Add a test for PR66655
[gcc.git] / gcc / cse.c
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2016 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "cfghooks.h"
28 #include "df.h"
29 #include "tm_p.h"
30 #include "insn-config.h"
31 #include "regs.h"
32 #include "emit-rtl.h"
33 #include "recog.h"
34 #include "cfgrtl.h"
35 #include "cfganal.h"
36 #include "cfgcleanup.h"
37 #include "alias.h"
38 #include "toplev.h"
39 #include "params.h"
40 #include "rtlhooks-def.h"
41 #include "tree-pass.h"
42 #include "dbgcnt.h"
43 #include "rtl-iter.h"
44
45 #ifndef LOAD_EXTEND_OP
46 #define LOAD_EXTEND_OP(M) UNKNOWN
47 #endif
48
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.
53
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.
59
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.
63
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.
67
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.
76
77 Registers and "quantity numbers":
78
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.
87
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.
91
92 Quantity numbers below zero do not exist and none of the `qty_table'
93 entries should be referenced with a negative index.
94
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.
98
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.
102
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.
106
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.
111
112 Constants and quantity numbers
113
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.
117
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.
121
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'.
125
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.
131
132 Other expressions:
133
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.
139
140 Other chains through the same elements connect the elements which
141 currently have equivalent values.
142
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.
147
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.
151
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.
156
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.
164
165 2. If the value changing is a register, all expressions
166 containing references to that register, and only those,
167 must be removed.
168
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.
176
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.
187
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.
195
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.
199
200 Related expressions:
201
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. */
208
209 /* Length of qty_table vector. We know in advance we will not need
210 a quantity number this big. */
211
212 static int max_qty;
213
214 /* Next quantity number to be allocated.
215 This is 1 + the largest number needed so far. */
216
217 static int next_qty;
218
219 /* Per-qty information tracking.
220
221 `first_reg' and `last_reg' track the head and tail of the
222 chain of registers which currently contain this quantity.
223
224 `mode' contains the machine mode of this quantity.
225
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.
231
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. */
243
244 struct qty_table_elem
245 {
246 rtx const_rtx;
247 rtx_insn *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;
255 };
256
257 /* The table of all qtys, indexed by qty number. */
258 static struct qty_table_elem *qty_table;
259
260 /* For machines that have a CC0, we do not record its value in the hash
261 table since its use is guaranteed to be the insn immediately following
262 its definition and any other insn is presumed to invalidate it.
263
264 Instead, we store below the current and last value assigned to CC0.
265 If it should happen to be a constant, it is stored in preference
266 to the actual assigned value. In case it is a constant, we store
267 the mode in which the constant should be interpreted. */
268
269 static rtx this_insn_cc0, prev_insn_cc0;
270 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
271
272 /* Insn being scanned. */
273
274 static rtx_insn *this_insn;
275 static bool optimize_this_for_speed_p;
276
277 /* Index by register number, gives the number of the next (or
278 previous) register in the chain of registers sharing the same
279 value.
