2008-08-17 Paul Thomas <pault@gcc.gnu.org>
[official-gcc.git] / gcc / cse.c
blob5f83892c79fe6aecaf224d4e769fb9b7d3b6cf56
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
4 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 /* stdio.h must precede rtl.h for FFS. */
24 #include "system.h"
25 #include "coretypes.h"
26 #include "tm.h"
27 #include "rtl.h"
28 #include "tm_p.h"
29 #include "hard-reg-set.h"
30 #include "regs.h"
31 #include "basic-block.h"
32 #include "flags.h"
33 #include "real.h"
34 #include "insn-config.h"
35 #include "recog.h"
36 #include "function.h"
37 #include "expr.h"
38 #include "toplev.h"
39 #include "output.h"
40 #include "ggc.h"
41 #include "timevar.h"
42 #include "except.h"
43 #include "target.h"
44 #include "params.h"
45 #include "rtlhooks-def.h"
46 #include "tree-pass.h"
47 #include "df.h"
48 #include "dbgcnt.h"
50 /* The basic idea of common subexpression elimination is to go
51 through the code, keeping a record of expressions that would
52 have the same value at the current scan point, and replacing
53 expressions encountered with the cheapest equivalent expression.
55 It is too complicated to keep track of the different possibilities
56 when control paths merge in this code; so, at each label, we forget all
57 that is known and start fresh. This can be described as processing each
58 extended basic block separately. We have a separate pass to perform
59 global CSE.
61 Note CSE can turn a conditional or computed jump into a nop or
62 an unconditional jump. When this occurs we arrange to run the jump
63 optimizer after CSE to delete the unreachable code.
65 We use two data structures to record the equivalent expressions:
66 a hash table for most expressions, and a vector of "quantity
67 numbers" to record equivalent (pseudo) registers.
69 The use of the special data structure for registers is desirable
70 because it is faster. It is possible because registers references
71 contain a fairly small number, the register number, taken from
72 a contiguously allocated series, and two register references are
73 identical if they have the same number. General expressions
74 do not have any such thing, so the only way to retrieve the
75 information recorded on an expression other than a register
76 is to keep it in a hash table.
78 Registers and "quantity numbers":
80 At the start of each basic block, all of the (hardware and pseudo)
81 registers used in the function are given distinct quantity
82 numbers to indicate their contents. During scan, when the code
83 copies one register into another, we copy the quantity number.
84 When a register is loaded in any other way, we allocate a new
85 quantity number to describe the value generated by this operation.
86 `REG_QTY (N)' records what quantity register N is currently thought
87 of as containing.
89 All real quantity numbers are greater than or equal to zero.
90 If register N has not been assigned a quantity, `REG_QTY (N)' will
91 equal -N - 1, which is always negative.
93 Quantity numbers below zero do not exist and none of the `qty_table'
94 entries should be referenced with a negative index.
96 We also maintain a bidirectional chain of registers for each
97 quantity number. The `qty_table` members `first_reg' and `last_reg',
98 and `reg_eqv_table' members `next' and `prev' hold these chains.
100 The first register in a chain is the one whose lifespan is least local.
101 Among equals, it is the one that was seen first.
102 We replace any equivalent register with that one.
104 If two registers have the same quantity number, it must be true that
105 REG expressions with qty_table `mode' must be in the hash table for both
106 registers and must be in the same class.
108 The converse is not true. Since hard registers may be referenced in
109 any mode, two REG expressions might be equivalent in the hash table
110 but not have the same quantity number if the quantity number of one
111 of the registers is not the same mode as those expressions.
113 Constants and quantity numbers
115 When a quantity has a known constant value, that value is stored
116 in the appropriate qty_table `const_rtx'. This is in addition to
117 putting the constant in the hash table as is usual for non-regs.
119 Whether a reg or a constant is preferred is determined by the configuration
120 macro CONST_COSTS and will often depend on the constant value. In any
121 event, expressions containing constants can be simplified, by fold_rtx.
123 When a quantity has a known nearly constant value (such as an address
124 of a stack slot), that value is stored in the appropriate qty_table
125 `const_rtx'.
127 Integer constants don't have a machine mode. However, cse
128 determines the intended machine mode from the destination
129 of the instruction that moves the constant. The machine mode
130 is recorded in the hash table along with the actual RTL
131 constant expression so that different modes are kept separate.
133 Other expressions:
135 To record known equivalences among expressions in general
136 we use a hash table called `table'. It has a fixed number of buckets
137 that contain chains of `struct table_elt' elements for expressions.
138 These chains connect the elements whose expressions have the same
139 hash codes.
141 Other chains through the same elements connect the elements which
142 currently have equivalent values.
144 Register references in an expression are canonicalized before hashing
145 the expression. This is done using `reg_qty' and qty_table `first_reg'.
146 The hash code of a register reference is computed using the quantity
147 number, not the register number.
149 When the value of an expression changes, it is necessary to remove from the
150 hash table not just that expression but all expressions whose values
151 could be different as a result.
153 1. If the value changing is in memory, except in special cases
154 ANYTHING referring to memory could be changed. That is because
155 nobody knows where a pointer does not point.
156 The function `invalidate_memory' removes what is necessary.
158 The special cases are when the address is constant or is
159 a constant plus a fixed register such as the frame pointer
160 or a static chain pointer. When such addresses are stored in,
161 we can tell exactly which other such addresses must be invalidated
162 due to overlap. `invalidate' does this.
163 All expressions that refer to non-constant
164 memory addresses are also invalidated. `invalidate_memory' does this.
166 2. If the value changing is a register, all expressions
167 containing references to that register, and only those,
168 must be removed.
170 Because searching the entire hash table for expressions that contain
171 a register is very slow, we try to figure out when it isn't necessary.
172 Precisely, this is necessary only when expressions have been
173 entered in the hash table using this register, and then the value has
174 changed, and then another expression wants to be added to refer to
175 the register's new value. This sequence of circumstances is rare
176 within any one basic block.
178 `REG_TICK' and `REG_IN_TABLE', accessors for members of
179 cse_reg_info, are used to detect this case. REG_TICK (i) is
180 incremented whenever a value is stored in register i.
181 REG_IN_TABLE (i) holds -1 if no references to register i have been
182 entered in the table; otherwise, it contains the value REG_TICK (i)
183 had when the references were entered. If we want to enter a
184 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
185 remove old references. Until we want to enter a new entry, the
186 mere fact that the two vectors don't match makes the entries be
187 ignored if anyone tries to match them.
189 Registers themselves are entered in the hash table as well as in
190 the equivalent-register chains. However, `REG_TICK' and
191 `REG_IN_TABLE' do not apply to expressions which are simple
192 register references. These expressions are removed from the table
193 immediately when they become invalid, and this can be done even if
194 we do not immediately search for all the expressions that refer to
195 the register.
197 A CLOBBER rtx in an instruction invalidates its operand for further
198 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
199 invalidates everything that resides in memory.
201 Related expressions:
203 Constant expressions that differ only by an additive integer
204 are called related. When a constant expression is put in
205 the table, the related expression with no constant term
206 is also entered. These are made to point at each other
207 so that it is possible to find out if there exists any
208 register equivalent to an expression related to a given expression. */
210 /* Length of qty_table vector. We know in advance we will not need
211 a quantity number this big. */
213 static int max_qty;
215 /* Next quantity number to be allocated.
216 This is 1 + the largest number needed so far. */
218 static int next_qty;
220 /* Per-qty information tracking.
222 `first_reg' and `last_reg' track the head and tail of the
223 chain of registers which currently contain this quantity.
225 `mode' contains the machine mode of this quantity.
227 `const_rtx' holds the rtx of the constant value of this
228 quantity, if known. A summations of the frame/arg pointer
229 and a constant can also be entered here. When this holds
230 a known value, `const_insn' is the insn which stored the
231 constant value.
233 `comparison_{code,const,qty}' are used to track when a
234 comparison between a quantity and some constant or register has
235 been passed. In such a case, we know the results of the comparison
236 in case we see it again. These members record a comparison that
237 is known to be true. `comparison_code' holds the rtx code of such
238 a comparison, else it is set to UNKNOWN and the other two
239 comparison members are undefined. `comparison_const' holds
240 the constant being compared against, or zero if the comparison
241 is not against a constant. `comparison_qty' holds the quantity
242 being compared against when the result is known. If the comparison
243 is not with a register, `comparison_qty' is -1. */
245 struct qty_table_elem
247 rtx const_rtx;
248 rtx const_insn;
249 rtx comparison_const;
250 int comparison_qty;
251 unsigned int first_reg, last_reg;
252 /* The sizes of these fields should match the sizes of the
253 code and mode fields of struct rtx_def (see rtl.h). */
254 ENUM_BITFIELD(rtx_code) comparison_code : 16;
255 ENUM_BITFIELD(machine_mode) mode : 8;
258 /* The table of all qtys, indexed by qty number. */
259 static struct qty_table_elem *qty_table;
261 /* Structure used to pass arguments via for_each_rtx to function
262 cse_change_cc_mode. */
263 struct change_cc_mode_args
265 rtx insn;
266 rtx newreg;
269 #ifdef HAVE_cc0
270 /* For machines that have a CC0, we do not record its value in the hash
271 table since its use is guaranteed to be the insn immediately following
272 its definition and any other insn is presumed to invalidate it.
274 Instead, we store below the current and last value assigned to CC0.
275 If it should happen to be a constant, it is stored in preference
276 to the actual assigned value. In case it is a constant, we store
277 the mode in which the constant should be interpreted. */
279 static rtx this_insn_cc0, prev_insn_cc0;
280 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
281 #endif
283 /* Insn being scanned. */
285 static rtx this_insn;
286 static bool optimize_this_for_speed_p;
288 /* Index by register number, gives the number of the next (or
289 previous) register in the chain of registers sharing the same
290 value.
292 Or -1 if this register is at the end of the chain.
294 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
296 /* Per-register equivalence chain. */
297 struct reg_eqv_elem
299 int next, prev;
302 /* The table of all register equivalence chains. */
303 static struct reg_eqv_elem *reg_eqv_table;
305 struct cse_reg_info
307 /* The timestamp at which this register is initialized. */
308 unsigned int timestamp;
310 /* The quantity number of the register's current contents. */
311 int reg_qty;
313 /* The number of times the register has been altered in the current
314 basic block. */
315 int reg_tick;
317 /* The REG_TICK value at which rtx's containing this register are
318 valid in the hash table. If this does not equal the current
319 reg_tick value, such expressions existing in the hash table are
320 invalid. */
321 int reg_in_table;
323 /* The SUBREG that was set when REG_TICK was last incremented. Set
324 to -1 if the last store was to the whole register, not a subreg. */
325 unsigned int subreg_ticked;
328 /* A table of cse_reg_info indexed by register numbers. */
329 static struct cse_reg_info *cse_reg_info_table;
331 /* The size of the above table. */
332 static unsigned int cse_reg_info_table_size;
334 /* The index of the first entry that has not been initialized. */
335 static unsigned int cse_reg_info_table_first_uninitialized;
337 /* The timestamp at the beginning of the current run of
338 cse_extended_basic_block. We increment this variable at the beginning of
339 the current run of cse_extended_basic_block. The timestamp field of a
340 cse_reg_info entry matches the value of this variable if and only
341 if the entry has been initialized during the current run of
342 cse_extended_basic_block. */
343 static unsigned int cse_reg_info_timestamp;
345 /* A HARD_REG_SET containing all the hard registers for which there is
346 currently a REG expression in the hash table. Note the difference
347 from the above variables, which indicate if the REG is mentioned in some
348 expression in the table. */
350 static HARD_REG_SET hard_regs_in_table;
352 /* True if CSE has altered the CFG. */
353 static bool cse_cfg_altered;
355 /* True if CSE has altered conditional jump insns in such a way
356 that jump optimization should be redone. */
357 static bool cse_jumps_altered;
359 /* True if we put a LABEL_REF into the hash table for an INSN
360 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
361 to put in the note. */
362 static bool recorded_label_ref;
364 /* canon_hash stores 1 in do_not_record
365 if it notices a reference to CC0, PC, or some other volatile
366 subexpression. */
368 static int do_not_record;
370 /* canon_hash stores 1 in hash_arg_in_memory
371 if it notices a reference to memory within the expression being hashed. */
373 static int hash_arg_in_memory;
375 /* The hash table contains buckets which are chains of `struct table_elt's,
376 each recording one expression's information.
377 That expression is in the `exp' field.
379 The canon_exp field contains a canonical (from the point of view of
380 alias analysis) version of the `exp' field.
382 Those elements with the same hash code are chained in both directions
383 through the `next_same_hash' and `prev_same_hash' fields.
385 Each set of expressions with equivalent values
386 are on a two-way chain through the `next_same_value'
387 and `prev_same_value' fields, and all point with
388 the `first_same_value' field at the first element in
389 that chain. The chain is in order of increasing cost.
390 Each element's cost value is in its `cost' field.
392 The `in_memory' field is nonzero for elements that
393 involve any reference to memory. These elements are removed
394 whenever a write is done to an unidentified location in memory.
395 To be safe, we assume that a memory address is unidentified unless
396 the address is either a symbol constant or a constant plus
397 the frame pointer or argument pointer.
399 The `related_value' field is used to connect related expressions
400 (that differ by adding an integer).
401 The related expressions are chained in a circular fashion.
402 `related_value' is zero for expressions for which this
403 chain is not useful.
405 The `cost' field stores the cost of this element's expression.
406 The `regcost' field stores the value returned by approx_reg_cost for
407 this element's expression.
409 The `is_const' flag is set if the element is a constant (including
410 a fixed address).
412 The `flag' field is used as a temporary during some search routines.
414 The `mode' field is usually the same as GET_MODE (`exp'), but
415 if `exp' is a CONST_INT and has no machine mode then the `mode'
416 field is the mode it was being used as. Each constant is
417 recorded separately for each mode it is used with. */
419 struct table_elt
421 rtx exp;
422 rtx canon_exp;
423 struct table_elt *next_same_hash;
424 struct table_elt *prev_same_hash;
425 struct table_elt *next_same_value;
426 struct table_elt *prev_same_value;
427 struct table_elt *first_same_value;
428 struct table_elt *related_value;
429 int cost;
430 int regcost;
431 /* The size of this field should match the size
432 of the mode field of struct rtx_def (see rtl.h). */
433 ENUM_BITFIELD(machine_mode) mode : 8;
434 char in_memory;
435 char is_const;
436 char flag;
439 /* We don't want a lot of buckets, because we rarely have very many
440 things stored in the hash table, and a lot of buckets slows
441 down a lot of loops that happen frequently. */
442 #define HASH_SHIFT 5
443 #define HASH_SIZE (1 << HASH_SHIFT)
444 #define HASH_MASK (HASH_SIZE - 1)
446 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
447 register (hard registers may require `do_not_record' to be set). */
449 #define HASH(X, M) \
450 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
451 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
452 : canon_hash (X, M)) & HASH_MASK)
454 /* Like HASH, but without side-effects. */
455 #define SAFE_HASH(X, M) \
456 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
457 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
458 : safe_hash (X, M)) & HASH_MASK)
460 /* Determine whether register number N is considered a fixed register for the
461 purpose of approximating register costs.
462 It is desirable to replace other regs with fixed regs, to reduce need for
463 non-fixed hard regs.
464 A reg wins if it is either the frame pointer or designated as fixed. */
465 #define FIXED_REGNO_P(N) \
466 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
467 || fixed_regs[N] || global_regs[N])
469 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
470 hard registers and pointers into the frame are the cheapest with a cost
471 of 0. Next come pseudos with a cost of one and other hard registers with
472 a cost of 2. Aside from these special cases, call `rtx_cost'. */
474 #define CHEAP_REGNO(N) \
475 (REGNO_PTR_FRAME_P(N) \
476 || (HARD_REGISTER_NUM_P (N) \
477 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
479 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET))
480 #define COST_IN(X,OUTER) (REG_P (X) ? 0 : notreg_cost (X, OUTER))
482 /* Get the number of times this register has been updated in this
483 basic block. */
485 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
487 /* Get the point at which REG was recorded in the table. */
489 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
491 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
492 SUBREG). */
494 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
496 /* Get the quantity number for REG. */
498 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
500 /* Determine if the quantity number for register X represents a valid index
501 into the qty_table. */
503 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
505 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
507 #define CHEAPER(X, Y) \
508 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
510 static struct table_elt *table[HASH_SIZE];
512 /* Chain of `struct table_elt's made so far for this function
513 but currently removed from the table. */
515 static struct table_elt *free_element_chain;
517 /* Set to the cost of a constant pool reference if one was found for a
518 symbolic constant. If this was found, it means we should try to
519 convert constants into constant pool entries if they don't fit in
520 the insn. */
522 static int constant_pool_entries_cost;
523 static int constant_pool_entries_regcost;
525 /* Trace a patch through the CFG. */
527 struct branch_path
529 /* The basic block for this path entry. */
530 basic_block bb;
533 /* This data describes a block that will be processed by
534 cse_extended_basic_block. */
536 struct cse_basic_block_data
538 /* Total number of SETs in block. */
539 int nsets;
540 /* Size of current branch path, if any. */
541 int path_size;
542 /* Current path, indicating which basic_blocks will be processed. */
543 struct branch_path *path;
547 /* Pointers to the live in/live out bitmaps for the boundaries of the
548 current EBB. */
549 static bitmap cse_ebb_live_in, cse_ebb_live_out;
551 /* A simple bitmap to track which basic blocks have been visited
552 already as part of an already processed extended basic block. */
553 static sbitmap cse_visited_basic_blocks;
555 static bool fixed_base_plus_p (rtx x);
556 static int notreg_cost (rtx, enum rtx_code);
557 static int approx_reg_cost_1 (rtx *, void *);
558 static int approx_reg_cost (rtx);
559 static int preferable (int, int, int, int);
560 static void new_basic_block (void);
561 static void make_new_qty (unsigned int, enum machine_mode);
562 static void make_regs_eqv (unsigned int, unsigned int);
563 static void delete_reg_equiv (unsigned int);
564 static int mention_regs (rtx);
565 static int insert_regs (rtx, struct table_elt *, int);
566 static void remove_from_table (struct table_elt *, unsigned);
567 static void remove_pseudo_from_table (rtx, unsigned);
568 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
569 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
570 static rtx lookup_as_function (rtx, enum rtx_code);
571 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
572 enum machine_mode, int, int);
573 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
574 enum machine_mode);
575 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
576 static void invalidate (rtx, enum machine_mode);
577 static bool cse_rtx_varies_p (const_rtx, bool);
578 static void remove_invalid_refs (unsigned int);
579 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
580 enum machine_mode);
581 static void rehash_using_reg (rtx);
582 static void invalidate_memory (void);
583 static void invalidate_for_call (void);
584 static rtx use_related_value (rtx, struct table_elt *);
586 static inline unsigned canon_hash (rtx, enum machine_mode);
587 static inline unsigned safe_hash (rtx, enum machine_mode);
588 static inline unsigned hash_rtx_string (const char *);
590 static rtx canon_reg (rtx, rtx);
591 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
592 enum machine_mode *,
593 enum machine_mode *);
594 static rtx fold_rtx (rtx, rtx);
595 static rtx equiv_constant (rtx);
596 static void record_jump_equiv (rtx, bool);
597 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
598 int);
599 static void cse_insn (rtx);
600 static void cse_prescan_path (struct cse_basic_block_data *);
601 static void invalidate_from_clobbers (rtx);
602 static rtx cse_process_notes (rtx, rtx, bool *);
603 static void cse_extended_basic_block (struct cse_basic_block_data *);
604 static void count_reg_usage (rtx, int *, rtx, int);
605 static int check_for_label_ref (rtx *, void *);
606 extern void dump_class (struct table_elt*);
607 static void get_cse_reg_info_1 (unsigned int regno);
608 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
609 static int check_dependence (rtx *, void *);
611 static void flush_hash_table (void);
612 static bool insn_live_p (rtx, int *);
613 static bool set_live_p (rtx, rtx, int *);
614 static int cse_change_cc_mode (rtx *, void *);
615 static void cse_change_cc_mode_insn (rtx, rtx);
616 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
617 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
618 bool);
621 #undef RTL_HOOKS_GEN_LOWPART
622 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
624 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
626 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
627 virtual regs here because the simplify_*_operation routines are called
628 by integrate.c, which is called before virtual register instantiation. */
630 static bool
631 fixed_base_plus_p (rtx x)
633 switch (GET_CODE (x))
635 case REG:
636 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
637 return true;
638 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
639 return true;
640 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
641 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
642 return true;
643 return false;
645 case PLUS:
646 if (!CONST_INT_P (XEXP (x, 1)))
647 return false;
648 return fixed_base_plus_p (XEXP (x, 0));
650 default:
651 return false;
655 /* Dump the expressions in the equivalence class indicated by CLASSP.
656 This function is used only for debugging. */
657 void
658 dump_class (struct table_elt *classp)
660 struct table_elt *elt;
662 fprintf (stderr, "Equivalence chain for ");
663 print_rtl (stderr, classp->exp);
664 fprintf (stderr, ": \n");
666 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
668 print_rtl (stderr, elt->exp);
669 fprintf (stderr, "\n");
673 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
675 static int
676 approx_reg_cost_1 (rtx *xp, void *data)
678 rtx x = *xp;
679 int *cost_p = (int *) data;
681 if (x && REG_P (x))
683 unsigned int regno = REGNO (x);
685 if (! CHEAP_REGNO (regno))
687 if (regno < FIRST_PSEUDO_REGISTER)
689 if (SMALL_REGISTER_CLASSES)
690 return 1;
691 *cost_p += 2;
693 else
694 *cost_p += 1;
698 return 0;
701 /* Return an estimate of the cost of the registers used in an rtx.
702 This is mostly the number of different REG expressions in the rtx;
703 however for some exceptions like fixed registers we use a cost of
704 0. If any other hard register reference occurs, return MAX_COST. */
706 static int
707 approx_reg_cost (rtx x)
709 int cost = 0;
711 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
712 return MAX_COST;
714 return cost;
717 /* Return a negative value if an rtx A, whose costs are given by COST_A
718 and REGCOST_A, is more desirable than an rtx B.
719 Return a positive value if A is less desirable, or 0 if the two are
720 equally good. */
721 static int
722 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
724 /* First, get rid of cases involving expressions that are entirely
725 unwanted. */
726 if (cost_a != cost_b)
728 if (cost_a == MAX_COST)
729 return 1;
730 if (cost_b == MAX_COST)
731 return -1;
734 /* Avoid extending lifetimes of hardregs. */
735 if (regcost_a != regcost_b)
737 if (regcost_a == MAX_COST)
738 return 1;
739 if (regcost_b == MAX_COST)
740 return -1;
743 /* Normal operation costs take precedence. */
744 if (cost_a != cost_b)
745 return cost_a - cost_b;
746 /* Only if these are identical consider effects on register pressure. */
747 if (regcost_a != regcost_b)
748 return regcost_a - regcost_b;
749 return 0;
752 /* Internal function, to compute cost when X is not a register; called
753 from COST macro to keep it simple. */
755 static int
756 notreg_cost (rtx x, enum rtx_code outer)
758 return ((GET_CODE (x) == SUBREG
759 && REG_P (SUBREG_REG (x))
760 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
761 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
762 && (GET_MODE_SIZE (GET_MODE (x))
763 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
764 && subreg_lowpart_p (x)
765 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (x)),
766 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x)))))
768 : rtx_cost (x, outer, optimize_this_for_speed_p) * 2);
772 /* Initialize CSE_REG_INFO_TABLE. */
774 static void
775 init_cse_reg_info (unsigned int nregs)
777 /* Do we need to grow the table? */
778 if (nregs > cse_reg_info_table_size)
780 unsigned int new_size;
782 if (cse_reg_info_table_size < 2048)
784 /* Compute a new size that is a power of 2 and no smaller
785 than the large of NREGS and 64. */
786 new_size = (cse_reg_info_table_size
787 ? cse_reg_info_table_size : 64);
789 while (new_size < nregs)
790 new_size *= 2;
792 else
794 /* If we need a big table, allocate just enough to hold
795 NREGS registers. */
796 new_size = nregs;
799 /* Reallocate the table with NEW_SIZE entries. */
800 if (cse_reg_info_table)
801 free (cse_reg_info_table);
802 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
803 cse_reg_info_table_size = new_size;
804 cse_reg_info_table_first_uninitialized = 0;
807 /* Do we have all of the first NREGS entries initialized? */
808 if (cse_reg_info_table_first_uninitialized < nregs)
810 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
811 unsigned int i;
813 /* Put the old timestamp on newly allocated entries so that they
814 will all be considered out of date. We do not touch those
815 entries beyond the first NREGS entries to be nice to the
816 virtual memory. */
817 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
818 cse_reg_info_table[i].timestamp = old_timestamp;
820 cse_reg_info_table_first_uninitialized = nregs;
824 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
826 static void
827 get_cse_reg_info_1 (unsigned int regno)
829 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
830 entry will be considered to have been initialized. */
831 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
833 /* Initialize the rest of the entry. */
834 cse_reg_info_table[regno].reg_tick = 1;
835 cse_reg_info_table[regno].reg_in_table = -1;
836 cse_reg_info_table[regno].subreg_ticked = -1;
837 cse_reg_info_table[regno].reg_qty = -regno - 1;
840 /* Find a cse_reg_info entry for REGNO. */
842 static inline struct cse_reg_info *
843 get_cse_reg_info (unsigned int regno)
845 struct cse_reg_info *p = &cse_reg_info_table[regno];
847 /* If this entry has not been initialized, go ahead and initialize
848 it. */
849 if (p->timestamp != cse_reg_info_timestamp)
850 get_cse_reg_info_1 (regno);
852 return p;
855 /* Clear the hash table and initialize each register with its own quantity,
856 for a new basic block. */
858 static void
859 new_basic_block (void)
861 int i;
863 next_qty = 0;
865 /* Invalidate cse_reg_info_table. */
866 cse_reg_info_timestamp++;
868 /* Clear out hash table state for this pass. */
869 CLEAR_HARD_REG_SET (hard_regs_in_table);
871 /* The per-quantity values used to be initialized here, but it is
872 much faster to initialize each as it is made in `make_new_qty'. */
874 for (i = 0; i < HASH_SIZE; i++)
876 struct table_elt *first;
878 first = table[i];
879 if (first != NULL)
881 struct table_elt *last = first;
883 table[i] = NULL;
885 while (last->next_same_hash != NULL)
886 last = last->next_same_hash;
888 /* Now relink this hash entire chain into
889 the free element list. */
891 last->next_same_hash = free_element_chain;
892 free_element_chain = first;
896 #ifdef HAVE_cc0
897 prev_insn_cc0 = 0;
898 #endif
901 /* Say that register REG contains a quantity in mode MODE not in any
902 register before and initialize that quantity. */
904 static void
905 make_new_qty (unsigned int reg, enum machine_mode mode)
907 int q;
908 struct qty_table_elem *ent;
909 struct reg_eqv_elem *eqv;
911 gcc_assert (next_qty < max_qty);
913 q = REG_QTY (reg) = next_qty++;
914 ent = &qty_table[q];
915 ent->first_reg = reg;
916 ent->last_reg = reg;
917 ent->mode = mode;
918 ent->const_rtx = ent->const_insn = NULL_RTX;
919 ent->comparison_code = UNKNOWN;
921 eqv = &reg_eqv_table[reg];
922 eqv->next = eqv->prev = -1;
925 /* Make reg NEW equivalent to reg OLD.
