Merge trunk version 188721 into gupc branch.
[official-gcc.git] / gcc / cse.c
blob9d5e32eee39d00733e4a08e08f7ebdd0804ef009
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
4 2011 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "regs.h"
30 #include "basic-block.h"
31 #include "flags.h"
32 #include "insn-config.h"
33 #include "recog.h"
34 #include "function.h"
35 #include "expr.h"
36 #include "diagnostic-core.h"
37 #include "toplev.h"
38 #include "ggc.h"
39 #include "timevar.h"
40 #include "except.h"
41 #include "target.h"
42 #include "params.h"
43 #include "rtlhooks-def.h"
44 #include "tree-pass.h"
45 #include "df.h"
46 #include "dbgcnt.h"
48 /* The basic idea of common subexpression elimination is to go
49 through the code, keeping a record of expressions that would
50 have the same value at the current scan point, and replacing
51 expressions encountered with the cheapest equivalent expression.
53 It is too complicated to keep track of the different possibilities
54 when control paths merge in this code; so, at each label, we forget all
55 that is known and start fresh. This can be described as processing each
56 extended basic block separately. We have a separate pass to perform
57 global CSE.
59 Note CSE can turn a conditional or computed jump into a nop or
60 an unconditional jump. When this occurs we arrange to run the jump
61 optimizer after CSE to delete the unreachable code.
63 We use two data structures to record the equivalent expressions:
64 a hash table for most expressions, and a vector of "quantity
65 numbers" to record equivalent (pseudo) registers.
67 The use of the special data structure for registers is desirable
68 because it is faster. It is possible because registers references
69 contain a fairly small number, the register number, taken from
70 a contiguously allocated series, and two register references are
71 identical if they have the same number. General expressions
72 do not have any such thing, so the only way to retrieve the
73 information recorded on an expression other than a register
74 is to keep it in a hash table.
76 Registers and "quantity numbers":
78 At the start of each basic block, all of the (hardware and pseudo)
79 registers used in the function are given distinct quantity
80 numbers to indicate their contents. During scan, when the code
81 copies one register into another, we copy the quantity number.
82 When a register is loaded in any other way, we allocate a new
83 quantity number to describe the value generated by this operation.
84 `REG_QTY (N)' records what quantity register N is currently thought
85 of as containing.
87 All real quantity numbers are greater than or equal to zero.
88 If register N has not been assigned a quantity, `REG_QTY (N)' will
89 equal -N - 1, which is always negative.
91 Quantity numbers below zero do not exist and none of the `qty_table'
92 entries should be referenced with a negative index.
94 We also maintain a bidirectional chain of registers for each
95 quantity number. The `qty_table` members `first_reg' and `last_reg',
96 and `reg_eqv_table' members `next' and `prev' hold these chains.
98 The first register in a chain is the one whose lifespan is least local.
99 Among equals, it is the one that was seen first.
100 We replace any equivalent register with that one.
102 If two registers have the same quantity number, it must be true that
103 REG expressions with qty_table `mode' must be in the hash table for both
104 registers and must be in the same class.
106 The converse is not true. Since hard registers may be referenced in
107 any mode, two REG expressions might be equivalent in the hash table
108 but not have the same quantity number if the quantity number of one
109 of the registers is not the same mode as those expressions.
111 Constants and quantity numbers
113 When a quantity has a known constant value, that value is stored
114 in the appropriate qty_table `const_rtx'. This is in addition to
115 putting the constant in the hash table as is usual for non-regs.
117 Whether a reg or a constant is preferred is determined by the configuration
118 macro CONST_COSTS and will often depend on the constant value. In any
119 event, expressions containing constants can be simplified, by fold_rtx.
121 When a quantity has a known nearly constant value (such as an address
122 of a stack slot), that value is stored in the appropriate qty_table
123 `const_rtx'.
125 Integer constants don't have a machine mode. However, cse
126 determines the intended machine mode from the destination
127 of the instruction that moves the constant. The machine mode
128 is recorded in the hash table along with the actual RTL
129 constant expression so that different modes are kept separate.
131 Other expressions:
133 To record known equivalences among expressions in general
134 we use a hash table called `table'. It has a fixed number of buckets
135 that contain chains of `struct table_elt' elements for expressions.
136 These chains connect the elements whose expressions have the same
137 hash codes.
139 Other chains through the same elements connect the elements which
140 currently have equivalent values.
142 Register references in an expression are canonicalized before hashing
143 the expression. This is done using `reg_qty' and qty_table `first_reg'.
144 The hash code of a register reference is computed using the quantity
145 number, not the register number.
147 When the value of an expression changes, it is necessary to remove from the
148 hash table not just that expression but all expressions whose values
149 could be different as a result.
151 1. If the value changing is in memory, except in special cases
152 ANYTHING referring to memory could be changed. That is because
153 nobody knows where a pointer does not point.
154 The function `invalidate_memory' removes what is necessary.
156 The special cases are when the address is constant or is
157 a constant plus a fixed register such as the frame pointer
158 or a static chain pointer. When such addresses are stored in,
159 we can tell exactly which other such addresses must be invalidated
160 due to overlap. `invalidate' does this.
161 All expressions that refer to non-constant
162 memory addresses are also invalidated. `invalidate_memory' does this.
164 2. If the value changing is a register, all expressions
165 containing references to that register, and only those,
166 must be removed.
168 Because searching the entire hash table for expressions that contain
169 a register is very slow, we try to figure out when it isn't necessary.
170 Precisely, this is necessary only when expressions have been
171 entered in the hash table using this register, and then the value has
172 changed, and then another expression wants to be added to refer to
173 the register's new value. This sequence of circumstances is rare
174 within any one basic block.
176 `REG_TICK' and `REG_IN_TABLE', accessors for members of
177 cse_reg_info, are used to detect this case. REG_TICK (i) is
178 incremented whenever a value is stored in register i.
179 REG_IN_TABLE (i) holds -1 if no references to register i have been
180 entered in the table; otherwise, it contains the value REG_TICK (i)
181 had when the references were entered. If we want to enter a
182 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
183 remove old references. Until we want to enter a new entry, the
184 mere fact that the two vectors don't match makes the entries be
185 ignored if anyone tries to match them.
187 Registers themselves are entered in the hash table as well as in
188 the equivalent-register chains. However, `REG_TICK' and
189 `REG_IN_TABLE' do not apply to expressions which are simple
190 register references. These expressions are removed from the table
191 immediately when they become invalid, and this can be done even if
192 we do not immediately search for all the expressions that refer to
193 the register.
195 A CLOBBER rtx in an instruction invalidates its operand for further
196 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
197 invalidates everything that resides in memory.
199 Related expressions:
201 Constant expressions that differ only by an additive integer
202 are called related. When a constant expression is put in
203 the table, the related expression with no constant term
204 is also entered. These are made to point at each other
205 so that it is possible to find out if there exists any
206 register equivalent to an expression related to a given expression. */
208 /* Length of qty_table vector. We know in advance we will not need
209 a quantity number this big. */
211 static int max_qty;
213 /* Next quantity number to be allocated.
214 This is 1 + the largest number needed so far. */
216 static int next_qty;
218 /* Per-qty information tracking.
220 `first_reg' and `last_reg' track the head and tail of the
221 chain of registers which currently contain this quantity.
223 `mode' contains the machine mode of this quantity.
225 `const_rtx' holds the rtx of the constant value of this
226 quantity, if known. A summations of the frame/arg pointer
227 and a constant can also be entered here. When this holds
228 a known value, `const_insn' is the insn which stored the
229 constant value.
231 `comparison_{code,const,qty}' are used to track when a
232 comparison between a quantity and some constant or register has
233 been passed. In such a case, we know the results of the comparison
234 in case we see it again. These members record a comparison that
235 is known to be true. `comparison_code' holds the rtx code of such
236 a comparison, else it is set to UNKNOWN and the other two
237 comparison members are undefined. `comparison_const' holds
238 the constant being compared against, or zero if the comparison
239 is not against a constant. `comparison_qty' holds the quantity
240 being compared against when the result is known. If the comparison
241 is not with a register, `comparison_qty' is -1. */
243 struct qty_table_elem
245 rtx const_rtx;
246 rtx const_insn;
247 rtx comparison_const;
248 int comparison_qty;
249 unsigned int first_reg, last_reg;
250 /* The sizes of these fields should match the sizes of the
251 code and mode fields of struct rtx_def (see rtl.h). */
252 ENUM_BITFIELD(rtx_code) comparison_code : 16;
253 ENUM_BITFIELD(machine_mode) mode : 8;
256 /* The table of all qtys, indexed by qty number. */
257 static struct qty_table_elem *qty_table;
259 /* Structure used to pass arguments via for_each_rtx to function
260 cse_change_cc_mode. */
261 struct change_cc_mode_args
263 rtx insn;
264 rtx newreg;
267 #ifdef HAVE_cc0
268 /* For machines that have a CC0, we do not record its value in the hash
269 table since its use is guaranteed to be the insn immediately following
270 its definition and any other insn is presumed to invalidate it.
272 Instead, we store below the current and last value assigned to CC0.
273 If it should happen to be a constant, it is stored in preference
274 to the actual assigned value. In case it is a constant, we store
275 the mode in which the constant should be interpreted. */
277 static rtx this_insn_cc0, prev_insn_cc0;
278 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
279 #endif
281 /* Insn being scanned. */
283 static rtx this_insn;
284 static bool optimize_this_for_speed_p;
286 /* Index by register number, gives the number of the next (or
287 previous) register in the chain of registers sharing the same
288 value.
290 Or -1 if this register is at the end of the chain.
292 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
294 /* Per-register equivalence chain. */
295 struct reg_eqv_elem
297 int next, prev;
300 /* The table of all register equivalence chains. */
301 static struct reg_eqv_elem *reg_eqv_table;
303 struct cse_reg_info
305 /* The timestamp at which this register is initialized. */
306 unsigned int timestamp;
308 /* The quantity number of the register's current contents. */
309 int reg_qty;
311 /* The number of times the register has been altered in the current
312 basic block. */
313 int reg_tick;
315 /* The REG_TICK value at which rtx's containing this register are
316 valid in the hash table. If this does not equal the current
317 reg_tick value, such expressions existing in the hash table are
318 invalid. */
319 int reg_in_table;
321 /* The SUBREG that was set when REG_TICK was last incremented. Set
322 to -1 if the last store was to the whole register, not a subreg. */
323 unsigned int subreg_ticked;
326 /* A table of cse_reg_info indexed by register numbers. */
327 static struct cse_reg_info *cse_reg_info_table;
329 /* The size of the above table. */
330 static unsigned int cse_reg_info_table_size;
332 /* The index of the first entry that has not been initialized. */
333 static unsigned int cse_reg_info_table_first_uninitialized;
335 /* The timestamp at the beginning of the current run of
336 cse_extended_basic_block. We increment this variable at the beginning of
337 the current run of cse_extended_basic_block. The timestamp field of a
338 cse_reg_info entry matches the value of this variable if and only
339 if the entry has been initialized during the current run of
340 cse_extended_basic_block. */
341 static unsigned int cse_reg_info_timestamp;
343 /* A HARD_REG_SET containing all the hard registers for which there is
344 currently a REG expression in the hash table. Note the difference
345 from the above variables, which indicate if the REG is mentioned in some
346 expression in the table. */
348 static HARD_REG_SET hard_regs_in_table;
350 /* True if CSE has altered the CFG. */
351 static bool cse_cfg_altered;
353 /* True if CSE has altered conditional jump insns in such a way
354 that jump optimization should be redone. */
355 static bool cse_jumps_altered;
357 /* True if we put a LABEL_REF into the hash table for an INSN
358 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
359 to put in the note. */
360 static bool recorded_label_ref;
362 /* canon_hash stores 1 in do_not_record
363 if it notices a reference to CC0, PC, or some other volatile
364 subexpression. */
366 static int do_not_record;
368 /* canon_hash stores 1 in hash_arg_in_memory
369 if it notices a reference to memory within the expression being hashed. */
371 static int hash_arg_in_memory;
373 /* The hash table contains buckets which are chains of `struct table_elt's,
374 each recording one expression's information.
375 That expression is in the `exp' field.
377 The canon_exp field contains a canonical (from the point of view of
378 alias analysis) version of the `exp' field.
380 Those elements with the same hash code are chained in both directions
381 through the `next_same_hash' and `prev_same_hash' fields.
383 Each set of expressions with equivalent values
384 are on a two-way chain through the `next_same_value'
385 and `prev_same_value' fields, and all point with
386 the `first_same_value' field at the first element in
387 that chain. The chain is in order of increasing cost.
388 Each element's cost value is in its `cost' field.
390 The `in_memory' field is nonzero for elements that
391 involve any reference to memory. These elements are removed
392 whenever a write is done to an unidentified location in memory.
393 To be safe, we assume that a memory address is unidentified unless
394 the address is either a symbol constant or a constant plus
395 the frame pointer or argument pointer.
397 The `related_value' field is used to connect related expressions
398 (that differ by adding an integer).
399 The related expressions are chained in a circular fashion.
400 `related_value' is zero for expressions for which this
401 chain is not useful.
403 The `cost' field stores the cost of this element's expression.
404 The `regcost' field stores the value returned by approx_reg_cost for
405 this element's expression.
407 The `is_const' flag is set if the element is a constant (including
408 a fixed address).
410 The `flag' field is used as a temporary during some search routines.
412 The `mode' field is usually the same as GET_MODE (`exp'), but
413 if `exp' is a CONST_INT and has no machine mode then the `mode'
414 field is the mode it was being used as. Each constant is
415 recorded separately for each mode it is used with. */
417 struct table_elt
419 rtx exp;
420 rtx canon_exp;
421 struct table_elt *next_same_hash;
422 struct table_elt *prev_same_hash;
423 struct table_elt *next_same_value;
424 struct table_elt *prev_same_value;
425 struct table_elt *first_same_value;
426 struct table_elt *related_value;
427 int cost;
428 int regcost;
429 /* The size of this field should match the size
430 of the mode field of struct rtx_def (see rtl.h). */
431 ENUM_BITFIELD(machine_mode) mode : 8;
432 char in_memory;
433 char is_const;
434 char flag;
437 /* We don't want a lot of buckets, because we rarely have very many
438 things stored in the hash table, and a lot of buckets slows
439 down a lot of loops that happen frequently. */
440 #define HASH_SHIFT 5
441 #define HASH_SIZE (1 << HASH_SHIFT)
442 #define HASH_MASK (HASH_SIZE - 1)
444 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
445 register (hard registers may require `do_not_record' to be set). */
447 #define HASH(X, M) \
448 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
449 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
450 : canon_hash (X, M)) & HASH_MASK)
452 /* Like HASH, but without side-effects. */
453 #define SAFE_HASH(X, M) \
454 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
455 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
456 : safe_hash (X, M)) & HASH_MASK)
458 /* Determine whether register number N is considered a fixed register for the
459 purpose of approximating register costs.
460 It is desirable to replace other regs with fixed regs, to reduce need for
461 non-fixed hard regs.
462 A reg wins if it is either the frame pointer or designated as fixed. */
463 #define FIXED_REGNO_P(N) \
464 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
465 || fixed_regs[N] || global_regs[N])
467 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
468 hard registers and pointers into the frame are the cheapest with a cost
469 of 0. Next come pseudos with a cost of one and other hard registers with
470 a cost of 2. Aside from these special cases, call `rtx_cost'. */
472 #define CHEAP_REGNO(N) \
473 (REGNO_PTR_FRAME_P(N) \
474 || (HARD_REGISTER_NUM_P (N) \
475 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
477 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET, 1))
478 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
480 /* Get the number of times this register has been updated in this
481 basic block. */
483 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
485 /* Get the point at which REG was recorded in the table. */
487 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
489 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
490 SUBREG). */
492 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
494 /* Get the quantity number for REG. */
496 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
498 /* Determine if the quantity number for register X represents a valid index
499 into the qty_table. */
501 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
503 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
505 #define CHEAPER(X, Y) \
506 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
508 static struct table_elt *table[HASH_SIZE];
510 /* Chain of `struct table_elt's made so far for this function
511 but currently removed from the table. */
513 static struct table_elt *free_element_chain;
515 /* Set to the cost of a constant pool reference if one was found for a
516 symbolic constant. If this was found, it means we should try to
517 convert constants into constant pool entries if they don't fit in
518 the insn. */
520 static int constant_pool_entries_cost;
521 static int constant_pool_entries_regcost;
523 /* Trace a patch through the CFG. */
525 struct branch_path
527 /* The basic block for this path entry. */
528 basic_block bb;
531 /* This data describes a block that will be processed by
532 cse_extended_basic_block. */
534 struct cse_basic_block_data
536 /* Total number of SETs in block. */
537 int nsets;
538 /* Size of current branch path, if any. */
539 int path_size;
540 /* Current path, indicating which basic_blocks will be processed. */
541 struct branch_path *path;
545 /* Pointers to the live in/live out bitmaps for the boundaries of the
546 current EBB. */
547 static bitmap cse_ebb_live_in, cse_ebb_live_out;
549 /* A simple bitmap to track which basic blocks have been visited
550 already as part of an already processed extended basic block. */
551 static sbitmap cse_visited_basic_blocks;
553 static bool fixed_base_plus_p (rtx x);
554 static int notreg_cost (rtx, enum rtx_code, int);
555 static int approx_reg_cost_1 (rtx *, void *);
556 static int approx_reg_cost (rtx);
557 static int preferable (int, int, int, int);
558 static void new_basic_block (void);
559 static void make_new_qty (unsigned int, enum machine_mode);
560 static void make_regs_eqv (unsigned int, unsigned int);
561 static void delete_reg_equiv (unsigned int);
562 static int mention_regs (rtx);
563 static int insert_regs (rtx, struct table_elt *, int);
564 static void remove_from_table (struct table_elt *, unsigned);
565 static void remove_pseudo_from_table (rtx, unsigned);
566 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
567 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
568 static rtx lookup_as_function (rtx, enum rtx_code);
569 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
570 enum machine_mode, int, int);
571 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
572 enum machine_mode);
573 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
574 static void invalidate (rtx, enum machine_mode);
575 static void remove_invalid_refs (unsigned int);
576 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
577 enum machine_mode);
578 static void rehash_using_reg (rtx);
579 static void invalidate_memory (void);
580 static void invalidate_for_call (void);
581 static rtx use_related_value (rtx, struct table_elt *);
583 static inline unsigned canon_hash (rtx, enum machine_mode);
584 static inline unsigned safe_hash (rtx, enum machine_mode);
585 static inline unsigned hash_rtx_string (const char *);
587 static rtx canon_reg (rtx, rtx);
588 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
589 enum machine_mode *,
590 enum machine_mode *);
591 static rtx fold_rtx (rtx, rtx);
592 static rtx equiv_constant (rtx);
593 static void record_jump_equiv (rtx, bool);
594 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
595 int);
596 static void cse_insn (rtx);
597 static void cse_prescan_path (struct cse_basic_block_data *);
598 static void invalidate_from_clobbers (rtx);
599 static void invalidate_from_sets_and_clobbers (rtx);
600 static rtx cse_process_notes (rtx, rtx, bool *);
601 static void cse_extended_basic_block (struct cse_basic_block_data *);
602 static int check_for_label_ref (rtx *, void *);
603 extern void dump_class (struct table_elt*);
604 static void get_cse_reg_info_1 (unsigned int regno);
605 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
606 static int check_dependence (rtx *, void *);
608 static void flush_hash_table (void);
609 static bool insn_live_p (rtx, int *);
610 static bool set_live_p (rtx, rtx, int *);
611 static int cse_change_cc_mode (rtx *, void *);
612 static void cse_change_cc_mode_insn (rtx, rtx);
613 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
614 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
615 bool);
618 #undef RTL_HOOKS_GEN_LOWPART
619 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
621 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
623 /* Nonzero if X has the form (PLUS frame-pointer integer). */
625 static bool
626 fixed_base_plus_p (rtx x)
628 switch (GET_CODE (x))
630 case REG:
631 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
632 return true;
633 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
634 return true;
635 return false;
637 case PLUS:
638 if (!CONST_INT_P (XEXP (x, 1)))
639 return false;
640 return fixed_base_plus_p (XEXP (x, 0));
642 default:
643 return false;
647 /* Dump the expressions in the equivalence class indicated by CLASSP.
648 This function is used only for debugging. */
649 DEBUG_FUNCTION void
650 dump_class (struct table_elt *classp)
652 struct table_elt *elt;
654 fprintf (stderr, "Equivalence chain for ");
655 print_rtl (stderr, classp->exp);
656 fprintf (stderr, ": \n");
658 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
660 print_rtl (stderr, elt->exp);
661 fprintf (stderr, "\n");
665 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
667 static int
668 approx_reg_cost_1 (rtx *xp, void *data)
670 rtx x = *xp;
671 int *cost_p = (int *) data;
673 if (x && REG_P (x))
675 unsigned int regno = REGNO (x);
677 if (! CHEAP_REGNO (regno))
679 if (regno < FIRST_PSEUDO_REGISTER)
681 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
682 return 1;
683 *cost_p += 2;
685 else
686 *cost_p += 1;
690 return 0;
693 /* Return an estimate of the cost of the registers used in an rtx.
694 This is mostly the number of different REG expressions in the rtx;
695 however for some exceptions like fixed registers we use a cost of
696 0. If any other hard register reference occurs, return MAX_COST. */
698 static int
699 approx_reg_cost (rtx x)
701 int cost = 0;
703 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
704 return MAX_COST;
706 return cost;
709 /* Return a negative value if an rtx A, whose costs are given by COST_A
710 and REGCOST_A, is more desirable than an rtx B.
711 Return a positive value if A is less desirable, or 0 if the two are
712 equally good. */
713 static int
714 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
716 /* First, get rid of cases involving expressions that are entirely
717 unwanted. */
718 if (cost_a != cost_b)
720 if (cost_a == MAX_COST)
721 return 1;
722 if (cost_b == MAX_COST)
723 return -1;
726 /* Avoid extending lifetimes of hardregs. */
727 if (regcost_a != regcost_b)
729 if (regcost_a == MAX_COST)
730 return 1;
731 if (regcost_b == MAX_COST)
732 return -1;
735 /* Normal operation costs take precedence. */
736 if (cost_a != cost_b)
737 return cost_a - cost_b;
738 /* Only if these are identical consider effects on register pressure. */
739 if (regcost_a != regcost_b)
740 return regcost_a - regcost_b;
741 return 0;
744 /* Internal function, to compute cost when X is not a register; called
745 from COST macro to keep it simple. */
747 static int
748 notreg_cost (rtx x, enum rtx_code outer, int opno)
750 return ((GET_CODE (x) == SUBREG
751 && REG_P (SUBREG_REG (x))
752 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
753 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
754 && (GET_MODE_SIZE (GET_MODE (x))
755 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
756 && subreg_lowpart_p (x)
757 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
758 GET_MODE (SUBREG_REG (x))))
760 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
764 /* Initialize CSE_REG_INFO_TABLE. */
766 static void
767 init_cse_reg_info (unsigned int nregs)
769 /* Do we need to grow the table? */
770 if (nregs > cse_reg_info_table_size)
772 unsigned int new_size;
774 if (cse_reg_info_table_size < 2048)
776 /* Compute a new size that is a power of 2 and no smaller
777 than the large of NREGS and 64. */
778 new_size = (cse_reg_info_table_size
779 ? cse_reg_info_table_size : 64);
781 while (new_size < nregs)
782 new_size *= 2;
784 else
786 /* If we need a big table, allocate just enough to hold
787 NREGS registers. */
788 new_size = nregs;
791 /* Reallocate the table with NEW_SIZE entries. */
792 free (cse_reg_info_table);
793 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
794 cse_reg_info_table_size = new_size;
795 cse_reg_info_table_first_uninitialized = 0;
798 /* Do we have all of the first NREGS entries initialized? */
799 if (cse_reg_info_table_first_uninitialized < nregs)
801 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
802 unsigned int i;
804 /* Put the old timestamp on newly allocated entries so that they
805 will all be considered out of date. We do not touch those
806 entries beyond the first NREGS entries to be nice to the
807 virtual memory. */
808 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
809 cse_reg_info_table[i].timestamp = old_timestamp;
811 cse_reg_info_table_first_uninitialized = nregs;
815 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
817 static void
818 get_cse_reg_info_1 (unsigned int regno)
820 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
821 entry will be considered to have been initialized. */
822 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
824 /* Initialize the rest of the entry. */
825 cse_reg_info_table[regno].reg_tick = 1;
826 cse_reg_info_table[regno].reg_in_table = -1;
827 cse_reg_info_table[regno].subreg_ticked = -1;
828 cse_reg_info_table[regno].reg_qty = -regno - 1;
831 /* Find a cse_reg_info entry for REGNO. */
833 static inline struct cse_reg_info *
834 get_cse_reg_info (unsigned int regno)
836 struct cse_reg_info *p = &cse_reg_info_table[regno];
838 /* If this entry has not been initialized, go ahead and initialize
839 it. */
840 if (p->timestamp != cse_reg_info_timestamp)
841 get_cse_reg_info_1 (regno);
843 return p;
846 /* Clear the hash table and initialize each register with its own quantity,
847 for a new basic block. */
849 static void
850 new_basic_block (void)
852 int i;
854 next_qty = 0;
856 /* Invalidate cse_reg_info_table. */
857 cse_reg_info_timestamp++;
859 /* Clear out hash table state for this pass. */
860 CLEAR_HARD_REG_SET (hard_regs_in_table);
862 /* The per-quantity values used to be initialized here, but it is
863 much faster to initialize each as it is made in `make_new_qty'. */
865 for (i = 0; i < HASH_SIZE; i++)
867 struct table_elt *first;
869 first = table[i];
870 if (first != NULL)
872 struct table_elt *last = first;
874 table[i] = NULL;
876 while (last->next_same_hash != NULL)
877 last = last->next_same_hash;
879 /* Now relink this hash entire chain into
880 the free element list. */
882 last->next_same_hash = free_element_chain;
883 free_element_chain = first;
887 #ifdef HAVE_cc0
888 prev_insn_cc0 = 0;
889 #endif
892 /* Say that register REG contains a quantity in mode MODE not in any
893 register before and initialize that quantity. */
895 static void
896 make_new_qty (unsigned int reg, enum machine_mode mode)
898 int q;
899 struct qty_table_elem *ent;
900 struct reg_eqv_elem *eqv;
902 gcc_assert (next_qty < max_qty);
904 q = REG_QTY (reg) = next_qty++;
905 ent = &qty_table[q];
906 ent->first_reg = reg;
907 ent->last_reg = reg;
908 ent->mode = mode;
909 ent->const_rtx = ent->const_insn = NULL_RTX;
910 ent->comparison_code = UNKNOWN;
912 eqv = &reg_eqv_table[reg];
913 eqv->next = eqv->prev = -1;
916 /* Make reg NEW equivalent to reg OLD.