280
281 Or -1 if this register is at the end of the chain.
282
283 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
284
285 /* Per-register equivalence chain. */
286 struct reg_eqv_elem
287 {
288 int next, prev;
289 };
290
291 /* The table of all register equivalence chains. */
292 static struct reg_eqv_elem *reg_eqv_table;
293
294 struct cse_reg_info
295 {
296 /* The timestamp at which this register is initialized. */
297 unsigned int timestamp;
298
299 /* The quantity number of the register's current contents. */
300 int reg_qty;
301
302 /* The number of times the register has been altered in the current
303 basic block. */
304 int reg_tick;
305
306 /* The REG_TICK value at which rtx's containing this register are
307 valid in the hash table. If this does not equal the current
308 reg_tick value, such expressions existing in the hash table are
309 invalid. */
310 int reg_in_table;
311
312 /* The SUBREG that was set when REG_TICK was last incremented. Set
313 to -1 if the last store was to the whole register, not a subreg. */
314 unsigned int subreg_ticked;
315 };
316
317 /* A table of cse_reg_info indexed by register numbers. */
318 static struct cse_reg_info *cse_reg_info_table;
319
320 /* The size of the above table. */
321 static unsigned int cse_reg_info_table_size;
322
323 /* The index of the first entry that has not been initialized. */
324 static unsigned int cse_reg_info_table_first_uninitialized;
325
326 /* The timestamp at the beginning of the current run of
327 cse_extended_basic_block. We increment this variable at the beginning of
328 the current run of cse_extended_basic_block. The timestamp field of a
329 cse_reg_info entry matches the value of this variable if and only
330 if the entry has been initialized during the current run of
331 cse_extended_basic_block. */
332 static unsigned int cse_reg_info_timestamp;
333
334 /* A HARD_REG_SET containing all the hard registers for which there is
335 currently a REG expression in the hash table. Note the difference
336 from the above variables, which indicate if the REG is mentioned in some
337 expression in the table. */
338
339 static HARD_REG_SET hard_regs_in_table;
340
341 /* True if CSE has altered the CFG. */
342 static bool cse_cfg_altered;
343
344 /* True if CSE has altered conditional jump insns in such a way
345 that jump optimization should be redone. */
346 static bool cse_jumps_altered;
347
348 /* True if we put a LABEL_REF into the hash table for an INSN
349 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
350 to put in the note. */
351 static bool recorded_label_ref;
352
353 /* canon_hash stores 1 in do_not_record
354 if it notices a reference to CC0, PC, or some other volatile
355 subexpression. */
356
357 static int do_not_record;
358
359 /* canon_hash stores 1 in hash_arg_in_memory
360 if it notices a reference to memory within the expression being hashed. */
361
362 static int hash_arg_in_memory;
363
364 /* The hash table contains buckets which are chains of `struct table_elt's,
365 each recording one expression's information.
366 That expression is in the `exp' field.
367
368 The canon_exp field contains a canonical (from the point of view of
369 alias analysis) version of the `exp' field.
370
371 Those elements with the same hash code are chained in both directions
372 through the `next_same_hash' and `prev_same_hash' fields.
373
374 Each set of expressions with equivalent values
375 are on a two-way chain through the `next_same_value'
376 and `prev_same_value' fields, and all point with
377 the `first_same_value' field at the first element in
378 that chain. The chain is in order of increasing cost.
379 Each element's cost value is in its `cost' field.
380
381 The `in_memory' field is nonzero for elements that
382 involve any reference to memory. These elements are removed
383 whenever a write is done to an unidentified location in memory.
384 To be safe, we assume that a memory address is unidentified unless
385 the address is either a symbol constant or a constant plus
386 the frame pointer or argument pointer.
387
388 The `related_value' field is used to connect related expressions
389 (that differ by adding an integer).
390 The related expressions are chained in a circular fashion.
391 `related_value' is zero for expressions for which this
392 chain is not useful.
393
394 The `cost' field stores the cost of this element's expression.
395 The `regcost' field stores the value returned by approx_reg_cost for
396 this element's expression.
397
398 The `is_const' flag is set if the element is a constant (including
399 a fixed address).