926 OLD is not changing; NEW is. */
928 static void
929 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
931 unsigned int lastr, firstr;
932 int q = REG_QTY (old_reg);
933 struct qty_table_elem *ent;
935 ent = &qty_table[q];
937 /* Nothing should become eqv until it has a "non-invalid" qty number. */
938 gcc_assert (REGNO_QTY_VALID_P (old_reg));
940 REG_QTY (new_reg) = q;
941 firstr = ent->first_reg;
942 lastr = ent->last_reg;
944 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
945 hard regs. Among pseudos, if NEW will live longer than any other reg
946 of the same qty, and that is beyond the current basic block,
947 make it the new canonical replacement for this qty. */
948 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
949 /* Certain fixed registers might be of the class NO_REGS. This means
950 that not only can they not be allocated by the compiler, but
951 they cannot be used in substitutions or canonicalizations
952 either. */
953 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
954 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
955 || (new_reg >= FIRST_PSEUDO_REGISTER
956 && (firstr < FIRST_PSEUDO_REGISTER
957 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
958 && !bitmap_bit_p (cse_ebb_live_out, firstr))
959 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
960 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
962 reg_eqv_table[firstr].prev = new_reg;
963 reg_eqv_table[new_reg].next = firstr;
964 reg_eqv_table[new_reg].prev = -1;
965 ent->first_reg = new_reg;
967 else
969 /* If NEW is a hard reg (known to be non-fixed), insert at end.
970 Otherwise, insert before any non-fixed hard regs that are at the
971 end. Registers of class NO_REGS cannot be used as an
972 equivalent for anything. */
973 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
974 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
975 && new_reg >= FIRST_PSEUDO_REGISTER)
976 lastr = reg_eqv_table[lastr].prev;
977 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
978 if (reg_eqv_table[lastr].next >= 0)
979 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
980 else
981 qty_table[q].last_reg = new_reg;
982 reg_eqv_table[lastr].next = new_reg;
983 reg_eqv_table[new_reg].prev = lastr;
987 /* Remove REG from its equivalence class. */
989 static void
990 delete_reg_equiv (unsigned int reg)
992 struct qty_table_elem *ent;
993 int q = REG_QTY (reg);
994 int p, n;
996 /* If invalid, do nothing. */
997 if (! REGNO_QTY_VALID_P (reg))
998 return;
1000 ent = &qty_table[q];
1002 p = reg_eqv_table[reg].prev;
1003 n = reg_eqv_table[reg].next;
1005 if (n != -1)
1006 reg_eqv_table[n].prev = p;
1007 else
1008 ent->last_reg = p;
1009 if (p != -1)
1010 reg_eqv_table[p].next = n;
1011 else
1012 ent->first_reg = n;
1014 REG_QTY (reg) = -reg - 1;
1017 /* Remove any invalid expressions from the hash table
1018 that refer to any of the registers contained in expression X.
1020 Make sure that newly inserted references to those registers
1021 as subexpressions will be considered valid.
1023 mention_regs is not called when a register itself
1024 is being stored in the table.
1026 Return 1 if we have done something that may have changed the hash code
1027 of X. */
1029 static int
1030 mention_regs (rtx x)
1032 enum rtx_code code;
1033 int i, j;
1034 const char *fmt;
1035 int changed = 0;
1037 if (x == 0)
1038 return 0;
1040 code = GET_CODE (x);
1041 if (code == REG)
1043 unsigned int regno = REGNO (x);
1044 unsigned int endregno = END_REGNO (x);
1045 unsigned int i;
1047 for (i = regno; i < endregno; i++)
1049 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1050 remove_invalid_refs (i);
1052 REG_IN_TABLE (i) = REG_TICK (i);
1053 SUBREG_TICKED (i) = -1;
1056 return 0;
1059 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1060 pseudo if they don't use overlapping words. We handle only pseudos
1061 here for simplicity. */
1062 if (code == SUBREG && REG_P (SUBREG_REG (x))
1063 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1065 unsigned int i = REGNO (SUBREG_REG (x));
1067 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1069 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1070 the last store to this register really stored into this
1071 subreg, then remove the memory of this subreg.
1072 Otherwise, remove any memory of the entire register and
1073 all its subregs from the table. */
1074 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1075 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1076 remove_invalid_refs (i);
1077 else
1078 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1081 REG_IN_TABLE (i) = REG_TICK (i);
1082 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1083 return 0;
1086 /* If X is a comparison or a COMPARE and either operand is a register
1087 that does not have a quantity, give it one. This is so that a later
1088 call to record_jump_equiv won't cause X to be assigned a different
1089 hash code and not found in the table after that call.
1091 It is not necessary to do this here, since rehash_using_reg can
1092 fix up the table later, but doing this here eliminates the need to
1093 call that expensive function in the most common case where the only
1094 use of the register is in the comparison. */
1096 if (code == COMPARE || COMPARISON_P (x))
1098 if (REG_P (XEXP (x, 0))
1099 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1100 if (insert_regs (XEXP (x, 0), NULL, 0))
1102 rehash_using_reg (XEXP (x, 0));
1103 changed = 1;
1106 if (REG_P (XEXP (x, 1))
1107 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1108 if (insert_regs (XEXP (x, 1), NULL, 0))
1110 rehash_using_reg (XEXP (x, 1));
1111 changed = 1;
1115 fmt = GET_RTX_FORMAT (code);
1116 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1117 if (fmt[i] == 'e')
1118 changed |= mention_regs (XEXP (x, i));
1119 else if (fmt[i] == 'E')
1120 for (j = 0; j < XVECLEN (x, i); j++)
1121 changed |= mention_regs (XVECEXP (x, i, j));
1123 return changed;
1126 /* Update the register quantities for inserting X into the hash table
1127 with a value equivalent to CLASSP.
1128 (If the class does not contain a REG, it is irrelevant.)
1129 If MODIFIED is nonzero, X is a destination; it is being modified.
1130 Note that delete_reg_equiv should be called on a register
1131 before insert_regs is done on that register with MODIFIED != 0.
1133 Nonzero value means that elements of reg_qty have changed
1134 so X's hash code may be different. */
1136 static int
1137 insert_regs (rtx x, struct table_elt *classp, int modified)
1139 if (REG_P (x))
1141 unsigned int regno = REGNO (x);
1142 int qty_valid;
1144 /* If REGNO is in the equivalence table already but is of the
1145 wrong mode for that equivalence, don't do anything here. */
1147 qty_valid = REGNO_QTY_VALID_P (regno);
1148 if (qty_valid)
1150 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1152 if (ent->mode != GET_MODE (x))
1153 return 0;
1156 if (modified || ! qty_valid)
1158 if (classp)
1159 for (classp = classp->first_same_value;
1160 classp != 0;
1161 classp = classp->next_same_value)
1162 if (REG_P (classp->exp)
1163 && GET_MODE (classp->exp) == GET_MODE (x))
1165 unsigned c_regno = REGNO (classp->exp);
1167 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1169 /* Suppose that 5 is hard reg and 100 and 101 are
1170 pseudos. Consider
1172 (set (reg:si 100) (reg:si 5))
1173 (set (reg:si 5) (reg:si 100))
1174 (set (reg:di 101) (reg:di 5))
1176 We would now set REG_QTY (101) = REG_QTY (5), but the
1177 entry for 5 is in SImode. When we use this later in
1178 copy propagation, we get the register in wrong mode. */
1179 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1180 continue;
1182 make_regs_eqv (regno, c_regno);
1183 return 1;
1186 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1187 than REG_IN_TABLE to find out if there was only a single preceding
1188 invalidation - for the SUBREG - or another one, which would be
1189 for the full register. However, if we find here that REG_TICK
1190 indicates that the register is invalid, it means that it has
1191 been invalidated in a separate operation. The SUBREG might be used
1192 now (then this is a recursive call), or we might use the full REG
1193 now and a SUBREG of it later. So bump up REG_TICK so that
1194 mention_regs will do the right thing. */
1195 if (! modified
1196 && REG_IN_TABLE (regno) >= 0
1197 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1198 REG_TICK (regno)++;
1199 make_new_qty (regno, GET_MODE (x));
1200 return 1;
1203 return 0;
1206 /* If X is a SUBREG, we will likely be inserting the inner register in the
1207 table. If that register doesn't have an assigned quantity number at
1208 this point but does later, the insertion that we will be doing now will
1209 not be accessible because its hash code will have changed. So assign
1210 a quantity number now. */
1212 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1213 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1215 insert_regs (SUBREG_REG (x), NULL, 0);
1216 mention_regs (x);
1217 return 1;
1219 else
1220 return mention_regs (x);
1224 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1225 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1226 CST is equal to an anchor. */
1228 static bool
1229 compute_const_anchors (rtx cst,
1230 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1231 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1233 HOST_WIDE_INT n = INTVAL (cst);
1235 *lower_base = n & ~(targetm.const_anchor - 1);
1236 if (*lower_base == n)
1237 return false;
1239 *upper_base =
1240 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1241 *upper_offs = n - *upper_base;
1242 *lower_offs = n - *lower_base;
1243 return true;
1246 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1248 static void
1249 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1250 enum machine_mode mode)
1252 struct table_elt *elt;
1253 unsigned hash;
1254 rtx anchor_exp;
1255 rtx exp;
1257 anchor_exp = GEN_INT (anchor);
1258 hash = HASH (anchor_exp, mode);
1259 elt = lookup (anchor_exp, hash, mode);
1260 if (!elt)
1261 elt = insert (anchor_exp, NULL, hash, mode);
1263 exp = plus_constant (reg, offs);
1264 /* REG has just been inserted and the hash codes recomputed. */
1265 mention_regs (exp);
1266 hash = HASH (exp, mode);
1268 /* Use the cost of the register rather than the whole expression. When
1269 looking up constant anchors we will further offset the corresponding
1270 expression therefore it does not make sense to prefer REGs over
1271 reg-immediate additions. Prefer instead the oldest expression. Also
1272 don't prefer pseudos over hard regs so that we derive constants in
1273 argument registers from other argument registers rather than from the
1274 original pseudo that was used to synthesize the constant. */
1275 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1278 /* The constant CST is equivalent to the register REG. Create
1279 equivalences between the two anchors of CST and the corresponding
1280 register-offset expressions using REG. */
1282 static void
1283 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1285 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1287 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1288 &upper_base, &upper_offs))
1289 return;
1291 /* Ignore anchors of value 0. Constants accessible from zero are
1292 simple. */
1293 if (lower_base != 0)
1294 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1296 if (upper_base != 0)
1297 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1300 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1301 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1302 valid expression. Return the cheapest and oldest of such expressions. In
1303 *OLD, return how old the resulting expression is compared to the other
1304 equivalent expressions. */
1306 static rtx
1307 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1308 unsigned *old)
1310 struct table_elt *elt;
1311 unsigned idx;
1312 struct table_elt *match_elt;
1313 rtx match;
1315 /* Find the cheapest and *oldest* expression to maximize the chance of
1316 reusing the same pseudo. */
1318 match_elt = NULL;
1319 match = NULL_RTX;
1320 for (elt = anchor_elt->first_same_value, idx = 0;
1321 elt;
1322 elt = elt->next_same_value, idx++)
1324 if (match_elt && CHEAPER (match_elt, elt))
1325 return match;
1327 if (REG_P (elt->exp)
1328 || (GET_CODE (elt->exp) == PLUS
1329 && REG_P (XEXP (elt->exp, 0))
1330 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1332 rtx x;
1334 /* Ignore expressions that are no longer valid. */
1335 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1336 continue;
1338 x = plus_constant (elt->exp, offs);
1339 if (REG_P (x)
1340 || (GET_CODE (x) == PLUS
1341 && IN_RANGE (INTVAL (XEXP (x, 1)),
1342 -targetm.const_anchor,
1343 targetm.const_anchor - 1)))
1345 match = x;
1346 match_elt = elt;
1347 *old = idx;
1352 return match;
1355 /* Try to express the constant SRC_CONST using a register+offset expression
1356 derived from a constant anchor. Return it if successful or NULL_RTX,
1357 otherwise. */
1359 static rtx
1360 try_const_anchors (rtx src_const, enum machine_mode mode)
1362 struct table_elt *lower_elt, *upper_elt;
1363 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1364 rtx lower_anchor_rtx, upper_anchor_rtx;
1365 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1366 unsigned lower_old, upper_old;
1368 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1369 &upper_base, &upper_offs))
1370 return NULL_RTX;
1372 lower_anchor_rtx = GEN_INT (lower_base);
1373 upper_anchor_rtx = GEN_INT (upper_base);
1374 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1375 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1377 if (lower_elt)
1378 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1379 if (upper_elt)
1380 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1382 if (!lower_exp)
1383 return upper_exp;
1384 if (!upper_exp)
1385 return lower_exp;
1387 /* Return the older expression. */
1388 return (upper_old > lower_old ? upper_exp : lower_exp);
1391 /* Look in or update the hash table. */
1393 /* Remove table element ELT from use in the table.
1394 HASH is its hash code, made using the HASH macro.
1395 It's an argument because often that is known in advance
1396 and we save much time not recomputing it. */
1398 static void
1399 remove_from_table (struct table_elt *elt, unsigned int hash)
1401 if (elt == 0)
1402 return;
1404 /* Mark this element as removed. See cse_insn. */
1405 elt->first_same_value = 0;
1407 /* Remove the table element from its equivalence class. */
1410 struct table_elt *prev = elt->prev_same_value;
1411 struct table_elt *next = elt->next_same_value;
1413 if (next)
1414 next->prev_same_value = prev;
1416 if (prev)
1417 prev->next_same_value = next;
1418 else
1420 struct table_elt *newfirst = next;
1421 while (next)
1423 next->first_same_value = newfirst;
1424 next = next->next_same_value;
1429 /* Remove the table element from its hash bucket. */
1432 struct table_elt *prev = elt->prev_same_hash;
1433 struct table_elt *next = elt->next_same_hash;
1435 if (next)
1436 next->prev_same_hash = prev;
1438 if (prev)
1439 prev->next_same_hash = next;
1440 else if (table[hash] == elt)
1441 table[hash] = next;
1442 else
1444 /* This entry is not in the proper hash bucket. This can happen
1445 when two classes were merged by `merge_equiv_classes'. Search
1446 for the hash bucket that it heads. This happens only very
1447 rarely, so the cost is acceptable. */
1448 for (hash = 0; hash < HASH_SIZE; hash++)
1449 if (table[hash] == elt)
1450 table[hash] = next;
1454 /* Remove the table element from its related-value circular chain. */
1456 if (elt->related_value != 0 && elt->related_value != elt)
1458 struct table_elt *p = elt->related_value;
1460 while (p->related_value != elt)
1461 p = p->related_value;
1462 p->related_value = elt->related_value;
1463 if (p->related_value == p)
1464 p->related_value = 0;
1467 /* Now add it to the free element chain. */
1468 elt->next_same_hash = free_element_chain;
1469 free_element_chain = elt;
1472 /* Same as above, but X is a pseudo-register. */
1474 static void
1475 remove_pseudo_from_table (rtx x, unsigned int hash)
1477 struct table_elt *elt;
1479 /* Because a pseudo-register can be referenced in more than one
1480 mode, we might have to remove more than one table entry. */
1481 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1482 remove_from_table (elt, hash);
1485 /* Look up X in the hash table and return its table element,
1486 or 0 if X is not in the table.
1488 MODE is the machine-mode of X, or if X is an integer constant
1489 with VOIDmode then MODE is the mode with which X will be used.
1491 Here we are satisfied to find an expression whose tree structure
1492 looks like X. */
1494 static struct table_elt *
1495 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1497 struct table_elt *p;
1499 for (p = table[hash]; p; p = p->next_same_hash)
1500 if (mode == p->mode && ((x == p->exp && REG_P (x))
1501 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1502 return p;
1504 return 0;
1507 /* Like `lookup' but don't care whether the table element uses invalid regs.
1508 Also ignore discrepancies in the machine mode of a register. */
1510 static struct table_elt *
1511 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1513 struct table_elt *p;
1515 if (REG_P (x))
1517 unsigned int regno = REGNO (x);
1519 /* Don't check the machine mode when comparing registers;
1520 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1521 for (p = table[hash]; p; p = p->next_same_hash)
1522 if (REG_P (p->exp)
1523 && REGNO (p->exp) == regno)
1524 return p;
1526 else
1528 for (p = table[hash]; p; p = p->next_same_hash)
1529 if (mode == p->mode
1530 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1531 return p;
1534 return 0;
1537 /* Look for an expression equivalent to X and with code CODE.
1538 If one is found, return that expression. */
1540 static rtx
1541 lookup_as_function (rtx x, enum rtx_code code)
1543 struct table_elt *p
1544 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1546 if (p == 0)
1547 return 0;
1549 for (p = p->first_same_value; p; p = p->next_same_value)
1550 if (GET_CODE (p->exp) == code
1551 /* Make sure this is a valid entry in the table. */
1552 && exp_equiv_p (p->exp, p->exp, 1, false))
1553 return p->exp;
1555 return 0;
1558 /* Insert X in the hash table, assuming HASH is its hash code and
1559 CLASSP is an element of the class it should go in (or 0 if a new
1560 class should be made). COST is the code of X and reg_cost is the
1561 cost of registers in X. It is inserted at the proper position to
1562 keep the class in the order cheapest first.
1564 MODE is the machine-mode of X, or if X is an integer constant
1565 with VOIDmode then MODE is the mode with which X will be used.
1567 For elements of equal cheapness, the most recent one
1568 goes in front, except that the first element in the list
1569 remains first unless a cheaper element is added. The order of
1570 pseudo-registers does not matter, as canon_reg will be called to
1571 find the cheapest when a register is retrieved from the table.
1573 The in_memory field in the hash table element is set to 0.
1574 The caller must set it nonzero if appropriate.
1576 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1577 and if insert_regs returns a nonzero value
1578 you must then recompute its hash code before calling here.
1580 If necessary, update table showing constant values of quantities. */
1582 static struct table_elt *
1583 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1584 enum machine_mode mode, int cost, int reg_cost)
1586 struct table_elt *elt;
1588 /* If X is a register and we haven't made a quantity for it,
1589 something is wrong. */
1590 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1592 /* If X is a hard register, show it is being put in the table. */
1593 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1594 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1596 /* Put an element for X into the right hash bucket. */
1598 elt = free_element_chain;
1599 if (elt)
1600 free_element_chain = elt->next_same_hash;
1601 else
1602 elt = XNEW (struct table_elt);
1604 elt->exp = x;
1605 elt->canon_exp = NULL_RTX;
1606 elt->cost = cost;
1607 elt->regcost = reg_cost;
1608 elt->next_same_value = 0;
1609 elt->prev_same_value = 0;
1610 elt->next_same_hash = table[hash];
1611 elt->prev_same_hash = 0;
1612 elt->related_value = 0;
1613 elt->in_memory = 0;
1614 elt->mode = mode;
1615 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1617 if (table[hash])
1618 table[hash]->prev_same_hash = elt;
1619 table[hash] = elt;
1621 /* Put it into the proper value-class. */
1622 if (classp)
1624 classp = classp->first_same_value;
1625 if (CHEAPER (elt, classp))
1626 /* Insert at the head of the class. */
1628 struct table_elt *p;
1629 elt->next_same_value = classp;
1630 classp->prev_same_value = elt;
1631 elt->first_same_value = elt;
1633 for (p = classp; p; p = p->next_same_value)
1634 p->first_same_value = elt;
1636 else
1638 /* Insert not at head of the class. */
1639 /* Put it after the last element cheaper than X. */
1640 struct table_elt *p, *next;
1642 for (p = classp; (next = p->next_same_value) && CHEAPER (next, elt);
1643 p = next);
1645 /* Put it after P and before NEXT. */
1646 elt->next_same_value = next;
1647 if (next)
1648 next->prev_same_value = elt;
1650 elt->prev_same_value = p;
1651 p->next_same_value = elt;
1652 elt->first_same_value = classp;
1655 else
1656 elt->first_same_value = elt;
1658 /* If this is a constant being set equivalent to a register or a register
1659 being set equivalent to a constant, note the constant equivalence.
1661 If this is a constant, it cannot be equivalent to a different constant,
1662 and a constant is the only thing that can be cheaper than a register. So
1663 we know the register is the head of the class (before the constant was
1664 inserted).
1666 If this is a register that is not already known equivalent to a
1667 constant, we must check the entire class.
1669 If this is a register that is already known equivalent to an insn,
1670 update the qtys `const_insn' to show that `this_insn' is the latest
1671 insn making that quantity equivalent to the constant. */
1673 if (elt->is_const && classp && REG_P (classp->exp)
1674 && !REG_P (x))
1676 int exp_q = REG_QTY (REGNO (classp->exp));
1677 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1679 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1680 exp_ent->const_insn = this_insn;
1683 else if (REG_P (x)
1684 && classp
1685 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1686 && ! elt->is_const)
1688 struct table_elt *p;
1690 for (p = classp; p != 0; p = p->next_same_value)
1692 if (p->is_const && !REG_P (p->exp))
1694 int x_q = REG_QTY (REGNO (x));
1695 struct qty_table_elem *x_ent = &qty_table[x_q];
1697 x_ent->const_rtx
1698 = gen_lowpart (GET_MODE (x), p->exp);
1699 x_ent->const_insn = this_insn;
1700 break;
1705 else if (REG_P (x)
1706 && qty_table[REG_QTY (REGNO (x))].const_rtx
1707 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1708 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1710 /* If this is a constant with symbolic value,
1711 and it has a term with an explicit integer value,
1712 link it up with related expressions. */
1713 if (GET_CODE (x) == CONST)
1715 rtx subexp = get_related_value (x);
1716 unsigned subhash;
1717 struct table_elt *subelt, *subelt_prev;
1719 if (subexp != 0)
1721 /* Get the integer-free subexpression in the hash table. */
1722 subhash = SAFE_HASH (subexp, mode);
1723 subelt = lookup (subexp, subhash, mode);
1724 if (subelt == 0)
1725 subelt = insert (subexp, NULL, subhash, mode);
1726 /* Initialize SUBELT's circular chain if it has none. */
1727 if (subelt->related_value == 0)
1728 subelt->related_value = subelt;
1729 /* Find the element in the circular chain that precedes SUBELT. */
1730 subelt_prev = subelt;
1731 while (subelt_prev->related_value != subelt)
1732 subelt_prev = subelt_prev->related_value;
1733 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1734 This way the element that follows SUBELT is the oldest one. */
1735 elt->related_value = subelt_prev->related_value;
1736 subelt_prev->related_value = elt;
1740 return elt;
1743 /* Wrap insert_with_costs by passing the default costs. */
1745 static struct table_elt *
1746 insert (rtx x, struct table_elt *classp, unsigned int hash,
1747 enum machine_mode mode)
1749 return
1750 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1754 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1755 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1756 the two classes equivalent.
1758 CLASS1 will be the surviving class; CLASS2 should not be used after this
1759 call.
1761 Any invalid entries in CLASS2 will not be copied. */
1763 static void
1764 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1766 struct table_elt *elt, *next, *new_elt;
1768 /* Ensure we start with the head of the classes. */
1769 class1 = class1->first_same_value;
1770 class2 = class2->first_same_value;
1772 /* If they were already equal, forget it. */
1773 if (class1 == class2)
1774 return;
1776 for (elt = class2; elt; elt = next)
1778 unsigned int hash;
1779 rtx exp = elt->exp;
1780 enum machine_mode mode = elt->mode;
1782 next = elt->next_same_value;
1784 /* Remove old entry, make a new one in CLASS1's class.
1785 Don't do this for invalid entries as we cannot find their
1786 hash code (it also isn't necessary). */
1787 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1789 bool need_rehash = false;
1791 hash_arg_in_memory = 0;
1792 hash = HASH (exp, mode);
1794 if (REG_P (exp))
1796 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1797 delete_reg_equiv (REGNO (exp));
1800 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1801 remove_pseudo_from_table (exp, hash);
1802 else
1803 remove_from_table (elt, hash);
1805 if (insert_regs (exp, class1, 0) || need_rehash)
1807 rehash_using_reg (exp);
1808 hash = HASH (exp, mode);
1810 new_elt = insert (exp, class1, hash, mode);
1811 new_elt->in_memory = hash_arg_in_memory;
1816 /* Flush the entire hash table. */
1818 static void
1819 flush_hash_table (void)
1821 int i;
1822 struct table_elt *p;
1824 for (i = 0; i < HASH_SIZE; i++)
1825 for (p = table[i]; p; p = table[i])
1827 /* Note that invalidate can remove elements
1828 after P in the current hash chain. */
1829 if (REG_P (p->exp))
1830 invalidate (p->exp, VOIDmode);
1831 else
1832 remove_from_table (p, i);
1836 /* Function called for each rtx to check whether true dependence exist. */
1837 struct check_dependence_data
1839 enum machine_mode mode;
1840 rtx exp;
1841 rtx addr;
1844 static int
1845 check_dependence (rtx *x, void *data)
1847 struct check_dependence_data *d = (struct check_dependence_data *) data;
1848 if (*x && MEM_P (*x))
1849 return canon_true_dependence (d->exp, d->mode, d->addr, *x, NULL_RTX,
1850 cse_rtx_varies_p);
1851 else
1852 return 0;
1855 /* Remove from the hash table, or mark as invalid, all expressions whose
1856 values could be altered by storing in X. X is a register, a subreg, or
1857 a memory reference with nonvarying address (because, when a memory
1858 reference with a varying address is stored in, all memory references are
1859 removed by invalidate_memory so specific invalidation is superfluous).