917 OLD is not changing; NEW is. */
919 static void
920 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
922 unsigned int lastr, firstr;
923 int q = REG_QTY (old_reg);
924 struct qty_table_elem *ent;
926 ent = &qty_table[q];
928 /* Nothing should become eqv until it has a "non-invalid" qty number. */
929 gcc_assert (REGNO_QTY_VALID_P (old_reg));
931 REG_QTY (new_reg) = q;
932 firstr = ent->first_reg;
933 lastr = ent->last_reg;
935 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
936 hard regs. Among pseudos, if NEW will live longer than any other reg
937 of the same qty, and that is beyond the current basic block,
938 make it the new canonical replacement for this qty. */
939 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
940 /* Certain fixed registers might be of the class NO_REGS. This means
941 that not only can they not be allocated by the compiler, but
942 they cannot be used in substitutions or canonicalizations
943 either. */
944 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
945 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
946 || (new_reg >= FIRST_PSEUDO_REGISTER
947 && (firstr < FIRST_PSEUDO_REGISTER
948 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
949 && !bitmap_bit_p (cse_ebb_live_out, firstr))
950 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
951 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
953 reg_eqv_table[firstr].prev = new_reg;
954 reg_eqv_table[new_reg].next = firstr;
955 reg_eqv_table[new_reg].prev = -1;
956 ent->first_reg = new_reg;
958 else
960 /* If NEW is a hard reg (known to be non-fixed), insert at end.
961 Otherwise, insert before any non-fixed hard regs that are at the
962 end. Registers of class NO_REGS cannot be used as an
963 equivalent for anything. */
964 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
965 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
966 && new_reg >= FIRST_PSEUDO_REGISTER)
967 lastr = reg_eqv_table[lastr].prev;
968 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
969 if (reg_eqv_table[lastr].next >= 0)
970 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
971 else
972 qty_table[q].last_reg = new_reg;
973 reg_eqv_table[lastr].next = new_reg;
974 reg_eqv_table[new_reg].prev = lastr;
978 /* Remove REG from its equivalence class. */
980 static void
981 delete_reg_equiv (unsigned int reg)
983 struct qty_table_elem *ent;
984 int q = REG_QTY (reg);
985 int p, n;
987 /* If invalid, do nothing. */
988 if (! REGNO_QTY_VALID_P (reg))
989 return;
991 ent = &qty_table[q];
993 p = reg_eqv_table[reg].prev;
994 n = reg_eqv_table[reg].next;
996 if (n != -1)
997 reg_eqv_table[n].prev = p;
998 else
999 ent->last_reg = p;
1000 if (p != -1)
1001 reg_eqv_table[p].next = n;
1002 else
1003 ent->first_reg = n;
1005 REG_QTY (reg) = -reg - 1;
1008 /* Remove any invalid expressions from the hash table
1009 that refer to any of the registers contained in expression X.
1011 Make sure that newly inserted references to those registers
1012 as subexpressions will be considered valid.
1014 mention_regs is not called when a register itself
1015 is being stored in the table.
1017 Return 1 if we have done something that may have changed the hash code
1018 of X. */
1020 static int
1021 mention_regs (rtx x)
1023 enum rtx_code code;
1024 int i, j;
1025 const char *fmt;
1026 int changed = 0;
1028 if (x == 0)
1029 return 0;
1031 code = GET_CODE (x);
1032 if (code == REG)
1034 unsigned int regno = REGNO (x);
1035 unsigned int endregno = END_REGNO (x);
1036 unsigned int i;
1038 for (i = regno; i < endregno; i++)
1040 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1041 remove_invalid_refs (i);
1043 REG_IN_TABLE (i) = REG_TICK (i);
1044 SUBREG_TICKED (i) = -1;
1047 return 0;
1050 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1051 pseudo if they don't use overlapping words. We handle only pseudos
1052 here for simplicity. */
1053 if (code == SUBREG && REG_P (SUBREG_REG (x))
1054 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1056 unsigned int i = REGNO (SUBREG_REG (x));
1058 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1060 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1061 the last store to this register really stored into this
1062 subreg, then remove the memory of this subreg.
1063 Otherwise, remove any memory of the entire register and
1064 all its subregs from the table. */
1065 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1066 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1067 remove_invalid_refs (i);
1068 else
1069 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1072 REG_IN_TABLE (i) = REG_TICK (i);
1073 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1074 return 0;
1077 /* If X is a comparison or a COMPARE and either operand is a register
1078 that does not have a quantity, give it one. This is so that a later
1079 call to record_jump_equiv won't cause X to be assigned a different
1080 hash code and not found in the table after that call.
1082 It is not necessary to do this here, since rehash_using_reg can
1083 fix up the table later, but doing this here eliminates the need to
1084 call that expensive function in the most common case where the only
1085 use of the register is in the comparison. */
1087 if (code == COMPARE || COMPARISON_P (x))
1089 if (REG_P (XEXP (x, 0))
1090 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1091 if (insert_regs (XEXP (x, 0), NULL, 0))
1093 rehash_using_reg (XEXP (x, 0));
1094 changed = 1;
1097 if (REG_P (XEXP (x, 1))
1098 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1099 if (insert_regs (XEXP (x, 1), NULL, 0))
1101 rehash_using_reg (XEXP (x, 1));
1102 changed = 1;
1106 fmt = GET_RTX_FORMAT (code);
1107 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1108 if (fmt[i] == 'e')
1109 changed |= mention_regs (XEXP (x, i));
1110 else if (fmt[i] == 'E')
1111 for (j = 0; j < XVECLEN (x, i); j++)
1112 changed |= mention_regs (XVECEXP (x, i, j));
1114 return changed;
1117 /* Update the register quantities for inserting X into the hash table
1118 with a value equivalent to CLASSP.
1119 (If the class does not contain a REG, it is irrelevant.)
1120 If MODIFIED is nonzero, X is a destination; it is being modified.
1121 Note that delete_reg_equiv should be called on a register
1122 before insert_regs is done on that register with MODIFIED != 0.
1124 Nonzero value means that elements of reg_qty have changed
1125 so X's hash code may be different. */
1127 static int
1128 insert_regs (rtx x, struct table_elt *classp, int modified)
1130 if (REG_P (x))
1132 unsigned int regno = REGNO (x);
1133 int qty_valid;
1135 /* If REGNO is in the equivalence table already but is of the
1136 wrong mode for that equivalence, don't do anything here. */
1138 qty_valid = REGNO_QTY_VALID_P (regno);
1139 if (qty_valid)
1141 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1143 if (ent->mode != GET_MODE (x))
1144 return 0;
1147 if (modified || ! qty_valid)
1149 if (classp)
1150 for (classp = classp->first_same_value;
1151 classp != 0;
1152 classp = classp->next_same_value)
1153 if (REG_P (classp->exp)
1154 && GET_MODE (classp->exp) == GET_MODE (x))
1156 unsigned c_regno = REGNO (classp->exp);
1158 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1160 /* Suppose that 5 is hard reg and 100 and 101 are
1161 pseudos. Consider
1163 (set (reg:si 100) (reg:si 5))
1164 (set (reg:si 5) (reg:si 100))
1165 (set (reg:di 101) (reg:di 5))
1167 We would now set REG_QTY (101) = REG_QTY (5), but the
1168 entry for 5 is in SImode. When we use this later in
1169 copy propagation, we get the register in wrong mode. */
1170 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1171 continue;
1173 make_regs_eqv (regno, c_regno);
1174 return 1;
1177 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1178 than REG_IN_TABLE to find out if there was only a single preceding
1179 invalidation - for the SUBREG - or another one, which would be
1180 for the full register. However, if we find here that REG_TICK
1181 indicates that the register is invalid, it means that it has
1182 been invalidated in a separate operation. The SUBREG might be used
1183 now (then this is a recursive call), or we might use the full REG
1184 now and a SUBREG of it later. So bump up REG_TICK so that
1185 mention_regs will do the right thing. */
1186 if (! modified
1187 && REG_IN_TABLE (regno) >= 0
1188 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1189 REG_TICK (regno)++;
1190 make_new_qty (regno, GET_MODE (x));
1191 return 1;
1194 return 0;
1197 /* If X is a SUBREG, we will likely be inserting the inner register in the
1198 table. If that register doesn't have an assigned quantity number at
1199 this point but does later, the insertion that we will be doing now will
1200 not be accessible because its hash code will have changed. So assign
1201 a quantity number now. */
1203 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1204 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1206 insert_regs (SUBREG_REG (x), NULL, 0);
1207 mention_regs (x);
1208 return 1;
1210 else
1211 return mention_regs (x);
1215 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1216 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1217 CST is equal to an anchor. */
1219 static bool
1220 compute_const_anchors (rtx cst,
1221 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1222 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1224 HOST_WIDE_INT n = INTVAL (cst);
1226 *lower_base = n & ~(targetm.const_anchor - 1);
1227 if (*lower_base == n)
1228 return false;
1230 *upper_base =
1231 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1232 *upper_offs = n - *upper_base;
1233 *lower_offs = n - *lower_base;
1234 return true;
1237 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1239 static void
1240 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1241 enum machine_mode mode)
1243 struct table_elt *elt;
1244 unsigned hash;
1245 rtx anchor_exp;
1246 rtx exp;
1248 anchor_exp = GEN_INT (anchor);
1249 hash = HASH (anchor_exp, mode);
1250 elt = lookup (anchor_exp, hash, mode);
1251 if (!elt)
1252 elt = insert (anchor_exp, NULL, hash, mode);
1254 exp = plus_constant (mode, reg, offs);
1255 /* REG has just been inserted and the hash codes recomputed. */
1256 mention_regs (exp);
1257 hash = HASH (exp, mode);
1259 /* Use the cost of the register rather than the whole expression. When
1260 looking up constant anchors we will further offset the corresponding
1261 expression therefore it does not make sense to prefer REGs over
1262 reg-immediate additions. Prefer instead the oldest expression. Also
1263 don't prefer pseudos over hard regs so that we derive constants in
1264 argument registers from other argument registers rather than from the
1265 original pseudo that was used to synthesize the constant. */
1266 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1269 /* The constant CST is equivalent to the register REG. Create
1270 equivalences between the two anchors of CST and the corresponding
1271 register-offset expressions using REG. */
1273 static void
1274 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1276 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1278 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1279 &upper_base, &upper_offs))
1280 return;
1282 /* Ignore anchors of value 0. Constants accessible from zero are
1283 simple. */
1284 if (lower_base != 0)
1285 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1287 if (upper_base != 0)
1288 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1291 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1292 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1293 valid expression. Return the cheapest and oldest of such expressions. In
1294 *OLD, return how old the resulting expression is compared to the other
1295 equivalent expressions. */
1297 static rtx
1298 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1299 unsigned *old)
1301 struct table_elt *elt;
1302 unsigned idx;
1303 struct table_elt *match_elt;
1304 rtx match;
1306 /* Find the cheapest and *oldest* expression to maximize the chance of
1307 reusing the same pseudo. */
1309 match_elt = NULL;
1310 match = NULL_RTX;
1311 for (elt = anchor_elt->first_same_value, idx = 0;
1312 elt;
1313 elt = elt->next_same_value, idx++)
1315 if (match_elt && CHEAPER (match_elt, elt))
1316 return match;
1318 if (REG_P (elt->exp)
1319 || (GET_CODE (elt->exp) == PLUS
1320 && REG_P (XEXP (elt->exp, 0))
1321 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1323 rtx x;
1325 /* Ignore expressions that are no longer valid. */
1326 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1327 continue;
1329 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1330 if (REG_P (x)
1331 || (GET_CODE (x) == PLUS
1332 && IN_RANGE (INTVAL (XEXP (x, 1)),
1333 -targetm.const_anchor,
1334 targetm.const_anchor - 1)))
1336 match = x;
1337 match_elt = elt;
1338 *old = idx;
1343 return match;
1346 /* Try to express the constant SRC_CONST using a register+offset expression
1347 derived from a constant anchor. Return it if successful or NULL_RTX,
1348 otherwise. */
1350 static rtx
1351 try_const_anchors (rtx src_const, enum machine_mode mode)
1353 struct table_elt *lower_elt, *upper_elt;
1354 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1355 rtx lower_anchor_rtx, upper_anchor_rtx;
1356 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1357 unsigned lower_old, upper_old;
1359 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1360 &upper_base, &upper_offs))
1361 return NULL_RTX;
1363 lower_anchor_rtx = GEN_INT (lower_base);
1364 upper_anchor_rtx = GEN_INT (upper_base);
1365 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1366 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1368 if (lower_elt)
1369 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1370 if (upper_elt)
1371 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1373 if (!lower_exp)
1374 return upper_exp;
1375 if (!upper_exp)
1376 return lower_exp;
1378 /* Return the older expression. */
1379 return (upper_old > lower_old ? upper_exp : lower_exp);
1382 /* Look in or update the hash table. */
1384 /* Remove table element ELT from use in the table.
1385 HASH is its hash code, made using the HASH macro.
1386 It's an argument because often that is known in advance
1387 and we save much time not recomputing it. */
1389 static void
1390 remove_from_table (struct table_elt *elt, unsigned int hash)
1392 if (elt == 0)
1393 return;
1395 /* Mark this element as removed. See cse_insn. */
1396 elt->first_same_value = 0;
1398 /* Remove the table element from its equivalence class. */
1401 struct table_elt *prev = elt->prev_same_value;
1402 struct table_elt *next = elt->next_same_value;
1404 if (next)
1405 next->prev_same_value = prev;
1407 if (prev)
1408 prev->next_same_value = next;
1409 else
1411 struct table_elt *newfirst = next;
1412 while (next)
1414 next->first_same_value = newfirst;
1415 next = next->next_same_value;
1420 /* Remove the table element from its hash bucket. */
1423 struct table_elt *prev = elt->prev_same_hash;
1424 struct table_elt *next = elt->next_same_hash;
1426 if (next)
1427 next->prev_same_hash = prev;
1429 if (prev)
1430 prev->next_same_hash = next;
1431 else if (table[hash] == elt)
1432 table[hash] = next;
1433 else
1435 /* This entry is not in the proper hash bucket. This can happen
1436 when two classes were merged by `merge_equiv_classes'. Search
1437 for the hash bucket that it heads. This happens only very
1438 rarely, so the cost is acceptable. */
1439 for (hash = 0; hash < HASH_SIZE; hash++)
1440 if (table[hash] == elt)
1441 table[hash] = next;
1445 /* Remove the table element from its related-value circular chain. */
1447 if (elt->related_value != 0 && elt->related_value != elt)
1449 struct table_elt *p = elt->related_value;
1451 while (p->related_value != elt)
1452 p = p->related_value;
1453 p->related_value = elt->related_value;
1454 if (p->related_value == p)
1455 p->related_value = 0;
1458 /* Now add it to the free element chain. */
1459 elt->next_same_hash = free_element_chain;
1460 free_element_chain = elt;
1463 /* Same as above, but X is a pseudo-register. */
1465 static void
1466 remove_pseudo_from_table (rtx x, unsigned int hash)
1468 struct table_elt *elt;
1470 /* Because a pseudo-register can be referenced in more than one
1471 mode, we might have to remove more than one table entry. */
1472 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1473 remove_from_table (elt, hash);
1476 /* Look up X in the hash table and return its table element,
1477 or 0 if X is not in the table.
1479 MODE is the machine-mode of X, or if X is an integer constant
1480 with VOIDmode then MODE is the mode with which X will be used.
1482 Here we are satisfied to find an expression whose tree structure
1483 looks like X. */
1485 static struct table_elt *
1486 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1488 struct table_elt *p;
1490 for (p = table[hash]; p; p = p->next_same_hash)
1491 if (mode == p->mode && ((x == p->exp && REG_P (x))
1492 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1493 return p;
1495 return 0;
1498 /* Like `lookup' but don't care whether the table element uses invalid regs.
1499 Also ignore discrepancies in the machine mode of a register. */
1501 static struct table_elt *
1502 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1504 struct table_elt *p;
1506 if (REG_P (x))
1508 unsigned int regno = REGNO (x);
1510 /* Don't check the machine mode when comparing registers;
1511 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1512 for (p = table[hash]; p; p = p->next_same_hash)
1513 if (REG_P (p->exp)
1514 && REGNO (p->exp) == regno)
1515 return p;
1517 else
1519 for (p = table[hash]; p; p = p->next_same_hash)
1520 if (mode == p->mode
1521 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1522 return p;
1525 return 0;
1528 /* Look for an expression equivalent to X and with code CODE.
1529 If one is found, return that expression. */
1531 static rtx
1532 lookup_as_function (rtx x, enum rtx_code code)
1534 struct table_elt *p
1535 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1537 if (p == 0)
1538 return 0;
1540 for (p = p->first_same_value; p; p = p->next_same_value)
1541 if (GET_CODE (p->exp) == code
1542 /* Make sure this is a valid entry in the table. */
1543 && exp_equiv_p (p->exp, p->exp, 1, false))
1544 return p->exp;
1546 return 0;
1549 /* Insert X in the hash table, assuming HASH is its hash code and
1550 CLASSP is an element of the class it should go in (or 0 if a new
1551 class should be made). COST is the code of X and reg_cost is the
1552 cost of registers in X. It is inserted at the proper position to
1553 keep the class in the order cheapest first.
1555 MODE is the machine-mode of X, or if X is an integer constant
1556 with VOIDmode then MODE is the mode with which X will be used.
1558 For elements of equal cheapness, the most recent one
1559 goes in front, except that the first element in the list
1560 remains first unless a cheaper element is added. The order of
1561 pseudo-registers does not matter, as canon_reg will be called to
1562 find the cheapest when a register is retrieved from the table.
1564 The in_memory field in the hash table element is set to 0.
1565 The caller must set it nonzero if appropriate.
1567 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1568 and if insert_regs returns a nonzero value
1569 you must then recompute its hash code before calling here.
1571 If necessary, update table showing constant values of quantities. */
1573 static struct table_elt *
1574 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1575 enum machine_mode mode, int cost, int reg_cost)
1577 struct table_elt *elt;
1579 /* If X is a register and we haven't made a quantity for it,
1580 something is wrong. */
1581 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1583 /* If X is a hard register, show it is being put in the table. */
1584 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1585 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1587 /* Put an element for X into the right hash bucket. */
1589 elt = free_element_chain;
1590 if (elt)
1591 free_element_chain = elt->next_same_hash;
1592 else
1593 elt = XNEW (struct table_elt);
1595 elt->exp = x;
1596 elt->canon_exp = NULL_RTX;
1597 elt->cost = cost;
1598 elt->regcost = reg_cost;
1599 elt->next_same_value = 0;
1600 elt->prev_same_value = 0;
1601 elt->next_same_hash = table[hash];
1602 elt->prev_same_hash = 0;
1603 elt->related_value = 0;
1604 elt->in_memory = 0;
1605 elt->mode = mode;
1606 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1608 if (table[hash])
1609 table[hash]->prev_same_hash = elt;
1610 table[hash] = elt;
1612 /* Put it into the proper value-class. */
1613 if (classp)
1615 classp = classp->first_same_value;
1616 if (CHEAPER (elt, classp))
1617 /* Insert at the head of the class. */
1619 struct table_elt *p;
1620 elt->next_same_value = classp;
1621 classp->prev_same_value = elt;
1622 elt->first_same_value = elt;
1624 for (p = classp; p; p = p->next_same_value)
1625 p->first_same_value = elt;
1627 else
1629 /* Insert not at head of the class. */
1630 /* Put it after the last element cheaper than X. */
1631 struct table_elt *p, *next;
1633 for (p = classp;
1634 (next = p->next_same_value) && CHEAPER (next, elt);
1635 p = next)
1638 /* Put it after P and before NEXT. */
1639 elt->next_same_value = next;
1640 if (next)
1641 next->prev_same_value = elt;
1643 elt->prev_same_value = p;
1644 p->next_same_value = elt;
1645 elt->first_same_value = classp;
1648 else
1649 elt->first_same_value = elt;
1651 /* If this is a constant being set equivalent to a register or a register
1652 being set equivalent to a constant, note the constant equivalence.
1654 If this is a constant, it cannot be equivalent to a different constant,
1655 and a constant is the only thing that can be cheaper than a register. So
1656 we know the register is the head of the class (before the constant was
1657 inserted).
1659 If this is a register that is not already known equivalent to a
1660 constant, we must check the entire class.
1662 If this is a register that is already known equivalent to an insn,
1663 update the qtys `const_insn' to show that `this_insn' is the latest
1664 insn making that quantity equivalent to the constant. */
1666 if (elt->is_const && classp && REG_P (classp->exp)
1667 && !REG_P (x))
1669 int exp_q = REG_QTY (REGNO (classp->exp));
1670 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1672 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1673 exp_ent->const_insn = this_insn;
1676 else if (REG_P (x)
1677 && classp
1678 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1679 && ! elt->is_const)
1681 struct table_elt *p;
1683 for (p = classp; p != 0; p = p->next_same_value)
1685 if (p->is_const && !REG_P (p->exp))
1687 int x_q = REG_QTY (REGNO (x));
1688 struct qty_table_elem *x_ent = &qty_table[x_q];
1690 x_ent->const_rtx
1691 = gen_lowpart (GET_MODE (x), p->exp);
1692 x_ent->const_insn = this_insn;
1693 break;
1698 else if (REG_P (x)
1699 && qty_table[REG_QTY (REGNO (x))].const_rtx
1700 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1701 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1703 /* If this is a constant with symbolic value,
1704 and it has a term with an explicit integer value,
1705 link it up with related expressions. */
1706 if (GET_CODE (x) == CONST)
1708 rtx subexp = get_related_value (x);
1709 unsigned subhash;
1710 struct table_elt *subelt, *subelt_prev;
1712 if (subexp != 0)
1714 /* Get the integer-free subexpression in the hash table. */
1715 subhash = SAFE_HASH (subexp, mode);
1716 subelt = lookup (subexp, subhash, mode);
1717 if (subelt == 0)
1718 subelt = insert (subexp, NULL, subhash, mode);
1719 /* Initialize SUBELT's circular chain if it has none. */
1720 if (subelt->related_value == 0)
1721 subelt->related_value = subelt;
1722 /* Find the element in the circular chain that precedes SUBELT. */
1723 subelt_prev = subelt;
1724 while (subelt_prev->related_value != subelt)
1725 subelt_prev = subelt_prev->related_value;
1726 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1727 This way the element that follows SUBELT is the oldest one. */
1728 elt->related_value = subelt_prev->related_value;
1729 subelt_prev->related_value = elt;
1733 return elt;
1736 /* Wrap insert_with_costs by passing the default costs. */
1738 static struct table_elt *
1739 insert (rtx x, struct table_elt *classp, unsigned int hash,
1740 enum machine_mode mode)
1742 return
1743 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1747 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1748 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1749 the two classes equivalent.
1751 CLASS1 will be the surviving class; CLASS2 should not be used after this
1752 call.
1754 Any invalid entries in CLASS2 will not be copied. */
1756 static void
1757 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1759 struct table_elt *elt, *next, *new_elt;
1761 /* Ensure we start with the head of the classes. */
1762 class1 = class1->first_same_value;
1763 class2 = class2->first_same_value;
1765 /* If they were already equal, forget it. */
1766 if (class1 == class2)
1767 return;
1769 for (elt = class2; elt; elt = next)
1771 unsigned int hash;
1772 rtx exp = elt->exp;
1773 enum machine_mode mode = elt->mode;
1775 next = elt->next_same_value;
1777 /* Remove old entry, make a new one in CLASS1's class.
1778 Don't do this for invalid entries as we cannot find their
1779 hash code (it also isn't necessary). */
1780 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1782 bool need_rehash = false;
1784 hash_arg_in_memory = 0;
1785 hash = HASH (exp, mode);
1787 if (REG_P (exp))
1789 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1790 delete_reg_equiv (REGNO (exp));
1793 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1794 remove_pseudo_from_table (exp, hash);
1795 else
1796 remove_from_table (elt, hash);
1798 if (insert_regs (exp, class1, 0) || need_rehash)
1800 rehash_using_reg (exp);
1801 hash = HASH (exp, mode);
1803 new_elt = insert (exp, class1, hash, mode);
1804 new_elt->in_memory = hash_arg_in_memory;
1809 /* Flush the entire hash table. */
1811 static void
1812 flush_hash_table (void)
1814 int i;
1815 struct table_elt *p;
1817 for (i = 0; i < HASH_SIZE; i++)
1818 for (p = table[i]; p; p = table[i])
1820 /* Note that invalidate can remove elements
1821 after P in the current hash chain. */
1822 if (REG_P (p->exp))
1823 invalidate (p->exp, VOIDmode);
1824 else
1825 remove_from_table (p, i);
1829 /* Function called for each rtx to check whether true dependence exist. */
1830 struct check_dependence_data
1832 enum machine_mode mode;
1833 rtx exp;
1834 rtx addr;
1837 static int
1838 check_dependence (rtx *x, void *data)
1840 struct check_dependence_data *d = (struct check_dependence_data *) data;
1841 if (*x && MEM_P (*x))
1842 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX);
1843 else
1844 return 0;
1847 /* Remove from the hash table, or mark as invalid, all expressions whose
1848 values could be altered by storing in X. X is a register, a subreg, or
1849 a memory reference with nonvarying address (because, when a memory
1850 reference with a varying address is stored in, all memory references are
1851 removed by invalidate_memory so specific invalidation is superfluous).
1852 FULL_MODE, if not VOIDmode, indicates that this much should be
1853 invalidated instead of just the amount indicated by the mode of X. This
1854 is only used for bitfield stores into memory.
