, if nonzero, each contain additional equivalent
4989 expressions. Prune these latter expressions by deleting expressions
4990 already in the equivalence class.
4992 Check for an equivalent identical to the destination. If found,
4993 this is the preferred equivalent since it will likely lead to
4994 elimination of the insn. Indicate this by placing it in
4995 `src_related'. */
4997 if (elt)
4998 elt = elt->first_same_value;
4999 for (p = elt; p; p = p->next_same_value)
5001 enum rtx_code code = GET_CODE (p->exp);
5003 /* If the expression is not valid, ignore it. Then we do not
5004 have to check for validity below. In most cases, we can use
5005 `rtx_equal_p', since canonicalization has already been done. */
5006 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
5007 continue;
5009 /* Also skip paradoxical subregs, unless that's what we're
5010 looking for. */
5011 if (paradoxical_subreg_p (p->exp)
5012 && ! (src != 0
5013 && GET_CODE (src) == SUBREG
5014 && GET_MODE (src) == GET_MODE (p->exp)
5015 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5016 GET_MODE (SUBREG_REG (p->exp)))))
5017 continue;
5019 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5020 src = 0;
5021 else if (src_folded && GET_CODE (src_folded) == code
5022 && rtx_equal_p (src_folded, p->exp))
5023 src_folded = 0;
5024 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5025 && rtx_equal_p (src_eqv_here, p->exp))
5026 src_eqv_here = 0;
5027 else if (src_related && GET_CODE (src_related) == code
5028 && rtx_equal_p (src_related, p->exp))
5029 src_related = 0;
5031 /* This is the same as the destination of the insns, we want
5032 to prefer it. Copy it to src_related. The code below will
5033 then give it a negative cost. */
5034 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5035 src_related = dest;
5038 /* Find the cheapest valid equivalent, trying all the available
5039 possibilities. Prefer items not in the hash table to ones
5040 that are when they are equal cost. Note that we can never
5041 worsen an insn as the current contents will also succeed.
5042 If we find an equivalent identical to the destination, use it as best,
5043 since this insn will probably be eliminated in that case. */
5044 if (src)
5046 if (rtx_equal_p (src, dest))
5047 src_cost = src_regcost = -1;
5048 else
5050 src_cost = COST (src, mode);
5051 src_regcost = approx_reg_cost (src);
5055 if (src_eqv_here)
5057 if (rtx_equal_p (src_eqv_here, dest))
5058 src_eqv_cost = src_eqv_regcost = -1;
5059 else
5061 src_eqv_cost = COST (src_eqv_here, mode);
5062 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5066 if (src_folded)
5068 if (rtx_equal_p (src_folded, dest))
5069 src_folded_cost = src_folded_regcost = -1;
5070 else
5072 src_folded_cost = COST (src_folded, mode);
5073 src_folded_regcost = approx_reg_cost (src_folded);
5077 if (src_related)
5079 if (rtx_equal_p (src_related, dest))
5080 src_related_cost = src_related_regcost = -1;
5081 else
5083 src_related_cost = COST (src_related, mode);
5084 src_related_regcost = approx_reg_cost (src_related);
5086 /* If a const-anchor is used to synthesize a constant that
5087 normally requires multiple instructions then slightly prefer
5088 it over the original sequence. These instructions are likely
5089 to become redundant now. We can't compare against the cost
5090 of src_eqv_here because, on MIPS for example, multi-insn
5091 constants have zero cost; they are assumed to be hoisted from
5092 loops. */
5093 if (src_related_is_const_anchor
5094 && src_related_cost == src_cost
5095 && src_eqv_here)
5096 src_related_cost--;
5100 /* If this was an indirect jump insn, a known label will really be
5101 cheaper even though it looks more expensive. */
5102 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5103 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5105 /* Terminate loop when replacement made. This must terminate since
5106 the current contents will be tested and will always be valid. */
5107 while (1)
5109 rtx trial;
5111 /* Skip invalid entries. */
5112 while (elt && !REG_P (elt->exp)
5113 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5114 elt = elt->next_same_value;
5116 /* A paradoxical subreg would be bad here: it'll be the right
5117 size, but later may be adjusted so that the upper bits aren't
5118 what we want. So reject it. */
5119 if (elt != 0
5120 && paradoxical_subreg_p (elt->exp)
5121 /* It is okay, though, if the rtx we're trying to match
5122 will ignore any of the bits we can't predict. */
5123 && ! (src != 0
5124 && GET_CODE (src) == SUBREG
5125 && GET_MODE (src) == GET_MODE (elt->exp)
5126 && partial_subreg_p (GET_MODE (SUBREG_REG (src)),
5127 GET_MODE (SUBREG_REG (elt->exp)))))
5129 elt = elt->next_same_value;
5130 continue;
5133 if (elt)
5135 src_elt_cost = elt->cost;
5136 src_elt_regcost = elt->regcost;
5139 /* Find cheapest and skip it for the next time. For items
5140 of equal cost, use this order:
5141 src_folded, src, src_eqv, src_related and hash table entry. */
5142 if (src_folded
5143 && preferable (src_folded_cost, src_folded_regcost,
5144 src_cost, src_regcost) <= 0
5145 && preferable (src_folded_cost, src_folded_regcost,
5146 src_eqv_cost, src_eqv_regcost) <= 0
5147 && preferable (src_folded_cost, src_folded_regcost,
5148 src_related_cost, src_related_regcost) <= 0
5149 && preferable (src_folded_cost, src_folded_regcost,
5150 src_elt_cost, src_elt_regcost) <= 0)
5152 trial = src_folded, src_folded_cost = MAX_COST;
5153 if (src_folded_force_flag)
5155 rtx forced = force_const_mem (mode, trial);
5156 if (forced)
5157 trial = forced;
5160 else if (src
5161 && preferable (src_cost, src_regcost,
5162 src_eqv_cost, src_eqv_regcost) <= 0
5163 && preferable (src_cost, src_regcost,
5164 src_related_cost, src_related_regcost) <= 0
5165 && preferable (src_cost, src_regcost,
5166 src_elt_cost, src_elt_regcost) <= 0)
5167 trial = src, src_cost = MAX_COST;
5168 else if (src_eqv_here
5169 && preferable (src_eqv_cost, src_eqv_regcost,
5170 src_related_cost, src_related_regcost) <= 0
5171 && preferable (src_eqv_cost, src_eqv_regcost,
5172 src_elt_cost, src_elt_regcost) <= 0)
5173 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5174 else if (src_related
5175 && preferable (src_related_cost, src_related_regcost,
5176 src_elt_cost, src_elt_regcost) <= 0)
5177 trial = src_related, src_related_cost = MAX_COST;
5178 else
5180 trial = elt->exp;
5181 elt = elt->next_same_value;
5182 src_elt_cost = MAX_COST;
5185 /* Avoid creation of overlapping memory moves. */
5186 if (MEM_P (trial) && MEM_P (dest) && !rtx_equal_p (trial, dest))
5188 rtx src, dest;
5190 /* BLKmode moves are not handled by cse anyway. */
5191 if (GET_MODE (trial) == BLKmode)
5192 break;
5194 src = canon_rtx (trial);
5195 dest = canon_rtx (SET_DEST (sets[i].rtl));
5197 if (!MEM_P (src) || !MEM_P (dest)
5198 || !nonoverlapping_memrefs_p (src, dest, false))
5199 break;
5202 /* Try to optimize
5203 (set (reg:M N) (const_int A))
5204 (set (reg:M2 O) (const_int B))
5205 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5206 (reg:M2 O)). */
5207 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5208 && CONST_INT_P (trial)
5209 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5210 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5211 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5212 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5213 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5214 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5215 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5216 <= HOST_BITS_PER_WIDE_INT))
5218 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5219 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5220 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5221 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5222 struct table_elt *dest_elt
5223 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5224 rtx dest_cst = NULL;
5226 if (dest_elt)
5227 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5228 if (p->is_const && CONST_INT_P (p->exp))
5230 dest_cst = p->exp;
5231 break;
5233 if (dest_cst)
5235 HOST_WIDE_INT val = INTVAL (dest_cst);
5236 HOST_WIDE_INT mask;
5237 unsigned int shift;
5238 /* This is the mode of DEST_CST as well. */
5239 scalar_int_mode dest_mode
5240 = as_a <scalar_int_mode> (GET_MODE (dest_reg));
5241 if (BITS_BIG_ENDIAN)
5242 shift = GET_MODE_PRECISION (dest_mode)
5243 - INTVAL (pos) - INTVAL (width);
5244 else
5245 shift = INTVAL (pos);
5246 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5247 mask = HOST_WIDE_INT_M1;
5248 else
5249 mask = (HOST_WIDE_INT_1 << INTVAL (width)) - 1;
5250 val &= ~(mask << shift);
5251 val |= (INTVAL (trial) & mask) << shift;
5252 val = trunc_int_for_mode (val, dest_mode);
5253 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5254 dest_reg, 1);
5255 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5256 GEN_INT (val), 1);
5257 if (apply_change_group ())
5259 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5260 if (note)
5262 remove_note (insn, note);
5263 df_notes_rescan (insn);
5265 src_eqv = NULL_RTX;
5266 src_eqv_elt = NULL;
5267 src_eqv_volatile = 0;
5268 src_eqv_in_memory = 0;
5269 src_eqv_hash = 0;
5270 repeat = true;
5271 break;
5276 /* We don't normally have an insn matching (set (pc) (pc)), so
5277 check for this separately here. We will delete such an
5278 insn below.