1860 FULL_MODE, if not VOIDmode, indicates that this much should be
1861 invalidated instead of just the amount indicated by the mode of X. This
1862 is only used for bitfield stores into memory.
1864 A nonvarying address may be just a register or just a symbol reference,
1865 or it may be either of those plus a numeric offset. */
1867 static void
1868 invalidate (rtx x, enum machine_mode full_mode)
1870 int i;
1871 struct table_elt *p;
1872 rtx addr;
1874 switch (GET_CODE (x))
1876 case REG:
1878 /* If X is a register, dependencies on its contents are recorded
1879 through the qty number mechanism. Just change the qty number of
1880 the register, mark it as invalid for expressions that refer to it,
1881 and remove it itself. */
1882 unsigned int regno = REGNO (x);
1883 unsigned int hash = HASH (x, GET_MODE (x));
1885 /* Remove REGNO from any quantity list it might be on and indicate
1886 that its value might have changed. If it is a pseudo, remove its
1887 entry from the hash table.
1889 For a hard register, we do the first two actions above for any
1890 additional hard registers corresponding to X. Then, if any of these
1891 registers are in the table, we must remove any REG entries that
1892 overlap these registers. */
1894 delete_reg_equiv (regno);
1895 REG_TICK (regno)++;
1896 SUBREG_TICKED (regno) = -1;
1898 if (regno >= FIRST_PSEUDO_REGISTER)
1899 remove_pseudo_from_table (x, hash);
1900 else
1902 HOST_WIDE_INT in_table
1903 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1904 unsigned int endregno = END_HARD_REGNO (x);
1905 unsigned int tregno, tendregno, rn;
1906 struct table_elt *p, *next;
1908 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1910 for (rn = regno + 1; rn < endregno; rn++)
1912 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1913 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1914 delete_reg_equiv (rn);
1915 REG_TICK (rn)++;
1916 SUBREG_TICKED (rn) = -1;
1919 if (in_table)
1920 for (hash = 0; hash < HASH_SIZE; hash++)
1921 for (p = table[hash]; p; p = next)
1923 next = p->next_same_hash;
1925 if (!REG_P (p->exp)
1926 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1927 continue;
1929 tregno = REGNO (p->exp);
1930 tendregno = END_HARD_REGNO (p->exp);
1931 if (tendregno > regno && tregno < endregno)
1932 remove_from_table (p, hash);
1936 return;
1938 case SUBREG:
1939 invalidate (SUBREG_REG (x), VOIDmode);
1940 return;
1942 case PARALLEL:
1943 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1944 invalidate (XVECEXP (x, 0, i), VOIDmode);
1945 return;
1947 case EXPR_LIST:
1948 /* This is part of a disjoint return value; extract the location in
1949 question ignoring the offset. */
1950 invalidate (XEXP (x, 0), VOIDmode);
1951 return;
1953 case MEM:
1954 addr = canon_rtx (get_addr (XEXP (x, 0)));
1955 /* Calculate the canonical version of X here so that
1956 true_dependence doesn't generate new RTL for X on each call. */
1957 x = canon_rtx (x);
1959 /* Remove all hash table elements that refer to overlapping pieces of
1960 memory. */
1961 if (full_mode == VOIDmode)
1962 full_mode = GET_MODE (x);
1964 for (i = 0; i < HASH_SIZE; i++)
1966 struct table_elt *next;
1968 for (p = table[i]; p; p = next)
1970 next = p->next_same_hash;
1971 if (p->in_memory)
1973 struct check_dependence_data d;
1975 /* Just canonicalize the expression once;
1976 otherwise each time we call invalidate
1977 true_dependence will canonicalize the
1978 expression again. */
1979 if (!p->canon_exp)
1980 p->canon_exp = canon_rtx (p->exp);
1981 d.exp = x;
1982 d.addr = addr;
1983 d.mode = full_mode;
1984 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1985 remove_from_table (p, i);
1989 return;
1991 default:
1992 gcc_unreachable ();
1996 /* Remove all expressions that refer to register REGNO,
1997 since they are already invalid, and we are about to
1998 mark that register valid again and don't want the old
1999 expressions to reappear as valid. */
2001 static void
2002 remove_invalid_refs (unsigned int regno)
2004 unsigned int i;
2005 struct table_elt *p, *next;
2007 for (i = 0; i < HASH_SIZE; i++)
2008 for (p = table[i]; p; p = next)
2010 next = p->next_same_hash;
2011 if (!REG_P (p->exp)
2012 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2013 remove_from_table (p, i);
2017 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2018 and mode MODE. */
2019 static void
2020 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2021 enum machine_mode mode)
2023 unsigned int i;
2024 struct table_elt *p, *next;
2025 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2027 for (i = 0; i < HASH_SIZE; i++)
2028 for (p = table[i]; p; p = next)
2030 rtx exp = p->exp;
2031 next = p->next_same_hash;
2033 if (!REG_P (exp)
2034 && (GET_CODE (exp) != SUBREG
2035 || !REG_P (SUBREG_REG (exp))
2036 || REGNO (SUBREG_REG (exp)) != regno
2037 || (((SUBREG_BYTE (exp)
2038 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2039 && SUBREG_BYTE (exp) <= end))
2040 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2041 remove_from_table (p, i);
2045 /* Recompute the hash codes of any valid entries in the hash table that
2046 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2048 This is called when we make a jump equivalence. */
2050 static void
2051 rehash_using_reg (rtx x)
2053 unsigned int i;
2054 struct table_elt *p, *next;
2055 unsigned hash;
2057 if (GET_CODE (x) == SUBREG)
2058 x = SUBREG_REG (x);
2060 /* If X is not a register or if the register is known not to be in any
2061 valid entries in the table, we have no work to do. */
2063 if (!REG_P (x)
2064 || REG_IN_TABLE (REGNO (x)) < 0
2065 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2066 return;
2068 /* Scan all hash chains looking for valid entries that mention X.
2069 If we find one and it is in the wrong hash chain, move it. */
2071 for (i = 0; i < HASH_SIZE; i++)
2072 for (p = table[i]; p; p = next)
2074 next = p->next_same_hash;
2075 if (reg_mentioned_p (x, p->exp)
2076 && exp_equiv_p (p->exp, p->exp, 1, false)
2077 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2079 if (p->next_same_hash)
2080 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2082 if (p->prev_same_hash)
2083 p->prev_same_hash->next_same_hash = p->next_same_hash;
2084 else
2085 table[i] = p->next_same_hash;
2087 p->next_same_hash = table[hash];
2088 p->prev_same_hash = 0;
2089 if (table[hash])
2090 table[hash]->prev_same_hash = p;
2091 table[hash] = p;
2096 /* Remove from the hash table any expression that is a call-clobbered
2097 register. Also update their TICK values. */
2099 static void
2100 invalidate_for_call (void)
2102 unsigned int regno, endregno;
2103 unsigned int i;
2104 unsigned hash;
2105 struct table_elt *p, *next;
2106 int in_table = 0;
2108 /* Go through all the hard registers. For each that is clobbered in
2109 a CALL_INSN, remove the register from quantity chains and update
2110 reg_tick if defined. Also see if any of these registers is currently
2111 in the table. */
2113 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2114 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2116 delete_reg_equiv (regno);
2117 if (REG_TICK (regno) >= 0)
2119 REG_TICK (regno)++;
2120 SUBREG_TICKED (regno) = -1;
2123 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2126 /* In the case where we have no call-clobbered hard registers in the
2127 table, we are done. Otherwise, scan the table and remove any
2128 entry that overlaps a call-clobbered register. */
2130 if (in_table)
2131 for (hash = 0; hash < HASH_SIZE; hash++)
2132 for (p = table[hash]; p; p = next)
2134 next = p->next_same_hash;
2136 if (!REG_P (p->exp)
2137 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2138 continue;
2140 regno = REGNO (p->exp);
2141 endregno = END_HARD_REGNO (p->exp);
2143 for (i = regno; i < endregno; i++)
2144 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2146 remove_from_table (p, hash);
2147 break;
2152 /* Given an expression X of type CONST,
2153 and ELT which is its table entry (or 0 if it
2154 is not in the hash table),
2155 return an alternate expression for X as a register plus integer.
2156 If none can be found, return 0. */
2158 static rtx
2159 use_related_value (rtx x, struct table_elt *elt)
2161 struct table_elt *relt = 0;
2162 struct table_elt *p, *q;
2163 HOST_WIDE_INT offset;
2165 /* First, is there anything related known?
2166 If we have a table element, we can tell from that.
2167 Otherwise, must look it up. */
2169 if (elt != 0 && elt->related_value != 0)
2170 relt = elt;
2171 else if (elt == 0 && GET_CODE (x) == CONST)
2173 rtx subexp = get_related_value (x);
2174 if (subexp != 0)
2175 relt = lookup (subexp,
2176 SAFE_HASH (subexp, GET_MODE (subexp)),
2177 GET_MODE (subexp));
2180 if (relt == 0)
2181 return 0;
2183 /* Search all related table entries for one that has an
2184 equivalent register. */
2186 p = relt;
2187 while (1)
2189 /* This loop is strange in that it is executed in two different cases.
2190 The first is when X is already in the table. Then it is searching
2191 the RELATED_VALUE list of X's class (RELT). The second case is when
2192 X is not in the table. Then RELT points to a class for the related
2193 value.
2195 Ensure that, whatever case we are in, that we ignore classes that have
2196 the same value as X. */
2198 if (rtx_equal_p (x, p->exp))
2199 q = 0;
2200 else
2201 for (q = p->first_same_value; q; q = q->next_same_value)
2202 if (REG_P (q->exp))
2203 break;
2205 if (q)
2206 break;
2208 p = p->related_value;
2210 /* We went all the way around, so there is nothing to be found.
2211 Alternatively, perhaps RELT was in the table for some other reason
2212 and it has no related values recorded. */
2213 if (p == relt || p == 0)
2214 break;
2217 if (q == 0)
2218 return 0;
2220 offset = (get_integer_term (x) - get_integer_term (p->exp));
2221 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2222 return plus_constant (q->exp, offset);
2226 /* Hash a string. Just add its bytes up. */
2227 static inline unsigned
2228 hash_rtx_string (const char *ps)
2230 unsigned hash = 0;
2231 const unsigned char *p = (const unsigned char *) ps;
2233 if (p)
2234 while (*p)
2235 hash += *p++;
2237 return hash;
2240 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2241 When the callback returns true, we continue with the new rtx. */
2243 unsigned
2244 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2245 int *do_not_record_p, int *hash_arg_in_memory_p,
2246 bool have_reg_qty, hash_rtx_callback_function cb)
2248 int i, j;
2249 unsigned hash = 0;
2250 enum rtx_code code;
2251 const char *fmt;
2252 enum machine_mode newmode;
2253 rtx newx;
2255 /* Used to turn recursion into iteration. We can't rely on GCC's
2256 tail-recursion elimination since we need to keep accumulating values
2257 in HASH. */
2258 repeat:
2259 if (x == 0)
2260 return hash;
2262 /* Invoke the callback first. */
2263 if (cb != NULL
2264 && ((*cb) (x, mode, &newx, &newmode)))
2266 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2267 hash_arg_in_memory_p, have_reg_qty, cb);
2268 return hash;
2271 code = GET_CODE (x);
2272 switch (code)
2274 case REG:
2276 unsigned int regno = REGNO (x);
2278 if (do_not_record_p && !reload_completed)
2280 /* On some machines, we can't record any non-fixed hard register,
2281 because extending its life will cause reload problems. We
2282 consider ap, fp, sp, gp to be fixed for this purpose.
2284 We also consider CCmode registers to be fixed for this purpose;
2285 failure to do so leads to failure to simplify 0<100 type of
2286 conditionals.
2288 On all machines, we can't record any global registers.
2289 Nor should we record any register that is in a small
2290 class, as defined by CLASS_LIKELY_SPILLED_P. */
2291 bool record;
2293 if (regno >= FIRST_PSEUDO_REGISTER)
2294 record = true;
2295 else if (x == frame_pointer_rtx
2296 || x == hard_frame_pointer_rtx
2297 || x == arg_pointer_rtx
2298 || x == stack_pointer_rtx
2299 || x == pic_offset_table_rtx)
2300 record = true;
2301 else if (global_regs[regno])
2302 record = false;
2303 else if (fixed_regs[regno])
2304 record = true;
2305 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2306 record = true;
2307 else if (SMALL_REGISTER_CLASSES)
2308 record = false;
2309 else if (CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno)))
2310 record = false;
2311 else
2312 record = true;
2314 if (!record)
2316 *do_not_record_p = 1;
2317 return 0;
2321 hash += ((unsigned int) REG << 7);
2322 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2323 return hash;
2326 /* We handle SUBREG of a REG specially because the underlying
2327 reg changes its hash value with every value change; we don't
2328 want to have to forget unrelated subregs when one subreg changes. */
2329 case SUBREG:
2331 if (REG_P (SUBREG_REG (x)))
2333 hash += (((unsigned int) SUBREG << 7)
2334 + REGNO (SUBREG_REG (x))
2335 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2336 return hash;
2338 break;
2341 case CONST_INT:
2342 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2343 + (unsigned int) INTVAL (x));
2344 return hash;
2346 case CONST_DOUBLE:
2347 /* This is like the general case, except that it only counts
2348 the integers representing the constant. */
2349 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2350 if (GET_MODE (x) != VOIDmode)
2351 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2352 else
2353 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2354 + (unsigned int) CONST_DOUBLE_HIGH (x));
2355 return hash;
2357 case CONST_FIXED:
2358 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2359 hash += fixed_hash (CONST_FIXED_VALUE (x));
2360 return hash;
2362 case CONST_VECTOR:
2364 int units;
2365 rtx elt;
2367 units = CONST_VECTOR_NUNITS (x);
2369 for (i = 0; i < units; ++i)
2371 elt = CONST_VECTOR_ELT (x, i);
2372 hash += hash_rtx_cb (elt, GET_MODE (elt),
2373 do_not_record_p, hash_arg_in_memory_p,
2374 have_reg_qty, cb);
2377 return hash;
2380 /* Assume there is only one rtx object for any given label. */
2381 case LABEL_REF:
2382 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2383 differences and differences between each stage's debugging dumps. */
2384 hash += (((unsigned int) LABEL_REF << 7)
2385 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2386 return hash;
2388 case SYMBOL_REF:
2390 /* Don't hash on the symbol's address to avoid bootstrap differences.
2391 Different hash values may cause expressions to be recorded in
2392 different orders and thus different registers to be used in the
2393 final assembler. This also avoids differences in the dump files
2394 between various stages. */
2395 unsigned int h = 0;
2396 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2398 while (*p)
2399 h += (h << 7) + *p++; /* ??? revisit */
2401 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2402 return hash;
2405 case MEM:
2406 /* We don't record if marked volatile or if BLKmode since we don't
2407 know the size of the move. */
2408 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2410 *do_not_record_p = 1;
2411 return 0;
2413 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2414 *hash_arg_in_memory_p = 1;
2416 /* Now that we have already found this special case,
2417 might as well speed it up as much as possible. */
2418 hash += (unsigned) MEM;
2419 x = XEXP (x, 0);
2420 goto repeat;
2422 case USE:
2423 /* A USE that mentions non-volatile memory needs special
2424 handling since the MEM may be BLKmode which normally
2425 prevents an entry from being made. Pure calls are
2426 marked by a USE which mentions BLKmode memory.
2427 See calls.c:emit_call_1. */
2428 if (MEM_P (XEXP (x, 0))
2429 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2431 hash += (unsigned) USE;
2432 x = XEXP (x, 0);
2434 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2435 *hash_arg_in_memory_p = 1;
2437 /* Now that we have already found this special case,
2438 might as well speed it up as much as possible. */
2439 hash += (unsigned) MEM;
2440 x = XEXP (x, 0);
2441 goto repeat;
2443 break;
2445 case PRE_DEC:
2446 case PRE_INC:
2447 case POST_DEC:
2448 case POST_INC:
2449 case PRE_MODIFY:
2450 case POST_MODIFY:
2451 case PC:
2452 case CC0:
2453 case CALL:
2454 case UNSPEC_VOLATILE:
2455 if (do_not_record_p) {
2456 *do_not_record_p = 1;
2457 return 0;
2459 else
2460 return hash;
2461 break;
2463 case ASM_OPERANDS:
2464 if (do_not_record_p && MEM_VOLATILE_P (x))
2466 *do_not_record_p = 1;
2467 return 0;
2469 else
2471 /* We don't want to take the filename and line into account. */
2472 hash += (unsigned) code + (unsigned) GET_MODE (x)
2473 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2474 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2475 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2477 if (ASM_OPERANDS_INPUT_LENGTH (x))
2479 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2481 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2482 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2483 do_not_record_p, hash_arg_in_memory_p,
2484 have_reg_qty, cb)
2485 + hash_rtx_string
2486 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2489 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2490 x = ASM_OPERANDS_INPUT (x, 0);
2491 mode = GET_MODE (x);
2492 goto repeat;
2495 return hash;
2497 break;
2499 default:
2500 break;
2503 i = GET_RTX_LENGTH (code) - 1;
2504 hash += (unsigned) code + (unsigned) GET_MODE (x);
2505 fmt = GET_RTX_FORMAT (code);
2506 for (; i >= 0; i--)
2508 switch (fmt[i])
2510 case 'e':
2511 /* If we are about to do the last recursive call
2512 needed at this level, change it into iteration.
2513 This function is called enough to be worth it. */
2514 if (i == 0)
2516 x = XEXP (x, i);
2517 goto repeat;
2520 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2521 hash_arg_in_memory_p,
2522 have_reg_qty, cb);
2523 break;
2525 case 'E':
2526 for (j = 0; j < XVECLEN (x, i); j++)
2527 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2528 hash_arg_in_memory_p,
2529 have_reg_qty, cb);
2530 break;
2532 case 's':
2533 hash += hash_rtx_string (XSTR (x, i));
2534 break;
2536 case 'i':
2537 hash += (unsigned int) XINT (x, i);
2538 break;
2540 case '0': case 't':
2541 /* Unused. */
2542 break;
2544 default:
2545 gcc_unreachable ();
2549 return hash;
2552 /* Hash an rtx. We are careful to make sure the value is never negative.
2553 Equivalent registers hash identically.
2554 MODE is used in hashing for CONST_INTs only;
2555 otherwise the mode of X is used.
2557 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2559 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2560 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2562 Note that cse_insn knows that the hash code of a MEM expression
2563 is just (int) MEM plus the hash code of the address. */
2565 unsigned
2566 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2567 int *hash_arg_in_memory_p, bool have_reg_qty)
2569 return hash_rtx_cb (x, mode, do_not_record_p,
2570 hash_arg_in_memory_p, have_reg_qty, NULL);
2573 /* Hash an rtx X for cse via hash_rtx.
2574 Stores 1 in do_not_record if any subexpression is volatile.
2575 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2576 does not have the RTX_UNCHANGING_P bit set. */
2578 static inline unsigned
2579 canon_hash (rtx x, enum machine_mode mode)
2581 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2584 /* Like canon_hash but with no side effects, i.e. do_not_record
2585 and hash_arg_in_memory are not changed. */
2587 static inline unsigned
2588 safe_hash (rtx x, enum machine_mode mode)
2590 int dummy_do_not_record;
2591 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2594 /* Return 1 iff X and Y would canonicalize into the same thing,
2595 without actually constructing the canonicalization of either one.
2596 If VALIDATE is nonzero,
2597 we assume X is an expression being processed from the rtl
2598 and Y was found in the hash table. We check register refs
2599 in Y for being marked as valid.
2601 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2604 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2606 int i, j;
2607 enum rtx_code code;
2608 const char *fmt;
2610 /* Note: it is incorrect to assume an expression is equivalent to itself
2611 if VALIDATE is nonzero. */
2612 if (x == y && !validate)
2613 return 1;
2615 if (x == 0 || y == 0)
2616 return x == y;
2618 code = GET_CODE (x);
2619 if (code != GET_CODE (y))
2620 return 0;
2622 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2623 if (GET_MODE (x) != GET_MODE (y))
2624 return 0;
2626 switch (code)
2628 case PC:
2629 case CC0:
2630 case CONST_INT:
2631 case CONST_DOUBLE:
2632 case CONST_FIXED:
2633 return x == y;
2635 case LABEL_REF:
2636 return XEXP (x, 0) == XEXP (y, 0);
2638 case SYMBOL_REF:
2639 return XSTR (x, 0) == XSTR (y, 0);
2641 case REG:
2642 if (for_gcse)
2643 return REGNO (x) == REGNO (y);
2644 else
2646 unsigned int regno = REGNO (y);
2647 unsigned int i;
2648 unsigned int endregno = END_REGNO (y);
2650 /* If the quantities are not the same, the expressions are not
2651 equivalent. If there are and we are not to validate, they
2652 are equivalent. Otherwise, ensure all regs are up-to-date. */
2654 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2655 return 0;
2657 if (! validate)
2658 return 1;
2660 for (i = regno; i < endregno; i++)
2661 if (REG_IN_TABLE (i) != REG_TICK (i))
2662 return 0;
2664 return 1;
2667 case MEM:
2668 if (for_gcse)
2670 /* A volatile mem should not be considered equivalent to any
2671 other. */
2672 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2673 return 0;
2675 /* Can't merge two expressions in different alias sets, since we
2676 can decide that the expression is transparent in a block when
2677 it isn't, due to it being set with the different alias set.
2679 Also, can't merge two expressions with different MEM_ATTRS.
2680 They could e.g. be two different entities allocated into the
2681 same space on the stack (see e.g. PR25130). In that case, the
2682 MEM addresses can be the same, even though the two MEMs are
2683 absolutely not equivalent.
2685 But because really all MEM attributes should be the same for
2686 equivalent MEMs, we just use the invariant that MEMs that have
2687 the same attributes share the same mem_attrs data structure. */
2688 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2689 return 0;
2691 break;
2693 /* For commutative operations, check both orders. */
2694 case PLUS:
2695 case MULT:
2696 case AND:
2697 case IOR:
2698 case XOR:
2699 case NE:
2700 case EQ:
2701 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2702 validate, for_gcse)
2703 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2704 validate, for_gcse))
2705 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2706 validate, for_gcse)
2707 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2708 validate, for_gcse)));
2710 case ASM_OPERANDS:
2711 /* We don't use the generic code below because we want to
2712 disregard filename and line numbers. */
2714 /* A volatile asm isn't equivalent to any other. */
2715 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2716 return 0;
2718 if (GET_MODE (x) != GET_MODE (y)
2719 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2720 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2721 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2722 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2723 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2724 return 0;
2726 if (ASM_OPERANDS_INPUT_LENGTH (x))
2728 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2729 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2730 ASM_OPERANDS_INPUT (y, i),
2731 validate, for_gcse)
2732 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2733 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2734 return 0;
2737 return 1;
2739 default:
2740 break;
2743 /* Compare the elements. If any pair of corresponding elements
2744 fail to match, return 0 for the whole thing. */
2746 fmt = GET_RTX_FORMAT (code);
2747 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2749 switch (fmt[i])
2751 case 'e':
2752 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2753 validate, for_gcse))
2754 return 0;
2755 break;
2757 case 'E':
2758 if (XVECLEN (x, i) != XVECLEN (y, i))
2759 return 0;
2760 for (j = 0; j < XVECLEN (x, i); j++)
2761 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2762 validate, for_gcse))
2763 return 0;
2764 break;
2766 case 's':
2767 if (strcmp (XSTR (x, i), XSTR (y, i)))
2768 return 0;
2769 break;
2771 case 'i':
2772 if (XINT (x, i) != XINT (y, i))
2773 return 0;
2774 break;
2776 case 'w':
2777 if (XWINT (x, i) != XWINT (y, i))
2778 return 0;
2779 break;
2781 case '0':
2782 case 't':
2783 break;
2785 default:
2786 gcc_unreachable ();
2790 return 1;
2793 /* Return 1 if X has a value that can vary even between two
2794 executions of the program. 0 means X can be compared reliably
2795 against certain constants or near-constants. */
2797 static bool
2798 cse_rtx_varies_p (const_rtx x, bool from_alias)
2800 /* We need not check for X and the equivalence class being of the same
2801 mode because if X is equivalent to a constant in some mode, it
2802 doesn't vary in any mode. */
2804 if (REG_P (x)
2805 && REGNO_QTY_VALID_P (REGNO (x)))
2807 int x_q = REG_QTY (REGNO (x));
2808 struct qty_table_elem *x_ent = &qty_table[x_q];
2810 if (GET_MODE (x) == x_ent->mode
2811 && x_ent->const_rtx != NULL_RTX)
2812 return 0;
2815 if (GET_CODE (x) == PLUS
2816 && CONST_INT_P (XEXP (x, 1))
2817 && REG_P (XEXP (x, 0))
2818 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
2820 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2821 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2823 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2824 && x0_ent->const_rtx != NULL_RTX)
2825 return 0;
2828 /* This can happen as the result of virtual register instantiation, if
2829 the initial constant is too large to be a valid address. This gives
2830 us a three instruction sequence, load large offset into a register,
2831 load fp minus a constant into a register, then a MEM which is the
2832 sum of the two `constant' registers. */
2833 if (GET_CODE (x) == PLUS
2834 && REG_P (XEXP (x, 0))
2835 && REG_P (XEXP (x, 1))
2836 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 0)))
2837 && REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
2839 int x0_q = REG_QTY (REGNO (XEXP (x, 0)));
2840 int x1_q = REG_QTY (REGNO (XEXP (x, 1)));
2841 struct qty_table_elem *x0_ent = &qty_table[x0_q];
2842 struct qty_table_elem *x1_ent = &qty_table[x1_q];
2844 if ((GET_MODE (XEXP (x, 0)) == x0_ent->mode)
2845 && x0_ent->const_rtx != NULL_RTX
2846 && (GET_MODE (XEXP (x, 1)) == x1_ent->mode)
2847 && x1_ent->const_rtx != NULL_RTX)
2848 return 0;
2851 return rtx_varies_p (x, from_alias);
2854 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2855 the result if necessary. INSN is as for canon_reg. */
2857 static void
2858 validate_canon_reg (rtx *xloc, rtx insn)
2860 if (*xloc)
2862 rtx new_rtx = canon_reg (*xloc, insn);
2864 /* If replacing pseudo with hard reg or vice versa, ensure the
2865 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2866 gcc_assert (insn && new_rtx);
2867 validate_change (insn, xloc, new_rtx, 1);
2871 /* Canonicalize an expression:
2872 replace each register reference inside it
2873 with the "oldest" equivalent register.