400
401 The `flag' field is used as a temporary during some search routines.
402
403 The `mode' field is usually the same as GET_MODE (`exp'), but
404 if `exp' is a CONST_INT and has no machine mode then the `mode'
405 field is the mode it was being used as. Each constant is
406 recorded separately for each mode it is used with. */
407
408 struct table_elt
409 {
410 rtx exp;
411 rtx canon_exp;
412 struct table_elt *next_same_hash;
413 struct table_elt *prev_same_hash;
414 struct table_elt *next_same_value;
415 struct table_elt *prev_same_value;
416 struct table_elt *first_same_value;
417 struct table_elt *related_value;
418 int cost;
419 int regcost;
420 /* The size of this field should match the size
421 of the mode field of struct rtx_def (see rtl.h). */
422 ENUM_BITFIELD(machine_mode) mode : 8;
423 char in_memory;
424 char is_const;
425 char flag;
426 };
427
428 /* We don't want a lot of buckets, because we rarely have very many
429 things stored in the hash table, and a lot of buckets slows
430 down a lot of loops that happen frequently. */
431 #define HASH_SHIFT 5
432 #define HASH_SIZE (1 << HASH_SHIFT)
433 #define HASH_MASK (HASH_SIZE - 1)
434
435 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
436 register (hard registers may require `do_not_record' to be set). */
437
438 #define HASH(X, M) \
439 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
440 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
441 : canon_hash (X, M)) & HASH_MASK)
442
443 /* Like HASH, but without side-effects. */
444 #define SAFE_HASH(X, M) \
445 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
446 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
447 : safe_hash (X, M)) & HASH_MASK)
448
449 /* Determine whether register number N is considered a fixed register for the
450 purpose of approximating register costs.
451 It is desirable to replace other regs with fixed regs, to reduce need for
452 non-fixed hard regs.
453 A reg wins if it is either the frame pointer or designated as fixed. */
454 #define FIXED_REGNO_P(N) \
455 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
456 || fixed_regs[N] || global_regs[N])
457
458 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
459 hard registers and pointers into the frame are the cheapest with a cost
460 of 0. Next come pseudos with a cost of one and other hard registers with
461 a cost of 2. Aside from these special cases, call `rtx_cost'. */
462
463 #define CHEAP_REGNO(N) \
464 (REGNO_PTR_FRAME_P (N) \
465 || (HARD_REGISTER_NUM_P (N) \
466 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
467
468 #define COST(X, MODE) \
469 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
470 #define COST_IN(X, MODE, OUTER, OPNO) \
471 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
472
473 /* Get the number of times this register has been updated in this
474 basic block. */
475
476 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
477
478 /* Get the point at which REG was recorded in the table. */
479
480 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
481
482 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
483 SUBREG). */
484
485 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
486
487 /* Get the quantity number for REG. */
488
489 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
490
491 /* Determine if the quantity number for register X represents a valid index
492 into the qty_table. */
493
494 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
495
496 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
497
498 #define CHEAPER(X, Y) \
499 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
500
501 static struct table_elt *table[HASH_SIZE];
502
503 /* Chain of `struct table_elt's made so far for this function
504 but currently removed from the table. */
505
506 static struct table_elt *free_element_chain;
507
508 /* Set to the cost of a constant pool reference if one was found for a
509 symbolic constant. If this was found, it means we should try to
510 convert constants into constant pool entries if they don't fit in
511 the insn. */
512
513 static int constant_pool_entries_cost;
514 static int constant_pool_entries_regcost;
515
516 /* Trace a patch through the CFG. */
517
518 struct branch_path
519 {
520 /* The basic block for this path entry. */
521 basic_block bb;
522 };
523
524 /* This data describes a block that will be processed by
525 cse_extended_basic_block. */
526
527 struct cse_basic_block_data
528 {
529 /* Total number of SETs in block. */
530 int nsets;
531 /* Size of current branch path, if any. */
532 int path_size;
533 /* Current path, indicating which basic_blocks will be processed. */
534 struct branch_path *path;
535 };
536
537
538 /* Pointers to the live in/live out bitmaps for the boundaries of the
539 current EBB. */
540 static bitmap cse_ebb_live_in, cse_ebb_live_out;
541
542 /* A simple bitmap to track which basic blocks have been visited