5280 For other cases such as a table jump or conditional jump
5281 where we know the ultimate target, go ahead and replace the
5282 operand. While that may not make a valid insn, we will
5283 reemit the jump below (and also insert any necessary
5284 barriers). */
5285 if (n_sets == 1 && dest == pc_rtx
5286 && (trial == pc_rtx
5287 || (GET_CODE (trial) == LABEL_REF
5288 && ! condjump_p (insn))))
5290 /* Don't substitute non-local labels, this confuses CFG. */
5291 if (GET_CODE (trial) == LABEL_REF
5292 && LABEL_REF_NONLOCAL_P (trial))
5293 continue;
5295 SET_SRC (sets[i].rtl) = trial;
5296 cse_jumps_altered = true;
5297 break;
5300 /* Similarly, lots of targets don't allow no-op
5301 (set (mem x) (mem x)) moves. */
5302 else if (n_sets == 1
5303 && MEM_P (trial)
5304 && MEM_P (dest)
5305 && rtx_equal_p (trial, dest)
5306 && !side_effects_p (dest)
5307 && (cfun->can_delete_dead_exceptions
5308 || insn_nothrow_p (insn)))
5310 SET_SRC (sets[i].rtl) = trial;
5311 mem_noop_insn = true;
5312 break;
5315 /* Reject certain invalid forms of CONST that we create. */
5316 else if (CONSTANT_P (trial)
5317 && GET_CODE (trial) == CONST
5318 /* Reject cases that will cause decode_rtx_const to
5319 die. On the alpha when simplifying a switch, we
5320 get (const (truncate (minus (label_ref)
5321 (label_ref)))). */
5322 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5323 /* Likewise on IA-64, except without the
5324 truncate. */
5325 || (GET_CODE (XEXP (trial, 0)) == MINUS
5326 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5327 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5328 /* Do nothing for this case. */
5331 /* Look for a substitution that makes a valid insn. */
5332 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5333 trial, 0))
5335 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5337 /* The result of apply_change_group can be ignored; see
5338 canon_reg. */
5340 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5341 apply_change_group ();
5343 break;
5346 /* If we previously found constant pool entries for
5347 constants and this is a constant, try making a
5348 pool entry. Put it in src_folded unless we already have done
5349 this since that is where it likely came from. */
5351 else if (constant_pool_entries_cost
5352 && CONSTANT_P (trial)
5353 && (src_folded == 0
5354 || (!MEM_P (src_folded)
5355 && ! src_folded_force_flag))
5356 && GET_MODE_CLASS (mode) != MODE_CC
5357 && mode != VOIDmode)
5359 src_folded_force_flag = 1;
5360 src_folded = trial;
5361 src_folded_cost = constant_pool_entries_cost;
5362 src_folded_regcost = constant_pool_entries_regcost;
5366 /* If we changed the insn too much, handle this set from scratch. */
5367 if (repeat)
5369 i--;
5370 continue;
5373 src = SET_SRC (sets[i].rtl);
5375 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5376 However, there is an important exception: If both are registers
5377 that are not the head of their equivalence class, replace SET_SRC
5378 with the head of the class. If we do not do this, we will have
5379 both registers live over a portion of the basic block. This way,
5380 their lifetimes will likely abut instead of overlapping. */
5381 if (REG_P (dest)
5382 && REGNO_QTY_VALID_P (REGNO (dest)))
5384 int dest_q = REG_QTY (REGNO (dest));
5385 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5387 if (dest_ent->mode == GET_MODE (dest)
5388 && dest_ent->first_reg != REGNO (dest)
5389 && REG_P (src) && REGNO (src) == REGNO (dest)
5390 /* Don't do this if the original insn had a hard reg as
5391 SET_SRC or SET_DEST. */
5392 && (!REG_P (sets[i].src)
5393 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5394 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5395 /* We can't call canon_reg here because it won't do anything if
5396 SRC is a hard register. */
5398 int src_q = REG_QTY (REGNO (src));
5399 struct qty_table_elem *src_ent = &qty_table[src_q];
5400 int first = src_ent->first_reg;
5401 rtx new_src
5402 = (first >= FIRST_PSEUDO_REGISTER
5403 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5405 /* We must use validate-change even for this, because this
5406 might be a special no-op instruction, suitable only to
5407 tag notes onto. */
5408 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5410 src = new_src;
5411 /* If we had a constant that is cheaper than what we are now
5412 setting SRC to, use that constant. We ignored it when we
5413 thought we could make this into a no-op. */
5414 if (src_const && COST (src_const, mode) < COST (src, mode)
5415 && validate_change (insn, &SET_SRC (sets[i].rtl),
5416 src_const, 0))
5417 src = src_const;
5422 /* If we made a change, recompute SRC values. */
5423 if (src != sets[i].src)
5425 do_not_record = 0;
5426 hash_arg_in_memory = 0;
5427 sets[i].src = src;
5428 sets[i].src_hash = HASH (src, mode);
5429 sets[i].src_volatile = do_not_record;
5430 sets[i].src_in_memory = hash_arg_in_memory;
5431 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5434 /* If this is a single SET, we are setting a register, and we have an
5435 equivalent constant, we want to add a REG_EQUAL note if the constant
5436 is different from the source. We don't want to do it for a constant
5437 pseudo since verifying that this pseudo hasn't been eliminated is a
5438 pain; moreover such a note won't help anything.
5440 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5441 which can be created for a reference to a compile time computable
5442 entry in a jump table. */
5443 if (n_sets == 1
5444 && REG_P (dest)
5445 && src_const
5446 && !REG_P (src_const)
5447 && !(GET_CODE (src_const) == SUBREG
5448 && REG_P (SUBREG_REG (src_const)))
5449 && !(GET_CODE (src_const) == CONST
5450 && GET_CODE (XEXP (src_const, 0)) == MINUS
5451 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5452 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5453 && !rtx_equal_p (src, src_const))
5455 /* Make sure that the rtx is not shared. */
5456 src_const = copy_rtx (src_const);
5458 /* Record the actual constant value in a REG_EQUAL note,
5459 making a new one if one does not already exist. */
5460 set_unique_reg_note (insn, REG_EQUAL, src_const);
5461 df_notes_rescan (insn);
5464 /* Now deal with the destination. */
5465 do_not_record = 0;
5467 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5468 while (GET_CODE (dest) == SUBREG
5469 || GET_CODE (dest) == ZERO_EXTRACT
5470 || GET_CODE (dest) == STRICT_LOW_PART)
5471 dest = XEXP (dest, 0);
5473 sets[i].inner_dest = dest;
5475 if (MEM_P (dest))
5477 #ifdef PUSH_ROUNDING
5478 /* Stack pushes invalidate the stack pointer. */
5479 rtx addr = XEXP (dest, 0);
5480 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5481 && XEXP (addr, 0) == stack_pointer_rtx)
5482 invalidate (stack_pointer_rtx, VOIDmode);
5483 #endif
5484 dest = fold_rtx (dest, insn);
5487 /* Compute the hash code of the destination now,
5488 before the effects of this instruction are recorded,
5489 since the register values used in the address computation
5490 are those before this instruction. */
5491 sets[i].dest_hash = HASH (dest, mode);
5493 /* Don't enter a bit-field in the hash table
5494 because the value in it after the store
5495 may not equal what was stored, due to truncation. */
5497 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5499 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5501 if (src_const != 0 && CONST_INT_P (src_const)
5502 && CONST_INT_P (width)
5503 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5504 && ! (INTVAL (src_const)
5505 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5506 /* Exception: if the value is constant,
5507 and it won't be truncated, record it. */
5509 else
5511 /* This is chosen so that the destination will be invalidated
5512 but no new value will be recorded.
5513 We must invalidate because sometimes constant
5514 values can be recorded for bitfields. */
5515 sets[i].src_elt = 0;
5516 sets[i].src_volatile = 1;
5517 src_eqv = 0;
5518 src_eqv_elt = 0;
5522 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5523 the insn. */
5524 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5526 /* One less use of the label this insn used to jump to. */
5527 cse_cfg_altered |= delete_insn_and_edges (insn);
5528 cse_jumps_altered = true;
5529 /* No more processing for this set. */
5530 sets[i].rtl = 0;
5533 /* Similarly for no-op MEM moves. */
5534 else if (mem_noop_insn)
5536 if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
5537 cse_cfg_altered = true;
5538 cse_cfg_altered |= delete_insn_and_edges (insn);
5539 /* No more processing for this set. */
5540 sets[i].rtl = 0;
5543 /* If this SET is now setting PC to a label, we know it used to
5544 be a conditional or computed branch. */
5545 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5546 && !LABEL_REF_NONLOCAL_P (src))
5548 /* We reemit the jump in as many cases as possible just in
5549 case the form of an unconditional jump is significantly
5550 different than a computed jump or conditional jump.
5552 If this insn has multiple sets, then reemitting the
5553 jump is nontrivial. So instead we just force rerecognition
5554 and hope for the best. */
5555 if (n_sets == 1)
5557 rtx_jump_insn *new_rtx;
5558 rtx note;
5560 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5561 new_rtx = emit_jump_insn_before (seq, insn);
5562 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5563 LABEL_NUSES (XEXP (src, 0))++;
5565 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5566 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5567 if (note)
5569 XEXP (note, 1) = NULL_RTX;
5570 REG_NOTES (new_rtx) = note;
5573 cse_cfg_altered |= delete_insn_and_edges (insn);
5574 insn = new_rtx;
5576 else
5577 INSN_CODE (insn) = -1;
5579 /* Do not bother deleting any unreachable code, let jump do it. */
5580 cse_jumps_altered = true;
5581 sets[i].rtl = 0;
5584 /* If destination is volatile, invalidate it and then do no further
5585 processing for this assignment. */
5587 else if (do_not_record)
5589 invalidate_dest (dest);
5590 sets[i].rtl = 0;
5593 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5595 do_not_record = 0;
5596 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5597 if (do_not_record)
5599 invalidate_dest (SET_DEST (sets[i].rtl));
5600 sets[i].rtl = 0;
5604 /* If setting CC0, record what it was set to, or a constant, if it
5605 is equivalent to a constant. If it is being set to a floating-point
5606 value, make a COMPARE with the appropriate constant of 0. If we
5607 don't do this, later code can interpret this as a test against
5608 const0_rtx, which can cause problems if we try to put it into an
5609 insn as a floating-point operand. */
5610 if (dest == cc0_rtx)
5612 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5613 this_insn_cc0_mode = mode;
5614 if (FLOAT_MODE_P (mode))
5615 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5616 CONST0_RTX (mode));
5620 /* Now enter all non-volatile source expressions in the hash table
5621 if they are not already present.