2875 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2876 after we make our substitution. The calls are made with IN_GROUP nonzero
2877 so apply_change_group must be called upon the outermost return from this
2878 function (unless INSN is zero). The result of apply_change_group can
2879 generally be discarded since the changes we are making are optional. */
2881 static rtx
2882 canon_reg (rtx x, rtx insn)
2884 int i;
2885 enum rtx_code code;
2886 const char *fmt;
2888 if (x == 0)
2889 return x;
2891 code = GET_CODE (x);
2892 switch (code)
2894 case PC:
2895 case CC0:
2896 case CONST:
2897 case CONST_INT:
2898 case CONST_DOUBLE:
2899 case CONST_FIXED:
2900 case CONST_VECTOR:
2901 case SYMBOL_REF:
2902 case LABEL_REF:
2903 case ADDR_VEC:
2904 case ADDR_DIFF_VEC:
2905 return x;
2907 case REG:
2909 int first;
2910 int q;
2911 struct qty_table_elem *ent;
2913 /* Never replace a hard reg, because hard regs can appear
2914 in more than one machine mode, and we must preserve the mode
2915 of each occurrence. Also, some hard regs appear in
2916 MEMs that are shared and mustn't be altered. Don't try to
2917 replace any reg that maps to a reg of class NO_REGS. */
2918 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2919 || ! REGNO_QTY_VALID_P (REGNO (x)))
2920 return x;
2922 q = REG_QTY (REGNO (x));
2923 ent = &qty_table[q];
2924 first = ent->first_reg;
2925 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2926 : REGNO_REG_CLASS (first) == NO_REGS ? x
2927 : gen_rtx_REG (ent->mode, first));
2930 default:
2931 break;
2934 fmt = GET_RTX_FORMAT (code);
2935 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2937 int j;
2939 if (fmt[i] == 'e')
2940 validate_canon_reg (&XEXP (x, i), insn);
2941 else if (fmt[i] == 'E')
2942 for (j = 0; j < XVECLEN (x, i); j++)
2943 validate_canon_reg (&XVECEXP (x, i, j), insn);
2946 return x;
2949 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2950 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2951 what values are being compared.
2953 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2954 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2955 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2956 compared to produce cc0.
2958 The return value is the comparison operator and is either the code of
2959 A or the code corresponding to the inverse of the comparison. */
2961 static enum rtx_code
2962 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2963 enum machine_mode *pmode1, enum machine_mode *pmode2)
2965 rtx arg1, arg2;
2967 arg1 = *parg1, arg2 = *parg2;
2969 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2971 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2973 /* Set nonzero when we find something of interest. */
2974 rtx x = 0;
2975 int reverse_code = 0;
2976 struct table_elt *p = 0;
2978 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2979 On machines with CC0, this is the only case that can occur, since
2980 fold_rtx will return the COMPARE or item being compared with zero
2981 when given CC0. */
2983 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2984 x = arg1;
2986 /* If ARG1 is a comparison operator and CODE is testing for
2987 STORE_FLAG_VALUE, get the inner arguments. */
2989 else if (COMPARISON_P (arg1))
2991 #ifdef FLOAT_STORE_FLAG_VALUE
2992 REAL_VALUE_TYPE fsfv;
2993 #endif
2995 if (code == NE
2996 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2997 && code == LT && STORE_FLAG_VALUE == -1)
2998 #ifdef FLOAT_STORE_FLAG_VALUE
2999 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3000 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3001 REAL_VALUE_NEGATIVE (fsfv)))
3002 #endif
3004 x = arg1;
3005 else if (code == EQ
3006 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
3007 && code == GE && STORE_FLAG_VALUE == -1)
3008 #ifdef FLOAT_STORE_FLAG_VALUE
3009 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
3010 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3011 REAL_VALUE_NEGATIVE (fsfv)))
3012 #endif
3014 x = arg1, reverse_code = 1;
3017 /* ??? We could also check for
3019 (ne (and (eq (...) (const_int 1))) (const_int 0))
3021 and related forms, but let's wait until we see them occurring. */
3023 if (x == 0)
3024 /* Look up ARG1 in the hash table and see if it has an equivalence
3025 that lets us see what is being compared. */
3026 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
3027 if (p)
3029 p = p->first_same_value;
3031 /* If what we compare is already known to be constant, that is as
3032 good as it gets.
3033 We need to break the loop in this case, because otherwise we
3034 can have an infinite loop when looking at a reg that is known
3035 to be a constant which is the same as a comparison of a reg
3036 against zero which appears later in the insn stream, which in
3037 turn is constant and the same as the comparison of the first reg
3038 against zero... */
3039 if (p->is_const)
3040 break;
3043 for (; p; p = p->next_same_value)
3045 enum machine_mode inner_mode = GET_MODE (p->exp);
3046 #ifdef FLOAT_STORE_FLAG_VALUE
3047 REAL_VALUE_TYPE fsfv;
3048 #endif
3050 /* If the entry isn't valid, skip it. */
3051 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3052 continue;
3054 if (GET_CODE (p->exp) == COMPARE
3055 /* Another possibility is that this machine has a compare insn
3056 that includes the comparison code. In that case, ARG1 would
3057 be equivalent to a comparison operation that would set ARG1 to
3058 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3059 ORIG_CODE is the actual comparison being done; if it is an EQ,
3060 we must reverse ORIG_CODE. On machine with a negative value
3061 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3062 || ((code == NE
3063 || (code == LT
3064 && GET_MODE_CLASS (inner_mode) == MODE_INT
3065 && (GET_MODE_BITSIZE (inner_mode)
3066 <= HOST_BITS_PER_WIDE_INT)
3067 && (STORE_FLAG_VALUE
3068 & ((HOST_WIDE_INT) 1
3069 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3070 #ifdef FLOAT_STORE_FLAG_VALUE
3071 || (code == LT
3072 && SCALAR_FLOAT_MODE_P (inner_mode)
3073 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3074 REAL_VALUE_NEGATIVE (fsfv)))
3075 #endif
3077 && COMPARISON_P (p->exp)))
3079 x = p->exp;
3080 break;
3082 else if ((code == EQ
3083 || (code == GE
3084 && GET_MODE_CLASS (inner_mode) == MODE_INT
3085 && (GET_MODE_BITSIZE (inner_mode)
3086 <= HOST_BITS_PER_WIDE_INT)
3087 && (STORE_FLAG_VALUE
3088 & ((HOST_WIDE_INT) 1
3089 << (GET_MODE_BITSIZE (inner_mode) - 1))))
3090 #ifdef FLOAT_STORE_FLAG_VALUE
3091 || (code == GE
3092 && SCALAR_FLOAT_MODE_P (inner_mode)
3093 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3094 REAL_VALUE_NEGATIVE (fsfv)))
3095 #endif
3097 && COMPARISON_P (p->exp))
3099 reverse_code = 1;
3100 x = p->exp;
3101 break;
3104 /* If this non-trapping address, e.g. fp + constant, the
3105 equivalent is a better operand since it may let us predict
3106 the value of the comparison. */
3107 else if (!rtx_addr_can_trap_p (p->exp))
3109 arg1 = p->exp;
3110 continue;
3114 /* If we didn't find a useful equivalence for ARG1, we are done.
3115 Otherwise, set up for the next iteration. */
3116 if (x == 0)
3117 break;
3119 /* If we need to reverse the comparison, make sure that that is
3120 possible -- we can't necessarily infer the value of GE from LT
3121 with floating-point operands. */
3122 if (reverse_code)
3124 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3125 if (reversed == UNKNOWN)
3126 break;
3127 else
3128 code = reversed;
3130 else if (COMPARISON_P (x))
3131 code = GET_CODE (x);
3132 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3135 /* Return our results. Return the modes from before fold_rtx
3136 because fold_rtx might produce const_int, and then it's too late. */
3137 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3138 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3140 return code;
3143 /* If X is a nontrivial arithmetic operation on an argument for which
3144 a constant value can be determined, return the result of operating
3145 on that value, as a constant. Otherwise, return X, possibly with
3146 one or more operands changed to a forward-propagated constant.
3148 If X is a register whose contents are known, we do NOT return
3149 those contents here; equiv_constant is called to perform that task.
3150 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3152 INSN is the insn that we may be modifying. If it is 0, make a copy
3153 of X before modifying it. */
3155 static rtx
3156 fold_rtx (rtx x, rtx insn)
3158 enum rtx_code code;
3159 enum machine_mode mode;
3160 const char *fmt;
3161 int i;
3162 rtx new_rtx = 0;
3163 int changed = 0;
3165 /* Operands of X. */
3166 rtx folded_arg0;
3167 rtx folded_arg1;
3169 /* Constant equivalents of first three operands of X;
3170 0 when no such equivalent is known. */
3171 rtx const_arg0;
3172 rtx const_arg1;
3173 rtx const_arg2;
3175 /* The mode of the first operand of X. We need this for sign and zero
3176 extends. */
3177 enum machine_mode mode_arg0;
3179 if (x == 0)
3180 return x;
3182 /* Try to perform some initial simplifications on X. */
3183 code = GET_CODE (x);
3184 switch (code)
3186 case MEM:
3187 case SUBREG:
3188 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3189 return new_rtx;
3190 return x;
3192 case CONST:
3193 case CONST_INT:
3194 case CONST_DOUBLE:
3195 case CONST_FIXED:
3196 case CONST_VECTOR:
3197 case SYMBOL_REF:
3198 case LABEL_REF:
3199 case REG:
3200 case PC:
3201 /* No use simplifying an EXPR_LIST
3202 since they are used only for lists of args
3203 in a function call's REG_EQUAL note. */
3204 case EXPR_LIST:
3205 return x;
3207 #ifdef HAVE_cc0
3208 case CC0:
3209 return prev_insn_cc0;
3210 #endif
3212 case ASM_OPERANDS:
3213 if (insn)
3215 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3216 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3217 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3219 return x;
3221 #ifdef NO_FUNCTION_CSE
3222 case CALL:
3223 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3224 return x;
3225 break;
3226 #endif
3228 /* Anything else goes through the loop below. */
3229 default:
3230 break;
3233 mode = GET_MODE (x);
3234 const_arg0 = 0;
3235 const_arg1 = 0;
3236 const_arg2 = 0;
3237 mode_arg0 = VOIDmode;
3239 /* Try folding our operands.
3240 Then see which ones have constant values known. */
3242 fmt = GET_RTX_FORMAT (code);
3243 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3244 if (fmt[i] == 'e')
3246 rtx folded_arg = XEXP (x, i), const_arg;
3247 enum machine_mode mode_arg = GET_MODE (folded_arg);
3249 switch (GET_CODE (folded_arg))
3251 case MEM:
3252 case REG:
3253 case SUBREG:
3254 const_arg = equiv_constant (folded_arg);
3255 break;
3257 case CONST:
3258 case CONST_INT:
3259 case SYMBOL_REF:
3260 case LABEL_REF:
3261 case CONST_DOUBLE:
3262 case CONST_FIXED:
3263 case CONST_VECTOR:
3264 const_arg = folded_arg;
3265 break;
3267 #ifdef HAVE_cc0
3268 case CC0:
3269 folded_arg = prev_insn_cc0;
3270 mode_arg = prev_insn_cc0_mode;
3271 const_arg = equiv_constant (folded_arg);
3272 break;
3273 #endif
3275 default:
3276 folded_arg = fold_rtx (folded_arg, insn);
3277 const_arg = equiv_constant (folded_arg);
3278 break;
3281 /* For the first three operands, see if the operand
3282 is constant or equivalent to a constant. */
3283 switch (i)
3285 case 0:
3286 folded_arg0 = folded_arg;
3287 const_arg0 = const_arg;
3288 mode_arg0 = mode_arg;
3289 break;
3290 case 1:
3291 folded_arg1 = folded_arg;
3292 const_arg1 = const_arg;
3293 break;
3294 case 2:
3295 const_arg2 = const_arg;
3296 break;
3299 /* Pick the least expensive of the argument and an equivalent constant
3300 argument. */
3301 if (const_arg != 0
3302 && const_arg != folded_arg
3303 && COST_IN (const_arg, code) <= COST_IN (folded_arg, code)
3305 /* It's not safe to substitute the operand of a conversion
3306 operator with a constant, as the conversion's identity
3307 depends upon the mode of its operand. This optimization
3308 is handled by the call to simplify_unary_operation. */
3309 && (GET_RTX_CLASS (code) != RTX_UNARY
3310 || GET_MODE (const_arg) == mode_arg0
3311 || (code != ZERO_EXTEND
3312 && code != SIGN_EXTEND
3313 && code != TRUNCATE
3314 && code != FLOAT_TRUNCATE
3315 && code != FLOAT_EXTEND
3316 && code != FLOAT
3317 && code != FIX
3318 && code != UNSIGNED_FLOAT
3319 && code != UNSIGNED_FIX)))
3320 folded_arg = const_arg;
3322 if (folded_arg == XEXP (x, i))
3323 continue;
3325 if (insn == NULL_RTX && !changed)
3326 x = copy_rtx (x);
3327 changed = 1;
3328 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3331 if (changed)
3333 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3334 consistent with the order in X. */
3335 if (canonicalize_change_group (insn, x))
3337 rtx tem;
3338 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3339 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3342 apply_change_group ();
3345 /* If X is an arithmetic operation, see if we can simplify it. */
3347 switch (GET_RTX_CLASS (code))
3349 case RTX_UNARY:
3351 /* We can't simplify extension ops unless we know the
3352 original mode. */
3353 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3354 && mode_arg0 == VOIDmode)
3355 break;
3357 new_rtx = simplify_unary_operation (code, mode,
3358 const_arg0 ? const_arg0 : folded_arg0,
3359 mode_arg0);
3361 break;
3363 case RTX_COMPARE:
3364 case RTX_COMM_COMPARE:
3365 /* See what items are actually being compared and set FOLDED_ARG[01]
3366 to those values and CODE to the actual comparison code. If any are
3367 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3368 do anything if both operands are already known to be constant. */
3370 /* ??? Vector mode comparisons are not supported yet. */
3371 if (VECTOR_MODE_P (mode))
3372 break;
3374 if (const_arg0 == 0 || const_arg1 == 0)
3376 struct table_elt *p0, *p1;
3377 rtx true_rtx, false_rtx;
3378 enum machine_mode mode_arg1;
3380 if (SCALAR_FLOAT_MODE_P (mode))
3382 #ifdef FLOAT_STORE_FLAG_VALUE
3383 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3384 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3385 #else
3386 true_rtx = NULL_RTX;
3387 #endif
3388 false_rtx = CONST0_RTX (mode);
3390 else
3392 true_rtx = const_true_rtx;
3393 false_rtx = const0_rtx;
3396 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3397 &mode_arg0, &mode_arg1);
3399 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3400 what kinds of things are being compared, so we can't do
3401 anything with this comparison. */
3403 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3404 break;
3406 const_arg0 = equiv_constant (folded_arg0);
3407 const_arg1 = equiv_constant (folded_arg1);
3409 /* If we do not now have two constants being compared, see
3410 if we can nevertheless deduce some things about the
3411 comparison. */
3412 if (const_arg0 == 0 || const_arg1 == 0)
3414 if (const_arg1 != NULL)
3416 rtx cheapest_simplification;
3417 int cheapest_cost;
3418 rtx simp_result;
3419 struct table_elt *p;
3421 /* See if we can find an equivalent of folded_arg0
3422 that gets us a cheaper expression, possibly a
3423 constant through simplifications. */
3424 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3425 mode_arg0);
3427 if (p != NULL)
3429 cheapest_simplification = x;
3430 cheapest_cost = COST (x);
3432 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3434 int cost;
3436 /* If the entry isn't valid, skip it. */
3437 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3438 continue;
3440 /* Try to simplify using this equivalence. */
3441 simp_result
3442 = simplify_relational_operation (code, mode,
3443 mode_arg0,
3444 p->exp,
3445 const_arg1);
3447 if (simp_result == NULL)
3448 continue;
3450 cost = COST (simp_result);
3451 if (cost < cheapest_cost)
3453 cheapest_cost = cost;
3454 cheapest_simplification = simp_result;
3458 /* If we have a cheaper expression now, use that
3459 and try folding it further, from the top. */
3460 if (cheapest_simplification != x)
3461 return fold_rtx (copy_rtx (cheapest_simplification),
3462 insn);
3466 /* See if the two operands are the same. */
3468 if ((REG_P (folded_arg0)
3469 && REG_P (folded_arg1)
3470 && (REG_QTY (REGNO (folded_arg0))
3471 == REG_QTY (REGNO (folded_arg1))))
3472 || ((p0 = lookup (folded_arg0,
3473 SAFE_HASH (folded_arg0, mode_arg0),
3474 mode_arg0))
3475 && (p1 = lookup (folded_arg1,
3476 SAFE_HASH (folded_arg1, mode_arg0),
3477 mode_arg0))
3478 && p0->first_same_value == p1->first_same_value))
3479 folded_arg1 = folded_arg0;
3481 /* If FOLDED_ARG0 is a register, see if the comparison we are
3482 doing now is either the same as we did before or the reverse
3483 (we only check the reverse if not floating-point). */
3484 else if (REG_P (folded_arg0))
3486 int qty = REG_QTY (REGNO (folded_arg0));
3488 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3490 struct qty_table_elem *ent = &qty_table[qty];
3492 if ((comparison_dominates_p (ent->comparison_code, code)
3493 || (! FLOAT_MODE_P (mode_arg0)
3494 && comparison_dominates_p (ent->comparison_code,
3495 reverse_condition (code))))
3496 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3497 || (const_arg1
3498 && rtx_equal_p (ent->comparison_const,
3499 const_arg1))
3500 || (REG_P (folded_arg1)
3501 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3503 if (comparison_dominates_p (ent->comparison_code, code))
3505 if (true_rtx)
3506 return true_rtx;
3507 else
3508 break;
3510 else
3511 return false_rtx;
3518 /* If we are comparing against zero, see if the first operand is
3519 equivalent to an IOR with a constant. If so, we may be able to
3520 determine the result of this comparison. */
3521 if (const_arg1 == const0_rtx && !const_arg0)
3523 rtx y = lookup_as_function (folded_arg0, IOR);
3524 rtx inner_const;
3526 if (y != 0
3527 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3528 && CONST_INT_P (inner_const)
3529 && INTVAL (inner_const) != 0)
3530 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3534 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3535 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3536 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3538 break;
3540 case RTX_BIN_ARITH:
3541 case RTX_COMM_ARITH:
3542 switch (code)
3544 case PLUS:
3545 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3546 with that LABEL_REF as its second operand. If so, the result is
3547 the first operand of that MINUS. This handles switches with an
3548 ADDR_DIFF_VEC table. */
3549 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3551 rtx y
3552 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3553 : lookup_as_function (folded_arg0, MINUS);
3555 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3556 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3557 return XEXP (y, 0);
3559 /* Now try for a CONST of a MINUS like the above. */
3560 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3561 : lookup_as_function (folded_arg0, CONST))) != 0
3562 && GET_CODE (XEXP (y, 0)) == MINUS
3563 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3564 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3565 return XEXP (XEXP (y, 0), 0);
3568 /* Likewise if the operands are in the other order. */
3569 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3571 rtx y
3572 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3573 : lookup_as_function (folded_arg1, MINUS);
3575 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3576 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3577 return XEXP (y, 0);
3579 /* Now try for a CONST of a MINUS like the above. */
3580 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3581 : lookup_as_function (folded_arg1, CONST))) != 0
3582 && GET_CODE (XEXP (y, 0)) == MINUS
3583 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3584 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3585 return XEXP (XEXP (y, 0), 0);
3588 /* If second operand is a register equivalent to a negative
3589 CONST_INT, see if we can find a register equivalent to the
3590 positive constant. Make a MINUS if so. Don't do this for
3591 a non-negative constant since we might then alternate between
3592 choosing positive and negative constants. Having the positive
3593 constant previously-used is the more common case. Be sure
3594 the resulting constant is non-negative; if const_arg1 were
3595 the smallest negative number this would overflow: depending
3596 on the mode, this would either just be the same value (and
3597 hence not save anything) or be incorrect. */
3598 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3599 && INTVAL (const_arg1) < 0
3600 /* This used to test
3602 -INTVAL (const_arg1) >= 0
3604 But The Sun V5.0 compilers mis-compiled that test. So
3605 instead we test for the problematic value in a more direct
3606 manner and hope the Sun compilers get it correct. */
3607 && INTVAL (const_arg1) !=
3608 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3609 && REG_P (folded_arg1))
3611 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3612 struct table_elt *p
3613 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3615 if (p)
3616 for (p = p->first_same_value; p; p = p->next_same_value)
3617 if (REG_P (p->exp))
3618 return simplify_gen_binary (MINUS, mode, folded_arg0,
3619 canon_reg (p->exp, NULL_RTX));
3621 goto from_plus;
3623 case MINUS:
3624 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3625 If so, produce (PLUS Z C2-C). */
3626 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3628 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3629 if (y && CONST_INT_P (XEXP (y, 1)))
3630 return fold_rtx (plus_constant (copy_rtx (y),
3631 -INTVAL (const_arg1)),
3632 NULL_RTX);
3635 /* Fall through. */
3637 from_plus:
3638 case SMIN: case SMAX: case UMIN: case UMAX:
3639 case IOR: case AND: case XOR:
3640 case MULT:
3641 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3642 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3643 is known to be of similar form, we may be able to replace the
3644 operation with a combined operation. This may eliminate the
3645 intermediate operation if every use is simplified in this way.
3646 Note that the similar optimization done by combine.c only works
3647 if the intermediate operation's result has only one reference. */
3649 if (REG_P (folded_arg0)
3650 && const_arg1 && CONST_INT_P (const_arg1))
3652 int is_shift
3653 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3654 rtx y, inner_const, new_const;
3655 rtx canon_const_arg1 = const_arg1;
3656 enum rtx_code associate_code;
3658 if (is_shift
3659 && (INTVAL (const_arg1) >= GET_MODE_BITSIZE (mode)
3660 || INTVAL (const_arg1) < 0))
3662 if (SHIFT_COUNT_TRUNCATED)
3663 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3664 & (GET_MODE_BITSIZE (mode)
3665 - 1));
3666 else
3667 break;
3670 y = lookup_as_function (folded_arg0, code);
3671 if (y == 0)
3672 break;
3674 /* If we have compiled a statement like
3675 "if (x == (x & mask1))", and now are looking at
3676 "x & mask2", we will have a case where the first operand
3677 of Y is the same as our first operand. Unless we detect
3678 this case, an infinite loop will result. */
3679 if (XEXP (y, 0) == folded_arg0)
3680 break;
3682 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3683 if (!inner_const || !CONST_INT_P (inner_const))
3684 break;
3686 /* Don't associate these operations if they are a PLUS with the
3687 same constant and it is a power of two. These might be doable
3688 with a pre- or post-increment. Similarly for two subtracts of
3689 identical powers of two with post decrement. */
3691 if (code == PLUS && const_arg1 == inner_const
3692 && ((HAVE_PRE_INCREMENT
3693 && exact_log2 (INTVAL (const_arg1)) >= 0)
3694 || (HAVE_POST_INCREMENT
3695 && exact_log2 (INTVAL (const_arg1)) >= 0)
3696 || (HAVE_PRE_DECREMENT
3697 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3698 || (HAVE_POST_DECREMENT
3699 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3700 break;
3702 /* ??? Vector mode shifts by scalar
3703 shift operand are not supported yet. */
3704 if (is_shift && VECTOR_MODE_P (mode))
3705 break;
3707 if (is_shift
3708 && (INTVAL (inner_const) >= GET_MODE_BITSIZE (mode)
3709 || INTVAL (inner_const) < 0))
3711 if (SHIFT_COUNT_TRUNCATED)
3712 inner_const = GEN_INT (INTVAL (inner_const)
3713 & (GET_MODE_BITSIZE (mode) - 1));
3714 else
3715 break;
3718 /* Compute the code used to compose the constants. For example,
3719 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3721 associate_code = (is_shift || code == MINUS ? PLUS : code);
3723 new_const = simplify_binary_operation (associate_code, mode,
3724 canon_const_arg1,
3725 inner_const);
3727 if (new_const == 0)
3728 break;
3730 /* If we are associating shift operations, don't let this
3731 produce a shift of the size of the object or larger.
3732 This could occur when we follow a sign-extend by a right
3733 shift on a machine that does a sign-extend as a pair
3734 of shifts. */
3736 if (is_shift
3737 && CONST_INT_P (new_const)
3738 && INTVAL (new_const) >= GET_MODE_BITSIZE (mode))
3740 /* As an exception, we can turn an ASHIFTRT of this
3741 form into a shift of the number of bits - 1. */
3742 if (code == ASHIFTRT)
3743 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3744 else if (!side_effects_p (XEXP (y, 0)))
3745 return CONST0_RTX (mode);
3746 else
3747 break;
3750 y = copy_rtx (XEXP (y, 0));
3752 /* If Y contains our first operand (the most common way this
3753 can happen is if Y is a MEM), we would do into an infinite
3754 loop if we tried to fold it. So don't in that case. */
3756 if (! reg_mentioned_p (folded_arg0, y))
3757 y = fold_rtx (y, insn);
3759 return simplify_gen_binary (code, mode, y, new_const);
3761 break;
3763 case DIV: case UDIV:
3764 /* ??? The associative optimization performed immediately above is
3765 also possible for DIV and UDIV using associate_code of MULT.
