aix.h (FP_SAVE_INLINE, [...]): Delete.
[official-gcc.git] / gcc / cse.c
blob9f4e97932eda5511565b19e2d3f06476ac2a7e17
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010,
4 2011 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "regs.h"
30 #include "basic-block.h"
31 #include "flags.h"
32 #include "insn-config.h"
33 #include "recog.h"
34 #include "function.h"
35 #include "expr.h"
36 #include "diagnostic-core.h"
37 #include "toplev.h"
38 #include "output.h"
39 #include "ggc.h"
40 #include "timevar.h"
41 #include "except.h"
42 #include "target.h"
43 #include "params.h"
44 #include "rtlhooks-def.h"
45 #include "tree-pass.h"
46 #include "df.h"
47 #include "dbgcnt.h"
49 /* The basic idea of common subexpression elimination is to go
50 through the code, keeping a record of expressions that would
51 have the same value at the current scan point, and replacing
52 expressions encountered with the cheapest equivalent expression.
54 It is too complicated to keep track of the different possibilities
55 when control paths merge in this code; so, at each label, we forget all
56 that is known and start fresh. This can be described as processing each
57 extended basic block separately. We have a separate pass to perform
58 global CSE.
60 Note CSE can turn a conditional or computed jump into a nop or
61 an unconditional jump. When this occurs we arrange to run the jump
62 optimizer after CSE to delete the unreachable code.
64 We use two data structures to record the equivalent expressions:
65 a hash table for most expressions, and a vector of "quantity
66 numbers" to record equivalent (pseudo) registers.
68 The use of the special data structure for registers is desirable
69 because it is faster. It is possible because registers references
70 contain a fairly small number, the register number, taken from
71 a contiguously allocated series, and two register references are
72 identical if they have the same number. General expressions
73 do not have any such thing, so the only way to retrieve the
74 information recorded on an expression other than a register
75 is to keep it in a hash table.
77 Registers and "quantity numbers":
79 At the start of each basic block, all of the (hardware and pseudo)
80 registers used in the function are given distinct quantity
81 numbers to indicate their contents. During scan, when the code
82 copies one register into another, we copy the quantity number.
83 When a register is loaded in any other way, we allocate a new
84 quantity number to describe the value generated by this operation.
85 `REG_QTY (N)' records what quantity register N is currently thought
86 of as containing.
88 All real quantity numbers are greater than or equal to zero.
89 If register N has not been assigned a quantity, `REG_QTY (N)' will
90 equal -N - 1, which is always negative.
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
95 We also maintain a bidirectional chain of registers for each
96 quantity number. The `qty_table` members `first_reg' and `last_reg',
97 and `reg_eqv_table' members `next' and `prev' hold these chains.
99 The first register in a chain is the one whose lifespan is least local.
100 Among equals, it is the one that was seen first.
101 We replace any equivalent register with that one.
103 If two registers have the same quantity number, it must be true that
104 REG expressions with qty_table `mode' must be in the hash table for both
105 registers and must be in the same class.
107 The converse is not true. Since hard registers may be referenced in
108 any mode, two REG expressions might be equivalent in the hash table
109 but not have the same quantity number if the quantity number of one
110 of the registers is not the same mode as those expressions.
112 Constants and quantity numbers
114 When a quantity has a known constant value, that value is stored
115 in the appropriate qty_table `const_rtx'. This is in addition to
116 putting the constant in the hash table as is usual for non-regs.
118 Whether a reg or a constant is preferred is determined by the configuration
119 macro CONST_COSTS and will often depend on the constant value. In any
120 event, expressions containing constants can be simplified, by fold_rtx.
122 When a quantity has a known nearly constant value (such as an address
123 of a stack slot), that value is stored in the appropriate qty_table
124 `const_rtx'.
126 Integer constants don't have a machine mode. However, cse
127 determines the intended machine mode from the destination
128 of the instruction that moves the constant. The machine mode
129 is recorded in the hash table along with the actual RTL
130 constant expression so that different modes are kept separate.
132 Other expressions:
134 To record known equivalences among expressions in general
135 we use a hash table called `table'. It has a fixed number of buckets
136 that contain chains of `struct table_elt' elements for expressions.
137 These chains connect the elements whose expressions have the same
138 hash codes.
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
143 Register references in an expression are canonicalized before hashing
144 the expression. This is done using `reg_qty' and qty_table `first_reg'.
145 The hash code of a register reference is computed using the quantity
146 number, not the register number.
148 When the value of an expression changes, it is necessary to remove from the
149 hash table not just that expression but all expressions whose values
150 could be different as a result.
152 1. If the value changing is in memory, except in special cases
153 ANYTHING referring to memory could be changed. That is because
154 nobody knows where a pointer does not point.
155 The function `invalidate_memory' removes what is necessary.
157 The special cases are when the address is constant or is
158 a constant plus a fixed register such as the frame pointer
159 or a static chain pointer. When such addresses are stored in,
160 we can tell exactly which other such addresses must be invalidated
161 due to overlap. `invalidate' does this.
162 All expressions that refer to non-constant
163 memory addresses are also invalidated. `invalidate_memory' does this.
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
169 Because searching the entire hash table for expressions that contain
170 a register is very slow, we try to figure out when it isn't necessary.
171 Precisely, this is necessary only when expressions have been
172 entered in the hash table using this register, and then the value has
173 changed, and then another expression wants to be added to refer to
174 the register's new value. This sequence of circumstances is rare
175 within any one basic block.
177 `REG_TICK' and `REG_IN_TABLE', accessors for members of
178 cse_reg_info, are used to detect this case. REG_TICK (i) is
179 incremented whenever a value is stored in register i.
180 REG_IN_TABLE (i) holds -1 if no references to register i have been
181 entered in the table; otherwise, it contains the value REG_TICK (i)
182 had when the references were entered. If we want to enter a
183 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
184 remove old references. Until we want to enter a new entry, the
185 mere fact that the two vectors don't match makes the entries be
186 ignored if anyone tries to match them.
188 Registers themselves are entered in the hash table as well as in
189 the equivalent-register chains. However, `REG_TICK' and
190 `REG_IN_TABLE' do not apply to expressions which are simple
191 register references. These expressions are removed from the table
192 immediately when they become invalid, and this can be done even if
193 we do not immediately search for all the expressions that refer to
194 the register.
196 A CLOBBER rtx in an instruction invalidates its operand for further
197 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
198 invalidates everything that resides in memory.
200 Related expressions:
202 Constant expressions that differ only by an additive integer
203 are called related. When a constant expression is put in
204 the table, the related expression with no constant term
205 is also entered. These are made to point at each other
206 so that it is possible to find out if there exists any
207 register equivalent to an expression related to a given expression. */
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
212 static int max_qty;
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
217 static int next_qty;
219 /* Per-qty information tracking.
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
224 `mode' contains the machine mode of this quantity.
226 `const_rtx' holds the rtx of the constant value of this
227 quantity, if known. A summations of the frame/arg pointer
228 and a constant can also be entered here. When this holds
229 a known value, `const_insn' is the insn which stored the
230 constant value.
232 `comparison_{code,const,qty}' are used to track when a
233 comparison between a quantity and some constant or register has
234 been passed. In such a case, we know the results of the comparison
235 in case we see it again. These members record a comparison that
236 is known to be true. `comparison_code' holds the rtx code of such
237 a comparison, else it is set to UNKNOWN and the other two
238 comparison members are undefined. `comparison_const' holds
239 the constant being compared against, or zero if the comparison
240 is not against a constant. `comparison_qty' holds the quantity
241 being compared against when the result is known. If the comparison
242 is not with a register, `comparison_qty' is -1. */
244 struct qty_table_elem
246 rtx const_rtx;
247 rtx const_insn;
248 rtx comparison_const;
249 int comparison_qty;
250 unsigned int first_reg, last_reg;
251 /* The sizes of these fields should match the sizes of the
252 code and mode fields of struct rtx_def (see rtl.h). */
253 ENUM_BITFIELD(rtx_code) comparison_code : 16;
254 ENUM_BITFIELD(machine_mode) mode : 8;
257 /* The table of all qtys, indexed by qty number. */
258 static struct qty_table_elem *qty_table;
260 /* Structure used to pass arguments via for_each_rtx to function
261 cse_change_cc_mode. */
262 struct change_cc_mode_args
264 rtx insn;
265 rtx newreg;
268 #ifdef HAVE_cc0
269 /* For machines that have a CC0, we do not record its value in the hash
270 table since its use is guaranteed to be the insn immediately following
271 its definition and any other insn is presumed to invalidate it.
273 Instead, we store below the current and last value assigned to CC0.
274 If it should happen to be a constant, it is stored in preference
275 to the actual assigned value. In case it is a constant, we store
276 the mode in which the constant should be interpreted. */
278 static rtx this_insn_cc0, prev_insn_cc0;
279 static enum machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
280 #endif
282 /* Insn being scanned. */
284 static rtx this_insn;
285 static bool optimize_this_for_speed_p;
287 /* Index by register number, gives the number of the next (or
288 previous) register in the chain of registers sharing the same
289 value.
291 Or -1 if this register is at the end of the chain.
293 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
295 /* Per-register equivalence chain. */
296 struct reg_eqv_elem
298 int next, prev;
301 /* The table of all register equivalence chains. */
302 static struct reg_eqv_elem *reg_eqv_table;
304 struct cse_reg_info
306 /* The timestamp at which this register is initialized. */
307 unsigned int timestamp;
309 /* The quantity number of the register's current contents. */
310 int reg_qty;
312 /* The number of times the register has been altered in the current
313 basic block. */
314 int reg_tick;
316 /* The REG_TICK value at which rtx's containing this register are
317 valid in the hash table. If this does not equal the current
318 reg_tick value, such expressions existing in the hash table are
319 invalid. */
320 int reg_in_table;
322 /* The SUBREG that was set when REG_TICK was last incremented. Set
323 to -1 if the last store was to the whole register, not a subreg. */
324 unsigned int subreg_ticked;
327 /* A table of cse_reg_info indexed by register numbers. */
328 static struct cse_reg_info *cse_reg_info_table;
330 /* The size of the above table. */
331 static unsigned int cse_reg_info_table_size;
333 /* The index of the first entry that has not been initialized. */
334 static unsigned int cse_reg_info_table_first_uninitialized;
336 /* The timestamp at the beginning of the current run of
337 cse_extended_basic_block. We increment this variable at the beginning of
338 the current run of cse_extended_basic_block. The timestamp field of a
339 cse_reg_info entry matches the value of this variable if and only
340 if the entry has been initialized during the current run of
341 cse_extended_basic_block. */
342 static unsigned int cse_reg_info_timestamp;
344 /* A HARD_REG_SET containing all the hard registers for which there is
345 currently a REG expression in the hash table. Note the difference
346 from the above variables, which indicate if the REG is mentioned in some
347 expression in the table. */
349 static HARD_REG_SET hard_regs_in_table;
351 /* True if CSE has altered the CFG. */
352 static bool cse_cfg_altered;
354 /* True if CSE has altered conditional jump insns in such a way
355 that jump optimization should be redone. */
356 static bool cse_jumps_altered;
358 /* True if we put a LABEL_REF into the hash table for an INSN
359 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
360 to put in the note. */
361 static bool recorded_label_ref;
363 /* canon_hash stores 1 in do_not_record
364 if it notices a reference to CC0, PC, or some other volatile
365 subexpression. */
367 static int do_not_record;
369 /* canon_hash stores 1 in hash_arg_in_memory
370 if it notices a reference to memory within the expression being hashed. */
372 static int hash_arg_in_memory;
374 /* The hash table contains buckets which are chains of `struct table_elt's,
375 each recording one expression's information.
376 That expression is in the `exp' field.
378 The canon_exp field contains a canonical (from the point of view of
379 alias analysis) version of the `exp' field.
381 Those elements with the same hash code are chained in both directions
382 through the `next_same_hash' and `prev_same_hash' fields.
384 Each set of expressions with equivalent values
385 are on a two-way chain through the `next_same_value'
386 and `prev_same_value' fields, and all point with
387 the `first_same_value' field at the first element in
388 that chain. The chain is in order of increasing cost.
389 Each element's cost value is in its `cost' field.
391 The `in_memory' field is nonzero for elements that
392 involve any reference to memory. These elements are removed
393 whenever a write is done to an unidentified location in memory.
394 To be safe, we assume that a memory address is unidentified unless
395 the address is either a symbol constant or a constant plus
396 the frame pointer or argument pointer.
398 The `related_value' field is used to connect related expressions
399 (that differ by adding an integer).
400 The related expressions are chained in a circular fashion.
401 `related_value' is zero for expressions for which this
402 chain is not useful.
404 The `cost' field stores the cost of this element's expression.
405 The `regcost' field stores the value returned by approx_reg_cost for
406 this element's expression.
408 The `is_const' flag is set if the element is a constant (including
409 a fixed address).
411 The `flag' field is used as a temporary during some search routines.
413 The `mode' field is usually the same as GET_MODE (`exp'), but
414 if `exp' is a CONST_INT and has no machine mode then the `mode'
415 field is the mode it was being used as. Each constant is
416 recorded separately for each mode it is used with. */
418 struct table_elt
420 rtx exp;
421 rtx canon_exp;
422 struct table_elt *next_same_hash;
423 struct table_elt *prev_same_hash;
424 struct table_elt *next_same_value;
425 struct table_elt *prev_same_value;
426 struct table_elt *first_same_value;
427 struct table_elt *related_value;
428 int cost;
429 int regcost;
430 /* The size of this field should match the size
431 of the mode field of struct rtx_def (see rtl.h). */
432 ENUM_BITFIELD(machine_mode) mode : 8;
433 char in_memory;
434 char is_const;
435 char flag;
438 /* We don't want a lot of buckets, because we rarely have very many
439 things stored in the hash table, and a lot of buckets slows
440 down a lot of loops that happen frequently. */
441 #define HASH_SHIFT 5
442 #define HASH_SIZE (1 << HASH_SHIFT)
443 #define HASH_MASK (HASH_SIZE - 1)
445 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
446 register (hard registers may require `do_not_record' to be set). */
448 #define HASH(X, M) \
449 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
450 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
451 : canon_hash (X, M)) & HASH_MASK)
453 /* Like HASH, but without side-effects. */
454 #define SAFE_HASH(X, M) \
455 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
456 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
457 : safe_hash (X, M)) & HASH_MASK)
459 /* Determine whether register number N is considered a fixed register for the
460 purpose of approximating register costs.
461 It is desirable to replace other regs with fixed regs, to reduce need for
462 non-fixed hard regs.
463 A reg wins if it is either the frame pointer or designated as fixed. */
464 #define FIXED_REGNO_P(N) \
465 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
466 || fixed_regs[N] || global_regs[N])
468 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
469 hard registers and pointers into the frame are the cheapest with a cost
470 of 0. Next come pseudos with a cost of one and other hard registers with
471 a cost of 2. Aside from these special cases, call `rtx_cost'. */
473 #define CHEAP_REGNO(N) \
474 (REGNO_PTR_FRAME_P(N) \
475 || (HARD_REGISTER_NUM_P (N) \
476 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
478 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET, 1))
479 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
481 /* Get the number of times this register has been updated in this
482 basic block. */
484 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
486 /* Get the point at which REG was recorded in the table. */
488 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
490 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
491 SUBREG). */
493 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
495 /* Get the quantity number for REG. */
497 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
499 /* Determine if the quantity number for register X represents a valid index
500 into the qty_table. */
502 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
504 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
506 #define CHEAPER(X, Y) \
507 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
509 static struct table_elt *table[HASH_SIZE];
511 /* Chain of `struct table_elt's made so far for this function
512 but currently removed from the table. */
514 static struct table_elt *free_element_chain;
516 /* Set to the cost of a constant pool reference if one was found for a
517 symbolic constant. If this was found, it means we should try to
518 convert constants into constant pool entries if they don't fit in
519 the insn. */
521 static int constant_pool_entries_cost;
522 static int constant_pool_entries_regcost;
524 /* Trace a patch through the CFG. */
526 struct branch_path
528 /* The basic block for this path entry. */
529 basic_block bb;
532 /* This data describes a block that will be processed by
533 cse_extended_basic_block. */
535 struct cse_basic_block_data
537 /* Total number of SETs in block. */
538 int nsets;
539 /* Size of current branch path, if any. */
540 int path_size;
541 /* Current path, indicating which basic_blocks will be processed. */
542 struct branch_path *path;
546 /* Pointers to the live in/live out bitmaps for the boundaries of the
547 current EBB. */
548 static bitmap cse_ebb_live_in, cse_ebb_live_out;
550 /* A simple bitmap to track which basic blocks have been visited
551 already as part of an already processed extended basic block. */
552 static sbitmap cse_visited_basic_blocks;
554 static bool fixed_base_plus_p (rtx x);
555 static int notreg_cost (rtx, enum rtx_code, int);
556 static int approx_reg_cost_1 (rtx *, void *);
557 static int approx_reg_cost (rtx);
558 static int preferable (int, int, int, int);
559 static void new_basic_block (void);
560 static void make_new_qty (unsigned int, enum machine_mode);
561 static void make_regs_eqv (unsigned int, unsigned int);
562 static void delete_reg_equiv (unsigned int);
563 static int mention_regs (rtx);
564 static int insert_regs (rtx, struct table_elt *, int);
565 static void remove_from_table (struct table_elt *, unsigned);
566 static void remove_pseudo_from_table (rtx, unsigned);
567 static struct table_elt *lookup (rtx, unsigned, enum machine_mode);
568 static struct table_elt *lookup_for_remove (rtx, unsigned, enum machine_mode);
569 static rtx lookup_as_function (rtx, enum rtx_code);
570 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
571 enum machine_mode, int, int);
572 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
573 enum machine_mode);
574 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
575 static void invalidate (rtx, enum machine_mode);
576 static void remove_invalid_refs (unsigned int);
577 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
578 enum machine_mode);
579 static void rehash_using_reg (rtx);
580 static void invalidate_memory (void);
581 static void invalidate_for_call (void);
582 static rtx use_related_value (rtx, struct table_elt *);
584 static inline unsigned canon_hash (rtx, enum machine_mode);
585 static inline unsigned safe_hash (rtx, enum machine_mode);
586 static inline unsigned hash_rtx_string (const char *);
588 static rtx canon_reg (rtx, rtx);
589 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
590 enum machine_mode *,
591 enum machine_mode *);
592 static rtx fold_rtx (rtx, rtx);
593 static rtx equiv_constant (rtx);
594 static void record_jump_equiv (rtx, bool);
595 static void record_jump_cond (enum rtx_code, enum machine_mode, rtx, rtx,
596 int);
597 static void cse_insn (rtx);
598 static void cse_prescan_path (struct cse_basic_block_data *);
599 static void invalidate_from_clobbers (rtx);
600 static void invalidate_from_sets_and_clobbers (rtx);
601 static rtx cse_process_notes (rtx, rtx, bool *);
602 static void cse_extended_basic_block (struct cse_basic_block_data *);
603 static void count_reg_usage (rtx, int *, rtx, int);
604 static int check_for_label_ref (rtx *, void *);
605 extern void dump_class (struct table_elt*);
606 static void get_cse_reg_info_1 (unsigned int regno);
607 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
608 static int check_dependence (rtx *, void *);
610 static void flush_hash_table (void);
611 static bool insn_live_p (rtx, int *);
612 static bool set_live_p (rtx, rtx, int *);
613 static int cse_change_cc_mode (rtx *, void *);
614 static void cse_change_cc_mode_insn (rtx, rtx);
615 static void cse_change_cc_mode_insns (rtx, rtx, rtx);
616 static enum machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
617 bool);
620 #undef RTL_HOOKS_GEN_LOWPART
621 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
623 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
625 /* Nonzero if X has the form (PLUS frame-pointer integer). We check for
626 virtual regs here because the simplify_*_operation routines are called
627 by integrate.c, which is called before virtual register instantiation. */
629 static bool
630 fixed_base_plus_p (rtx x)
632 switch (GET_CODE (x))
634 case REG:
635 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
636 return true;
637 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
638 return true;
639 if (REGNO (x) >= FIRST_VIRTUAL_REGISTER
640 && REGNO (x) <= LAST_VIRTUAL_REGISTER)
641 return true;
642 return false;
644 case PLUS:
645 if (!CONST_INT_P (XEXP (x, 1)))
646 return false;
647 return fixed_base_plus_p (XEXP (x, 0));
649 default:
650 return false;
654 /* Dump the expressions in the equivalence class indicated by CLASSP.
655 This function is used only for debugging. */
656 DEBUG_FUNCTION void
657 dump_class (struct table_elt *classp)
659 struct table_elt *elt;
661 fprintf (stderr, "Equivalence chain for ");
662 print_rtl (stderr, classp->exp);
663 fprintf (stderr, ": \n");
665 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
667 print_rtl (stderr, elt->exp);
668 fprintf (stderr, "\n");
672 /* Subroutine of approx_reg_cost; called through for_each_rtx. */
674 static int
675 approx_reg_cost_1 (rtx *xp, void *data)
677 rtx x = *xp;
678 int *cost_p = (int *) data;
680 if (x && REG_P (x))
682 unsigned int regno = REGNO (x);
684 if (! CHEAP_REGNO (regno))
686 if (regno < FIRST_PSEUDO_REGISTER)
688 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
689 return 1;
690 *cost_p += 2;
692 else
693 *cost_p += 1;
697 return 0;
700 /* Return an estimate of the cost of the registers used in an rtx.
701 This is mostly the number of different REG expressions in the rtx;
702 however for some exceptions like fixed registers we use a cost of
703 0. If any other hard register reference occurs, return MAX_COST. */
705 static int
706 approx_reg_cost (rtx x)
708 int cost = 0;
710 if (for_each_rtx (&x, approx_reg_cost_1, (void *) &cost))
711 return MAX_COST;
713 return cost;
716 /* Return a negative value if an rtx A, whose costs are given by COST_A
717 and REGCOST_A, is more desirable than an rtx B.
718 Return a positive value if A is less desirable, or 0 if the two are
719 equally good. */
720 static int
721 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
723 /* First, get rid of cases involving expressions that are entirely
724 unwanted. */
725 if (cost_a != cost_b)
727 if (cost_a == MAX_COST)
728 return 1;
729 if (cost_b == MAX_COST)
730 return -1;
733 /* Avoid extending lifetimes of hardregs. */
734 if (regcost_a != regcost_b)
736 if (regcost_a == MAX_COST)
737 return 1;
738 if (regcost_b == MAX_COST)
739 return -1;
742 /* Normal operation costs take precedence. */
743 if (cost_a != cost_b)
744 return cost_a - cost_b;
745 /* Only if these are identical consider effects on register pressure. */
746 if (regcost_a != regcost_b)
747 return regcost_a - regcost_b;
748 return 0;
751 /* Internal function, to compute cost when X is not a register; called
752 from COST macro to keep it simple. */
754 static int
755 notreg_cost (rtx x, enum rtx_code outer, int opno)
757 return ((GET_CODE (x) == SUBREG
758 && REG_P (SUBREG_REG (x))
759 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
760 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
761 && (GET_MODE_SIZE (GET_MODE (x))
762 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
763 && subreg_lowpart_p (x)
764 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
765 GET_MODE (SUBREG_REG (x))))
767 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
771 /* Initialize CSE_REG_INFO_TABLE. */
773 static void
774 init_cse_reg_info (unsigned int nregs)
776 /* Do we need to grow the table? */
777 if (nregs > cse_reg_info_table_size)
779 unsigned int new_size;
781 if (cse_reg_info_table_size < 2048)
783 /* Compute a new size that is a power of 2 and no smaller
784 than the large of NREGS and 64. */
785 new_size = (cse_reg_info_table_size
786 ? cse_reg_info_table_size : 64);
788 while (new_size < nregs)
789 new_size *= 2;
791 else
793 /* If we need a big table, allocate just enough to hold
794 NREGS registers. */
795 new_size = nregs;
798 /* Reallocate the table with NEW_SIZE entries. */
799 free (cse_reg_info_table);
800 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
801 cse_reg_info_table_size = new_size;
802 cse_reg_info_table_first_uninitialized = 0;
805 /* Do we have all of the first NREGS entries initialized? */
806 if (cse_reg_info_table_first_uninitialized < nregs)
808 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
809 unsigned int i;
811 /* Put the old timestamp on newly allocated entries so that they
812 will all be considered out of date. We do not touch those
813 entries beyond the first NREGS entries to be nice to the
814 virtual memory. */
815 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
816 cse_reg_info_table[i].timestamp = old_timestamp;
818 cse_reg_info_table_first_uninitialized = nregs;
822 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
824 static void
825 get_cse_reg_info_1 (unsigned int regno)
827 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
828 entry will be considered to have been initialized. */
829 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
831 /* Initialize the rest of the entry. */
832 cse_reg_info_table[regno].reg_tick = 1;
833 cse_reg_info_table[regno].reg_in_table = -1;
834 cse_reg_info_table[regno].subreg_ticked = -1;
835 cse_reg_info_table[regno].reg_qty = -regno - 1;
838 /* Find a cse_reg_info entry for REGNO. */
840 static inline struct cse_reg_info *
841 get_cse_reg_info (unsigned int regno)
843 struct cse_reg_info *p = &cse_reg_info_table[regno];
845 /* If this entry has not been initialized, go ahead and initialize
846 it. */
847 if (p->timestamp != cse_reg_info_timestamp)
848 get_cse_reg_info_1 (regno);
850 return p;
853 /* Clear the hash table and initialize each register with its own quantity,
854 for a new basic block. */
856 static void
857 new_basic_block (void)
859 int i;
861 next_qty = 0;
863 /* Invalidate cse_reg_info_table. */
864 cse_reg_info_timestamp++;
866 /* Clear out hash table state for this pass. */
867 CLEAR_HARD_REG_SET (hard_regs_in_table);
869 /* The per-quantity values used to be initialized here, but it is
870 much faster to initialize each as it is made in `make_new_qty'. */
872 for (i = 0; i < HASH_SIZE; i++)
874 struct table_elt *first;
876 first = table[i];
877 if (first != NULL)
879 struct table_elt *last = first;
881 table[i] = NULL;
883 while (last->next_same_hash != NULL)
884 last = last->next_same_hash;
886 /* Now relink this hash entire chain into
887 the free element list. */
889 last->next_same_hash = free_element_chain;
890 free_element_chain = first;
894 #ifdef HAVE_cc0
895 prev_insn_cc0 = 0;
896 #endif
899 /* Say that register REG contains a quantity in mode MODE not in any
900 register before and initialize that quantity. */
902 static void
903 make_new_qty (unsigned int reg, enum machine_mode mode)
905 int q;
906 struct qty_table_elem *ent;
907 struct reg_eqv_elem *eqv;
909 gcc_assert (next_qty < max_qty);
911 q = REG_QTY (reg) = next_qty++;
912 ent = &qty_table[q];
913 ent->first_reg = reg;
914 ent->last_reg = reg;
915 ent->mode = mode;
916 ent->const_rtx = ent->const_insn = NULL_RTX;
917 ent->comparison_code = UNKNOWN;
919 eqv = &reg_eqv_table[reg];
920 eqv->next = eqv->prev = -1;
923 /* Make reg NEW equivalent to reg OLD.