5622 Record their equivalence classes in src_elt.
5623 This way we can insert the corresponding destinations into
5624 the same classes even if the actual sources are no longer in them
5625 (having been invalidated). */
5627 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5628 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5630 struct table_elt *elt;
5631 struct table_elt *classp = sets[0].src_elt;
5632 rtx dest = SET_DEST (sets[0].rtl);
5633 machine_mode eqvmode = GET_MODE (dest);
5635 if (GET_CODE (dest) == STRICT_LOW_PART)
5637 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5638 classp = 0;
5640 if (insert_regs (src_eqv, classp, 0))
5642 rehash_using_reg (src_eqv);
5643 src_eqv_hash = HASH (src_eqv, eqvmode);
5645 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5646 elt->in_memory = src_eqv_in_memory;
5647 src_eqv_elt = elt;
5649 /* Check to see if src_eqv_elt is the same as a set source which
5650 does not yet have an elt, and if so set the elt of the set source
5651 to src_eqv_elt. */
5652 for (i = 0; i < n_sets; i++)
5653 if (sets[i].rtl && sets[i].src_elt == 0
5654 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5655 sets[i].src_elt = src_eqv_elt;
5658 for (i = 0; i < n_sets; i++)
5659 if (sets[i].rtl && ! sets[i].src_volatile
5660 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5662 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5664 /* REG_EQUAL in setting a STRICT_LOW_PART
5665 gives an equivalent for the entire destination register,
5666 not just for the subreg being stored in now.
5667 This is a more interesting equivalence, so we arrange later
5668 to treat the entire reg as the destination. */
5669 sets[i].src_elt = src_eqv_elt;
5670 sets[i].src_hash = src_eqv_hash;
5672 else
5674 /* Insert source and constant equivalent into hash table, if not
5675 already present. */
5676 struct table_elt *classp = src_eqv_elt;
5677 rtx src = sets[i].src;
5678 rtx dest = SET_DEST (sets[i].rtl);
5679 machine_mode mode
5680 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5682 /* It's possible that we have a source value known to be
5683 constant but don't have a REG_EQUAL note on the insn.
5684 Lack of a note will mean src_eqv_elt will be NULL. This
5685 can happen where we've generated a SUBREG to access a
5686 CONST_INT that is already in a register in a wider mode.
5687 Ensure that the source expression is put in the proper
5688 constant class. */
5689 if (!classp)
5690 classp = sets[i].src_const_elt;
5692 if (sets[i].src_elt == 0)
5694 struct table_elt *elt;
5696 /* Note that these insert_regs calls cannot remove
5697 any of the src_elt's, because they would have failed to
5698 match if not still valid. */
5699 if (insert_regs (src, classp, 0))
5701 rehash_using_reg (src);
5702 sets[i].src_hash = HASH (src, mode);
5704 elt = insert (src, classp, sets[i].src_hash, mode);
5705 elt->in_memory = sets[i].src_in_memory;
5706 /* If inline asm has any clobbers, ensure we only reuse
5707 existing inline asms and never try to put the ASM_OPERANDS
5708 into an insn that isn't inline asm. */
5709 if (GET_CODE (src) == ASM_OPERANDS
5710 && GET_CODE (x) == PARALLEL)
5711 elt->cost = MAX_COST;
5712 sets[i].src_elt = classp = elt;
5714 if (sets[i].src_const && sets[i].src_const_elt == 0
5715 && src != sets[i].src_const
5716 && ! rtx_equal_p (sets[i].src_const, src))
5717 sets[i].src_elt = insert (sets[i].src_const, classp,
5718 sets[i].src_const_hash, mode);
5721 else if (sets[i].src_elt == 0)
5722 /* If we did not insert the source into the hash table (e.g., it was
5723 volatile), note the equivalence class for the REG_EQUAL value, if any,
5724 so that the destination goes into that class. */
5725 sets[i].src_elt = src_eqv_elt;
5727 /* Record destination addresses in the hash table. This allows us to
5728 check if they are invalidated by other sets. */
5729 for (i = 0; i < n_sets; i++)
5731 if (sets[i].rtl)
5733 rtx x = sets[i].inner_dest;
5734 struct table_elt *elt;
5735 machine_mode mode;
5736 unsigned hash;
5738 if (MEM_P (x))
5740 x = XEXP (x, 0);
5741 mode = GET_MODE (x);
5742 hash = HASH (x, mode);
5743 elt = lookup (x, hash, mode);
5744 if (!elt)
5746 if (insert_regs (x, NULL, 0))
5748 rtx dest = SET_DEST (sets[i].rtl);
5750 rehash_using_reg (x);
5751 hash = HASH (x, mode);
5752 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5754 elt = insert (x, NULL, hash, mode);
5757 sets[i].dest_addr_elt = elt;
5759 else
5760 sets[i].dest_addr_elt = NULL;
5764 invalidate_from_clobbers (insn);
5766 /* Some registers are invalidated by subroutine calls. Memory is
5767 invalidated by non-constant calls. */
5769 if (CALL_P (insn))
5771 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5772 invalidate_memory ();
5773 else
5774 /* For const/pure calls, invalidate any argument slots, because
5775 those are owned by the callee. */
5776 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
5777 if (GET_CODE (XEXP (tem, 0)) == USE
5778 && MEM_P (XEXP (XEXP (tem, 0), 0)))
5779 invalidate (XEXP (XEXP (tem, 0), 0), VOIDmode);
5780 invalidate_for_call ();
5783 /* Now invalidate everything set by this instruction.
5784 If a SUBREG or other funny destination is being set,
5785 sets[i].rtl is still nonzero, so here we invalidate the reg
5786 a part of which is being set. */
5788 for (i = 0; i < n_sets; i++)
5789 if (sets[i].rtl)
5791 /* We can't use the inner dest, because the mode associated with
5792 a ZERO_EXTRACT is significant. */
5793 rtx dest = SET_DEST (sets[i].rtl);
5795 /* Needed for registers to remove the register from its
5796 previous quantity's chain.
5797 Needed for memory if this is a nonvarying address, unless
5798 we have just done an invalidate_memory that covers even those. */
5799 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5800 invalidate (dest, VOIDmode);
5801 else if (MEM_P (dest))
5802 invalidate (dest, VOIDmode);
5803 else if (GET_CODE (dest) == STRICT_LOW_PART
5804 || GET_CODE (dest) == ZERO_EXTRACT)
5805 invalidate (XEXP (dest, 0), GET_MODE (dest));
5808 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5809 the regs restored by the longjmp come from a later time
5810 than the setjmp. */
5811 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5813 flush_hash_table ();
5814 goto done;
5817 /* Make sure registers mentioned in destinations
5818 are safe for use in an expression to be inserted.
5819 This removes from the hash table
5820 any invalid entry that refers to one of these registers.
5822 We don't care about the return value from mention_regs because
5823 we are going to hash the SET_DEST values unconditionally. */
5825 for (i = 0; i < n_sets; i++)
5827 if (sets[i].rtl)
5829 rtx x = SET_DEST (sets[i].rtl);
5831 if (!REG_P (x))
5832 mention_regs (x);
5833 else
5835 /* We used to rely on all references to a register becoming
5836 inaccessible when a register changes to a new quantity,
5837 since that changes the hash code. However, that is not
5838 safe, since after HASH_SIZE new quantities we get a
5839 hash 'collision' of a register with its own invalid
5840 entries. And since SUBREGs have been changed not to
5841 change their hash code with the hash code of the register,
5842 it wouldn't work any longer at all. So we have to check
5843 for any invalid references lying around now.
5844 This code is similar to the REG case in mention_regs,
5845 but it knows that reg_tick has been incremented, and
5846 it leaves reg_in_table as -1 . */
5847 unsigned int regno = REGNO (x);
5848 unsigned int endregno = END_REGNO (x);
5849 unsigned int i;
5851 for (i = regno; i < endregno; i++)
5853 if (REG_IN_TABLE (i) >= 0)
5855 remove_invalid_refs (i);
5856 REG_IN_TABLE (i) = -1;
5863 /* We may have just removed some of the src_elt's from the hash table.
5864 So replace each one with the current head of the same class.
5865 Also check if destination addresses have been removed. */
5867 for (i = 0; i < n_sets; i++)
5868 if (sets[i].rtl)
5870 if (sets[i].dest_addr_elt
5871 && sets[i].dest_addr_elt->first_same_value == 0)
5873 /* The elt was removed, which means this destination is not
5874 valid after this instruction. */
5875 sets[i].rtl = NULL_RTX;
5877 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5878 /* If elt was removed, find current head of same class,
5879 or 0 if nothing remains of that class. */
5881 struct table_elt *elt = sets[i].src_elt;
5883 while (elt && elt->prev_same_value)
5884 elt = elt->prev_same_value;
5886 while (elt && elt->first_same_value == 0)
5887 elt = elt->next_same_value;
5888 sets[i].src_elt = elt ? elt->first_same_value : 0;
5892 /* Now insert the destinations into their equivalence classes. */
5894 for (i = 0; i < n_sets; i++)
5895 if (sets[i].rtl)
5897 rtx dest = SET_DEST (sets[i].rtl);
5898 struct table_elt *elt;
5900 /* Don't record value if we are not supposed to risk allocating
5901 floating-point values in registers that might be wider than
5902 memory. */
5903 if ((flag_float_store
5904 && MEM_P (dest)
5905 && FLOAT_MODE_P (GET_MODE (dest)))
5906 /* Don't record BLKmode values, because we don't know the
5907 size of it, and can't be sure that other BLKmode values
5908 have the same or smaller size. */
5909 || GET_MODE (dest) == BLKmode
5910 /* If we didn't put a REG_EQUAL value or a source into the hash
5911 table, there is no point is recording DEST. */
5912 || sets[i].src_elt == 0)
5913 continue;
5915 /* STRICT_LOW_PART isn't part of the value BEING set,
5916 and neither is the SUBREG inside it.