543 already as part of an already processed extended basic block. */
544 static sbitmap cse_visited_basic_blocks;
545
546 static bool fixed_base_plus_p (rtx x);
547 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
548 static int preferable (int, int, int, int);
549 static void new_basic_block (void);
550 static void make_new_qty (unsigned int, machine_mode);
551 static void make_regs_eqv (unsigned int, unsigned int);
552 static void delete_reg_equiv (unsigned int);
553 static int mention_regs (rtx);
554 static int insert_regs (rtx, struct table_elt *, int);
555 static void remove_from_table (struct table_elt *, unsigned);
556 static void remove_pseudo_from_table (rtx, unsigned);
557 static struct table_elt *lookup (rtx, unsigned, machine_mode);
558 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
559 static rtx lookup_as_function (rtx, enum rtx_code);
560 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
561 machine_mode, int, int);
562 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
563 machine_mode);
564 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
565 static void invalidate (rtx, machine_mode);
566 static void remove_invalid_refs (unsigned int);
567 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
568 machine_mode);
569 static void rehash_using_reg (rtx);
570 static void invalidate_memory (void);
571 static void invalidate_for_call (void);
572 static rtx use_related_value (rtx, struct table_elt *);
573
574 static inline unsigned canon_hash (rtx, machine_mode);
575 static inline unsigned safe_hash (rtx, machine_mode);
576 static inline unsigned hash_rtx_string (const char *);
577
578 static rtx canon_reg (rtx, rtx_insn *);
579 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
580 machine_mode *,
581 machine_mode *);
582 static rtx fold_rtx (rtx, rtx_insn *);
583 static rtx equiv_constant (rtx);
584 static void record_jump_equiv (rtx_insn *, bool);
585 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
586 int);
587 static void cse_insn (rtx_insn *);
588 static void cse_prescan_path (struct cse_basic_block_data *);
589 static void invalidate_from_clobbers (rtx_insn *);
590 static void invalidate_from_sets_and_clobbers (rtx_insn *);
591 static rtx cse_process_notes (rtx, rtx, bool *);
592 static void cse_extended_basic_block (struct cse_basic_block_data *);
593 extern void dump_class (struct table_elt*);
594 static void get_cse_reg_info_1 (unsigned int regno);
595 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
596
597 static void flush_hash_table (void);
598 static bool insn_live_p (rtx_insn *, int *);
599 static bool set_live_p (rtx, rtx_insn *, int *);
600 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
601 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
602 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
603 bool);
604 \f
605
606 #undef RTL_HOOKS_GEN_LOWPART
607 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
608
609 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
610 \f
611 /* Nonzero if X has the form (PLUS frame-pointer integer). */
612
613 static bool
614 fixed_base_plus_p (rtx x)
615 {
616 switch (GET_CODE (x))
617 {
618 case REG:
619 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
620 return true;
621 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
622 return true;
623 return false;
624
625 case PLUS:
626 if (!CONST_INT_P (XEXP (x, 1)))
627 return false;
628 return fixed_base_plus_p (XEXP (x, 0));
629
630 default:
631 return false;
632 }
633 }
634
635 /* Dump the expressions in the equivalence class indicated by CLASSP.
636 This function is used only for debugging. */
637 DEBUG_FUNCTION void
638 dump_class (struct table_elt *classp)
639 {
640 struct table_elt *elt;
641
642 fprintf (stderr, "Equivalence chain for ");
643 print_rtl (stderr, classp->exp);
644 fprintf (stderr, ": \n");
645
646 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
647 {
648 print_rtl (stderr, elt->exp);
649 fprintf (stderr, "\n");
650 }
651 }
652
653 /* Return an estimate of the cost of the registers used in an rtx.
654 This is mostly the number of different REG expressions in the rtx;
655 however for some exceptions like fixed registers we use a cost of
656 0. If any other hard register reference occurs, return MAX_COST. */
657
658 static int
659 approx_reg_cost (const_rtx x)
660 {
661 int cost = 0;
662 subrtx_iterator::array_type array;
663 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
664 {
665 const_rtx x = *iter;
666 if (REG_P (x))
667 {
668 unsigned int regno = REGNO (x);
669 if (!CHEAP_REGNO (regno))
670 {
671 if (regno < FIRST_PSEUDO_REGISTER)
672 {
673 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
674 return MAX_COST;
675 cost += 2;
676 }
677 else
678 cost += 1;
679 }
680 }
681 }
682 return cost;
683 }
684
685 /* Return a negative value if an rtx A, whose costs are given by COST_A
686 and REGCOST_A, is more desirable than an rtx B.