924 OLD is not changing; NEW is. */
926 static void
927 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
929 unsigned int lastr, firstr;
930 int q = REG_QTY (old_reg);
931 struct qty_table_elem *ent;
933 ent = &qty_table[q];
935 /* Nothing should become eqv until it has a "non-invalid" qty number. */
936 gcc_assert (REGNO_QTY_VALID_P (old_reg));
938 REG_QTY (new_reg) = q;
939 firstr = ent->first_reg;
940 lastr = ent->last_reg;
942 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
943 hard regs. Among pseudos, if NEW will live longer than any other reg
944 of the same qty, and that is beyond the current basic block,
945 make it the new canonical replacement for this qty. */
946 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
947 /* Certain fixed registers might be of the class NO_REGS. This means
948 that not only can they not be allocated by the compiler, but
949 they cannot be used in substitutions or canonicalizations
950 either. */
951 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
952 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
953 || (new_reg >= FIRST_PSEUDO_REGISTER
954 && (firstr < FIRST_PSEUDO_REGISTER
955 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
956 && !bitmap_bit_p (cse_ebb_live_out, firstr))
957 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
958 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
960 reg_eqv_table[firstr].prev = new_reg;
961 reg_eqv_table[new_reg].next = firstr;
962 reg_eqv_table[new_reg].prev = -1;
963 ent->first_reg = new_reg;
965 else
967 /* If NEW is a hard reg (known to be non-fixed), insert at end.
968 Otherwise, insert before any non-fixed hard regs that are at the
969 end. Registers of class NO_REGS cannot be used as an
970 equivalent for anything. */
971 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
972 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
973 && new_reg >= FIRST_PSEUDO_REGISTER)
974 lastr = reg_eqv_table[lastr].prev;
975 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
976 if (reg_eqv_table[lastr].next >= 0)
977 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
978 else
979 qty_table[q].last_reg = new_reg;
980 reg_eqv_table[lastr].next = new_reg;
981 reg_eqv_table[new_reg].prev = lastr;
985 /* Remove REG from its equivalence class. */
987 static void
988 delete_reg_equiv (unsigned int reg)
990 struct qty_table_elem *ent;
991 int q = REG_QTY (reg);
992 int p, n;
994 /* If invalid, do nothing. */
995 if (! REGNO_QTY_VALID_P (reg))
996 return;
998 ent = &qty_table[q];
1000 p = reg_eqv_table[reg].prev;
1001 n = reg_eqv_table[reg].next;
1003 if (n != -1)
1004 reg_eqv_table[n].prev = p;
1005 else
1006 ent->last_reg = p;
1007 if (p != -1)
1008 reg_eqv_table[p].next = n;
1009 else
1010 ent->first_reg = n;
1012 REG_QTY (reg) = -reg - 1;
1015 /* Remove any invalid expressions from the hash table
1016 that refer to any of the registers contained in expression X.
1018 Make sure that newly inserted references to those registers
1019 as subexpressions will be considered valid.
1021 mention_regs is not called when a register itself
1022 is being stored in the table.
1024 Return 1 if we have done something that may have changed the hash code
1025 of X. */
1027 static int
1028 mention_regs (rtx x)
1030 enum rtx_code code;
1031 int i, j;
1032 const char *fmt;
1033 int changed = 0;
1035 if (x == 0)
1036 return 0;
1038 code = GET_CODE (x);
1039 if (code == REG)
1041 unsigned int regno = REGNO (x);
1042 unsigned int endregno = END_REGNO (x);
1043 unsigned int i;
1045 for (i = regno; i < endregno; i++)
1047 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1048 remove_invalid_refs (i);
1050 REG_IN_TABLE (i) = REG_TICK (i);
1051 SUBREG_TICKED (i) = -1;
1054 return 0;
1057 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1058 pseudo if they don't use overlapping words. We handle only pseudos
1059 here for simplicity. */
1060 if (code == SUBREG && REG_P (SUBREG_REG (x))
1061 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1063 unsigned int i = REGNO (SUBREG_REG (x));
1065 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1067 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1068 the last store to this register really stored into this
1069 subreg, then remove the memory of this subreg.
1070 Otherwise, remove any memory of the entire register and
1071 all its subregs from the table. */
1072 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1073 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1074 remove_invalid_refs (i);
1075 else
1076 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1079 REG_IN_TABLE (i) = REG_TICK (i);
1080 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1081 return 0;
1084 /* If X is a comparison or a COMPARE and either operand is a register
1085 that does not have a quantity, give it one. This is so that a later
1086 call to record_jump_equiv won't cause X to be assigned a different
1087 hash code and not found in the table after that call.
1089 It is not necessary to do this here, since rehash_using_reg can
1090 fix up the table later, but doing this here eliminates the need to
1091 call that expensive function in the most common case where the only
1092 use of the register is in the comparison. */
1094 if (code == COMPARE || COMPARISON_P (x))
1096 if (REG_P (XEXP (x, 0))
1097 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1098 if (insert_regs (XEXP (x, 0), NULL, 0))
1100 rehash_using_reg (XEXP (x, 0));
1101 changed = 1;
1104 if (REG_P (XEXP (x, 1))
1105 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1106 if (insert_regs (XEXP (x, 1), NULL, 0))
1108 rehash_using_reg (XEXP (x, 1));
1109 changed = 1;
1113 fmt = GET_RTX_FORMAT (code);
1114 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1115 if (fmt[i] == 'e')
1116 changed |= mention_regs (XEXP (x, i));
1117 else if (fmt[i] == 'E')
1118 for (j = 0; j < XVECLEN (x, i); j++)
1119 changed |= mention_regs (XVECEXP (x, i, j));
1121 return changed;
1124 /* Update the register quantities for inserting X into the hash table
1125 with a value equivalent to CLASSP.
1126 (If the class does not contain a REG, it is irrelevant.)
1127 If MODIFIED is nonzero, X is a destination; it is being modified.
1128 Note that delete_reg_equiv should be called on a register
1129 before insert_regs is done on that register with MODIFIED != 0.
1131 Nonzero value means that elements of reg_qty have changed
1132 so X's hash code may be different. */
1134 static int
1135 insert_regs (rtx x, struct table_elt *classp, int modified)
1137 if (REG_P (x))
1139 unsigned int regno = REGNO (x);
1140 int qty_valid;
1142 /* If REGNO is in the equivalence table already but is of the
1143 wrong mode for that equivalence, don't do anything here. */
1145 qty_valid = REGNO_QTY_VALID_P (regno);
1146 if (qty_valid)
1148 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1150 if (ent->mode != GET_MODE (x))
1151 return 0;
1154 if (modified || ! qty_valid)
1156 if (classp)
1157 for (classp = classp->first_same_value;
1158 classp != 0;
1159 classp = classp->next_same_value)
1160 if (REG_P (classp->exp)
1161 && GET_MODE (classp->exp) == GET_MODE (x))
1163 unsigned c_regno = REGNO (classp->exp);
1165 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1167 /* Suppose that 5 is hard reg and 100 and 101 are
1168 pseudos. Consider
1170 (set (reg:si 100) (reg:si 5))
1171 (set (reg:si 5) (reg:si 100))
1172 (set (reg:di 101) (reg:di 5))
1174 We would now set REG_QTY (101) = REG_QTY (5), but the
1175 entry for 5 is in SImode. When we use this later in
1176 copy propagation, we get the register in wrong mode. */
1177 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1178 continue;
1180 make_regs_eqv (regno, c_regno);
1181 return 1;
1184 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1185 than REG_IN_TABLE to find out if there was only a single preceding
1186 invalidation - for the SUBREG - or another one, which would be
1187 for the full register. However, if we find here that REG_TICK
1188 indicates that the register is invalid, it means that it has
1189 been invalidated in a separate operation. The SUBREG might be used
1190 now (then this is a recursive call), or we might use the full REG
1191 now and a SUBREG of it later. So bump up REG_TICK so that
1192 mention_regs will do the right thing. */
1193 if (! modified
1194 && REG_IN_TABLE (regno) >= 0
1195 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1196 REG_TICK (regno)++;
1197 make_new_qty (regno, GET_MODE (x));
1198 return 1;
1201 return 0;
1204 /* If X is a SUBREG, we will likely be inserting the inner register in the
1205 table. If that register doesn't have an assigned quantity number at
1206 this point but does later, the insertion that we will be doing now will
1207 not be accessible because its hash code will have changed. So assign
1208 a quantity number now. */
1210 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1211 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1213 insert_regs (SUBREG_REG (x), NULL, 0);
1214 mention_regs (x);
1215 return 1;
1217 else
1218 return mention_regs (x);
1222 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1223 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1224 CST is equal to an anchor. */
1226 static bool
1227 compute_const_anchors (rtx cst,
1228 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1229 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1231 HOST_WIDE_INT n = INTVAL (cst);
1233 *lower_base = n & ~(targetm.const_anchor - 1);
1234 if (*lower_base == n)
1235 return false;
1237 *upper_base =
1238 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1239 *upper_offs = n - *upper_base;
1240 *lower_offs = n - *lower_base;
1241 return true;
1244 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1246 static void
1247 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1248 enum machine_mode mode)
1250 struct table_elt *elt;
1251 unsigned hash;
1252 rtx anchor_exp;
1253 rtx exp;
1255 anchor_exp = GEN_INT (anchor);
1256 hash = HASH (anchor_exp, mode);
1257 elt = lookup (anchor_exp, hash, mode);
1258 if (!elt)
1259 elt = insert (anchor_exp, NULL, hash, mode);
1261 exp = plus_constant (mode, reg, offs);
1262 /* REG has just been inserted and the hash codes recomputed. */
1263 mention_regs (exp);
1264 hash = HASH (exp, mode);
1266 /* Use the cost of the register rather than the whole expression. When
1267 looking up constant anchors we will further offset the corresponding
1268 expression therefore it does not make sense to prefer REGs over
1269 reg-immediate additions. Prefer instead the oldest expression. Also
1270 don't prefer pseudos over hard regs so that we derive constants in
1271 argument registers from other argument registers rather than from the
1272 original pseudo that was used to synthesize the constant. */
1273 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1276 /* The constant CST is equivalent to the register REG. Create
1277 equivalences between the two anchors of CST and the corresponding
1278 register-offset expressions using REG. */
1280 static void
1281 insert_const_anchors (rtx reg, rtx cst, enum machine_mode mode)
1283 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1285 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1286 &upper_base, &upper_offs))
1287 return;
1289 /* Ignore anchors of value 0. Constants accessible from zero are
1290 simple. */
1291 if (lower_base != 0)
1292 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1294 if (upper_base != 0)
1295 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1298 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1299 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1300 valid expression. Return the cheapest and oldest of such expressions. In
1301 *OLD, return how old the resulting expression is compared to the other
1302 equivalent expressions. */
1304 static rtx
1305 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1306 unsigned *old)
1308 struct table_elt *elt;
1309 unsigned idx;
1310 struct table_elt *match_elt;
1311 rtx match;
1313 /* Find the cheapest and *oldest* expression to maximize the chance of
1314 reusing the same pseudo. */
1316 match_elt = NULL;
1317 match = NULL_RTX;
1318 for (elt = anchor_elt->first_same_value, idx = 0;
1319 elt;
1320 elt = elt->next_same_value, idx++)
1322 if (match_elt && CHEAPER (match_elt, elt))
1323 return match;
1325 if (REG_P (elt->exp)
1326 || (GET_CODE (elt->exp) == PLUS
1327 && REG_P (XEXP (elt->exp, 0))
1328 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1330 rtx x;
1332 /* Ignore expressions that are no longer valid. */
1333 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1334 continue;
1336 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1337 if (REG_P (x)
1338 || (GET_CODE (x) == PLUS
1339 && IN_RANGE (INTVAL (XEXP (x, 1)),
1340 -targetm.const_anchor,
1341 targetm.const_anchor - 1)))
1343 match = x;
1344 match_elt = elt;
1345 *old = idx;
1350 return match;
1353 /* Try to express the constant SRC_CONST using a register+offset expression
1354 derived from a constant anchor. Return it if successful or NULL_RTX,
1355 otherwise. */
1357 static rtx
1358 try_const_anchors (rtx src_const, enum machine_mode mode)
1360 struct table_elt *lower_elt, *upper_elt;
1361 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1362 rtx lower_anchor_rtx, upper_anchor_rtx;
1363 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1364 unsigned lower_old, upper_old;
1366 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1367 &upper_base, &upper_offs))
1368 return NULL_RTX;
1370 lower_anchor_rtx = GEN_INT (lower_base);
1371 upper_anchor_rtx = GEN_INT (upper_base);
1372 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1373 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1375 if (lower_elt)
1376 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1377 if (upper_elt)
1378 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1380 if (!lower_exp)
1381 return upper_exp;
1382 if (!upper_exp)
1383 return lower_exp;
1385 /* Return the older expression. */
1386 return (upper_old > lower_old ? upper_exp : lower_exp);
1389 /* Look in or update the hash table. */
1391 /* Remove table element ELT from use in the table.
1392 HASH is its hash code, made using the HASH macro.
1393 It's an argument because often that is known in advance
1394 and we save much time not recomputing it. */
1396 static void
1397 remove_from_table (struct table_elt *elt, unsigned int hash)
1399 if (elt == 0)
1400 return;
1402 /* Mark this element as removed. See cse_insn. */
1403 elt->first_same_value = 0;
1405 /* Remove the table element from its equivalence class. */
1408 struct table_elt *prev = elt->prev_same_value;
1409 struct table_elt *next = elt->next_same_value;
1411 if (next)
1412 next->prev_same_value = prev;
1414 if (prev)
1415 prev->next_same_value = next;
1416 else
1418 struct table_elt *newfirst = next;
1419 while (next)
1421 next->first_same_value = newfirst;
1422 next = next->next_same_value;
1427 /* Remove the table element from its hash bucket. */
1430 struct table_elt *prev = elt->prev_same_hash;
1431 struct table_elt *next = elt->next_same_hash;
1433 if (next)
1434 next->prev_same_hash = prev;
1436 if (prev)
1437 prev->next_same_hash = next;
1438 else if (table[hash] == elt)
1439 table[hash] = next;
1440 else
1442 /* This entry is not in the proper hash bucket. This can happen
1443 when two classes were merged by `merge_equiv_classes'. Search
1444 for the hash bucket that it heads. This happens only very
1445 rarely, so the cost is acceptable. */
1446 for (hash = 0; hash < HASH_SIZE; hash++)
1447 if (table[hash] == elt)
1448 table[hash] = next;
1452 /* Remove the table element from its related-value circular chain. */
1454 if (elt->related_value != 0 && elt->related_value != elt)
1456 struct table_elt *p = elt->related_value;
1458 while (p->related_value != elt)
1459 p = p->related_value;
1460 p->related_value = elt->related_value;
1461 if (p->related_value == p)
1462 p->related_value = 0;
1465 /* Now add it to the free element chain. */
1466 elt->next_same_hash = free_element_chain;
1467 free_element_chain = elt;
1470 /* Same as above, but X is a pseudo-register. */
1472 static void
1473 remove_pseudo_from_table (rtx x, unsigned int hash)
1475 struct table_elt *elt;
1477 /* Because a pseudo-register can be referenced in more than one
1478 mode, we might have to remove more than one table entry. */
1479 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1480 remove_from_table (elt, hash);
1483 /* Look up X in the hash table and return its table element,
1484 or 0 if X is not in the table.
1486 MODE is the machine-mode of X, or if X is an integer constant
1487 with VOIDmode then MODE is the mode with which X will be used.
1489 Here we are satisfied to find an expression whose tree structure
1490 looks like X. */
1492 static struct table_elt *
1493 lookup (rtx x, unsigned int hash, enum machine_mode mode)
1495 struct table_elt *p;
1497 for (p = table[hash]; p; p = p->next_same_hash)
1498 if (mode == p->mode && ((x == p->exp && REG_P (x))
1499 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1500 return p;
1502 return 0;
1505 /* Like `lookup' but don't care whether the table element uses invalid regs.
1506 Also ignore discrepancies in the machine mode of a register. */
1508 static struct table_elt *
1509 lookup_for_remove (rtx x, unsigned int hash, enum machine_mode mode)
1511 struct table_elt *p;
1513 if (REG_P (x))
1515 unsigned int regno = REGNO (x);
1517 /* Don't check the machine mode when comparing registers;
1518 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1519 for (p = table[hash]; p; p = p->next_same_hash)
1520 if (REG_P (p->exp)
1521 && REGNO (p->exp) == regno)
1522 return p;
1524 else
1526 for (p = table[hash]; p; p = p->next_same_hash)
1527 if (mode == p->mode
1528 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1529 return p;
1532 return 0;
1535 /* Look for an expression equivalent to X and with code CODE.
1536 If one is found, return that expression. */
1538 static rtx
1539 lookup_as_function (rtx x, enum rtx_code code)
1541 struct table_elt *p
1542 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1544 if (p == 0)
1545 return 0;
1547 for (p = p->first_same_value; p; p = p->next_same_value)
1548 if (GET_CODE (p->exp) == code
1549 /* Make sure this is a valid entry in the table. */
1550 && exp_equiv_p (p->exp, p->exp, 1, false))
1551 return p->exp;
1553 return 0;
1556 /* Insert X in the hash table, assuming HASH is its hash code and
1557 CLASSP is an element of the class it should go in (or 0 if a new
1558 class should be made). COST is the code of X and reg_cost is the
1559 cost of registers in X. It is inserted at the proper position to
1560 keep the class in the order cheapest first.
1562 MODE is the machine-mode of X, or if X is an integer constant
1563 with VOIDmode then MODE is the mode with which X will be used.
1565 For elements of equal cheapness, the most recent one
1566 goes in front, except that the first element in the list
1567 remains first unless a cheaper element is added. The order of
1568 pseudo-registers does not matter, as canon_reg will be called to
1569 find the cheapest when a register is retrieved from the table.
1571 The in_memory field in the hash table element is set to 0.
1572 The caller must set it nonzero if appropriate.
1574 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1575 and if insert_regs returns a nonzero value
1576 you must then recompute its hash code before calling here.
1578 If necessary, update table showing constant values of quantities. */
1580 static struct table_elt *
1581 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1582 enum machine_mode mode, int cost, int reg_cost)
1584 struct table_elt *elt;
1586 /* If X is a register and we haven't made a quantity for it,
1587 something is wrong. */
1588 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1590 /* If X is a hard register, show it is being put in the table. */
1591 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1592 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1594 /* Put an element for X into the right hash bucket. */
1596 elt = free_element_chain;
1597 if (elt)
1598 free_element_chain = elt->next_same_hash;
1599 else
1600 elt = XNEW (struct table_elt);
1602 elt->exp = x;
1603 elt->canon_exp = NULL_RTX;
1604 elt->cost = cost;
1605 elt->regcost = reg_cost;
1606 elt->next_same_value = 0;
1607 elt->prev_same_value = 0;
1608 elt->next_same_hash = table[hash];
1609 elt->prev_same_hash = 0;
1610 elt->related_value = 0;
1611 elt->in_memory = 0;
1612 elt->mode = mode;
1613 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1615 if (table[hash])
1616 table[hash]->prev_same_hash = elt;
1617 table[hash] = elt;
1619 /* Put it into the proper value-class. */
1620 if (classp)
1622 classp = classp->first_same_value;
1623 if (CHEAPER (elt, classp))
1624 /* Insert at the head of the class. */
1626 struct table_elt *p;
1627 elt->next_same_value = classp;
1628 classp->prev_same_value = elt;
1629 elt->first_same_value = elt;
1631 for (p = classp; p; p = p->next_same_value)
1632 p->first_same_value = elt;
1634 else
1636 /* Insert not at head of the class. */
1637 /* Put it after the last element cheaper than X. */
1638 struct table_elt *p, *next;
1640 for (p = classp;
1641 (next = p->next_same_value) && CHEAPER (next, elt);
1642 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 else
1851 return 0;
1854 /* Remove from the hash table, or mark as invalid, all expressions whose
1855 values could be altered by storing in X. X is a register, a subreg, or
1856 a memory reference with nonvarying address (because, when a memory
1857 reference with a varying address is stored in, all memory references are
1858 removed by invalidate_memory so specific invalidation is superfluous).
1859 FULL_MODE, if not VOIDmode, indicates that this much should be
1860 invalidated instead of just the amount indicated by the mode of X. This
1861 is only used for bitfield stores into memory.
1863 A nonvarying address may be just a register or just a symbol reference,
1864 or it may be either of those plus a numeric offset. */
1866 static void
1867 invalidate (rtx x, enum machine_mode full_mode)
1869 int i;
1870 struct table_elt *p;
1871 rtx addr;
1873 switch (GET_CODE (x))
1875 case REG:
1877 /* If X is a register, dependencies on its contents are recorded
1878 through the qty number mechanism. Just change the qty number of
1879 the register, mark it as invalid for expressions that refer to it,
1880 and remove it itself. */
1881 unsigned int regno = REGNO (x);
1882 unsigned int hash = HASH (x, GET_MODE (x));
1884 /* Remove REGNO from any quantity list it might be on and indicate
1885 that its value might have changed. If it is a pseudo, remove its
1886 entry from the hash table.
1888 For a hard register, we do the first two actions above for any
1889 additional hard registers corresponding to X. Then, if any of these
1890 registers are in the table, we must remove any REG entries that
1891 overlap these registers. */
1893 delete_reg_equiv (regno);
1894 REG_TICK (regno)++;
1895 SUBREG_TICKED (regno) = -1;
1897 if (regno >= FIRST_PSEUDO_REGISTER)
1898 remove_pseudo_from_table (x, hash);
1899 else
1901 HOST_WIDE_INT in_table
1902 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1903 unsigned int endregno = END_HARD_REGNO (x);
1904 unsigned int tregno, tendregno, rn;
1905 struct table_elt *p, *next;
1907 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1909 for (rn = regno + 1; rn < endregno; rn++)
1911 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1912 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1913 delete_reg_equiv (rn);
1914 REG_TICK (rn)++;
1915 SUBREG_TICKED (rn) = -1;
1918 if (in_table)
1919 for (hash = 0; hash < HASH_SIZE; hash++)
1920 for (p = table[hash]; p; p = next)
1922 next = p->next_same_hash;
1924 if (!REG_P (p->exp)
1925 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1926 continue;
1928 tregno = REGNO (p->exp);
1929 tendregno = END_HARD_REGNO (p->exp);
1930 if (tendregno > regno && tregno < endregno)
1931 remove_from_table (p, hash);
1935 return;
1937 case SUBREG:
1938 invalidate (SUBREG_REG (x), VOIDmode);
1939 return;
1941 case PARALLEL:
1942 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1943 invalidate (XVECEXP (x, 0, i), VOIDmode);
1944 return;
1946 case EXPR_LIST:
1947 /* This is part of a disjoint return value; extract the location in
1948 question ignoring the offset. */
1949 invalidate (XEXP (x, 0), VOIDmode);
1950 return;
1952 case MEM:
1953 addr = canon_rtx (get_addr (XEXP (x, 0)));
1954 /* Calculate the canonical version of X here so that
1955 true_dependence doesn't generate new RTL for X on each call. */
1956 x = canon_rtx (x);
1958 /* Remove all hash table elements that refer to overlapping pieces of
1959 memory. */
1960 if (full_mode == VOIDmode)
1961 full_mode = GET_MODE (x);
1963 for (i = 0; i < HASH_SIZE; i++)
1965 struct table_elt *next;
1967 for (p = table[i]; p; p = next)
1969 next = p->next_same_hash;
1970 if (p->in_memory)
1972 struct check_dependence_data d;
1974 /* Just canonicalize the expression once;
1975 otherwise each time we call invalidate
1976 true_dependence will canonicalize the
1977 expression again. */
1978 if (!p->canon_exp)
1979 p->canon_exp = canon_rtx (p->exp);
1980 d.exp = x;
1981 d.addr = addr;
1982 d.mode = full_mode;
1983 if (for_each_rtx (&p->canon_exp, check_dependence, &d))
1984 remove_from_table (p, i);
1988 return;
1990 default:
1991 gcc_unreachable ();
1995 /* Remove all expressions that refer to register REGNO,
1996 since they are already invalid, and we are about to
1997 mark that register valid again and don't want the old
1998 expressions to reappear as valid. */
2000 static void
2001 remove_invalid_refs (unsigned int regno)
2003 unsigned int i;
2004 struct table_elt *p, *next;
2006 for (i = 0; i < HASH_SIZE; i++)
2007 for (p = table[i]; p; p = next)
2009 next = p->next_same_hash;
2010 if (!REG_P (p->exp)
2011 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2012 remove_from_table (p, i);
2016 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2017 and mode MODE. */
2018 static void
2019 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2020 enum machine_mode mode)
2022 unsigned int i;
2023 struct table_elt *p, *next;
2024 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2026 for (i = 0; i < HASH_SIZE; i++)
2027 for (p = table[i]; p; p = next)
2029 rtx exp = p->exp;
2030 next = p->next_same_hash;
2032 if (!REG_P (exp)
2033 && (GET_CODE (exp) != SUBREG
2034 || !REG_P (SUBREG_REG (exp))
2035 || REGNO (SUBREG_REG (exp)) != regno
2036 || (((SUBREG_BYTE (exp)
2037 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2038 && SUBREG_BYTE (exp) <= end))
2039 && refers_to_regno_p (regno, regno + 1, p->exp, (rtx *) 0))
2040 remove_from_table (p, i);
2044 /* Recompute the hash codes of any valid entries in the hash table that
2045 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2047 This is called when we make a jump equivalence. */
2049 static void
2050 rehash_using_reg (rtx x)
2052 unsigned int i;
2053 struct table_elt *p, *next;
2054 unsigned hash;
2056 if (GET_CODE (x) == SUBREG)
2057 x = SUBREG_REG (x);
2059 /* If X is not a register or if the register is known not to be in any
2060 valid entries in the table, we have no work to do. */
2062 if (!REG_P (x)
2063 || REG_IN_TABLE (REGNO (x)) < 0
2064 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2065 return;
2067 /* Scan all hash chains looking for valid entries that mention X.