5917 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5918 if (GET_CODE (dest) == STRICT_LOW_PART)
5919 dest = SUBREG_REG (XEXP (dest, 0));
5921 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5922 /* Registers must also be inserted into chains for quantities. */
5923 if (insert_regs (dest, sets[i].src_elt, 1))
5925 /* If `insert_regs' changes something, the hash code must be
5926 recalculated. */
5927 rehash_using_reg (dest);
5928 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5931 /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
5932 outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
5933 if (paradoxical_subreg_p (dest))
5934 continue;
5936 elt = insert (dest, sets[i].src_elt,
5937 sets[i].dest_hash, GET_MODE (dest));
5939 /* If this is a constant, insert the constant anchors with the
5940 equivalent register-offset expressions using register DEST. */
5941 if (targetm.const_anchor
5942 && REG_P (dest)
5943 && SCALAR_INT_MODE_P (GET_MODE (dest))
5944 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5945 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5947 elt->in_memory = (MEM_P (sets[i].inner_dest)
5948 && !MEM_READONLY_P (sets[i].inner_dest));
5950 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5951 narrower than M2, and both M1 and M2 are the same number of words,
5952 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5953 make that equivalence as well.
5955 However, BAR may have equivalences for which gen_lowpart
5956 will produce a simpler value than gen_lowpart applied to
5957 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5958 BAR's equivalences. If we don't get a simplified form, make
5959 the SUBREG. It will not be used in an equivalence, but will
5960 cause two similar assignments to be detected.
5962 Note the loop below will find SUBREG_REG (DEST) since we have
5963 already entered SRC and DEST of the SET in the table. */
5965 if (GET_CODE (dest) == SUBREG
5966 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5967 / UNITS_PER_WORD)
5968 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5969 && !partial_subreg_p (dest)
5970 && sets[i].src_elt != 0)
5972 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5973 struct table_elt *elt, *classp = 0;
5975 for (elt = sets[i].src_elt->first_same_value; elt;
5976 elt = elt->next_same_value)
5978 rtx new_src = 0;
5979 unsigned src_hash;
5980 struct table_elt *src_elt;
5982 /* Ignore invalid entries. */
5983 if (!REG_P (elt->exp)
5984 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5985 continue;
5987 /* We may have already been playing subreg games. If the
5988 mode is already correct for the destination, use it. */
5989 if (GET_MODE (elt->exp) == new_mode)
5990 new_src = elt->exp;
5991 else
5993 unsigned int byte
5994 = subreg_lowpart_offset (new_mode, GET_MODE (dest));
5995 new_src = simplify_gen_subreg (new_mode, elt->exp,
5996 GET_MODE (dest), byte);
5999 /* The call to simplify_gen_subreg fails if the value
6000 is VOIDmode, yet we can't do any simplification, e.g.
6001 for EXPR_LISTs denoting function call results.
6002 It is invalid to construct a SUBREG with a VOIDmode
6003 SUBREG_REG, hence a zero new_src means we can't do
6004 this substitution. */
6005 if (! new_src)
6006 continue;
6008 src_hash = HASH (new_src, new_mode);
6009 src_elt = lookup (new_src, src_hash, new_mode);
6011 /* Put the new source in the hash table is if isn't
6012 already. */
6013 if (src_elt == 0)
6015 if (insert_regs (new_src, classp, 0))
6017 rehash_using_reg (new_src);
6018 src_hash = HASH (new_src, new_mode);
6020 src_elt = insert (new_src, classp, src_hash, new_mode);
6021 src_elt->in_memory = elt->in_memory;
6022 if (GET_CODE (new_src) == ASM_OPERANDS
6023 && elt->cost == MAX_COST)
6024 src_elt->cost = MAX_COST;
6026 else if (classp && classp != src_elt->first_same_value)
6027 /* Show that two things that we've seen before are
6028 actually the same. */
6029 merge_equiv_classes (src_elt, classp);
6031 classp = src_elt->first_same_value;
6032 /* Ignore invalid entries. */
6033 while (classp
6034 && !REG_P (classp->exp)
6035 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
6036 classp = classp->next_same_value;
6041 /* Special handling for (set REG0 REG1) where REG0 is the
6042 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6043 be used in the sequel, so (if easily done) change this insn to
6044 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6045 that computed their value. Then REG1 will become a dead store
6046 and won't cloud the situation for later optimizations.
6048 Do not make this change if REG1 is a hard register, because it will
6049 then be used in the sequel and we may be changing a two-operand insn
6050 into a three-operand insn.
6052 Also do not do this if we are operating on a copy of INSN. */
6054 if (n_sets == 1 && sets[0].rtl)
6055 try_back_substitute_reg (sets[0].rtl, insn);
6057 done:;
6060 /* Remove from the hash table all expressions that reference memory. */
6062 static void
6063 invalidate_memory (void)
6065 int i;
6066 struct table_elt *p, *next;
6068 for (i = 0; i < HASH_SIZE; i++)
6069 for (p = table[i]; p; p = next)
6071 next = p->next_same_hash;
6072 if (p->in_memory)
6073 remove_from_table (p, i);
6077 /* Perform invalidation on the basis of everything about INSN,
6078 except for invalidating the actual places that are SET in it.
6079 This includes the places CLOBBERed, and anything that might
6080 alias with something that is SET or CLOBBERed. */
6082 static void
6083 invalidate_from_clobbers (rtx_insn *insn)
6085 rtx x = PATTERN (insn);
6087 if (GET_CODE (x) == CLOBBER)
6089 rtx ref = XEXP (x, 0);
6090 if (ref)
6092 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6093 || MEM_P (ref))
6094 invalidate (ref, VOIDmode);
6095 else if (GET_CODE (ref) == STRICT_LOW_PART
6096 || GET_CODE (ref) == ZERO_EXTRACT)
6097 invalidate (XEXP (ref, 0), GET_MODE (ref));
6100 else if (GET_CODE (x) == PARALLEL)
6102 int i;
6103 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6105 rtx y = XVECEXP (x, 0, i);
6106 if (GET_CODE (y) == CLOBBER)
6108 rtx ref = XEXP (y, 0);
6109 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6110 || MEM_P (ref))
6111 invalidate (ref, VOIDmode);
6112 else if (GET_CODE (ref) == STRICT_LOW_PART
6113 || GET_CODE (ref) == ZERO_EXTRACT)
6114 invalidate (XEXP (ref, 0), GET_MODE (ref));
6120 /* Perform invalidation on the basis of everything about INSN.
6121 This includes the places CLOBBERed, and anything that might
6122 alias with something that is SET or CLOBBERed. */
6124 static void
6125 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6127 rtx tem;
6128 rtx x = PATTERN (insn);
6130 if (CALL_P (insn))
6132 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6133 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6134 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6137 /* Ensure we invalidate the destination register of a CALL insn.
6138 This is necessary for machines where this register is a fixed_reg,
6139 because no other code would invalidate it. */
6140 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6141 invalidate (SET_DEST (x), VOIDmode);
6143 else if (GET_CODE (x) == PARALLEL)
6145 int i;
6147 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6149 rtx y = XVECEXP (x, 0, i);
6150 if (GET_CODE (y) == CLOBBER)
6152 rtx clobbered = XEXP (y, 0);
6154 if (REG_P (clobbered)
6155 || GET_CODE (clobbered) == SUBREG)
6156 invalidate (clobbered, VOIDmode);
6157 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6158 || GET_CODE (clobbered) == ZERO_EXTRACT)
6159 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6161 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6162 invalidate (SET_DEST (y), VOIDmode);
6167 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6168 and replace any registers in them with either an equivalent constant
6169 or the canonical form of the register. If we are inside an address,
6170 only do this if the address remains valid.
6172 OBJECT is 0 except when within a MEM in which case it is the MEM.
6174 Return the replacement for X. */
6176 static rtx
6177 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6179 enum rtx_code code = GET_CODE (x);
6180 const char *fmt = GET_RTX_FORMAT (code);
6181 int i;
6183 switch (code)
6185 case CONST:
6186 case SYMBOL_REF:
6187 case LABEL_REF:
6188 CASE_CONST_ANY:
6189 case PC:
6190 case CC0:
6191 case LO_SUM:
6192 return x;
6194 case MEM:
6195 validate_change (x, &XEXP (x, 0),
6196 cse_process_notes (XEXP (x, 0), x, changed), 0);
6197 return x;
6199 case EXPR_LIST:
6200 if (REG_NOTE_KIND (x) == REG_EQUAL)
6201 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6202 /* Fall through. */
6204 case INSN_LIST:
6205 case INT_LIST:
6206 if (XEXP (x, 1))
6207 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6208 return x;
6210 case SIGN_EXTEND:
6211 case ZERO_EXTEND:
6212 case SUBREG:
6214 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6215 /* We don't substitute VOIDmode constants into these rtx,
6216 since they would impede folding. */
6217 if (GET_MODE (new_rtx) != VOIDmode)
6218 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6219 return x;
6222 case UNSIGNED_FLOAT:
6224 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6225 /* We don't substitute negative VOIDmode constants into these rtx,
6226 since they would impede folding. */
6227 if (GET_MODE (new_rtx) != VOIDmode
6228 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6229 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6230 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6231 return x;
6234 case REG:
6235 i = REG_QTY (REGNO (x));
6237 /* Return a constant or a constant register. */
6238 if (REGNO_QTY_VALID_P (REGNO (x)))
6240 struct qty_table_elem *ent = &qty_table[i];
6242 if (ent->const_rtx != NULL_RTX
6243 && (CONSTANT_P (ent->const_rtx)
6244 || REG_P (ent->const_rtx)))
6246 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6247 if (new_rtx)
6248 return copy_rtx (new_rtx);
6252 /* Otherwise, canonicalize this register. */
6253 return canon_reg (x, NULL);
6255 default:
6256 break;
6259 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6260 if (fmt[i] == 'e')
6261 validate_change (object, &XEXP (x, i),
6262 cse_process_notes (XEXP (x, i), object, changed), 0);
6264 return x;
6267 static rtx
6268 cse_process_notes (rtx x, rtx object, bool *changed)
6270 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6271 if (new_rtx != x)
6272 *changed = true;
6273 return new_rtx;
6277 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6279 DATA is a pointer to a struct cse_basic_block_data, that is used to
6280 describe the path.