1856 A nonvarying address may be just a register or just a symbol reference,
1857 or it may be either of those plus a numeric offset. */
1859 static void
1860 invalidate (rtx x, enum machine_mode full_mode)
1862 int i;
1863 struct table_elt *p;
1864 rtx addr;
1866 switch (GET_CODE (x))
1868 case REG:
1870 /* If X is a register, dependencies on its contents are recorded
1871 through the qty number mechanism. Just change the qty number of
1872 the register, mark it as invalid for expressions that refer to it,
1873 and remove it itself. */
1874 unsigned int regno = REGNO (x);
1875 unsigned int hash = HASH (x, GET_MODE (x));
1877 /* Remove REGNO from any quantity list it might be on and indicate
1878 that its value might have changed. If it is a pseudo, remove its
1879 entry from the hash table.
1881 For a hard register, we do the first two actions above for any
1882 additional hard registers corresponding to X. Then, if any of these
1883 registers are in the table, we must remove any REG entries that
1884 overlap these registers. */
1886 delete_reg_equiv (regno);
1887 REG_TICK (regno)++;
1888 SUBREG_TICKED (regno) = -1;
1890 if (regno >= FIRST_PSEUDO_REGISTER)
1891 remove_pseudo_from_table (x, hash);
1892 else
1894 HOST_WIDE_INT in_table
1895 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1896 unsigned int endregno = END_HARD_REGNO (x);
1897 unsigned int tregno, tendregno, rn;
1898 struct table_elt *p, *next;
1900 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1902 for (rn = regno + 1; rn < endregno; rn++)
1904 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1905 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1906 delete_reg_equiv (rn);
1907 REG_TICK (rn)++;
1908 SUBREG_TICKED (rn) = -1;
1911 if (in_table)
1912 for (hash = 0; hash < HASH_SIZE; hash++)
1913 for (p = table[hash]; p; p = next)
1915 next = p->next_same_hash;
1917 if (!REG_P (p->exp)
1918 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1919 continue;
1921 tregno = REGNO (p->exp);
1922 tendregno = END_HARD_REGNO (p->exp);
1923 if (tendregno > regno && tregno < endregno)
1924 remove_from_table (p, hash);
1928 return;
1930 case SUBREG:
1931 invalidate (SUBREG_REG (x), VOIDmode);
1932 return;
1934 case PARALLEL:
1935 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1936 invalidate (XVECEXP (x, 0, i), VOIDmode);
1937 return;
1939 case EXPR_LIST:
1940 /* This is part of a disjoint return value; extract the location in
1941 question ignoring the offset. */
1942 invalidate (XEXP (x, 0), VOIDmode);
1943 return;
1945 case MEM:
1946 addr = canon_rtx (get_addr (XEXP (x, 0)));
1947 /* Calculate the canonical version of X here so that
1948 true_dependence doesn't generate new RTL for X on each call. */
1949 x = canon_rtx (x);
1951 /* Remove all hash table elements that refer to overlapping pieces of
1952 memory. */
1953 if (full_mode == VOIDmode)
1954 full_mode = GET_MODE (x);
1956 for (i = 0; i < HASH_SIZE; i++)
1958 struct table_elt *next;
1960 for (p = table[i]; p; p = next)
1962 next = p->next_same_hash;
1963 if (p->in_memory)
1965 struct check_dependence_data d;
1967 /* Just canonicalize the expression once;
1968 otherwise each time we call invalidate
1969 true_dependence will canonicalize the
1970 expression again. */
1971 if (!p->canon_exp)
1972 p->canon_exp = canon_rtx (p->exp);
1973 d.exp = x;
1974 d.addr = addr;
1975 d.mode = full_mode;
1976 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1977 remove_from_table (p, i);
1981 return;
1983 default:
1984 gcc_unreachable ();
1988 /* Remove all expressions that refer to register REGNO,
1989 since they are already invalid, and we are about to
1990 mark that register valid again and don't want the old
1991 expressions to reappear as valid. */
1993 static void
1994 remove_invalid_refs (unsigned int regno)
1996 unsigned int i;
1997 struct table_elt *p, *next;
1999 for (i = 0; i < HASH_SIZE; i++)
2000 for (p = table[i]; p; p = next)
2002 next = p->next_same_hash;
2003 if (!REG_P (p->exp)
2004 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2005 remove_from_table (p, i);
2009 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2010 and mode MODE. */
2011 static void
2012 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2013 enum machine_mode mode)
2015 unsigned int i;
2016 struct table_elt *p, *next;
2017 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2019 for (i = 0; i < HASH_SIZE; i++)
2020 for (p = table[i]; p; p = next)
2022 rtx exp = p->exp;
2023 next = p->next_same_hash;
2025 if (!REG_P (exp)
2026 && (GET_CODE (exp) != SUBREG
2027 || !REG_P (SUBREG_REG (exp))
2028 || REGNO (SUBREG_REG (exp)) != regno
2029 || (((SUBREG_BYTE (exp)
2030 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2031 && SUBREG_BYTE (exp) <= end))
2032 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2033 remove_from_table (p, i);
2037 /* Recompute the hash codes of any valid entries in the hash table that
2038 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2040 This is called when we make a jump equivalence. */
2042 static void
2043 rehash_using_reg (rtx x)
2045 unsigned int i;
2046 struct table_elt *p, *next;
2047 unsigned hash;
2049 if (GET_CODE (x) == SUBREG)
2050 x = SUBREG_REG (x);
2052 /* If X is not a register or if the register is known not to be in any
2053 valid entries in the table, we have no work to do. */
2055 if (!REG_P (x)
2056 || REG_IN_TABLE (REGNO (x)) < 0
2057 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2058 return;
2060 /* Scan all hash chains looking for valid entries that mention X.
2061 If we find one and it is in the wrong hash chain, move it. */
2063 for (i = 0; i < HASH_SIZE; i++)
2064 for (p = table[i]; p; p = next)
2066 next = p->next_same_hash;
2067 if (reg_mentioned_p (x, p->exp)
2068 && exp_equiv_p (p->exp, p->exp, 1, false)
2069 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2071 if (p->next_same_hash)
2072 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2074 if (p->prev_same_hash)
2075 p->prev_same_hash->next_same_hash = p->next_same_hash;
2076 else
2077 table[i] = p->next_same_hash;
2079 p->next_same_hash = table[hash];
2080 p->prev_same_hash = 0;
2081 if (table[hash])
2082 table[hash]->prev_same_hash = p;
2083 table[hash] = p;
2088 /* Remove from the hash table any expression that is a call-clobbered
2089 register. Also update their TICK values. */
2091 static void
2092 invalidate_for_call (void)
2094 unsigned int regno, endregno;
2095 unsigned int i;
2096 unsigned hash;
2097 struct table_elt *p, *next;
2098 int in_table = 0;
2100 /* Go through all the hard registers. For each that is clobbered in
2101 a CALL_INSN, remove the register from quantity chains and update
2102 reg_tick if defined. Also see if any of these registers is currently
2103 in the table. */
2105 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2106 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2108 delete_reg_equiv (regno);
2109 if (REG_TICK (regno) >= 0)
2111 REG_TICK (regno)++;
2112 SUBREG_TICKED (regno) = -1;
2115 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2118 /* In the case where we have no call-clobbered hard registers in the
2119 table, we are done. Otherwise, scan the table and remove any
2120 entry that overlaps a call-clobbered register. */
2122 if (in_table)
2123 for (hash = 0; hash < HASH_SIZE; hash++)
2124 for (p = table[hash]; p; p = next)
2126 next = p->next_same_hash;
2128 if (!REG_P (p->exp)
2129 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2130 continue;
2132 regno = REGNO (p->exp);
2133 endregno = END_HARD_REGNO (p->exp);
2135 for (i = regno; i < endregno; i++)
2136 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2138 remove_from_table (p, hash);
2139 break;
2144 /* Given an expression X of type CONST,
2145 and ELT which is its table entry (or 0 if it
2146 is not in the hash table),
2147 return an alternate expression for X as a register plus integer.
2148 If none can be found, return 0. */
2150 static rtx
2151 use_related_value (rtx x, struct table_elt *elt)
2153 struct table_elt *relt = 0;
2154 struct table_elt *p, *q;
2155 HOST_WIDE_INT offset;
2157 /* First, is there anything related known?
2158 If we have a table element, we can tell from that.
2159 Otherwise, must look it up. */
2161 if (elt != 0 && elt->related_value != 0)
2162 relt = elt;
2163 else if (elt == 0 && GET_CODE (x) == CONST)
2165 rtx subexp = get_related_value (x);
2166 if (subexp != 0)
2167 relt = lookup (subexp,
2168 SAFE_HASH (subexp, GET_MODE (subexp)),
2169 GET_MODE (subexp));
2172 if (relt == 0)
2173 return 0;
2175 /* Search all related table entries for one that has an
2176 equivalent register. */
2178 p = relt;
2179 while (1)
2181 /* This loop is strange in that it is executed in two different cases.
2182 The first is when X is already in the table. Then it is searching
2183 the RELATED_VALUE list of X's class (RELT). The second case is when
2184 X is not in the table. Then RELT points to a class for the related
2185 value.
2187 Ensure that, whatever case we are in, that we ignore classes that have
2188 the same value as X. */
2190 if (rtx_equal_p (x, p->exp))
2191 q = 0;
2192 else
2193 for (q = p->first_same_value; q; q = q->next_same_value)
2194 if (REG_P (q->exp))
2195 break;
2197 if (q)
2198 break;
2200 p = p->related_value;
2202 /* We went all the way around, so there is nothing to be found.
2203 Alternatively, perhaps RELT was in the table for some other reason
2204 and it has no related values recorded. */
2205 if (p == relt || p == 0)
2206 break;
2209 if (q == 0)
2210 return 0;
2212 offset = (get_integer_term (x) - get_integer_term (p->exp));
2213 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2214 return plus_constant (q->mode, q->exp, offset);
2218 /* Hash a string. Just add its bytes up. */
2219 static inline unsigned
2220 hash_rtx_string (const char *ps)
2222 unsigned hash = 0;
2223 const unsigned char *p = (const unsigned char *) ps;
2225 if (p)
2226 while (*p)
2227 hash += *p++;
2229 return hash;
2232 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2233 When the callback returns true, we continue with the new rtx. */
2235 unsigned
2236 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2237 int *do_not_record_p, int *hash_arg_in_memory_p,
2238 bool have_reg_qty, hash_rtx_callback_function cb)
2240 int i, j;
2241 unsigned hash = 0;
2242 enum rtx_code code;
2243 const char *fmt;
2244 enum machine_mode newmode;
2245 rtx newx;
2247 /* Used to turn recursion into iteration. We can't rely on GCC's
2248 tail-recursion elimination since we need to keep accumulating values
2249 in HASH. */
2250 repeat:
2251 if (x == 0)
2252 return hash;
2254 /* Invoke the callback first. */
2255 if (cb != NULL
2256 && ((*cb) (x, mode, &newx, &newmode)))
2258 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2259 hash_arg_in_memory_p, have_reg_qty, cb);
2260 return hash;
2263 code = GET_CODE (x);
2264 switch (code)
2266 case REG:
2268 unsigned int regno = REGNO (x);
2270 if (do_not_record_p && !reload_completed)
2272 /* On some machines, we can't record any non-fixed hard register,
2273 because extending its life will cause reload problems. We
2274 consider ap, fp, sp, gp to be fixed for this purpose.
2276 We also consider CCmode registers to be fixed for this purpose;
2277 failure to do so leads to failure to simplify 0<100 type of
2278 conditionals.
2280 On all machines, we can't record any global registers.
2281 Nor should we record any register that is in a small
2282 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2283 bool record;
2285 if (regno >= FIRST_PSEUDO_REGISTER)
2286 record = true;
2287 else if (x == frame_pointer_rtx
2288 || x == hard_frame_pointer_rtx
2289 || x == arg_pointer_rtx
2290 || x == stack_pointer_rtx
2291 || x == pic_offset_table_rtx)
2292 record = true;
2293 else if (global_regs[regno])
2294 record = false;
2295 else if (fixed_regs[regno])
2296 record = true;
2297 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2298 record = true;
2299 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2300 record = false;
2301 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2302 record = false;
2303 else
2304 record = true;
2306 if (!record)
2308 *do_not_record_p = 1;
2309 return 0;
2313 hash += ((unsigned int) REG << 7);
2314 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2315 return hash;
2318 /* We handle SUBREG of a REG specially because the underlying
2319 reg changes its hash value with every value change; we don't
2320 want to have to forget unrelated subregs when one subreg changes. */
2321 case SUBREG:
2323 if (REG_P (SUBREG_REG (x)))
2325 hash += (((unsigned int) SUBREG << 7)
2326 + REGNO (SUBREG_REG (x))
2327 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2328 return hash;
2330 break;
2333 case CONST_INT:
2334 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2335 + (unsigned int) INTVAL (x));
2336 return hash;
2338 case CONST_DOUBLE:
2339 /* This is like the general case, except that it only counts
2340 the integers representing the constant. */
2341 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2342 if (GET_MODE (x) != VOIDmode)
2343 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2344 else
2345 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2346 + (unsigned int) CONST_DOUBLE_HIGH (x));
2347 return hash;
2349 case CONST_FIXED:
2350 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2351 hash += fixed_hash (CONST_FIXED_VALUE (x));
2352 return hash;
2354 case CONST_VECTOR:
2356 int units;
2357 rtx elt;
2359 units = CONST_VECTOR_NUNITS (x);
2361 for (i = 0; i < units; ++i)
2363 elt = CONST_VECTOR_ELT (x, i);
2364 hash += hash_rtx_cb (elt, GET_MODE (elt),
2365 do_not_record_p, hash_arg_in_memory_p,
2366 have_reg_qty, cb);
2369 return hash;
2372 /* Assume there is only one rtx object for any given label. */
2373 case LABEL_REF:
2374 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2375 differences and differences between each stage's debugging dumps. */
2376 hash += (((unsigned int) LABEL_REF << 7)
2377 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2378 return hash;
2380 case SYMBOL_REF:
2382 /* Don't hash on the symbol's address to avoid bootstrap differences.
2383 Different hash values may cause expressions to be recorded in
2384 different orders and thus different registers to be used in the
2385 final assembler. This also avoids differences in the dump files
2386 between various stages. */
2387 unsigned int h = 0;
2388 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2390 while (*p)
2391 h += (h << 7) + *p++; /* ??? revisit */
2393 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2394 return hash;
2397 case MEM:
2398 /* We don't record if marked volatile or if BLKmode since we don't
2399 know the size of the move. */
2400 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2402 *do_not_record_p = 1;
2403 return 0;
2405 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2406 *hash_arg_in_memory_p = 1;
2408 /* Now that we have already found this special case,
2409 might as well speed it up as much as possible. */
2410 hash += (unsigned) MEM;
2411 x = XEXP (x, 0);
2412 goto repeat;
2414 case USE:
2415 /* A USE that mentions non-volatile memory needs special
2416 handling since the MEM may be BLKmode which normally
2417 prevents an entry from being made. Pure calls are
2418 marked by a USE which mentions BLKmode memory.
2419 See calls.c:emit_call_1. */
2420 if (MEM_P (XEXP (x, 0))
2421 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2423 hash += (unsigned) USE;
2424 x = XEXP (x, 0);
2426 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2427 *hash_arg_in_memory_p = 1;
2429 /* Now that we have already found this special case,
2430 might as well speed it up as much as possible. */
2431 hash += (unsigned) MEM;
2432 x = XEXP (x, 0);
2433 goto repeat;
2435 break;
2437 case PRE_DEC:
2438 case PRE_INC:
2439 case POST_DEC:
2440 case POST_INC:
2441 case PRE_MODIFY:
2442 case POST_MODIFY:
2443 case PC:
2444 case CC0:
2445 case CALL:
2446 case UNSPEC_VOLATILE:
2447 if (do_not_record_p) {
2448 *do_not_record_p = 1;
2449 return 0;
2451 else
2452 return hash;
2453 break;
2455 case ASM_OPERANDS:
2456 if (do_not_record_p && MEM_VOLATILE_P (x))
2458 *do_not_record_p = 1;
2459 return 0;
2461 else
2463 /* We don't want to take the filename and line into account. */
2464 hash += (unsigned) code + (unsigned) GET_MODE (x)
2465 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2466 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2467 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2469 if (ASM_OPERANDS_INPUT_LENGTH (x))
2471 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2473 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2474 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2475 do_not_record_p, hash_arg_in_memory_p,
2476 have_reg_qty, cb)
2477 + hash_rtx_string
2478 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2481 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2482 x = ASM_OPERANDS_INPUT (x, 0);
2483 mode = GET_MODE (x);
2484 goto repeat;
2487 return hash;
2489 break;
2491 default:
2492 break;
2495 i = GET_RTX_LENGTH (code) - 1;
2496 hash += (unsigned) code + (unsigned) GET_MODE (x);
2497 fmt = GET_RTX_FORMAT (code);
2498 for (; i >= 0; i--)
2500 switch (fmt[i])
2502 case 'e':
2503 /* If we are about to do the last recursive call
2504 needed at this level, change it into iteration.
2505 This function is called enough to be worth it. */
2506 if (i == 0)
2508 x = XEXP (x, i);
2509 goto repeat;
2512 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2513 hash_arg_in_memory_p,
2514 have_reg_qty, cb);
2515 break;
2517 case 'E':
2518 for (j = 0; j < XVECLEN (x, i); j++)
2519 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2520 hash_arg_in_memory_p,
2521 have_reg_qty, cb);
2522 break;
2524 case 's':
2525 hash += hash_rtx_string (XSTR (x, i));
2526 break;
2528 case 'i':
2529 hash += (unsigned int) XINT (x, i);
2530 break;
2532 case '0': case 't':
2533 /* Unused. */
2534 break;
2536 default:
2537 gcc_unreachable ();
2541 return hash;
2544 /* Hash an rtx. We are careful to make sure the value is never negative.
2545 Equivalent registers hash identically.
2546 MODE is used in hashing for CONST_INTs only;
2547 otherwise the mode of X is used.
2549 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2551 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2552 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2554 Note that cse_insn knows that the hash code of a MEM expression
2555 is just (int) MEM plus the hash code of the address. */
2557 unsigned
2558 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2559 int *hash_arg_in_memory_p, bool have_reg_qty)
2561 return hash_rtx_cb (x, mode, do_not_record_p,
2562 hash_arg_in_memory_p, have_reg_qty, NULL);
2565 /* Hash an rtx X for cse via hash_rtx.
2566 Stores 1 in do_not_record if any subexpression is volatile.
2567 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2568 does not have the RTX_UNCHANGING_P bit set. */
2570 static inline unsigned
2571 canon_hash (rtx x, enum machine_mode mode)
2573 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2576 /* Like canon_hash but with no side effects, i.e. do_not_record
2577 and hash_arg_in_memory are not changed. */
2579 static inline unsigned
2580 safe_hash (rtx x, enum machine_mode mode)
2582 int dummy_do_not_record;
2583 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2586 /* Return 1 iff X and Y would canonicalize into the same thing,
2587 without actually constructing the canonicalization of either one.
2588 If VALIDATE is nonzero,
2589 we assume X is an expression being processed from the rtl
2590 and Y was found in the hash table. We check register refs
2591 in Y for being marked as valid.
2593 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2596 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2598 int i, j;
2599 enum rtx_code code;
2600 const char *fmt;
2602 /* Note: it is incorrect to assume an expression is equivalent to itself
2603 if VALIDATE is nonzero. */
2604 if (x == y && !validate)
2605 return 1;
2607 if (x == 0 || y == 0)
2608 return x == y;
2610 code = GET_CODE (x);
2611 if (code != GET_CODE (y))
2612 return 0;
2614 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2615 if (GET_MODE (x) != GET_MODE (y))
2616 return 0;
2618 /* MEMs referring to different address space are not equivalent. */
2619 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2620 return 0;
2622 switch (code)
2624 case PC:
2625 case CC0:
2626 case CONST_INT:
2627 case CONST_DOUBLE:
2628 case CONST_FIXED:
2629 return x == y;
2631 case LABEL_REF:
2632 return XEXP (x, 0) == XEXP (y, 0);
2634 case SYMBOL_REF:
2635 return XSTR (x, 0) == XSTR (y, 0);
2637 case REG:
2638 if (for_gcse)
2639 return REGNO (x) == REGNO (y);
2640 else
2642 unsigned int regno = REGNO (y);
2643 unsigned int i;
2644 unsigned int endregno = END_REGNO (y);
2646 /* If the quantities are not the same, the expressions are not
2647 equivalent. If there are and we are not to validate, they
2648 are equivalent. Otherwise, ensure all regs are up-to-date. */
2650 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2651 return 0;
2653 if (! validate)
2654 return 1;
2656 for (i = regno; i < endregno; i++)
2657 if (REG_IN_TABLE (i) != REG_TICK (i))
2658 return 0;
2660 return 1;
2663 case MEM:
2664 if (for_gcse)
2666 /* A volatile mem should not be considered equivalent to any
2667 other. */
2668 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2669 return 0;
2671 /* Can't merge two expressions in different alias sets, since we
2672 can decide that the expression is transparent in a block when
2673 it isn't, due to it being set with the different alias set.
2675 Also, can't merge two expressions with different MEM_ATTRS.
2676 They could e.g. be two different entities allocated into the
2677 same space on the stack (see e.g. PR25130). In that case, the
2678 MEM addresses can be the same, even though the two MEMs are
2679 absolutely not equivalent.
2681 But because really all MEM attributes should be the same for
2682 equivalent MEMs, we just use the invariant that MEMs that have
2683 the same attributes share the same mem_attrs data structure. */
2684 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2685 return 0;
2687 break;
2689 /* For commutative operations, check both orders. */
2690 case PLUS:
2691 case MULT:
2692 case AND:
2693 case IOR:
2694 case XOR:
2695 case NE:
2696 case EQ:
2697 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2698 validate, for_gcse)
2699 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2700 validate, for_gcse))
2701 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2702 validate, for_gcse)
2703 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2704 validate, for_gcse)));
2706 case ASM_OPERANDS:
2707 /* We don't use the generic code below because we want to
2708 disregard filename and line numbers. */
2710 /* A volatile asm isn't equivalent to any other. */
2711 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2712 return 0;
2714 if (GET_MODE (x) != GET_MODE (y)
2715 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2716 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2717 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2718 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2719 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2720 return 0;
2722 if (ASM_OPERANDS_INPUT_LENGTH (x))
2724 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2725 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2726 ASM_OPERANDS_INPUT (y, i),
2727 validate, for_gcse)
2728 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2729 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2730 return 0;
2733 return 1;
2735 default:
2736 break;
2739 /* Compare the elements. If any pair of corresponding elements
2740 fail to match, return 0 for the whole thing. */
2742 fmt = GET_RTX_FORMAT (code);
2743 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2745 switch (fmt[i])
2747 case 'e':
2748 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2749 validate, for_gcse))
2750 return 0;
2751 break;
2753 case 'E':
2754 if (XVECLEN (x, i) != XVECLEN (y, i))
2755 return 0;
2756 for (j = 0; j < XVECLEN (x, i); j++)
2757 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2758 validate, for_gcse))
2759 return 0;
2760 break;
2762 case 's':
2763 if (strcmp (XSTR (x, i), XSTR (y, i)))
2764 return 0;
2765 break;
2767 case 'i':
2768 if (XINT (x, i) != XINT (y, i))
2769 return 0;
2770 break;
2772 case 'w':
2773 if (XWINT (x, i) != XWINT (y, i))
2774 return 0;
2775 break;
2777 case '0':
2778 case 't':
2779 break;
2781 default:
2782 gcc_unreachable ();
2786 return 1;
2789 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2790 the result if necessary. INSN is as for canon_reg. */
2792 static void
2793 validate_canon_reg (rtx *xloc, rtx insn)
2795 if (*xloc)
2797 rtx new_rtx = canon_reg (*xloc, insn);
2799 /* If replacing pseudo with hard reg or vice versa, ensure the
2800 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2801 gcc_assert (insn && new_rtx);
2802 validate_change (insn, xloc, new_rtx, 1);
2806 /* Canonicalize an expression:
2807 replace each register reference inside it
2808 with the "oldest" equivalent register.
2810 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2811 after we make our substitution. The calls are made with IN_GROUP nonzero
2812 so apply_change_group must be called upon the outermost return from this
2813 function (unless INSN is zero). The result of apply_change_group can
2814 generally be discarded since the changes we are making are optional. */
2816 static rtx
2817 canon_reg (rtx x, rtx insn)
2819 int i;
2820 enum rtx_code code;
2821 const char *fmt;
2823 if (x == 0)
2824 return x;
2826 code = GET_CODE (x);
2827 switch (code)
2829 case PC:
2830 case CC0:
2831 case CONST:
2832 case CONST_INT:
2833 case CONST_DOUBLE:
2834 case CONST_FIXED:
2835 case CONST_VECTOR:
2836 case SYMBOL_REF:
2837 case LABEL_REF:
2838 case ADDR_VEC:
2839 case ADDR_DIFF_VEC:
2840 return x;
2842 case REG:
2844 int first;
2845 int q;
2846 struct qty_table_elem *ent;
2848 /* Never replace a hard reg, because hard regs can appear
2849 in more than one machine mode, and we must preserve the mode
2850 of each occurrence. Also, some hard regs appear in
2851 MEMs that are shared and mustn't be altered. Don't try to
2852 replace any reg that maps to a reg of class NO_REGS. */
2853 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2854 || ! REGNO_QTY_VALID_P (REGNO (x)))
2855 return x;
2857 q = REG_QTY (REGNO (x));
2858 ent = &qty_table[q];
2859 first = ent->first_reg;
2860 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2861 : REGNO_REG_CLASS (first) == NO_REGS ? x
2862 : gen_rtx_REG (ent->mode, first));
2865 default:
2866 break;
2869 fmt = GET_RTX_FORMAT (code);
2870 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2872 int j;
2874 if (fmt[i] == 'e')
2875 validate_canon_reg (&XEXP (x, i), insn);
2876 else if (fmt[i] == 'E')
2877 for (j = 0; j < XVECLEN (x, i); j++)
2878 validate_canon_reg (&XVECEXP (x, i, j), insn);
2881 return x;
2884 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2885 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2886 what values are being compared.
2888 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2889 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2890 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2891 compared to produce cc0.