2068 If we find one and it is in the wrong hash chain, move it. */
2070 for (i = 0; i < HASH_SIZE; i++)
2071 for (p = table[i]; p; p = next)
2073 next = p->next_same_hash;
2074 if (reg_mentioned_p (x, p->exp)
2075 && exp_equiv_p (p->exp, p->exp, 1, false)
2076 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2078 if (p->next_same_hash)
2079 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2081 if (p->prev_same_hash)
2082 p->prev_same_hash->next_same_hash = p->next_same_hash;
2083 else
2084 table[i] = p->next_same_hash;
2086 p->next_same_hash = table[hash];
2087 p->prev_same_hash = 0;
2088 if (table[hash])
2089 table[hash]->prev_same_hash = p;
2090 table[hash] = p;
2095 /* Remove from the hash table any expression that is a call-clobbered
2096 register. Also update their TICK values. */
2098 static void
2099 invalidate_for_call (void)
2101 unsigned int regno, endregno;
2102 unsigned int i;
2103 unsigned hash;
2104 struct table_elt *p, *next;
2105 int in_table = 0;
2107 /* Go through all the hard registers. For each that is clobbered in
2108 a CALL_INSN, remove the register from quantity chains and update
2109 reg_tick if defined. Also see if any of these registers is currently
2110 in the table. */
2112 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2113 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2115 delete_reg_equiv (regno);
2116 if (REG_TICK (regno) >= 0)
2118 REG_TICK (regno)++;
2119 SUBREG_TICKED (regno) = -1;
2122 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2125 /* In the case where we have no call-clobbered hard registers in the
2126 table, we are done. Otherwise, scan the table and remove any
2127 entry that overlaps a call-clobbered register. */
2129 if (in_table)
2130 for (hash = 0; hash < HASH_SIZE; hash++)
2131 for (p = table[hash]; p; p = next)
2133 next = p->next_same_hash;
2135 if (!REG_P (p->exp)
2136 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2137 continue;
2139 regno = REGNO (p->exp);
2140 endregno = END_HARD_REGNO (p->exp);
2142 for (i = regno; i < endregno; i++)
2143 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2145 remove_from_table (p, hash);
2146 break;
2151 /* Given an expression X of type CONST,
2152 and ELT which is its table entry (or 0 if it
2153 is not in the hash table),
2154 return an alternate expression for X as a register plus integer.
2155 If none can be found, return 0. */
2157 static rtx
2158 use_related_value (rtx x, struct table_elt *elt)
2160 struct table_elt *relt = 0;
2161 struct table_elt *p, *q;
2162 HOST_WIDE_INT offset;
2164 /* First, is there anything related known?
2165 If we have a table element, we can tell from that.
2166 Otherwise, must look it up. */
2168 if (elt != 0 && elt->related_value != 0)
2169 relt = elt;
2170 else if (elt == 0 && GET_CODE (x) == CONST)
2172 rtx subexp = get_related_value (x);
2173 if (subexp != 0)
2174 relt = lookup (subexp,
2175 SAFE_HASH (subexp, GET_MODE (subexp)),
2176 GET_MODE (subexp));
2179 if (relt == 0)
2180 return 0;
2182 /* Search all related table entries for one that has an
2183 equivalent register. */
2185 p = relt;
2186 while (1)
2188 /* This loop is strange in that it is executed in two different cases.
2189 The first is when X is already in the table. Then it is searching
2190 the RELATED_VALUE list of X's class (RELT). The second case is when
2191 X is not in the table. Then RELT points to a class for the related
2192 value.
2194 Ensure that, whatever case we are in, that we ignore classes that have
2195 the same value as X. */
2197 if (rtx_equal_p (x, p->exp))
2198 q = 0;
2199 else
2200 for (q = p->first_same_value; q; q = q->next_same_value)
2201 if (REG_P (q->exp))
2202 break;
2204 if (q)
2205 break;
2207 p = p->related_value;
2209 /* We went all the way around, so there is nothing to be found.
2210 Alternatively, perhaps RELT was in the table for some other reason
2211 and it has no related values recorded. */
2212 if (p == relt || p == 0)
2213 break;
2216 if (q == 0)
2217 return 0;
2219 offset = (get_integer_term (x) - get_integer_term (p->exp));
2220 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2221 return plus_constant (q->mode, q->exp, offset);
2225 /* Hash a string. Just add its bytes up. */
2226 static inline unsigned
2227 hash_rtx_string (const char *ps)
2229 unsigned hash = 0;
2230 const unsigned char *p = (const unsigned char *) ps;
2232 if (p)
2233 while (*p)
2234 hash += *p++;
2236 return hash;
2239 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2240 When the callback returns true, we continue with the new rtx. */
2242 unsigned
2243 hash_rtx_cb (const_rtx x, enum machine_mode mode,
2244 int *do_not_record_p, int *hash_arg_in_memory_p,
2245 bool have_reg_qty, hash_rtx_callback_function cb)
2247 int i, j;
2248 unsigned hash = 0;
2249 enum rtx_code code;
2250 const char *fmt;
2251 enum machine_mode newmode;
2252 rtx newx;
2254 /* Used to turn recursion into iteration. We can't rely on GCC's
2255 tail-recursion elimination since we need to keep accumulating values
2256 in HASH. */
2257 repeat:
2258 if (x == 0)
2259 return hash;
2261 /* Invoke the callback first. */
2262 if (cb != NULL
2263 && ((*cb) (x, mode, &newx, &newmode)))
2265 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2266 hash_arg_in_memory_p, have_reg_qty, cb);
2267 return hash;
2270 code = GET_CODE (x);
2271 switch (code)
2273 case REG:
2275 unsigned int regno = REGNO (x);
2277 if (do_not_record_p && !reload_completed)
2279 /* On some machines, we can't record any non-fixed hard register,
2280 because extending its life will cause reload problems. We
2281 consider ap, fp, sp, gp to be fixed for this purpose.
2283 We also consider CCmode registers to be fixed for this purpose;
2284 failure to do so leads to failure to simplify 0<100 type of
2285 conditionals.
2287 On all machines, we can't record any global registers.
2288 Nor should we record any register that is in a small
2289 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2290 bool record;
2292 if (regno >= FIRST_PSEUDO_REGISTER)
2293 record = true;
2294 else if (x == frame_pointer_rtx
2295 || x == hard_frame_pointer_rtx
2296 || x == arg_pointer_rtx
2297 || x == stack_pointer_rtx
2298 || x == pic_offset_table_rtx)
2299 record = true;
2300 else if (global_regs[regno])
2301 record = false;
2302 else if (fixed_regs[regno])
2303 record = true;
2304 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2305 record = true;
2306 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2307 record = false;
2308 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2309 record = false;
2310 else
2311 record = true;
2313 if (!record)
2315 *do_not_record_p = 1;
2316 return 0;
2320 hash += ((unsigned int) REG << 7);
2321 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2322 return hash;
2325 /* We handle SUBREG of a REG specially because the underlying
2326 reg changes its hash value with every value change; we don't
2327 want to have to forget unrelated subregs when one subreg changes. */
2328 case SUBREG:
2330 if (REG_P (SUBREG_REG (x)))
2332 hash += (((unsigned int) SUBREG << 7)
2333 + REGNO (SUBREG_REG (x))
2334 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2335 return hash;
2337 break;
2340 case CONST_INT:
2341 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2342 + (unsigned int) INTVAL (x));
2343 return hash;
2345 case CONST_DOUBLE:
2346 /* This is like the general case, except that it only counts
2347 the integers representing the constant. */
2348 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2349 if (GET_MODE (x) != VOIDmode)
2350 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2351 else
2352 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2353 + (unsigned int) CONST_DOUBLE_HIGH (x));
2354 return hash;
2356 case CONST_FIXED:
2357 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2358 hash += fixed_hash (CONST_FIXED_VALUE (x));
2359 return hash;
2361 case CONST_VECTOR:
2363 int units;
2364 rtx elt;
2366 units = CONST_VECTOR_NUNITS (x);
2368 for (i = 0; i < units; ++i)
2370 elt = CONST_VECTOR_ELT (x, i);
2371 hash += hash_rtx_cb (elt, GET_MODE (elt),
2372 do_not_record_p, hash_arg_in_memory_p,
2373 have_reg_qty, cb);
2376 return hash;
2379 /* Assume there is only one rtx object for any given label. */
2380 case LABEL_REF:
2381 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2382 differences and differences between each stage's debugging dumps. */
2383 hash += (((unsigned int) LABEL_REF << 7)
2384 + CODE_LABEL_NUMBER (XEXP (x, 0)));
2385 return hash;
2387 case SYMBOL_REF:
2389 /* Don't hash on the symbol's address to avoid bootstrap differences.
2390 Different hash values may cause expressions to be recorded in
2391 different orders and thus different registers to be used in the
2392 final assembler. This also avoids differences in the dump files
2393 between various stages. */
2394 unsigned int h = 0;
2395 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2397 while (*p)
2398 h += (h << 7) + *p++; /* ??? revisit */
2400 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2401 return hash;
2404 case MEM:
2405 /* We don't record if marked volatile or if BLKmode since we don't
2406 know the size of the move. */
2407 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2409 *do_not_record_p = 1;
2410 return 0;
2412 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2413 *hash_arg_in_memory_p = 1;
2415 /* Now that we have already found this special case,
2416 might as well speed it up as much as possible. */
2417 hash += (unsigned) MEM;
2418 x = XEXP (x, 0);
2419 goto repeat;
2421 case USE:
2422 /* A USE that mentions non-volatile memory needs special
2423 handling since the MEM may be BLKmode which normally
2424 prevents an entry from being made. Pure calls are
2425 marked by a USE which mentions BLKmode memory.
2426 See calls.c:emit_call_1. */
2427 if (MEM_P (XEXP (x, 0))
2428 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2430 hash += (unsigned) USE;
2431 x = XEXP (x, 0);
2433 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2434 *hash_arg_in_memory_p = 1;
2436 /* Now that we have already found this special case,
2437 might as well speed it up as much as possible. */
2438 hash += (unsigned) MEM;
2439 x = XEXP (x, 0);
2440 goto repeat;
2442 break;
2444 case PRE_DEC:
2445 case PRE_INC:
2446 case POST_DEC:
2447 case POST_INC:
2448 case PRE_MODIFY:
2449 case POST_MODIFY:
2450 case PC:
2451 case CC0:
2452 case CALL:
2453 case UNSPEC_VOLATILE:
2454 if (do_not_record_p) {
2455 *do_not_record_p = 1;
2456 return 0;
2458 else
2459 return hash;
2460 break;
2462 case ASM_OPERANDS:
2463 if (do_not_record_p && MEM_VOLATILE_P (x))
2465 *do_not_record_p = 1;
2466 return 0;
2468 else
2470 /* We don't want to take the filename and line into account. */
2471 hash += (unsigned) code + (unsigned) GET_MODE (x)
2472 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2473 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2474 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2476 if (ASM_OPERANDS_INPUT_LENGTH (x))
2478 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2480 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2481 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2482 do_not_record_p, hash_arg_in_memory_p,
2483 have_reg_qty, cb)
2484 + hash_rtx_string
2485 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2488 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2489 x = ASM_OPERANDS_INPUT (x, 0);
2490 mode = GET_MODE (x);
2491 goto repeat;
2494 return hash;
2496 break;
2498 default:
2499 break;
2502 i = GET_RTX_LENGTH (code) - 1;
2503 hash += (unsigned) code + (unsigned) GET_MODE (x);
2504 fmt = GET_RTX_FORMAT (code);
2505 for (; i >= 0; i--)
2507 switch (fmt[i])
2509 case 'e':
2510 /* If we are about to do the last recursive call
2511 needed at this level, change it into iteration.
2512 This function is called enough to be worth it. */
2513 if (i == 0)
2515 x = XEXP (x, i);
2516 goto repeat;
2519 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2520 hash_arg_in_memory_p,
2521 have_reg_qty, cb);
2522 break;
2524 case 'E':
2525 for (j = 0; j < XVECLEN (x, i); j++)
2526 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2527 hash_arg_in_memory_p,
2528 have_reg_qty, cb);
2529 break;
2531 case 's':
2532 hash += hash_rtx_string (XSTR (x, i));
2533 break;
2535 case 'i':
2536 hash += (unsigned int) XINT (x, i);
2537 break;
2539 case '0': case 't':
2540 /* Unused. */
2541 break;
2543 default:
2544 gcc_unreachable ();
2548 return hash;
2551 /* Hash an rtx. We are careful to make sure the value is never negative.
2552 Equivalent registers hash identically.
2553 MODE is used in hashing for CONST_INTs only;
2554 otherwise the mode of X is used.
2556 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2558 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2559 a MEM rtx which does not have the RTX_UNCHANGING_P bit set.
2561 Note that cse_insn knows that the hash code of a MEM expression
2562 is just (int) MEM plus the hash code of the address. */
2564 unsigned
2565 hash_rtx (const_rtx x, enum machine_mode mode, int *do_not_record_p,
2566 int *hash_arg_in_memory_p, bool have_reg_qty)
2568 return hash_rtx_cb (x, mode, do_not_record_p,
2569 hash_arg_in_memory_p, have_reg_qty, NULL);
2572 /* Hash an rtx X for cse via hash_rtx.
2573 Stores 1 in do_not_record if any subexpression is volatile.
2574 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2575 does not have the RTX_UNCHANGING_P bit set. */
2577 static inline unsigned
2578 canon_hash (rtx x, enum machine_mode mode)
2580 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2583 /* Like canon_hash but with no side effects, i.e. do_not_record
2584 and hash_arg_in_memory are not changed. */
2586 static inline unsigned
2587 safe_hash (rtx x, enum machine_mode mode)
2589 int dummy_do_not_record;
2590 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2593 /* Return 1 iff X and Y would canonicalize into the same thing,
2594 without actually constructing the canonicalization of either one.
2595 If VALIDATE is nonzero,
2596 we assume X is an expression being processed from the rtl
2597 and Y was found in the hash table. We check register refs
2598 in Y for being marked as valid.
2600 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2603 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2605 int i, j;
2606 enum rtx_code code;
2607 const char *fmt;
2609 /* Note: it is incorrect to assume an expression is equivalent to itself
2610 if VALIDATE is nonzero. */
2611 if (x == y && !validate)
2612 return 1;
2614 if (x == 0 || y == 0)
2615 return x == y;
2617 code = GET_CODE (x);
2618 if (code != GET_CODE (y))
2619 return 0;
2621 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2622 if (GET_MODE (x) != GET_MODE (y))
2623 return 0;
2625 /* MEMs referring to different address space are not equivalent. */
2626 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2627 return 0;
2629 switch (code)
2631 case PC:
2632 case CC0:
2633 case CONST_INT:
2634 case CONST_DOUBLE:
2635 case CONST_FIXED:
2636 return x == y;
2638 case LABEL_REF:
2639 return XEXP (x, 0) == XEXP (y, 0);
2641 case SYMBOL_REF:
2642 return XSTR (x, 0) == XSTR (y, 0);
2644 case REG:
2645 if (for_gcse)
2646 return REGNO (x) == REGNO (y);
2647 else
2649 unsigned int regno = REGNO (y);
2650 unsigned int i;
2651 unsigned int endregno = END_REGNO (y);
2653 /* If the quantities are not the same, the expressions are not
2654 equivalent. If there are and we are not to validate, they
2655 are equivalent. Otherwise, ensure all regs are up-to-date. */
2657 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2658 return 0;
2660 if (! validate)
2661 return 1;
2663 for (i = regno; i < endregno; i++)
2664 if (REG_IN_TABLE (i) != REG_TICK (i))
2665 return 0;
2667 return 1;
2670 case MEM:
2671 if (for_gcse)
2673 /* A volatile mem should not be considered equivalent to any
2674 other. */
2675 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2676 return 0;
2678 /* Can't merge two expressions in different alias sets, since we
2679 can decide that the expression is transparent in a block when
2680 it isn't, due to it being set with the different alias set.
2682 Also, can't merge two expressions with different MEM_ATTRS.
2683 They could e.g. be two different entities allocated into the
2684 same space on the stack (see e.g. PR25130). In that case, the
2685 MEM addresses can be the same, even though the two MEMs are
2686 absolutely not equivalent.
2688 But because really all MEM attributes should be the same for
2689 equivalent MEMs, we just use the invariant that MEMs that have
2690 the same attributes share the same mem_attrs data structure. */
2691 if (MEM_ATTRS (x) != MEM_ATTRS (y))
2692 return 0;
2694 break;
2696 /* For commutative operations, check both orders. */
2697 case PLUS:
2698 case MULT:
2699 case AND:
2700 case IOR:
2701 case XOR:
2702 case NE:
2703 case EQ:
2704 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2705 validate, for_gcse)
2706 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2707 validate, for_gcse))
2708 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2709 validate, for_gcse)
2710 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2711 validate, for_gcse)));
2713 case ASM_OPERANDS:
2714 /* We don't use the generic code below because we want to
2715 disregard filename and line numbers. */
2717 /* A volatile asm isn't equivalent to any other. */
2718 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2719 return 0;
2721 if (GET_MODE (x) != GET_MODE (y)
2722 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2723 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2724 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2725 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2726 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2727 return 0;
2729 if (ASM_OPERANDS_INPUT_LENGTH (x))
2731 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2732 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2733 ASM_OPERANDS_INPUT (y, i),
2734 validate, for_gcse)
2735 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2736 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2737 return 0;
2740 return 1;
2742 default:
2743 break;
2746 /* Compare the elements. If any pair of corresponding elements
2747 fail to match, return 0 for the whole thing. */
2749 fmt = GET_RTX_FORMAT (code);
2750 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2752 switch (fmt[i])
2754 case 'e':
2755 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2756 validate, for_gcse))
2757 return 0;
2758 break;
2760 case 'E':
2761 if (XVECLEN (x, i) != XVECLEN (y, i))
2762 return 0;
2763 for (j = 0; j < XVECLEN (x, i); j++)
2764 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2765 validate, for_gcse))
2766 return 0;
2767 break;
2769 case 's':
2770 if (strcmp (XSTR (x, i), XSTR (y, i)))
2771 return 0;
2772 break;
2774 case 'i':
2775 if (XINT (x, i) != XINT (y, i))
2776 return 0;
2777 break;
2779 case 'w':
2780 if (XWINT (x, i) != XWINT (y, i))
2781 return 0;
2782 break;
2784 case '0':
2785 case 't':
2786 break;
2788 default:
2789 gcc_unreachable ();
2793 return 1;
2796 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2797 the result if necessary. INSN is as for canon_reg. */
2799 static void
2800 validate_canon_reg (rtx *xloc, rtx insn)
2802 if (*xloc)
2804 rtx new_rtx = canon_reg (*xloc, insn);
2806 /* If replacing pseudo with hard reg or vice versa, ensure the
2807 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2808 gcc_assert (insn && new_rtx);
2809 validate_change (insn, xloc, new_rtx, 1);
2813 /* Canonicalize an expression:
2814 replace each register reference inside it
2815 with the "oldest" equivalent register.
2817 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2818 after we make our substitution. The calls are made with IN_GROUP nonzero
2819 so apply_change_group must be called upon the outermost return from this
2820 function (unless INSN is zero). The result of apply_change_group can
2821 generally be discarded since the changes we are making are optional. */
2823 static rtx
2824 canon_reg (rtx x, rtx insn)
2826 int i;
2827 enum rtx_code code;
2828 const char *fmt;
2830 if (x == 0)
2831 return x;
2833 code = GET_CODE (x);
2834 switch (code)
2836 case PC:
2837 case CC0:
2838 case CONST:
2839 case CONST_INT:
2840 case CONST_DOUBLE:
2841 case CONST_FIXED:
2842 case CONST_VECTOR:
2843 case SYMBOL_REF:
2844 case LABEL_REF:
2845 case ADDR_VEC:
2846 case ADDR_DIFF_VEC:
2847 return x;
2849 case REG:
2851 int first;
2852 int q;
2853 struct qty_table_elem *ent;
2855 /* Never replace a hard reg, because hard regs can appear
2856 in more than one machine mode, and we must preserve the mode
2857 of each occurrence. Also, some hard regs appear in
2858 MEMs that are shared and mustn't be altered. Don't try to
2859 replace any reg that maps to a reg of class NO_REGS. */
2860 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2861 || ! REGNO_QTY_VALID_P (REGNO (x)))
2862 return x;
2864 q = REG_QTY (REGNO (x));
2865 ent = &qty_table[q];
2866 first = ent->first_reg;
2867 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2868 : REGNO_REG_CLASS (first) == NO_REGS ? x
2869 : gen_rtx_REG (ent->mode, first));
2872 default:
2873 break;
2876 fmt = GET_RTX_FORMAT (code);
2877 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2879 int j;
2881 if (fmt[i] == 'e')
2882 validate_canon_reg (&XEXP (x, i), insn);
2883 else if (fmt[i] == 'E')
2884 for (j = 0; j < XVECLEN (x, i); j++)
2885 validate_canon_reg (&XVECEXP (x, i, j), insn);
2888 return x;
2891 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2892 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2893 what values are being compared.
2895 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2896 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2897 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2898 compared to produce cc0.
2900 The return value is the comparison operator and is either the code of
2901 A or the code corresponding to the inverse of the comparison. */
2903 static enum rtx_code
2904 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2905 enum machine_mode *pmode1, enum machine_mode *pmode2)
2907 rtx arg1, arg2;
2909 arg1 = *parg1, arg2 = *parg2;
2911 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2913 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2915 /* Set nonzero when we find something of interest. */
2916 rtx x = 0;
2917 int reverse_code = 0;
2918 struct table_elt *p = 0;
2920 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2921 On machines with CC0, this is the only case that can occur, since
2922 fold_rtx will return the COMPARE or item being compared with zero
2923 when given CC0. */
2925 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2926 x = arg1;
2928 /* If ARG1 is a comparison operator and CODE is testing for
2929 STORE_FLAG_VALUE, get the inner arguments. */
2931 else if (COMPARISON_P (arg1))
2933 #ifdef FLOAT_STORE_FLAG_VALUE
2934 REAL_VALUE_TYPE fsfv;
2935 #endif
2937 if (code == NE
2938 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2939 && code == LT && STORE_FLAG_VALUE == -1)
2940 #ifdef FLOAT_STORE_FLAG_VALUE
2941 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2942 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2943 REAL_VALUE_NEGATIVE (fsfv)))
2944 #endif
2946 x = arg1;
2947 else if (code == EQ
2948 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2949 && code == GE && STORE_FLAG_VALUE == -1)
2950 #ifdef FLOAT_STORE_FLAG_VALUE
2951 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2952 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2953 REAL_VALUE_NEGATIVE (fsfv)))
2954 #endif
2956 x = arg1, reverse_code = 1;
2959 /* ??? We could also check for
2961 (ne (and (eq (...) (const_int 1))) (const_int 0))
2963 and related forms, but let's wait until we see them occurring. */
2965 if (x == 0)
2966 /* Look up ARG1 in the hash table and see if it has an equivalence
2967 that lets us see what is being compared. */
2968 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2969 if (p)
2971 p = p->first_same_value;
2973 /* If what we compare is already known to be constant, that is as
2974 good as it gets.
2975 We need to break the loop in this case, because otherwise we
2976 can have an infinite loop when looking at a reg that is known
2977 to be a constant which is the same as a comparison of a reg
2978 against zero which appears later in the insn stream, which in
2979 turn is constant and the same as the comparison of the first reg
2980 against zero... */
2981 if (p->is_const)
2982 break;
2985 for (; p; p = p->next_same_value)
2987 enum machine_mode inner_mode = GET_MODE (p->exp);
2988 #ifdef FLOAT_STORE_FLAG_VALUE
2989 REAL_VALUE_TYPE fsfv;
2990 #endif
2992 /* If the entry isn't valid, skip it. */
2993 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2994 continue;
2996 /* If it's the same comparison we're already looking at, skip it. */
2997 if (COMPARISON_P (p->exp)
2998 && XEXP (p->exp, 0) == arg1
2999 && XEXP (p->exp, 1) == arg2)
3000 continue;
3002 if (GET_CODE (p->exp) == COMPARE
3003 /* Another possibility is that this machine has a compare insn
3004 that includes the comparison code. In that case, ARG1 would
3005 be equivalent to a comparison operation that would set ARG1 to
3006 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3007 ORIG_CODE is the actual comparison being done; if it is an EQ,
3008 we must reverse ORIG_CODE. On machine with a negative value
3009 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3010 || ((code == NE
3011 || (code == LT
3012 && val_signbit_known_set_p (inner_mode,
3013 STORE_FLAG_VALUE))
3014 #ifdef FLOAT_STORE_FLAG_VALUE
3015 || (code == LT
3016 && SCALAR_FLOAT_MODE_P (inner_mode)
3017 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3018 REAL_VALUE_NEGATIVE (fsfv)))
3019 #endif
3021 && COMPARISON_P (p->exp)))
3023 x = p->exp;
3024 break;
3026 else if ((code == EQ
3027 || (code == GE
3028 && val_signbit_known_set_p (inner_mode,
3029 STORE_FLAG_VALUE))
3030 #ifdef FLOAT_STORE_FLAG_VALUE
3031 || (code == GE
3032 && SCALAR_FLOAT_MODE_P (inner_mode)
3033 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3034 REAL_VALUE_NEGATIVE (fsfv)))
3035 #endif
3037 && COMPARISON_P (p->exp))
3039 reverse_code = 1;
3040 x = p->exp;
3041 break;
3044 /* If this non-trapping address, e.g. fp + constant, the
3045 equivalent is a better operand since it may let us predict
3046 the value of the comparison. */
3047 else if (!rtx_addr_can_trap_p (p->exp))
3049 arg1 = p->exp;
3050 continue;
3054 /* If we didn't find a useful equivalence for ARG1, we are done.