6281 It is filled with a queue of basic blocks, starting with FIRST_BB
6282 and following a trace through the CFG.
6284 If all paths starting at FIRST_BB have been followed, or no new path
6285 starting at FIRST_BB can be constructed, this function returns FALSE.
6286 Otherwise, DATA->path is filled and the function returns TRUE indicating
6287 that a path to follow was found.
6289 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6290 block in the path will be FIRST_BB. */
6292 static bool
6293 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6294 int follow_jumps)
6296 basic_block bb;
6297 edge e;
6298 int path_size;
6300 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6302 /* See if there is a previous path. */
6303 path_size = data->path_size;
6305 /* There is a previous path. Make sure it started with FIRST_BB. */
6306 if (path_size)
6307 gcc_assert (data->path[0].bb == first_bb);
6309 /* There was only one basic block in the last path. Clear the path and
6310 return, so that paths starting at another basic block can be tried. */
6311 if (path_size == 1)
6313 path_size = 0;
6314 goto done;
6317 /* If the path was empty from the beginning, construct a new path. */
6318 if (path_size == 0)
6319 data->path[path_size++].bb = first_bb;
6320 else
6322 /* Otherwise, path_size must be equal to or greater than 2, because
6323 a previous path exists that is at least two basic blocks long.
6325 Update the previous branch path, if any. If the last branch was
6326 previously along the branch edge, take the fallthrough edge now. */
6327 while (path_size >= 2)
6329 basic_block last_bb_in_path, previous_bb_in_path;
6330 edge e;
6332 --path_size;
6333 last_bb_in_path = data->path[path_size].bb;
6334 previous_bb_in_path = data->path[path_size - 1].bb;
6336 /* If we previously followed a path along the branch edge, try
6337 the fallthru edge now. */
6338 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6339 && any_condjump_p (BB_END (previous_bb_in_path))
6340 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6341 && e == BRANCH_EDGE (previous_bb_in_path))
6343 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6344 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6345 && single_pred_p (bb)
6346 /* We used to assert here that we would only see blocks
6347 that we have not visited yet. But we may end up
6348 visiting basic blocks twice if the CFG has changed
6349 in this run of cse_main, because when the CFG changes
6350 the topological sort of the CFG also changes. A basic
6351 blocks that previously had more than two predecessors
6352 may now have a single predecessor, and become part of
6353 a path that starts at another basic block.
6355 We still want to visit each basic block only once, so
6356 halt the path here if we have already visited BB. */
6357 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6359 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6360 data->path[path_size++].bb = bb;
6361 break;
6365 data->path[path_size].bb = NULL;
6368 /* If only one block remains in the path, bail. */
6369 if (path_size == 1)
6371 path_size = 0;
6372 goto done;
6376 /* Extend the path if possible. */
6377 if (follow_jumps)
6379 bb = data->path[path_size - 1].bb;
6380 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6382 if (single_succ_p (bb))
6383 e = single_succ_edge (bb);
6384 else if (EDGE_COUNT (bb->succs) == 2
6385 && any_condjump_p (BB_END (bb)))
6387 /* First try to follow the branch. If that doesn't lead
6388 to a useful path, follow the fallthru edge. */
6389 e = BRANCH_EDGE (bb);
6390 if (!single_pred_p (e->dest))
6391 e = FALLTHRU_EDGE (bb);
6393 else
6394 e = NULL;
6396 if (e
6397 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6398 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6399 && single_pred_p (e->dest)
6400 /* Avoid visiting basic blocks twice. The large comment
6401 above explains why this can happen. */
6402 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6404 basic_block bb2 = e->dest;
6405 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6406 data->path[path_size++].bb = bb2;
6407 bb = bb2;
6409 else
6410 bb = NULL;
6414 done:
6415 data->path_size = path_size;
6416 return path_size != 0;
6419 /* Dump the path in DATA to file F. NSETS is the number of sets
6420 in the path. */
6422 static void
6423 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6425 int path_entry;
6427 fprintf (f, ";; Following path with %d sets: ", nsets);
6428 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6429 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6430 fputc ('\n', dump_file);
6431 fflush (f);
6435 /* Return true if BB has exception handling successor edges. */
6437 static bool
6438 have_eh_succ_edges (basic_block bb)
6440 edge e;
6441 edge_iterator ei;
6443 FOR_EACH_EDGE (e, ei, bb->succs)
6444 if (e->flags & EDGE_EH)
6445 return true;
6447 return false;
6451 /* Scan to the end of the path described by DATA. Return an estimate of
6452 the total number of SETs of all insns in the path. */
6454 static void
6455 cse_prescan_path (struct cse_basic_block_data *data)
6457 int nsets = 0;
6458 int path_size = data->path_size;
6459 int path_entry;
6461 /* Scan to end of each basic block in the path. */
6462 for (path_entry = 0; path_entry < path_size; path_entry++)
6464 basic_block bb;
6465 rtx_insn *insn;
6467 bb = data->path[path_entry].bb;
6469 FOR_BB_INSNS (bb, insn)
6471 if (!INSN_P (insn))
6472 continue;
6474 /* A PARALLEL can have lots of SETs in it,
6475 especially if it is really an ASM_OPERANDS. */
6476 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6477 nsets += XVECLEN (PATTERN (insn), 0);
6478 else
6479 nsets += 1;
6483 data->nsets = nsets;
6486 /* Return true if the pattern of INSN uses a LABEL_REF for which
6487 there isn't a REG_LABEL_OPERAND note. */
6489 static bool
6490 check_for_label_ref (rtx_insn *insn)
6492 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6493 note for it, we must rerun jump since it needs to place the note. If
6494 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6495 don't do this since no REG_LABEL_OPERAND will be added. */
6496 subrtx_iterator::array_type array;
6497 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6499 const_rtx x = *iter;
6500 if (GET_CODE (x) == LABEL_REF
6501 && !LABEL_REF_NONLOCAL_P (x)
6502 && (!JUMP_P (insn)
6503 || !label_is_jump_target_p (label_ref_label (x), insn))
6504 && LABEL_P (label_ref_label (x))
6505 && INSN_UID (label_ref_label (x)) != 0
6506 && !find_reg_note (insn, REG_LABEL_OPERAND, label_ref_label (x)))
6507 return true;
6509 return false;
6512 /* Process a single extended basic block described by EBB_DATA. */
6514 static void
6515 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6517 int path_size = ebb_data->path_size;
6518 int path_entry;
6519 int num_insns = 0;
6521 /* Allocate the space needed by qty_table. */
6522 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6524 new_basic_block ();
6525 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6526 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6527 for (path_entry = 0; path_entry < path_size; path_entry++)
6529 basic_block bb;
6530 rtx_insn *insn;
6532 bb = ebb_data->path[path_entry].bb;
6534 /* Invalidate recorded information for eh regs if there is an EH
6535 edge pointing to that bb. */
6536 if (bb_has_eh_pred (bb))
6538 df_ref def;
6540 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6541 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6542 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6545 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6546 FOR_BB_INSNS (bb, insn)
6548 /* If we have processed 1,000 insns, flush the hash table to
6549 avoid extreme quadratic behavior. We must not include NOTEs
6550 in the count since there may be more of them when generating
6551 debugging information. If we clear the table at different
6552 times, code generated with -g -O might be different than code
6553 generated with -O but not -g.
6555 FIXME: This is a real kludge and needs to be done some other
6556 way. */
6557 if (NONDEBUG_INSN_P (insn)
6558 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6560 flush_hash_table ();
6561 num_insns = 0;
6564 if (INSN_P (insn))
6566 /* Process notes first so we have all notes in canonical forms
6567 when looking for duplicate operations. */
6568 if (REG_NOTES (insn))
6570 bool changed = false;
6571 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6572 NULL_RTX, &changed);
6573 if (changed)
6574 df_notes_rescan (insn);
6577 cse_insn (insn);
6579 /* If we haven't already found an insn where we added a LABEL_REF,
6580 check this one. */
6581 if (INSN_P (insn) && !recorded_label_ref
6582 && check_for_label_ref (insn))
6583 recorded_label_ref = true;
6585 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6587 /* If the previous insn sets CC0 and this insn no
6588 longer references CC0, delete the previous insn.
6589 Here we use fact that nothing expects CC0 to be
6590 valid over an insn, which is true until the final
6591 pass. */
6592 rtx_insn *prev_insn;
6593 rtx tem;
6595 prev_insn = prev_nonnote_nondebug_insn (insn);
6596 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6597 && (tem = single_set (prev_insn)) != NULL_RTX
6598 && SET_DEST (tem) == cc0_rtx
6599 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6600 delete_insn (prev_insn);
6602 /* If this insn is not the last insn in the basic
6603 block, it will be PREV_INSN(insn) in the next
6604 iteration. If we recorded any CC0-related
6605 information for this insn, remember it. */
6606 if (insn != BB_END (bb))
6608 prev_insn_cc0 = this_insn_cc0;
6609 prev_insn_cc0_mode = this_insn_cc0_mode;
6615 /* With non-call exceptions, we are not always able to update
6616 the CFG properly inside cse_insn. So clean up possibly
6617 redundant EH edges here. */
6618 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6619 cse_cfg_altered |= purge_dead_edges (bb);
6621 /* If we changed a conditional jump, we may have terminated
6622 the path we are following. Check that by verifying that
6623 the edge we would take still exists. If the edge does
6624 not exist anymore, purge the remainder of the path.