687 Return a positive value if A is less desirable, or 0 if the two are
688 equally good. */
689 static int
690 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
691 {
692 /* First, get rid of cases involving expressions that are entirely
693 unwanted. */
694 if (cost_a != cost_b)
695 {
696 if (cost_a == MAX_COST)
697 return 1;
698 if (cost_b == MAX_COST)
699 return -1;
700 }
701
702 /* Avoid extending lifetimes of hardregs. */
703 if (regcost_a != regcost_b)
704 {
705 if (regcost_a == MAX_COST)
706 return 1;
707 if (regcost_b == MAX_COST)
708 return -1;
709 }
710
711 /* Normal operation costs take precedence. */
712 if (cost_a != cost_b)
713 return cost_a - cost_b;
714 /* Only if these are identical consider effects on register pressure. */
715 if (regcost_a != regcost_b)
716 return regcost_a - regcost_b;
717 return 0;
718 }
719
720 /* Internal function, to compute cost when X is not a register; called
721 from COST macro to keep it simple. */
722
723 static int
724 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
725 {
726 return ((GET_CODE (x) == SUBREG
727 && REG_P (SUBREG_REG (x))
728 && GET_MODE_CLASS (mode) == MODE_INT
729 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
730 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
731 && subreg_lowpart_p (x)
732 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x))))
733 ? 0
734 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
735 }
736
737 \f
738 /* Initialize CSE_REG_INFO_TABLE. */
739
740 static void
741 init_cse_reg_info (unsigned int nregs)
742 {
743 /* Do we need to grow the table? */
744 if (nregs > cse_reg_info_table_size)
745 {
746 unsigned int new_size;
747
748 if (cse_reg_info_table_size < 2048)
749 {
750 /* Compute a new size that is a power of 2 and no smaller
751 than the large of NREGS and 64. */
752 new_size = (cse_reg_info_table_size
753 ? cse_reg_info_table_size : 64);
754
755 while (new_size < nregs)
756 new_size *= 2;
757 }
758 else
759 {
760 /* If we need a big table, allocate just enough to hold
761 NREGS registers. */
762 new_size = nregs;
763 }
764
765 /* Reallocate the table with NEW_SIZE entries. */
766 free (cse_reg_info_table);
767 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
768 cse_reg_info_table_size = new_size;
769 cse_reg_info_table_first_uninitialized = 0;
770 }
771
772 /* Do we have all of the first NREGS entries initialized? */
773 if (cse_reg_info_table_first_uninitialized < nregs)
774 {
775 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
776 unsigned int i;
777
778 /* Put the old timestamp on newly allocated entries so that they
779 will all be considered out of date. We do not touch those
780 entries beyond the first NREGS entries to be nice to the
781 virtual memory. */
782 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
783 cse_reg_info_table[i].timestamp = old_timestamp;
784
785 cse_reg_info_table_first_uninitialized = nregs;
786 }
787 }
788
789 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
790
791 static void
792 get_cse_reg_info_1 (unsigned int regno)
793 {
794 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
795 entry will be considered to have been initialized. */
796 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
797
798 /* Initialize the rest of the entry. */
799 cse_reg_info_table[regno].reg_tick = 1;
800 cse_reg_info_table[regno].reg_in_table = -1;
801 cse_reg_info_table[regno].subreg_ticked = -1;
802 cse_reg_info_table[regno].reg_qty = -regno - 1;
803 }
804
805 /* Find a cse_reg_info entry for REGNO. */
806
807 static inline struct cse_reg_info *