3766 However, we would need extra code to verify that the
3767 multiplication does not overflow, that is, there is no overflow
3768 in the calculation of new_const. */
3769 break;
3771 default:
3772 break;
3775 new_rtx = simplify_binary_operation (code, mode,
3776 const_arg0 ? const_arg0 : folded_arg0,
3777 const_arg1 ? const_arg1 : folded_arg1);
3778 break;
3780 case RTX_OBJ:
3781 /* (lo_sum (high X) X) is simply X. */
3782 if (code == LO_SUM && const_arg0 != 0
3783 && GET_CODE (const_arg0) == HIGH
3784 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3785 return const_arg1;
3786 break;
3788 case RTX_TERNARY:
3789 case RTX_BITFIELD_OPS:
3790 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3791 const_arg0 ? const_arg0 : folded_arg0,
3792 const_arg1 ? const_arg1 : folded_arg1,
3793 const_arg2 ? const_arg2 : XEXP (x, 2));
3794 break;
3796 default:
3797 break;
3800 return new_rtx ? new_rtx : x;
3803 /* Return a constant value currently equivalent to X.
3804 Return 0 if we don't know one. */
3806 static rtx
3807 equiv_constant (rtx x)
3809 if (REG_P (x)
3810 && REGNO_QTY_VALID_P (REGNO (x)))
3812 int x_q = REG_QTY (REGNO (x));
3813 struct qty_table_elem *x_ent = &qty_table[x_q];
3815 if (x_ent->const_rtx)
3816 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3819 if (x == 0 || CONSTANT_P (x))
3820 return x;
3822 if (GET_CODE (x) == SUBREG)
3824 enum machine_mode mode = GET_MODE (x);
3825 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3826 rtx new_rtx;
3828 /* See if we previously assigned a constant value to this SUBREG. */
3829 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3830 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3831 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3832 return new_rtx;
3834 /* If we didn't and if doing so makes sense, see if we previously
3835 assigned a constant value to the enclosing word mode SUBREG. */
3836 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3837 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3839 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3840 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3842 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3843 new_rtx = lookup_as_function (y, CONST_INT);
3844 if (new_rtx)
3845 return gen_lowpart (mode, new_rtx);
3849 /* Otherwise see if we already have a constant for the inner REG. */
3850 if (REG_P (SUBREG_REG (x))
3851 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3852 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3854 return 0;
3857 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3858 the hash table in case its value was seen before. */
3860 if (MEM_P (x))
3862 struct table_elt *elt;
3864 x = avoid_constant_pool_reference (x);
3865 if (CONSTANT_P (x))
3866 return x;
3868 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3869 if (elt == 0)
3870 return 0;
3872 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3873 if (elt->is_const && CONSTANT_P (elt->exp))
3874 return elt->exp;
3877 return 0;
3880 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3881 "taken" branch.
3883 In certain cases, this can cause us to add an equivalence. For example,
3884 if we are following the taken case of
3885 if (i == 2)
3886 we can add the fact that `i' and '2' are now equivalent.
3888 In any case, we can record that this comparison was passed. If the same
3889 comparison is seen later, we will know its value. */
3891 static void
3892 record_jump_equiv (rtx insn, bool taken)
3894 int cond_known_true;
3895 rtx op0, op1;
3896 rtx set;
3897 enum machine_mode mode, mode0, mode1;
3898 int reversed_nonequality = 0;
3899 enum rtx_code code;
3901 /* Ensure this is the right kind of insn. */
3902 gcc_assert (any_condjump_p (insn));
3904 set = pc_set (insn);
3906 /* See if this jump condition is known true or false. */
3907 if (taken)
3908 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3909 else
3910 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3912 /* Get the type of comparison being done and the operands being compared.
3913 If we had to reverse a non-equality condition, record that fact so we
3914 know that it isn't valid for floating-point. */
3915 code = GET_CODE (XEXP (SET_SRC (set), 0));
3916 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3917 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3919 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3920 if (! cond_known_true)
3922 code = reversed_comparison_code_parts (code, op0, op1, insn);
3924 /* Don't remember if we can't find the inverse. */
3925 if (code == UNKNOWN)
3926 return;
3929 /* The mode is the mode of the non-constant. */
3930 mode = mode0;
3931 if (mode1 != VOIDmode)
3932 mode = mode1;
3934 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3937 /* Yet another form of subreg creation. In this case, we want something in
3938 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3940 static rtx
3941 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3943 enum machine_mode op_mode = GET_MODE (op);
3944 if (op_mode == mode || op_mode == VOIDmode)
3945 return op;
3946 return lowpart_subreg (mode, op, op_mode);
3949 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3950 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3951 Make any useful entries we can with that information. Called from
3952 above function and called recursively. */
3954 static void
3955 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3956 rtx op1, int reversed_nonequality)
3958 unsigned op0_hash, op1_hash;
3959 int op0_in_memory, op1_in_memory;
3960 struct table_elt *op0_elt, *op1_elt;
3962 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3963 we know that they are also equal in the smaller mode (this is also
3964 true for all smaller modes whether or not there is a SUBREG, but
3965 is not worth testing for with no SUBREG). */
3967 /* Note that GET_MODE (op0) may not equal MODE. */
3968 if (code == EQ && GET_CODE (op0) == SUBREG
3969 && (GET_MODE_SIZE (GET_MODE (op0))
3970 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3972 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3973 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3974 if (tem)
3975 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3976 reversed_nonequality);
3979 if (code == EQ && GET_CODE (op1) == SUBREG
3980 && (GET_MODE_SIZE (GET_MODE (op1))
3981 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3983 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3984 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3985 if (tem)
3986 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3987 reversed_nonequality);
3990 /* Similarly, if this is an NE comparison, and either is a SUBREG
3991 making a smaller mode, we know the whole thing is also NE. */
3993 /* Note that GET_MODE (op0) may not equal MODE;
3994 if we test MODE instead, we can get an infinite recursion
3995 alternating between two modes each wider than MODE. */
3997 if (code == NE && GET_CODE (op0) == SUBREG
3998 && subreg_lowpart_p (op0)
3999 && (GET_MODE_SIZE (GET_MODE (op0))
4000 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
4002 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
4003 rtx tem = record_jump_cond_subreg (inner_mode, op1);
4004 if (tem)
4005 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
4006 reversed_nonequality);
4009 if (code == NE && GET_CODE (op1) == SUBREG
4010 && subreg_lowpart_p (op1)
4011 && (GET_MODE_SIZE (GET_MODE (op1))
4012 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
4014 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
4015 rtx tem = record_jump_cond_subreg (inner_mode, op0);
4016 if (tem)
4017 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
4018 reversed_nonequality);
4021 /* Hash both operands. */
4023 do_not_record = 0;
4024 hash_arg_in_memory = 0;
4025 op0_hash = HASH (op0, mode);
4026 op0_in_memory = hash_arg_in_memory;
4028 if (do_not_record)
4029 return;
4031 do_not_record = 0;
4032 hash_arg_in_memory = 0;
4033 op1_hash = HASH (op1, mode);
4034 op1_in_memory = hash_arg_in_memory;
4036 if (do_not_record)
4037 return;
4039 /* Look up both operands. */
4040 op0_elt = lookup (op0, op0_hash, mode);
4041 op1_elt = lookup (op1, op1_hash, mode);
4043 /* If both operands are already equivalent or if they are not in the
4044 table but are identical, do nothing. */
4045 if ((op0_elt != 0 && op1_elt != 0
4046 && op0_elt->first_same_value == op1_elt->first_same_value)
4047 || op0 == op1 || rtx_equal_p (op0, op1))
4048 return;
4050 /* If we aren't setting two things equal all we can do is save this
4051 comparison. Similarly if this is floating-point. In the latter
4052 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4053 If we record the equality, we might inadvertently delete code
4054 whose intent was to change -0 to +0. */
4056 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4058 struct qty_table_elem *ent;
4059 int qty;
4061 /* If we reversed a floating-point comparison, if OP0 is not a
4062 register, or if OP1 is neither a register or constant, we can't
4063 do anything. */
4065 if (!REG_P (op1))
4066 op1 = equiv_constant (op1);
4068 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4069 || !REG_P (op0) || op1 == 0)
4070 return;
4072 /* Put OP0 in the hash table if it isn't already. This gives it a
4073 new quantity number. */
4074 if (op0_elt == 0)
4076 if (insert_regs (op0, NULL, 0))
4078 rehash_using_reg (op0);
4079 op0_hash = HASH (op0, mode);
4081 /* If OP0 is contained in OP1, this changes its hash code
4082 as well. Faster to rehash than to check, except
4083 for the simple case of a constant. */
4084 if (! CONSTANT_P (op1))
4085 op1_hash = HASH (op1,mode);
4088 op0_elt = insert (op0, NULL, op0_hash, mode);
4089 op0_elt->in_memory = op0_in_memory;
4092 qty = REG_QTY (REGNO (op0));
4093 ent = &qty_table[qty];
4095 ent->comparison_code = code;
4096 if (REG_P (op1))
4098 /* Look it up again--in case op0 and op1 are the same. */
4099 op1_elt = lookup (op1, op1_hash, mode);
4101 /* Put OP1 in the hash table so it gets a new quantity number. */
4102 if (op1_elt == 0)
4104 if (insert_regs (op1, NULL, 0))
4106 rehash_using_reg (op1);
4107 op1_hash = HASH (op1, mode);
4110 op1_elt = insert (op1, NULL, op1_hash, mode);
4111 op1_elt->in_memory = op1_in_memory;
4114 ent->comparison_const = NULL_RTX;
4115 ent->comparison_qty = REG_QTY (REGNO (op1));
4117 else
4119 ent->comparison_const = op1;
4120 ent->comparison_qty = -1;
4123 return;
4126 /* If either side is still missing an equivalence, make it now,
4127 then merge the equivalences. */
4129 if (op0_elt == 0)
4131 if (insert_regs (op0, NULL, 0))
4133 rehash_using_reg (op0);
4134 op0_hash = HASH (op0, mode);
4137 op0_elt = insert (op0, NULL, op0_hash, mode);
4138 op0_elt->in_memory = op0_in_memory;
4141 if (op1_elt == 0)
4143 if (insert_regs (op1, NULL, 0))
4145 rehash_using_reg (op1);
4146 op1_hash = HASH (op1, mode);
4149 op1_elt = insert (op1, NULL, op1_hash, mode);
4150 op1_elt->in_memory = op1_in_memory;
4153 merge_equiv_classes (op0_elt, op1_elt);
4156 /* CSE processing for one instruction.
4157 First simplify sources and addresses of all assignments
4158 in the instruction, using previously-computed equivalents values.
4159 Then install the new sources and destinations in the table
4160 of available values. */
4162 /* Data on one SET contained in the instruction. */
4164 struct set
4166 /* The SET rtx itself. */
4167 rtx rtl;
4168 /* The SET_SRC of the rtx (the original value, if it is changing). */
4169 rtx src;
4170 /* The hash-table element for the SET_SRC of the SET. */
4171 struct table_elt *src_elt;
4172 /* Hash value for the SET_SRC. */
4173 unsigned src_hash;
4174 /* Hash value for the SET_DEST. */
4175 unsigned dest_hash;
4176 /* The SET_DEST, with SUBREG, etc., stripped. */
4177 rtx inner_dest;
4178 /* Nonzero if the SET_SRC is in memory. */
4179 char src_in_memory;
4180 /* Nonzero if the SET_SRC contains something
4181 whose value cannot be predicted and understood. */
4182 char src_volatile;
4183 /* Original machine mode, in case it becomes a CONST_INT.
4184 The size of this field should match the size of the mode
4185 field of struct rtx_def (see rtl.h). */
4186 ENUM_BITFIELD(machine_mode) mode : 8;
4187 /* A constant equivalent for SET_SRC, if any. */
4188 rtx src_const;
4189 /* Hash value of constant equivalent for SET_SRC. */
4190 unsigned src_const_hash;
4191 /* Table entry for constant equivalent for SET_SRC, if any. */
4192 struct table_elt *src_const_elt;
4193 /* Table entry for the destination address. */
4194 struct table_elt *dest_addr_elt;
4197 static void
4198 cse_insn (rtx insn)
4200 rtx x = PATTERN (insn);
4201 int i;
4202 rtx tem;
4203 int n_sets = 0;
4205 rtx src_eqv = 0;
4206 struct table_elt *src_eqv_elt = 0;
4207 int src_eqv_volatile = 0;
4208 int src_eqv_in_memory = 0;
4209 unsigned src_eqv_hash = 0;
4211 struct set *sets = (struct set *) 0;
4213 this_insn = insn;
4214 #ifdef HAVE_cc0
4215 /* Records what this insn does to set CC0. */
4216 this_insn_cc0 = 0;
4217 this_insn_cc0_mode = VOIDmode;
4218 #endif
4220 /* Find all the SETs and CLOBBERs in this instruction.
4221 Record all the SETs in the array `set' and count them.
4222 Also determine whether there is a CLOBBER that invalidates
4223 all memory references, or all references at varying addresses. */
4225 if (CALL_P (insn))
4227 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4229 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
4230 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
4231 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4235 if (GET_CODE (x) == SET)
4237 sets = XALLOCA (struct set);
4238 sets[0].rtl = x;
4240 /* Ignore SETs that are unconditional jumps.
4241 They never need cse processing, so this does not hurt.
4242 The reason is not efficiency but rather
4243 so that we can test at the end for instructions
4244 that have been simplified to unconditional jumps
4245 and not be misled by unchanged instructions
4246 that were unconditional jumps to begin with. */
4247 if (SET_DEST (x) == pc_rtx
4248 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4251 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4252 The hard function value register is used only once, to copy to
4253 someplace else, so it isn't worth cse'ing (and on 80386 is unsafe)!
4254 Ensure we invalidate the destination register. On the 80386 no
4255 other code would invalidate it since it is a fixed_reg.
4256 We need not check the return of apply_change_group; see canon_reg. */
4258 else if (GET_CODE (SET_SRC (x)) == CALL)
4260 canon_reg (SET_SRC (x), insn);
4261 apply_change_group ();
4262 fold_rtx (SET_SRC (x), insn);
4263 invalidate (SET_DEST (x), VOIDmode);
4265 else
4266 n_sets = 1;
4268 else if (GET_CODE (x) == PARALLEL)
4270 int lim = XVECLEN (x, 0);
4272 sets = XALLOCAVEC (struct set, lim);
4274 /* Find all regs explicitly clobbered in this insn,
4275 and ensure they are not replaced with any other regs
4276 elsewhere in this insn.
4277 When a reg that is clobbered is also used for input,
4278 we should presume that that is for a reason,
4279 and we should not substitute some other register
4280 which is not supposed to be clobbered.
4281 Therefore, this loop cannot be merged into the one below
4282 because a CALL may precede a CLOBBER and refer to the
4283 value clobbered. We must not let a canonicalization do
4284 anything in that case. */
4285 for (i = 0; i < lim; i++)
4287 rtx y = XVECEXP (x, 0, i);
4288 if (GET_CODE (y) == CLOBBER)
4290 rtx clobbered = XEXP (y, 0);
4292 if (REG_P (clobbered)
4293 || GET_CODE (clobbered) == SUBREG)
4294 invalidate (clobbered, VOIDmode);
4295 else if (GET_CODE (clobbered) == STRICT_LOW_PART
4296 || GET_CODE (clobbered) == ZERO_EXTRACT)
4297 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
4301 for (i = 0; i < lim; i++)
4303 rtx y = XVECEXP (x, 0, i);
4304 if (GET_CODE (y) == SET)
4306 /* As above, we ignore unconditional jumps and call-insns and
4307 ignore the result of apply_change_group. */
4308 if (GET_CODE (SET_SRC (y)) == CALL)
4310 canon_reg (SET_SRC (y), insn);
4311 apply_change_group ();
4312 fold_rtx (SET_SRC (y), insn);
4313 invalidate (SET_DEST (y), VOIDmode);
4315 else if (SET_DEST (y) == pc_rtx
4316 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4318 else
4319 sets[n_sets++].rtl = y;
4321 else if (GET_CODE (y) == CLOBBER)
4323 /* If we clobber memory, canon the address.
4324 This does nothing when a register is clobbered
4325 because we have already invalidated the reg. */
4326 if (MEM_P (XEXP (y, 0)))
4327 canon_reg (XEXP (y, 0), insn);
4329 else if (GET_CODE (y) == USE
4330 && ! (REG_P (XEXP (y, 0))
4331 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4332 canon_reg (y, insn);
4333 else if (GET_CODE (y) == CALL)
4335 /* The result of apply_change_group can be ignored; see
4336 canon_reg. */
4337 canon_reg (y, insn);
4338 apply_change_group ();
4339 fold_rtx (y, insn);
4343 else if (GET_CODE (x) == CLOBBER)
4345 if (MEM_P (XEXP (x, 0)))
4346 canon_reg (XEXP (x, 0), insn);
4349 /* Canonicalize a USE of a pseudo register or memory location. */
4350 else if (GET_CODE (x) == USE
4351 && ! (REG_P (XEXP (x, 0))
4352 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4353 canon_reg (XEXP (x, 0), insn);
4354 else if (GET_CODE (x) == CALL)
4356 /* The result of apply_change_group can be ignored; see canon_reg. */
4357 canon_reg (x, insn);
4358 apply_change_group ();
4359 fold_rtx (x, insn);
4362 /* Store the equivalent value in SRC_EQV, if different, or if the DEST
4363 is a STRICT_LOW_PART. The latter condition is necessary because SRC_EQV
4364 is handled specially for this case, and if it isn't set, then there will
4365 be no equivalence for the destination. */
4366 if (n_sets == 1 && REG_NOTES (insn) != 0
4367 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4368 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4369 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4371 /* The result of apply_change_group can be ignored; see canon_reg. */
4372 canon_reg (XEXP (tem, 0), insn);
4373 apply_change_group ();
4374 src_eqv = fold_rtx (XEXP (tem, 0), insn);
4375 XEXP (tem, 0) = copy_rtx (src_eqv);
4376 df_notes_rescan (insn);
4379 /* Canonicalize sources and addresses of destinations.
4380 We do this in a separate pass to avoid problems when a MATCH_DUP is
4381 present in the insn pattern. In that case, we want to ensure that
4382 we don't break the duplicate nature of the pattern. So we will replace
4383 both operands at the same time. Otherwise, we would fail to find an
4384 equivalent substitution in the loop calling validate_change below.
4386 We used to suppress canonicalization of DEST if it appears in SRC,
4387 but we don't do this any more. */
4389 for (i = 0; i < n_sets; i++)
4391 rtx dest = SET_DEST (sets[i].rtl);
4392 rtx src = SET_SRC (sets[i].rtl);
4393 rtx new_rtx = canon_reg (src, insn);
4395 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4397 if (GET_CODE (dest) == ZERO_EXTRACT)
4399 validate_change (insn, &XEXP (dest, 1),
4400 canon_reg (XEXP (dest, 1), insn), 1);
4401 validate_change (insn, &XEXP (dest, 2),
4402 canon_reg (XEXP (dest, 2), insn), 1);
4405 while (GET_CODE (dest) == SUBREG
4406 || GET_CODE (dest) == ZERO_EXTRACT
4407 || GET_CODE (dest) == STRICT_LOW_PART)
4408 dest = XEXP (dest, 0);
4410 if (MEM_P (dest))
4411 canon_reg (dest, insn);
4414 /* Now that we have done all the replacements, we can apply the change
4415 group and see if they all work. Note that this will cause some
4416 canonicalizations that would have worked individually not to be applied
4417 because some other canonicalization didn't work, but this should not
4418 occur often.
4420 The result of apply_change_group can be ignored; see canon_reg. */
4422 apply_change_group ();
4424 /* Set sets[i].src_elt to the class each source belongs to.
4425 Detect assignments from or to volatile things
4426 and set set[i] to zero so they will be ignored
4427 in the rest of this function.
4429 Nothing in this loop changes the hash table or the register chains. */
4431 for (i = 0; i < n_sets; i++)
4433 rtx src, dest;
4434 rtx src_folded;
4435 struct table_elt *elt = 0, *p;
4436 enum machine_mode mode;
4437 rtx src_eqv_here;
4438 rtx src_const = 0;
4439 rtx src_related = 0;
4440 bool src_related_is_const_anchor = false;
4441 struct table_elt *src_const_elt = 0;
4442 int src_cost = MAX_COST;
4443 int src_eqv_cost = MAX_COST;
4444 int src_folded_cost = MAX_COST;
4445 int src_related_cost = MAX_COST;
4446 int src_elt_cost = MAX_COST;
4447 int src_regcost = MAX_COST;
4448 int src_eqv_regcost = MAX_COST;
4449 int src_folded_regcost = MAX_COST;
4450 int src_related_regcost = MAX_COST;
4451 int src_elt_regcost = MAX_COST;
4452 /* Set nonzero if we need to call force_const_mem on with the
4453 contents of src_folded before using it. */
4454 int src_folded_force_flag = 0;
4456 dest = SET_DEST (sets[i].rtl);
4457 src = SET_SRC (sets[i].rtl);
4459 /* If SRC is a constant that has no machine mode,
4460 hash it with the destination's machine mode.
4461 This way we can keep different modes separate. */
4463 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4464 sets[i].mode = mode;
4466 if (src_eqv)
4468 enum machine_mode eqvmode = mode;
4469 if (GET_CODE (dest) == STRICT_LOW_PART)
4470 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4471 do_not_record = 0;
4472 hash_arg_in_memory = 0;
4473 src_eqv_hash = HASH (src_eqv, eqvmode);
4475 /* Find the equivalence class for the equivalent expression. */
4477 if (!do_not_record)
4478 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4480 src_eqv_volatile = do_not_record;
4481 src_eqv_in_memory = hash_arg_in_memory;
4484 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4485 value of the INNER register, not the destination. So it is not
4486 a valid substitution for the source. But save it for later. */
4487 if (GET_CODE (dest) == STRICT_LOW_PART)
4488 src_eqv_here = 0;
4489 else
4490 src_eqv_here = src_eqv;
4492 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4493 simplified result, which may not necessarily be valid. */
4494 src_folded = fold_rtx (src, insn);
4496 #if 0
4497 /* ??? This caused bad code to be generated for the m68k port with -O2.
4498 Suppose src is (CONST_INT -1), and that after truncation src_folded
4499 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4500 At the end we will add src and src_const to the same equivalence
4501 class. We now have 3 and -1 on the same equivalence class. This
4502 causes later instructions to be mis-optimized. */
4503 /* If storing a constant in a bitfield, pre-truncate the constant
4504 so we will be able to record it later. */
4505 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4507 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4509 if (CONST_INT_P (src)
4510 && CONST_INT_P (width)
4511 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4512 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4513 src_folded
4514 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4515 << INTVAL (width)) - 1));
4517 #endif
4519 /* Compute SRC's hash code, and also notice if it
4520 should not be recorded at all. In that case,
4521 prevent any further processing of this assignment. */
4522 do_not_record = 0;
4523 hash_arg_in_memory = 0;
4525 sets[i].src = src;
4526 sets[i].src_hash = HASH (src, mode);
4527 sets[i].src_volatile = do_not_record;
4528 sets[i].src_in_memory = hash_arg_in_memory;
4530 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4531 a pseudo, do not record SRC. Using SRC as a replacement for
4532 anything else will be incorrect in that situation. Note that
4533 this usually occurs only for stack slots, in which case all the
4534 RTL would be referring to SRC, so we don't lose any optimization
4535 opportunities by not having SRC in the hash table. */
4537 if (MEM_P (src)
4538 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4539 && REG_P (dest)
4540 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4541 sets[i].src_volatile = 1;
4543 #if 0
4544 /* It is no longer clear why we used to do this, but it doesn't
4545 appear to still be needed. So let's try without it since this
4546 code hurts cse'ing widened ops. */
4547 /* If source is a paradoxical subreg (such as QI treated as an SI),
4548 treat it as volatile. It may do the work of an SI in one context
4549 where the extra bits are not being used, but cannot replace an SI
4550 in general. */
4551 if (GET_CODE (src) == SUBREG
4552 && (GET_MODE_SIZE (GET_MODE (src))
4553 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
4554 sets[i].src_volatile = 1;
4555 #endif
4557 /* Locate all possible equivalent forms for SRC. Try to replace
4558 SRC in the insn with each cheaper equivalent.
4560 We have the following types of equivalents: SRC itself, a folded
4561 version, a value given in a REG_EQUAL note, or a value related
4562 to a constant.
4564 Each of these equivalents may be part of an additional class
4565 of equivalents (if more than one is in the table, they must be in
4566 the same class; we check for this).
4568 If the source is volatile, we don't do any table lookups.
4570 We note any constant equivalent for possible later use in a
4571 REG_NOTE. */
4573 if (!sets[i].src_volatile)
4574 elt = lookup (src, sets[i].src_hash, mode);
4576 sets[i].src_elt = elt;
4578 if (elt && src_eqv_here && src_eqv_elt)
4580 if (elt->first_same_value != src_eqv_elt->first_same_value)
4582 /* The REG_EQUAL is indicating that two formerly distinct
4583 classes are now equivalent. So merge them. */
4584 merge_equiv_classes (elt, src_eqv_elt);
4585 src_eqv_hash = HASH (src_eqv, elt->mode);
4586 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4589 src_eqv_here = 0;
4592 else if (src_eqv_elt)
4593 elt = src_eqv_elt;
4595 /* Try to find a constant somewhere and record it in `src_const'.