6625 Note that this will cause us to return to the caller. */
6626 if (path_entry < path_size - 1)
6628 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6629 if (!find_edge (bb, next_bb))
6633 path_size--;
6635 /* If we truncate the path, we must also reset the
6636 visited bit on the remaining blocks in the path,
6637 or we will never visit them at all. */
6638 bitmap_clear_bit (cse_visited_basic_blocks,
6639 ebb_data->path[path_size].bb->index);
6640 ebb_data->path[path_size].bb = NULL;
6642 while (path_size - 1 != path_entry);
6643 ebb_data->path_size = path_size;
6647 /* If this is a conditional jump insn, record any known
6648 equivalences due to the condition being tested. */
6649 insn = BB_END (bb);
6650 if (path_entry < path_size - 1
6651 && EDGE_COUNT (bb->succs) == 2
6652 && JUMP_P (insn)
6653 && single_set (insn)
6654 && any_condjump_p (insn))
6656 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6657 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6658 record_jump_equiv (insn, taken);
6661 /* Clear the CC0-tracking related insns, they can't provide
6662 useful information across basic block boundaries. */
6663 prev_insn_cc0 = 0;
6666 gcc_assert (next_qty <= max_qty);
6668 free (qty_table);
6672 /* Perform cse on the instructions of a function.
6673 F is the first instruction.
6674 NREGS is one plus the highest pseudo-reg number used in the instruction.
6676 Return 2 if jump optimizations should be redone due to simplifications
6677 in conditional jump instructions.
6678 Return 1 if the CFG should be cleaned up because it has been modified.
6679 Return 0 otherwise. */
6681 static int
6682 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6684 struct cse_basic_block_data ebb_data;
6685 basic_block bb;
6686 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6687 int i, n_blocks;
6689 /* CSE doesn't use dominane info but can invalidate it in different ways.
6690 For simplicity free dominance info here. */
6691 free_dominance_info (CDI_DOMINATORS);
6693 df_set_flags (DF_LR_RUN_DCE);
6694 df_note_add_problem ();
6695 df_analyze ();
6696 df_set_flags (DF_DEFER_INSN_RESCAN);
6698 reg_scan (get_insns (), max_reg_num ());
6699 init_cse_reg_info (nregs);
6701 ebb_data.path = XNEWVEC (struct branch_path,
6702 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6704 cse_cfg_altered = false;
6705 cse_jumps_altered = false;
6706 recorded_label_ref = false;
6707 constant_pool_entries_cost = 0;
6708 constant_pool_entries_regcost = 0;
6709 ebb_data.path_size = 0;
6710 ebb_data.nsets = 0;
6711 rtl_hooks = cse_rtl_hooks;
6713 init_recog ();
6714 init_alias_analysis ();
6716 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6718 /* Set up the table of already visited basic blocks. */
6719 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6720 bitmap_clear (cse_visited_basic_blocks);
6722 /* Loop over basic blocks in reverse completion order (RPO),
6723 excluding the ENTRY and EXIT blocks. */
6724 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6725 i = 0;
6726 while (i < n_blocks)
6728 /* Find the first block in the RPO queue that we have not yet
6729 processed before. */
6732 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6734 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6735 && i < n_blocks);
6737 /* Find all paths starting with BB, and process them. */
6738 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6740 /* Pre-scan the path. */
6741 cse_prescan_path (&ebb_data);
6743 /* If this basic block has no sets, skip it. */
6744 if (ebb_data.nsets == 0)
6745 continue;
6747 /* Get a reasonable estimate for the maximum number of qty's
6748 needed for this path. For this, we take the number of sets
6749 and multiply that by MAX_RECOG_OPERANDS. */
6750 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6752 /* Dump the path we're about to process. */
6753 if (dump_file)
6754 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6756 cse_extended_basic_block (&ebb_data);
6760 /* Clean up. */
6761 end_alias_analysis ();
6762 free (reg_eqv_table);
6763 free (ebb_data.path);
6764 sbitmap_free (cse_visited_basic_blocks);
6765 free (rc_order);
6766 rtl_hooks = general_rtl_hooks;
6768 if (cse_jumps_altered || recorded_label_ref)
6769 return 2;
6770 else if (cse_cfg_altered)
6771 return 1;
6772 else
6773 return 0;
6776 /* Count the number of times registers are used (not set) in X.
6777 COUNTS is an array in which we accumulate the count, INCR is how much
6778 we count each register usage.
6780 Don't count a usage of DEST, which is the SET_DEST of a SET which
6781 contains X in its SET_SRC. This is because such a SET does not
6782 modify the liveness of DEST.
6783 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6784 We must then count uses of a SET_DEST regardless, because the insn can't be
6785 deleted here. */
6787 static void
6788 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6790 enum rtx_code code;
6791 rtx note;
6792 const char *fmt;
6793 int i, j;
6795 if (x == 0)
6796 return;
6798 switch (code = GET_CODE (x))
6800 case REG:
6801 if (x != dest)
6802 counts[REGNO (x)] += incr;
6803 return;
6805 case PC:
6806 case CC0:
6807 case CONST:
6808 CASE_CONST_ANY:
6809 case SYMBOL_REF:
6810 case LABEL_REF:
6811 return;
6813 case CLOBBER:
6814 /* If we are clobbering a MEM, mark any registers inside the address
6815 as being used. */
6816 if (MEM_P (XEXP (x, 0)))
6817 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6818 return;
6820 case SET:
6821 /* Unless we are setting a REG, count everything in SET_DEST. */
6822 if (!REG_P (SET_DEST (x)))
6823 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6824 count_reg_usage (SET_SRC (x), counts,
6825 dest ? dest : SET_DEST (x),
6826 incr);
6827 return;
6829 case DEBUG_INSN:
6830 return;
6832 case CALL_INSN:
6833 case INSN:
6834 case JUMP_INSN:
6835 /* We expect dest to be NULL_RTX here. If the insn may throw,
6836 or if it cannot be deleted due to side-effects, mark this fact
6837 by setting DEST to pc_rtx. */
6838 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6839 || side_effects_p (PATTERN (x)))
6840 dest = pc_rtx;
6841 if (code == CALL_INSN)
6842 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6843 count_reg_usage (PATTERN (x), counts, dest, incr);
6845 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6846 use them. */
6848 note = find_reg_equal_equiv_note (x);
6849 if (note)
6851 rtx eqv = XEXP (note, 0);
6853 if (GET_CODE (eqv) == EXPR_LIST)
6854 /* This REG_EQUAL note describes the result of a function call.
6855 Process all the arguments. */
6858 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6859 eqv = XEXP (eqv, 1);
6861 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6862 else
6863 count_reg_usage (eqv, counts, dest, incr);
6865 return;
6867 case EXPR_LIST:
6868 if (REG_NOTE_KIND (x) == REG_EQUAL
6869 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6870 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6871 involving registers in the address. */
6872 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6873 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6875 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6876 return;
6878 case ASM_OPERANDS:
6879 /* Iterate over just the inputs, not the constraints as well. */
6880 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6881 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6882 return;
6884 case INSN_LIST:
6885 case INT_LIST:
6886 gcc_unreachable ();
6888 default:
6889 break;
6892 fmt = GET_RTX_FORMAT (code);
6893 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6895 if (fmt[i] == 'e')
6896 count_reg_usage (XEXP (x, i), counts, dest, incr);
6897 else if (fmt[i] == 'E')
6898 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6899 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6903 /* Return true if X is a dead register. */
6905 static inline int
6906 is_dead_reg (const_rtx x, int *counts)
6908 return (REG_P (x)
6909 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6910 && counts[REGNO (x)] == 0);
6913 /* Return true if set is live. */
6914 static bool
6915 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6916 int *counts)
6918 rtx_insn *tem;
6920 if (set_noop_p (set))
6923 else if (GET_CODE (SET_DEST (set)) == CC0
6924 && !side_effects_p (SET_SRC (set))
6925 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6926 || !INSN_P (tem)
6927 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6928 return false;
6929 else if (!is_dead_reg (SET_DEST (set), counts)
6930 || side_effects_p (SET_SRC (set)))
6931 return true;
6932 return false;
6935 /* Return true if insn is live. */
6937 static bool
6938 insn_live_p (rtx_insn *insn, int *counts)
6940 int i;
6941 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6942 return true;
6943 else if (GET_CODE (PATTERN (insn)) == SET)
6944 return set_live_p (PATTERN (insn), insn, counts);
6945 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6947 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6949 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6951 if (GET_CODE (elt) == SET)
6953 if (set_live_p (elt, insn, counts))
6954 return true;
6956 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6957 return true;
6959 return false;
6961 else if (DEBUG_INSN_P (insn))
6963 rtx_insn *next;
6965 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6966 if (NOTE_P (next))
6967 continue;
6968 else if (!DEBUG_INSN_P (next))
6969 return true;
6970 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6971 return false;
6973 return true;
6975 else
6976 return true;
6979 /* Count the number of stores into pseudo. Callback for note_stores. */
6981 static void
6982 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6984 int *counts = (int *) data;
6985 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6986 counts[REGNO (x)]++;
6989 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6990 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6991 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6992 Set *SEEN_REPL to true if we see a dead register that does have
6993 a replacement. */
6995 static bool
6996 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6997 bool *seen_repl)
6999 subrtx_iterator::array_type array;
7000 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
7002 const_rtx x = *iter;
7003 if (is_dead_reg (x, counts))
7005 if (replacements && replacements[REGNO (x)] != NULL_RTX)
7006 *seen_repl = true;
7007 else
7008 return true;
7011 return false;
7014 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
7015 Callback for simplify_replace_fn_rtx. */
7017 static rtx
7018 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
7020 rtx *replacements = (rtx *) data;
7022 if (REG_P (x)
7023 && REGNO (x) >= FIRST_PSEUDO_REGISTER
7024 && replacements[REGNO (x)] != NULL_RTX)
7026 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
7027 return replacements[REGNO (x)];
7028 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
7029 GET_MODE (replacements[REGNO (x)]));
7031 return NULL_RTX;
7034 /* Scan all the insns and delete any that are dead; i.e., they store a register
7035 that is never used or they copy a register to itself.