2893 The return value is the comparison operator and is either the code of
2894 A or the code corresponding to the inverse of the comparison. */
2896 static enum rtx_code
2897 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2898 enum machine_mode *pmode1, enum machine_mode *pmode2)
2900 rtx arg1, arg2;
2902 arg1 = *parg1, arg2 = *parg2;
2904 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2906 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2908 /* Set nonzero when we find something of interest. */
2909 rtx x = 0;
2910 int reverse_code = 0;
2911 struct table_elt *p = 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 enum 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 the same comparison we're already looking at, skip it. */
2990 if (COMPARISON_P (p->exp)
2991 && XEXP (p->exp, 0) == arg1
2992 && XEXP (p->exp, 1) == arg2)
2993 continue;
2995 if (GET_CODE (p->exp) == COMPARE
2996 /* Another possibility is that this machine has a compare insn
2997 that includes the comparison code. In that case, ARG1 would
2998 be equivalent to a comparison operation that would set ARG1 to
2999 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3000 ORIG_CODE is the actual comparison being done; if it is an EQ,
3001 we must reverse ORIG_CODE. On machine with a negative value
3002 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3003 || ((code == NE
3004 || (code == LT
3005 && val_signbit_known_set_p (inner_mode,
3006 STORE_FLAG_VALUE))
3007 #ifdef FLOAT_STORE_FLAG_VALUE
3008 || (code == LT
3009 && SCALAR_FLOAT_MODE_P (inner_mode)
3010 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3011 REAL_VALUE_NEGATIVE (fsfv)))
3012 #endif
3014 && COMPARISON_P (p->exp)))
3016 x = p->exp;
3017 break;
3019 else if ((code == EQ
3020 || (code == GE
3021 && val_signbit_known_set_p (inner_mode,
3022 STORE_FLAG_VALUE))
3023 #ifdef FLOAT_STORE_FLAG_VALUE
3024 || (code == GE
3025 && SCALAR_FLOAT_MODE_P (inner_mode)
3026 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3027 REAL_VALUE_NEGATIVE (fsfv)))
3028 #endif
3030 && COMPARISON_P (p->exp))
3032 reverse_code = 1;
3033 x = p->exp;
3034 break;
3037 /* If this non-trapping address, e.g. fp + constant, the
3038 equivalent is a better operand since it may let us predict
3039 the value of the comparison. */
3040 else if (!rtx_addr_can_trap_p (p->exp))
3042 arg1 = p->exp;
3043 continue;
3047 /* If we didn't find a useful equivalence for ARG1, we are done.
3048 Otherwise, set up for the next iteration. */
3049 if (x == 0)
3050 break;
3052 /* If we need to reverse the comparison, make sure that that is
3053 possible -- we can't necessarily infer the value of GE from LT
3054 with floating-point operands. */
3055 if (reverse_code)
3057 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3058 if (reversed == UNKNOWN)
3059 break;
3060 else
3061 code = reversed;
3063 else if (COMPARISON_P (x))
3064 code = GET_CODE (x);
3065 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3068 /* Return our results. Return the modes from before fold_rtx
3069 because fold_rtx might produce const_int, and then it's too late. */
3070 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3071 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3073 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)
3091 enum rtx_code code;
3092 enum machine_mode mode;
3093 const char *fmt;
3094 int i;
3095 rtx new_rtx = 0;
3096 int changed = 0;
3098 /* Operands of X. */
3099 rtx folded_arg0;
3100 rtx folded_arg1;
3102 /* Constant equivalents of first three operands of X;
3103 0 when no such equivalent is known. */
3104 rtx const_arg0;
3105 rtx const_arg1;
3106 rtx const_arg2;
3108 /* The mode of the first operand of X. We need this for sign and zero
3109 extends. */
3110 enum machine_mode mode_arg0;
3112 if (x == 0)
3113 return x;
3115 /* Try to perform some initial simplifications on X. */
3116 code = GET_CODE (x);
3117 switch (code)
3119 case MEM:
3120 case SUBREG:
3121 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3122 return new_rtx;
3123 return x;
3125 case CONST:
3126 case CONST_INT:
3127 case CONST_DOUBLE:
3128 case CONST_FIXED:
3129 case CONST_VECTOR:
3130 case SYMBOL_REF:
3131 case LABEL_REF:
3132 case REG:
3133 case PC:
3134 /* No use simplifying an EXPR_LIST
3135 since they are used only for lists of args
3136 in a function call's REG_EQUAL note. */
3137 case EXPR_LIST:
3138 return x;
3140 #ifdef HAVE_cc0
3141 case CC0:
3142 return prev_insn_cc0;
3143 #endif
3145 case ASM_OPERANDS:
3146 if (insn)
3148 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3149 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3150 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3152 return x;
3154 #ifdef NO_FUNCTION_CSE
3155 case CALL:
3156 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3157 return x;
3158 break;
3159 #endif
3161 /* Anything else goes through the loop below. */
3162 default:
3163 break;
3166 mode = GET_MODE (x);
3167 const_arg0 = 0;
3168 const_arg1 = 0;
3169 const_arg2 = 0;
3170 mode_arg0 = VOIDmode;
3172 /* Try folding our operands.
3173 Then see which ones have constant values known. */
3175 fmt = GET_RTX_FORMAT (code);
3176 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3177 if (fmt[i] == 'e')
3179 rtx folded_arg = XEXP (x, i), const_arg;
3180 enum machine_mode mode_arg = GET_MODE (folded_arg);
3182 switch (GET_CODE (folded_arg))
3184 case MEM:
3185 case REG:
3186 case SUBREG:
3187 const_arg = equiv_constant (folded_arg);
3188 break;
3190 case CONST:
3191 case CONST_INT:
3192 case SYMBOL_REF:
3193 case LABEL_REF:
3194 case CONST_DOUBLE:
3195 case CONST_FIXED:
3196 case CONST_VECTOR:
3197 const_arg = folded_arg;
3198 break;
3200 #ifdef HAVE_cc0
3201 case CC0:
3202 folded_arg = prev_insn_cc0;
3203 mode_arg = prev_insn_cc0_mode;
3204 const_arg = equiv_constant (folded_arg);
3205 break;
3206 #endif
3208 default:
3209 folded_arg = fold_rtx (folded_arg, insn);
3210 const_arg = equiv_constant (folded_arg);
3211 break;
3214 /* For the first three operands, see if the operand
3215 is constant or equivalent to a constant. */
3216 switch (i)
3218 case 0:
3219 folded_arg0 = folded_arg;
3220 const_arg0 = const_arg;
3221 mode_arg0 = mode_arg;
3222 break;
3223 case 1:
3224 folded_arg1 = folded_arg;
3225 const_arg1 = const_arg;
3226 break;
3227 case 2:
3228 const_arg2 = const_arg;
3229 break;
3232 /* Pick the least expensive of the argument and an equivalent constant
3233 argument. */
3234 if (const_arg != 0
3235 && const_arg != folded_arg
3236 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3238 /* It's not safe to substitute the operand of a conversion
3239 operator with a constant, as the conversion's identity
3240 depends upon the mode of its operand. This optimization
3241 is handled by the call to simplify_unary_operation. */
3242 && (GET_RTX_CLASS (code) != RTX_UNARY
3243 || GET_MODE (const_arg) == mode_arg0
3244 || (code != ZERO_EXTEND
3245 && code != SIGN_EXTEND
3246 && code != TRUNCATE
3247 && code != FLOAT_TRUNCATE
3248 && code != FLOAT_EXTEND
3249 && code != FLOAT
3250 && code != FIX
3251 && code != UNSIGNED_FLOAT
3252 && code != UNSIGNED_FIX)))
3253 folded_arg = const_arg;
3255 if (folded_arg == XEXP (x, i))
3256 continue;
3258 if (insn == NULL_RTX && !changed)
3259 x = copy_rtx (x);
3260 changed = 1;
3261 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3264 if (changed)
3266 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3267 consistent with the order in X. */
3268 if (canonicalize_change_group (insn, x))
3270 rtx tem;
3271 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3272 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3275 apply_change_group ();
3278 /* If X is an arithmetic operation, see if we can simplify it. */
3280 switch (GET_RTX_CLASS (code))
3282 case RTX_UNARY:
3284 /* We can't simplify extension ops unless we know the
3285 original mode. */
3286 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3287 && mode_arg0 == VOIDmode)
3288 break;
3290 new_rtx = simplify_unary_operation (code, mode,
3291 const_arg0 ? const_arg0 : folded_arg0,
3292 mode_arg0);
3294 break;
3296 case RTX_COMPARE:
3297 case RTX_COMM_COMPARE:
3298 /* See what items are actually being compared and set FOLDED_ARG[01]
3299 to those values and CODE to the actual comparison code. If any are
3300 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3301 do anything if both operands are already known to be constant. */
3303 /* ??? Vector mode comparisons are not supported yet. */
3304 if (VECTOR_MODE_P (mode))
3305 break;
3307 if (const_arg0 == 0 || const_arg1 == 0)
3309 struct table_elt *p0, *p1;
3310 rtx true_rtx, false_rtx;
3311 enum machine_mode mode_arg1;
3313 if (SCALAR_FLOAT_MODE_P (mode))
3315 #ifdef FLOAT_STORE_FLAG_VALUE
3316 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3317 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3318 #else
3319 true_rtx = NULL_RTX;
3320 #endif
3321 false_rtx = CONST0_RTX (mode);
3323 else
3325 true_rtx = const_true_rtx;
3326 false_rtx = const0_rtx;
3329 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3330 &mode_arg0, &mode_arg1);
3332 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3333 what kinds of things are being compared, so we can't do
3334 anything with this comparison. */
3336 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3337 break;
3339 const_arg0 = equiv_constant (folded_arg0);
3340 const_arg1 = equiv_constant (folded_arg1);
3342 /* If we do not now have two constants being compared, see
3343 if we can nevertheless deduce some things about the
3344 comparison. */
3345 if (const_arg0 == 0 || const_arg1 == 0)
3347 if (const_arg1 != NULL)
3349 rtx cheapest_simplification;
3350 int cheapest_cost;
3351 rtx simp_result;
3352 struct table_elt *p;
3354 /* See if we can find an equivalent of folded_arg0
3355 that gets us a cheaper expression, possibly a
3356 constant through simplifications. */
3357 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3358 mode_arg0);
3360 if (p != NULL)
3362 cheapest_simplification = x;
3363 cheapest_cost = COST (x);
3365 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3367 int cost;
3369 /* If the entry isn't valid, skip it. */
3370 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3371 continue;
3373 /* Try to simplify using this equivalence. */
3374 simp_result
3375 = simplify_relational_operation (code, mode,
3376 mode_arg0,
3377 p->exp,
3378 const_arg1);
3380 if (simp_result == NULL)
3381 continue;
3383 cost = COST (simp_result);
3384 if (cost < cheapest_cost)
3386 cheapest_cost = cost;
3387 cheapest_simplification = simp_result;
3391 /* If we have a cheaper expression now, use that
3392 and try folding it further, from the top. */
3393 if (cheapest_simplification != x)
3394 return fold_rtx (copy_rtx (cheapest_simplification),
3395 insn);
3399 /* See if the two operands are the same. */
3401 if ((REG_P (folded_arg0)
3402 && REG_P (folded_arg1)
3403 && (REG_QTY (REGNO (folded_arg0))
3404 == REG_QTY (REGNO (folded_arg1))))
3405 || ((p0 = lookup (folded_arg0,
3406 SAFE_HASH (folded_arg0, mode_arg0),
3407 mode_arg0))
3408 && (p1 = lookup (folded_arg1,
3409 SAFE_HASH (folded_arg1, mode_arg0),
3410 mode_arg0))
3411 && p0->first_same_value == p1->first_same_value))
3412 folded_arg1 = folded_arg0;
3414 /* If FOLDED_ARG0 is a register, see if the comparison we are
3415 doing now is either the same as we did before or the reverse
3416 (we only check the reverse if not floating-point). */
3417 else if (REG_P (folded_arg0))
3419 int qty = REG_QTY (REGNO (folded_arg0));
3421 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3423 struct qty_table_elem *ent = &qty_table[qty];
3425 if ((comparison_dominates_p (ent->comparison_code, code)
3426 || (! FLOAT_MODE_P (mode_arg0)
3427 && comparison_dominates_p (ent->comparison_code,
3428 reverse_condition (code))))
3429 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3430 || (const_arg1
3431 && rtx_equal_p (ent->comparison_const,
3432 const_arg1))
3433 || (REG_P (folded_arg1)
3434 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3436 if (comparison_dominates_p (ent->comparison_code, code))
3438 if (true_rtx)
3439 return true_rtx;
3440 else
3441 break;
3443 else
3444 return false_rtx;
3451 /* If we are comparing against zero, see if the first operand is
3452 equivalent to an IOR with a constant. If so, we may be able to
3453 determine the result of this comparison. */
3454 if (const_arg1 == const0_rtx && !const_arg0)
3456 rtx y = lookup_as_function (folded_arg0, IOR);
3457 rtx inner_const;
3459 if (y != 0
3460 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3461 && CONST_INT_P (inner_const)
3462 && INTVAL (inner_const) != 0)
3463 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3467 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3468 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3469 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3471 break;
3473 case RTX_BIN_ARITH:
3474 case RTX_COMM_ARITH:
3475 switch (code)
3477 case PLUS:
3478 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3479 with that LABEL_REF as its second operand. If so, the result is
3480 the first operand of that MINUS. This handles switches with an
3481 ADDR_DIFF_VEC table. */
3482 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3484 rtx y
3485 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3486 : lookup_as_function (folded_arg0, MINUS);
3488 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3489 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3490 return XEXP (y, 0);
3492 /* Now try for a CONST of a MINUS like the above. */
3493 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3494 : lookup_as_function (folded_arg0, CONST))) != 0
3495 && GET_CODE (XEXP (y, 0)) == MINUS
3496 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3497 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3498 return XEXP (XEXP (y, 0), 0);
3501 /* Likewise if the operands are in the other order. */
3502 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3504 rtx y
3505 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3506 : lookup_as_function (folded_arg1, MINUS);
3508 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3509 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3510 return XEXP (y, 0);
3512 /* Now try for a CONST of a MINUS like the above. */
3513 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3514 : lookup_as_function (folded_arg1, CONST))) != 0
3515 && GET_CODE (XEXP (y, 0)) == MINUS
3516 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3517 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3518 return XEXP (XEXP (y, 0), 0);
3521 /* If second operand is a register equivalent to a negative
3522 CONST_INT, see if we can find a register equivalent to the
3523 positive constant. Make a MINUS if so. Don't do this for
3524 a non-negative constant since we might then alternate between
3525 choosing positive and negative constants. Having the positive
3526 constant previously-used is the more common case. Be sure
3527 the resulting constant is non-negative; if const_arg1 were
3528 the smallest negative number this would overflow: depending
3529 on the mode, this would either just be the same value (and
3530 hence not save anything) or be incorrect. */
3531 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3532 && INTVAL (const_arg1) < 0
3533 /* This used to test
3535 -INTVAL (const_arg1) >= 0
3537 But The Sun V5.0 compilers mis-compiled that test. So
3538 instead we test for the problematic value in a more direct
3539 manner and hope the Sun compilers get it correct. */
3540 && INTVAL (const_arg1) !=
3541 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3542 && REG_P (folded_arg1))
3544 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3545 struct table_elt *p
3546 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3548 if (p)
3549 for (p = p->first_same_value; p; p = p->next_same_value)
3550 if (REG_P (p->exp))
3551 return simplify_gen_binary (MINUS, mode, folded_arg0,
3552 canon_reg (p->exp, NULL_RTX));
3554 goto from_plus;
3556 case MINUS:
3557 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3558 If so, produce (PLUS Z C2-C). */
3559 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3561 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3562 if (y && CONST_INT_P (XEXP (y, 1)))
3563 return fold_rtx (plus_constant (mode, copy_rtx (y),
3564 -INTVAL (const_arg1)),
3565 NULL_RTX);
3568 /* Fall through. */
3570 from_plus:
3571 case SMIN: case SMAX: case UMIN: case UMAX:
3572 case IOR: case AND: case XOR:
3573 case MULT:
3574 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3575 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3576 is known to be of similar form, we may be able to replace the
3577 operation with a combined operation. This may eliminate the
3578 intermediate operation if every use is simplified in this way.
3579 Note that the similar optimization done by combine.c only works
3580 if the intermediate operation's result has only one reference. */
3582 if (REG_P (folded_arg0)
3583 && const_arg1 && CONST_INT_P (const_arg1))
3585 int is_shift
3586 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3587 rtx y, inner_const, new_const;
3588 rtx canon_const_arg1 = const_arg1;
3589 enum rtx_code associate_code;
3591 if (is_shift
3592 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3593 || INTVAL (const_arg1) < 0))
3595 if (SHIFT_COUNT_TRUNCATED)
3596 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3597 & (GET_MODE_BITSIZE (mode)
3598 - 1));
3599 else
3600 break;
3603 y = lookup_as_function (folded_arg0, code);
3604 if (y == 0)
3605 break;
3607 /* If we have compiled a statement like
3608 "if (x == (x & mask1))", and now are looking at
3609 "x & mask2", we will have a case where the first operand
3610 of Y is the same as our first operand. Unless we detect
3611 this case, an infinite loop will result. */
3612 if (XEXP (y, 0) == folded_arg0)
3613 break;
3615 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3616 if (!inner_const || !CONST_INT_P (inner_const))
3617 break;
3619 /* Don't associate these operations if they are a PLUS with the
3620 same constant and it is a power of two. These might be doable
3621 with a pre- or post-increment. Similarly for two subtracts of
3622 identical powers of two with post decrement. */
3624 if (code == PLUS && const_arg1 == inner_const
3625 && ((HAVE_PRE_INCREMENT
3626 && exact_log2 (INTVAL (const_arg1)) >= 0)
3627 || (HAVE_POST_INCREMENT
3628 && exact_log2 (INTVAL (const_arg1)) >= 0)
3629 || (HAVE_PRE_DECREMENT
3630 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3631 || (HAVE_POST_DECREMENT
3632 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3633 break;
3635 /* ??? Vector mode shifts by scalar
3636 shift operand are not supported yet. */
3637 if (is_shift && VECTOR_MODE_P (mode))
3638 break;
3640 if (is_shift
3641 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3642 || INTVAL (inner_const) < 0))
3644 if (SHIFT_COUNT_TRUNCATED)
3645 inner_const = GEN_INT (INTVAL (inner_const)
3646 & (GET_MODE_BITSIZE (mode) - 1));
3647 else
3648 break;
3651 /* Compute the code used to compose the constants. For example,
3652 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3654 associate_code = (is_shift || code == MINUS ? PLUS : code);
3656 new_const = simplify_binary_operation (associate_code, mode,
3657 canon_const_arg1,
3658 inner_const);
3660 if (new_const == 0)
3661 break;
3663 /* If we are associating shift operations, don't let this
3664 produce a shift of the size of the object or larger.
3665 This could occur when we follow a sign-extend by a right
3666 shift on a machine that does a sign-extend as a pair
3667 of shifts. */
3669 if (is_shift
3670 && CONST_INT_P (new_const)
3671 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3673 /* As an exception, we can turn an ASHIFTRT of this
3674 form into a shift of the number of bits - 1. */
3675 if (code == ASHIFTRT)
3676 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3677 else if (!side_effects_p (XEXP (y, 0)))
3678 return CONST0_RTX (mode);
3679 else
3680 break;
3683 y = copy_rtx (XEXP (y, 0));
3685 /* If Y contains our first operand (the most common way this
3686 can happen is if Y is a MEM), we would do into an infinite
3687 loop if we tried to fold it. So don't in that case. */
3689 if (! reg_mentioned_p (folded_arg0, y))
3690 y = fold_rtx (y, insn);
3692 return simplify_gen_binary (code, mode, y, new_const);
3694 break;
3696 case DIV: case UDIV:
3697 /* ??? The associative optimization performed immediately above is
3698 also possible for DIV and UDIV using associate_code of MULT.
3699 However, we would need extra code to verify that the
3700 multiplication does not overflow, that is, there is no overflow
3701 in the calculation of new_const. */
3702 break;
3704 default:
3705 break;
3708 new_rtx = simplify_binary_operation (code, mode,
3709 const_arg0 ? const_arg0 : folded_arg0,
3710 const_arg1 ? const_arg1 : folded_arg1);
3711 break;
3713 case RTX_OBJ:
3714 /* (lo_sum (high X) X) is simply X. */
3715 if (code == LO_SUM && const_arg0 != 0
3716 && GET_CODE (const_arg0) == HIGH
3717 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3718 return const_arg1;
3719 break;
3721 case RTX_TERNARY:
3722 case RTX_BITFIELD_OPS:
3723 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3724 const_arg0 ? const_arg0 : folded_arg0,
3725 const_arg1 ? const_arg1 : folded_arg1,
3726 const_arg2 ? const_arg2 : XEXP (x, 2));
3727 break;
3729 default:
3730 break;
3733 return new_rtx ? new_rtx : x;
3736 /* Return a constant value currently equivalent to X.
3737 Return 0 if we don't know one. */
3739 static rtx
3740 equiv_constant (rtx x)
3742 if (REG_P (x)
3743 && REGNO_QTY_VALID_P (REGNO (x)))
3745 int x_q = REG_QTY (REGNO (x));
3746 struct qty_table_elem *x_ent = &qty_table[x_q];
3748 if (x_ent->const_rtx)
3749 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3752 if (x == 0 || CONSTANT_P (x))
3753 return x;
3755 if (GET_CODE (x) == SUBREG)
3757 enum machine_mode mode = GET_MODE (x);
3758 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3759 rtx new_rtx;
3761 /* See if we previously assigned a constant value to this SUBREG. */
3762 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3763 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3764 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3765 return new_rtx;
3767 /* If we didn't and if doing so makes sense, see if we previously
3768 assigned a constant value to the enclosing word mode SUBREG. */
3769 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3770 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3772 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3773 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3775 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3776 new_rtx = lookup_as_function (y, CONST_INT);
3777 if (new_rtx)
3778 return gen_lowpart (mode, new_rtx);
3782 /* Otherwise see if we already have a constant for the inner REG,
3783 and if that is enough to calculate an equivalent constant for
3784 the subreg. Note that the upper bits of paradoxical subregs
3785 are undefined, so they cannot be said to equal anything. */
3786 if (REG_P (SUBREG_REG (x))
3787 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3788 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3789 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3791 return 0;
3794 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3795 the hash table in case its value was seen before. */
3797 if (MEM_P (x))
3799 struct table_elt *elt;
3801 x = avoid_constant_pool_reference (x);
3802 if (CONSTANT_P (x))
3803 return x;
3805 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3806 if (elt == 0)
3807 return 0;
3809 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3810 if (elt->is_const && CONSTANT_P (elt->exp))
3811 return elt->exp;
3814 return 0;
3817 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3818 "taken" branch.
3820 In certain cases, this can cause us to add an equivalence. For example,
3821 if we are following the taken case of
3822 if (i == 2)
3823 we can add the fact that `i' and '2' are now equivalent.
3825 In any case, we can record that this comparison was passed. If the same
3826 comparison is seen later, we will know its value. */
3828 static void
3829 record_jump_equiv (rtx insn, bool taken)
3831 int cond_known_true;
3832 rtx op0, op1;
3833 rtx set;
3834 enum machine_mode mode, mode0, mode1;
3835 int reversed_nonequality = 0;
3836 enum rtx_code code;
3838 /* Ensure this is the right kind of insn. */
3839 gcc_assert (any_condjump_p (insn));
3841 set = pc_set (insn);
3843 /* See if this jump condition is known true or false. */
3844 if (taken)
3845 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3846 else
3847 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3849 /* Get the type of comparison being done and the operands being compared.