808 get_cse_reg_info (unsigned int regno)
809 {
810 struct cse_reg_info *p = &cse_reg_info_table[regno];
811
812 /* If this entry has not been initialized, go ahead and initialize
813 it. */
814 if (p->timestamp != cse_reg_info_timestamp)
815 get_cse_reg_info_1 (regno);
816
817 return p;
818 }
819
820 /* Clear the hash table and initialize each register with its own quantity,
821 for a new basic block. */
822
823 static void
824 new_basic_block (void)
825 {
826 int i;
827
828 next_qty = 0;
829
830 /* Invalidate cse_reg_info_table. */
831 cse_reg_info_timestamp++;
832
833 /* Clear out hash table state for this pass. */
834 CLEAR_HARD_REG_SET (hard_regs_in_table);
835
836 /* The per-quantity values used to be initialized here, but it is
837 much faster to initialize each as it is made in `make_new_qty'. */
838
839 for (i = 0; i < HASH_SIZE; i++)
840 {
841 struct table_elt *first;
842
843 first = table[i];
844 if (first != NULL)
845 {
846 struct table_elt *last = first;
847
848 table[i] = NULL;
849
850 while (last->next_same_hash != NULL)
851 last = last->next_same_hash;
852
853 /* Now relink this hash entire chain into
854 the free element list. */
855
856 last->next_same_hash = free_element_chain;
857 free_element_chain = first;
858 }
859 }
860
861 prev_insn_cc0 = 0;
862 }
863
864 /* Say that register REG contains a quantity in mode MODE not in any
865 register before and initialize that quantity. */
866
867 static void
868 make_new_qty (unsigned int reg, machine_mode mode)
869 {
870 int q;
871 struct qty_table_elem *ent;
872 struct reg_eqv_elem *eqv;
873
874 gcc_assert (next_qty < max_qty);
875
876 q = REG_QTY (reg) = next_qty++;
877 ent = &qty_table[q];
878 ent->first_reg = reg;
879 ent->last_reg = reg;
880 ent->mode = mode;
881 ent->const_rtx = ent->const_insn = NULL;
882 ent->comparison_code = UNKNOWN;
883
884 eqv = &reg_eqv_table[reg];
885 eqv->next = eqv->prev = -1;
886 }
887
888 /* Make reg NEW equivalent to reg OLD.
889 OLD is not changing; NEW is. */
890
891 static void
892 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
893 {
894 unsigned int lastr, firstr;
895 int q = REG_QTY (old_reg);
896 struct qty_table_elem *ent;
897
898 ent = &qty_table[q];
899
900 /* Nothing should become eqv until it has a "non-invalid" qty number. */
901 gcc_assert (REGNO_QTY_VALID_P (old_reg));
902
903 REG_QTY (new_reg) = q;
904 firstr = ent->first_reg;
905 lastr = ent->last_reg;
906
907 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
908 hard regs. Among pseudos, if NEW will live longer than any other reg
909 of the same qty, and that is beyond the current basic block,
910 make it the new canonical replacement for this qty. */
911 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
912 /* Certain fixed registers might be of the class NO_REGS. This means
913 that not only can they not be allocated by the compiler, but
914 they cannot be used in substitutions or canonicalizations
915 either. */
916 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
917 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
918 || (new_reg >= FIRST_PSEUDO_REGISTER
919 && (firstr < FIRST_PSEUDO_REGISTER
920 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
921 && !bitmap_bit_p (cse_ebb_live_out, firstr))
922 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
923 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
924 {
925 reg_eqv_table[firstr].prev = new_reg;
926 reg_eqv_table[new_reg].next = firstr;
927 reg_eqv_table[new_reg].prev = -1;
928 ent->first_reg = new_reg;
929 }
930 else
931 {
932 /* If NEW is a hard reg (known to be non-fixed), insert at end.