7037 This is used to remove insns made obviously dead by cse, loop or other
7038 optimizations. It improves the heuristics in loop since it won't try to
7039 move dead invariants out of loops or make givs for dead quantities. The
7040 remaining passes of the compilation are also sped up. */
7043 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
7045 int *counts;
7046 rtx_insn *insn, *prev;
7047 rtx *replacements = NULL;
7048 int ndead = 0;
7050 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7051 /* First count the number of times each register is used. */
7052 if (MAY_HAVE_DEBUG_INSNS)
7054 counts = XCNEWVEC (int, nreg * 3);
7055 for (insn = insns; insn; insn = NEXT_INSN (insn))
7056 if (DEBUG_INSN_P (insn))
7057 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7058 NULL_RTX, 1);
7059 else if (INSN_P (insn))
7061 count_reg_usage (insn, counts, NULL_RTX, 1);
7062 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7064 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7065 First one counts how many times each pseudo is used outside
7066 of debug insns, second counts how many times each pseudo is
7067 used in debug insns and third counts how many times a pseudo
7068 is stored. */
7070 else
7072 counts = XCNEWVEC (int, nreg);
7073 for (insn = insns; insn; insn = NEXT_INSN (insn))
7074 if (INSN_P (insn))
7075 count_reg_usage (insn, counts, NULL_RTX, 1);
7076 /* If no debug insns can be present, COUNTS is just an array
7077 which counts how many times each pseudo is used. */
7079 /* Pseudo PIC register should be considered as used due to possible
7080 new usages generated. */
7081 if (!reload_completed
7082 && pic_offset_table_rtx
7083 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7084 counts[REGNO (pic_offset_table_rtx)]++;
7085 /* Go from the last insn to the first and delete insns that only set unused
7086 registers or copy a register to itself. As we delete an insn, remove
7087 usage counts for registers it uses.
7089 The first jump optimization pass may leave a real insn as the last
7090 insn in the function. We must not skip that insn or we may end
7091 up deleting code that is not really dead.
7093 If some otherwise unused register is only used in DEBUG_INSNs,
7094 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7095 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7096 has been created for the unused register, replace it with
7097 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7098 for (insn = get_last_insn (); insn; insn = prev)
7100 int live_insn = 0;
7102 prev = PREV_INSN (insn);
7103 if (!INSN_P (insn))
7104 continue;
7106 live_insn = insn_live_p (insn, counts);
7108 /* If this is a dead insn, delete it and show registers in it aren't
7109 being used. */
7111 if (! live_insn && dbg_cnt (delete_trivial_dead))
7113 if (DEBUG_INSN_P (insn))
7114 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7115 NULL_RTX, -1);
7116 else
7118 rtx set;
7119 if (MAY_HAVE_DEBUG_INSNS
7120 && (set = single_set (insn)) != NULL_RTX
7121 && is_dead_reg (SET_DEST (set), counts)
7122 /* Used at least once in some DEBUG_INSN. */
7123 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7124 /* And set exactly once. */
7125 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7126 && !side_effects_p (SET_SRC (set))
7127 && asm_noperands (PATTERN (insn)) < 0)
7129 rtx dval, bind_var_loc;
7130 rtx_insn *bind;
7132 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7133 dval = make_debug_expr_from_rtl (SET_DEST (set));
7135 /* Emit a debug bind insn before the insn in which
7136 reg dies. */
7137 bind_var_loc =
7138 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7139 DEBUG_EXPR_TREE_DECL (dval),
7140 SET_SRC (set),
7141 VAR_INIT_STATUS_INITIALIZED);
7142 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7144 bind = emit_debug_insn_before (bind_var_loc, insn);
7145 df_insn_rescan (bind);
7147 if (replacements == NULL)
7148 replacements = XCNEWVEC (rtx, nreg);
7149 replacements[REGNO (SET_DEST (set))] = dval;
7152 count_reg_usage (insn, counts, NULL_RTX, -1);
7153 ndead++;
7155 cse_cfg_altered |= delete_insn_and_edges (insn);
7159 if (MAY_HAVE_DEBUG_INSNS)
7161 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7162 if (DEBUG_INSN_P (insn))
7164 /* If this debug insn references a dead register that wasn't replaced
7165 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7166 bool seen_repl = false;
7167 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7168 counts, replacements, &seen_repl))
7170 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7171 df_insn_rescan (insn);
7173 else if (seen_repl)
7175 INSN_VAR_LOCATION_LOC (insn)
7176 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7177 NULL_RTX, replace_dead_reg,
7178 replacements);
7179 df_insn_rescan (insn);
7182 free (replacements);
7185 if (dump_file && ndead)
7186 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7187 ndead);
7188 /* Clean up. */
7189 free (counts);
7190 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7191 return ndead;
7194 /* If LOC contains references to NEWREG in a different mode, change them
7195 to use NEWREG instead. */
7197 static void
7198 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7199 rtx *loc, rtx_insn *insn, rtx newreg)
7201 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7203 rtx *loc = *iter;
7204 rtx x = *loc;
7205 if (x
7206 && REG_P (x)
7207 && REGNO (x) == REGNO (newreg)
7208 && GET_MODE (x) != GET_MODE (newreg))
7210 validate_change (insn, loc, newreg, 1);
7211 iter.skip_subrtxes ();
7216 /* Change the mode of any reference to the register REGNO (NEWREG) to
7217 GET_MODE (NEWREG) in INSN. */
7219 static void
7220 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7222 int success;
7224 if (!INSN_P (insn))
7225 return;
7227 subrtx_ptr_iterator::array_type array;
7228 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7229 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7231 /* If the following assertion was triggered, there is most probably
7232 something wrong with the cc_modes_compatible back end function.
7233 CC modes only can be considered compatible if the insn - with the mode
7234 replaced by any of the compatible modes - can still be recognized. */
7235 success = apply_change_group ();
7236 gcc_assert (success);
7239 /* Change the mode of any reference to the register REGNO (NEWREG) to
7240 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7241 any instruction which modifies NEWREG. */
7243 static void
7244 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7246 rtx_insn *insn;
7248 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7250 if (! INSN_P (insn))
7251 continue;
7253 if (reg_set_p (newreg, insn))
7254 return;
7256 cse_change_cc_mode_insn (insn, newreg);
7260 /* BB is a basic block which finishes with CC_REG as a condition code
7261 register which is set to CC_SRC. Look through the successors of BB
7262 to find blocks which have a single predecessor (i.e., this one),
7263 and look through those blocks for an assignment to CC_REG which is
7264 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7265 permitted to change the mode of CC_SRC to a compatible mode. This
7266 returns VOIDmode if no equivalent assignments were found.
7267 Otherwise it returns the mode which CC_SRC should wind up with.
7268 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7269 but is passed unmodified down to recursive calls in order to prevent
7270 endless recursion.
7272 The main complexity in this function is handling the mode issues.
7273 We may have more than one duplicate which we can eliminate, and we
7274 try to find a mode which will work for multiple duplicates. */
7276 static machine_mode
7277 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7278 bool can_change_mode)
7280 bool found_equiv;
7281 machine_mode mode;
7282 unsigned int insn_count;
7283 edge e;
7284 rtx_insn *insns[2];
7285 machine_mode modes[2];
7286 rtx_insn *last_insns[2];
7287 unsigned int i;
7288 rtx newreg;
7289 edge_iterator ei;
7291 /* We expect to have two successors. Look at both before picking
7292 the final mode for the comparison. If we have more successors
7293 (i.e., some sort of table jump, although that seems unlikely),
7294 then we require all beyond the first two to use the same
7295 mode. */
7297 found_equiv = false;
7298 mode = GET_MODE (cc_src);
7299 insn_count = 0;
7300 FOR_EACH_EDGE (e, ei, bb->succs)
7302 rtx_insn *insn;
7303 rtx_insn *end;
7305 if (e->flags & EDGE_COMPLEX)
7306 continue;
7308 if (EDGE_COUNT (e->dest->preds) != 1
7309 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7310 /* Avoid endless recursion on unreachable blocks. */
7311 || e->dest == orig_bb)
7312 continue;
7314 end = NEXT_INSN (BB_END (e->dest));
7315 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7317 rtx set;
7319 if (! INSN_P (insn))
7320 continue;
7322 /* If CC_SRC is modified, we have to stop looking for
7323 something which uses it. */
7324 if (modified_in_p (cc_src, insn))
7325 break;
7327 /* Check whether INSN sets CC_REG to CC_SRC. */
7328 set = single_set (insn);
7329 if (set
7330 && REG_P (SET_DEST (set))
7331 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7333 bool found;
7334 machine_mode set_mode;
7335 machine_mode comp_mode;
7337 found = false;
7338 set_mode = GET_MODE (SET_SRC (set));
7339 comp_mode = set_mode;
7340 if (rtx_equal_p (cc_src, SET_SRC (set)))
7341 found = true;
7342 else if (GET_CODE (cc_src) == COMPARE
7343 && GET_CODE (SET_SRC (set)) == COMPARE
7344 && mode != set_mode
7345 && rtx_equal_p (XEXP (cc_src, 0),
7346 XEXP (SET_SRC (set), 0))
7347 && rtx_equal_p (XEXP (cc_src, 1),
7348 XEXP (SET_SRC (set), 1)))
7351 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7352 if (comp_mode != VOIDmode
7353 && (can_change_mode || comp_mode == mode))
7354 found = true;
7357 if (found)
7359 found_equiv = true;
7360 if (insn_count < ARRAY_SIZE (insns))
7362 insns[insn_count] = insn;
7363 modes[insn_count] = set_mode;
7364 last_insns[insn_count] = end;
7365 ++insn_count;
7367 if (mode != comp_mode)