3850 If we had to reverse a non-equality condition, record that fact so we
3851 know that it isn't valid for floating-point. */
3852 code = GET_CODE (XEXP (SET_SRC (set), 0));
3853 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3854 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3856 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3857 if (! cond_known_true)
3859 code = reversed_comparison_code_parts (code, op0, op1, insn);
3861 /* Don't remember if we can't find the inverse. */
3862 if (code == UNKNOWN)
3863 return;
3866 /* The mode is the mode of the non-constant. */
3867 mode = mode0;
3868 if (mode1 != VOIDmode)
3869 mode = mode1;
3871 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3874 /* Yet another form of subreg creation. In this case, we want something in
3875 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3877 static rtx
3878 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3880 enum machine_mode op_mode = GET_MODE (op);
3881 if (op_mode == mode || op_mode == VOIDmode)
3882 return op;
3883 return lowpart_subreg (mode, op, op_mode);
3886 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3887 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3888 Make any useful entries we can with that information. Called from
3889 above function and called recursively. */
3891 static void
3892 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3893 rtx op1, int reversed_nonequality)
3895 unsigned op0_hash, op1_hash;
3896 int op0_in_memory, op1_in_memory;
3897 struct table_elt *op0_elt, *op1_elt;
3899 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3900 we know that they are also equal in the smaller mode (this is also
3901 true for all smaller modes whether or not there is a SUBREG, but
3902 is not worth testing for with no SUBREG). */
3904 /* Note that GET_MODE (op0) may not equal MODE. */
3905 if (code == EQ && paradoxical_subreg_p (op0))
3907 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3908 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3909 if (tem)
3910 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3911 reversed_nonequality);
3914 if (code == EQ && paradoxical_subreg_p (op1))
3916 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3917 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3918 if (tem)
3919 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3920 reversed_nonequality);
3923 /* Similarly, if this is an NE comparison, and either is a SUBREG
3924 making a smaller mode, we know the whole thing is also NE. */
3926 /* Note that GET_MODE (op0) may not equal MODE;
3927 if we test MODE instead, we can get an infinite recursion
3928 alternating between two modes each wider than MODE. */
3930 if (code == NE && GET_CODE (op0) == SUBREG
3931 && subreg_lowpart_p (op0)
3932 && (GET_MODE_SIZE (GET_MODE (op0))
3933 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3935 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3936 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3937 if (tem)
3938 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3939 reversed_nonequality);
3942 if (code == NE && GET_CODE (op1) == SUBREG
3943 && subreg_lowpart_p (op1)
3944 && (GET_MODE_SIZE (GET_MODE (op1))
3945 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3947 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3948 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3949 if (tem)
3950 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3951 reversed_nonequality);
3954 /* Hash both operands. */
3956 do_not_record = 0;
3957 hash_arg_in_memory = 0;
3958 op0_hash = HASH (op0, mode);
3959 op0_in_memory = hash_arg_in_memory;
3961 if (do_not_record)
3962 return;
3964 do_not_record = 0;
3965 hash_arg_in_memory = 0;
3966 op1_hash = HASH (op1, mode);
3967 op1_in_memory = hash_arg_in_memory;
3969 if (do_not_record)
3970 return;
3972 /* Look up both operands. */
3973 op0_elt = lookup (op0, op0_hash, mode);
3974 op1_elt = lookup (op1, op1_hash, mode);
3976 /* If both operands are already equivalent or if they are not in the
3977 table but are identical, do nothing. */
3978 if ((op0_elt != 0 && op1_elt != 0
3979 && op0_elt->first_same_value == op1_elt->first_same_value)
3980 || op0 == op1 || rtx_equal_p (op0, op1))
3981 return;
3983 /* If we aren't setting two things equal all we can do is save this
3984 comparison. Similarly if this is floating-point. In the latter
3985 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
3986 If we record the equality, we might inadvertently delete code
3987 whose intent was to change -0 to +0. */
3989 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
3991 struct qty_table_elem *ent;
3992 int qty;
3994 /* If we reversed a floating-point comparison, if OP0 is not a
3995 register, or if OP1 is neither a register or constant, we can't
3996 do anything. */
3998 if (!REG_P (op1))
3999 op1 = equiv_constant (op1);
4001 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4002 || !REG_P (op0) || op1 == 0)
4003 return;
4005 /* Put OP0 in the hash table if it isn't already. This gives it a
4006 new quantity number. */
4007 if (op0_elt == 0)
4009 if (insert_regs (op0, NULL, 0))
4011 rehash_using_reg (op0);
4012 op0_hash = HASH (op0, mode);
4014 /* If OP0 is contained in OP1, this changes its hash code
4015 as well. Faster to rehash than to check, except
4016 for the simple case of a constant. */
4017 if (! CONSTANT_P (op1))
4018 op1_hash = HASH (op1,mode);
4021 op0_elt = insert (op0, NULL, op0_hash, mode);
4022 op0_elt->in_memory = op0_in_memory;
4025 qty = REG_QTY (REGNO (op0));
4026 ent = &qty_table[qty];
4028 ent->comparison_code = code;
4029 if (REG_P (op1))
4031 /* Look it up again--in case op0 and op1 are the same. */
4032 op1_elt = lookup (op1, op1_hash, mode);
4034 /* Put OP1 in the hash table so it gets a new quantity number. */
4035 if (op1_elt == 0)
4037 if (insert_regs (op1, NULL, 0))
4039 rehash_using_reg (op1);
4040 op1_hash = HASH (op1, mode);
4043 op1_elt = insert (op1, NULL, op1_hash, mode);
4044 op1_elt->in_memory = op1_in_memory;
4047 ent->comparison_const = NULL_RTX;
4048 ent->comparison_qty = REG_QTY (REGNO (op1));
4050 else
4052 ent->comparison_const = op1;
4053 ent->comparison_qty = -1;
4056 return;
4059 /* If either side is still missing an equivalence, make it now,
4060 then merge the equivalences. */
4062 if (op0_elt == 0)
4064 if (insert_regs (op0, NULL, 0))
4066 rehash_using_reg (op0);
4067 op0_hash = HASH (op0, mode);
4070 op0_elt = insert (op0, NULL, op0_hash, mode);
4071 op0_elt->in_memory = op0_in_memory;
4074 if (op1_elt == 0)
4076 if (insert_regs (op1, NULL, 0))
4078 rehash_using_reg (op1);
4079 op1_hash = HASH (op1, mode);
4082 op1_elt = insert (op1, NULL, op1_hash, mode);
4083 op1_elt->in_memory = op1_in_memory;
4086 merge_equiv_classes (op0_elt, op1_elt);
4089 /* CSE processing for one instruction.
4091 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4092 but the few that "leak through" are cleaned up by cse_insn, and complex
4093 addressing modes are often formed here.
4095 The main function is cse_insn, and between here and that function
4096 a couple of helper functions is defined to keep the size of cse_insn
4097 within reasonable proportions.
4099 Data is shared between the main and helper functions via STRUCT SET,
4100 that contains all data related for every set in the instruction that
4101 is being processed.
4103 Note that cse_main processes all sets in the instruction. Most
4104 passes in GCC only process simple SET insns or single_set insns, but
4105 CSE processes insns with multiple sets as well. */
4107 /* Data on one SET contained in the instruction. */
4109 struct set
4111 /* The SET rtx itself. */
4112 rtx rtl;
4113 /* The SET_SRC of the rtx (the original value, if it is changing). */
4114 rtx src;
4115 /* The hash-table element for the SET_SRC of the SET. */
4116 struct table_elt *src_elt;
4117 /* Hash value for the SET_SRC. */
4118 unsigned src_hash;
4119 /* Hash value for the SET_DEST. */
4120 unsigned dest_hash;
4121 /* The SET_DEST, with SUBREG, etc., stripped. */
4122 rtx inner_dest;
4123 /* Nonzero if the SET_SRC is in memory. */
4124 char src_in_memory;
4125 /* Nonzero if the SET_SRC contains something
4126 whose value cannot be predicted and understood. */
4127 char src_volatile;
4128 /* Original machine mode, in case it becomes a CONST_INT.
4129 The size of this field should match the size of the mode
4130 field of struct rtx_def (see rtl.h). */
4131 ENUM_BITFIELD(machine_mode) mode : 8;
4132 /* A constant equivalent for SET_SRC, if any. */
4133 rtx src_const;
4134 /* Hash value of constant equivalent for SET_SRC. */
4135 unsigned src_const_hash;
4136 /* Table entry for constant equivalent for SET_SRC, if any. */
4137 struct table_elt *src_const_elt;
4138 /* Table entry for the destination address. */
4139 struct table_elt *dest_addr_elt;
4142 /* Special handling for (set REG0 REG1) where REG0 is the
4143 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4144 be used in the sequel, so (if easily done) change this insn to
4145 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4146 that computed their value. Then REG1 will become a dead store
4147 and won't cloud the situation for later optimizations.
4149 Do not make this change if REG1 is a hard register, because it will
4150 then be used in the sequel and we may be changing a two-operand insn
4151 into a three-operand insn.
4153 This is the last transformation that cse_insn will try to do. */
4155 static void
4156 try_back_substitute_reg (rtx set, rtx insn)
4158 rtx dest = SET_DEST (set);
4159 rtx src = SET_SRC (set);
4161 if (REG_P (dest)
4162 && REG_P (src) && ! HARD_REGISTER_P (src)
4163 && REGNO_QTY_VALID_P (REGNO (src)))
4165 int src_q = REG_QTY (REGNO (src));
4166 struct qty_table_elem *src_ent = &qty_table[src_q];
4168 if (src_ent->first_reg == REGNO (dest))
4170 /* Scan for the previous nonnote insn, but stop at a basic
4171 block boundary. */
4172 rtx prev = insn;
4173 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4176 prev = PREV_INSN (prev);
4178 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4180 /* Do not swap the registers around if the previous instruction
4181 attaches a REG_EQUIV note to REG1.
4183 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4184 from the pseudo that originally shadowed an incoming argument
4185 to another register. Some uses of REG_EQUIV might rely on it
4186 being attached to REG1 rather than REG2.
4188 This section previously turned the REG_EQUIV into a REG_EQUAL
4189 note. We cannot do that because REG_EQUIV may provide an
4190 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4191 if (NONJUMP_INSN_P (prev)
4192 && GET_CODE (PATTERN (prev)) == SET
4193 && SET_DEST (PATTERN (prev)) == src
4194 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4196 rtx note;
4198 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4199 validate_change (insn, &SET_DEST (set), src, 1);
4200 validate_change (insn, &SET_SRC (set), dest, 1);
4201 apply_change_group ();
4203 /* If INSN has a REG_EQUAL note, and this note mentions
4204 REG0, then we must delete it, because the value in
4205 REG0 has changed. If the note's value is REG1, we must
4206 also delete it because that is now this insn's dest. */
4207 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4208 if (note != 0
4209 && (reg_mentioned_p (dest, XEXP (note, 0))
4210 || rtx_equal_p (src, XEXP (note, 0))))
4211 remove_note (insn, note);
4217 /* Record all the SETs in this instruction into SETS_PTR,
4218 and return the number of recorded sets. */
4219 static int
4220 find_sets_in_insn (rtx insn, struct set **psets)
4222 struct set *sets = *psets;
4223 int n_sets = 0;
4224 rtx x = PATTERN (insn);
4226 if (GET_CODE (x) == SET)
4228 /* Ignore SETs that are unconditional jumps.
4229 They never need cse processing, so this does not hurt.
4230 The reason is not efficiency but rather
4231 so that we can test at the end for instructions
4232 that have been simplified to unconditional jumps
4233 and not be misled by unchanged instructions
4234 that were unconditional jumps to begin with. */
4235 if (SET_DEST (x) == pc_rtx
4236 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4238 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4239 The hard function value register is used only once, to copy to
4240 someplace else, so it isn't worth cse'ing. */
4241 else if (GET_CODE (SET_SRC (x)) == CALL)
4243 else
4244 sets[n_sets++].rtl = x;
4246 else if (GET_CODE (x) == PARALLEL)
4248 int i, lim = XVECLEN (x, 0);
4250 /* Go over the epressions of the PARALLEL in forward order, to
4251 put them in the same order in the SETS array. */
4252 for (i = 0; i < lim; i++)
4254 rtx y = XVECEXP (x, 0, i);
4255 if (GET_CODE (y) == SET)
4257 /* As above, we ignore unconditional jumps and call-insns and
4258 ignore the result of apply_change_group. */
4259 if (SET_DEST (y) == pc_rtx
4260 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4262 else if (GET_CODE (SET_SRC (y)) == CALL)
4264 else
4265 sets[n_sets++].rtl = y;
4270 return n_sets;
4273 /* Where possible, substitute every register reference in the N_SETS
4274 number of SETS in INSN with the the canonical register.
4276 Register canonicalization propagatest the earliest register (i.e.
4277 one that is set before INSN) with the same value. This is a very
4278 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4279 to RTL. For instance, a CONST for an address is usually expanded
4280 multiple times to loads into different registers, thus creating many
4281 subexpressions of the form:
4283 (set (reg1) (some_const))
4284 (set (mem (... reg1 ...) (thing)))
4285 (set (reg2) (some_const))
4286 (set (mem (... reg2 ...) (thing)))
4288 After canonicalizing, the code takes the following form:
4290 (set (reg1) (some_const))
4291 (set (mem (... reg1 ...) (thing)))
4292 (set (reg2) (some_const))
4293 (set (mem (... reg1 ...) (thing)))
4295 The set to reg2 is now trivially dead, and the memory reference (or
4296 address, or whatever) may be a candidate for further CSEing.
4298 In this function, the result of apply_change_group can be ignored;
4299 see canon_reg. */
4301 static void
4302 canonicalize_insn (rtx insn, struct set **psets, int n_sets)
4304 struct set *sets = *psets;
4305 rtx tem;
4306 rtx x = PATTERN (insn);
4307 int i;
4309 if (CALL_P (insn))
4311 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4312 if (GET_CODE (XEXP (tem, 0)) != SET)
4313 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4316 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4318 canon_reg (SET_SRC (x), insn);
4319 apply_change_group ();
4320 fold_rtx (SET_SRC (x), insn);
4322 else if (GET_CODE (x) == CLOBBER)
4324 /* If we clobber memory, canon the address.
4325 This does nothing when a register is clobbered
4326 because we have already invalidated the reg. */
4327 if (MEM_P (XEXP (x, 0)))
4328 canon_reg (XEXP (x, 0), insn);
4330 else if (GET_CODE (x) == USE
4331 && ! (REG_P (XEXP (x, 0))
4332 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4333 /* Canonicalize a USE of a pseudo register or memory location. */
4334 canon_reg (x, insn);
4335 else if (GET_CODE (x) == ASM_OPERANDS)
4337 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4339 rtx input = ASM_OPERANDS_INPUT (x, i);
4340 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4342 input = canon_reg (input, insn);
4343 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4347 else if (GET_CODE (x) == CALL)
4349 canon_reg (x, insn);
4350 apply_change_group ();
4351 fold_rtx (x, insn);
4353 else if (DEBUG_INSN_P (insn))
4354 canon_reg (PATTERN (insn), insn);
4355 else if (GET_CODE (x) == PARALLEL)
4357 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4359 rtx y = XVECEXP (x, 0, i);
4360 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4362 canon_reg (SET_SRC (y), insn);
4363 apply_change_group ();
4364 fold_rtx (SET_SRC (y), insn);
4366 else if (GET_CODE (y) == CLOBBER)
4368 if (MEM_P (XEXP (y, 0)))
4369 canon_reg (XEXP (y, 0), insn);
4371 else if (GET_CODE (y) == USE
4372 && ! (REG_P (XEXP (y, 0))
4373 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4374 canon_reg (y, insn);
4375 else if (GET_CODE (y) == CALL)
4377 canon_reg (y, insn);
4378 apply_change_group ();
4379 fold_rtx (y, insn);
4384 if (n_sets == 1 && REG_NOTES (insn) != 0
4385 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4387 /* We potentially will process this insn many times. Therefore,
4388 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4389 unique set in INSN.
4391 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4392 because cse_insn handles those specially. */
4393 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4394 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4395 remove_note (insn, tem);
4396 else
4398 canon_reg (XEXP (tem, 0), insn);
4399 apply_change_group ();
4400 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4401 df_notes_rescan (insn);
4405 /* Canonicalize sources and addresses of destinations.
4406 We do this in a separate pass to avoid problems when a MATCH_DUP is
4407 present in the insn pattern. In that case, we want to ensure that
4408 we don't break the duplicate nature of the pattern. So we will replace
4409 both operands at the same time. Otherwise, we would fail to find an
4410 equivalent substitution in the loop calling validate_change below.
4412 We used to suppress canonicalization of DEST if it appears in SRC,
4413 but we don't do this any more. */
4415 for (i = 0; i < n_sets; i++)
4417 rtx dest = SET_DEST (sets[i].rtl);
4418 rtx src = SET_SRC (sets[i].rtl);
4419 rtx new_rtx = canon_reg (src, insn);
4421 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4423 if (GET_CODE (dest) == ZERO_EXTRACT)
4425 validate_change (insn, &XEXP (dest, 1),
4426 canon_reg (XEXP (dest, 1), insn), 1);
4427 validate_change (insn, &XEXP (dest, 2),
4428 canon_reg (XEXP (dest, 2), insn), 1);
4431 while (GET_CODE (dest) == SUBREG
4432 || GET_CODE (dest) == ZERO_EXTRACT
4433 || GET_CODE (dest) == STRICT_LOW_PART)
4434 dest = XEXP (dest, 0);
4436 if (MEM_P (dest))
4437 canon_reg (dest, insn);
4440 /* Now that we have done all the replacements, we can apply the change
4441 group and see if they all work. Note that this will cause some
4442 canonicalizations that would have worked individually not to be applied
4443 because some other canonicalization didn't work, but this should not
4444 occur often.
4446 The result of apply_change_group can be ignored; see canon_reg. */
4448 apply_change_group ();
4451 /* Main function of CSE.
4452 First simplify sources and addresses of all assignments
4453 in the instruction, using previously-computed equivalents values.
4454 Then install the new sources and destinations in the table
4455 of available values. */
4457 static void
4458 cse_insn (rtx insn)
4460 rtx x = PATTERN (insn);
4461 int i;
4462 rtx tem;
4463 int n_sets = 0;
4465 rtx src_eqv = 0;
4466 struct table_elt *src_eqv_elt = 0;
4467 int src_eqv_volatile = 0;
4468 int src_eqv_in_memory = 0;
4469 unsigned src_eqv_hash = 0;
4471 struct set *sets = (struct set *) 0;
4473 if (GET_CODE (x) == SET)
4474 sets = XALLOCA (struct set);
4475 else if (GET_CODE (x) == PARALLEL)
4476 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4478 this_insn = insn;
4479 #ifdef HAVE_cc0
4480 /* Records what this insn does to set CC0. */
4481 this_insn_cc0 = 0;
4482 this_insn_cc0_mode = VOIDmode;
4483 #endif
4485 /* Find all regs explicitly clobbered in this insn,
4486 to ensure they are not replaced with any other regs
4487 elsewhere in this insn. */
4488 invalidate_from_sets_and_clobbers (insn);
4490 /* Record all the SETs in this instruction. */
4491 n_sets = find_sets_in_insn (insn, &sets);
4493 /* Substitute the canonical register where possible. */
4494 canonicalize_insn (insn, &sets, n_sets);
4496 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4497 if different, or if the DEST is a STRICT_LOW_PART. The latter condition
4498 is necessary because SRC_EQV is handled specially for this case, and if
4499 it isn't set, then there will be no equivalence for the destination. */
4500 if (n_sets == 1 && REG_NOTES (insn) != 0
4501 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4502 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4503 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4504 src_eqv = copy_rtx (XEXP (tem, 0));
4506 /* Set sets[i].src_elt to the class each source belongs to.
4507 Detect assignments from or to volatile things
4508 and set set[i] to zero so they will be ignored
4509 in the rest of this function.
4511 Nothing in this loop changes the hash table or the register chains. */
4513 for (i = 0; i < n_sets; i++)
4515 bool repeat = false;
4516 rtx src, dest;
4517 rtx src_folded;
4518 struct table_elt *elt = 0, *p;
4519 enum machine_mode mode;
4520 rtx src_eqv_here;
4521 rtx src_const = 0;
4522 rtx src_related = 0;
4523 bool src_related_is_const_anchor = false;
4524 struct table_elt *src_const_elt = 0;
4525 int src_cost = MAX_COST;
4526 int src_eqv_cost = MAX_COST;
4527 int src_folded_cost = MAX_COST;
4528 int src_related_cost = MAX_COST;
4529 int src_elt_cost = MAX_COST;
4530 int src_regcost = MAX_COST;
4531 int src_eqv_regcost = MAX_COST;
4532 int src_folded_regcost = MAX_COST;
4533 int src_related_regcost = MAX_COST;
4534 int src_elt_regcost = MAX_COST;
4535 /* Set nonzero if we need to call force_const_mem on with the
4536 contents of src_folded before using it. */
4537 int src_folded_force_flag = 0;
4539 dest = SET_DEST (sets[i].rtl);
4540 src = SET_SRC (sets[i].rtl);
4542 /* If SRC is a constant that has no machine mode,
4543 hash it with the destination's machine mode.
4544 This way we can keep different modes separate. */
4546 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4547 sets[i].mode = mode;
4549 if (src_eqv)
4551 enum machine_mode eqvmode = mode;
4552 if (GET_CODE (dest) == STRICT_LOW_PART)
4553 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4554 do_not_record = 0;
4555 hash_arg_in_memory = 0;
4556 src_eqv_hash = HASH (src_eqv, eqvmode);
4558 /* Find the equivalence class for the equivalent expression. */
4560 if (!do_not_record)
4561 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4563 src_eqv_volatile = do_not_record;
4564 src_eqv_in_memory = hash_arg_in_memory;
4567 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4568 value of the INNER register, not the destination. So it is not
4569 a valid substitution for the source. But save it for later. */
4570 if (GET_CODE (dest) == STRICT_LOW_PART)
4571 src_eqv_here = 0;
4572 else
4573 src_eqv_here = src_eqv;
4575 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4576 simplified result, which may not necessarily be valid. */
4577 src_folded = fold_rtx (src, insn);
4579 #if 0
4580 /* ??? This caused bad code to be generated for the m68k port with -O2.
4581 Suppose src is (CONST_INT -1), and that after truncation src_folded
4582 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4583 At the end we will add src and src_const to the same equivalence
4584 class. We now have 3 and -1 on the same equivalence class. This
4585 causes later instructions to be mis-optimized. */
4586 /* If storing a constant in a bitfield, pre-truncate the constant
4587 so we will be able to record it later. */
4588 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4590 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4592 if (CONST_INT_P (src)
4593 && CONST_INT_P (width)
4594 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4595 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4596 src_folded
4597 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4598 << INTVAL (width)) - 1));
4600 #endif
4602 /* Compute SRC's hash code, and also notice if it
4603 should not be recorded at all. In that case,
4604 prevent any further processing of this assignment. */
4605 do_not_record = 0;
4606 hash_arg_in_memory = 0;
4608 sets[i].src = src;
4609 sets[i].src_hash = HASH (src, mode);
4610 sets[i].src_volatile = do_not_record;
4611 sets[i].src_in_memory = hash_arg_in_memory;
4613 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4614 a pseudo, do not record SRC. Using SRC as a replacement for
4615 anything else will be incorrect in that situation. Note that
4616 this usually occurs only for stack slots, in which case all the
4617 RTL would be referring to SRC, so we don't lose any optimization
4618 opportunities by not having SRC in the hash table. */
4620 if (MEM_P (src)
4621 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4622 && REG_P (dest)
4623 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4624 sets[i].src_volatile = 1;
4626 #if 0
4627 /* It is no longer clear why we used to do this, but it doesn't
4628 appear to still be needed. So let's try without it since this
4629 code hurts cse'ing widened ops. */
4630 /* If source is a paradoxical subreg (such as QI treated as an SI),
4631 treat it as volatile. It may do the work of an SI in one context
4632 where the extra bits are not being used, but cannot replace an SI
4633 in general. */
4634 if (paradoxical_subreg_p (src))
4635 sets[i].src_volatile = 1;
4636 #endif
4638 /* Locate all possible equivalent forms for SRC. Try to replace
4639 SRC in the insn with each cheaper equivalent.
4641 We have the following types of equivalents: SRC itself, a folded
4642 version, a value given in a REG_EQUAL note, or a value related
4643 to a constant.
4645 Each of these equivalents may be part of an additional class
4646 of equivalents (if more than one is in the table, they must be in
4647 the same class; we check for this).
4649 If the source is volatile, we don't do any table lookups.
4651 We note any constant equivalent for possible later use in a
4652 REG_NOTE. */
4654 if (!sets[i].src_volatile)
4655 elt = lookup (src, sets[i].src_hash, mode);
4657 sets[i].src_elt = elt;
4659 if (elt && src_eqv_here && src_eqv_elt)
4661 if (elt->first_same_value != src_eqv_elt->first_same_value)
4663 /* The REG_EQUAL is indicating that two formerly distinct
4664 classes are now equivalent. So merge them. */
4665 merge_equiv_classes (elt, src_eqv_elt);
4666 src_eqv_hash = HASH (src_eqv, elt->mode);
4667 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4670 src_eqv_here = 0;
4673 else if (src_eqv_elt)
4674 elt = src_eqv_elt;
4676 /* Try to find a constant somewhere and record it in `src_const'.
4677 Record its table element, if any, in `src_const_elt'. Look in
4678 any known equivalences first. (If the constant is not in the
4679 table, also set `sets[i].src_const_hash'). */
4680 if (elt)
4681 for (p = elt->first_same_value; p; p = p->next_same_value)
4682 if (p->is_const)
4684 src_const = p->exp;
4685 src_const_elt = elt;
4686 break;
4689 if (src_const == 0
4690 && (CONSTANT_P (src_folded)
4691 /* Consider (minus (label_ref L1) (label_ref L2)) as
4692 "constant" here so we will record it. This allows us
4693 to fold switch statements when an ADDR_DIFF_VEC is used. */
4694 || (GET_CODE (src_folded) == MINUS
4695 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4696 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4697 src_const = src_folded, src_const_elt = elt;
4698 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4699 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4701 /* If we don't know if the constant is in the table, get its
4702 hash code and look it up. */
4703 if (src_const && src_const_elt == 0)
4705 sets[i].src_const_hash = HASH (src_const, mode);
4706 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4709 sets[i].src_const = src_const;
4710 sets[i].src_const_elt = src_const_elt;
4712 /* If the constant and our source are both in the table, mark them as
4713 equivalent. Otherwise, if a constant is in the table but the source
4714 isn't, set ELT to it. */
4715 if (src_const_elt && elt
4716 && src_const_elt->first_same_value != elt->first_same_value)
4717 merge_equiv_classes (elt, src_const_elt);
4718 else if (src_const_elt && elt == 0)
4719 elt = src_const_elt;
4721 /* See if there is a register linearly related to a constant
4722 equivalent of SRC. */
4723 if (src_const
4724 && (GET_CODE (src_const) == CONST
4725 || (src_const_elt && src_const_elt->related_value != 0)))
4727 src_related = use_related_value (src_const, src_const_elt);
4728 if (src_related)
4730 struct table_elt *src_related_elt
4731 = lookup (src_related, HASH (src_related, mode), mode);
4732 if (src_related_elt && elt)
4734 if (elt->first_same_value
4735 != src_related_elt->first_same_value)
4736 /* This can occur when we previously saw a CONST
4737 involving a SYMBOL_REF and then see the SYMBOL_REF
4738 twice. Merge the involved classes. */
4739 merge_equiv_classes (elt, src_related_elt);
4741 src_related = 0;
4742 src_related_elt = 0;
4744 else if (src_related_elt && elt == 0)
4745 elt = src_related_elt;
4749 /* See if we have a CONST_INT that is already in a register in a
4750 wider mode. */
4752 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4753 && GET_MODE_CLASS (mode) == MODE_INT
4754 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4756 enum machine_mode wider_mode;
4758 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4759 wider_mode != VOIDmode
4760 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4761 && src_related == 0;
4762 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4764 struct table_elt *const_elt
4765 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4767 if (const_elt == 0)
4768 continue;
4770 for (const_elt = const_elt->first_same_value;
4771 const_elt; const_elt = const_elt->next_same_value)
4772 if (REG_P (const_elt->exp))
4774 src_related = gen_lowpart (mode, const_elt->exp);
4775 break;
4780 /* Another possibility is that we have an AND with a constant in
4781 a mode narrower than a word. If so, it might have been generated
4782 as part of an "if" which would narrow the AND. If we already
4783 have done the AND in a wider mode, we can use a SUBREG of that
4784 value. */
4786 if (flag_expensive_optimizations && ! src_related
4787 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4788 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4790 enum machine_mode tmode;
4791 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4793 for (tmode = GET_MODE_WIDER_MODE (mode);
4794 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4795 tmode = GET_MODE_WIDER_MODE (tmode))
4797 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4798 struct table_elt *larger_elt;
4800 if (inner)
4802 PUT_MODE (new_and, tmode);
4803 XEXP (new_and, 0) = inner;
4804 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4805 if (larger_elt == 0)
4806 continue;
4808 for (larger_elt = larger_elt->first_same_value;
4809 larger_elt; larger_elt = larger_elt->next_same_value)
4810 if (REG_P (larger_elt->exp))
4812 src_related
4813 = gen_lowpart (mode, larger_elt->exp);
4814 break;
4817 if (src_related)
4818 break;
4823 #ifdef LOAD_EXTEND_OP
4824 /* See if a MEM has already been loaded with a widening operation;
4825 if it has, we can use a subreg of that. Many CISC machines
4826 also have such operations, but this is only likely to be
4827 beneficial on these machines. */
4829 if (flag_expensive_optimizations && src_related == 0
4830 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4831 && GET_MODE_CLASS (mode) == MODE_INT
4832 && MEM_P (src) && ! do_not_record
4833 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4835 struct rtx_def memory_extend_buf;
4836 rtx memory_extend_rtx = &memory_extend_buf;
4837 enum machine_mode tmode;
4839 /* Set what we are trying to extend and the operation it might
4840 have been extended with. */
4841 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4842 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4843 XEXP (memory_extend_rtx, 0) = src;
4845 for (tmode = GET_MODE_WIDER_MODE (mode);
4846 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4847 tmode = GET_MODE_WIDER_MODE (tmode))
4849 struct table_elt *larger_elt;
4851 PUT_MODE (memory_extend_rtx, tmode);
4852 larger_elt = lookup (memory_extend_rtx,
4853 HASH (memory_extend_rtx, tmode), tmode);
4854 if (larger_elt == 0)
4855 continue;
4857 for (larger_elt = larger_elt->first_same_value;
4858 larger_elt; larger_elt = larger_elt->next_same_value)
4859 if (REG_P (larger_elt->exp))
4861 src_related = gen_lowpart (mode, larger_elt->exp);
4862 break;
4865 if (src_related)
4866 break;
4869 #endif /* LOAD_EXTEND_OP */
4871 /* Try to express the constant using a register+offset expression
4872 derived from a constant anchor. */
4874 if (targetm.const_anchor
4875 && !src_related
4876 && src_const
4877 && GET_CODE (src_const) == CONST_INT)
4879 src_related = try_const_anchors (src_const, mode);
4880 src_related_is_const_anchor = src_related != NULL_RTX;
4884 if (src == src_folded)
4885 src_folded = 0;
4887 /* At this point, ELT, if nonzero, points to a class of expressions
4888 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4889 and SRC_RELATED, if nonzero, each contain additional equivalent
4890 expressions. Prune these latter expressions by deleting expressions
4891 already in the equivalence class.