3055 Otherwise, set up for the next iteration. */
3056 if (x == 0)
3057 break;
3059 /* If we need to reverse the comparison, make sure that that is
3060 possible -- we can't necessarily infer the value of GE from LT
3061 with floating-point operands. */
3062 if (reverse_code)
3064 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3065 if (reversed == UNKNOWN)
3066 break;
3067 else
3068 code = reversed;
3070 else if (COMPARISON_P (x))
3071 code = GET_CODE (x);
3072 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3075 /* Return our results. Return the modes from before fold_rtx
3076 because fold_rtx might produce const_int, and then it's too late. */
3077 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3078 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3080 return code;
3083 /* If X is a nontrivial arithmetic operation on an argument for which
3084 a constant value can be determined, return the result of operating
3085 on that value, as a constant. Otherwise, return X, possibly with
3086 one or more operands changed to a forward-propagated constant.
3088 If X is a register whose contents are known, we do NOT return
3089 those contents here; equiv_constant is called to perform that task.
3090 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3092 INSN is the insn that we may be modifying. If it is 0, make a copy
3093 of X before modifying it. */
3095 static rtx
3096 fold_rtx (rtx x, rtx insn)
3098 enum rtx_code code;
3099 enum machine_mode mode;
3100 const char *fmt;
3101 int i;
3102 rtx new_rtx = 0;
3103 int changed = 0;
3105 /* Operands of X. */
3106 rtx folded_arg0;
3107 rtx folded_arg1;
3109 /* Constant equivalents of first three operands of X;
3110 0 when no such equivalent is known. */
3111 rtx const_arg0;
3112 rtx const_arg1;
3113 rtx const_arg2;
3115 /* The mode of the first operand of X. We need this for sign and zero
3116 extends. */
3117 enum machine_mode mode_arg0;
3119 if (x == 0)
3120 return x;
3122 /* Try to perform some initial simplifications on X. */
3123 code = GET_CODE (x);
3124 switch (code)
3126 case MEM:
3127 case SUBREG:
3128 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3129 return new_rtx;
3130 return x;
3132 case CONST:
3133 case CONST_INT:
3134 case CONST_DOUBLE:
3135 case CONST_FIXED:
3136 case CONST_VECTOR:
3137 case SYMBOL_REF:
3138 case LABEL_REF:
3139 case REG:
3140 case PC:
3141 /* No use simplifying an EXPR_LIST
3142 since they are used only for lists of args
3143 in a function call's REG_EQUAL note. */
3144 case EXPR_LIST:
3145 return x;
3147 #ifdef HAVE_cc0
3148 case CC0:
3149 return prev_insn_cc0;
3150 #endif
3152 case ASM_OPERANDS:
3153 if (insn)
3155 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3156 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3157 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3159 return x;
3161 #ifdef NO_FUNCTION_CSE
3162 case CALL:
3163 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3164 return x;
3165 break;
3166 #endif
3168 /* Anything else goes through the loop below. */
3169 default:
3170 break;
3173 mode = GET_MODE (x);
3174 const_arg0 = 0;
3175 const_arg1 = 0;
3176 const_arg2 = 0;
3177 mode_arg0 = VOIDmode;
3179 /* Try folding our operands.
3180 Then see which ones have constant values known. */
3182 fmt = GET_RTX_FORMAT (code);
3183 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3184 if (fmt[i] == 'e')
3186 rtx folded_arg = XEXP (x, i), const_arg;
3187 enum machine_mode mode_arg = GET_MODE (folded_arg);
3189 switch (GET_CODE (folded_arg))
3191 case MEM:
3192 case REG:
3193 case SUBREG:
3194 const_arg = equiv_constant (folded_arg);
3195 break;
3197 case CONST:
3198 case CONST_INT:
3199 case SYMBOL_REF:
3200 case LABEL_REF:
3201 case CONST_DOUBLE:
3202 case CONST_FIXED:
3203 case CONST_VECTOR:
3204 const_arg = folded_arg;
3205 break;
3207 #ifdef HAVE_cc0
3208 case CC0:
3209 folded_arg = prev_insn_cc0;
3210 mode_arg = prev_insn_cc0_mode;
3211 const_arg = equiv_constant (folded_arg);
3212 break;
3213 #endif
3215 default:
3216 folded_arg = fold_rtx (folded_arg, insn);
3217 const_arg = equiv_constant (folded_arg);
3218 break;
3221 /* For the first three operands, see if the operand
3222 is constant or equivalent to a constant. */
3223 switch (i)
3225 case 0:
3226 folded_arg0 = folded_arg;
3227 const_arg0 = const_arg;
3228 mode_arg0 = mode_arg;
3229 break;
3230 case 1:
3231 folded_arg1 = folded_arg;
3232 const_arg1 = const_arg;
3233 break;
3234 case 2:
3235 const_arg2 = const_arg;
3236 break;
3239 /* Pick the least expensive of the argument and an equivalent constant
3240 argument. */
3241 if (const_arg != 0
3242 && const_arg != folded_arg
3243 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3245 /* It's not safe to substitute the operand of a conversion
3246 operator with a constant, as the conversion's identity
3247 depends upon the mode of its operand. This optimization
3248 is handled by the call to simplify_unary_operation. */
3249 && (GET_RTX_CLASS (code) != RTX_UNARY
3250 || GET_MODE (const_arg) == mode_arg0
3251 || (code != ZERO_EXTEND
3252 && code != SIGN_EXTEND
3253 && code != TRUNCATE
3254 && code != FLOAT_TRUNCATE
3255 && code != FLOAT_EXTEND
3256 && code != FLOAT
3257 && code != FIX
3258 && code != UNSIGNED_FLOAT
3259 && code != UNSIGNED_FIX)))
3260 folded_arg = const_arg;
3262 if (folded_arg == XEXP (x, i))
3263 continue;
3265 if (insn == NULL_RTX && !changed)
3266 x = copy_rtx (x);
3267 changed = 1;
3268 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3271 if (changed)
3273 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3274 consistent with the order in X. */
3275 if (canonicalize_change_group (insn, x))
3277 rtx tem;
3278 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3279 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3282 apply_change_group ();
3285 /* If X is an arithmetic operation, see if we can simplify it. */
3287 switch (GET_RTX_CLASS (code))
3289 case RTX_UNARY:
3291 /* We can't simplify extension ops unless we know the
3292 original mode. */
3293 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3294 && mode_arg0 == VOIDmode)
3295 break;
3297 new_rtx = simplify_unary_operation (code, mode,
3298 const_arg0 ? const_arg0 : folded_arg0,
3299 mode_arg0);
3301 break;
3303 case RTX_COMPARE:
3304 case RTX_COMM_COMPARE:
3305 /* See what items are actually being compared and set FOLDED_ARG[01]
3306 to those values and CODE to the actual comparison code. If any are
3307 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3308 do anything if both operands are already known to be constant. */
3310 /* ??? Vector mode comparisons are not supported yet. */
3311 if (VECTOR_MODE_P (mode))
3312 break;
3314 if (const_arg0 == 0 || const_arg1 == 0)
3316 struct table_elt *p0, *p1;
3317 rtx true_rtx, false_rtx;
3318 enum machine_mode mode_arg1;
3320 if (SCALAR_FLOAT_MODE_P (mode))
3322 #ifdef FLOAT_STORE_FLAG_VALUE
3323 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3324 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3325 #else
3326 true_rtx = NULL_RTX;
3327 #endif
3328 false_rtx = CONST0_RTX (mode);
3330 else
3332 true_rtx = const_true_rtx;
3333 false_rtx = const0_rtx;
3336 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3337 &mode_arg0, &mode_arg1);
3339 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3340 what kinds of things are being compared, so we can't do
3341 anything with this comparison. */
3343 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3344 break;
3346 const_arg0 = equiv_constant (folded_arg0);
3347 const_arg1 = equiv_constant (folded_arg1);
3349 /* If we do not now have two constants being compared, see
3350 if we can nevertheless deduce some things about the
3351 comparison. */
3352 if (const_arg0 == 0 || const_arg1 == 0)
3354 if (const_arg1 != NULL)
3356 rtx cheapest_simplification;
3357 int cheapest_cost;
3358 rtx simp_result;
3359 struct table_elt *p;
3361 /* See if we can find an equivalent of folded_arg0
3362 that gets us a cheaper expression, possibly a
3363 constant through simplifications. */
3364 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3365 mode_arg0);
3367 if (p != NULL)
3369 cheapest_simplification = x;
3370 cheapest_cost = COST (x);
3372 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3374 int cost;
3376 /* If the entry isn't valid, skip it. */
3377 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3378 continue;
3380 /* Try to simplify using this equivalence. */
3381 simp_result
3382 = simplify_relational_operation (code, mode,
3383 mode_arg0,
3384 p->exp,
3385 const_arg1);
3387 if (simp_result == NULL)
3388 continue;
3390 cost = COST (simp_result);
3391 if (cost < cheapest_cost)
3393 cheapest_cost = cost;
3394 cheapest_simplification = simp_result;
3398 /* If we have a cheaper expression now, use that
3399 and try folding it further, from the top. */
3400 if (cheapest_simplification != x)
3401 return fold_rtx (copy_rtx (cheapest_simplification),
3402 insn);
3406 /* See if the two operands are the same. */
3408 if ((REG_P (folded_arg0)
3409 && REG_P (folded_arg1)
3410 && (REG_QTY (REGNO (folded_arg0))
3411 == REG_QTY (REGNO (folded_arg1))))
3412 || ((p0 = lookup (folded_arg0,
3413 SAFE_HASH (folded_arg0, mode_arg0),
3414 mode_arg0))
3415 && (p1 = lookup (folded_arg1,
3416 SAFE_HASH (folded_arg1, mode_arg0),
3417 mode_arg0))
3418 && p0->first_same_value == p1->first_same_value))
3419 folded_arg1 = folded_arg0;
3421 /* If FOLDED_ARG0 is a register, see if the comparison we are
3422 doing now is either the same as we did before or the reverse
3423 (we only check the reverse if not floating-point). */
3424 else if (REG_P (folded_arg0))
3426 int qty = REG_QTY (REGNO (folded_arg0));
3428 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3430 struct qty_table_elem *ent = &qty_table[qty];
3432 if ((comparison_dominates_p (ent->comparison_code, code)
3433 || (! FLOAT_MODE_P (mode_arg0)
3434 && comparison_dominates_p (ent->comparison_code,
3435 reverse_condition (code))))
3436 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3437 || (const_arg1
3438 && rtx_equal_p (ent->comparison_const,
3439 const_arg1))
3440 || (REG_P (folded_arg1)
3441 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3443 if (comparison_dominates_p (ent->comparison_code, code))
3445 if (true_rtx)
3446 return true_rtx;
3447 else
3448 break;
3450 else
3451 return false_rtx;
3458 /* If we are comparing against zero, see if the first operand is
3459 equivalent to an IOR with a constant. If so, we may be able to
3460 determine the result of this comparison. */
3461 if (const_arg1 == const0_rtx && !const_arg0)
3463 rtx y = lookup_as_function (folded_arg0, IOR);
3464 rtx inner_const;
3466 if (y != 0
3467 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3468 && CONST_INT_P (inner_const)
3469 && INTVAL (inner_const) != 0)
3470 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3474 rtx op0 = const_arg0 ? const_arg0 : folded_arg0;
3475 rtx op1 = const_arg1 ? const_arg1 : folded_arg1;
3476 new_rtx = simplify_relational_operation (code, mode, mode_arg0, op0, op1);
3478 break;
3480 case RTX_BIN_ARITH:
3481 case RTX_COMM_ARITH:
3482 switch (code)
3484 case PLUS:
3485 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3486 with that LABEL_REF as its second operand. If so, the result is
3487 the first operand of that MINUS. This handles switches with an
3488 ADDR_DIFF_VEC table. */
3489 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3491 rtx y
3492 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3493 : lookup_as_function (folded_arg0, MINUS);
3495 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3496 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg1, 0))
3497 return XEXP (y, 0);
3499 /* Now try for a CONST of a MINUS like the above. */
3500 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3501 : lookup_as_function (folded_arg0, CONST))) != 0
3502 && GET_CODE (XEXP (y, 0)) == MINUS
3503 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3504 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg1, 0))
3505 return XEXP (XEXP (y, 0), 0);
3508 /* Likewise if the operands are in the other order. */
3509 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3511 rtx y
3512 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3513 : lookup_as_function (folded_arg1, MINUS);
3515 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3516 && XEXP (XEXP (y, 1), 0) == XEXP (const_arg0, 0))
3517 return XEXP (y, 0);
3519 /* Now try for a CONST of a MINUS like the above. */
3520 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3521 : lookup_as_function (folded_arg1, CONST))) != 0
3522 && GET_CODE (XEXP (y, 0)) == MINUS
3523 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3524 && XEXP (XEXP (XEXP (y, 0), 1), 0) == XEXP (const_arg0, 0))
3525 return XEXP (XEXP (y, 0), 0);
3528 /* If second operand is a register equivalent to a negative
3529 CONST_INT, see if we can find a register equivalent to the
3530 positive constant. Make a MINUS if so. Don't do this for
3531 a non-negative constant since we might then alternate between
3532 choosing positive and negative constants. Having the positive
3533 constant previously-used is the more common case. Be sure
3534 the resulting constant is non-negative; if const_arg1 were
3535 the smallest negative number this would overflow: depending
3536 on the mode, this would either just be the same value (and
3537 hence not save anything) or be incorrect. */
3538 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3539 && INTVAL (const_arg1) < 0
3540 /* This used to test
3542 -INTVAL (const_arg1) >= 0
3544 But The Sun V5.0 compilers mis-compiled that test. So
3545 instead we test for the problematic value in a more direct
3546 manner and hope the Sun compilers get it correct. */
3547 && INTVAL (const_arg1) !=
3548 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3549 && REG_P (folded_arg1))
3551 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3552 struct table_elt *p
3553 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3555 if (p)
3556 for (p = p->first_same_value; p; p = p->next_same_value)
3557 if (REG_P (p->exp))
3558 return simplify_gen_binary (MINUS, mode, folded_arg0,
3559 canon_reg (p->exp, NULL_RTX));
3561 goto from_plus;
3563 case MINUS:
3564 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3565 If so, produce (PLUS Z C2-C). */
3566 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3568 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3569 if (y && CONST_INT_P (XEXP (y, 1)))
3570 return fold_rtx (plus_constant (mode, copy_rtx (y),
3571 -INTVAL (const_arg1)),
3572 NULL_RTX);
3575 /* Fall through. */
3577 from_plus:
3578 case SMIN: case SMAX: case UMIN: case UMAX:
3579 case IOR: case AND: case XOR:
3580 case MULT:
3581 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3582 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3583 is known to be of similar form, we may be able to replace the
3584 operation with a combined operation. This may eliminate the
3585 intermediate operation if every use is simplified in this way.
3586 Note that the similar optimization done by combine.c only works
3587 if the intermediate operation's result has only one reference. */
3589 if (REG_P (folded_arg0)
3590 && const_arg1 && CONST_INT_P (const_arg1))
3592 int is_shift
3593 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3594 rtx y, inner_const, new_const;
3595 rtx canon_const_arg1 = const_arg1;
3596 enum rtx_code associate_code;
3598 if (is_shift
3599 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3600 || INTVAL (const_arg1) < 0))
3602 if (SHIFT_COUNT_TRUNCATED)
3603 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3604 & (GET_MODE_BITSIZE (mode)
3605 - 1));
3606 else
3607 break;
3610 y = lookup_as_function (folded_arg0, code);
3611 if (y == 0)
3612 break;
3614 /* If we have compiled a statement like
3615 "if (x == (x & mask1))", and now are looking at
3616 "x & mask2", we will have a case where the first operand
3617 of Y is the same as our first operand. Unless we detect
3618 this case, an infinite loop will result. */
3619 if (XEXP (y, 0) == folded_arg0)
3620 break;
3622 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3623 if (!inner_const || !CONST_INT_P (inner_const))
3624 break;
3626 /* Don't associate these operations if they are a PLUS with the
3627 same constant and it is a power of two. These might be doable
3628 with a pre- or post-increment. Similarly for two subtracts of
3629 identical powers of two with post decrement. */
3631 if (code == PLUS && const_arg1 == inner_const
3632 && ((HAVE_PRE_INCREMENT
3633 && exact_log2 (INTVAL (const_arg1)) >= 0)
3634 || (HAVE_POST_INCREMENT
3635 && exact_log2 (INTVAL (const_arg1)) >= 0)
3636 || (HAVE_PRE_DECREMENT
3637 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3638 || (HAVE_POST_DECREMENT
3639 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3640 break;
3642 /* ??? Vector mode shifts by scalar
3643 shift operand are not supported yet. */
3644 if (is_shift && VECTOR_MODE_P (mode))
3645 break;
3647 if (is_shift
3648 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3649 || INTVAL (inner_const) < 0))
3651 if (SHIFT_COUNT_TRUNCATED)
3652 inner_const = GEN_INT (INTVAL (inner_const)
3653 & (GET_MODE_BITSIZE (mode) - 1));
3654 else
3655 break;
3658 /* Compute the code used to compose the constants. For example,
3659 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3661 associate_code = (is_shift || code == MINUS ? PLUS : code);
3663 new_const = simplify_binary_operation (associate_code, mode,
3664 canon_const_arg1,
3665 inner_const);
3667 if (new_const == 0)
3668 break;
3670 /* If we are associating shift operations, don't let this
3671 produce a shift of the size of the object or larger.
3672 This could occur when we follow a sign-extend by a right
3673 shift on a machine that does a sign-extend as a pair
3674 of shifts. */
3676 if (is_shift
3677 && CONST_INT_P (new_const)
3678 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3680 /* As an exception, we can turn an ASHIFTRT of this
3681 form into a shift of the number of bits - 1. */
3682 if (code == ASHIFTRT)
3683 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3684 else if (!side_effects_p (XEXP (y, 0)))
3685 return CONST0_RTX (mode);
3686 else
3687 break;
3690 y = copy_rtx (XEXP (y, 0));
3692 /* If Y contains our first operand (the most common way this
3693 can happen is if Y is a MEM), we would do into an infinite
3694 loop if we tried to fold it. So don't in that case. */
3696 if (! reg_mentioned_p (folded_arg0, y))
3697 y = fold_rtx (y, insn);
3699 return simplify_gen_binary (code, mode, y, new_const);
3701 break;
3703 case DIV: case UDIV:
3704 /* ??? The associative optimization performed immediately above is
3705 also possible for DIV and UDIV using associate_code of MULT.
3706 However, we would need extra code to verify that the
3707 multiplication does not overflow, that is, there is no overflow
3708 in the calculation of new_const. */
3709 break;
3711 default:
3712 break;
3715 new_rtx = simplify_binary_operation (code, mode,
3716 const_arg0 ? const_arg0 : folded_arg0,
3717 const_arg1 ? const_arg1 : folded_arg1);
3718 break;
3720 case RTX_OBJ:
3721 /* (lo_sum (high X) X) is simply X. */
3722 if (code == LO_SUM && const_arg0 != 0
3723 && GET_CODE (const_arg0) == HIGH
3724 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3725 return const_arg1;
3726 break;
3728 case RTX_TERNARY:
3729 case RTX_BITFIELD_OPS:
3730 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3731 const_arg0 ? const_arg0 : folded_arg0,
3732 const_arg1 ? const_arg1 : folded_arg1,
3733 const_arg2 ? const_arg2 : XEXP (x, 2));
3734 break;
3736 default:
3737 break;
3740 return new_rtx ? new_rtx : x;
3743 /* Return a constant value currently equivalent to X.
3744 Return 0 if we don't know one. */
3746 static rtx
3747 equiv_constant (rtx x)
3749 if (REG_P (x)
3750 && REGNO_QTY_VALID_P (REGNO (x)))
3752 int x_q = REG_QTY (REGNO (x));
3753 struct qty_table_elem *x_ent = &qty_table[x_q];
3755 if (x_ent->const_rtx)
3756 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3759 if (x == 0 || CONSTANT_P (x))
3760 return x;
3762 if (GET_CODE (x) == SUBREG)
3764 enum machine_mode mode = GET_MODE (x);
3765 enum machine_mode imode = GET_MODE (SUBREG_REG (x));
3766 rtx new_rtx;
3768 /* See if we previously assigned a constant value to this SUBREG. */
3769 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3770 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3771 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3772 return new_rtx;
3774 /* If we didn't and if doing so makes sense, see if we previously
3775 assigned a constant value to the enclosing word mode SUBREG. */
3776 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3777 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3779 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3780 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3782 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3783 new_rtx = lookup_as_function (y, CONST_INT);
3784 if (new_rtx)
3785 return gen_lowpart (mode, new_rtx);
3789 /* Otherwise see if we already have a constant for the inner REG,
3790 and if that is enough to calculate an equivalent constant for
3791 the subreg. Note that the upper bits of paradoxical subregs
3792 are undefined, so they cannot be said to equal anything. */
3793 if (REG_P (SUBREG_REG (x))
3794 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3795 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3796 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3798 return 0;
3801 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3802 the hash table in case its value was seen before. */
3804 if (MEM_P (x))
3806 struct table_elt *elt;
3808 x = avoid_constant_pool_reference (x);
3809 if (CONSTANT_P (x))
3810 return x;
3812 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3813 if (elt == 0)
3814 return 0;
3816 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3817 if (elt->is_const && CONSTANT_P (elt->exp))
3818 return elt->exp;
3821 return 0;
3824 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3825 "taken" branch.
3827 In certain cases, this can cause us to add an equivalence. For example,
3828 if we are following the taken case of
3829 if (i == 2)
3830 we can add the fact that `i' and '2' are now equivalent.
3832 In any case, we can record that this comparison was passed. If the same
3833 comparison is seen later, we will know its value. */
3835 static void
3836 record_jump_equiv (rtx insn, bool taken)
3838 int cond_known_true;
3839 rtx op0, op1;
3840 rtx set;
3841 enum machine_mode mode, mode0, mode1;
3842 int reversed_nonequality = 0;
3843 enum rtx_code code;
3845 /* Ensure this is the right kind of insn. */
3846 gcc_assert (any_condjump_p (insn));
3848 set = pc_set (insn);
3850 /* See if this jump condition is known true or false. */
3851 if (taken)
3852 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3853 else
3854 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3856 /* Get the type of comparison being done and the operands being compared.
3857 If we had to reverse a non-equality condition, record that fact so we
3858 know that it isn't valid for floating-point. */
3859 code = GET_CODE (XEXP (SET_SRC (set), 0));
3860 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3861 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3863 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3864 if (! cond_known_true)
3866 code = reversed_comparison_code_parts (code, op0, op1, insn);
3868 /* Don't remember if we can't find the inverse. */
3869 if (code == UNKNOWN)
3870 return;
3873 /* The mode is the mode of the non-constant. */
3874 mode = mode0;
3875 if (mode1 != VOIDmode)
3876 mode = mode1;
3878 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3881 /* Yet another form of subreg creation. In this case, we want something in
3882 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3884 static rtx
3885 record_jump_cond_subreg (enum machine_mode mode, rtx op)
3887 enum machine_mode op_mode = GET_MODE (op);
3888 if (op_mode == mode || op_mode == VOIDmode)
3889 return op;
3890 return lowpart_subreg (mode, op, op_mode);
3893 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3894 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3895 Make any useful entries we can with that information. Called from
3896 above function and called recursively. */
3898 static void
3899 record_jump_cond (enum rtx_code code, enum machine_mode mode, rtx op0,
3900 rtx op1, int reversed_nonequality)
3902 unsigned op0_hash, op1_hash;
3903 int op0_in_memory, op1_in_memory;
3904 struct table_elt *op0_elt, *op1_elt;
3906 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3907 we know that they are also equal in the smaller mode (this is also
3908 true for all smaller modes whether or not there is a SUBREG, but
3909 is not worth testing for with no SUBREG). */
3911 /* Note that GET_MODE (op0) may not equal MODE. */
3912 if (code == EQ && paradoxical_subreg_p (op0))
3914 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3915 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3916 if (tem)
3917 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3918 reversed_nonequality);
3921 if (code == EQ && paradoxical_subreg_p (op1))
3923 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3924 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3925 if (tem)
3926 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3927 reversed_nonequality);
3930 /* Similarly, if this is an NE comparison, and either is a SUBREG
3931 making a smaller mode, we know the whole thing is also NE. */
3933 /* Note that GET_MODE (op0) may not equal MODE;
3934 if we test MODE instead, we can get an infinite recursion
3935 alternating between two modes each wider than MODE. */
3937 if (code == NE && GET_CODE (op0) == SUBREG
3938 && subreg_lowpart_p (op0)
3939 && (GET_MODE_SIZE (GET_MODE (op0))
3940 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3942 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3943 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3944 if (tem)
3945 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3946 reversed_nonequality);
3949 if (code == NE && GET_CODE (op1) == SUBREG
3950 && subreg_lowpart_p (op1)
3951 && (GET_MODE_SIZE (GET_MODE (op1))
3952 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3954 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3955 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3956 if (tem)
3957 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3958 reversed_nonequality);
3961 /* Hash both operands. */
3963 do_not_record = 0;
3964 hash_arg_in_memory = 0;
3965 op0_hash = HASH (op0, mode);
3966 op0_in_memory = hash_arg_in_memory;
3968 if (do_not_record)
3969 return;
3971 do_not_record = 0;
3972 hash_arg_in_memory = 0;
3973 op1_hash = HASH (op1, mode);
3974 op1_in_memory = hash_arg_in_memory;
3976 if (do_not_record)
3977 return;
3979 /* Look up both operands. */
3980 op0_elt = lookup (op0, op0_hash, mode);
3981 op1_elt = lookup (op1, op1_hash, mode);
3983 /* If both operands are already equivalent or if they are not in the
3984 table but are identical, do nothing. */
3985 if ((op0_elt != 0 && op1_elt != 0
3986 && op0_elt->first_same_value == op1_elt->first_same_value)
3987 || op0 == op1 || rtx_equal_p (op0, op1))
3988 return;
3990 /* If we aren't setting two things equal all we can do is save this
3991 comparison. Similarly if this is floating-point. In the latter
3992 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
3993 If we record the equality, we might inadvertently delete code
3994 whose intent was to change -0 to +0. */
3996 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
3998 struct qty_table_elem *ent;
3999 int qty;
4001 /* If we reversed a floating-point comparison, if OP0 is not a
4002 register, or if OP1 is neither a register or constant, we can't
4003 do anything. */
4005 if (!REG_P (op1))
4006 op1 = equiv_constant (op1);
4008 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4009 || !REG_P (op0) || op1 == 0)
4010 return;
4012 /* Put OP0 in the hash table if it isn't already. This gives it a
4013 new quantity number. */
4014 if (op0_elt == 0)
4016 if (insert_regs (op0, NULL, 0))
4018 rehash_using_reg (op0);
4019 op0_hash = HASH (op0, mode);
4021 /* If OP0 is contained in OP1, this changes its hash code
4022 as well. Faster to rehash than to check, except
4023 for the simple case of a constant. */
4024 if (! CONSTANT_P (op1))
4025 op1_hash = HASH (op1,mode);
4028 op0_elt = insert (op0, NULL, op0_hash, mode);
4029 op0_elt->in_memory = op0_in_memory;
4032 qty = REG_QTY (REGNO (op0));
4033 ent = &qty_table[qty];
4035 ent->comparison_code = code;
4036 if (REG_P (op1))
4038 /* Look it up again--in case op0 and op1 are the same. */
4039 op1_elt = lookup (op1, op1_hash, mode);
4041 /* Put OP1 in the hash table so it gets a new quantity number. */
4042 if (op1_elt == 0)
4044 if (insert_regs (op1, NULL, 0))
4046 rehash_using_reg (op1);
4047 op1_hash = HASH (op1, mode);
4050 op1_elt = insert (op1, NULL, op1_hash, mode);
4051 op1_elt->in_memory = op1_in_memory;
4054 ent->comparison_const = NULL_RTX;
4055 ent->comparison_qty = REG_QTY (REGNO (op1));
4057 else
4059 ent->comparison_const = op1;
4060 ent->comparison_qty = -1;
4063 return;
4066 /* If either side is still missing an equivalence, make it now,
4067 then merge the equivalences. */
4069 if (op0_elt == 0)
4071 if (insert_regs (op0, NULL, 0))
4073 rehash_using_reg (op0);
4074 op0_hash = HASH (op0, mode);
4077 op0_elt = insert (op0, NULL, op0_hash, mode);
4078 op0_elt->in_memory = op0_in_memory;
4081 if (op1_elt == 0)
4083 if (insert_regs (op1, NULL, 0))
4085 rehash_using_reg (op1);
4086 op1_hash = HASH (op1, mode);
4089 op1_elt = insert (op1, NULL, op1_hash, mode);
4090 op1_elt->in_memory = op1_in_memory;
4093 merge_equiv_classes (op0_elt, op1_elt);
4096 /* CSE processing for one instruction.