7369 gcc_assert (can_change_mode);
7370 mode = comp_mode;
7372 /* The modified insn will be re-recognized later. */
7373 PUT_MODE (cc_src, mode);
7376 else
7378 if (set_mode != mode)
7380 /* We found a matching expression in the
7381 wrong mode, but we don't have room to
7382 store it in the array. Punt. This case
7383 should be rare. */
7384 break;
7386 /* INSN sets CC_REG to a value equal to CC_SRC
7387 with the right mode. We can simply delete
7388 it. */
7389 delete_insn (insn);
7392 /* We found an instruction to delete. Keep looking,
7393 in the hopes of finding a three-way jump. */
7394 continue;
7397 /* We found an instruction which sets the condition
7398 code, so don't look any farther. */
7399 break;
7402 /* If INSN sets CC_REG in some other way, don't look any
7403 farther. */
7404 if (reg_set_p (cc_reg, insn))
7405 break;
7408 /* If we fell off the bottom of the block, we can keep looking
7409 through successors. We pass CAN_CHANGE_MODE as false because
7410 we aren't prepared to handle compatibility between the
7411 further blocks and this block. */
7412 if (insn == end)
7414 machine_mode submode;
7416 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7417 if (submode != VOIDmode)
7419 gcc_assert (submode == mode);
7420 found_equiv = true;
7421 can_change_mode = false;
7426 if (! found_equiv)
7427 return VOIDmode;
7429 /* Now INSN_COUNT is the number of instructions we found which set
7430 CC_REG to a value equivalent to CC_SRC. The instructions are in
7431 INSNS. The modes used by those instructions are in MODES. */
7433 newreg = NULL_RTX;
7434 for (i = 0; i < insn_count; ++i)
7436 if (modes[i] != mode)
7438 /* We need to change the mode of CC_REG in INSNS[i] and
7439 subsequent instructions. */
7440 if (! newreg)
7442 if (GET_MODE (cc_reg) == mode)
7443 newreg = cc_reg;
7444 else
7445 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7447 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7448 newreg);
7451 cse_cfg_altered |= delete_insn_and_edges (insns[i]);
7454 return mode;
7457 /* If we have a fixed condition code register (or two), walk through
7458 the instructions and try to eliminate duplicate assignments. */
7460 static void
7461 cse_condition_code_reg (void)
7463 unsigned int cc_regno_1;
7464 unsigned int cc_regno_2;
7465 rtx cc_reg_1;
7466 rtx cc_reg_2;
7467 basic_block bb;
7469 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7470 return;
7472 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7473 if (cc_regno_2 != INVALID_REGNUM)
7474 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7475 else
7476 cc_reg_2 = NULL_RTX;
7478 FOR_EACH_BB_FN (bb, cfun)
7480 rtx_insn *last_insn;
7481 rtx cc_reg;
7482 rtx_insn *insn;
7483 rtx_insn *cc_src_insn;
7484 rtx cc_src;
7485 machine_mode mode;
7486 machine_mode orig_mode;
7488 /* Look for blocks which end with a conditional jump based on a
7489 condition code register. Then look for the instruction which
7490 sets the condition code register. Then look through the
7491 successor blocks for instructions which set the condition
7492 code register to the same value. There are other possible
7493 uses of the condition code register, but these are by far the
7494 most common and the ones which we are most likely to be able
7495 to optimize. */
7497 last_insn = BB_END (bb);
7498 if (!JUMP_P (last_insn))
7499 continue;
7501 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7502 cc_reg = cc_reg_1;
7503 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7504 cc_reg = cc_reg_2;
7505 else
7506 continue;
7508 cc_src_insn = NULL;
7509 cc_src = NULL_RTX;
7510 for (insn = PREV_INSN (last_insn);
7511 insn && insn != PREV_INSN (BB_HEAD (bb));
7512 insn = PREV_INSN (insn))
7514 rtx set;
7516 if (! INSN_P (insn))
7517 continue;
7518 set = single_set (insn);
7519 if (set
7520 && REG_P (SET_DEST (set))
7521 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7523 cc_src_insn = insn;
7524 cc_src = SET_SRC (set);
7525 break;
7527 else if (reg_set_p (cc_reg, insn))
7528 break;
7531 if (! cc_src_insn)
7532 continue;
7534 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7535 continue;
7537 /* Now CC_REG is a condition code register used for a
7538 conditional jump at the end of the block, and CC_SRC, in
7539 CC_SRC_INSN, is the value to which that condition code
7540 register is set, and CC_SRC is still meaningful at the end of
7541 the basic block. */
7543 orig_mode = GET_MODE (cc_src);
7544 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7545 if (mode != VOIDmode)
7547 gcc_assert (mode == GET_MODE (cc_src));
7548 if (mode != orig_mode)
7550 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7552 cse_change_cc_mode_insn (cc_src_insn, newreg);
7554 /* Do the same in the following insns that use the
7555 current value of CC_REG within BB. */
7556 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7557 NEXT_INSN (last_insn),
7558 newreg);
7565 /* Perform common subexpression elimination. Nonzero value from
7566 `cse_main' means that jumps were simplified and some code may now
7567 be unreachable, so do jump optimization again. */
7568 static unsigned int
7569 rest_of_handle_cse (void)
7571 int tem;
7573 if (dump_file)
7574 dump_flow_info (dump_file, dump_flags);
7576 tem = cse_main (get_insns (), max_reg_num ());
7578 /* If we are not running more CSE passes, then we are no longer
7579 expecting CSE to be run. But always rerun it in a cheap mode. */
7580 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7582 if (tem == 2)
7584 timevar_push (TV_JUMP);
7585 rebuild_jump_labels (get_insns ());
7586 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7587 timevar_pop (TV_JUMP);
7589 else if (tem == 1 || optimize > 1)
7590 cse_cfg_altered |= cleanup_cfg (0);
7592 return 0;
7595 namespace {
7597 const pass_data pass_data_cse =
7599 RTL_PASS, /* type */
7600 "cse1", /* name */
7601 OPTGROUP_NONE, /* optinfo_flags */
7602 TV_CSE, /* tv_id */
7603 0, /* properties_required */
7604 0, /* properties_provided */
7605 0, /* properties_destroyed */
7606 0, /* todo_flags_start */
7607 TODO_df_finish, /* todo_flags_finish */
7610 class pass_cse : public rtl_opt_pass
7612 public:
7613 pass_cse (gcc::context *ctxt)
7614 : rtl_opt_pass (pass_data_cse, ctxt)
7617 /* opt_pass methods: */
7618 virtual bool gate (function *) { return optimize > 0; }
7619 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7621 }; // class pass_cse
7623 } // anon namespace
7625 rtl_opt_pass *
7626 make_pass_cse (gcc::context *ctxt)
7628 return new pass_cse (ctxt);
7632 /* Run second CSE pass after loop optimizations. */
7633 static unsigned int
7634 rest_of_handle_cse2 (void)
7636 int tem;
7638 if (dump_file)
7639 dump_flow_info (dump_file, dump_flags);
7641 tem = cse_main (get_insns (), max_reg_num ());
7643 /* Run a pass to eliminate duplicated assignments to condition code
7644 registers. We have to run this after bypass_jumps, because it
7645 makes it harder for that pass to determine whether a jump can be
7646 bypassed safely. */
7647 cse_condition_code_reg ();
7649 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7651 if (tem == 2)
7653 timevar_push (TV_JUMP);
7654 rebuild_jump_labels (get_insns ());
7655 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7656 timevar_pop (TV_JUMP);
7658 else if (tem == 1)
7659 cse_cfg_altered |= cleanup_cfg (0);
7661 cse_not_expected = 1;
7662 return 0;
7666 namespace {
7668 const pass_data pass_data_cse2 =
7670 RTL_PASS, /* type */
7671 "cse2", /* name */
7672 OPTGROUP_NONE, /* optinfo_flags */
7673 TV_CSE2, /* tv_id */
7674 0, /* properties_required */
7675 0, /* properties_provided */
7676 0, /* properties_destroyed */
7677 0, /* todo_flags_start */
7678 TODO_df_finish, /* todo_flags_finish */
7681 class pass_cse2 : public rtl_opt_pass
7683 public:
7684 pass_cse2 (gcc::context *ctxt)
7685 : rtl_opt_pass (pass_data_cse2, ctxt)
7688 /* opt_pass methods: */
7689 virtual bool gate (function *)
7691 return optimize > 0 && flag_rerun_cse_after_loop;
7694 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7696 }; // class pass_cse2
7698 } // anon namespace
7700 rtl_opt_pass *
7701 make_pass_cse2 (gcc::context *ctxt)
7703 return new pass_cse2 (ctxt);
7706 /* Run second CSE pass after loop optimizations. */
7707 static unsigned int
7708 rest_of_handle_cse_after_global_opts (void)
7710 int save_cfj;
7711 int tem;
7713 /* We only want to do local CSE, so don't follow jumps. */
7714 save_cfj = flag_cse_follow_jumps;
7715 flag_cse_follow_jumps = 0;
7717 rebuild_jump_labels (get_insns ());
7718 tem = cse_main (get_insns (), max_reg_num ());
7719 cse_cfg_altered |= purge_all_dead_edges ();
7720 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7722 cse_not_expected = !flag_rerun_cse_after_loop;
7724 /* If cse altered any jumps, rerun jump opts to clean things up. */
7725 if (tem == 2)
7727 timevar_push (TV_JUMP);
7728 rebuild_jump_labels (get_insns ());
7729 cse_cfg_altered |= cleanup_cfg (CLEANUP_CFG_CHANGED);
7730 timevar_pop (TV_JUMP);
7732 else if (tem == 1)
7733 cse_cfg_altered |= cleanup_cfg (0);
7735 flag_cse_follow_jumps = save_cfj;
7736 return 0;
7739 namespace {
7741 const pass_data pass_data_cse_after_global_opts =
7743 RTL_PASS, /* type */
7744 "cse_local", /* name */
7745 OPTGROUP_NONE, /* optinfo_flags */
7746 TV_CSE, /* tv_id */
7747 0, /* properties_required */
7748 0, /* properties_provided */
7749 0, /* properties_destroyed */
7750 0, /* todo_flags_start */
7751 TODO_df_finish, /* todo_flags_finish */
7754 class pass_cse_after_global_opts : public rtl_opt_pass
7756 public:
7757 pass_cse_after_global_opts (gcc::context *ctxt)
7758 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7761 /* opt_pass methods: */
7762 virtual bool gate (function *)
7764 return optimize > 0 && flag_rerun_cse_after_global_opts;
7767 virtual unsigned int execute (function *)
7769 return rest_of_handle_cse_after_global_opts ();
7772 }; // class pass_cse_after_global_opts
7774 } // anon namespace
7776 rtl_opt_pass *
7777 make_pass_cse_after_global_opts (gcc::context *ctxt)
7779 return new pass_cse_after_global_opts (ctxt);