4893 Check for an equivalent identical to the destination. If found,
4894 this is the preferred equivalent since it will likely lead to
4895 elimination of the insn. Indicate this by placing it in
4896 `src_related'. */
4898 if (elt)
4899 elt = elt->first_same_value;
4900 for (p = elt; p; p = p->next_same_value)
4902 enum rtx_code code = GET_CODE (p->exp);
4904 /* If the expression is not valid, ignore it. Then we do not
4905 have to check for validity below. In most cases, we can use
4906 `rtx_equal_p', since canonicalization has already been done. */
4907 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4908 continue;
4910 /* Also skip paradoxical subregs, unless that's what we're
4911 looking for. */
4912 if (paradoxical_subreg_p (p->exp)
4913 && ! (src != 0
4914 && GET_CODE (src) == SUBREG
4915 && GET_MODE (src) == GET_MODE (p->exp)
4916 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4917 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4918 continue;
4920 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4921 src = 0;
4922 else if (src_folded && GET_CODE (src_folded) == code
4923 && rtx_equal_p (src_folded, p->exp))
4924 src_folded = 0;
4925 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4926 && rtx_equal_p (src_eqv_here, p->exp))
4927 src_eqv_here = 0;
4928 else if (src_related && GET_CODE (src_related) == code
4929 && rtx_equal_p (src_related, p->exp))
4930 src_related = 0;
4932 /* This is the same as the destination of the insns, we want
4933 to prefer it. Copy it to src_related. The code below will
4934 then give it a negative cost. */
4935 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4936 src_related = dest;
4939 /* Find the cheapest valid equivalent, trying all the available
4940 possibilities. Prefer items not in the hash table to ones
4941 that are when they are equal cost. Note that we can never
4942 worsen an insn as the current contents will also succeed.
4943 If we find an equivalent identical to the destination, use it as best,
4944 since this insn will probably be eliminated in that case. */
4945 if (src)
4947 if (rtx_equal_p (src, dest))
4948 src_cost = src_regcost = -1;
4949 else
4951 src_cost = COST (src);
4952 src_regcost = approx_reg_cost (src);
4956 if (src_eqv_here)
4958 if (rtx_equal_p (src_eqv_here, dest))
4959 src_eqv_cost = src_eqv_regcost = -1;
4960 else
4962 src_eqv_cost = COST (src_eqv_here);
4963 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4967 if (src_folded)
4969 if (rtx_equal_p (src_folded, dest))
4970 src_folded_cost = src_folded_regcost = -1;
4971 else
4973 src_folded_cost = COST (src_folded);
4974 src_folded_regcost = approx_reg_cost (src_folded);
4978 if (src_related)
4980 if (rtx_equal_p (src_related, dest))
4981 src_related_cost = src_related_regcost = -1;
4982 else
4984 src_related_cost = COST (src_related);
4985 src_related_regcost = approx_reg_cost (src_related);
4987 /* If a const-anchor is used to synthesize a constant that
4988 normally requires multiple instructions then slightly prefer
4989 it over the original sequence. These instructions are likely
4990 to become redundant now. We can't compare against the cost
4991 of src_eqv_here because, on MIPS for example, multi-insn
4992 constants have zero cost; they are assumed to be hoisted from
4993 loops. */
4994 if (src_related_is_const_anchor
4995 && src_related_cost == src_cost
4996 && src_eqv_here)
4997 src_related_cost--;
5001 /* If this was an indirect jump insn, a known label will really be
5002 cheaper even though it looks more expensive. */
5003 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5004 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5006 /* Terminate loop when replacement made. This must terminate since
5007 the current contents will be tested and will always be valid. */
5008 while (1)
5010 rtx trial;
5012 /* Skip invalid entries. */
5013 while (elt && !REG_P (elt->exp)
5014 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5015 elt = elt->next_same_value;
5017 /* A paradoxical subreg would be bad here: it'll be the right
5018 size, but later may be adjusted so that the upper bits aren't
5019 what we want. So reject it. */
5020 if (elt != 0
5021 && paradoxical_subreg_p (elt->exp)
5022 /* It is okay, though, if the rtx we're trying to match
5023 will ignore any of the bits we can't predict. */
5024 && ! (src != 0
5025 && GET_CODE (src) == SUBREG
5026 && GET_MODE (src) == GET_MODE (elt->exp)
5027 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5028 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5030 elt = elt->next_same_value;
5031 continue;
5034 if (elt)
5036 src_elt_cost = elt->cost;
5037 src_elt_regcost = elt->regcost;
5040 /* Find cheapest and skip it for the next time. For items
5041 of equal cost, use this order:
5042 src_folded, src, src_eqv, src_related and hash table entry. */
5043 if (src_folded
5044 && preferable (src_folded_cost, src_folded_regcost,
5045 src_cost, src_regcost) <= 0
5046 && preferable (src_folded_cost, src_folded_regcost,
5047 src_eqv_cost, src_eqv_regcost) <= 0
5048 && preferable (src_folded_cost, src_folded_regcost,
5049 src_related_cost, src_related_regcost) <= 0
5050 && preferable (src_folded_cost, src_folded_regcost,
5051 src_elt_cost, src_elt_regcost) <= 0)
5053 trial = src_folded, src_folded_cost = MAX_COST;
5054 if (src_folded_force_flag)
5056 rtx forced = force_const_mem (mode, trial);
5057 if (forced)
5058 trial = forced;
5061 else if (src
5062 && preferable (src_cost, src_regcost,
5063 src_eqv_cost, src_eqv_regcost) <= 0
5064 && preferable (src_cost, src_regcost,
5065 src_related_cost, src_related_regcost) <= 0
5066 && preferable (src_cost, src_regcost,
5067 src_elt_cost, src_elt_regcost) <= 0)
5068 trial = src, src_cost = MAX_COST;
5069 else if (src_eqv_here
5070 && preferable (src_eqv_cost, src_eqv_regcost,
5071 src_related_cost, src_related_regcost) <= 0
5072 && preferable (src_eqv_cost, src_eqv_regcost,
5073 src_elt_cost, src_elt_regcost) <= 0)
5074 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5075 else if (src_related
5076 && preferable (src_related_cost, src_related_regcost,
5077 src_elt_cost, src_elt_regcost) <= 0)
5078 trial = src_related, src_related_cost = MAX_COST;
5079 else
5081 trial = elt->exp;
5082 elt = elt->next_same_value;
5083 src_elt_cost = MAX_COST;
5086 /* Avoid creation of overlapping memory moves. */
5087 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5089 rtx src, dest;
5091 /* BLKmode moves are not handled by cse anyway. */
5092 if (GET_MODE (trial) == BLKmode)
5093 break;
5095 src = canon_rtx (trial);
5096 dest = canon_rtx (SET_DEST (sets[i].rtl));
5098 if (!MEM_P (src) || !MEM_P (dest)
5099 || !nonoverlapping_memrefs_p (src, dest, false))
5100 break;
5103 /* Try to optimize
5104 (set (reg:M N) (const_int A))
5105 (set (reg:M2 O) (const_int B))
5106 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5107 (reg:M2 O)). */
5108 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5109 && CONST_INT_P (trial)
5110 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5111 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5112 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5113 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5114 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5115 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5116 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5117 <= HOST_BITS_PER_WIDE_INT))
5119 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5120 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5121 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5122 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5123 struct table_elt *dest_elt
5124 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5125 rtx dest_cst = NULL;
5127 if (dest_elt)
5128 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5129 if (p->is_const && CONST_INT_P (p->exp))
5131 dest_cst = p->exp;
5132 break;
5134 if (dest_cst)
5136 HOST_WIDE_INT val = INTVAL (dest_cst);
5137 HOST_WIDE_INT mask;
5138 unsigned int shift;
5139 if (BITS_BIG_ENDIAN)
5140 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5141 - INTVAL (pos) - INTVAL (width);
5142 else
5143 shift = INTVAL (pos);
5144 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5145 mask = ~(HOST_WIDE_INT) 0;
5146 else
5147 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5148 val &= ~(mask << shift);
5149 val |= (INTVAL (trial) & mask) << shift;
5150 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5151 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5152 dest_reg, 1);
5153 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5154 GEN_INT (val), 1);
5155 if (apply_change_group ())
5157 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5158 if (note)
5160 remove_note (insn, note);
5161 df_notes_rescan (insn);
5163 src_eqv = NULL_RTX;
5164 src_eqv_elt = NULL;
5165 src_eqv_volatile = 0;
5166 src_eqv_in_memory = 0;
5167 src_eqv_hash = 0;
5168 repeat = true;
5169 break;
5174 /* We don't normally have an insn matching (set (pc) (pc)), so
5175 check for this separately here. We will delete such an
5176 insn below.
5178 For other cases such as a table jump or conditional jump
5179 where we know the ultimate target, go ahead and replace the
5180 operand. While that may not make a valid insn, we will
5181 reemit the jump below (and also insert any necessary
5182 barriers). */
5183 if (n_sets == 1 && dest == pc_rtx
5184 && (trial == pc_rtx
5185 || (GET_CODE (trial) == LABEL_REF
5186 && ! condjump_p (insn))))
5188 /* Don't substitute non-local labels, this confuses CFG. */
5189 if (GET_CODE (trial) == LABEL_REF
5190 && LABEL_REF_NONLOCAL_P (trial))
5191 continue;
5193 SET_SRC (sets[i].rtl) = trial;
5194 cse_jumps_altered = true;
5195 break;
5198 /* Reject certain invalid forms of CONST that we create. */
5199 else if (CONSTANT_P (trial)
5200 && GET_CODE (trial) == CONST
5201 /* Reject cases that will cause decode_rtx_const to
5202 die. On the alpha when simplifying a switch, we
5203 get (const (truncate (minus (label_ref)
5204 (label_ref)))). */
5205 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5206 /* Likewise on IA-64, except without the
5207 truncate. */
5208 || (GET_CODE (XEXP (trial, 0)) == MINUS
5209 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5210 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5211 /* Do nothing for this case. */
5214 /* Look for a substitution that makes a valid insn. */
5215 else if (validate_unshare_change
5216 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5218 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5220 /* The result of apply_change_group can be ignored; see
5221 canon_reg. */
5223 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5224 apply_change_group ();
5226 break;
5229 /* If we previously found constant pool entries for
5230 constants and this is a constant, try making a
5231 pool entry. Put it in src_folded unless we already have done
5232 this since that is where it likely came from. */
5234 else if (constant_pool_entries_cost
5235 && CONSTANT_P (trial)
5236 && (src_folded == 0
5237 || (!MEM_P (src_folded)
5238 && ! src_folded_force_flag))
5239 && GET_MODE_CLASS (mode) != MODE_CC
5240 && mode != VOIDmode)
5242 src_folded_force_flag = 1;
5243 src_folded = trial;
5244 src_folded_cost = constant_pool_entries_cost;
5245 src_folded_regcost = constant_pool_entries_regcost;
5249 /* If we changed the insn too much, handle this set from scratch. */
5250 if (repeat)
5252 i--;
5253 continue;
5256 src = SET_SRC (sets[i].rtl);
5258 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5259 However, there is an important exception: If both are registers
5260 that are not the head of their equivalence class, replace SET_SRC
5261 with the head of the class. If we do not do this, we will have
5262 both registers live over a portion of the basic block. This way,
5263 their lifetimes will likely abut instead of overlapping. */
5264 if (REG_P (dest)
5265 && REGNO_QTY_VALID_P (REGNO (dest)))
5267 int dest_q = REG_QTY (REGNO (dest));
5268 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5270 if (dest_ent->mode == GET_MODE (dest)
5271 && dest_ent->first_reg != REGNO (dest)
5272 && REG_P (src) && REGNO (src) == REGNO (dest)
5273 /* Don't do this if the original insn had a hard reg as
5274 SET_SRC or SET_DEST. */
5275 && (!REG_P (sets[i].src)
5276 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5277 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5278 /* We can't call canon_reg here because it won't do anything if
5279 SRC is a hard register. */
5281 int src_q = REG_QTY (REGNO (src));
5282 struct qty_table_elem *src_ent = &qty_table[src_q];
5283 int first = src_ent->first_reg;
5284 rtx new_src
5285 = (first >= FIRST_PSEUDO_REGISTER
5286 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5288 /* We must use validate-change even for this, because this
5289 might be a special no-op instruction, suitable only to
5290 tag notes onto. */
5291 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5293 src = new_src;
5294 /* If we had a constant that is cheaper than what we are now
5295 setting SRC to, use that constant. We ignored it when we
5296 thought we could make this into a no-op. */
5297 if (src_const && COST (src_const) < COST (src)
5298 && validate_change (insn, &SET_SRC (sets[i].rtl),
5299 src_const, 0))
5300 src = src_const;
5305 /* If we made a change, recompute SRC values. */
5306 if (src != sets[i].src)
5308 do_not_record = 0;
5309 hash_arg_in_memory = 0;
5310 sets[i].src = src;
5311 sets[i].src_hash = HASH (src, mode);
5312 sets[i].src_volatile = do_not_record;
5313 sets[i].src_in_memory = hash_arg_in_memory;
5314 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5317 /* If this is a single SET, we are setting a register, and we have an
5318 equivalent constant, we want to add a REG_NOTE. We don't want
5319 to write a REG_EQUAL note for a constant pseudo since verifying that
5320 that pseudo hasn't been eliminated is a pain. Such a note also
5321 won't help anything.
5323 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5324 which can be created for a reference to a compile time computable
5325 entry in a jump table. */
5327 if (n_sets == 1 && src_const && REG_P (dest)
5328 && !REG_P (src_const)
5329 && ! (GET_CODE (src_const) == CONST
5330 && GET_CODE (XEXP (src_const, 0)) == MINUS
5331 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5332 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5334 /* We only want a REG_EQUAL note if src_const != src. */
5335 if (! rtx_equal_p (src, src_const))
5337 /* Make sure that the rtx is not shared. */
5338 src_const = copy_rtx (src_const);
5340 /* Record the actual constant value in a REG_EQUAL note,
5341 making a new one if one does not already exist. */
5342 set_unique_reg_note (insn, REG_EQUAL, src_const);
5343 df_notes_rescan (insn);
5347 /* Now deal with the destination. */
5348 do_not_record = 0;
5350 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5351 while (GET_CODE (dest) == SUBREG
5352 || GET_CODE (dest) == ZERO_EXTRACT
5353 || GET_CODE (dest) == STRICT_LOW_PART)
5354 dest = XEXP (dest, 0);
5356 sets[i].inner_dest = dest;
5358 if (MEM_P (dest))
5360 #ifdef PUSH_ROUNDING
5361 /* Stack pushes invalidate the stack pointer. */
5362 rtx addr = XEXP (dest, 0);
5363 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5364 && XEXP (addr, 0) == stack_pointer_rtx)
5365 invalidate (stack_pointer_rtx, VOIDmode);
5366 #endif
5367 dest = fold_rtx (dest, insn);
5370 /* Compute the hash code of the destination now,
5371 before the effects of this instruction are recorded,
5372 since the register values used in the address computation
5373 are those before this instruction. */
5374 sets[i].dest_hash = HASH (dest, mode);
5376 /* Don't enter a bit-field in the hash table
5377 because the value in it after the store
5378 may not equal what was stored, due to truncation. */
5380 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5382 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5384 if (src_const != 0 && CONST_INT_P (src_const)
5385 && CONST_INT_P (width)
5386 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5387 && ! (INTVAL (src_const)
5388 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5389 /* Exception: if the value is constant,
5390 and it won't be truncated, record it. */
5392 else
5394 /* This is chosen so that the destination will be invalidated
5395 but no new value will be recorded.
5396 We must invalidate because sometimes constant
5397 values can be recorded for bitfields. */
5398 sets[i].src_elt = 0;
5399 sets[i].src_volatile = 1;
5400 src_eqv = 0;
5401 src_eqv_elt = 0;
5405 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5406 the insn. */
5407 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5409 /* One less use of the label this insn used to jump to. */
5410 delete_insn_and_edges (insn);
5411 cse_jumps_altered = true;
5412 /* No more processing for this set. */
5413 sets[i].rtl = 0;
5416 /* If this SET is now setting PC to a label, we know it used to
5417 be a conditional or computed branch. */
5418 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5419 && !LABEL_REF_NONLOCAL_P (src))
5421 /* We reemit the jump in as many cases as possible just in
5422 case the form of an unconditional jump is significantly
5423 different than a computed jump or conditional jump.
5425 If this insn has multiple sets, then reemitting the
5426 jump is nontrivial. So instead we just force rerecognition
5427 and hope for the best. */
5428 if (n_sets == 1)
5430 rtx new_rtx, note;
5432 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5433 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5434 LABEL_NUSES (XEXP (src, 0))++;
5436 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5437 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5438 if (note)
5440 XEXP (note, 1) = NULL_RTX;
5441 REG_NOTES (new_rtx) = note;
5444 delete_insn_and_edges (insn);
5445 insn = new_rtx;
5447 else
5448 INSN_CODE (insn) = -1;
5450 /* Do not bother deleting any unreachable code, let jump do it. */
5451 cse_jumps_altered = true;
5452 sets[i].rtl = 0;
5455 /* If destination is volatile, invalidate it and then do no further
5456 processing for this assignment. */
5458 else if (do_not_record)
5460 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5461 invalidate (dest, VOIDmode);
5462 else if (MEM_P (dest))
5463 invalidate (dest, VOIDmode);
5464 else if (GET_CODE (dest) == STRICT_LOW_PART
5465 || GET_CODE (dest) == ZERO_EXTRACT)
5466 invalidate (XEXP (dest, 0), GET_MODE (dest));
5467 sets[i].rtl = 0;
5470 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5471 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5473 #ifdef HAVE_cc0
5474 /* If setting CC0, record what it was set to, or a constant, if it
5475 is equivalent to a constant. If it is being set to a floating-point
5476 value, make a COMPARE with the appropriate constant of 0. If we
5477 don't do this, later code can interpret this as a test against
5478 const0_rtx, which can cause problems if we try to put it into an
5479 insn as a floating-point operand. */
5480 if (dest == cc0_rtx)
5482 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5483 this_insn_cc0_mode = mode;
5484 if (FLOAT_MODE_P (mode))
5485 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5486 CONST0_RTX (mode));
5488 #endif
5491 /* Now enter all non-volatile source expressions in the hash table
5492 if they are not already present.
5493 Record their equivalence classes in src_elt.
5494 This way we can insert the corresponding destinations into
5495 the same classes even if the actual sources are no longer in them
5496 (having been invalidated). */
5498 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5499 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5501 struct table_elt *elt;
5502 struct table_elt *classp = sets[0].src_elt;
5503 rtx dest = SET_DEST (sets[0].rtl);
5504 enum machine_mode eqvmode = GET_MODE (dest);
5506 if (GET_CODE (dest) == STRICT_LOW_PART)
5508 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5509 classp = 0;
5511 if (insert_regs (src_eqv, classp, 0))
5513 rehash_using_reg (src_eqv);
5514 src_eqv_hash = HASH (src_eqv, eqvmode);
5516 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5517 elt->in_memory = src_eqv_in_memory;
5518 src_eqv_elt = elt;
5520 /* Check to see if src_eqv_elt is the same as a set source which
5521 does not yet have an elt, and if so set the elt of the set source
5522 to src_eqv_elt. */
5523 for (i = 0; i < n_sets; i++)
5524 if (sets[i].rtl && sets[i].src_elt == 0
5525 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5526 sets[i].src_elt = src_eqv_elt;
5529 for (i = 0; i < n_sets; i++)
5530 if (sets[i].rtl && ! sets[i].src_volatile
5531 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5533 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5535 /* REG_EQUAL in setting a STRICT_LOW_PART
5536 gives an equivalent for the entire destination register,
5537 not just for the subreg being stored in now.
5538 This is a more interesting equivalence, so we arrange later
5539 to treat the entire reg as the destination. */
5540 sets[i].src_elt = src_eqv_elt;
5541 sets[i].src_hash = src_eqv_hash;
5543 else
5545 /* Insert source and constant equivalent into hash table, if not
5546 already present. */
5547 struct table_elt *classp = src_eqv_elt;
5548 rtx src = sets[i].src;
5549 rtx dest = SET_DEST (sets[i].rtl);
5550 enum machine_mode mode
5551 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5553 /* It's possible that we have a source value known to be
5554 constant but don't have a REG_EQUAL note on the insn.
5555 Lack of a note will mean src_eqv_elt will be NULL. This
5556 can happen where we've generated a SUBREG to access a
5557 CONST_INT that is already in a register in a wider mode.
5558 Ensure that the source expression is put in the proper
5559 constant class. */
5560 if (!classp)
5561 classp = sets[i].src_const_elt;
5563 if (sets[i].src_elt == 0)
5565 struct table_elt *elt;
5567 /* Note that these insert_regs calls cannot remove
5568 any of the src_elt's, because they would have failed to
5569 match if not still valid. */
5570 if (insert_regs (src, classp, 0))
5572 rehash_using_reg (src);
5573 sets[i].src_hash = HASH (src, mode);
5575 elt = insert (src, classp, sets[i].src_hash, mode);
5576 elt->in_memory = sets[i].src_in_memory;
5577 sets[i].src_elt = classp = elt;
5579 if (sets[i].src_const && sets[i].src_const_elt == 0
5580 && src != sets[i].src_const
5581 && ! rtx_equal_p (sets[i].src_const, src))
5582 sets[i].src_elt = insert (sets[i].src_const, classp,
5583 sets[i].src_const_hash, mode);
5586 else if (sets[i].src_elt == 0)
5587 /* If we did not insert the source into the hash table (e.g., it was
5588 volatile), note the equivalence class for the REG_EQUAL value, if any,
5589 so that the destination goes into that class. */
5590 sets[i].src_elt = src_eqv_elt;
5592 /* Record destination addresses in the hash table. This allows us to
5593 check if they are invalidated by other sets. */
5594 for (i = 0; i < n_sets; i++)
5596 if (sets[i].rtl)
5598 rtx x = sets[i].inner_dest;
5599 struct table_elt *elt;
5600 enum machine_mode mode;
5601 unsigned hash;
5603 if (MEM_P (x))
5605 x = XEXP (x, 0);
5606 mode = GET_MODE (x);
5607 hash = HASH (x, mode);
5608 elt = lookup (x, hash, mode);
5609 if (!elt)
5611 if (insert_regs (x, NULL, 0))
5613 rtx dest = SET_DEST (sets[i].rtl);
5615 rehash_using_reg (x);
5616 hash = HASH (x, mode);
5617 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5619 elt = insert (x, NULL, hash, mode);
5622 sets[i].dest_addr_elt = elt;
5624 else
5625 sets[i].dest_addr_elt = NULL;
5629 invalidate_from_clobbers (insn);
5631 /* Some registers are invalidated by subroutine calls. Memory is
5632 invalidated by non-constant calls. */
5634 if (CALL_P (insn))
5636 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5637 invalidate_memory ();
5638 invalidate_for_call ();
5641 /* Now invalidate everything set by this instruction.
5642 If a SUBREG or other funny destination is being set,
5643 sets[i].rtl is still nonzero, so here we invalidate the reg
5644 a part of which is being set. */
5646 for (i = 0; i < n_sets; i++)
5647 if (sets[i].rtl)
5649 /* We can't use the inner dest, because the mode associated with
5650 a ZERO_EXTRACT is significant. */
5651 rtx dest = SET_DEST (sets[i].rtl);
5653 /* Needed for registers to remove the register from its
5654 previous quantity's chain.
5655 Needed for memory if this is a nonvarying address, unless
5656 we have just done an invalidate_memory that covers even those. */
5657 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5658 invalidate (dest, VOIDmode);
5659 else if (MEM_P (dest))
5660 invalidate (dest, VOIDmode);
5661 else if (GET_CODE (dest) == STRICT_LOW_PART
5662 || GET_CODE (dest) == ZERO_EXTRACT)
5663 invalidate (XEXP (dest, 0), GET_MODE (dest));
5666 /* A volatile ASM invalidates everything. */
5667 if (NONJUMP_INSN_P (insn)
5668 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5669 && MEM_VOLATILE_P (PATTERN (insn)))
5670 flush_hash_table ();
5672 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5673 the regs restored by the longjmp come from a later time
5674 than the setjmp. */
5675 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5677 flush_hash_table ();
5678 goto done;
5681 /* Make sure registers mentioned in destinations
5682 are safe for use in an expression to be inserted.