4596 Record its table element, if any, in `src_const_elt'. Look in
4597 any known equivalences first. (If the constant is not in the
4598 table, also set `sets[i].src_const_hash'). */
4599 if (elt)
4600 for (p = elt->first_same_value; p; p = p->next_same_value)
4601 if (p->is_const)
4603 src_const = p->exp;
4604 src_const_elt = elt;
4605 break;
4608 if (src_const == 0
4609 && (CONSTANT_P (src_folded)
4610 /* Consider (minus (label_ref L1) (label_ref L2)) as
4611 "constant" here so we will record it. This allows us
4612 to fold switch statements when an ADDR_DIFF_VEC is used. */
4613 || (GET_CODE (src_folded) == MINUS
4614 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4615 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4616 src_const = src_folded, src_const_elt = elt;
4617 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4618 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4620 /* If we don't know if the constant is in the table, get its
4621 hash code and look it up. */
4622 if (src_const && src_const_elt == 0)
4624 sets[i].src_const_hash = HASH (src_const, mode);
4625 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4628 sets[i].src_const = src_const;
4629 sets[i].src_const_elt = src_const_elt;
4631 /* If the constant and our source are both in the table, mark them as
4632 equivalent. Otherwise, if a constant is in the table but the source
4633 isn't, set ELT to it. */
4634 if (src_const_elt && elt
4635 && src_const_elt->first_same_value != elt->first_same_value)
4636 merge_equiv_classes (elt, src_const_elt);
4637 else if (src_const_elt && elt == 0)
4638 elt = src_const_elt;
4640 /* See if there is a register linearly related to a constant
4641 equivalent of SRC. */
4642 if (src_const
4643 && (GET_CODE (src_const) == CONST
4644 || (src_const_elt && src_const_elt->related_value != 0)))
4646 src_related = use_related_value (src_const, src_const_elt);
4647 if (src_related)
4649 struct table_elt *src_related_elt
4650 = lookup (src_related, HASH (src_related, mode), mode);
4651 if (src_related_elt && elt)
4653 if (elt->first_same_value
4654 != src_related_elt->first_same_value)
4655 /* This can occur when we previously saw a CONST
4656 involving a SYMBOL_REF and then see the SYMBOL_REF
4657 twice. Merge the involved classes. */
4658 merge_equiv_classes (elt, src_related_elt);
4660 src_related = 0;
4661 src_related_elt = 0;
4663 else if (src_related_elt && elt == 0)
4664 elt = src_related_elt;
4668 /* See if we have a CONST_INT that is already in a register in a
4669 wider mode. */
4671 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4672 && GET_MODE_CLASS (mode) == MODE_INT
4673 && GET_MODE_BITSIZE (mode) < BITS_PER_WORD)
4675 enum machine_mode wider_mode;
4677 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4678 wider_mode != VOIDmode
4679 && GET_MODE_BITSIZE (wider_mode) <= BITS_PER_WORD
4680 && src_related == 0;
4681 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4683 struct table_elt *const_elt
4684 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4686 if (const_elt == 0)
4687 continue;
4689 for (const_elt = const_elt->first_same_value;
4690 const_elt; const_elt = const_elt->next_same_value)
4691 if (REG_P (const_elt->exp))
4693 src_related = gen_lowpart (mode, const_elt->exp);
4694 break;
4699 /* Another possibility is that we have an AND with a constant in
4700 a mode narrower than a word. If so, it might have been generated
4701 as part of an "if" which would narrow the AND. If we already
4702 have done the AND in a wider mode, we can use a SUBREG of that
4703 value. */
4705 if (flag_expensive_optimizations && ! src_related
4706 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4707 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4709 enum machine_mode tmode;
4710 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4712 for (tmode = GET_MODE_WIDER_MODE (mode);
4713 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4714 tmode = GET_MODE_WIDER_MODE (tmode))
4716 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4717 struct table_elt *larger_elt;
4719 if (inner)
4721 PUT_MODE (new_and, tmode);
4722 XEXP (new_and, 0) = inner;
4723 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4724 if (larger_elt == 0)
4725 continue;
4727 for (larger_elt = larger_elt->first_same_value;
4728 larger_elt; larger_elt = larger_elt->next_same_value)
4729 if (REG_P (larger_elt->exp))
4731 src_related
4732 = gen_lowpart (mode, larger_elt->exp);
4733 break;
4736 if (src_related)
4737 break;
4742 #ifdef LOAD_EXTEND_OP
4743 /* See if a MEM has already been loaded with a widening operation;
4744 if it has, we can use a subreg of that. Many CISC machines
4745 also have such operations, but this is only likely to be
4746 beneficial on these machines. */
4748 if (flag_expensive_optimizations && src_related == 0
4749 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4750 && GET_MODE_CLASS (mode) == MODE_INT
4751 && MEM_P (src) && ! do_not_record
4752 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4754 struct rtx_def memory_extend_buf;
4755 rtx memory_extend_rtx = &memory_extend_buf;
4756 enum machine_mode tmode;
4758 /* Set what we are trying to extend and the operation it might
4759 have been extended with. */
4760 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4761 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4762 XEXP (memory_extend_rtx, 0) = src;
4764 for (tmode = GET_MODE_WIDER_MODE (mode);
4765 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4766 tmode = GET_MODE_WIDER_MODE (tmode))
4768 struct table_elt *larger_elt;
4770 PUT_MODE (memory_extend_rtx, tmode);
4771 larger_elt = lookup (memory_extend_rtx,
4772 HASH (memory_extend_rtx, tmode), tmode);
4773 if (larger_elt == 0)
4774 continue;
4776 for (larger_elt = larger_elt->first_same_value;
4777 larger_elt; larger_elt = larger_elt->next_same_value)
4778 if (REG_P (larger_elt->exp))
4780 src_related = gen_lowpart (mode, larger_elt->exp);
4781 break;
4784 if (src_related)
4785 break;
4788 #endif /* LOAD_EXTEND_OP */
4790 /* Try to express the constant using a register+offset expression
4791 derived from a constant anchor. */
4793 if (targetm.const_anchor
4794 && !src_related
4795 && src_const
4796 && GET_CODE (src_const) == CONST_INT)
4798 src_related = try_const_anchors (src_const, mode);
4799 src_related_is_const_anchor = src_related != NULL_RTX;
4803 if (src == src_folded)
4804 src_folded = 0;
4806 /* At this point, ELT, if nonzero, points to a class of expressions
4807 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4808 and SRC_RELATED, if nonzero, each contain additional equivalent
4809 expressions. Prune these latter expressions by deleting expressions
4810 already in the equivalence class.
4812 Check for an equivalent identical to the destination. If found,
4813 this is the preferred equivalent since it will likely lead to
4814 elimination of the insn. Indicate this by placing it in
4815 `src_related'. */
4817 if (elt)
4818 elt = elt->first_same_value;
4819 for (p = elt; p; p = p->next_same_value)
4821 enum rtx_code code = GET_CODE (p->exp);
4823 /* If the expression is not valid, ignore it. Then we do not
4824 have to check for validity below. In most cases, we can use
4825 `rtx_equal_p', since canonicalization has already been done. */
4826 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4827 continue;
4829 /* Also skip paradoxical subregs, unless that's what we're
4830 looking for. */
4831 if (code == SUBREG
4832 && (GET_MODE_SIZE (GET_MODE (p->exp))
4833 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))
4834 && ! (src != 0
4835 && GET_CODE (src) == SUBREG
4836 && GET_MODE (src) == GET_MODE (p->exp)
4837 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4838 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4839 continue;
4841 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4842 src = 0;
4843 else if (src_folded && GET_CODE (src_folded) == code
4844 && rtx_equal_p (src_folded, p->exp))
4845 src_folded = 0;
4846 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4847 && rtx_equal_p (src_eqv_here, p->exp))
4848 src_eqv_here = 0;
4849 else if (src_related && GET_CODE (src_related) == code
4850 && rtx_equal_p (src_related, p->exp))
4851 src_related = 0;
4853 /* This is the same as the destination of the insns, we want
4854 to prefer it. Copy it to src_related. The code below will
4855 then give it a negative cost. */
4856 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4857 src_related = dest;
4860 /* Find the cheapest valid equivalent, trying all the available
4861 possibilities. Prefer items not in the hash table to ones
4862 that are when they are equal cost. Note that we can never
4863 worsen an insn as the current contents will also succeed.
4864 If we find an equivalent identical to the destination, use it as best,
4865 since this insn will probably be eliminated in that case. */
4866 if (src)
4868 if (rtx_equal_p (src, dest))
4869 src_cost = src_regcost = -1;
4870 else
4872 src_cost = COST (src);
4873 src_regcost = approx_reg_cost (src);
4877 if (src_eqv_here)
4879 if (rtx_equal_p (src_eqv_here, dest))
4880 src_eqv_cost = src_eqv_regcost = -1;
4881 else
4883 src_eqv_cost = COST (src_eqv_here);
4884 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4888 if (src_folded)
4890 if (rtx_equal_p (src_folded, dest))
4891 src_folded_cost = src_folded_regcost = -1;
4892 else
4894 src_folded_cost = COST (src_folded);
4895 src_folded_regcost = approx_reg_cost (src_folded);
4899 if (src_related)
4901 if (rtx_equal_p (src_related, dest))
4902 src_related_cost = src_related_regcost = -1;
4903 else
4905 src_related_cost = COST (src_related);
4906 src_related_regcost = approx_reg_cost (src_related);
4908 /* If a const-anchor is used to synthesize a constant that
4909 normally requires multiple instructions then slightly prefer
4910 it over the original sequence. These instructions are likely
4911 to become redundant now. We can't compare against the cost
4912 of src_eqv_here because, on MIPS for example, multi-insn
4913 constants have zero cost; they are assumed to be hoisted from
4914 loops. */
4915 if (src_related_is_const_anchor
4916 && src_related_cost == src_cost
4917 && src_eqv_here)
4918 src_related_cost--;
4922 /* If this was an indirect jump insn, a known label will really be
4923 cheaper even though it looks more expensive. */
4924 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
4925 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
4927 /* Terminate loop when replacement made. This must terminate since
4928 the current contents will be tested and will always be valid. */
4929 while (1)
4931 rtx trial;
4933 /* Skip invalid entries. */
4934 while (elt && !REG_P (elt->exp)
4935 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
4936 elt = elt->next_same_value;
4938 /* A paradoxical subreg would be bad here: it'll be the right
4939 size, but later may be adjusted so that the upper bits aren't
4940 what we want. So reject it. */
4941 if (elt != 0
4942 && GET_CODE (elt->exp) == SUBREG
4943 && (GET_MODE_SIZE (GET_MODE (elt->exp))
4944 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))
4945 /* It is okay, though, if the rtx we're trying to match
4946 will ignore any of the bits we can't predict. */
4947 && ! (src != 0
4948 && GET_CODE (src) == SUBREG
4949 && GET_MODE (src) == GET_MODE (elt->exp)
4950 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4951 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
4953 elt = elt->next_same_value;
4954 continue;
4957 if (elt)
4959 src_elt_cost = elt->cost;
4960 src_elt_regcost = elt->regcost;
4963 /* Find cheapest and skip it for the next time. For items
4964 of equal cost, use this order:
4965 src_folded, src, src_eqv, src_related and hash table entry. */
4966 if (src_folded
4967 && preferable (src_folded_cost, src_folded_regcost,
4968 src_cost, src_regcost) <= 0
4969 && preferable (src_folded_cost, src_folded_regcost,
4970 src_eqv_cost, src_eqv_regcost) <= 0
4971 && preferable (src_folded_cost, src_folded_regcost,
4972 src_related_cost, src_related_regcost) <= 0
4973 && preferable (src_folded_cost, src_folded_regcost,
4974 src_elt_cost, src_elt_regcost) <= 0)
4976 trial = src_folded, src_folded_cost = MAX_COST;
4977 if (src_folded_force_flag)
4979 rtx forced = force_const_mem (mode, trial);
4980 if (forced)
4981 trial = forced;
4984 else if (src
4985 && preferable (src_cost, src_regcost,
4986 src_eqv_cost, src_eqv_regcost) <= 0
4987 && preferable (src_cost, src_regcost,
4988 src_related_cost, src_related_regcost) <= 0
4989 && preferable (src_cost, src_regcost,
4990 src_elt_cost, src_elt_regcost) <= 0)
4991 trial = src, src_cost = MAX_COST;
4992 else if (src_eqv_here
4993 && preferable (src_eqv_cost, src_eqv_regcost,
4994 src_related_cost, src_related_regcost) <= 0
4995 && preferable (src_eqv_cost, src_eqv_regcost,
4996 src_elt_cost, src_elt_regcost) <= 0)
4997 trial = src_eqv_here, src_eqv_cost = MAX_COST;
4998 else if (src_related
4999 && preferable (src_related_cost, src_related_regcost,
5000 src_elt_cost, src_elt_regcost) <= 0)
5001 trial = src_related, src_related_cost = MAX_COST;
5002 else
5004 trial = elt->exp;
5005 elt = elt->next_same_value;
5006 src_elt_cost = MAX_COST;
5009 /* Avoid creation of overlapping memory moves. */
5010 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5012 rtx src, dest;
5014 /* BLKmode moves are not handled by cse anyway. */
5015 if (GET_MODE (trial) == BLKmode)
5016 break;
5018 src = canon_rtx (trial);
5019 dest = canon_rtx (SET_DEST (sets[i].rtl));
5021 if (!MEM_P (src) || !MEM_P (dest)
5022 || !nonoverlapping_memrefs_p (src, dest))
5023 break;
5026 /* We don't normally have an insn matching (set (pc) (pc)), so
5027 check for this separately here. We will delete such an
5028 insn below.
5030 For other cases such as a table jump or conditional jump
5031 where we know the ultimate target, go ahead and replace the
5032 operand. While that may not make a valid insn, we will
5033 reemit the jump below (and also insert any necessary
5034 barriers). */
5035 if (n_sets == 1 && dest == pc_rtx
5036 && (trial == pc_rtx
5037 || (GET_CODE (trial) == LABEL_REF
5038 && ! condjump_p (insn))))
5040 /* Don't substitute non-local labels, this confuses CFG. */
5041 if (GET_CODE (trial) == LABEL_REF
5042 && LABEL_REF_NONLOCAL_P (trial))
5043 continue;
5045 SET_SRC (sets[i].rtl) = trial;
5046 cse_jumps_altered = true;
5047 break;
5050 /* Reject certain invalid forms of CONST that we create. */
5051 else if (CONSTANT_P (trial)
5052 && GET_CODE (trial) == CONST
5053 /* Reject cases that will cause decode_rtx_const to
5054 die. On the alpha when simplifying a switch, we
5055 get (const (truncate (minus (label_ref)
5056 (label_ref)))). */
5057 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5058 /* Likewise on IA-64, except without the
5059 truncate. */
5060 || (GET_CODE (XEXP (trial, 0)) == MINUS
5061 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5062 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5063 /* Do nothing for this case. */
5066 /* Look for a substitution that makes a valid insn. */
5067 else if (validate_unshare_change
5068 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5070 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5072 /* The result of apply_change_group can be ignored; see
5073 canon_reg. */
5075 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5076 apply_change_group ();
5078 break;
5081 /* If we previously found constant pool entries for
5082 constants and this is a constant, try making a
5083 pool entry. Put it in src_folded unless we already have done
5084 this since that is where it likely came from. */
5086 else if (constant_pool_entries_cost
5087 && CONSTANT_P (trial)
5088 && (src_folded == 0
5089 || (!MEM_P (src_folded)
5090 && ! src_folded_force_flag))
5091 && GET_MODE_CLASS (mode) != MODE_CC
5092 && mode != VOIDmode)
5094 src_folded_force_flag = 1;
5095 src_folded = trial;
5096 src_folded_cost = constant_pool_entries_cost;
5097 src_folded_regcost = constant_pool_entries_regcost;
5101 src = SET_SRC (sets[i].rtl);
5103 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5104 However, there is an important exception: If both are registers
5105 that are not the head of their equivalence class, replace SET_SRC
5106 with the head of the class. If we do not do this, we will have
5107 both registers live over a portion of the basic block. This way,
5108 their lifetimes will likely abut instead of overlapping. */
5109 if (REG_P (dest)
5110 && REGNO_QTY_VALID_P (REGNO (dest)))
5112 int dest_q = REG_QTY (REGNO (dest));
5113 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5115 if (dest_ent->mode == GET_MODE (dest)
5116 && dest_ent->first_reg != REGNO (dest)
5117 && REG_P (src) && REGNO (src) == REGNO (dest)
5118 /* Don't do this if the original insn had a hard reg as
5119 SET_SRC or SET_DEST. */
5120 && (!REG_P (sets[i].src)
5121 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5122 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5123 /* We can't call canon_reg here because it won't do anything if
5124 SRC is a hard register. */
5126 int src_q = REG_QTY (REGNO (src));
5127 struct qty_table_elem *src_ent = &qty_table[src_q];
5128 int first = src_ent->first_reg;
5129 rtx new_src
5130 = (first >= FIRST_PSEUDO_REGISTER
5131 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5133 /* We must use validate-change even for this, because this
5134 might be a special no-op instruction, suitable only to
5135 tag notes onto. */
5136 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5138 src = new_src;
5139 /* If we had a constant that is cheaper than what we are now
5140 setting SRC to, use that constant. We ignored it when we
5141 thought we could make this into a no-op. */
5142 if (src_const && COST (src_const) < COST (src)
5143 && validate_change (insn, &SET_SRC (sets[i].rtl),
5144 src_const, 0))
5145 src = src_const;
5150 /* If we made a change, recompute SRC values. */
5151 if (src != sets[i].src)
5153 do_not_record = 0;
5154 hash_arg_in_memory = 0;
5155 sets[i].src = src;
5156 sets[i].src_hash = HASH (src, mode);
5157 sets[i].src_volatile = do_not_record;
5158 sets[i].src_in_memory = hash_arg_in_memory;
5159 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5162 /* If this is a single SET, we are setting a register, and we have an
5163 equivalent constant, we want to add a REG_NOTE. We don't want
5164 to write a REG_EQUAL note for a constant pseudo since verifying that
5165 that pseudo hasn't been eliminated is a pain. Such a note also
5166 won't help anything.
5168 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5169 which can be created for a reference to a compile time computable
5170 entry in a jump table. */
5172 if (n_sets == 1 && src_const && REG_P (dest)
5173 && !REG_P (src_const)
5174 && ! (GET_CODE (src_const) == CONST
5175 && GET_CODE (XEXP (src_const, 0)) == MINUS
5176 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5177 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5179 /* We only want a REG_EQUAL note if src_const != src. */
5180 if (! rtx_equal_p (src, src_const))
5182 /* Make sure that the rtx is not shared. */
5183 src_const = copy_rtx (src_const);
5185 /* Record the actual constant value in a REG_EQUAL note,
5186 making a new one if one does not already exist. */
5187 set_unique_reg_note (insn, REG_EQUAL, src_const);
5188 df_notes_rescan (insn);
5192 /* Now deal with the destination. */
5193 do_not_record = 0;
5195 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5196 while (GET_CODE (dest) == SUBREG
5197 || GET_CODE (dest) == ZERO_EXTRACT
5198 || GET_CODE (dest) == STRICT_LOW_PART)
5199 dest = XEXP (dest, 0);
5201 sets[i].inner_dest = dest;
5203 if (MEM_P (dest))
5205 #ifdef PUSH_ROUNDING
5206 /* Stack pushes invalidate the stack pointer. */
5207 rtx addr = XEXP (dest, 0);
5208 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5209 && XEXP (addr, 0) == stack_pointer_rtx)
5210 invalidate (stack_pointer_rtx, VOIDmode);
5211 #endif
5212 dest = fold_rtx (dest, insn);
5215 /* Compute the hash code of the destination now,
5216 before the effects of this instruction are recorded,
5217 since the register values used in the address computation
5218 are those before this instruction. */
5219 sets[i].dest_hash = HASH (dest, mode);
5221 /* Don't enter a bit-field in the hash table
5222 because the value in it after the store
5223 may not equal what was stored, due to truncation. */
5225 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5227 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5229 if (src_const != 0 && CONST_INT_P (src_const)
5230 && CONST_INT_P (width)
5231 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5232 && ! (INTVAL (src_const)
5233 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5234 /* Exception: if the value is constant,
5235 and it won't be truncated, record it. */
5237 else
5239 /* This is chosen so that the destination will be invalidated
5240 but no new value will be recorded.
5241 We must invalidate because sometimes constant
5242 values can be recorded for bitfields. */
5243 sets[i].src_elt = 0;
5244 sets[i].src_volatile = 1;
5245 src_eqv = 0;
5246 src_eqv_elt = 0;
5250 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5251 the insn. */
5252 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5254 /* One less use of the label this insn used to jump to. */
5255 delete_insn_and_edges (insn);
5256 cse_jumps_altered = true;
5257 /* No more processing for this set. */
5258 sets[i].rtl = 0;
5261 /* If this SET is now setting PC to a label, we know it used to
5262 be a conditional or computed branch. */
5263 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5264 && !LABEL_REF_NONLOCAL_P (src))
5266 /* We reemit the jump in as many cases as possible just in
5267 case the form of an unconditional jump is significantly
5268 different than a computed jump or conditional jump.
5270 If this insn has multiple sets, then reemitting the
5271 jump is nontrivial. So instead we just force rerecognition
5272 and hope for the best. */
5273 if (n_sets == 1)
5275 rtx new_rtx, note;
5277 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5278 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5279 LABEL_NUSES (XEXP (src, 0))++;
5281 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5282 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5283 if (note)
5285 XEXP (note, 1) = NULL_RTX;
5286 REG_NOTES (new_rtx) = note;
5289 delete_insn_and_edges (insn);
5290 insn = new_rtx;
5292 else
5293 INSN_CODE (insn) = -1;
5295 /* Do not bother deleting any unreachable code, let jump do it. */
5296 cse_jumps_altered = true;
5297 sets[i].rtl = 0;
5300 /* If destination is volatile, invalidate it and then do no further
5301 processing for this assignment. */
5303 else if (do_not_record)
5305 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5306 invalidate (dest, VOIDmode);
5307 else if (MEM_P (dest))
5308 invalidate (dest, VOIDmode);
5309 else if (GET_CODE (dest) == STRICT_LOW_PART
5310 || GET_CODE (dest) == ZERO_EXTRACT)
5311 invalidate (XEXP (dest, 0), GET_MODE (dest));
5312 sets[i].rtl = 0;
5315 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5316 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5318 #ifdef HAVE_cc0
5319 /* If setting CC0, record what it was set to, or a constant, if it
5320 is equivalent to a constant. If it is being set to a floating-point
5321 value, make a COMPARE with the appropriate constant of 0. If we
5322 don't do this, later code can interpret this as a test against
5323 const0_rtx, which can cause problems if we try to put it into an
5324 insn as a floating-point operand. */
5325 if (dest == cc0_rtx)
5327 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5328 this_insn_cc0_mode = mode;
5329 if (FLOAT_MODE_P (mode))
5330 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5331 CONST0_RTX (mode));
5333 #endif
5336 /* Now enter all non-volatile source expressions in the hash table
5337 if they are not already present.
5338 Record their equivalence classes in src_elt.
5339 This way we can insert the corresponding destinations into
5340 the same classes even if the actual sources are no longer in them
5341 (having been invalidated). */
5343 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5344 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5346 struct table_elt *elt;
5347 struct table_elt *classp = sets[0].src_elt;
5348 rtx dest = SET_DEST (sets[0].rtl);
5349 enum machine_mode eqvmode = GET_MODE (dest);
5351 if (GET_CODE (dest) == STRICT_LOW_PART)
5353 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5354 classp = 0;
5356 if (insert_regs (src_eqv, classp, 0))
5358 rehash_using_reg (src_eqv);
5359 src_eqv_hash = HASH (src_eqv, eqvmode);
5361 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5362 elt->in_memory = src_eqv_in_memory;
5363 src_eqv_elt = elt;
5365 /* Check to see if src_eqv_elt is the same as a set source which
5366 does not yet have an elt, and if so set the elt of the set source
5367 to src_eqv_elt. */
5368 for (i = 0; i < n_sets; i++)
5369 if (sets[i].rtl && sets[i].src_elt == 0
5370 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5371 sets[i].src_elt = src_eqv_elt;
5374 for (i = 0; i < n_sets; i++)
5375 if (sets[i].rtl && ! sets[i].src_volatile
5376 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5378 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5380 /* REG_EQUAL in setting a STRICT_LOW_PART
5381 gives an equivalent for the entire destination register,
5382 not just for the subreg being stored in now.
5383 This is a more interesting equivalence, so we arrange later
5384 to treat the entire reg as the destination. */
5385 sets[i].src_elt = src_eqv_elt;
5386 sets[i].src_hash = src_eqv_hash;
5388 else
5390 /* Insert source and constant equivalent into hash table, if not
5391 already present. */
5392 struct table_elt *classp = src_eqv_elt;
5393 rtx src = sets[i].src;
5394 rtx dest = SET_DEST (sets[i].rtl);
5395 enum machine_mode mode
5396 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5398 /* It's possible that we have a source value known to be
5399 constant but don't have a REG_EQUAL note on the insn.
5400 Lack of a note will mean src_eqv_elt will be NULL. This
5401 can happen where we've generated a SUBREG to access a
5402 CONST_INT that is already in a register in a wider mode.
5403 Ensure that the source expression is put in the proper
5404 constant class. */
5405 if (!classp)
5406 classp = sets[i].src_const_elt;
5408 if (sets[i].src_elt == 0)
5410 struct table_elt *elt;
5412 /* Note that these insert_regs calls cannot remove
5413 any of the src_elt's, because they would have failed to
5414 match if not still valid. */
5415 if (insert_regs (src, classp, 0))
5417 rehash_using_reg (src);
5418 sets[i].src_hash = HASH (src, mode);
5420 elt = insert (src, classp, sets[i].src_hash, mode);
5421 elt->in_memory = sets[i].src_in_memory;
5422 sets[i].src_elt = classp = elt;
5424 if (sets[i].src_const && sets[i].src_const_elt == 0
5425 && src != sets[i].src_const
5426 && ! rtx_equal_p (sets[i].src_const, src))
5427 sets[i].src_elt = insert (sets[i].src_const, classp,
5428 sets[i].src_const_hash, mode);
5431 else if (sets[i].src_elt == 0)
5432 /* If we did not insert the source into the hash table (e.g., it was
5433 volatile), note the equivalence class for the REG_EQUAL value, if any,
5434 so that the destination goes into that class. */
5435 sets[i].src_elt = src_eqv_elt;
5437 /* Record destination addresses in the hash table. This allows us to
5438 check if they are invalidated by other sets. */
5439 for (i = 0; i < n_sets; i++)
5441 if (sets[i].rtl)
5443 rtx x = sets[i].inner_dest;
5444 struct table_elt *elt;
5445 enum machine_mode mode;
5446 unsigned hash;
5448 if (MEM_P (x))
5450 x = XEXP (x, 0);
5451 mode = GET_MODE (x);
5452 hash = HASH (x, mode);
5453 elt = lookup (x, hash, mode);
5454 if (!elt)
5456 if (insert_regs (x, NULL, 0))
5458 rtx dest = SET_DEST (sets[i].rtl);
5460 rehash_using_reg (x);
5461 hash = HASH (x, mode);
5462 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5464 elt = insert (x, NULL, hash, mode);
5467 sets[i].dest_addr_elt = elt;
5469 else
5470 sets[i].dest_addr_elt = NULL;
5474 invalidate_from_clobbers (x);
5476 /* Some registers are invalidated by subroutine calls. Memory is
5477 invalidated by non-constant calls. */
5479 if (CALL_P (insn))
5481 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5482 invalidate_memory ();
5483 invalidate_for_call ();
5486 /* Now invalidate everything set by this instruction.