4098 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4099 but the few that "leak through" are cleaned up by cse_insn, and complex
4100 addressing modes are often formed here.
4102 The main function is cse_insn, and between here and that function
4103 a couple of helper functions is defined to keep the size of cse_insn
4104 within reasonable proportions.
4106 Data is shared between the main and helper functions via STRUCT SET,
4107 that contains all data related for every set in the instruction that
4108 is being processed.
4110 Note that cse_main processes all sets in the instruction. Most
4111 passes in GCC only process simple SET insns or single_set insns, but
4112 CSE processes insns with multiple sets as well. */
4114 /* Data on one SET contained in the instruction. */
4116 struct set
4118 /* The SET rtx itself. */
4119 rtx rtl;
4120 /* The SET_SRC of the rtx (the original value, if it is changing). */
4121 rtx src;
4122 /* The hash-table element for the SET_SRC of the SET. */
4123 struct table_elt *src_elt;
4124 /* Hash value for the SET_SRC. */
4125 unsigned src_hash;
4126 /* Hash value for the SET_DEST. */
4127 unsigned dest_hash;
4128 /* The SET_DEST, with SUBREG, etc., stripped. */
4129 rtx inner_dest;
4130 /* Nonzero if the SET_SRC is in memory. */
4131 char src_in_memory;
4132 /* Nonzero if the SET_SRC contains something
4133 whose value cannot be predicted and understood. */
4134 char src_volatile;
4135 /* Original machine mode, in case it becomes a CONST_INT.
4136 The size of this field should match the size of the mode
4137 field of struct rtx_def (see rtl.h). */
4138 ENUM_BITFIELD(machine_mode) mode : 8;
4139 /* A constant equivalent for SET_SRC, if any. */
4140 rtx src_const;
4141 /* Hash value of constant equivalent for SET_SRC. */
4142 unsigned src_const_hash;
4143 /* Table entry for constant equivalent for SET_SRC, if any. */
4144 struct table_elt *src_const_elt;
4145 /* Table entry for the destination address. */
4146 struct table_elt *dest_addr_elt;
4149 /* Special handling for (set REG0 REG1) where REG0 is the
4150 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4151 be used in the sequel, so (if easily done) change this insn to
4152 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4153 that computed their value. Then REG1 will become a dead store
4154 and won't cloud the situation for later optimizations.
4156 Do not make this change if REG1 is a hard register, because it will
4157 then be used in the sequel and we may be changing a two-operand insn
4158 into a three-operand insn.
4160 This is the last transformation that cse_insn will try to do. */
4162 static void
4163 try_back_substitute_reg (rtx set, rtx insn)
4165 rtx dest = SET_DEST (set);
4166 rtx src = SET_SRC (set);
4168 if (REG_P (dest)
4169 && REG_P (src) && ! HARD_REGISTER_P (src)
4170 && REGNO_QTY_VALID_P (REGNO (src)))
4172 int src_q = REG_QTY (REGNO (src));
4173 struct qty_table_elem *src_ent = &qty_table[src_q];
4175 if (src_ent->first_reg == REGNO (dest))
4177 /* Scan for the previous nonnote insn, but stop at a basic
4178 block boundary. */
4179 rtx prev = insn;
4180 rtx bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4183 prev = PREV_INSN (prev);
4185 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4187 /* Do not swap the registers around if the previous instruction
4188 attaches a REG_EQUIV note to REG1.
4190 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4191 from the pseudo that originally shadowed an incoming argument
4192 to another register. Some uses of REG_EQUIV might rely on it
4193 being attached to REG1 rather than REG2.
4195 This section previously turned the REG_EQUIV into a REG_EQUAL
4196 note. We cannot do that because REG_EQUIV may provide an
4197 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4198 if (NONJUMP_INSN_P (prev)
4199 && GET_CODE (PATTERN (prev)) == SET
4200 && SET_DEST (PATTERN (prev)) == src
4201 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4203 rtx note;
4205 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4206 validate_change (insn, &SET_DEST (set), src, 1);
4207 validate_change (insn, &SET_SRC (set), dest, 1);
4208 apply_change_group ();
4210 /* If INSN has a REG_EQUAL note, and this note mentions
4211 REG0, then we must delete it, because the value in
4212 REG0 has changed. If the note's value is REG1, we must
4213 also delete it because that is now this insn's dest. */
4214 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4215 if (note != 0
4216 && (reg_mentioned_p (dest, XEXP (note, 0))
4217 || rtx_equal_p (src, XEXP (note, 0))))
4218 remove_note (insn, note);
4224 /* Record all the SETs in this instruction into SETS_PTR,
4225 and return the number of recorded sets. */
4226 static int
4227 find_sets_in_insn (rtx insn, struct set **psets)
4229 struct set *sets = *psets;
4230 int n_sets = 0;
4231 rtx x = PATTERN (insn);
4233 if (GET_CODE (x) == SET)
4235 /* Ignore SETs that are unconditional jumps.
4236 They never need cse processing, so this does not hurt.
4237 The reason is not efficiency but rather
4238 so that we can test at the end for instructions
4239 that have been simplified to unconditional jumps
4240 and not be misled by unchanged instructions
4241 that were unconditional jumps to begin with. */
4242 if (SET_DEST (x) == pc_rtx
4243 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4245 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4246 The hard function value register is used only once, to copy to
4247 someplace else, so it isn't worth cse'ing. */
4248 else if (GET_CODE (SET_SRC (x)) == CALL)
4250 else
4251 sets[n_sets++].rtl = x;
4253 else if (GET_CODE (x) == PARALLEL)
4255 int i, lim = XVECLEN (x, 0);
4257 /* Go over the epressions of the PARALLEL in forward order, to
4258 put them in the same order in the SETS array. */
4259 for (i = 0; i < lim; i++)
4261 rtx y = XVECEXP (x, 0, i);
4262 if (GET_CODE (y) == SET)
4264 /* As above, we ignore unconditional jumps and call-insns and
4265 ignore the result of apply_change_group. */
4266 if (SET_DEST (y) == pc_rtx
4267 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4269 else if (GET_CODE (SET_SRC (y)) == CALL)
4271 else
4272 sets[n_sets++].rtl = y;
4277 return n_sets;
4280 /* Where possible, substitute every register reference in the N_SETS
4281 number of SETS in INSN with the the canonical register.
4283 Register canonicalization propagatest the earliest register (i.e.
4284 one that is set before INSN) with the same value. This is a very
4285 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4286 to RTL. For instance, a CONST for an address is usually expanded
4287 multiple times to loads into different registers, thus creating many
4288 subexpressions of the form:
4290 (set (reg1) (some_const))
4291 (set (mem (... reg1 ...) (thing)))
4292 (set (reg2) (some_const))
4293 (set (mem (... reg2 ...) (thing)))
4295 After canonicalizing, the code takes the following form:
4297 (set (reg1) (some_const))
4298 (set (mem (... reg1 ...) (thing)))
4299 (set (reg2) (some_const))
4300 (set (mem (... reg1 ...) (thing)))
4302 The set to reg2 is now trivially dead, and the memory reference (or
4303 address, or whatever) may be a candidate for further CSEing.
4305 In this function, the result of apply_change_group can be ignored;
4306 see canon_reg. */
4308 static void
4309 canonicalize_insn (rtx insn, struct set **psets, int n_sets)
4311 struct set *sets = *psets;
4312 rtx tem;
4313 rtx x = PATTERN (insn);
4314 int i;
4316 if (CALL_P (insn))
4318 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4319 if (GET_CODE (XEXP (tem, 0)) != SET)
4320 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4323 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4325 canon_reg (SET_SRC (x), insn);
4326 apply_change_group ();
4327 fold_rtx (SET_SRC (x), insn);
4329 else if (GET_CODE (x) == CLOBBER)
4331 /* If we clobber memory, canon the address.
4332 This does nothing when a register is clobbered
4333 because we have already invalidated the reg. */
4334 if (MEM_P (XEXP (x, 0)))
4335 canon_reg (XEXP (x, 0), insn);
4337 else if (GET_CODE (x) == USE
4338 && ! (REG_P (XEXP (x, 0))
4339 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4340 /* Canonicalize a USE of a pseudo register or memory location. */
4341 canon_reg (x, insn);
4342 else if (GET_CODE (x) == ASM_OPERANDS)
4344 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4346 rtx input = ASM_OPERANDS_INPUT (x, i);
4347 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4349 input = canon_reg (input, insn);
4350 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4354 else if (GET_CODE (x) == CALL)
4356 canon_reg (x, insn);
4357 apply_change_group ();
4358 fold_rtx (x, insn);
4360 else if (DEBUG_INSN_P (insn))
4361 canon_reg (PATTERN (insn), insn);
4362 else if (GET_CODE (x) == PARALLEL)
4364 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4366 rtx y = XVECEXP (x, 0, i);
4367 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4369 canon_reg (SET_SRC (y), insn);
4370 apply_change_group ();
4371 fold_rtx (SET_SRC (y), insn);
4373 else if (GET_CODE (y) == CLOBBER)
4375 if (MEM_P (XEXP (y, 0)))
4376 canon_reg (XEXP (y, 0), insn);
4378 else if (GET_CODE (y) == USE
4379 && ! (REG_P (XEXP (y, 0))
4380 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4381 canon_reg (y, insn);
4382 else if (GET_CODE (y) == CALL)
4384 canon_reg (y, insn);
4385 apply_change_group ();
4386 fold_rtx (y, insn);
4391 if (n_sets == 1 && REG_NOTES (insn) != 0
4392 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4394 /* We potentially will process this insn many times. Therefore,
4395 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4396 unique set in INSN.
4398 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4399 because cse_insn handles those specially. */
4400 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4401 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4402 remove_note (insn, tem);
4403 else
4405 canon_reg (XEXP (tem, 0), insn);
4406 apply_change_group ();
4407 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4408 df_notes_rescan (insn);
4412 /* Canonicalize sources and addresses of destinations.
4413 We do this in a separate pass to avoid problems when a MATCH_DUP is
4414 present in the insn pattern. In that case, we want to ensure that
4415 we don't break the duplicate nature of the pattern. So we will replace
4416 both operands at the same time. Otherwise, we would fail to find an
4417 equivalent substitution in the loop calling validate_change below.
4419 We used to suppress canonicalization of DEST if it appears in SRC,
4420 but we don't do this any more. */
4422 for (i = 0; i < n_sets; i++)
4424 rtx dest = SET_DEST (sets[i].rtl);
4425 rtx src = SET_SRC (sets[i].rtl);
4426 rtx new_rtx = canon_reg (src, insn);
4428 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4430 if (GET_CODE (dest) == ZERO_EXTRACT)
4432 validate_change (insn, &XEXP (dest, 1),
4433 canon_reg (XEXP (dest, 1), insn), 1);
4434 validate_change (insn, &XEXP (dest, 2),
4435 canon_reg (XEXP (dest, 2), insn), 1);
4438 while (GET_CODE (dest) == SUBREG
4439 || GET_CODE (dest) == ZERO_EXTRACT
4440 || GET_CODE (dest) == STRICT_LOW_PART)
4441 dest = XEXP (dest, 0);
4443 if (MEM_P (dest))
4444 canon_reg (dest, insn);
4447 /* Now that we have done all the replacements, we can apply the change
4448 group and see if they all work. Note that this will cause some
4449 canonicalizations that would have worked individually not to be applied
4450 because some other canonicalization didn't work, but this should not
4451 occur often.
4453 The result of apply_change_group can be ignored; see canon_reg. */
4455 apply_change_group ();
4458 /* Main function of CSE.
4459 First simplify sources and addresses of all assignments
4460 in the instruction, using previously-computed equivalents values.
4461 Then install the new sources and destinations in the table
4462 of available values. */
4464 static void
4465 cse_insn (rtx insn)
4467 rtx x = PATTERN (insn);
4468 int i;
4469 rtx tem;
4470 int n_sets = 0;
4472 rtx src_eqv = 0;
4473 struct table_elt *src_eqv_elt = 0;
4474 int src_eqv_volatile = 0;
4475 int src_eqv_in_memory = 0;
4476 unsigned src_eqv_hash = 0;
4478 struct set *sets = (struct set *) 0;
4480 if (GET_CODE (x) == SET)
4481 sets = XALLOCA (struct set);
4482 else if (GET_CODE (x) == PARALLEL)
4483 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4485 this_insn = insn;
4486 #ifdef HAVE_cc0
4487 /* Records what this insn does to set CC0. */
4488 this_insn_cc0 = 0;
4489 this_insn_cc0_mode = VOIDmode;
4490 #endif
4492 /* Find all regs explicitly clobbered in this insn,
4493 to ensure they are not replaced with any other regs
4494 elsewhere in this insn. */
4495 invalidate_from_sets_and_clobbers (insn);
4497 /* Record all the SETs in this instruction. */
4498 n_sets = find_sets_in_insn (insn, &sets);
4500 /* Substitute the canonical register where possible. */
4501 canonicalize_insn (insn, &sets, n_sets);
4503 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4504 if different, or if the DEST is a STRICT_LOW_PART. The latter condition
4505 is necessary because SRC_EQV is handled specially for this case, and if
4506 it isn't set, then there will be no equivalence for the destination. */
4507 if (n_sets == 1 && REG_NOTES (insn) != 0
4508 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4509 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4510 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4511 src_eqv = copy_rtx (XEXP (tem, 0));
4513 /* Set sets[i].src_elt to the class each source belongs to.
4514 Detect assignments from or to volatile things
4515 and set set[i] to zero so they will be ignored
4516 in the rest of this function.
4518 Nothing in this loop changes the hash table or the register chains. */
4520 for (i = 0; i < n_sets; i++)
4522 bool repeat = false;
4523 rtx src, dest;
4524 rtx src_folded;
4525 struct table_elt *elt = 0, *p;
4526 enum machine_mode mode;
4527 rtx src_eqv_here;
4528 rtx src_const = 0;
4529 rtx src_related = 0;
4530 bool src_related_is_const_anchor = false;
4531 struct table_elt *src_const_elt = 0;
4532 int src_cost = MAX_COST;
4533 int src_eqv_cost = MAX_COST;
4534 int src_folded_cost = MAX_COST;
4535 int src_related_cost = MAX_COST;
4536 int src_elt_cost = MAX_COST;
4537 int src_regcost = MAX_COST;
4538 int src_eqv_regcost = MAX_COST;
4539 int src_folded_regcost = MAX_COST;
4540 int src_related_regcost = MAX_COST;
4541 int src_elt_regcost = MAX_COST;
4542 /* Set nonzero if we need to call force_const_mem on with the
4543 contents of src_folded before using it. */
4544 int src_folded_force_flag = 0;
4546 dest = SET_DEST (sets[i].rtl);
4547 src = SET_SRC (sets[i].rtl);
4549 /* If SRC is a constant that has no machine mode,
4550 hash it with the destination's machine mode.
4551 This way we can keep different modes separate. */
4553 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4554 sets[i].mode = mode;
4556 if (src_eqv)
4558 enum machine_mode eqvmode = mode;
4559 if (GET_CODE (dest) == STRICT_LOW_PART)
4560 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4561 do_not_record = 0;
4562 hash_arg_in_memory = 0;
4563 src_eqv_hash = HASH (src_eqv, eqvmode);
4565 /* Find the equivalence class for the equivalent expression. */
4567 if (!do_not_record)
4568 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4570 src_eqv_volatile = do_not_record;
4571 src_eqv_in_memory = hash_arg_in_memory;
4574 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4575 value of the INNER register, not the destination. So it is not
4576 a valid substitution for the source. But save it for later. */
4577 if (GET_CODE (dest) == STRICT_LOW_PART)
4578 src_eqv_here = 0;
4579 else
4580 src_eqv_here = src_eqv;
4582 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4583 simplified result, which may not necessarily be valid. */
4584 src_folded = fold_rtx (src, insn);
4586 #if 0
4587 /* ??? This caused bad code to be generated for the m68k port with -O2.
4588 Suppose src is (CONST_INT -1), and that after truncation src_folded
4589 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4590 At the end we will add src and src_const to the same equivalence
4591 class. We now have 3 and -1 on the same equivalence class. This
4592 causes later instructions to be mis-optimized. */
4593 /* If storing a constant in a bitfield, pre-truncate the constant
4594 so we will be able to record it later. */
4595 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4597 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4599 if (CONST_INT_P (src)
4600 && CONST_INT_P (width)
4601 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4602 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4603 src_folded
4604 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4605 << INTVAL (width)) - 1));
4607 #endif
4609 /* Compute SRC's hash code, and also notice if it
4610 should not be recorded at all. In that case,
4611 prevent any further processing of this assignment. */
4612 do_not_record = 0;
4613 hash_arg_in_memory = 0;
4615 sets[i].src = src;
4616 sets[i].src_hash = HASH (src, mode);
4617 sets[i].src_volatile = do_not_record;
4618 sets[i].src_in_memory = hash_arg_in_memory;
4620 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4621 a pseudo, do not record SRC. Using SRC as a replacement for
4622 anything else will be incorrect in that situation. Note that
4623 this usually occurs only for stack slots, in which case all the
4624 RTL would be referring to SRC, so we don't lose any optimization
4625 opportunities by not having SRC in the hash table. */
4627 if (MEM_P (src)
4628 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4629 && REG_P (dest)
4630 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4631 sets[i].src_volatile = 1;
4633 #if 0
4634 /* It is no longer clear why we used to do this, but it doesn't
4635 appear to still be needed. So let's try without it since this
4636 code hurts cse'ing widened ops. */
4637 /* If source is a paradoxical subreg (such as QI treated as an SI),
4638 treat it as volatile. It may do the work of an SI in one context
4639 where the extra bits are not being used, but cannot replace an SI
4640 in general. */
4641 if (paradoxical_subreg_p (src))
4642 sets[i].src_volatile = 1;
4643 #endif
4645 /* Locate all possible equivalent forms for SRC. Try to replace
4646 SRC in the insn with each cheaper equivalent.
4648 We have the following types of equivalents: SRC itself, a folded
4649 version, a value given in a REG_EQUAL note, or a value related
4650 to a constant.
4652 Each of these equivalents may be part of an additional class
4653 of equivalents (if more than one is in the table, they must be in
4654 the same class; we check for this).
4656 If the source is volatile, we don't do any table lookups.
4658 We note any constant equivalent for possible later use in a
4659 REG_NOTE. */
4661 if (!sets[i].src_volatile)
4662 elt = lookup (src, sets[i].src_hash, mode);
4664 sets[i].src_elt = elt;
4666 if (elt && src_eqv_here && src_eqv_elt)
4668 if (elt->first_same_value != src_eqv_elt->first_same_value)
4670 /* The REG_EQUAL is indicating that two formerly distinct
4671 classes are now equivalent. So merge them. */
4672 merge_equiv_classes (elt, src_eqv_elt);
4673 src_eqv_hash = HASH (src_eqv, elt->mode);
4674 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4677 src_eqv_here = 0;
4680 else if (src_eqv_elt)
4681 elt = src_eqv_elt;
4683 /* Try to find a constant somewhere and record it in `src_const'.
4684 Record its table element, if any, in `src_const_elt'. Look in
4685 any known equivalences first. (If the constant is not in the
4686 table, also set `sets[i].src_const_hash'). */
4687 if (elt)
4688 for (p = elt->first_same_value; p; p = p->next_same_value)
4689 if (p->is_const)
4691 src_const = p->exp;
4692 src_const_elt = elt;
4693 break;
4696 if (src_const == 0
4697 && (CONSTANT_P (src_folded)
4698 /* Consider (minus (label_ref L1) (label_ref L2)) as
4699 "constant" here so we will record it. This allows us
4700 to fold switch statements when an ADDR_DIFF_VEC is used. */
4701 || (GET_CODE (src_folded) == MINUS
4702 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4703 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4704 src_const = src_folded, src_const_elt = elt;
4705 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4706 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4708 /* If we don't know if the constant is in the table, get its
4709 hash code and look it up. */
4710 if (src_const && src_const_elt == 0)
4712 sets[i].src_const_hash = HASH (src_const, mode);
4713 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4716 sets[i].src_const = src_const;
4717 sets[i].src_const_elt = src_const_elt;
4719 /* If the constant and our source are both in the table, mark them as
4720 equivalent. Otherwise, if a constant is in the table but the source
4721 isn't, set ELT to it. */
4722 if (src_const_elt && elt
4723 && src_const_elt->first_same_value != elt->first_same_value)
4724 merge_equiv_classes (elt, src_const_elt);
4725 else if (src_const_elt && elt == 0)
4726 elt = src_const_elt;
4728 /* See if there is a register linearly related to a constant
4729 equivalent of SRC. */
4730 if (src_const
4731 && (GET_CODE (src_const) == CONST
4732 || (src_const_elt && src_const_elt->related_value != 0)))
4734 src_related = use_related_value (src_const, src_const_elt);
4735 if (src_related)
4737 struct table_elt *src_related_elt
4738 = lookup (src_related, HASH (src_related, mode), mode);
4739 if (src_related_elt && elt)
4741 if (elt->first_same_value
4742 != src_related_elt->first_same_value)
4743 /* This can occur when we previously saw a CONST
4744 involving a SYMBOL_REF and then see the SYMBOL_REF
4745 twice. Merge the involved classes. */
4746 merge_equiv_classes (elt, src_related_elt);
4748 src_related = 0;
4749 src_related_elt = 0;
4751 else if (src_related_elt && elt == 0)
4752 elt = src_related_elt;
4756 /* See if we have a CONST_INT that is already in a register in a
4757 wider mode. */
4759 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4760 && GET_MODE_CLASS (mode) == MODE_INT
4761 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4763 enum machine_mode wider_mode;
4765 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4766 wider_mode != VOIDmode
4767 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4768 && src_related == 0;
4769 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4771 struct table_elt *const_elt
4772 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4774 if (const_elt == 0)
4775 continue;
4777 for (const_elt = const_elt->first_same_value;
4778 const_elt; const_elt = const_elt->next_same_value)
4779 if (REG_P (const_elt->exp))
4781 src_related = gen_lowpart (mode, const_elt->exp);
4782 break;
4787 /* Another possibility is that we have an AND with a constant in
4788 a mode narrower than a word. If so, it might have been generated
4789 as part of an "if" which would narrow the AND. If we already
4790 have done the AND in a wider mode, we can use a SUBREG of that
4791 value. */
4793 if (flag_expensive_optimizations && ! src_related
4794 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4795 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4797 enum machine_mode tmode;
4798 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4800 for (tmode = GET_MODE_WIDER_MODE (mode);
4801 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4802 tmode = GET_MODE_WIDER_MODE (tmode))
4804 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4805 struct table_elt *larger_elt;
4807 if (inner)
4809 PUT_MODE (new_and, tmode);
4810 XEXP (new_and, 0) = inner;
4811 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4812 if (larger_elt == 0)
4813 continue;
4815 for (larger_elt = larger_elt->first_same_value;
4816 larger_elt; larger_elt = larger_elt->next_same_value)
4817 if (REG_P (larger_elt->exp))
4819 src_related
4820 = gen_lowpart (mode, larger_elt->exp);
4821 break;
4824 if (src_related)
4825 break;
4830 #ifdef LOAD_EXTEND_OP
4831 /* See if a MEM has already been loaded with a widening operation;
4832 if it has, we can use a subreg of that. Many CISC machines
4833 also have such operations, but this is only likely to be
4834 beneficial on these machines. */
4836 if (flag_expensive_optimizations && src_related == 0
4837 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4838 && GET_MODE_CLASS (mode) == MODE_INT
4839 && MEM_P (src) && ! do_not_record
4840 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4842 struct rtx_def memory_extend_buf;
4843 rtx memory_extend_rtx = &memory_extend_buf;
4844 enum machine_mode tmode;
4846 /* Set what we are trying to extend and the operation it might
4847 have been extended with. */
4848 memset (memory_extend_rtx, 0, sizeof(*memory_extend_rtx));
4849 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4850 XEXP (memory_extend_rtx, 0) = src;
4852 for (tmode = GET_MODE_WIDER_MODE (mode);
4853 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4854 tmode = GET_MODE_WIDER_MODE (tmode))
4856 struct table_elt *larger_elt;
4858 PUT_MODE (memory_extend_rtx, tmode);
4859 larger_elt = lookup (memory_extend_rtx,
4860 HASH (memory_extend_rtx, tmode), tmode);
4861 if (larger_elt == 0)
4862 continue;
4864 for (larger_elt = larger_elt->first_same_value;
4865 larger_elt; larger_elt = larger_elt->next_same_value)
4866 if (REG_P (larger_elt->exp))
4868 src_related = gen_lowpart (mode, larger_elt->exp);
4869 break;
4872 if (src_related)
4873 break;
4876 #endif /* LOAD_EXTEND_OP */
4878 /* Try to express the constant using a register+offset expression
4879 derived from a constant anchor. */
4881 if (targetm.const_anchor
4882 && !src_related
4883 && src_const
4884 && GET_CODE (src_const) == CONST_INT)
4886 src_related = try_const_anchors (src_const, mode);
4887 src_related_is_const_anchor = src_related != NULL_RTX;
4891 if (src == src_folded)
4892 src_folded = 0;
4894 /* At this point, ELT, if nonzero, points to a class of expressions
4895 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4896 and SRC_RELATED, if nonzero, each contain additional equivalent
4897 expressions. Prune these latter expressions by deleting expressions
4898 already in the equivalence class.