933 Otherwise, insert before any non-fixed hard regs that are at the
934 end. Registers of class NO_REGS cannot be used as an
935 equivalent for anything. */
936 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
937 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
938 && new_reg >= FIRST_PSEUDO_REGISTER)
939 lastr = reg_eqv_table[lastr].prev;
940 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
941 if (reg_eqv_table[lastr].next >= 0)
942 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
943 else
944 qty_table[q].last_reg = new_reg;
945 reg_eqv_table[lastr].next = new_reg;
946 reg_eqv_table[new_reg].prev = lastr;
947 }
948 }
949
950 /* Remove REG from its equivalence class. */
951
952 static void
953 delete_reg_equiv (unsigned int reg)
954 {
955 struct qty_table_elem *ent;
956 int q = REG_QTY (reg);
957 int p, n;
958
959 /* If invalid, do nothing. */
960 if (! REGNO_QTY_VALID_P (reg))
961 return;
962
963 ent = &qty_table[q];
964
965 p = reg_eqv_table[reg].prev;
966 n = reg_eqv_table[reg].next;
967
968 if (n != -1)
969 reg_eqv_table[n].prev = p;
970 else
971 ent->last_reg = p;
972 if (p != -1)
973 reg_eqv_table[p].next = n;
974 else
975 ent->first_reg = n;
976
977 REG_QTY (reg) = -reg - 1;
978 }
979
980 /* Remove any invalid expressions from the hash table
981 that refer to any of the registers contained in expression X.
982
983 Make sure that newly inserted references to those registers
984 as subexpressions will be considered valid.
985
986 mention_regs is not called when a register itself
987 is being stored in the table.
988
989 Return 1 if we have done something that may have changed the hash code
990 of X. */
991
992 static int
993 mention_regs (rtx x)
994 {
995 enum rtx_code code;
996 int i, j;
997 const char *fmt;
998 int changed = 0;
999
1000 if (x == 0)
1001 return 0;
1002
1003 code = GET_CODE (x);
1004 if (code == REG)
1005 {
1006 unsigned int regno = REGNO (x);
1007 unsigned int endregno = END_REGNO (x);
1008 unsigned int i;
1009
1010 for (i = regno; i < endregno; i++)
1011 {
1012 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1013 remove_invalid_refs (i);
1014
1015 REG_IN_TABLE (i) = REG_TICK (i);
1016 SUBREG_TICKED (i) = -1;
1017 }
1018
1019 return 0;
1020 }
1021
1022 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1023 pseudo if they don't use overlapping words. We handle only pseudos
1024 here for simplicity. */
1025 if (code == SUBREG && REG_P (SUBREG_REG (x))
1026 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1027 {
1028 unsigned int i = REGNO (SUBREG_REG (x));
1029
1030 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1031 {
1032 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1033 the last store to this register really stored into this
1034 subreg, then remove the memory of this subreg.
1035 Otherwise, remove any memory of the entire register and
1036 all its subregs from the table. */
1037 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1038 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1039 remove_invalid_refs (i);
1040 else
1041 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1042 }
1043
1044 REG_IN_TABLE (i) = REG_TICK (i);
1045 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1046 return 0;
1047 }
1048
1049 /* If X is a comparison or a COMPARE and either operand is a register
1050 that does not have a quantity, give it one. This is so that a later
1051 call to record_jump_equiv won't cause X to be assigned a different
1052 hash code and not found in the table after that call.