5487 If a SUBREG or other funny destination is being set,
5488 sets[i].rtl is still nonzero, so here we invalidate the reg
5489 a part of which is being set. */
5491 for (i = 0; i < n_sets; i++)
5492 if (sets[i].rtl)
5494 /* We can't use the inner dest, because the mode associated with
5495 a ZERO_EXTRACT is significant. */
5496 rtx dest = SET_DEST (sets[i].rtl);
5498 /* Needed for registers to remove the register from its
5499 previous quantity's chain.
5500 Needed for memory if this is a nonvarying address, unless
5501 we have just done an invalidate_memory that covers even those. */
5502 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5503 invalidate (dest, VOIDmode);
5504 else if (MEM_P (dest))
5505 invalidate (dest, VOIDmode);
5506 else if (GET_CODE (dest) == STRICT_LOW_PART
5507 || GET_CODE (dest) == ZERO_EXTRACT)
5508 invalidate (XEXP (dest, 0), GET_MODE (dest));
5511 /* A volatile ASM invalidates everything. */
5512 if (NONJUMP_INSN_P (insn)
5513 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5514 && MEM_VOLATILE_P (PATTERN (insn)))
5515 flush_hash_table ();
5517 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5518 the regs restored by the longjmp come from a later time
5519 than the setjmp. */
5520 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5522 flush_hash_table ();
5523 goto done;
5526 /* Make sure registers mentioned in destinations
5527 are safe for use in an expression to be inserted.
5528 This removes from the hash table
5529 any invalid entry that refers to one of these registers.
5531 We don't care about the return value from mention_regs because
5532 we are going to hash the SET_DEST values unconditionally. */
5534 for (i = 0; i < n_sets; i++)
5536 if (sets[i].rtl)
5538 rtx x = SET_DEST (sets[i].rtl);
5540 if (!REG_P (x))
5541 mention_regs (x);
5542 else
5544 /* We used to rely on all references to a register becoming
5545 inaccessible when a register changes to a new quantity,
5546 since that changes the hash code. However, that is not
5547 safe, since after HASH_SIZE new quantities we get a
5548 hash 'collision' of a register with its own invalid
5549 entries. And since SUBREGs have been changed not to
5550 change their hash code with the hash code of the register,
5551 it wouldn't work any longer at all. So we have to check
5552 for any invalid references lying around now.
5553 This code is similar to the REG case in mention_regs,
5554 but it knows that reg_tick has been incremented, and
5555 it leaves reg_in_table as -1 . */
5556 unsigned int regno = REGNO (x);
5557 unsigned int endregno = END_REGNO (x);
5558 unsigned int i;
5560 for (i = regno; i < endregno; i++)
5562 if (REG_IN_TABLE (i) >= 0)
5564 remove_invalid_refs (i);
5565 REG_IN_TABLE (i) = -1;
5572 /* We may have just removed some of the src_elt's from the hash table.
5573 So replace each one with the current head of the same class.
5574 Also check if destination addresses have been removed. */
5576 for (i = 0; i < n_sets; i++)
5577 if (sets[i].rtl)
5579 if (sets[i].dest_addr_elt
5580 && sets[i].dest_addr_elt->first_same_value == 0)
5582 /* The elt was removed, which means this destination is not
5583 valid after this instruction. */
5584 sets[i].rtl = NULL_RTX;
5586 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5587 /* If elt was removed, find current head of same class,
5588 or 0 if nothing remains of that class. */
5590 struct table_elt *elt = sets[i].src_elt;
5592 while (elt && elt->prev_same_value)
5593 elt = elt->prev_same_value;
5595 while (elt && elt->first_same_value == 0)
5596 elt = elt->next_same_value;
5597 sets[i].src_elt = elt ? elt->first_same_value : 0;
5601 /* Now insert the destinations into their equivalence classes. */
5603 for (i = 0; i < n_sets; i++)
5604 if (sets[i].rtl)
5606 rtx dest = SET_DEST (sets[i].rtl);
5607 struct table_elt *elt;
5609 /* Don't record value if we are not supposed to risk allocating
5610 floating-point values in registers that might be wider than
5611 memory. */
5612 if ((flag_float_store
5613 && MEM_P (dest)
5614 && FLOAT_MODE_P (GET_MODE (dest)))
5615 /* Don't record BLKmode values, because we don't know the
5616 size of it, and can't be sure that other BLKmode values
5617 have the same or smaller size. */
5618 || GET_MODE (dest) == BLKmode
5619 /* If we didn't put a REG_EQUAL value or a source into the hash
5620 table, there is no point is recording DEST. */
5621 || sets[i].src_elt == 0
5622 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5623 or SIGN_EXTEND, don't record DEST since it can cause
5624 some tracking to be wrong.
5626 ??? Think about this more later. */
5627 || (GET_CODE (dest) == SUBREG
5628 && (GET_MODE_SIZE (GET_MODE (dest))
5629 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5630 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5631 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5632 continue;
5634 /* STRICT_LOW_PART isn't part of the value BEING set,
5635 and neither is the SUBREG inside it.
5636 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5637 if (GET_CODE (dest) == STRICT_LOW_PART)
5638 dest = SUBREG_REG (XEXP (dest, 0));
5640 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5641 /* Registers must also be inserted into chains for quantities. */
5642 if (insert_regs (dest, sets[i].src_elt, 1))
5644 /* If `insert_regs' changes something, the hash code must be
5645 recalculated. */
5646 rehash_using_reg (dest);
5647 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5650 elt = insert (dest, sets[i].src_elt,
5651 sets[i].dest_hash, GET_MODE (dest));
5653 /* If this is a constant, insert the constant anchors with the
5654 equivalent register-offset expressions using register DEST. */
5655 if (targetm.const_anchor
5656 && REG_P (dest)
5657 && SCALAR_INT_MODE_P (GET_MODE (dest))
5658 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5659 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5661 elt->in_memory = (MEM_P (sets[i].inner_dest)
5662 && !MEM_READONLY_P (sets[i].inner_dest));
5664 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5665 narrower than M2, and both M1 and M2 are the same number of words,
5666 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5667 make that equivalence as well.
5669 However, BAR may have equivalences for which gen_lowpart
5670 will produce a simpler value than gen_lowpart applied to
5671 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5672 BAR's equivalences. If we don't get a simplified form, make
5673 the SUBREG. It will not be used in an equivalence, but will
5674 cause two similar assignments to be detected.
5676 Note the loop below will find SUBREG_REG (DEST) since we have
5677 already entered SRC and DEST of the SET in the table. */
5679 if (GET_CODE (dest) == SUBREG
5680 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5681 / UNITS_PER_WORD)
5682 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5683 && (GET_MODE_SIZE (GET_MODE (dest))
5684 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5685 && sets[i].src_elt != 0)
5687 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5688 struct table_elt *elt, *classp = 0;
5690 for (elt = sets[i].src_elt->first_same_value; elt;
5691 elt = elt->next_same_value)
5693 rtx new_src = 0;
5694 unsigned src_hash;
5695 struct table_elt *src_elt;
5696 int byte = 0;
5698 /* Ignore invalid entries. */
5699 if (!REG_P (elt->exp)
5700 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5701 continue;
5703 /* We may have already been playing subreg games. If the
5704 mode is already correct for the destination, use it. */
5705 if (GET_MODE (elt->exp) == new_mode)
5706 new_src = elt->exp;
5707 else
5709 /* Calculate big endian correction for the SUBREG_BYTE.
5710 We have already checked that M1 (GET_MODE (dest))
5711 is not narrower than M2 (new_mode). */
5712 if (BYTES_BIG_ENDIAN)
5713 byte = (GET_MODE_SIZE (GET_MODE (dest))
5714 - GET_MODE_SIZE (new_mode));
5716 new_src = simplify_gen_subreg (new_mode, elt->exp,
5717 GET_MODE (dest), byte);
5720 /* The call to simplify_gen_subreg fails if the value
5721 is VOIDmode, yet we can't do any simplification, e.g.
5722 for EXPR_LISTs denoting function call results.
5723 It is invalid to construct a SUBREG with a VOIDmode
5724 SUBREG_REG, hence a zero new_src means we can't do
5725 this substitution. */
5726 if (! new_src)
5727 continue;
5729 src_hash = HASH (new_src, new_mode);
5730 src_elt = lookup (new_src, src_hash, new_mode);
5732 /* Put the new source in the hash table is if isn't
5733 already. */
5734 if (src_elt == 0)
5736 if (insert_regs (new_src, classp, 0))
5738 rehash_using_reg (new_src);
5739 src_hash = HASH (new_src, new_mode);
5741 src_elt = insert (new_src, classp, src_hash, new_mode);
5742 src_elt->in_memory = elt->in_memory;
5744 else if (classp && classp != src_elt->first_same_value)
5745 /* Show that two things that we've seen before are
5746 actually the same. */
5747 merge_equiv_classes (src_elt, classp);
5749 classp = src_elt->first_same_value;
5750 /* Ignore invalid entries. */
5751 while (classp
5752 && !REG_P (classp->exp)
5753 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5754 classp = classp->next_same_value;
5759 /* Special handling for (set REG0 REG1) where REG0 is the
5760 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5761 be used in the sequel, so (if easily done) change this insn to
5762 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5763 that computed their value. Then REG1 will become a dead store
5764 and won't cloud the situation for later optimizations.
5766 Do not make this change if REG1 is a hard register, because it will
5767 then be used in the sequel and we may be changing a two-operand insn
5768 into a three-operand insn.
5770 Also do not do this if we are operating on a copy of INSN. */
5772 if (n_sets == 1 && sets[0].rtl && REG_P (SET_DEST (sets[0].rtl))
5773 && NEXT_INSN (PREV_INSN (insn)) == insn
5774 && REG_P (SET_SRC (sets[0].rtl))
5775 && REGNO (SET_SRC (sets[0].rtl)) >= FIRST_PSEUDO_REGISTER
5776 && REGNO_QTY_VALID_P (REGNO (SET_SRC (sets[0].rtl))))
5778 int src_q = REG_QTY (REGNO (SET_SRC (sets[0].rtl)));
5779 struct qty_table_elem *src_ent = &qty_table[src_q];
5781 if (src_ent->first_reg == REGNO (SET_DEST (sets[0].rtl)))
5783 /* Scan for the previous nonnote insn, but stop at a basic
5784 block boundary. */
5785 rtx prev = insn;
5786 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
5789 prev = PREV_INSN (prev);
5791 while (prev != bb_head && NOTE_P (prev));
5793 /* Do not swap the registers around if the previous instruction
5794 attaches a REG_EQUIV note to REG1.
5796 ??? It's not entirely clear whether we can transfer a REG_EQUIV
5797 from the pseudo that originally shadowed an incoming argument
5798 to another register. Some uses of REG_EQUIV might rely on it
5799 being attached to REG1 rather than REG2.
5801 This section previously turned the REG_EQUIV into a REG_EQUAL
5802 note. We cannot do that because REG_EQUIV may provide an
5803 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
5804 if (NONJUMP_INSN_P (prev)
5805 && GET_CODE (PATTERN (prev)) == SET
5806 && SET_DEST (PATTERN (prev)) == SET_SRC (sets[0].rtl)
5807 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
5809 rtx dest = SET_DEST (sets[0].rtl);
5810 rtx src = SET_SRC (sets[0].rtl);
5811 rtx note;
5813 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
5814 validate_change (insn, &SET_DEST (sets[0].rtl), src, 1);
5815 validate_change (insn, &SET_SRC (sets[0].rtl), dest, 1);
5816 apply_change_group ();
5818 /* If INSN has a REG_EQUAL note, and this note mentions
5819 REG0, then we must delete it, because the value in
5820 REG0 has changed. If the note's value is REG1, we must
5821 also delete it because that is now this insn's dest. */
5822 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5823 if (note != 0
5824 && (reg_mentioned_p (dest, XEXP (note, 0))
5825 || rtx_equal_p (src, XEXP (note, 0))))
5826 remove_note (insn, note);
5831 done:;
5834 /* Remove from the hash table all expressions that reference memory. */
5836 static void
5837 invalidate_memory (void)
5839 int i;
5840 struct table_elt *p, *next;
5842 for (i = 0; i < HASH_SIZE; i++)
5843 for (p = table[i]; p; p = next)
5845 next = p->next_same_hash;
5846 if (p->in_memory)
5847 remove_from_table (p, i);
5851 /* Perform invalidation on the basis of everything about an insn
5852 except for invalidating the actual places that are SET in it.
5853 This includes the places CLOBBERed, and anything that might
5854 alias with something that is SET or CLOBBERed.
5856 X is the pattern of the insn. */
5858 static void
5859 invalidate_from_clobbers (rtx x)
5861 if (GET_CODE (x) == CLOBBER)
5863 rtx ref = XEXP (x, 0);
5864 if (ref)
5866 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5867 || MEM_P (ref))
5868 invalidate (ref, VOIDmode);
5869 else if (GET_CODE (ref) == STRICT_LOW_PART
5870 || GET_CODE (ref) == ZERO_EXTRACT)
5871 invalidate (XEXP (ref, 0), GET_MODE (ref));
5874 else if (GET_CODE (x) == PARALLEL)
5876 int i;
5877 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5879 rtx y = XVECEXP (x, 0, i);
5880 if (GET_CODE (y) == CLOBBER)
5882 rtx ref = XEXP (y, 0);
5883 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5884 || MEM_P (ref))
5885 invalidate (ref, VOIDmode);
5886 else if (GET_CODE (ref) == STRICT_LOW_PART
5887 || GET_CODE (ref) == ZERO_EXTRACT)
5888 invalidate (XEXP (ref, 0), GET_MODE (ref));
5894 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
5895 and replace any registers in them with either an equivalent constant
5896 or the canonical form of the register. If we are inside an address,
5897 only do this if the address remains valid.
5899 OBJECT is 0 except when within a MEM in which case it is the MEM.
5901 Return the replacement for X. */
5903 static rtx
5904 cse_process_notes_1 (rtx x, rtx object, bool *changed)
5906 enum rtx_code code = GET_CODE (x);
5907 const char *fmt = GET_RTX_FORMAT (code);
5908 int i;
5910 switch (code)
5912 case CONST_INT:
5913 case CONST:
5914 case SYMBOL_REF:
5915 case LABEL_REF:
5916 case CONST_DOUBLE:
5917 case CONST_FIXED:
5918 case CONST_VECTOR:
5919 case PC:
5920 case CC0:
5921 case LO_SUM:
5922 return x;
5924 case MEM:
5925 validate_change (x, &XEXP (x, 0),
5926 cse_process_notes (XEXP (x, 0), x, changed), 0);
5927 return x;
5929 case EXPR_LIST:
5930 case INSN_LIST:
5931 if (REG_NOTE_KIND (x) == REG_EQUAL)
5932 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
5933 if (XEXP (x, 1))
5934 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
5935 return x;
5937 case SIGN_EXTEND:
5938 case ZERO_EXTEND:
5939 case SUBREG:
5941 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
5942 /* We don't substitute VOIDmode constants into these rtx,
5943 since they would impede folding. */
5944 if (GET_MODE (new_rtx) != VOIDmode)
5945 validate_change (object, &XEXP (x, 0), new_rtx, 0);
5946 return x;
5949 case REG:
5950 i = REG_QTY (REGNO (x));
5952 /* Return a constant or a constant register. */
5953 if (REGNO_QTY_VALID_P (REGNO (x)))
5955 struct qty_table_elem *ent = &qty_table[i];
5957 if (ent->const_rtx != NULL_RTX
5958 && (CONSTANT_P (ent->const_rtx)
5959 || REG_P (ent->const_rtx)))
5961 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
5962 if (new_rtx)
5963 return copy_rtx (new_rtx);
5967 /* Otherwise, canonicalize this register. */
5968 return canon_reg (x, NULL_RTX);
5970 default:
5971 break;
5974 for (i = 0; i < GET_RTX_LENGTH (code); i++)
5975 if (fmt[i] == 'e')
5976 validate_change (object, &XEXP (x, i),
5977 cse_process_notes (XEXP (x, i), object, changed), 0);
5979 return x;
5982 static rtx
5983 cse_process_notes (rtx x, rtx object, bool *changed)
5985 rtx new_rtx = cse_process_notes_1 (x, object, changed);
5986 if (new_rtx != x)
5987 *changed = true;
5988 return new_rtx;
5992 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
5994 DATA is a pointer to a struct cse_basic_block_data, that is used to
5995 describe the path.
5996 It is filled with a queue of basic blocks, starting with FIRST_BB
5997 and following a trace through the CFG.
5999 If all paths starting at FIRST_BB have been followed, or no new path
6000 starting at FIRST_BB can be constructed, this function returns FALSE.
6001 Otherwise, DATA->path is filled and the function returns TRUE indicating
6002 that a path to follow was found.
6004 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6005 block in the path will be FIRST_BB. */
6007 static bool
6008 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6009 int follow_jumps)
6011 basic_block bb;
6012 edge e;
6013 int path_size;
6015 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6017 /* See if there is a previous path. */
6018 path_size = data->path_size;
6020 /* There is a previous path. Make sure it started with FIRST_BB. */
6021 if (path_size)
6022 gcc_assert (data->path[0].bb == first_bb);
6024 /* There was only one basic block in the last path. Clear the path and
6025 return, so that paths starting at another basic block can be tried. */
6026 if (path_size == 1)
6028 path_size = 0;
6029 goto done;
6032 /* If the path was empty from the beginning, construct a new path. */
6033 if (path_size == 0)
6034 data->path[path_size++].bb = first_bb;
6035 else
6037 /* Otherwise, path_size must be equal to or greater than 2, because
6038 a previous path exists that is at least two basic blocks long.
6040 Update the previous branch path, if any. If the last branch was
6041 previously along the branch edge, take the fallthrough edge now. */
6042 while (path_size >= 2)
6044 basic_block last_bb_in_path, previous_bb_in_path;
6045 edge e;
6047 --path_size;
6048 last_bb_in_path = data->path[path_size].bb;
6049 previous_bb_in_path = data->path[path_size - 1].bb;
6051 /* If we previously followed a path along the branch edge, try
6052 the fallthru edge now. */
6053 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6054 && any_condjump_p (BB_END (previous_bb_in_path))
6055 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6056 && e == BRANCH_EDGE (previous_bb_in_path))
6058 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6059 if (bb != EXIT_BLOCK_PTR
6060 && single_pred_p (bb)
6061 /* We used to assert here that we would only see blocks
6062 that we have not visited yet. But we may end up
6063 visiting basic blocks twice if the CFG has changed
6064 in this run of cse_main, because when the CFG changes
6065 the topological sort of the CFG also changes. A basic
6066 blocks that previously had more than two predecessors
6067 may now have a single predecessor, and become part of
6068 a path that starts at another basic block.
6070 We still want to visit each basic block only once, so
6071 halt the path here if we have already visited BB. */
6072 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6074 SET_BIT (cse_visited_basic_blocks, bb->index);
6075 data->path[path_size++].bb = bb;
6076 break;
6080 data->path[path_size].bb = NULL;
6083 /* If only one block remains in the path, bail. */
6084 if (path_size == 1)
6086 path_size = 0;
6087 goto done;
6091 /* Extend the path if possible. */
6092 if (follow_jumps)
6094 bb = data->path[path_size - 1].bb;
6095 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6097 if (single_succ_p (bb))
6098 e = single_succ_edge (bb);
6099 else if (EDGE_COUNT (bb->succs) == 2
6100 && any_condjump_p (BB_END (bb)))
6102 /* First try to follow the branch. If that doesn't lead
6103 to a useful path, follow the fallthru edge. */
6104 e = BRANCH_EDGE (bb);
6105 if (!single_pred_p (e->dest))
6106 e = FALLTHRU_EDGE (bb);
6108 else
6109 e = NULL;
6111 if (e && e->dest != EXIT_BLOCK_PTR
6112 && single_pred_p (e->dest)
6113 /* Avoid visiting basic blocks twice. The large comment
6114 above explains why this can happen. */
6115 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6117 basic_block bb2 = e->dest;
6118 SET_BIT (cse_visited_basic_blocks, bb2->index);
6119 data->path[path_size++].bb = bb2;
6120 bb = bb2;
6122 else
6123 bb = NULL;
6127 done:
6128 data->path_size = path_size;
6129 return path_size != 0;
6132 /* Dump the path in DATA to file F. NSETS is the number of sets
6133 in the path. */
6135 static void
6136 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6138 int path_entry;
6140 fprintf (f, ";; Following path with %d sets: ", nsets);
6141 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6142 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6143 fputc ('\n', dump_file);
6144 fflush (f);
6148 /* Return true if BB has exception handling successor edges. */
6150 static bool
6151 have_eh_succ_edges (basic_block bb)
6153 edge e;
6154 edge_iterator ei;
6156 FOR_EACH_EDGE (e, ei, bb->succs)
6157 if (e->flags & EDGE_EH)
6158 return true;
6160 return false;
6164 /* Scan to the end of the path described by DATA. Return an estimate of
6165 the total number of SETs of all insns in the path. */
6167 static void
6168 cse_prescan_path (struct cse_basic_block_data *data)
6170 int nsets = 0;
6171 int path_size = data->path_size;
6172 int path_entry;
6174 /* Scan to end of each basic block in the path. */
6175 for (path_entry = 0; path_entry < path_size; path_entry++)
6177 basic_block bb;
6178 rtx insn;
6180 bb = data->path[path_entry].bb;
6182 FOR_BB_INSNS (bb, insn)
6184 if (!INSN_P (insn))
6185 continue;
6187 /* A PARALLEL can have lots of SETs in it,
6188 especially if it is really an ASM_OPERANDS. */
6189 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6190 nsets += XVECLEN (PATTERN (insn), 0);
6191 else
6192 nsets += 1;
6196 data->nsets = nsets;
6199 /* Process a single extended basic block described by EBB_DATA. */
6201 static void
6202 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6204 int path_size = ebb_data->path_size;
6205 int path_entry;
6206 int num_insns = 0;
6208 /* Allocate the space needed by qty_table. */
6209 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6211 new_basic_block ();
6212 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6213 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6214 for (path_entry = 0; path_entry < path_size; path_entry++)
6216 basic_block bb;
6217 rtx insn;
6219 bb = ebb_data->path[path_entry].bb;
6221 /* Invalidate recorded information for eh regs if there is an EH
6222 edge pointing to that bb. */
6223 if (bb_has_eh_pred (bb))
6225 df_ref *def_rec;
6227 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6229 df_ref def = *def_rec;
6230 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6231 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6235 FOR_BB_INSNS (bb, insn)
6237 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6238 /* If we have processed 1,000 insns, flush the hash table to
6239 avoid extreme quadratic behavior. We must not include NOTEs
6240 in the count since there may be more of them when generating
6241 debugging information. If we clear the table at different
6242 times, code generated with -g -O might be different than code
6243 generated with -O but not -g.
6245 FIXME: This is a real kludge and needs to be done some other
6246 way. */
6247 if (INSN_P (insn)
6248 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6250 flush_hash_table ();
6251 num_insns = 0;
6254 if (INSN_P (insn))
6256 /* Process notes first so we have all notes in canonical forms
6257 when looking for duplicate operations. */
6258 if (REG_NOTES (insn))
6260 bool changed = false;
6261 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6262 NULL_RTX, &changed);
6263 if (changed)
6264 df_notes_rescan (insn);
6267 cse_insn (insn);
6269 /* If we haven't already found an insn where we added a LABEL_REF,
6270 check this one. */
6271 if (INSN_P (insn) && !recorded_label_ref
6272 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6273 (void *) insn))
6274 recorded_label_ref = true;
6276 #ifdef HAVE_cc0
6277 /* If the previous insn set CC0 and this insn no longer
6278 references CC0, delete the previous insn. Here we use
6279 fact that nothing expects CC0 to be valid over an insn,
6280 which is true until the final pass. */
6282 rtx prev_insn, tem;
6284 prev_insn = PREV_INSN (insn);
6285 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6286 && (tem = single_set (prev_insn)) != 0
6287 && SET_DEST (tem) == cc0_rtx
6288 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6289 delete_insn (prev_insn);
6292 /* If this insn is not the last insn in the basic block,
6293 it will be PREV_INSN(insn) in the next iteration. If
6294 we recorded any CC0-related information for this insn,
6295 remember it. */
6296 if (insn != BB_END (bb))
6298 prev_insn_cc0 = this_insn_cc0;
6299 prev_insn_cc0_mode = this_insn_cc0_mode;
6301 #endif
6305 /* With non-call exceptions, we are not always able to update
6306 the CFG properly inside cse_insn. So clean up possibly
6307 redundant EH edges here. */
6308 if (flag_non_call_exceptions && have_eh_succ_edges (bb))
6309 cse_cfg_altered |= purge_dead_edges (bb);
6311 /* If we changed a conditional jump, we may have terminated
6312 the path we are following. Check that by verifying that
6313 the edge we would take still exists. If the edge does
6314 not exist anymore, purge the remainder of the path.