4900 Check for an equivalent identical to the destination. If found,
4901 this is the preferred equivalent since it will likely lead to
4902 elimination of the insn. Indicate this by placing it in
4903 `src_related'. */
4905 if (elt)
4906 elt = elt->first_same_value;
4907 for (p = elt; p; p = p->next_same_value)
4909 enum rtx_code code = GET_CODE (p->exp);
4911 /* If the expression is not valid, ignore it. Then we do not
4912 have to check for validity below. In most cases, we can use
4913 `rtx_equal_p', since canonicalization has already been done. */
4914 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4915 continue;
4917 /* Also skip paradoxical subregs, unless that's what we're
4918 looking for. */
4919 if (paradoxical_subreg_p (p->exp)
4920 && ! (src != 0
4921 && GET_CODE (src) == SUBREG
4922 && GET_MODE (src) == GET_MODE (p->exp)
4923 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4924 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4925 continue;
4927 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4928 src = 0;
4929 else if (src_folded && GET_CODE (src_folded) == code
4930 && rtx_equal_p (src_folded, p->exp))
4931 src_folded = 0;
4932 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4933 && rtx_equal_p (src_eqv_here, p->exp))
4934 src_eqv_here = 0;
4935 else if (src_related && GET_CODE (src_related) == code
4936 && rtx_equal_p (src_related, p->exp))
4937 src_related = 0;
4939 /* This is the same as the destination of the insns, we want
4940 to prefer it. Copy it to src_related. The code below will
4941 then give it a negative cost. */
4942 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4943 src_related = dest;
4946 /* Find the cheapest valid equivalent, trying all the available
4947 possibilities. Prefer items not in the hash table to ones
4948 that are when they are equal cost. Note that we can never
4949 worsen an insn as the current contents will also succeed.
4950 If we find an equivalent identical to the destination, use it as best,
4951 since this insn will probably be eliminated in that case. */
4952 if (src)
4954 if (rtx_equal_p (src, dest))
4955 src_cost = src_regcost = -1;
4956 else
4958 src_cost = COST (src);
4959 src_regcost = approx_reg_cost (src);
4963 if (src_eqv_here)
4965 if (rtx_equal_p (src_eqv_here, dest))
4966 src_eqv_cost = src_eqv_regcost = -1;
4967 else
4969 src_eqv_cost = COST (src_eqv_here);
4970 src_eqv_regcost = approx_reg_cost (src_eqv_here);
4974 if (src_folded)
4976 if (rtx_equal_p (src_folded, dest))
4977 src_folded_cost = src_folded_regcost = -1;
4978 else
4980 src_folded_cost = COST (src_folded);
4981 src_folded_regcost = approx_reg_cost (src_folded);
4985 if (src_related)
4987 if (rtx_equal_p (src_related, dest))
4988 src_related_cost = src_related_regcost = -1;
4989 else
4991 src_related_cost = COST (src_related);
4992 src_related_regcost = approx_reg_cost (src_related);
4994 /* If a const-anchor is used to synthesize a constant that
4995 normally requires multiple instructions then slightly prefer
4996 it over the original sequence. These instructions are likely
4997 to become redundant now. We can't compare against the cost
4998 of src_eqv_here because, on MIPS for example, multi-insn
4999 constants have zero cost; they are assumed to be hoisted from
5000 loops. */
5001 if (src_related_is_const_anchor
5002 && src_related_cost == src_cost
5003 && src_eqv_here)
5004 src_related_cost--;
5008 /* If this was an indirect jump insn, a known label will really be
5009 cheaper even though it looks more expensive. */
5010 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5011 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5013 /* Terminate loop when replacement made. This must terminate since
5014 the current contents will be tested and will always be valid. */
5015 while (1)
5017 rtx trial;
5019 /* Skip invalid entries. */
5020 while (elt && !REG_P (elt->exp)
5021 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5022 elt = elt->next_same_value;
5024 /* A paradoxical subreg would be bad here: it'll be the right
5025 size, but later may be adjusted so that the upper bits aren't
5026 what we want. So reject it. */
5027 if (elt != 0
5028 && paradoxical_subreg_p (elt->exp)
5029 /* It is okay, though, if the rtx we're trying to match
5030 will ignore any of the bits we can't predict. */
5031 && ! (src != 0
5032 && GET_CODE (src) == SUBREG
5033 && GET_MODE (src) == GET_MODE (elt->exp)
5034 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5035 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5037 elt = elt->next_same_value;
5038 continue;
5041 if (elt)
5043 src_elt_cost = elt->cost;
5044 src_elt_regcost = elt->regcost;
5047 /* Find cheapest and skip it for the next time. For items
5048 of equal cost, use this order:
5049 src_folded, src, src_eqv, src_related and hash table entry. */
5050 if (src_folded
5051 && preferable (src_folded_cost, src_folded_regcost,
5052 src_cost, src_regcost) <= 0
5053 && preferable (src_folded_cost, src_folded_regcost,
5054 src_eqv_cost, src_eqv_regcost) <= 0
5055 && preferable (src_folded_cost, src_folded_regcost,
5056 src_related_cost, src_related_regcost) <= 0
5057 && preferable (src_folded_cost, src_folded_regcost,
5058 src_elt_cost, src_elt_regcost) <= 0)
5060 trial = src_folded, src_folded_cost = MAX_COST;
5061 if (src_folded_force_flag)
5063 rtx forced = force_const_mem (mode, trial);
5064 if (forced)
5065 trial = forced;
5068 else if (src
5069 && preferable (src_cost, src_regcost,
5070 src_eqv_cost, src_eqv_regcost) <= 0
5071 && preferable (src_cost, src_regcost,
5072 src_related_cost, src_related_regcost) <= 0
5073 && preferable (src_cost, src_regcost,
5074 src_elt_cost, src_elt_regcost) <= 0)
5075 trial = src, src_cost = MAX_COST;
5076 else if (src_eqv_here
5077 && preferable (src_eqv_cost, src_eqv_regcost,
5078 src_related_cost, src_related_regcost) <= 0
5079 && preferable (src_eqv_cost, src_eqv_regcost,
5080 src_elt_cost, src_elt_regcost) <= 0)
5081 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5082 else if (src_related
5083 && preferable (src_related_cost, src_related_regcost,
5084 src_elt_cost, src_elt_regcost) <= 0)
5085 trial = src_related, src_related_cost = MAX_COST;
5086 else
5088 trial = elt->exp;
5089 elt = elt->next_same_value;
5090 src_elt_cost = MAX_COST;
5093 /* Avoid creation of overlapping memory moves. */
5094 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5096 rtx src, dest;
5098 /* BLKmode moves are not handled by cse anyway. */
5099 if (GET_MODE (trial) == BLKmode)
5100 break;
5102 src = canon_rtx (trial);
5103 dest = canon_rtx (SET_DEST (sets[i].rtl));
5105 if (!MEM_P (src) || !MEM_P (dest)
5106 || !nonoverlapping_memrefs_p (src, dest, false))
5107 break;
5110 /* Try to optimize
5111 (set (reg:M N) (const_int A))
5112 (set (reg:M2 O) (const_int B))
5113 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5114 (reg:M2 O)). */
5115 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5116 && CONST_INT_P (trial)
5117 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5118 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5119 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5120 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5121 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5122 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5123 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5124 <= HOST_BITS_PER_WIDE_INT))
5126 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5127 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5128 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5129 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5130 struct table_elt *dest_elt
5131 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5132 rtx dest_cst = NULL;
5134 if (dest_elt)
5135 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5136 if (p->is_const && CONST_INT_P (p->exp))
5138 dest_cst = p->exp;
5139 break;
5141 if (dest_cst)
5143 HOST_WIDE_INT val = INTVAL (dest_cst);
5144 HOST_WIDE_INT mask;
5145 unsigned int shift;
5146 if (BITS_BIG_ENDIAN)
5147 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5148 - INTVAL (pos) - INTVAL (width);
5149 else
5150 shift = INTVAL (pos);
5151 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5152 mask = ~(HOST_WIDE_INT) 0;
5153 else
5154 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5155 val &= ~(mask << shift);
5156 val |= (INTVAL (trial) & mask) << shift;
5157 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5158 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5159 dest_reg, 1);
5160 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5161 GEN_INT (val), 1);
5162 if (apply_change_group ())
5164 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5165 if (note)
5167 remove_note (insn, note);
5168 df_notes_rescan (insn);
5170 src_eqv = NULL_RTX;
5171 src_eqv_elt = NULL;
5172 src_eqv_volatile = 0;
5173 src_eqv_in_memory = 0;
5174 src_eqv_hash = 0;
5175 repeat = true;
5176 break;
5181 /* We don't normally have an insn matching (set (pc) (pc)), so
5182 check for this separately here. We will delete such an
5183 insn below.
5185 For other cases such as a table jump or conditional jump
5186 where we know the ultimate target, go ahead and replace the
5187 operand. While that may not make a valid insn, we will
5188 reemit the jump below (and also insert any necessary
5189 barriers). */
5190 if (n_sets == 1 && dest == pc_rtx
5191 && (trial == pc_rtx
5192 || (GET_CODE (trial) == LABEL_REF
5193 && ! condjump_p (insn))))
5195 /* Don't substitute non-local labels, this confuses CFG. */
5196 if (GET_CODE (trial) == LABEL_REF
5197 && LABEL_REF_NONLOCAL_P (trial))
5198 continue;
5200 SET_SRC (sets[i].rtl) = trial;
5201 cse_jumps_altered = true;
5202 break;
5205 /* Reject certain invalid forms of CONST that we create. */
5206 else if (CONSTANT_P (trial)
5207 && GET_CODE (trial) == CONST
5208 /* Reject cases that will cause decode_rtx_const to
5209 die. On the alpha when simplifying a switch, we
5210 get (const (truncate (minus (label_ref)
5211 (label_ref)))). */
5212 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5213 /* Likewise on IA-64, except without the
5214 truncate. */
5215 || (GET_CODE (XEXP (trial, 0)) == MINUS
5216 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5217 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5218 /* Do nothing for this case. */
5221 /* Look for a substitution that makes a valid insn. */
5222 else if (validate_unshare_change
5223 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5225 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5227 /* The result of apply_change_group can be ignored; see
5228 canon_reg. */
5230 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5231 apply_change_group ();
5233 break;
5236 /* If we previously found constant pool entries for
5237 constants and this is a constant, try making a
5238 pool entry. Put it in src_folded unless we already have done
5239 this since that is where it likely came from. */
5241 else if (constant_pool_entries_cost
5242 && CONSTANT_P (trial)
5243 && (src_folded == 0
5244 || (!MEM_P (src_folded)
5245 && ! src_folded_force_flag))
5246 && GET_MODE_CLASS (mode) != MODE_CC
5247 && mode != VOIDmode)
5249 src_folded_force_flag = 1;
5250 src_folded = trial;
5251 src_folded_cost = constant_pool_entries_cost;
5252 src_folded_regcost = constant_pool_entries_regcost;
5256 /* If we changed the insn too much, handle this set from scratch. */
5257 if (repeat)
5259 i--;
5260 continue;
5263 src = SET_SRC (sets[i].rtl);
5265 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5266 However, there is an important exception: If both are registers
5267 that are not the head of their equivalence class, replace SET_SRC
5268 with the head of the class. If we do not do this, we will have
5269 both registers live over a portion of the basic block. This way,
5270 their lifetimes will likely abut instead of overlapping. */
5271 if (REG_P (dest)
5272 && REGNO_QTY_VALID_P (REGNO (dest)))
5274 int dest_q = REG_QTY (REGNO (dest));
5275 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5277 if (dest_ent->mode == GET_MODE (dest)
5278 && dest_ent->first_reg != REGNO (dest)
5279 && REG_P (src) && REGNO (src) == REGNO (dest)
5280 /* Don't do this if the original insn had a hard reg as
5281 SET_SRC or SET_DEST. */
5282 && (!REG_P (sets[i].src)
5283 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5284 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5285 /* We can't call canon_reg here because it won't do anything if
5286 SRC is a hard register. */
5288 int src_q = REG_QTY (REGNO (src));
5289 struct qty_table_elem *src_ent = &qty_table[src_q];
5290 int first = src_ent->first_reg;
5291 rtx new_src
5292 = (first >= FIRST_PSEUDO_REGISTER
5293 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5295 /* We must use validate-change even for this, because this
5296 might be a special no-op instruction, suitable only to
5297 tag notes onto. */
5298 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5300 src = new_src;
5301 /* If we had a constant that is cheaper than what we are now
5302 setting SRC to, use that constant. We ignored it when we
5303 thought we could make this into a no-op. */
5304 if (src_const && COST (src_const) < COST (src)
5305 && validate_change (insn, &SET_SRC (sets[i].rtl),
5306 src_const, 0))
5307 src = src_const;
5312 /* If we made a change, recompute SRC values. */
5313 if (src != sets[i].src)
5315 do_not_record = 0;
5316 hash_arg_in_memory = 0;
5317 sets[i].src = src;
5318 sets[i].src_hash = HASH (src, mode);
5319 sets[i].src_volatile = do_not_record;
5320 sets[i].src_in_memory = hash_arg_in_memory;
5321 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5324 /* If this is a single SET, we are setting a register, and we have an
5325 equivalent constant, we want to add a REG_NOTE. We don't want
5326 to write a REG_EQUAL note for a constant pseudo since verifying that
5327 that pseudo hasn't been eliminated is a pain. Such a note also
5328 won't help anything.
5330 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5331 which can be created for a reference to a compile time computable
5332 entry in a jump table. */
5334 if (n_sets == 1 && src_const && REG_P (dest)
5335 && !REG_P (src_const)
5336 && ! (GET_CODE (src_const) == CONST
5337 && GET_CODE (XEXP (src_const, 0)) == MINUS
5338 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5339 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF))
5341 /* We only want a REG_EQUAL note if src_const != src. */
5342 if (! rtx_equal_p (src, src_const))
5344 /* Make sure that the rtx is not shared. */
5345 src_const = copy_rtx (src_const);
5347 /* Record the actual constant value in a REG_EQUAL note,
5348 making a new one if one does not already exist. */
5349 set_unique_reg_note (insn, REG_EQUAL, src_const);
5350 df_notes_rescan (insn);
5354 /* Now deal with the destination. */
5355 do_not_record = 0;
5357 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5358 while (GET_CODE (dest) == SUBREG
5359 || GET_CODE (dest) == ZERO_EXTRACT
5360 || GET_CODE (dest) == STRICT_LOW_PART)
5361 dest = XEXP (dest, 0);
5363 sets[i].inner_dest = dest;
5365 if (MEM_P (dest))
5367 #ifdef PUSH_ROUNDING
5368 /* Stack pushes invalidate the stack pointer. */
5369 rtx addr = XEXP (dest, 0);
5370 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5371 && XEXP (addr, 0) == stack_pointer_rtx)
5372 invalidate (stack_pointer_rtx, VOIDmode);
5373 #endif
5374 dest = fold_rtx (dest, insn);
5377 /* Compute the hash code of the destination now,
5378 before the effects of this instruction are recorded,
5379 since the register values used in the address computation
5380 are those before this instruction. */
5381 sets[i].dest_hash = HASH (dest, mode);
5383 /* Don't enter a bit-field in the hash table
5384 because the value in it after the store
5385 may not equal what was stored, due to truncation. */
5387 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5389 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5391 if (src_const != 0 && CONST_INT_P (src_const)
5392 && CONST_INT_P (width)
5393 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5394 && ! (INTVAL (src_const)
5395 & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
5396 /* Exception: if the value is constant,
5397 and it won't be truncated, record it. */
5399 else
5401 /* This is chosen so that the destination will be invalidated
5402 but no new value will be recorded.
5403 We must invalidate because sometimes constant
5404 values can be recorded for bitfields. */
5405 sets[i].src_elt = 0;
5406 sets[i].src_volatile = 1;
5407 src_eqv = 0;
5408 src_eqv_elt = 0;
5412 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5413 the insn. */
5414 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5416 /* One less use of the label this insn used to jump to. */
5417 delete_insn_and_edges (insn);
5418 cse_jumps_altered = true;
5419 /* No more processing for this set. */
5420 sets[i].rtl = 0;
5423 /* If this SET is now setting PC to a label, we know it used to
5424 be a conditional or computed branch. */
5425 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5426 && !LABEL_REF_NONLOCAL_P (src))
5428 /* We reemit the jump in as many cases as possible just in
5429 case the form of an unconditional jump is significantly
5430 different than a computed jump or conditional jump.
5432 If this insn has multiple sets, then reemitting the
5433 jump is nontrivial. So instead we just force rerecognition
5434 and hope for the best. */
5435 if (n_sets == 1)
5437 rtx new_rtx, note;
5439 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5440 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5441 LABEL_NUSES (XEXP (src, 0))++;
5443 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5444 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5445 if (note)
5447 XEXP (note, 1) = NULL_RTX;
5448 REG_NOTES (new_rtx) = note;
5451 delete_insn_and_edges (insn);
5452 insn = new_rtx;
5454 else
5455 INSN_CODE (insn) = -1;
5457 /* Do not bother deleting any unreachable code, let jump do it. */
5458 cse_jumps_altered = true;
5459 sets[i].rtl = 0;
5462 /* If destination is volatile, invalidate it and then do no further
5463 processing for this assignment. */
5465 else if (do_not_record)
5467 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5468 invalidate (dest, VOIDmode);
5469 else if (MEM_P (dest))
5470 invalidate (dest, VOIDmode);
5471 else if (GET_CODE (dest) == STRICT_LOW_PART
5472 || GET_CODE (dest) == ZERO_EXTRACT)
5473 invalidate (XEXP (dest, 0), GET_MODE (dest));
5474 sets[i].rtl = 0;
5477 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5478 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5480 #ifdef HAVE_cc0
5481 /* If setting CC0, record what it was set to, or a constant, if it
5482 is equivalent to a constant. If it is being set to a floating-point
5483 value, make a COMPARE with the appropriate constant of 0. If we
5484 don't do this, later code can interpret this as a test against
5485 const0_rtx, which can cause problems if we try to put it into an
5486 insn as a floating-point operand. */
5487 if (dest == cc0_rtx)
5489 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5490 this_insn_cc0_mode = mode;
5491 if (FLOAT_MODE_P (mode))
5492 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5493 CONST0_RTX (mode));
5495 #endif
5498 /* Now enter all non-volatile source expressions in the hash table
5499 if they are not already present.
5500 Record their equivalence classes in src_elt.
5501 This way we can insert the corresponding destinations into
5502 the same classes even if the actual sources are no longer in them
5503 (having been invalidated). */
5505 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5506 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5508 struct table_elt *elt;
5509 struct table_elt *classp = sets[0].src_elt;
5510 rtx dest = SET_DEST (sets[0].rtl);
5511 enum machine_mode eqvmode = GET_MODE (dest);
5513 if (GET_CODE (dest) == STRICT_LOW_PART)
5515 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5516 classp = 0;
5518 if (insert_regs (src_eqv, classp, 0))
5520 rehash_using_reg (src_eqv);
5521 src_eqv_hash = HASH (src_eqv, eqvmode);
5523 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5524 elt->in_memory = src_eqv_in_memory;
5525 src_eqv_elt = elt;
5527 /* Check to see if src_eqv_elt is the same as a set source which
5528 does not yet have an elt, and if so set the elt of the set source
5529 to src_eqv_elt. */
5530 for (i = 0; i < n_sets; i++)
5531 if (sets[i].rtl && sets[i].src_elt == 0
5532 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5533 sets[i].src_elt = src_eqv_elt;
5536 for (i = 0; i < n_sets; i++)
5537 if (sets[i].rtl && ! sets[i].src_volatile
5538 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5540 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5542 /* REG_EQUAL in setting a STRICT_LOW_PART
5543 gives an equivalent for the entire destination register,
5544 not just for the subreg being stored in now.
5545 This is a more interesting equivalence, so we arrange later
5546 to treat the entire reg as the destination. */
5547 sets[i].src_elt = src_eqv_elt;
5548 sets[i].src_hash = src_eqv_hash;
5550 else
5552 /* Insert source and constant equivalent into hash table, if not
5553 already present. */
5554 struct table_elt *classp = src_eqv_elt;
5555 rtx src = sets[i].src;
5556 rtx dest = SET_DEST (sets[i].rtl);
5557 enum machine_mode mode
5558 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5560 /* It's possible that we have a source value known to be
5561 constant but don't have a REG_EQUAL note on the insn.
5562 Lack of a note will mean src_eqv_elt will be NULL. This
5563 can happen where we've generated a SUBREG to access a
5564 CONST_INT that is already in a register in a wider mode.
5565 Ensure that the source expression is put in the proper
5566 constant class. */
5567 if (!classp)
5568 classp = sets[i].src_const_elt;
5570 if (sets[i].src_elt == 0)
5572 struct table_elt *elt;
5574 /* Note that these insert_regs calls cannot remove
5575 any of the src_elt's, because they would have failed to
5576 match if not still valid. */
5577 if (insert_regs (src, classp, 0))
5579 rehash_using_reg (src);
5580 sets[i].src_hash = HASH (src, mode);
5582 elt = insert (src, classp, sets[i].src_hash, mode);
5583 elt->in_memory = sets[i].src_in_memory;
5584 sets[i].src_elt = classp = elt;
5586 if (sets[i].src_const && sets[i].src_const_elt == 0
5587 && src != sets[i].src_const
5588 && ! rtx_equal_p (sets[i].src_const, src))
5589 sets[i].src_elt = insert (sets[i].src_const, classp,
5590 sets[i].src_const_hash, mode);
5593 else if (sets[i].src_elt == 0)
5594 /* If we did not insert the source into the hash table (e.g., it was
5595 volatile), note the equivalence class for the REG_EQUAL value, if any,
5596 so that the destination goes into that class. */
5597 sets[i].src_elt = src_eqv_elt;
5599 /* Record destination addresses in the hash table. This allows us to
5600 check if they are invalidated by other sets. */
5601 for (i = 0; i < n_sets; i++)
5603 if (sets[i].rtl)
5605 rtx x = sets[i].inner_dest;
5606 struct table_elt *elt;
5607 enum machine_mode mode;
5608 unsigned hash;
5610 if (MEM_P (x))
5612 x = XEXP (x, 0);
5613 mode = GET_MODE (x);
5614 hash = HASH (x, mode);
5615 elt = lookup (x, hash, mode);
5616 if (!elt)
5618 if (insert_regs (x, NULL, 0))
5620 rtx dest = SET_DEST (sets[i].rtl);
5622 rehash_using_reg (x);
5623 hash = HASH (x, mode);
5624 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5626 elt = insert (x, NULL, hash, mode);
5629 sets[i].dest_addr_elt = elt;
5631 else
5632 sets[i].dest_addr_elt = NULL;
5636 invalidate_from_clobbers (insn);
5638 /* Some registers are invalidated by subroutine calls. Memory is
5639 invalidated by non-constant calls. */
5641 if (CALL_P (insn))
5643 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5644 invalidate_memory ();
5645 invalidate_for_call ();
5648 /* Now invalidate everything set by this instruction.
5649 If a SUBREG or other funny destination is being set,
5650 sets[i].rtl is still nonzero, so here we invalidate the reg
5651 a part of which is being set. */
5653 for (i = 0; i < n_sets; i++)
5654 if (sets[i].rtl)
5656 /* We can't use the inner dest, because the mode associated with
5657 a ZERO_EXTRACT is significant. */
5658 rtx dest = SET_DEST (sets[i].rtl);
5660 /* Needed for registers to remove the register from its
5661 previous quantity's chain.
5662 Needed for memory if this is a nonvarying address, unless
5663 we have just done an invalidate_memory that covers even those. */
5664 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5665 invalidate (dest, VOIDmode);
5666 else if (MEM_P (dest))
5667 invalidate (dest, VOIDmode);
5668 else if (GET_CODE (dest) == STRICT_LOW_PART
5669 || GET_CODE (dest) == ZERO_EXTRACT)
5670 invalidate (XEXP (dest, 0), GET_MODE (dest));
5673 /* A volatile ASM invalidates everything. */
5674 if (NONJUMP_INSN_P (insn)
5675 && GET_CODE (PATTERN (insn)) == ASM_OPERANDS
5676 && MEM_VOLATILE_P (PATTERN (insn)))
5677 flush_hash_table ();
5679 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5680 the regs restored by the longjmp come from a later time
5681 than the setjmp. */
5682 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5684 flush_hash_table ();
5685 goto done;
5688 /* Make sure registers mentioned in destinations
5689 are safe for use in an expression to be inserted.
5690 This removes from the hash table
5691 any invalid entry that refers to one of these registers.
5693 We don't care about the return value from mention_regs because
5694 we are going to hash the SET_DEST values unconditionally. */
5696 for (i = 0; i < n_sets; i++)
5698 if (sets[i].rtl)
5700 rtx x = SET_DEST (sets[i].rtl);
5702 if (!REG_P (x))
5703 mention_regs (x);
5704 else
5706 /* We used to rely on all references to a register becoming
5707 inaccessible when a register changes to a new quantity,
5708 since that changes the hash code. However, that is not
5709 safe, since after HASH_SIZE new quantities we get a
5710 hash 'collision' of a register with its own invalid
5711 entries. And since SUBREGs have been changed not to
5712 change their hash code with the hash code of the register,
5713 it wouldn't work any longer at all. So we have to check
5714 for any invalid references lying around now.