5683 This removes from the hash table
5684 any invalid entry that refers to one of these registers.
5686 We don't care about the return value from mention_regs because
5687 we are going to hash the SET_DEST values unconditionally. */
5689 for (i = 0; i < n_sets; i++)
5691 if (sets[i].rtl)
5693 rtx x = SET_DEST (sets[i].rtl);
5695 if (!REG_P (x))
5696 mention_regs (x);
5697 else
5699 /* We used to rely on all references to a register becoming
5700 inaccessible when a register changes to a new quantity,
5701 since that changes the hash code. However, that is not
5702 safe, since after HASH_SIZE new quantities we get a
5703 hash 'collision' of a register with its own invalid
5704 entries. And since SUBREGs have been changed not to
5705 change their hash code with the hash code of the register,
5706 it wouldn't work any longer at all. So we have to check
5707 for any invalid references lying around now.
5708 This code is similar to the REG case in mention_regs,
5709 but it knows that reg_tick has been incremented, and
5710 it leaves reg_in_table as -1 . */
5711 unsigned int regno = REGNO (x);
5712 unsigned int endregno = END_REGNO (x);
5713 unsigned int i;
5715 for (i = regno; i < endregno; i++)
5717 if (REG_IN_TABLE (i) >= 0)
5719 remove_invalid_refs (i);
5720 REG_IN_TABLE (i) = -1;
5727 /* We may have just removed some of the src_elt's from the hash table.
5728 So replace each one with the current head of the same class.
5729 Also check if destination addresses have been removed. */
5731 for (i = 0; i < n_sets; i++)
5732 if (sets[i].rtl)
5734 if (sets[i].dest_addr_elt
5735 && sets[i].dest_addr_elt->first_same_value == 0)
5737 /* The elt was removed, which means this destination is not
5738 valid after this instruction. */
5739 sets[i].rtl = NULL_RTX;
5741 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5742 /* If elt was removed, find current head of same class,
5743 or 0 if nothing remains of that class. */
5745 struct table_elt *elt = sets[i].src_elt;
5747 while (elt && elt->prev_same_value)
5748 elt = elt->prev_same_value;
5750 while (elt && elt->first_same_value == 0)
5751 elt = elt->next_same_value;
5752 sets[i].src_elt = elt ? elt->first_same_value : 0;
5756 /* Now insert the destinations into their equivalence classes. */
5758 for (i = 0; i < n_sets; i++)
5759 if (sets[i].rtl)
5761 rtx dest = SET_DEST (sets[i].rtl);
5762 struct table_elt *elt;
5764 /* Don't record value if we are not supposed to risk allocating
5765 floating-point values in registers that might be wider than
5766 memory. */
5767 if ((flag_float_store
5768 && MEM_P (dest)
5769 && FLOAT_MODE_P (GET_MODE (dest)))
5770 /* Don't record BLKmode values, because we don't know the
5771 size of it, and can't be sure that other BLKmode values
5772 have the same or smaller size. */
5773 || GET_MODE (dest) == BLKmode
5774 /* If we didn't put a REG_EQUAL value or a source into the hash
5775 table, there is no point is recording DEST. */
5776 || sets[i].src_elt == 0
5777 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5778 or SIGN_EXTEND, don't record DEST since it can cause
5779 some tracking to be wrong.
5781 ??? Think about this more later. */
5782 || (paradoxical_subreg_p (dest)
5783 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5784 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5785 continue;
5787 /* STRICT_LOW_PART isn't part of the value BEING set,
5788 and neither is the SUBREG inside it.
5789 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5790 if (GET_CODE (dest) == STRICT_LOW_PART)
5791 dest = SUBREG_REG (XEXP (dest, 0));
5793 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5794 /* Registers must also be inserted into chains for quantities. */
5795 if (insert_regs (dest, sets[i].src_elt, 1))
5797 /* If `insert_regs' changes something, the hash code must be
5798 recalculated. */
5799 rehash_using_reg (dest);
5800 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5803 elt = insert (dest, sets[i].src_elt,
5804 sets[i].dest_hash, GET_MODE (dest));
5806 /* If this is a constant, insert the constant anchors with the
5807 equivalent register-offset expressions using register DEST. */
5808 if (targetm.const_anchor
5809 && REG_P (dest)
5810 && SCALAR_INT_MODE_P (GET_MODE (dest))
5811 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5812 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5814 elt->in_memory = (MEM_P (sets[i].inner_dest)
5815 && !MEM_READONLY_P (sets[i].inner_dest));
5817 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5818 narrower than M2, and both M1 and M2 are the same number of words,
5819 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5820 make that equivalence as well.
5822 However, BAR may have equivalences for which gen_lowpart
5823 will produce a simpler value than gen_lowpart applied to
5824 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5825 BAR's equivalences. If we don't get a simplified form, make
5826 the SUBREG. It will not be used in an equivalence, but will
5827 cause two similar assignments to be detected.
5829 Note the loop below will find SUBREG_REG (DEST) since we have
5830 already entered SRC and DEST of the SET in the table. */
5832 if (GET_CODE (dest) == SUBREG
5833 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5834 / UNITS_PER_WORD)
5835 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5836 && (GET_MODE_SIZE (GET_MODE (dest))
5837 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5838 && sets[i].src_elt != 0)
5840 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5841 struct table_elt *elt, *classp = 0;
5843 for (elt = sets[i].src_elt->first_same_value; elt;
5844 elt = elt->next_same_value)
5846 rtx new_src = 0;
5847 unsigned src_hash;
5848 struct table_elt *src_elt;
5849 int byte = 0;
5851 /* Ignore invalid entries. */
5852 if (!REG_P (elt->exp)
5853 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5854 continue;
5856 /* We may have already been playing subreg games. If the
5857 mode is already correct for the destination, use it. */
5858 if (GET_MODE (elt->exp) == new_mode)
5859 new_src = elt->exp;
5860 else
5862 /* Calculate big endian correction for the SUBREG_BYTE.
5863 We have already checked that M1 (GET_MODE (dest))
5864 is not narrower than M2 (new_mode). */
5865 if (BYTES_BIG_ENDIAN)
5866 byte = (GET_MODE_SIZE (GET_MODE (dest))
5867 - GET_MODE_SIZE (new_mode));
5869 new_src = simplify_gen_subreg (new_mode, elt->exp,
5870 GET_MODE (dest), byte);
5873 /* The call to simplify_gen_subreg fails if the value
5874 is VOIDmode, yet we can't do any simplification, e.g.
5875 for EXPR_LISTs denoting function call results.
5876 It is invalid to construct a SUBREG with a VOIDmode
5877 SUBREG_REG, hence a zero new_src means we can't do
5878 this substitution. */
5879 if (! new_src)
5880 continue;
5882 src_hash = HASH (new_src, new_mode);
5883 src_elt = lookup (new_src, src_hash, new_mode);
5885 /* Put the new source in the hash table is if isn't
5886 already. */
5887 if (src_elt == 0)
5889 if (insert_regs (new_src, classp, 0))
5891 rehash_using_reg (new_src);
5892 src_hash = HASH (new_src, new_mode);
5894 src_elt = insert (new_src, classp, src_hash, new_mode);
5895 src_elt->in_memory = elt->in_memory;
5897 else if (classp && classp != src_elt->first_same_value)
5898 /* Show that two things that we've seen before are
5899 actually the same. */
5900 merge_equiv_classes (src_elt, classp);
5902 classp = src_elt->first_same_value;
5903 /* Ignore invalid entries. */
5904 while (classp
5905 && !REG_P (classp->exp)
5906 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5907 classp = classp->next_same_value;
5912 /* Special handling for (set REG0 REG1) where REG0 is the
5913 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5914 be used in the sequel, so (if easily done) change this insn to
5915 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5916 that computed their value. Then REG1 will become a dead store
5917 and won't cloud the situation for later optimizations.
5919 Do not make this change if REG1 is a hard register, because it will
5920 then be used in the sequel and we may be changing a two-operand insn
5921 into a three-operand insn.
5923 Also do not do this if we are operating on a copy of INSN. */
5925 if (n_sets == 1 && sets[0].rtl)
5926 try_back_substitute_reg (sets[0].rtl, insn);
5928 done:;
5931 /* Remove from the hash table all expressions that reference memory. */
5933 static void
5934 invalidate_memory (void)
5936 int i;
5937 struct table_elt *p, *next;
5939 for (i = 0; i < HASH_SIZE; i++)
5940 for (p = table[i]; p; p = next)
5942 next = p->next_same_hash;
5943 if (p->in_memory)
5944 remove_from_table (p, i);
5948 /* Perform invalidation on the basis of everything about INSN,
5949 except for invalidating the actual places that are SET in it.
5950 This includes the places CLOBBERed, and anything that might
5951 alias with something that is SET or CLOBBERed. */
5953 static void
5954 invalidate_from_clobbers (rtx insn)
5956 rtx x = PATTERN (insn);
5958 if (GET_CODE (x) == CLOBBER)
5960 rtx ref = XEXP (x, 0);
5961 if (ref)
5963 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5964 || MEM_P (ref))
5965 invalidate (ref, VOIDmode);
5966 else if (GET_CODE (ref) == STRICT_LOW_PART
5967 || GET_CODE (ref) == ZERO_EXTRACT)
5968 invalidate (XEXP (ref, 0), GET_MODE (ref));
5971 else if (GET_CODE (x) == PARALLEL)
5973 int i;
5974 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5976 rtx y = XVECEXP (x, 0, i);
5977 if (GET_CODE (y) == CLOBBER)
5979 rtx ref = XEXP (y, 0);
5980 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5981 || MEM_P (ref))
5982 invalidate (ref, VOIDmode);
5983 else if (GET_CODE (ref) == STRICT_LOW_PART
5984 || GET_CODE (ref) == ZERO_EXTRACT)
5985 invalidate (XEXP (ref, 0), GET_MODE (ref));
5991 /* Perform invalidation on the basis of everything about INSN.
5992 This includes the places CLOBBERed, and anything that might
5993 alias with something that is SET or CLOBBERed. */
5995 static void
5996 invalidate_from_sets_and_clobbers (rtx insn)
5998 rtx tem;
5999 rtx x = PATTERN (insn);
6001 if (CALL_P (insn))
6003 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6004 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6005 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6008 /* Ensure we invalidate the destination register of a CALL insn.
6009 This is necessary for machines where this register is a fixed_reg,
6010 because no other code would invalidate it. */
6011 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6012 invalidate (SET_DEST (x), VOIDmode);
6014 else if (GET_CODE (x) == PARALLEL)
6016 int i;
6018 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6020 rtx y = XVECEXP (x, 0, i);
6021 if (GET_CODE (y) == CLOBBER)
6023 rtx clobbered = XEXP (y, 0);
6025 if (REG_P (clobbered)
6026 || GET_CODE (clobbered) == SUBREG)
6027 invalidate (clobbered, VOIDmode);
6028 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6029 || GET_CODE (clobbered) == ZERO_EXTRACT)
6030 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6032 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6033 invalidate (SET_DEST (y), VOIDmode);
6038 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6039 and replace any registers in them with either an equivalent constant
6040 or the canonical form of the register. If we are inside an address,
6041 only do this if the address remains valid.
6043 OBJECT is 0 except when within a MEM in which case it is the MEM.
6045 Return the replacement for X. */
6047 static rtx
6048 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6050 enum rtx_code code = GET_CODE (x);
6051 const char *fmt = GET_RTX_FORMAT (code);
6052 int i;
6054 switch (code)
6056 case CONST_INT:
6057 case CONST:
6058 case SYMBOL_REF:
6059 case LABEL_REF:
6060 case CONST_DOUBLE:
6061 case CONST_FIXED:
6062 case CONST_VECTOR:
6063 case PC:
6064 case CC0:
6065 case LO_SUM:
6066 return x;
6068 case MEM:
6069 validate_change (x, &XEXP (x, 0),
6070 cse_process_notes (XEXP (x, 0), x, changed), 0);
6071 return x;
6073 case EXPR_LIST:
6074 case INSN_LIST:
6075 if (REG_NOTE_KIND (x) == REG_EQUAL)
6076 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6077 if (XEXP (x, 1))
6078 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6079 return x;
6081 case SIGN_EXTEND:
6082 case ZERO_EXTEND:
6083 case SUBREG:
6085 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6086 /* We don't substitute VOIDmode constants into these rtx,
6087 since they would impede folding. */
6088 if (GET_MODE (new_rtx) != VOIDmode)
6089 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6090 return x;
6093 case REG:
6094 i = REG_QTY (REGNO (x));
6096 /* Return a constant or a constant register. */
6097 if (REGNO_QTY_VALID_P (REGNO (x)))
6099 struct qty_table_elem *ent = &qty_table[i];
6101 if (ent->const_rtx != NULL_RTX
6102 && (CONSTANT_P (ent->const_rtx)
6103 || REG_P (ent->const_rtx)))
6105 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6106 if (new_rtx)
6107 return copy_rtx (new_rtx);
6111 /* Otherwise, canonicalize this register. */
6112 return canon_reg (x, NULL_RTX);
6114 default:
6115 break;
6118 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6119 if (fmt[i] == 'e')
6120 validate_change (object, &XEXP (x, i),
6121 cse_process_notes (XEXP (x, i), object, changed), 0);
6123 return x;
6126 static rtx
6127 cse_process_notes (rtx x, rtx object, bool *changed)
6129 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6130 if (new_rtx != x)
6131 *changed = true;
6132 return new_rtx;
6136 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6138 DATA is a pointer to a struct cse_basic_block_data, that is used to
6139 describe the path.
6140 It is filled with a queue of basic blocks, starting with FIRST_BB
6141 and following a trace through the CFG.
6143 If all paths starting at FIRST_BB have been followed, or no new path
6144 starting at FIRST_BB can be constructed, this function returns FALSE.
6145 Otherwise, DATA->path is filled and the function returns TRUE indicating
6146 that a path to follow was found.
6148 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6149 block in the path will be FIRST_BB. */
6151 static bool
6152 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6153 int follow_jumps)
6155 basic_block bb;
6156 edge e;
6157 int path_size;
6159 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6161 /* See if there is a previous path. */
6162 path_size = data->path_size;
6164 /* There is a previous path. Make sure it started with FIRST_BB. */
6165 if (path_size)
6166 gcc_assert (data->path[0].bb == first_bb);
6168 /* There was only one basic block in the last path. Clear the path and
6169 return, so that paths starting at another basic block can be tried. */
6170 if (path_size == 1)
6172 path_size = 0;
6173 goto done;
6176 /* If the path was empty from the beginning, construct a new path. */
6177 if (path_size == 0)
6178 data->path[path_size++].bb = first_bb;
6179 else
6181 /* Otherwise, path_size must be equal to or greater than 2, because
6182 a previous path exists that is at least two basic blocks long.
6184 Update the previous branch path, if any. If the last branch was
6185 previously along the branch edge, take the fallthrough edge now. */
6186 while (path_size >= 2)
6188 basic_block last_bb_in_path, previous_bb_in_path;
6189 edge e;
6191 --path_size;
6192 last_bb_in_path = data->path[path_size].bb;
6193 previous_bb_in_path = data->path[path_size - 1].bb;
6195 /* If we previously followed a path along the branch edge, try
6196 the fallthru edge now. */
6197 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6198 && any_condjump_p (BB_END (previous_bb_in_path))
6199 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6200 && e == BRANCH_EDGE (previous_bb_in_path))
6202 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6203 if (bb != EXIT_BLOCK_PTR
6204 && single_pred_p (bb)
6205 /* We used to assert here that we would only see blocks
6206 that we have not visited yet. But we may end up
6207 visiting basic blocks twice if the CFG has changed
6208 in this run of cse_main, because when the CFG changes
6209 the topological sort of the CFG also changes. A basic
6210 blocks that previously had more than two predecessors
6211 may now have a single predecessor, and become part of
6212 a path that starts at another basic block.
6214 We still want to visit each basic block only once, so
6215 halt the path here if we have already visited BB. */
6216 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6218 SET_BIT (cse_visited_basic_blocks, bb->index);
6219 data->path[path_size++].bb = bb;
6220 break;
6224 data->path[path_size].bb = NULL;
6227 /* If only one block remains in the path, bail. */
6228 if (path_size == 1)
6230 path_size = 0;
6231 goto done;
6235 /* Extend the path if possible. */
6236 if (follow_jumps)
6238 bb = data->path[path_size - 1].bb;
6239 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6241 if (single_succ_p (bb))
6242 e = single_succ_edge (bb);
6243 else if (EDGE_COUNT (bb->succs) == 2
6244 && any_condjump_p (BB_END (bb)))
6246 /* First try to follow the branch. If that doesn't lead
6247 to a useful path, follow the fallthru edge. */
6248 e = BRANCH_EDGE (bb);
6249 if (!single_pred_p (e->dest))
6250 e = FALLTHRU_EDGE (bb);
6252 else
6253 e = NULL;
6255 if (e
6256 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6257 && e->dest != EXIT_BLOCK_PTR
6258 && single_pred_p (e->dest)
6259 /* Avoid visiting basic blocks twice. The large comment
6260 above explains why this can happen. */
6261 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6263 basic_block bb2 = e->dest;
6264 SET_BIT (cse_visited_basic_blocks, bb2->index);
6265 data->path[path_size++].bb = bb2;
6266 bb = bb2;
6268 else
6269 bb = NULL;
6273 done:
6274 data->path_size = path_size;
6275 return path_size != 0;
6278 /* Dump the path in DATA to file F. NSETS is the number of sets
6279 in the path. */
6281 static void
6282 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6284 int path_entry;
6286 fprintf (f, ";; Following path with %d sets: ", nsets);
6287 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6288 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6289 fputc ('\n', dump_file);
6290 fflush (f);
6294 /* Return true if BB has exception handling successor edges. */
6296 static bool
6297 have_eh_succ_edges (basic_block bb)
6299 edge e;
6300 edge_iterator ei;
6302 FOR_EACH_EDGE (e, ei, bb->succs)
6303 if (e->flags & EDGE_EH)
6304 return true;
6306 return false;
6310 /* Scan to the end of the path described by DATA. Return an estimate of
6311 the total number of SETs of all insns in the path. */
6313 static void
6314 cse_prescan_path (struct cse_basic_block_data *data)
6316 int nsets = 0;
6317 int path_size = data->path_size;
6318 int path_entry;
6320 /* Scan to end of each basic block in the path. */
6321 for (path_entry = 0; path_entry < path_size; path_entry++)
6323 basic_block bb;
6324 rtx insn;
6326 bb = data->path[path_entry].bb;
6328 FOR_BB_INSNS (bb, insn)
6330 if (!INSN_P (insn))
6331 continue;
6333 /* A PARALLEL can have lots of SETs in it,
6334 especially if it is really an ASM_OPERANDS. */
6335 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6336 nsets += XVECLEN (PATTERN (insn), 0);
6337 else
6338 nsets += 1;
6342 data->nsets = nsets;
6345 /* Process a single extended basic block described by EBB_DATA. */
6347 static void
6348 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6350 int path_size = ebb_data->path_size;
6351 int path_entry;
6352 int num_insns = 0;
6354 /* Allocate the space needed by qty_table. */
6355 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6357 new_basic_block ();
6358 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6359 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6360 for (path_entry = 0; path_entry < path_size; path_entry++)
6362 basic_block bb;
6363 rtx insn;
6365 bb = ebb_data->path[path_entry].bb;
6367 /* Invalidate recorded information for eh regs if there is an EH
6368 edge pointing to that bb. */
6369 if (bb_has_eh_pred (bb))
6371 df_ref *def_rec;
6373 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6375 df_ref def = *def_rec;
6376 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6377 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6381 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6382 FOR_BB_INSNS (bb, insn)
6384 /* If we have processed 1,000 insns, flush the hash table to
6385 avoid extreme quadratic behavior. We must not include NOTEs
6386 in the count since there may be more of them when generating
6387 debugging information. If we clear the table at different
6388 times, code generated with -g -O might be different than code
6389 generated with -O but not -g.
6391 FIXME: This is a real kludge and needs to be done some other
6392 way. */
6393 if (NONDEBUG_INSN_P (insn)
6394 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6396 flush_hash_table ();
6397 num_insns = 0;
6400 if (INSN_P (insn))
6402 /* Process notes first so we have all notes in canonical forms
6403 when looking for duplicate operations. */
6404 if (REG_NOTES (insn))
6406 bool changed = false;
6407 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6408 NULL_RTX, &changed);
6409 if (changed)
6410 df_notes_rescan (insn);
6413 cse_insn (insn);
6415 /* If we haven't already found an insn where we added a LABEL_REF,
6416 check this one. */
6417 if (INSN_P (insn) && !recorded_label_ref
6418 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6419 (void *) insn))
6420 recorded_label_ref = true;
6422 #ifdef HAVE_cc0
6423 if (NONDEBUG_INSN_P (insn))
6425 /* If the previous insn sets CC0 and this insn no
6426 longer references CC0, delete the previous insn.
6427 Here we use fact that nothing expects CC0 to be
6428 valid over an insn, which is true until the final
6429 pass. */
6430 rtx prev_insn, tem;
6432 prev_insn = prev_nonnote_nondebug_insn (insn);
6433 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6434 && (tem = single_set (prev_insn)) != NULL_RTX
6435 && SET_DEST (tem) == cc0_rtx
6436 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6437 delete_insn (prev_insn);
6439 /* If this insn is not the last insn in the basic
6440 block, it will be PREV_INSN(insn) in the next
6441 iteration. If we recorded any CC0-related
6442 information for this insn, remember it. */
6443 if (insn != BB_END (bb))
6445 prev_insn_cc0 = this_insn_cc0;
6446 prev_insn_cc0_mode = this_insn_cc0_mode;
6449 #endif
6453 /* With non-call exceptions, we are not always able to update
6454 the CFG properly inside cse_insn. So clean up possibly
6455 redundant EH edges here. */
6456 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6457 cse_cfg_altered |= purge_dead_edges (bb);
6459 /* If we changed a conditional jump, we may have terminated
6460 the path we are following. Check that by verifying that
6461 the edge we would take still exists. If the edge does
6462 not exist anymore, purge the remainder of the path.
6463 Note that this will cause us to return to the caller. */
6464 if (path_entry < path_size - 1)
6466 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6467 if (!find_edge (bb, next_bb))
6471 path_size--;
6473 /* If we truncate the path, we must also reset the
6474 visited bit on the remaining blocks in the path,
6475 or we will never visit them at all. */
6476 RESET_BIT (cse_visited_basic_blocks,
6477 ebb_data->path[path_size].bb->index);
6478 ebb_data->path[path_size].bb = NULL;
6480 while (path_size - 1 != path_entry);
6481 ebb_data->path_size = path_size;
6485 /* If this is a conditional jump insn, record any known
6486 equivalences due to the condition being tested. */
6487 insn = BB_END (bb);
6488 if (path_entry < path_size - 1
6489 && JUMP_P (insn)
6490 && single_set (insn)
6491 && any_condjump_p (insn))
6493 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6494 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6495 record_jump_equiv (insn, taken);
6498 #ifdef HAVE_cc0
6499 /* Clear the CC0-tracking related insns, they can't provide
6500 useful information across basic block boundaries. */
6501 prev_insn_cc0 = 0;
6502 #endif
6505 gcc_assert (next_qty <= max_qty);
6507 free (qty_table);
6511 /* Perform cse on the instructions of a function.
6512 F is the first instruction.
6513 NREGS is one plus the highest pseudo-reg number used in the instruction.
6515 Return 2 if jump optimizations should be redone due to simplifications
6516 in conditional jump instructions.
6517 Return 1 if the CFG should be cleaned up because it has been modified.
6518 Return 0 otherwise. */
6520 static int
6521 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6523 struct cse_basic_block_data ebb_data;
6524 basic_block bb;
6525 int *rc_order = XNEWVEC (int, last_basic_block);
6526 int i, n_blocks;
6528 df_set_flags (DF_LR_RUN_DCE);
6529 df_analyze ();
6530 df_set_flags (DF_DEFER_INSN_RESCAN);
6532 reg_scan (get_insns (), max_reg_num ());
6533 init_cse_reg_info (nregs);
6535 ebb_data.path = XNEWVEC (struct branch_path,
6536 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6538 cse_cfg_altered = false;
6539 cse_jumps_altered = false;
6540 recorded_label_ref = false;
6541 constant_pool_entries_cost = 0;
6542 constant_pool_entries_regcost = 0;
6543 ebb_data.path_size = 0;
6544 ebb_data.nsets = 0;
6545 rtl_hooks = cse_rtl_hooks;
6547 init_recog ();
6548 init_alias_analysis ();
6550 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6552 /* Set up the table of already visited basic blocks. */
6553 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6554 sbitmap_zero (cse_visited_basic_blocks);
6556 /* Loop over basic blocks in reverse completion order (RPO),
6557 excluding the ENTRY and EXIT blocks. */
6558 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6559 i = 0;
6560 while (i < n_blocks)
6562 /* Find the first block in the RPO queue that we have not yet
6563 processed before. */
6566 bb = BASIC_BLOCK (rc_order[i++]);
6568 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6569 && i < n_blocks);
6571 /* Find all paths starting with BB, and process them. */
6572 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6574 /* Pre-scan the path. */
6575 cse_prescan_path (&ebb_data);
6577 /* If this basic block has no sets, skip it. */
6578 if (ebb_data.nsets == 0)
6579 continue;
6581 /* Get a reasonable estimate for the maximum number of qty's
6582 needed for this path. For this, we take the number of sets
6583 and multiply that by MAX_RECOG_OPERANDS. */
6584 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6586 /* Dump the path we're about to process. */
6587 if (dump_file)
6588 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6590 cse_extended_basic_block (&ebb_data);
6594 /* Clean up. */
6595 end_alias_analysis ();
6596 free (reg_eqv_table);
6597 free (ebb_data.path);
6598 sbitmap_free (cse_visited_basic_blocks);
6599 free (rc_order);
6600 rtl_hooks = general_rtl_hooks;
6602 if (cse_jumps_altered || recorded_label_ref)
6603 return 2;
6604 else if (cse_cfg_altered)
6605 return 1;
6606 else
6607 return 0;
6610 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6611 which there isn't a REG_LABEL_OPERAND note.