6315 Note that this will cause us to return to the caller. */
6316 if (path_entry < path_size - 1)
6318 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6319 if (!find_edge (bb, next_bb))
6323 path_size--;
6325 /* If we truncate the path, we must also reset the
6326 visited bit on the remaining blocks in the path,
6327 or we will never visit them at all. */
6328 RESET_BIT (cse_visited_basic_blocks,
6329 ebb_data->path[path_size].bb->index);
6330 ebb_data->path[path_size].bb = NULL;
6332 while (path_size - 1 != path_entry);
6333 ebb_data->path_size = path_size;
6337 /* If this is a conditional jump insn, record any known
6338 equivalences due to the condition being tested. */
6339 insn = BB_END (bb);
6340 if (path_entry < path_size - 1
6341 && JUMP_P (insn)
6342 && single_set (insn)
6343 && any_condjump_p (insn))
6345 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6346 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6347 record_jump_equiv (insn, taken);
6350 #ifdef HAVE_cc0
6351 /* Clear the CC0-tracking related insns, they can't provide
6352 useful information across basic block boundaries. */
6353 prev_insn_cc0 = 0;
6354 #endif
6357 gcc_assert (next_qty <= max_qty);
6359 free (qty_table);
6363 /* Perform cse on the instructions of a function.
6364 F is the first instruction.
6365 NREGS is one plus the highest pseudo-reg number used in the instruction.
6367 Return 2 if jump optimizations should be redone due to simplifications
6368 in conditional jump instructions.
6369 Return 1 if the CFG should be cleaned up because it has been modified.
6370 Return 0 otherwise. */
6373 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6375 struct cse_basic_block_data ebb_data;
6376 basic_block bb;
6377 int *rc_order = XNEWVEC (int, last_basic_block);
6378 int i, n_blocks;
6380 df_set_flags (DF_LR_RUN_DCE);
6381 df_analyze ();
6382 df_set_flags (DF_DEFER_INSN_RESCAN);
6384 reg_scan (get_insns (), max_reg_num ());
6385 init_cse_reg_info (nregs);
6387 ebb_data.path = XNEWVEC (struct branch_path,
6388 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6390 cse_cfg_altered = false;
6391 cse_jumps_altered = false;
6392 recorded_label_ref = false;
6393 constant_pool_entries_cost = 0;
6394 constant_pool_entries_regcost = 0;
6395 ebb_data.path_size = 0;
6396 ebb_data.nsets = 0;
6397 rtl_hooks = cse_rtl_hooks;
6399 init_recog ();
6400 init_alias_analysis ();
6402 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6404 /* Set up the table of already visited basic blocks. */
6405 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6406 sbitmap_zero (cse_visited_basic_blocks);
6408 /* Loop over basic blocks in reverse completion order (RPO),
6409 excluding the ENTRY and EXIT blocks. */
6410 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6411 i = 0;
6412 while (i < n_blocks)
6414 /* Find the first block in the RPO queue that we have not yet
6415 processed before. */
6418 bb = BASIC_BLOCK (rc_order[i++]);
6420 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6421 && i < n_blocks);
6423 /* Find all paths starting with BB, and process them. */
6424 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6426 /* Pre-scan the path. */
6427 cse_prescan_path (&ebb_data);
6429 /* If this basic block has no sets, skip it. */
6430 if (ebb_data.nsets == 0)
6431 continue;
6433 /* Get a reasonable estimate for the maximum number of qty's
6434 needed for this path. For this, we take the number of sets
6435 and multiply that by MAX_RECOG_OPERANDS. */
6436 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6438 /* Dump the path we're about to process. */
6439 if (dump_file)
6440 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6442 cse_extended_basic_block (&ebb_data);
6446 /* Clean up. */
6447 end_alias_analysis ();
6448 free (reg_eqv_table);
6449 free (ebb_data.path);
6450 sbitmap_free (cse_visited_basic_blocks);
6451 free (rc_order);
6452 rtl_hooks = general_rtl_hooks;
6454 if (cse_jumps_altered || recorded_label_ref)
6455 return 2;
6456 else if (cse_cfg_altered)
6457 return 1;
6458 else
6459 return 0;
6462 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6463 which there isn't a REG_LABEL_OPERAND note.
6464 Return one if so. DATA is the insn. */
6466 static int
6467 check_for_label_ref (rtx *rtl, void *data)
6469 rtx insn = (rtx) data;
6471 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6472 note for it, we must rerun jump since it needs to place the note. If
6473 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6474 don't do this since no REG_LABEL_OPERAND will be added. */
6475 return (GET_CODE (*rtl) == LABEL_REF
6476 && ! LABEL_REF_NONLOCAL_P (*rtl)
6477 && (!JUMP_P (insn)
6478 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6479 && LABEL_P (XEXP (*rtl, 0))
6480 && INSN_UID (XEXP (*rtl, 0)) != 0
6481 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6484 /* Count the number of times registers are used (not set) in X.
6485 COUNTS is an array in which we accumulate the count, INCR is how much
6486 we count each register usage.
6488 Don't count a usage of DEST, which is the SET_DEST of a SET which
6489 contains X in its SET_SRC. This is because such a SET does not
6490 modify the liveness of DEST.
6491 DEST is set to pc_rtx for a trapping insn, which means that we must count
6492 uses of a SET_DEST regardless because the insn can't be deleted here. */
6494 static void
6495 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6497 enum rtx_code code;
6498 rtx note;
6499 const char *fmt;
6500 int i, j;
6502 if (x == 0)
6503 return;
6505 switch (code = GET_CODE (x))
6507 case REG:
6508 if (x != dest)
6509 counts[REGNO (x)] += incr;
6510 return;
6512 case PC:
6513 case CC0:
6514 case CONST:
6515 case CONST_INT:
6516 case CONST_DOUBLE:
6517 case CONST_FIXED:
6518 case CONST_VECTOR:
6519 case SYMBOL_REF:
6520 case LABEL_REF:
6521 return;
6523 case CLOBBER:
6524 /* If we are clobbering a MEM, mark any registers inside the address
6525 as being used. */
6526 if (MEM_P (XEXP (x, 0)))
6527 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6528 return;
6530 case SET:
6531 /* Unless we are setting a REG, count everything in SET_DEST. */
6532 if (!REG_P (SET_DEST (x)))
6533 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6534 count_reg_usage (SET_SRC (x), counts,
6535 dest ? dest : SET_DEST (x),
6536 incr);
6537 return;
6539 case CALL_INSN:
6540 case INSN:
6541 case JUMP_INSN:
6542 /* We expect dest to be NULL_RTX here. If the insn may trap, mark
6543 this fact by setting DEST to pc_rtx. */
6544 if (flag_non_call_exceptions && may_trap_p (PATTERN (x)))
6545 dest = pc_rtx;
6546 if (code == CALL_INSN)
6547 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6548 count_reg_usage (PATTERN (x), counts, dest, incr);
6550 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6551 use them. */
6553 note = find_reg_equal_equiv_note (x);
6554 if (note)
6556 rtx eqv = XEXP (note, 0);
6558 if (GET_CODE (eqv) == EXPR_LIST)
6559 /* This REG_EQUAL note describes the result of a function call.
6560 Process all the arguments. */
6563 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6564 eqv = XEXP (eqv, 1);
6566 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6567 else
6568 count_reg_usage (eqv, counts, dest, incr);
6570 return;
6572 case EXPR_LIST:
6573 if (REG_NOTE_KIND (x) == REG_EQUAL
6574 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6575 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6576 involving registers in the address. */
6577 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6578 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6580 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6581 return;
6583 case ASM_OPERANDS:
6584 /* If the asm is volatile, then this insn cannot be deleted,
6585 and so the inputs *must* be live. */
6586 if (MEM_VOLATILE_P (x))
6587 dest = NULL_RTX;
6588 /* Iterate over just the inputs, not the constraints as well. */
6589 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6590 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6591 return;
6593 case INSN_LIST:
6594 gcc_unreachable ();
6596 default:
6597 break;
6600 fmt = GET_RTX_FORMAT (code);
6601 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6603 if (fmt[i] == 'e')
6604 count_reg_usage (XEXP (x, i), counts, dest, incr);
6605 else if (fmt[i] == 'E')
6606 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6607 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6611 /* Return true if set is live. */
6612 static bool
6613 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6614 int *counts)
6616 #ifdef HAVE_cc0
6617 rtx tem;
6618 #endif
6620 if (set_noop_p (set))
6623 #ifdef HAVE_cc0
6624 else if (GET_CODE (SET_DEST (set)) == CC0
6625 && !side_effects_p (SET_SRC (set))
6626 && ((tem = next_nonnote_insn (insn)) == 0
6627 || !INSN_P (tem)
6628 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6629 return false;
6630 #endif
6631 else if (!REG_P (SET_DEST (set))
6632 || REGNO (SET_DEST (set)) < FIRST_PSEUDO_REGISTER
6633 || counts[REGNO (SET_DEST (set))] != 0
6634 || side_effects_p (SET_SRC (set)))
6635 return true;
6636 return false;
6639 /* Return true if insn is live. */
6641 static bool
6642 insn_live_p (rtx insn, int *counts)
6644 int i;
6645 if (flag_non_call_exceptions && may_trap_p (PATTERN (insn)))
6646 return true;
6647 else if (GET_CODE (PATTERN (insn)) == SET)
6648 return set_live_p (PATTERN (insn), insn, counts);
6649 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6651 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6653 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6655 if (GET_CODE (elt) == SET)
6657 if (set_live_p (elt, insn, counts))
6658 return true;
6660 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6661 return true;
6663 return false;
6665 else
6666 return true;
6669 /* Scan all the insns and delete any that are dead; i.e., they store a register
6670 that is never used or they copy a register to itself.
6672 This is used to remove insns made obviously dead by cse, loop or other
6673 optimizations. It improves the heuristics in loop since it won't try to
6674 move dead invariants out of loops or make givs for dead quantities. The
6675 remaining passes of the compilation are also sped up. */
6678 delete_trivially_dead_insns (rtx insns, int nreg)
6680 int *counts;
6681 rtx insn, prev;
6682 int ndead = 0;
6684 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6685 /* First count the number of times each register is used. */
6686 counts = XCNEWVEC (int, nreg);
6687 for (insn = insns; insn; insn = NEXT_INSN (insn))
6688 if (INSN_P (insn))
6689 count_reg_usage (insn, counts, NULL_RTX, 1);
6691 /* Go from the last insn to the first and delete insns that only set unused
6692 registers or copy a register to itself. As we delete an insn, remove
6693 usage counts for registers it uses.
6695 The first jump optimization pass may leave a real insn as the last
6696 insn in the function. We must not skip that insn or we may end
6697 up deleting code that is not really dead. */
6698 for (insn = get_last_insn (); insn; insn = prev)
6700 int live_insn = 0;
6702 prev = PREV_INSN (insn);
6703 if (!INSN_P (insn))
6704 continue;
6706 live_insn = insn_live_p (insn, counts);
6708 /* If this is a dead insn, delete it and show registers in it aren't
6709 being used. */
6711 if (! live_insn && dbg_cnt (delete_trivial_dead))
6713 count_reg_usage (insn, counts, NULL_RTX, -1);
6714 delete_insn_and_edges (insn);
6715 ndead++;
6719 if (dump_file && ndead)
6720 fprintf (dump_file, "Deleted %i trivially dead insns\n",
6721 ndead);
6722 /* Clean up. */
6723 free (counts);
6724 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
6725 return ndead;
6728 /* This function is called via for_each_rtx. The argument, NEWREG, is
6729 a condition code register with the desired mode. If we are looking
6730 at the same register in a different mode, replace it with
6731 NEWREG. */
6733 static int
6734 cse_change_cc_mode (rtx *loc, void *data)
6736 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
6738 if (*loc
6739 && REG_P (*loc)
6740 && REGNO (*loc) == REGNO (args->newreg)
6741 && GET_MODE (*loc) != GET_MODE (args->newreg))
6743 validate_change (args->insn, loc, args->newreg, 1);
6745 return -1;
6747 return 0;
6750 /* Change the mode of any reference to the register REGNO (NEWREG) to
6751 GET_MODE (NEWREG) in INSN. */
6753 static void
6754 cse_change_cc_mode_insn (rtx insn, rtx newreg)
6756 struct change_cc_mode_args args;
6757 int success;
6759 if (!INSN_P (insn))
6760 return;
6762 args.insn = insn;
6763 args.newreg = newreg;
6765 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
6766 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
6768 /* If the following assertion was triggered, there is most probably
6769 something wrong with the cc_modes_compatible back end function.
6770 CC modes only can be considered compatible if the insn - with the mode
6771 replaced by any of the compatible modes - can still be recognized. */
6772 success = apply_change_group ();
6773 gcc_assert (success);
6776 /* Change the mode of any reference to the register REGNO (NEWREG) to
6777 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
6778 any instruction which modifies NEWREG. */
6780 static void
6781 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
6783 rtx insn;
6785 for (insn = start; insn != end; insn = NEXT_INSN (insn))
6787 if (! INSN_P (insn))
6788 continue;
6790 if (reg_set_p (newreg, insn))
6791 return;
6793 cse_change_cc_mode_insn (insn, newreg);
6797 /* BB is a basic block which finishes with CC_REG as a condition code
6798 register which is set to CC_SRC. Look through the successors of BB
6799 to find blocks which have a single predecessor (i.e., this one),
6800 and look through those blocks for an assignment to CC_REG which is
6801 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
6802 permitted to change the mode of CC_SRC to a compatible mode. This
6803 returns VOIDmode if no equivalent assignments were found.
6804 Otherwise it returns the mode which CC_SRC should wind up with.
6805 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
6806 but is passed unmodified down to recursive calls in order to prevent
6807 endless recursion.
6809 The main complexity in this function is handling the mode issues.
6810 We may have more than one duplicate which we can eliminate, and we
6811 try to find a mode which will work for multiple duplicates. */
6813 static enum machine_mode
6814 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
6815 bool can_change_mode)
6817 bool found_equiv;
6818 enum machine_mode mode;
6819 unsigned int insn_count;
6820 edge e;
6821 rtx insns[2];
6822 enum machine_mode modes[2];
6823 rtx last_insns[2];
6824 unsigned int i;
6825 rtx newreg;
6826 edge_iterator ei;
6828 /* We expect to have two successors. Look at both before picking
6829 the final mode for the comparison. If we have more successors
6830 (i.e., some sort of table jump, although that seems unlikely),
6831 then we require all beyond the first two to use the same
6832 mode. */
6834 found_equiv = false;
6835 mode = GET_MODE (cc_src);
6836 insn_count = 0;
6837 FOR_EACH_EDGE (e, ei, bb->succs)
6839 rtx insn;
6840 rtx end;
6842 if (e->flags & EDGE_COMPLEX)
6843 continue;
6845 if (EDGE_COUNT (e->dest->preds) != 1
6846 || e->dest == EXIT_BLOCK_PTR
6847 /* Avoid endless recursion on unreachable blocks. */
6848 || e->dest == orig_bb)
6849 continue;
6851 end = NEXT_INSN (BB_END (e->dest));
6852 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
6854 rtx set;
6856 if (! INSN_P (insn))
6857 continue;
6859 /* If CC_SRC is modified, we have to stop looking for
6860 something which uses it. */
6861 if (modified_in_p (cc_src, insn))
6862 break;
6864 /* Check whether INSN sets CC_REG to CC_SRC. */
6865 set = single_set (insn);
6866 if (set
6867 && REG_P (SET_DEST (set))
6868 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
6870 bool found;
6871 enum machine_mode set_mode;
6872 enum machine_mode comp_mode;
6874 found = false;
6875 set_mode = GET_MODE (SET_SRC (set));
6876 comp_mode = set_mode;
6877 if (rtx_equal_p (cc_src, SET_SRC (set)))
6878 found = true;
6879 else if (GET_CODE (cc_src) == COMPARE
6880 && GET_CODE (SET_SRC (set)) == COMPARE
6881 && mode != set_mode
6882 && rtx_equal_p (XEXP (cc_src, 0),
6883 XEXP (SET_SRC (set), 0))
6884 && rtx_equal_p (XEXP (cc_src, 1),
6885 XEXP (SET_SRC (set), 1)))
6888 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
6889 if (comp_mode != VOIDmode
6890 && (can_change_mode || comp_mode == mode))
6891 found = true;
6894 if (found)
6896 found_equiv = true;
6897 if (insn_count < ARRAY_SIZE (insns))
6899 insns[insn_count] = insn;
6900 modes[insn_count] = set_mode;
6901 last_insns[insn_count] = end;
6902 ++insn_count;
6904 if (mode != comp_mode)
6906 gcc_assert (can_change_mode);
6907 mode = comp_mode;
6909 /* The modified insn will be re-recognized later. */
6910 PUT_MODE (cc_src, mode);
6913 else
6915 if (set_mode != mode)
6917 /* We found a matching expression in the
6918 wrong mode, but we don't have room to
6919 store it in the array. Punt. This case
6920 should be rare. */
6921 break;
6923 /* INSN sets CC_REG to a value equal to CC_SRC
6924 with the right mode. We can simply delete
6925 it. */
6926 delete_insn (insn);
6929 /* We found an instruction to delete. Keep looking,
6930 in the hopes of finding a three-way jump. */
6931 continue;
6934 /* We found an instruction which sets the condition
6935 code, so don't look any farther. */
6936 break;
6939 /* If INSN sets CC_REG in some other way, don't look any
6940 farther. */
6941 if (reg_set_p (cc_reg, insn))
6942 break;
6945 /* If we fell off the bottom of the block, we can keep looking
6946 through successors. We pass CAN_CHANGE_MODE as false because
6947 we aren't prepared to handle compatibility between the
6948 further blocks and this block. */
6949 if (insn == end)
6951 enum machine_mode submode;
6953 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
6954 if (submode != VOIDmode)
6956 gcc_assert (submode == mode);
6957 found_equiv = true;
6958 can_change_mode = false;
6963 if (! found_equiv)
6964 return VOIDmode;
6966 /* Now INSN_COUNT is the number of instructions we found which set
6967 CC_REG to a value equivalent to CC_SRC. The instructions are in
6968 INSNS. The modes used by those instructions are in MODES. */
6970 newreg = NULL_RTX;
6971 for (i = 0; i < insn_count; ++i)
6973 if (modes[i] != mode)
6975 /* We need to change the mode of CC_REG in INSNS[i] and
6976 subsequent instructions. */
6977 if (! newreg)
6979 if (GET_MODE (cc_reg) == mode)
6980 newreg = cc_reg;
6981 else
6982 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
6984 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
6985 newreg);
6988 delete_insn_and_edges (insns[i]);
6991 return mode;
6994 /* If we have a fixed condition code register (or two), walk through
6995 the instructions and try to eliminate duplicate assignments. */
6997 static void
6998 cse_condition_code_reg (void)
7000 unsigned int cc_regno_1;
7001 unsigned int cc_regno_2;
7002 rtx cc_reg_1;
7003 rtx cc_reg_2;
7004 basic_block bb;
7006 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7007 return;
7009 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7010 if (cc_regno_2 != INVALID_REGNUM)
7011 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7012 else
7013 cc_reg_2 = NULL_RTX;
7015 FOR_EACH_BB (bb)
7017 rtx last_insn;
7018 rtx cc_reg;
7019 rtx insn;
7020 rtx cc_src_insn;
7021 rtx cc_src;
7022 enum machine_mode mode;
7023 enum machine_mode orig_mode;
7025 /* Look for blocks which end with a conditional jump based on a
7026 condition code register. Then look for the instruction which
7027 sets the condition code register. Then look through the
7028 successor blocks for instructions which set the condition
7029 code register to the same value. There are other possible
7030 uses of the condition code register, but these are by far the
7031 most common and the ones which we are most likely to be able
7032 to optimize. */
7034 last_insn = BB_END (bb);
7035 if (!JUMP_P (last_insn))
7036 continue;
7038 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7039 cc_reg = cc_reg_1;
7040 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7041 cc_reg = cc_reg_2;
7042 else
7043 continue;
7045 cc_src_insn = NULL_RTX;
7046 cc_src = NULL_RTX;
7047 for (insn = PREV_INSN (last_insn);
7048 insn && insn != PREV_INSN (BB_HEAD (bb));
7049 insn = PREV_INSN (insn))
7051 rtx set;
7053 if (! INSN_P (insn))
7054 continue;
7055 set = single_set (insn);
7056 if (set
7057 && REG_P (SET_DEST (set))
7058 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7060 cc_src_insn = insn;
7061 cc_src = SET_SRC (set);
7062 break;
7064 else if (reg_set_p (cc_reg, insn))
7065 break;
7068 if (! cc_src_insn)
7069 continue;
7071 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7072 continue;
7074 /* Now CC_REG is a condition code register used for a
7075 conditional jump at the end of the block, and CC_SRC, in
7076 CC_SRC_INSN, is the value to which that condition code
7077 register is set, and CC_SRC is still meaningful at the end of
7078 the basic block. */
7080 orig_mode = GET_MODE (cc_src);
7081 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7082 if (mode != VOIDmode)
7084 gcc_assert (mode == GET_MODE (cc_src));
7085 if (mode != orig_mode)
7087 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7089 cse_change_cc_mode_insn (cc_src_insn, newreg);
7091 /* Do the same in the following insns that use the
7092 current value of CC_REG within BB. */
7093 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7094 NEXT_INSN (last_insn),
7095 newreg);
7102 /* Perform common subexpression elimination. Nonzero value from
7103 `cse_main' means that jumps were simplified and some code may now
7104 be unreachable, so do jump optimization again. */
7105 static bool
7106 gate_handle_cse (void)
7108 return optimize > 0;
7111 static unsigned int
7112 rest_of_handle_cse (void)
7114 int tem;
7116 if (dump_file)
7117 dump_flow_info (dump_file, dump_flags);
7119 tem = cse_main (get_insns (), max_reg_num ());
7121 /* If we are not running more CSE passes, then we are no longer
7122 expecting CSE to be run. But always rerun it in a cheap mode. */
7123 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7125 if (tem == 2)
7127 timevar_push (TV_JUMP);
7128 rebuild_jump_labels (get_insns ());
7129 cleanup_cfg (0);
7130 timevar_pop (TV_JUMP);
7132 else if (tem == 1 || optimize > 1)
7133 cleanup_cfg (0);
7135 return 0;
7138 struct rtl_opt_pass pass_cse =
7141 RTL_PASS,
7142 "cse1", /* name */
7143 gate_handle_cse, /* gate */
7144 rest_of_handle_cse, /* execute */
7145 NULL, /* sub */
7146 NULL, /* next */
7147 0, /* static_pass_number */
7148 TV_CSE, /* tv_id */
7149 0, /* properties_required */
7150 0, /* properties_provided */
7151 0, /* properties_destroyed */
7152 0, /* todo_flags_start */
7153 TODO_df_finish | TODO_verify_rtl_sharing |
7154 TODO_dump_func |
7155 TODO_ggc_collect |
7156 TODO_verify_flow, /* todo_flags_finish */
7161 static bool
7162 gate_handle_cse2 (void)
7164 return optimize > 0 && flag_rerun_cse_after_loop;
7167 /* Run second CSE pass after loop optimizations. */
7168 static unsigned int
7169 rest_of_handle_cse2 (void)
7171 int tem;
7173 if (dump_file)
7174 dump_flow_info (dump_file, dump_flags);
7176 tem = cse_main (get_insns (), max_reg_num ());
7178 /* Run a pass to eliminate duplicated assignments to condition code
7179 registers. We have to run this after bypass_jumps, because it
7180 makes it harder for that pass to determine whether a jump can be
7181 bypassed safely. */
7182 cse_condition_code_reg ();
7184 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7186 if (tem == 2)
7188 timevar_push (TV_JUMP);
7189 rebuild_jump_labels (get_insns ());
7190 cleanup_cfg (0);
7191 timevar_pop (TV_JUMP);
7193 else if (tem == 1)
7194 cleanup_cfg (0);
7196 cse_not_expected = 1;
7197 return 0;
7201 struct rtl_opt_pass pass_cse2 =
7204 RTL_PASS,
7205 "cse2", /* name */
7206 gate_handle_cse2, /* gate */
7207 rest_of_handle_cse2, /* execute */
7208 NULL, /* sub */
7209 NULL, /* next */
7210 0, /* static_pass_number */
7211 TV_CSE2, /* tv_id */
7212 0, /* properties_required */
7213 0, /* properties_provided */
7214 0, /* properties_destroyed */
7215 0, /* todo_flags_start */
7216 TODO_df_finish | TODO_verify_rtl_sharing |
7217 TODO_dump_func |
7218 TODO_ggc_collect |
7219 TODO_verify_flow /* todo_flags_finish */
7223 static bool
7224 gate_handle_cse_after_global_opts (void)
7226 return optimize > 0 && flag_rerun_cse_after_global_opts;
7229 /* Run second CSE pass after loop optimizations. */
7230 static unsigned int
7231 rest_of_handle_cse_after_global_opts (void)
7233 int save_cfj;
7234 int tem;
7236 /* We only want to do local CSE, so don't follow jumps. */
7237 save_cfj = flag_cse_follow_jumps;
7238 flag_cse_follow_jumps = 0;
7240 rebuild_jump_labels (get_insns ());
7241 tem = cse_main (get_insns (), max_reg_num ());
7242 purge_all_dead_edges ();
7243 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7245 cse_not_expected = !flag_rerun_cse_after_loop;
7247 /* If cse altered any jumps, rerun jump opts to clean things up. */
7248 if (tem == 2)
7250 timevar_push (TV_JUMP);
7251 rebuild_jump_labels (get_insns ());
7252 cleanup_cfg (0);
7253 timevar_pop (TV_JUMP);
7255 else if (tem == 1)
7256 cleanup_cfg (0);
7258 flag_cse_follow_jumps = save_cfj;
7259 return 0;
7262 struct rtl_opt_pass pass_cse_after_global_opts =
7265 RTL_PASS,
7266 "cse_local", /* name */
7267 gate_handle_cse_after_global_opts, /* gate */
7268 rest_of_handle_cse_after_global_opts, /* execute */
7269 NULL, /* sub */
7270 NULL, /* next */
7271 0, /* static_pass_number */
7272 TV_CSE, /* tv_id */
7273 0, /* properties_required */
7274 0, /* properties_provided */
7275 0, /* properties_destroyed */
7276 0, /* todo_flags_start */
7277 TODO_df_finish | TODO_verify_rtl_sharing |
7278 TODO_dump_func |
7279 TODO_ggc_collect |
7280 TODO_verify_flow /* todo_flags_finish */