5715 This code is similar to the REG case in mention_regs,
5716 but it knows that reg_tick has been incremented, and
5717 it leaves reg_in_table as -1 . */
5718 unsigned int regno = REGNO (x);
5719 unsigned int endregno = END_REGNO (x);
5720 unsigned int i;
5722 for (i = regno; i < endregno; i++)
5724 if (REG_IN_TABLE (i) >= 0)
5726 remove_invalid_refs (i);
5727 REG_IN_TABLE (i) = -1;
5734 /* We may have just removed some of the src_elt's from the hash table.
5735 So replace each one with the current head of the same class.
5736 Also check if destination addresses have been removed. */
5738 for (i = 0; i < n_sets; i++)
5739 if (sets[i].rtl)
5741 if (sets[i].dest_addr_elt
5742 && sets[i].dest_addr_elt->first_same_value == 0)
5744 /* The elt was removed, which means this destination is not
5745 valid after this instruction. */
5746 sets[i].rtl = NULL_RTX;
5748 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5749 /* If elt was removed, find current head of same class,
5750 or 0 if nothing remains of that class. */
5752 struct table_elt *elt = sets[i].src_elt;
5754 while (elt && elt->prev_same_value)
5755 elt = elt->prev_same_value;
5757 while (elt && elt->first_same_value == 0)
5758 elt = elt->next_same_value;
5759 sets[i].src_elt = elt ? elt->first_same_value : 0;
5763 /* Now insert the destinations into their equivalence classes. */
5765 for (i = 0; i < n_sets; i++)
5766 if (sets[i].rtl)
5768 rtx dest = SET_DEST (sets[i].rtl);
5769 struct table_elt *elt;
5771 /* Don't record value if we are not supposed to risk allocating
5772 floating-point values in registers that might be wider than
5773 memory. */
5774 if ((flag_float_store
5775 && MEM_P (dest)
5776 && FLOAT_MODE_P (GET_MODE (dest)))
5777 /* Don't record BLKmode values, because we don't know the
5778 size of it, and can't be sure that other BLKmode values
5779 have the same or smaller size. */
5780 || GET_MODE (dest) == BLKmode
5781 /* If we didn't put a REG_EQUAL value or a source into the hash
5782 table, there is no point is recording DEST. */
5783 || sets[i].src_elt == 0
5784 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5785 or SIGN_EXTEND, don't record DEST since it can cause
5786 some tracking to be wrong.
5788 ??? Think about this more later. */
5789 || (paradoxical_subreg_p (dest)
5790 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5791 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5792 continue;
5794 /* STRICT_LOW_PART isn't part of the value BEING set,
5795 and neither is the SUBREG inside it.
5796 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5797 if (GET_CODE (dest) == STRICT_LOW_PART)
5798 dest = SUBREG_REG (XEXP (dest, 0));
5800 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5801 /* Registers must also be inserted into chains for quantities. */
5802 if (insert_regs (dest, sets[i].src_elt, 1))
5804 /* If `insert_regs' changes something, the hash code must be
5805 recalculated. */
5806 rehash_using_reg (dest);
5807 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5810 elt = insert (dest, sets[i].src_elt,
5811 sets[i].dest_hash, GET_MODE (dest));
5813 /* If this is a constant, insert the constant anchors with the
5814 equivalent register-offset expressions using register DEST. */
5815 if (targetm.const_anchor
5816 && REG_P (dest)
5817 && SCALAR_INT_MODE_P (GET_MODE (dest))
5818 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5819 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5821 elt->in_memory = (MEM_P (sets[i].inner_dest)
5822 && !MEM_READONLY_P (sets[i].inner_dest));
5824 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5825 narrower than M2, and both M1 and M2 are the same number of words,
5826 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5827 make that equivalence as well.
5829 However, BAR may have equivalences for which gen_lowpart
5830 will produce a simpler value than gen_lowpart applied to
5831 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5832 BAR's equivalences. If we don't get a simplified form, make
5833 the SUBREG. It will not be used in an equivalence, but will
5834 cause two similar assignments to be detected.
5836 Note the loop below will find SUBREG_REG (DEST) since we have
5837 already entered SRC and DEST of the SET in the table. */
5839 if (GET_CODE (dest) == SUBREG
5840 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5841 / UNITS_PER_WORD)
5842 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5843 && (GET_MODE_SIZE (GET_MODE (dest))
5844 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5845 && sets[i].src_elt != 0)
5847 enum machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5848 struct table_elt *elt, *classp = 0;
5850 for (elt = sets[i].src_elt->first_same_value; elt;
5851 elt = elt->next_same_value)
5853 rtx new_src = 0;
5854 unsigned src_hash;
5855 struct table_elt *src_elt;
5856 int byte = 0;
5858 /* Ignore invalid entries. */
5859 if (!REG_P (elt->exp)
5860 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5861 continue;
5863 /* We may have already been playing subreg games. If the
5864 mode is already correct for the destination, use it. */
5865 if (GET_MODE (elt->exp) == new_mode)
5866 new_src = elt->exp;
5867 else
5869 /* Calculate big endian correction for the SUBREG_BYTE.
5870 We have already checked that M1 (GET_MODE (dest))
5871 is not narrower than M2 (new_mode). */
5872 if (BYTES_BIG_ENDIAN)
5873 byte = (GET_MODE_SIZE (GET_MODE (dest))
5874 - GET_MODE_SIZE (new_mode));
5876 new_src = simplify_gen_subreg (new_mode, elt->exp,
5877 GET_MODE (dest), byte);
5880 /* The call to simplify_gen_subreg fails if the value
5881 is VOIDmode, yet we can't do any simplification, e.g.
5882 for EXPR_LISTs denoting function call results.
5883 It is invalid to construct a SUBREG with a VOIDmode
5884 SUBREG_REG, hence a zero new_src means we can't do
5885 this substitution. */
5886 if (! new_src)
5887 continue;
5889 src_hash = HASH (new_src, new_mode);
5890 src_elt = lookup (new_src, src_hash, new_mode);
5892 /* Put the new source in the hash table is if isn't
5893 already. */
5894 if (src_elt == 0)
5896 if (insert_regs (new_src, classp, 0))
5898 rehash_using_reg (new_src);
5899 src_hash = HASH (new_src, new_mode);
5901 src_elt = insert (new_src, classp, src_hash, new_mode);
5902 src_elt->in_memory = elt->in_memory;
5904 else if (classp && classp != src_elt->first_same_value)
5905 /* Show that two things that we've seen before are
5906 actually the same. */
5907 merge_equiv_classes (src_elt, classp);
5909 classp = src_elt->first_same_value;
5910 /* Ignore invalid entries. */
5911 while (classp
5912 && !REG_P (classp->exp)
5913 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5914 classp = classp->next_same_value;
5919 /* Special handling for (set REG0 REG1) where REG0 is the
5920 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5921 be used in the sequel, so (if easily done) change this insn to
5922 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5923 that computed their value. Then REG1 will become a dead store
5924 and won't cloud the situation for later optimizations.
5926 Do not make this change if REG1 is a hard register, because it will
5927 then be used in the sequel and we may be changing a two-operand insn
5928 into a three-operand insn.
5930 Also do not do this if we are operating on a copy of INSN. */
5932 if (n_sets == 1 && sets[0].rtl)
5933 try_back_substitute_reg (sets[0].rtl, insn);
5935 done:;
5938 /* Remove from the hash table all expressions that reference memory. */
5940 static void
5941 invalidate_memory (void)
5943 int i;
5944 struct table_elt *p, *next;
5946 for (i = 0; i < HASH_SIZE; i++)
5947 for (p = table[i]; p; p = next)
5949 next = p->next_same_hash;
5950 if (p->in_memory)
5951 remove_from_table (p, i);
5955 /* Perform invalidation on the basis of everything about INSN,
5956 except for invalidating the actual places that are SET in it.
5957 This includes the places CLOBBERed, and anything that might
5958 alias with something that is SET or CLOBBERed. */
5960 static void
5961 invalidate_from_clobbers (rtx insn)
5963 rtx x = PATTERN (insn);
5965 if (GET_CODE (x) == CLOBBER)
5967 rtx ref = XEXP (x, 0);
5968 if (ref)
5970 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5971 || MEM_P (ref))
5972 invalidate (ref, VOIDmode);
5973 else if (GET_CODE (ref) == STRICT_LOW_PART
5974 || GET_CODE (ref) == ZERO_EXTRACT)
5975 invalidate (XEXP (ref, 0), GET_MODE (ref));
5978 else if (GET_CODE (x) == PARALLEL)
5980 int i;
5981 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
5983 rtx y = XVECEXP (x, 0, i);
5984 if (GET_CODE (y) == CLOBBER)
5986 rtx ref = XEXP (y, 0);
5987 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5988 || MEM_P (ref))
5989 invalidate (ref, VOIDmode);
5990 else if (GET_CODE (ref) == STRICT_LOW_PART
5991 || GET_CODE (ref) == ZERO_EXTRACT)
5992 invalidate (XEXP (ref, 0), GET_MODE (ref));
5998 /* Perform invalidation on the basis of everything about INSN.
5999 This includes the places CLOBBERed, and anything that might
6000 alias with something that is SET or CLOBBERed. */
6002 static void
6003 invalidate_from_sets_and_clobbers (rtx insn)
6005 rtx tem;
6006 rtx x = PATTERN (insn);
6008 if (CALL_P (insn))
6010 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6011 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6012 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6015 /* Ensure we invalidate the destination register of a CALL insn.
6016 This is necessary for machines where this register is a fixed_reg,
6017 because no other code would invalidate it. */
6018 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6019 invalidate (SET_DEST (x), VOIDmode);
6021 else if (GET_CODE (x) == PARALLEL)
6023 int i;
6025 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6027 rtx y = XVECEXP (x, 0, i);
6028 if (GET_CODE (y) == CLOBBER)
6030 rtx clobbered = XEXP (y, 0);
6032 if (REG_P (clobbered)
6033 || GET_CODE (clobbered) == SUBREG)
6034 invalidate (clobbered, VOIDmode);
6035 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6036 || GET_CODE (clobbered) == ZERO_EXTRACT)
6037 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6039 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6040 invalidate (SET_DEST (y), VOIDmode);
6045 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6046 and replace any registers in them with either an equivalent constant
6047 or the canonical form of the register. If we are inside an address,
6048 only do this if the address remains valid.
6050 OBJECT is 0 except when within a MEM in which case it is the MEM.
6052 Return the replacement for X. */
6054 static rtx
6055 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6057 enum rtx_code code = GET_CODE (x);
6058 const char *fmt = GET_RTX_FORMAT (code);
6059 int i;
6061 switch (code)
6063 case CONST_INT:
6064 case CONST:
6065 case SYMBOL_REF:
6066 case LABEL_REF:
6067 case CONST_DOUBLE:
6068 case CONST_FIXED:
6069 case CONST_VECTOR:
6070 case PC:
6071 case CC0:
6072 case LO_SUM:
6073 return x;
6075 case MEM:
6076 validate_change (x, &XEXP (x, 0),
6077 cse_process_notes (XEXP (x, 0), x, changed), 0);
6078 return x;
6080 case EXPR_LIST:
6081 case INSN_LIST:
6082 if (REG_NOTE_KIND (x) == REG_EQUAL)
6083 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6084 if (XEXP (x, 1))
6085 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6086 return x;
6088 case SIGN_EXTEND:
6089 case ZERO_EXTEND:
6090 case SUBREG:
6092 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6093 /* We don't substitute VOIDmode constants into these rtx,
6094 since they would impede folding. */
6095 if (GET_MODE (new_rtx) != VOIDmode)
6096 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6097 return x;
6100 case REG:
6101 i = REG_QTY (REGNO (x));
6103 /* Return a constant or a constant register. */
6104 if (REGNO_QTY_VALID_P (REGNO (x)))
6106 struct qty_table_elem *ent = &qty_table[i];
6108 if (ent->const_rtx != NULL_RTX
6109 && (CONSTANT_P (ent->const_rtx)
6110 || REG_P (ent->const_rtx)))
6112 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6113 if (new_rtx)
6114 return copy_rtx (new_rtx);
6118 /* Otherwise, canonicalize this register. */
6119 return canon_reg (x, NULL_RTX);
6121 default:
6122 break;
6125 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6126 if (fmt[i] == 'e')
6127 validate_change (object, &XEXP (x, i),
6128 cse_process_notes (XEXP (x, i), object, changed), 0);
6130 return x;
6133 static rtx
6134 cse_process_notes (rtx x, rtx object, bool *changed)
6136 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6137 if (new_rtx != x)
6138 *changed = true;
6139 return new_rtx;
6143 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6145 DATA is a pointer to a struct cse_basic_block_data, that is used to
6146 describe the path.
6147 It is filled with a queue of basic blocks, starting with FIRST_BB
6148 and following a trace through the CFG.
6150 If all paths starting at FIRST_BB have been followed, or no new path
6151 starting at FIRST_BB can be constructed, this function returns FALSE.
6152 Otherwise, DATA->path is filled and the function returns TRUE indicating
6153 that a path to follow was found.
6155 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6156 block in the path will be FIRST_BB. */
6158 static bool
6159 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6160 int follow_jumps)
6162 basic_block bb;
6163 edge e;
6164 int path_size;
6166 SET_BIT (cse_visited_basic_blocks, first_bb->index);
6168 /* See if there is a previous path. */
6169 path_size = data->path_size;
6171 /* There is a previous path. Make sure it started with FIRST_BB. */
6172 if (path_size)
6173 gcc_assert (data->path[0].bb == first_bb);
6175 /* There was only one basic block in the last path. Clear the path and
6176 return, so that paths starting at another basic block can be tried. */
6177 if (path_size == 1)
6179 path_size = 0;
6180 goto done;
6183 /* If the path was empty from the beginning, construct a new path. */
6184 if (path_size == 0)
6185 data->path[path_size++].bb = first_bb;
6186 else
6188 /* Otherwise, path_size must be equal to or greater than 2, because
6189 a previous path exists that is at least two basic blocks long.
6191 Update the previous branch path, if any. If the last branch was
6192 previously along the branch edge, take the fallthrough edge now. */
6193 while (path_size >= 2)
6195 basic_block last_bb_in_path, previous_bb_in_path;
6196 edge e;
6198 --path_size;
6199 last_bb_in_path = data->path[path_size].bb;
6200 previous_bb_in_path = data->path[path_size - 1].bb;
6202 /* If we previously followed a path along the branch edge, try
6203 the fallthru edge now. */
6204 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6205 && any_condjump_p (BB_END (previous_bb_in_path))
6206 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6207 && e == BRANCH_EDGE (previous_bb_in_path))
6209 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6210 if (bb != EXIT_BLOCK_PTR
6211 && single_pred_p (bb)
6212 /* We used to assert here that we would only see blocks
6213 that we have not visited yet. But we may end up
6214 visiting basic blocks twice if the CFG has changed
6215 in this run of cse_main, because when the CFG changes
6216 the topological sort of the CFG also changes. A basic
6217 blocks that previously had more than two predecessors
6218 may now have a single predecessor, and become part of
6219 a path that starts at another basic block.
6221 We still want to visit each basic block only once, so
6222 halt the path here if we have already visited BB. */
6223 && !TEST_BIT (cse_visited_basic_blocks, bb->index))
6225 SET_BIT (cse_visited_basic_blocks, bb->index);
6226 data->path[path_size++].bb = bb;
6227 break;
6231 data->path[path_size].bb = NULL;
6234 /* If only one block remains in the path, bail. */
6235 if (path_size == 1)
6237 path_size = 0;
6238 goto done;
6242 /* Extend the path if possible. */
6243 if (follow_jumps)
6245 bb = data->path[path_size - 1].bb;
6246 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6248 if (single_succ_p (bb))
6249 e = single_succ_edge (bb);
6250 else if (EDGE_COUNT (bb->succs) == 2
6251 && any_condjump_p (BB_END (bb)))
6253 /* First try to follow the branch. If that doesn't lead
6254 to a useful path, follow the fallthru edge. */
6255 e = BRANCH_EDGE (bb);
6256 if (!single_pred_p (e->dest))
6257 e = FALLTHRU_EDGE (bb);
6259 else
6260 e = NULL;
6262 if (e
6263 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6264 && e->dest != EXIT_BLOCK_PTR
6265 && single_pred_p (e->dest)
6266 /* Avoid visiting basic blocks twice. The large comment
6267 above explains why this can happen. */
6268 && !TEST_BIT (cse_visited_basic_blocks, e->dest->index))
6270 basic_block bb2 = e->dest;
6271 SET_BIT (cse_visited_basic_blocks, bb2->index);
6272 data->path[path_size++].bb = bb2;
6273 bb = bb2;
6275 else
6276 bb = NULL;
6280 done:
6281 data->path_size = path_size;
6282 return path_size != 0;
6285 /* Dump the path in DATA to file F. NSETS is the number of sets
6286 in the path. */
6288 static void
6289 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6291 int path_entry;
6293 fprintf (f, ";; Following path with %d sets: ", nsets);
6294 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6295 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6296 fputc ('\n', dump_file);
6297 fflush (f);
6301 /* Return true if BB has exception handling successor edges. */
6303 static bool
6304 have_eh_succ_edges (basic_block bb)
6306 edge e;
6307 edge_iterator ei;
6309 FOR_EACH_EDGE (e, ei, bb->succs)
6310 if (e->flags & EDGE_EH)
6311 return true;
6313 return false;
6317 /* Scan to the end of the path described by DATA. Return an estimate of
6318 the total number of SETs of all insns in the path. */
6320 static void
6321 cse_prescan_path (struct cse_basic_block_data *data)
6323 int nsets = 0;
6324 int path_size = data->path_size;
6325 int path_entry;
6327 /* Scan to end of each basic block in the path. */
6328 for (path_entry = 0; path_entry < path_size; path_entry++)
6330 basic_block bb;
6331 rtx insn;
6333 bb = data->path[path_entry].bb;
6335 FOR_BB_INSNS (bb, insn)
6337 if (!INSN_P (insn))
6338 continue;
6340 /* A PARALLEL can have lots of SETs in it,
6341 especially if it is really an ASM_OPERANDS. */
6342 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6343 nsets += XVECLEN (PATTERN (insn), 0);
6344 else
6345 nsets += 1;
6349 data->nsets = nsets;
6352 /* Process a single extended basic block described by EBB_DATA. */
6354 static void
6355 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6357 int path_size = ebb_data->path_size;
6358 int path_entry;
6359 int num_insns = 0;
6361 /* Allocate the space needed by qty_table. */
6362 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6364 new_basic_block ();
6365 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6366 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6367 for (path_entry = 0; path_entry < path_size; path_entry++)
6369 basic_block bb;
6370 rtx insn;
6372 bb = ebb_data->path[path_entry].bb;
6374 /* Invalidate recorded information for eh regs if there is an EH
6375 edge pointing to that bb. */
6376 if (bb_has_eh_pred (bb))
6378 df_ref *def_rec;
6380 for (def_rec = df_get_artificial_defs (bb->index); *def_rec; def_rec++)
6382 df_ref def = *def_rec;
6383 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6384 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6388 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6389 FOR_BB_INSNS (bb, insn)
6391 /* If we have processed 1,000 insns, flush the hash table to
6392 avoid extreme quadratic behavior. We must not include NOTEs
6393 in the count since there may be more of them when generating
6394 debugging information. If we clear the table at different
6395 times, code generated with -g -O might be different than code
6396 generated with -O but not -g.
6398 FIXME: This is a real kludge and needs to be done some other
6399 way. */
6400 if (NONDEBUG_INSN_P (insn)
6401 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6403 flush_hash_table ();
6404 num_insns = 0;
6407 if (INSN_P (insn))
6409 /* Process notes first so we have all notes in canonical forms
6410 when looking for duplicate operations. */
6411 if (REG_NOTES (insn))
6413 bool changed = false;
6414 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6415 NULL_RTX, &changed);
6416 if (changed)
6417 df_notes_rescan (insn);
6420 cse_insn (insn);
6422 /* If we haven't already found an insn where we added a LABEL_REF,
6423 check this one. */
6424 if (INSN_P (insn) && !recorded_label_ref
6425 && for_each_rtx (&PATTERN (insn), check_for_label_ref,
6426 (void *) insn))
6427 recorded_label_ref = true;
6429 #ifdef HAVE_cc0
6430 if (NONDEBUG_INSN_P (insn))
6432 /* If the previous insn sets CC0 and this insn no
6433 longer references CC0, delete the previous insn.
6434 Here we use fact that nothing expects CC0 to be
6435 valid over an insn, which is true until the final
6436 pass. */
6437 rtx prev_insn, tem;
6439 prev_insn = prev_nonnote_nondebug_insn (insn);
6440 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6441 && (tem = single_set (prev_insn)) != NULL_RTX
6442 && SET_DEST (tem) == cc0_rtx
6443 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6444 delete_insn (prev_insn);
6446 /* If this insn is not the last insn in the basic
6447 block, it will be PREV_INSN(insn) in the next
6448 iteration. If we recorded any CC0-related
6449 information for this insn, remember it. */
6450 if (insn != BB_END (bb))
6452 prev_insn_cc0 = this_insn_cc0;
6453 prev_insn_cc0_mode = this_insn_cc0_mode;
6456 #endif
6460 /* With non-call exceptions, we are not always able to update
6461 the CFG properly inside cse_insn. So clean up possibly
6462 redundant EH edges here. */
6463 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6464 cse_cfg_altered |= purge_dead_edges (bb);
6466 /* If we changed a conditional jump, we may have terminated
6467 the path we are following. Check that by verifying that
6468 the edge we would take still exists. If the edge does
6469 not exist anymore, purge the remainder of the path.
6470 Note that this will cause us to return to the caller. */
6471 if (path_entry < path_size - 1)
6473 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6474 if (!find_edge (bb, next_bb))
6478 path_size--;
6480 /* If we truncate the path, we must also reset the
6481 visited bit on the remaining blocks in the path,
6482 or we will never visit them at all. */
6483 RESET_BIT (cse_visited_basic_blocks,
6484 ebb_data->path[path_size].bb->index);
6485 ebb_data->path[path_size].bb = NULL;
6487 while (path_size - 1 != path_entry);
6488 ebb_data->path_size = path_size;
6492 /* If this is a conditional jump insn, record any known
6493 equivalences due to the condition being tested. */
6494 insn = BB_END (bb);
6495 if (path_entry < path_size - 1
6496 && JUMP_P (insn)
6497 && single_set (insn)
6498 && any_condjump_p (insn))
6500 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6501 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6502 record_jump_equiv (insn, taken);
6505 #ifdef HAVE_cc0
6506 /* Clear the CC0-tracking related insns, they can't provide
6507 useful information across basic block boundaries. */
6508 prev_insn_cc0 = 0;
6509 #endif
6512 gcc_assert (next_qty <= max_qty);
6514 free (qty_table);
6518 /* Perform cse on the instructions of a function.
6519 F is the first instruction.
6520 NREGS is one plus the highest pseudo-reg number used in the instruction.
6522 Return 2 if jump optimizations should be redone due to simplifications
6523 in conditional jump instructions.
6524 Return 1 if the CFG should be cleaned up because it has been modified.
6525 Return 0 otherwise. */
6527 static int
6528 cse_main (rtx f ATTRIBUTE_UNUSED, int nregs)
6530 struct cse_basic_block_data ebb_data;
6531 basic_block bb;
6532 int *rc_order = XNEWVEC (int, last_basic_block);
6533 int i, n_blocks;
6535 df_set_flags (DF_LR_RUN_DCE);
6536 df_analyze ();
6537 df_set_flags (DF_DEFER_INSN_RESCAN);
6539 reg_scan (get_insns (), max_reg_num ());
6540 init_cse_reg_info (nregs);
6542 ebb_data.path = XNEWVEC (struct branch_path,
6543 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6545 cse_cfg_altered = false;
6546 cse_jumps_altered = false;
6547 recorded_label_ref = false;
6548 constant_pool_entries_cost = 0;
6549 constant_pool_entries_regcost = 0;
6550 ebb_data.path_size = 0;
6551 ebb_data.nsets = 0;
6552 rtl_hooks = cse_rtl_hooks;
6554 init_recog ();
6555 init_alias_analysis ();
6557 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6559 /* Set up the table of already visited basic blocks. */
6560 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block);
6561 sbitmap_zero (cse_visited_basic_blocks);
6563 /* Loop over basic blocks in reverse completion order (RPO),
6564 excluding the ENTRY and EXIT blocks. */
6565 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6566 i = 0;
6567 while (i < n_blocks)
6569 /* Find the first block in the RPO queue that we have not yet
6570 processed before. */
6573 bb = BASIC_BLOCK (rc_order[i++]);
6575 while (TEST_BIT (cse_visited_basic_blocks, bb->index)
6576 && i < n_blocks);
6578 /* Find all paths starting with BB, and process them. */
6579 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6581 /* Pre-scan the path. */
6582 cse_prescan_path (&ebb_data);
6584 /* If this basic block has no sets, skip it. */
6585 if (ebb_data.nsets == 0)
6586 continue;
6588 /* Get a reasonable estimate for the maximum number of qty's
6589 needed for this path. For this, we take the number of sets
6590 and multiply that by MAX_RECOG_OPERANDS. */
6591 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6593 /* Dump the path we're about to process. */
6594 if (dump_file)
6595 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6597 cse_extended_basic_block (&ebb_data);
6601 /* Clean up. */
6602 end_alias_analysis ();
6603 free (reg_eqv_table);
6604 free (ebb_data.path);
6605 sbitmap_free (cse_visited_basic_blocks);
6606 free (rc_order);
6607 rtl_hooks = general_rtl_hooks;
6609 if (cse_jumps_altered || recorded_label_ref)
6610 return 2;
6611 else if (cse_cfg_altered)
6612 return 1;
6613 else
6614 return 0;
6617 /* Called via for_each_rtx to see if an insn is using a LABEL_REF for
6618 which there isn't a REG_LABEL_OPERAND note.
6619 Return one if so. DATA is the insn. */
6621 static int
6622 check_for_label_ref (rtx *rtl, void *data)
6624 rtx insn = (rtx) data;
6626 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6627 note for it, we must rerun jump since it needs to place the note. If
6628 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6629 don't do this since no REG_LABEL_OPERAND will be added. */
6630 return (GET_CODE (*rtl) == LABEL_REF
6631 && ! LABEL_REF_NONLOCAL_P (*rtl)
6632 && (!JUMP_P (insn)
6633 || !label_is_jump_target_p (XEXP (*rtl, 0), insn))
6634 && LABEL_P (XEXP (*rtl, 0))
6635 && INSN_UID (XEXP (*rtl, 0)) != 0
6636 && ! find_reg_note (insn, REG_LABEL_OPERAND, XEXP (*rtl, 0)));
6639 /* Count the number of times registers are used (not set) in X.