6612 Return one if so. DATA is the insn. */
6614 static int
6615 check_for_label_ref (rtx *rtl, void *data)
6617 rtx insn = (rtx) data;
6619 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6620 note for it, we must rerun jump since it needs to place the note. If
6621 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6622 don't do this since no REG_LABEL_OPERAND will be added. */
6623 return (GET_CODE (*rtl) == LABEL_REF
6624 && ! LABEL_REF_NONLOCAL_P (*rtl)
6625 && (!JUMP_P (insn)
6626 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6627 && LABEL_P (XEXP (*rtl, 0))
6628 && INSN_UID (XEXP (*rtl, 0)) != 0
6629 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6632 /* Count the number of times registers are used (not set) in X.
6633 COUNTS is an array in which we accumulate the count, INCR is how much
6634 we count each register usage.
6636 Don't count a usage of DEST, which is the SET_DEST of a SET which
6637 contains X in its SET_SRC. This is because such a SET does not
6638 modify the liveness of DEST.
6639 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6640 We must then count uses of a SET_DEST regardless, because the insn can't be
6641 deleted here. */
6643 static void
6644 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6646 enum rtx_code code;
6647 rtx note;
6648 const char *fmt;
6649 int i, j;
6651 if (x == 0)
6652 return;
6654 switch (code = GET_CODE (x))
6656 case REG:
6657 if (x != dest)
6658 counts[REGNO (x)] += incr;
6659 return;
6661 case PC:
6662 case CC0:
6663 case CONST:
6664 case CONST_INT:
6665 case CONST_DOUBLE:
6666 case CONST_FIXED:
6667 case CONST_VECTOR:
6668 case SYMBOL_REF:
6669 case LABEL_REF:
6670 return;
6672 case CLOBBER:
6673 /* If we are clobbering a MEM, mark any registers inside the address
6674 as being used. */
6675 if (MEM_P (XEXP (x, 0)))
6676 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6677 return;
6679 case SET:
6680 /* Unless we are setting a REG, count everything in SET_DEST. */
6681 if (!REG_P (SET_DEST (x)))
6682 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6683 count_reg_usage (SET_SRC (x), counts,
6684 dest ? dest : SET_DEST (x),
6685 incr);
6686 return;
6688 case DEBUG_INSN:
6689 return;
6691 case CALL_INSN:
6692 case INSN:
6693 case JUMP_INSN:
6694 /* We expect dest to be NULL_RTX here. If the insn may throw,
6695 or if it cannot be deleted due to side-effects, mark this fact
6696 by setting DEST to pc_rtx. */
6697 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6698 || side_effects_p (PATTERN (x)))
6699 dest = pc_rtx;
6700 if (code == CALL_INSN)
6701 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6702 count_reg_usage (PATTERN (x), counts, dest, incr);
6704 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6705 use them. */
6707 note = find_reg_equal_equiv_note (x);
6708 if (note)
6710 rtx eqv = XEXP (note, 0);
6712 if (GET_CODE (eqv) == EXPR_LIST)
6713 /* This REG_EQUAL note describes the result of a function call.
6714 Process all the arguments. */
6717 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6718 eqv = XEXP (eqv, 1);
6720 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6721 else
6722 count_reg_usage (eqv, counts, dest, incr);
6724 return;
6726 case EXPR_LIST:
6727 if (REG_NOTE_KIND (x) == REG_EQUAL
6728 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6729 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6730 involving registers in the address. */
6731 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6732 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6734 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6735 return;
6737 case ASM_OPERANDS:
6738 /* Iterate over just the inputs, not the constraints as well. */
6739 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6740 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6741 return;
6743 case INSN_LIST:
6744 gcc_unreachable ();
6746 default:
6747 break;
6750 fmt = GET_RTX_FORMAT (code);
6751 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6753 if (fmt[i] == 'e')
6754 count_reg_usage (XEXP (x, i), counts, dest, incr);
6755 else if (fmt[i] == 'E')
6756 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6757 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6761 /* Return true if X is a dead register. */
6763 static inline int
6764 is_dead_reg (rtx x, int *counts)
6766 return (REG_P (x)
6767 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6768 && counts[REGNO (x)] == 0);
6771 /* Return true if set is live. */
6772 static bool
6773 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6774 int *counts)
6776 #ifdef HAVE_cc0
6777 rtx tem;
6778 #endif
6780 if (set_noop_p (set))
6783 #ifdef HAVE_cc0
6784 else if (GET_CODE (SET_DEST (set)) == CC0
6785 && !side_effects_p (SET_SRC (set))
6786 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6787 || !INSN_P (tem)
6788 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6789 return false;
6790 #endif
6791 else if (!is_dead_reg (SET_DEST (set), counts)
6792 || side_effects_p (SET_SRC (set)))
6793 return true;
6794 return false;
6797 /* Return true if insn is live. */
6799 static bool
6800 insn_live_p (rtx insn, int *counts)
6802 int i;
6803 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6804 return true;
6805 else if (GET_CODE (PATTERN (insn)) == SET)
6806 return set_live_p (PATTERN (insn), insn, counts);
6807 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6809 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6811 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6813 if (GET_CODE (elt) == SET)
6815 if (set_live_p (elt, insn, counts))
6816 return true;
6818 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6819 return true;
6821 return false;
6823 else if (DEBUG_INSN_P (insn))
6825 rtx next;
6827 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6828 if (NOTE_P (next))
6829 continue;
6830 else if (!DEBUG_INSN_P (next))
6831 return true;
6832 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6833 return false;
6835 return true;
6837 else
6838 return true;
6841 /* Count the number of stores into pseudo. Callback for note_stores. */
6843 static void
6844 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6846 int *counts = (int *) data;
6847 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6848 counts[REGNO (x)]++;
6851 struct dead_debug_insn_data
6853 int *counts;
6854 rtx *replacements;
6855 bool seen_repl;
6858 /* Return if a DEBUG_INSN needs to be reset because some dead
6859 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6861 static int
6862 is_dead_debug_insn (rtx *loc, void *data)
6864 rtx x = *loc;
6865 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6867 if (is_dead_reg (x, ddid->counts))
6869 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6870 ddid->seen_repl = true;
6871 else
6872 return 1;
6874 return 0;
6877 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6878 Callback for simplify_replace_fn_rtx. */
6880 static rtx
6881 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6883 rtx *replacements = (rtx *) data;
6885 if (REG_P (x)
6886 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6887 && replacements[REGNO (x)] != NULL_RTX)
6889 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6890 return replacements[REGNO (x)];
6891 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6892 GET_MODE (replacements[REGNO (x)]));
6894 return NULL_RTX;
6897 /* Scan all the insns and delete any that are dead; i.e., they store a register
6898 that is never used or they copy a register to itself.
6900 This is used to remove insns made obviously dead by cse, loop or other
6901 optimizations. It improves the heuristics in loop since it won't try to
6902 move dead invariants out of loops or make givs for dead quantities. The
6903 remaining passes of the compilation are also sped up. */
6906 delete_trivially_dead_insns (rtx insns, int nreg)
6908 int *counts;
6909 rtx insn, prev;
6910 rtx *replacements = NULL;
6911 int ndead = 0;
6913 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6914 /* First count the number of times each register is used. */
6915 if (MAY_HAVE_DEBUG_INSNS)
6917 counts = XCNEWVEC (int, nreg * 3);
6918 for (insn = insns; insn; insn = NEXT_INSN (insn))
6919 if (DEBUG_INSN_P (insn))
6920 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6921 NULL_RTX, 1);
6922 else if (INSN_P (insn))
6924 count_reg_usage (insn, counts, NULL_RTX, 1);
6925 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6927 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6928 First one counts how many times each pseudo is used outside
6929 of debug insns, second counts how many times each pseudo is
6930 used in debug insns and third counts how many times a pseudo
6931 is stored. */
6933 else
6935 counts = XCNEWVEC (int, nreg);
6936 for (insn = insns; insn; insn = NEXT_INSN (insn))
6937 if (INSN_P (insn))
6938 count_reg_usage (insn, counts, NULL_RTX, 1);
6939 /* If no debug insns can be present, COUNTS is just an array
6940 which counts how many times each pseudo is used. */
6942 /* Go from the last insn to the first and delete insns that only set unused
6943 registers or copy a register to itself. As we delete an insn, remove
6944 usage counts for registers it uses.
6946 The first jump optimization pass may leave a real insn as the last
6947 insn in the function. We must not skip that insn or we may end
6948 up deleting code that is not really dead.
6950 If some otherwise unused register is only used in DEBUG_INSNs,
6951 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6952 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6953 has been created for the unused register, replace it with
6954 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6955 for (insn = get_last_insn (); insn; insn = prev)
6957 int live_insn = 0;
6959 prev = PREV_INSN (insn);
6960 if (!INSN_P (insn))
6961 continue;
6963 live_insn = insn_live_p (insn, counts);
6965 /* If this is a dead insn, delete it and show registers in it aren't
6966 being used. */
6968 if (! live_insn && dbg_cnt (delete_trivial_dead))
6970 if (DEBUG_INSN_P (insn))
6971 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6972 NULL_RTX, -1);
6973 else
6975 rtx set;
6976 if (MAY_HAVE_DEBUG_INSNS
6977 && (set = single_set (insn)) != NULL_RTX
6978 && is_dead_reg (SET_DEST (set), counts)
6979 /* Used at least once in some DEBUG_INSN. */
6980 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6981 /* And set exactly once. */
6982 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6983 && !side_effects_p (SET_SRC (set))
6984 && asm_noperands (PATTERN (insn)) < 0)
6986 rtx dval, bind;
6988 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6989 dval = make_debug_expr_from_rtl (SET_DEST (set));
6991 /* Emit a debug bind insn before the insn in which
6992 reg dies. */
6993 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
6994 DEBUG_EXPR_TREE_DECL (dval),
6995 SET_SRC (set),
6996 VAR_INIT_STATUS_INITIALIZED);
6997 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
6999 bind = emit_debug_insn_before (bind, insn);
7000 df_insn_rescan (bind);
7002 if (replacements == NULL)
7003 replacements = XCNEWVEC (rtx, nreg);
7004 replacements[REGNO (SET_DEST (set))] = dval;
7007 count_reg_usage (insn, counts, NULL_RTX, -1);
7008 ndead++;
7010 delete_insn_and_edges (insn);
7014 if (MAY_HAVE_DEBUG_INSNS)
7016 struct dead_debug_insn_data ddid;
7017 ddid.counts = counts;
7018 ddid.replacements = replacements;
7019 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7020 if (DEBUG_INSN_P (insn))
7022 /* If this debug insn references a dead register that wasn't replaced
7023 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7024 ddid.seen_repl = false;
7025 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
7026 is_dead_debug_insn, &ddid))
7028 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7029 df_insn_rescan (insn);
7031 else if (ddid.seen_repl)
7033 INSN_VAR_LOCATION_LOC (insn)
7034 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7035 NULL_RTX, replace_dead_reg,
7036 replacements);
7037 df_insn_rescan (insn);
7040 free (replacements);
7043 if (dump_file && ndead)
7044 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7045 ndead);
7046 /* Clean up. */
7047 free (counts);
7048 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7049 return ndead;
7052 /* This function is called via for_each_rtx. The argument, NEWREG, is
7053 a condition code register with the desired mode. If we are looking
7054 at the same register in a different mode, replace it with
7055 NEWREG. */
7057 static int
7058 cse_change_cc_mode (rtx *loc, void *data)
7060 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7062 if (*loc
7063 && REG_P (*loc)
7064 && REGNO (*loc) == REGNO (args->newreg)
7065 && GET_MODE (*loc) != GET_MODE (args->newreg))
7067 validate_change (args->insn, loc, args->newreg, 1);
7069 return -1;
7071 return 0;
7074 /* Change the mode of any reference to the register REGNO (NEWREG) to
7075 GET_MODE (NEWREG) in INSN. */
7077 static void
7078 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7080 struct change_cc_mode_args args;
7081 int success;
7083 if (!INSN_P (insn))
7084 return;
7086 args.insn = insn;
7087 args.newreg = newreg;
7089 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7090 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7092 /* If the following assertion was triggered, there is most probably
7093 something wrong with the cc_modes_compatible back end function.
7094 CC modes only can be considered compatible if the insn - with the mode
7095 replaced by any of the compatible modes - can still be recognized. */
7096 success = apply_change_group ();
7097 gcc_assert (success);
7100 /* Change the mode of any reference to the register REGNO (NEWREG) to
7101 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7102 any instruction which modifies NEWREG. */
7104 static void
7105 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7107 rtx insn;
7109 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7111 if (! INSN_P (insn))
7112 continue;
7114 if (reg_set_p (newreg, insn))
7115 return;
7117 cse_change_cc_mode_insn (insn, newreg);
7121 /* BB is a basic block which finishes with CC_REG as a condition code
7122 register which is set to CC_SRC. Look through the successors of BB
7123 to find blocks which have a single predecessor (i.e., this one),
7124 and look through those blocks for an assignment to CC_REG which is
7125 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7126 permitted to change the mode of CC_SRC to a compatible mode. This
7127 returns VOIDmode if no equivalent assignments were found.
7128 Otherwise it returns the mode which CC_SRC should wind up with.
7129 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7130 but is passed unmodified down to recursive calls in order to prevent
7131 endless recursion.
7133 The main complexity in this function is handling the mode issues.
7134 We may have more than one duplicate which we can eliminate, and we
7135 try to find a mode which will work for multiple duplicates. */
7137 static enum machine_mode
7138 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7139 bool can_change_mode)
7141 bool found_equiv;
7142 enum machine_mode mode;
7143 unsigned int insn_count;
7144 edge e;
7145 rtx insns[2];
7146 enum machine_mode modes[2];
7147 rtx last_insns[2];
7148 unsigned int i;
7149 rtx newreg;
7150 edge_iterator ei;
7152 /* We expect to have two successors. Look at both before picking
7153 the final mode for the comparison. If we have more successors
7154 (i.e., some sort of table jump, although that seems unlikely),
7155 then we require all beyond the first two to use the same
7156 mode. */
7158 found_equiv = false;
7159 mode = GET_MODE (cc_src);
7160 insn_count = 0;
7161 FOR_EACH_EDGE (e, ei, bb->succs)
7163 rtx insn;
7164 rtx end;
7166 if (e->flags & EDGE_COMPLEX)
7167 continue;
7169 if (EDGE_COUNT (e->dest->preds) != 1
7170 || e->dest == EXIT_BLOCK_PTR
7171 /* Avoid endless recursion on unreachable blocks. */
7172 || e->dest == orig_bb)
7173 continue;
7175 end = NEXT_INSN (BB_END (e->dest));
7176 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7178 rtx set;
7180 if (! INSN_P (insn))
7181 continue;
7183 /* If CC_SRC is modified, we have to stop looking for
7184 something which uses it. */
7185 if (modified_in_p (cc_src, insn))
7186 break;
7188 /* Check whether INSN sets CC_REG to CC_SRC. */
7189 set = single_set (insn);
7190 if (set
7191 && REG_P (SET_DEST (set))
7192 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7194 bool found;
7195 enum machine_mode set_mode;
7196 enum machine_mode comp_mode;
7198 found = false;
7199 set_mode = GET_MODE (SET_SRC (set));
7200 comp_mode = set_mode;
7201 if (rtx_equal_p (cc_src, SET_SRC (set)))
7202 found = true;
7203 else if (GET_CODE (cc_src) == COMPARE
7204 && GET_CODE (SET_SRC (set)) == COMPARE
7205 && mode != set_mode
7206 && rtx_equal_p (XEXP (cc_src, 0),
7207 XEXP (SET_SRC (set), 0))
7208 && rtx_equal_p (XEXP (cc_src, 1),
7209 XEXP (SET_SRC (set), 1)))
7212 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7213 if (comp_mode != VOIDmode
7214 && (can_change_mode || comp_mode == mode))
7215 found = true;
7218 if (found)
7220 found_equiv = true;
7221 if (insn_count < ARRAY_SIZE (insns))
7223 insns[insn_count] = insn;
7224 modes[insn_count] = set_mode;
7225 last_insns[insn_count] = end;
7226 ++insn_count;
7228 if (mode != comp_mode)
7230 gcc_assert (can_change_mode);
7231 mode = comp_mode;
7233 /* The modified insn will be re-recognized later. */
7234 PUT_MODE (cc_src, mode);
7237 else
7239 if (set_mode != mode)
7241 /* We found a matching expression in the
7242 wrong mode, but we don't have room to
7243 store it in the array. Punt. This case
7244 should be rare. */
7245 break;
7247 /* INSN sets CC_REG to a value equal to CC_SRC
7248 with the right mode. We can simply delete
7249 it. */
7250 delete_insn (insn);
7253 /* We found an instruction to delete. Keep looking,
7254 in the hopes of finding a three-way jump. */
7255 continue;
7258 /* We found an instruction which sets the condition
7259 code, so don't look any farther. */
7260 break;
7263 /* If INSN sets CC_REG in some other way, don't look any
7264 farther. */
7265 if (reg_set_p (cc_reg, insn))
7266 break;
7269 /* If we fell off the bottom of the block, we can keep looking
7270 through successors. We pass CAN_CHANGE_MODE as false because
7271 we aren't prepared to handle compatibility between the
7272 further blocks and this block. */
7273 if (insn == end)
7275 enum machine_mode submode;
7277 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7278 if (submode != VOIDmode)
7280 gcc_assert (submode == mode);
7281 found_equiv = true;
7282 can_change_mode = false;
7287 if (! found_equiv)
7288 return VOIDmode;
7290 /* Now INSN_COUNT is the number of instructions we found which set
7291 CC_REG to a value equivalent to CC_SRC. The instructions are in
7292 INSNS. The modes used by those instructions are in MODES. */
7294 newreg = NULL_RTX;
7295 for (i = 0; i < insn_count; ++i)
7297 if (modes[i] != mode)
7299 /* We need to change the mode of CC_REG in INSNS[i] and
7300 subsequent instructions. */
7301 if (! newreg)
7303 if (GET_MODE (cc_reg) == mode)
7304 newreg = cc_reg;
7305 else
7306 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7308 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7309 newreg);
7312 delete_insn_and_edges (insns[i]);
7315 return mode;
7318 /* If we have a fixed condition code register (or two), walk through
7319 the instructions and try to eliminate duplicate assignments. */
7321 static void
7322 cse_condition_code_reg (void)
7324 unsigned int cc_regno_1;
7325 unsigned int cc_regno_2;
7326 rtx cc_reg_1;
7327 rtx cc_reg_2;
7328 basic_block bb;
7330 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7331 return;
7333 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7334 if (cc_regno_2 != INVALID_REGNUM)
7335 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7336 else
7337 cc_reg_2 = NULL_RTX;
7339 FOR_EACH_BB (bb)
7341 rtx last_insn;
7342 rtx cc_reg;
7343 rtx insn;
7344 rtx cc_src_insn;
7345 rtx cc_src;
7346 enum machine_mode mode;
7347 enum machine_mode orig_mode;
7349 /* Look for blocks which end with a conditional jump based on a
7350 condition code register. Then look for the instruction which
7351 sets the condition code register. Then look through the
7352 successor blocks for instructions which set the condition
7353 code register to the same value. There are other possible
7354 uses of the condition code register, but these are by far the
7355 most common and the ones which we are most likely to be able
7356 to optimize. */
7358 last_insn = BB_END (bb);
7359 if (!JUMP_P (last_insn))
7360 continue;
7362 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7363 cc_reg = cc_reg_1;
7364 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7365 cc_reg = cc_reg_2;
7366 else
7367 continue;
7369 cc_src_insn = NULL_RTX;
7370 cc_src = NULL_RTX;
7371 for (insn = PREV_INSN (last_insn);
7372 insn && insn != PREV_INSN (BB_HEAD (bb));
7373 insn = PREV_INSN (insn))
7375 rtx set;
7377 if (! INSN_P (insn))
7378 continue;
7379 set = single_set (insn);
7380 if (set
7381 && REG_P (SET_DEST (set))
7382 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7384 cc_src_insn = insn;
7385 cc_src = SET_SRC (set);
7386 break;
7388 else if (reg_set_p (cc_reg, insn))
7389 break;
7392 if (! cc_src_insn)
7393 continue;
7395 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7396 continue;
7398 /* Now CC_REG is a condition code register used for a
7399 conditional jump at the end of the block, and CC_SRC, in
7400 CC_SRC_INSN, is the value to which that condition code
7401 register is set, and CC_SRC is still meaningful at the end of
7402 the basic block. */
7404 orig_mode = GET_MODE (cc_src);
7405 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7406 if (mode != VOIDmode)
7408 gcc_assert (mode == GET_MODE (cc_src));
7409 if (mode != orig_mode)
7411 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7413 cse_change_cc_mode_insn (cc_src_insn, newreg);
7415 /* Do the same in the following insns that use the
7416 current value of CC_REG within BB. */
7417 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7418 NEXT_INSN (last_insn),
7419 newreg);
7426 /* Perform common subexpression elimination. Nonzero value from
7427 `cse_main' means that jumps were simplified and some code may now
7428 be unreachable, so do jump optimization again. */
7429 static bool
7430 gate_handle_cse (void)
7432 return optimize > 0;
7435 static unsigned int
7436 rest_of_handle_cse (void)
7438 int tem;
7440 if (dump_file)
7441 dump_flow_info (dump_file, dump_flags);
7443 tem = cse_main (get_insns (), max_reg_num ());
7445 /* If we are not running more CSE passes, then we are no longer
7446 expecting CSE to be run. But always rerun it in a cheap mode. */
7447 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7449 if (tem == 2)
7451 timevar_push (TV_JUMP);
7452 rebuild_jump_labels (get_insns ());
7453 cleanup_cfg (CLEANUP_CFG_CHANGED);
7454 timevar_pop (TV_JUMP);
7456 else if (tem == 1 || optimize > 1)
7457 cleanup_cfg (0);
7459 return 0;
7462 struct rtl_opt_pass pass_cse =
7465 RTL_PASS,
7466 "cse1", /* name */
7467 gate_handle_cse, /* gate */
7468 rest_of_handle_cse, /* execute */
7469 NULL, /* sub */
7470 NULL, /* next */
7471 0, /* static_pass_number */
7472 TV_CSE, /* tv_id */
7473 0, /* properties_required */
7474 0, /* properties_provided */
7475 0, /* properties_destroyed */
7476 0, /* todo_flags_start */
7477 TODO_df_finish | TODO_verify_rtl_sharing |
7478 TODO_ggc_collect |
7479 TODO_verify_flow, /* todo_flags_finish */
7484 static bool
7485 gate_handle_cse2 (void)
7487 return optimize > 0 && flag_rerun_cse_after_loop;
7490 /* Run second CSE pass after loop optimizations. */
7491 static unsigned int
7492 rest_of_handle_cse2 (void)
7494 int tem;
7496 if (dump_file)
7497 dump_flow_info (dump_file, dump_flags);
7499 tem = cse_main (get_insns (), max_reg_num ());
7501 /* Run a pass to eliminate duplicated assignments to condition code
7502 registers. We have to run this after bypass_jumps, because it
7503 makes it harder for that pass to determine whether a jump can be
7504 bypassed safely. */
7505 cse_condition_code_reg ();
7507 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7509 if (tem == 2)
7511 timevar_push (TV_JUMP);
7512 rebuild_jump_labels (get_insns ());
7513 cleanup_cfg (CLEANUP_CFG_CHANGED);
7514 timevar_pop (TV_JUMP);
7516 else if (tem == 1)
7517 cleanup_cfg (0);
7519 cse_not_expected = 1;
7520 return 0;
7524 struct rtl_opt_pass pass_cse2 =
7527 RTL_PASS,
7528 "cse2", /* name */
7529 gate_handle_cse2, /* gate */
7530 rest_of_handle_cse2, /* execute */
7531 NULL, /* sub */
7532 NULL, /* next */
7533 0, /* static_pass_number */
7534 TV_CSE2, /* tv_id */
7535 0, /* properties_required */
7536 0, /* properties_provided */
7537 0, /* properties_destroyed */
7538 0, /* todo_flags_start */
7539 TODO_df_finish | TODO_verify_rtl_sharing |
7540 TODO_ggc_collect |
7541 TODO_verify_flow /* todo_flags_finish */
7545 static bool
7546 gate_handle_cse_after_global_opts (void)
7548 return optimize > 0 && flag_rerun_cse_after_global_opts;
7551 /* Run second CSE pass after loop optimizations. */
7552 static unsigned int
7553 rest_of_handle_cse_after_global_opts (void)
7555 int save_cfj;
7556 int tem;
7558 /* We only want to do local CSE, so don't follow jumps. */
7559 save_cfj = flag_cse_follow_jumps;
7560 flag_cse_follow_jumps = 0;
7562 rebuild_jump_labels (get_insns ());
7563 tem = cse_main (get_insns (), max_reg_num ());
7564 purge_all_dead_edges ();
7565 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7567 cse_not_expected = !flag_rerun_cse_after_loop;
7569 /* If cse altered any jumps, rerun jump opts to clean things up. */
7570 if (tem == 2)
7572 timevar_push (TV_JUMP);
7573 rebuild_jump_labels (get_insns ());
7574 cleanup_cfg (CLEANUP_CFG_CHANGED);
7575 timevar_pop (TV_JUMP);
7577 else if (tem == 1)
7578 cleanup_cfg (0);
7580 flag_cse_follow_jumps = save_cfj;
7581 return 0;
7584 struct rtl_opt_pass pass_cse_after_global_opts =
7587 RTL_PASS,
7588 "cse_local", /* name */
7589 gate_handle_cse_after_global_opts, /* gate */
7590 rest_of_handle_cse_after_global_opts, /* execute */
7591 NULL, /* sub */
7592 NULL, /* next */
7593 0, /* static_pass_number */
7594 TV_CSE, /* tv_id */
7595 0, /* properties_required */
7596 0, /* properties_provided */
7597 0, /* properties_destroyed */
7598 0, /* todo_flags_start */
7599 TODO_df_finish | TODO_verify_rtl_sharing |
7600 TODO_ggc_collect |
7601 TODO_verify_flow /* todo_flags_finish */