6640 COUNTS is an array in which we accumulate the count, INCR is how much
6641 we count each register usage.
6643 Don't count a usage of DEST, which is the SET_DEST of a SET which
6644 contains X in its SET_SRC. This is because such a SET does not
6645 modify the liveness of DEST.
6646 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6647 We must then count uses of a SET_DEST regardless, because the insn can't be
6648 deleted here. */
6650 static void
6651 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6653 enum rtx_code code;
6654 rtx note;
6655 const char *fmt;
6656 int i, j;
6658 if (x == 0)
6659 return;
6661 switch (code = GET_CODE (x))
6663 case REG:
6664 if (x != dest)
6665 counts[REGNO (x)] += incr;
6666 return;
6668 case PC:
6669 case CC0:
6670 case CONST:
6671 case CONST_INT:
6672 case CONST_DOUBLE:
6673 case CONST_FIXED:
6674 case CONST_VECTOR:
6675 case SYMBOL_REF:
6676 case LABEL_REF:
6677 return;
6679 case CLOBBER:
6680 /* If we are clobbering a MEM, mark any registers inside the address
6681 as being used. */
6682 if (MEM_P (XEXP (x, 0)))
6683 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6684 return;
6686 case SET:
6687 /* Unless we are setting a REG, count everything in SET_DEST. */
6688 if (!REG_P (SET_DEST (x)))
6689 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6690 count_reg_usage (SET_SRC (x), counts,
6691 dest ? dest : SET_DEST (x),
6692 incr);
6693 return;
6695 case DEBUG_INSN:
6696 return;
6698 case CALL_INSN:
6699 case INSN:
6700 case JUMP_INSN:
6701 /* We expect dest to be NULL_RTX here. If the insn may trap,
6702 or if it cannot be deleted due to side-effects, mark this fact
6703 by setting DEST to pc_rtx. */
6704 if (insn_could_throw_p (x) || side_effects_p (PATTERN (x)))
6705 dest = pc_rtx;
6706 if (code == CALL_INSN)
6707 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6708 count_reg_usage (PATTERN (x), counts, dest, incr);
6710 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6711 use them. */
6713 note = find_reg_equal_equiv_note (x);
6714 if (note)
6716 rtx eqv = XEXP (note, 0);
6718 if (GET_CODE (eqv) == EXPR_LIST)
6719 /* This REG_EQUAL note describes the result of a function call.
6720 Process all the arguments. */
6723 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6724 eqv = XEXP (eqv, 1);
6726 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6727 else
6728 count_reg_usage (eqv, counts, dest, incr);
6730 return;
6732 case EXPR_LIST:
6733 if (REG_NOTE_KIND (x) == REG_EQUAL
6734 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6735 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6736 involving registers in the address. */
6737 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6738 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6740 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6741 return;
6743 case ASM_OPERANDS:
6744 /* Iterate over just the inputs, not the constraints as well. */
6745 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6746 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6747 return;
6749 case INSN_LIST:
6750 gcc_unreachable ();
6752 default:
6753 break;
6756 fmt = GET_RTX_FORMAT (code);
6757 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6759 if (fmt[i] == 'e')
6760 count_reg_usage (XEXP (x, i), counts, dest, incr);
6761 else if (fmt[i] == 'E')
6762 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6763 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6767 /* Return true if X is a dead register. */
6769 static inline int
6770 is_dead_reg (rtx x, int *counts)
6772 return (REG_P (x)
6773 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6774 && counts[REGNO (x)] == 0);
6777 /* Return true if set is live. */
6778 static bool
6779 set_live_p (rtx set, rtx insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6780 int *counts)
6782 #ifdef HAVE_cc0
6783 rtx tem;
6784 #endif
6786 if (set_noop_p (set))
6789 #ifdef HAVE_cc0
6790 else if (GET_CODE (SET_DEST (set)) == CC0
6791 && !side_effects_p (SET_SRC (set))
6792 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6793 || !INSN_P (tem)
6794 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6795 return false;
6796 #endif
6797 else if (!is_dead_reg (SET_DEST (set), counts)
6798 || side_effects_p (SET_SRC (set)))
6799 return true;
6800 return false;
6803 /* Return true if insn is live. */
6805 static bool
6806 insn_live_p (rtx insn, int *counts)
6808 int i;
6809 if (insn_could_throw_p (insn))
6810 return true;
6811 else if (GET_CODE (PATTERN (insn)) == SET)
6812 return set_live_p (PATTERN (insn), insn, counts);
6813 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6815 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6817 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6819 if (GET_CODE (elt) == SET)
6821 if (set_live_p (elt, insn, counts))
6822 return true;
6824 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6825 return true;
6827 return false;
6829 else if (DEBUG_INSN_P (insn))
6831 rtx next;
6833 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6834 if (NOTE_P (next))
6835 continue;
6836 else if (!DEBUG_INSN_P (next))
6837 return true;
6838 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6839 return false;
6841 return true;
6843 else
6844 return true;
6847 /* Count the number of stores into pseudo. Callback for note_stores. */
6849 static void
6850 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6852 int *counts = (int *) data;
6853 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6854 counts[REGNO (x)]++;
6857 struct dead_debug_insn_data
6859 int *counts;
6860 rtx *replacements;
6861 bool seen_repl;
6864 /* Return if a DEBUG_INSN needs to be reset because some dead
6865 pseudo doesn't have a replacement. Callback for for_each_rtx. */
6867 static int
6868 is_dead_debug_insn (rtx *loc, void *data)
6870 rtx x = *loc;
6871 struct dead_debug_insn_data *ddid = (struct dead_debug_insn_data *) data;
6873 if (is_dead_reg (x, ddid->counts))
6875 if (ddid->replacements && ddid->replacements[REGNO (x)] != NULL_RTX)
6876 ddid->seen_repl = true;
6877 else
6878 return 1;
6880 return 0;
6883 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6884 Callback for simplify_replace_fn_rtx. */
6886 static rtx
6887 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6889 rtx *replacements = (rtx *) data;
6891 if (REG_P (x)
6892 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6893 && replacements[REGNO (x)] != NULL_RTX)
6895 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6896 return replacements[REGNO (x)];
6897 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6898 GET_MODE (replacements[REGNO (x)]));
6900 return NULL_RTX;
6903 /* Scan all the insns and delete any that are dead; i.e., they store a register
6904 that is never used or they copy a register to itself.
6906 This is used to remove insns made obviously dead by cse, loop or other
6907 optimizations. It improves the heuristics in loop since it won't try to
6908 move dead invariants out of loops or make givs for dead quantities. The
6909 remaining passes of the compilation are also sped up. */
6912 delete_trivially_dead_insns (rtx insns, int nreg)
6914 int *counts;
6915 rtx insn, prev;
6916 rtx *replacements = NULL;
6917 int ndead = 0;
6919 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6920 /* First count the number of times each register is used. */
6921 if (MAY_HAVE_DEBUG_INSNS)
6923 counts = XCNEWVEC (int, nreg * 3);
6924 for (insn = insns; insn; insn = NEXT_INSN (insn))
6925 if (DEBUG_INSN_P (insn))
6926 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6927 NULL_RTX, 1);
6928 else if (INSN_P (insn))
6930 count_reg_usage (insn, counts, NULL_RTX, 1);
6931 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6933 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6934 First one counts how many times each pseudo is used outside
6935 of debug insns, second counts how many times each pseudo is
6936 used in debug insns and third counts how many times a pseudo
6937 is stored. */
6939 else
6941 counts = XCNEWVEC (int, nreg);
6942 for (insn = insns; insn; insn = NEXT_INSN (insn))
6943 if (INSN_P (insn))
6944 count_reg_usage (insn, counts, NULL_RTX, 1);
6945 /* If no debug insns can be present, COUNTS is just an array
6946 which counts how many times each pseudo is used. */
6948 /* Go from the last insn to the first and delete insns that only set unused
6949 registers or copy a register to itself. As we delete an insn, remove
6950 usage counts for registers it uses.
6952 The first jump optimization pass may leave a real insn as the last
6953 insn in the function. We must not skip that insn or we may end
6954 up deleting code that is not really dead.
6956 If some otherwise unused register is only used in DEBUG_INSNs,
6957 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6958 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
6959 has been created for the unused register, replace it with
6960 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
6961 for (insn = get_last_insn (); insn; insn = prev)
6963 int live_insn = 0;
6965 prev = PREV_INSN (insn);
6966 if (!INSN_P (insn))
6967 continue;
6969 live_insn = insn_live_p (insn, counts);
6971 /* If this is a dead insn, delete it and show registers in it aren't
6972 being used. */
6974 if (! live_insn && dbg_cnt (delete_trivial_dead))
6976 if (DEBUG_INSN_P (insn))
6977 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6978 NULL_RTX, -1);
6979 else
6981 rtx set;
6982 if (MAY_HAVE_DEBUG_INSNS
6983 && (set = single_set (insn)) != NULL_RTX
6984 && is_dead_reg (SET_DEST (set), counts)
6985 /* Used at least once in some DEBUG_INSN. */
6986 && counts[REGNO (SET_DEST (set)) + nreg] > 0
6987 /* And set exactly once. */
6988 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
6989 && !side_effects_p (SET_SRC (set))
6990 && asm_noperands (PATTERN (insn)) < 0)
6992 rtx dval, bind;
6994 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
6995 dval = make_debug_expr_from_rtl (SET_DEST (set));
6997 /* Emit a debug bind insn before the insn in which
6998 reg dies. */
6999 bind = gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7000 DEBUG_EXPR_TREE_DECL (dval),
7001 SET_SRC (set),
7002 VAR_INIT_STATUS_INITIALIZED);
7003 count_reg_usage (bind, counts + nreg, NULL_RTX, 1);
7005 bind = emit_debug_insn_before (bind, insn);
7006 df_insn_rescan (bind);
7008 if (replacements == NULL)
7009 replacements = XCNEWVEC (rtx, nreg);
7010 replacements[REGNO (SET_DEST (set))] = dval;
7013 count_reg_usage (insn, counts, NULL_RTX, -1);
7014 ndead++;
7016 delete_insn_and_edges (insn);
7020 if (MAY_HAVE_DEBUG_INSNS)
7022 struct dead_debug_insn_data ddid;
7023 ddid.counts = counts;
7024 ddid.replacements = replacements;
7025 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7026 if (DEBUG_INSN_P (insn))
7028 /* If this debug insn references a dead register that wasn't replaced
7029 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7030 ddid.seen_repl = false;
7031 if (for_each_rtx (&INSN_VAR_LOCATION_LOC (insn),
7032 is_dead_debug_insn, &ddid))
7034 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7035 df_insn_rescan (insn);
7037 else if (ddid.seen_repl)
7039 INSN_VAR_LOCATION_LOC (insn)
7040 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7041 NULL_RTX, replace_dead_reg,
7042 replacements);
7043 df_insn_rescan (insn);
7046 free (replacements);
7049 if (dump_file && ndead)
7050 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7051 ndead);
7052 /* Clean up. */
7053 free (counts);
7054 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7055 return ndead;
7058 /* This function is called via for_each_rtx. The argument, NEWREG, is
7059 a condition code register with the desired mode. If we are looking
7060 at the same register in a different mode, replace it with
7061 NEWREG. */
7063 static int
7064 cse_change_cc_mode (rtx *loc, void *data)
7066 struct change_cc_mode_args* args = (struct change_cc_mode_args*)data;
7068 if (*loc
7069 && REG_P (*loc)
7070 && REGNO (*loc) == REGNO (args->newreg)
7071 && GET_MODE (*loc) != GET_MODE (args->newreg))
7073 validate_change (args->insn, loc, args->newreg, 1);
7075 return -1;
7077 return 0;
7080 /* Change the mode of any reference to the register REGNO (NEWREG) to
7081 GET_MODE (NEWREG) in INSN. */
7083 static void
7084 cse_change_cc_mode_insn (rtx insn, rtx newreg)
7086 struct change_cc_mode_args args;
7087 int success;
7089 if (!INSN_P (insn))
7090 return;
7092 args.insn = insn;
7093 args.newreg = newreg;
7095 for_each_rtx (&PATTERN (insn), cse_change_cc_mode, &args);
7096 for_each_rtx (&REG_NOTES (insn), cse_change_cc_mode, &args);
7098 /* If the following assertion was triggered, there is most probably
7099 something wrong with the cc_modes_compatible back end function.
7100 CC modes only can be considered compatible if the insn - with the mode
7101 replaced by any of the compatible modes - can still be recognized. */
7102 success = apply_change_group ();
7103 gcc_assert (success);
7106 /* Change the mode of any reference to the register REGNO (NEWREG) to
7107 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7108 any instruction which modifies NEWREG. */
7110 static void
7111 cse_change_cc_mode_insns (rtx start, rtx end, rtx newreg)
7113 rtx insn;
7115 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7117 if (! INSN_P (insn))
7118 continue;
7120 if (reg_set_p (newreg, insn))
7121 return;
7123 cse_change_cc_mode_insn (insn, newreg);
7127 /* BB is a basic block which finishes with CC_REG as a condition code
7128 register which is set to CC_SRC. Look through the successors of BB
7129 to find blocks which have a single predecessor (i.e., this one),
7130 and look through those blocks for an assignment to CC_REG which is
7131 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7132 permitted to change the mode of CC_SRC to a compatible mode. This
7133 returns VOIDmode if no equivalent assignments were found.
7134 Otherwise it returns the mode which CC_SRC should wind up with.
7135 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7136 but is passed unmodified down to recursive calls in order to prevent
7137 endless recursion.
7139 The main complexity in this function is handling the mode issues.
7140 We may have more than one duplicate which we can eliminate, and we
7141 try to find a mode which will work for multiple duplicates. */
7143 static enum machine_mode
7144 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7145 bool can_change_mode)
7147 bool found_equiv;
7148 enum machine_mode mode;
7149 unsigned int insn_count;
7150 edge e;
7151 rtx insns[2];
7152 enum machine_mode modes[2];
7153 rtx last_insns[2];
7154 unsigned int i;
7155 rtx newreg;
7156 edge_iterator ei;
7158 /* We expect to have two successors. Look at both before picking
7159 the final mode for the comparison. If we have more successors
7160 (i.e., some sort of table jump, although that seems unlikely),
7161 then we require all beyond the first two to use the same
7162 mode. */
7164 found_equiv = false;
7165 mode = GET_MODE (cc_src);
7166 insn_count = 0;
7167 FOR_EACH_EDGE (e, ei, bb->succs)
7169 rtx insn;
7170 rtx end;
7172 if (e->flags & EDGE_COMPLEX)
7173 continue;
7175 if (EDGE_COUNT (e->dest->preds) != 1
7176 || e->dest == EXIT_BLOCK_PTR
7177 /* Avoid endless recursion on unreachable blocks. */
7178 || e->dest == orig_bb)
7179 continue;
7181 end = NEXT_INSN (BB_END (e->dest));
7182 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7184 rtx set;
7186 if (! INSN_P (insn))
7187 continue;
7189 /* If CC_SRC is modified, we have to stop looking for
7190 something which uses it. */
7191 if (modified_in_p (cc_src, insn))
7192 break;
7194 /* Check whether INSN sets CC_REG to CC_SRC. */
7195 set = single_set (insn);
7196 if (set
7197 && REG_P (SET_DEST (set))
7198 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7200 bool found;
7201 enum machine_mode set_mode;
7202 enum machine_mode comp_mode;
7204 found = false;
7205 set_mode = GET_MODE (SET_SRC (set));
7206 comp_mode = set_mode;
7207 if (rtx_equal_p (cc_src, SET_SRC (set)))
7208 found = true;
7209 else if (GET_CODE (cc_src) == COMPARE
7210 && GET_CODE (SET_SRC (set)) == COMPARE
7211 && mode != set_mode
7212 && rtx_equal_p (XEXP (cc_src, 0),
7213 XEXP (SET_SRC (set), 0))
7214 && rtx_equal_p (XEXP (cc_src, 1),
7215 XEXP (SET_SRC (set), 1)))
7218 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7219 if (comp_mode != VOIDmode
7220 && (can_change_mode || comp_mode == mode))
7221 found = true;
7224 if (found)
7226 found_equiv = true;
7227 if (insn_count < ARRAY_SIZE (insns))
7229 insns[insn_count] = insn;
7230 modes[insn_count] = set_mode;
7231 last_insns[insn_count] = end;
7232 ++insn_count;
7234 if (mode != comp_mode)
7236 gcc_assert (can_change_mode);
7237 mode = comp_mode;
7239 /* The modified insn will be re-recognized later. */
7240 PUT_MODE (cc_src, mode);
7243 else
7245 if (set_mode != mode)
7247 /* We found a matching expression in the
7248 wrong mode, but we don't have room to
7249 store it in the array. Punt. This case
7250 should be rare. */
7251 break;
7253 /* INSN sets CC_REG to a value equal to CC_SRC
7254 with the right mode. We can simply delete
7255 it. */
7256 delete_insn (insn);
7259 /* We found an instruction to delete. Keep looking,
7260 in the hopes of finding a three-way jump. */
7261 continue;
7264 /* We found an instruction which sets the condition
7265 code, so don't look any farther. */
7266 break;
7269 /* If INSN sets CC_REG in some other way, don't look any
7270 farther. */
7271 if (reg_set_p (cc_reg, insn))
7272 break;
7275 /* If we fell off the bottom of the block, we can keep looking
7276 through successors. We pass CAN_CHANGE_MODE as false because
7277 we aren't prepared to handle compatibility between the
7278 further blocks and this block. */
7279 if (insn == end)
7281 enum machine_mode submode;
7283 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7284 if (submode != VOIDmode)
7286 gcc_assert (submode == mode);
7287 found_equiv = true;
7288 can_change_mode = false;
7293 if (! found_equiv)
7294 return VOIDmode;
7296 /* Now INSN_COUNT is the number of instructions we found which set
7297 CC_REG to a value equivalent to CC_SRC. The instructions are in
7298 INSNS. The modes used by those instructions are in MODES. */
7300 newreg = NULL_RTX;
7301 for (i = 0; i < insn_count; ++i)
7303 if (modes[i] != mode)
7305 /* We need to change the mode of CC_REG in INSNS[i] and
7306 subsequent instructions. */
7307 if (! newreg)
7309 if (GET_MODE (cc_reg) == mode)
7310 newreg = cc_reg;
7311 else
7312 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7314 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7315 newreg);
7318 delete_insn_and_edges (insns[i]);
7321 return mode;
7324 /* If we have a fixed condition code register (or two), walk through
7325 the instructions and try to eliminate duplicate assignments. */
7327 static void
7328 cse_condition_code_reg (void)
7330 unsigned int cc_regno_1;
7331 unsigned int cc_regno_2;
7332 rtx cc_reg_1;
7333 rtx cc_reg_2;
7334 basic_block bb;
7336 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7337 return;
7339 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7340 if (cc_regno_2 != INVALID_REGNUM)
7341 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7342 else
7343 cc_reg_2 = NULL_RTX;
7345 FOR_EACH_BB (bb)
7347 rtx last_insn;
7348 rtx cc_reg;
7349 rtx insn;
7350 rtx cc_src_insn;
7351 rtx cc_src;
7352 enum machine_mode mode;
7353 enum machine_mode orig_mode;
7355 /* Look for blocks which end with a conditional jump based on a
7356 condition code register. Then look for the instruction which
7357 sets the condition code register. Then look through the
7358 successor blocks for instructions which set the condition
7359 code register to the same value. There are other possible
7360 uses of the condition code register, but these are by far the
7361 most common and the ones which we are most likely to be able
7362 to optimize. */
7364 last_insn = BB_END (bb);
7365 if (!JUMP_P (last_insn))
7366 continue;
7368 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7369 cc_reg = cc_reg_1;
7370 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7371 cc_reg = cc_reg_2;
7372 else
7373 continue;
7375 cc_src_insn = NULL_RTX;
7376 cc_src = NULL_RTX;
7377 for (insn = PREV_INSN (last_insn);
7378 insn && insn != PREV_INSN (BB_HEAD (bb));
7379 insn = PREV_INSN (insn))
7381 rtx set;
7383 if (! INSN_P (insn))
7384 continue;
7385 set = single_set (insn);
7386 if (set
7387 && REG_P (SET_DEST (set))
7388 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7390 cc_src_insn = insn;
7391 cc_src = SET_SRC (set);
7392 break;
7394 else if (reg_set_p (cc_reg, insn))
7395 break;
7398 if (! cc_src_insn)
7399 continue;
7401 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7402 continue;
7404 /* Now CC_REG is a condition code register used for a
7405 conditional jump at the end of the block, and CC_SRC, in
7406 CC_SRC_INSN, is the value to which that condition code
7407 register is set, and CC_SRC is still meaningful at the end of
7408 the basic block. */
7410 orig_mode = GET_MODE (cc_src);
7411 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7412 if (mode != VOIDmode)
7414 gcc_assert (mode == GET_MODE (cc_src));
7415 if (mode != orig_mode)
7417 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7419 cse_change_cc_mode_insn (cc_src_insn, newreg);
7421 /* Do the same in the following insns that use the
7422 current value of CC_REG within BB. */
7423 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7424 NEXT_INSN (last_insn),
7425 newreg);
7432 /* Perform common subexpression elimination. Nonzero value from
7433 `cse_main' means that jumps were simplified and some code may now
7434 be unreachable, so do jump optimization again. */
7435 static bool
7436 gate_handle_cse (void)
7438 return optimize > 0;
7441 static unsigned int
7442 rest_of_handle_cse (void)
7444 int tem;
7446 if (dump_file)
7447 dump_flow_info (dump_file, dump_flags);
7449 tem = cse_main (get_insns (), max_reg_num ());
7451 /* If we are not running more CSE passes, then we are no longer
7452 expecting CSE to be run. But always rerun it in a cheap mode. */
7453 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7455 if (tem == 2)
7457 timevar_push (TV_JUMP);
7458 rebuild_jump_labels (get_insns ());
7459 cleanup_cfg (CLEANUP_CFG_CHANGED);
7460 timevar_pop (TV_JUMP);
7462 else if (tem == 1 || optimize > 1)
7463 cleanup_cfg (0);
7465 return 0;
7468 struct rtl_opt_pass pass_cse =
7471 RTL_PASS,
7472 "cse1", /* name */
7473 gate_handle_cse, /* gate */
7474 rest_of_handle_cse, /* execute */
7475 NULL, /* sub */
7476 NULL, /* next */
7477 0, /* static_pass_number */
7478 TV_CSE, /* tv_id */
7479 0, /* properties_required */
7480 0, /* properties_provided */
7481 0, /* properties_destroyed */
7482 0, /* todo_flags_start */
7483 TODO_df_finish | TODO_verify_rtl_sharing |
7484 TODO_ggc_collect |
7485 TODO_verify_flow, /* todo_flags_finish */
7490 static bool
7491 gate_handle_cse2 (void)
7493 return optimize > 0 && flag_rerun_cse_after_loop;
7496 /* Run second CSE pass after loop optimizations. */
7497 static unsigned int
7498 rest_of_handle_cse2 (void)
7500 int tem;
7502 if (dump_file)
7503 dump_flow_info (dump_file, dump_flags);
7505 tem = cse_main (get_insns (), max_reg_num ());
7507 /* Run a pass to eliminate duplicated assignments to condition code
7508 registers. We have to run this after bypass_jumps, because it
7509 makes it harder for that pass to determine whether a jump can be
7510 bypassed safely. */
7511 cse_condition_code_reg ();
7513 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7515 if (tem == 2)
7517 timevar_push (TV_JUMP);
7518 rebuild_jump_labels (get_insns ());
7519 cleanup_cfg (CLEANUP_CFG_CHANGED);
7520 timevar_pop (TV_JUMP);
7522 else if (tem == 1)
7523 cleanup_cfg (0);
7525 cse_not_expected = 1;
7526 return 0;
7530 struct rtl_opt_pass pass_cse2 =
7533 RTL_PASS,
7534 "cse2", /* name */
7535 gate_handle_cse2, /* gate */
7536 rest_of_handle_cse2, /* execute */
7537 NULL, /* sub */
7538 NULL, /* next */
7539 0, /* static_pass_number */
7540 TV_CSE2, /* tv_id */
7541 0, /* properties_required */
7542 0, /* properties_provided */
7543 0, /* properties_destroyed */
7544 0, /* todo_flags_start */
7545 TODO_df_finish | TODO_verify_rtl_sharing |
7546 TODO_ggc_collect |
7547 TODO_verify_flow /* todo_flags_finish */
7551 static bool
7552 gate_handle_cse_after_global_opts (void)
7554 return optimize > 0 && flag_rerun_cse_after_global_opts;
7557 /* Run second CSE pass after loop optimizations. */
7558 static unsigned int
7559 rest_of_handle_cse_after_global_opts (void)
7561 int save_cfj;
7562 int tem;
7564 /* We only want to do local CSE, so don't follow jumps. */
7565 save_cfj = flag_cse_follow_jumps;
7566 flag_cse_follow_jumps = 0;
7568 rebuild_jump_labels (get_insns ());
7569 tem = cse_main (get_insns (), max_reg_num ());
7570 purge_all_dead_edges ();
7571 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7573 cse_not_expected = !flag_rerun_cse_after_loop;
7575 /* If cse altered any jumps, rerun jump opts to clean things up. */
7576 if (tem == 2)
7578 timevar_push (TV_JUMP);
7579 rebuild_jump_labels (get_insns ());
7580 cleanup_cfg (CLEANUP_CFG_CHANGED);
7581 timevar_pop (TV_JUMP);
7583 else if (tem == 1)
7584 cleanup_cfg (0);
7586 flag_cse_follow_jumps = save_cfj;
7587 return 0;
7590 struct rtl_opt_pass pass_cse_after_global_opts =
7593 RTL_PASS,
7594 "cse_local", /* name */
7595 gate_handle_cse_after_global_opts, /* gate */
7596 rest_of_handle_cse_after_global_opts, /* execute */
7597 NULL, /* sub */
7598 NULL, /* next */
7599 0, /* static_pass_number */
7600 TV_CSE, /* tv_id */
7601 0, /* properties_required */
7602 0, /* properties_provided */
7603 0, /* properties_destroyed */
7604 0, /* todo_flags_start */
7605 TODO_df_finish | TODO_verify_rtl_sharing |
7606 TODO_ggc_collect |
7607 TODO_verify_flow /* todo_flags_finish */