1053
1054 It is not necessary to do this here, since rehash_using_reg can
1055 fix up the table later, but doing this here eliminates the need to
1056 call that expensive function in the most common case where the only
1057 use of the register is in the comparison. */
1058
1059 if (code == COMPARE || COMPARISON_P (x))
1060 {
1061 if (REG_P (XEXP (x, 0))
1062 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1063 if (insert_regs (XEXP (x, 0), NULL, 0))
1064 {
1065 rehash_using_reg (XEXP (x, 0));
1066 changed = 1;
1067 }
1068
1069 if (REG_P (XEXP (x, 1))
1070 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1071 if (insert_regs (XEXP (x, 1), NULL, 0))
1072 {
1073 rehash_using_reg (XEXP (x, 1));
1074 changed = 1;
1075 }
1076 }
1077
1078 fmt = GET_RTX_FORMAT (code);
1079 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1080 if (fmt[i] == 'e')
1081 changed |= mention_regs (XEXP (x, i));
1082 else if (fmt[i] == 'E')
1083 for (j = 0; j < XVECLEN (x, i); j++)
1084 changed |= mention_regs (XVECEXP (x, i, j));
1085
1086 return changed;
1087 }
1088
1089 /* Update the register quantities for inserting X into the hash table
1090 with a value equivalent to CLASSP.
1091 (If the class does not contain a REG, it is irrelevant.)
1092 If MODIFIED is nonzero, X is a destination; it is being modified.
1093 Note that delete_reg_equiv should be called on a register
1094 before insert_regs is done on that register with MODIFIED != 0.
1095
1096 Nonzero value means that elements of reg_qty have changed
1097 so X's hash code may be different. */
1098
1099 static int
1100 insert_regs (rtx x, struct table_elt *classp, int modified)
1101 {
1102 if (REG_P (x))
1103 {
1104 unsigned int regno = REGNO (x);
1105 int qty_valid;
1106
1107 /* If REGNO is in the equivalence table already but is of the
1108 wrong mode for that equivalence, don't do anything here. */
1109
1110 qty_valid = REGNO_QTY_VALID_P (regno);
1111 if (qty_valid)
1112 {
1113 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1114
1115 if (ent->mode != GET_MODE (x))
1116 return 0;
1117 }
1118
1119 if (modified || ! qty_valid)
1120 {
1121 if (classp)
1122 for (classp = classp->first_same_value;
1123 classp != 0;
1124 classp = classp->next_same_value)
1125 if (REG_P (classp->exp)
1126 && GET_MODE (classp->exp) == GET_MODE (x))
1127 {
1128 unsigned c_regno = REGNO (classp->exp);
1129
1130 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1131
1132 /* Suppose that 5 is hard reg and 100 and 101 are
1133 pseudos. Consider
1134
1135 (set (reg:si 100) (reg:si 5))
1136 (set (reg:si 5) (reg:si 100))
1137 (set (reg:di 101) (reg:di 5))
1138
1139 We would now set REG_QTY (101) = REG_QTY (5), but the
1140 entry for 5 is in SImode. When we use this later in
1141 copy propagation, we get the register in wrong mode. */
1142 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1143 continue;
1144
1145 make_regs_eqv (regno, c_regno);
1146 return 1;
1147 }
1148
1149 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1150 than REG_IN_TABLE to find out if there was only a single preceding
1151 invalidation - for the SUBREG - or another one, which would be
1152 for the full register. However, if we find here that REG_TICK
1153 indicates that the register is invalid, it means that it has
1154 been invalidated in a separate operation. The SUBREG might be used
1155 now (then this is a recursive call), or we might use the full REG
1156 now and a SUBREG of it later. So bump up REG_TICK so that
1157 mention_regs will do the right thing. */
1158 if (! modified
1159 && REG_IN_TABLE (regno) >= 0
1160 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1161 REG_TICK (regno)++;
1162 make_new_qty (regno, GET_MODE (x));
1163 return 1;
1164 }
1165
1166 return 0;
1167 }
1168
1169 /* If X is a SUBREG, we will likely be inserting the inner register in the
1170 table. If that register doesn't have an assigned quantity number at
1171 this point but does later, the insertion that we will be doing now will
1172 not be accessible because its hash code will have changed. So assign
1173 a quantity number now. */
1174
1175 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1176 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1177 {
1178 insert_regs (SUBREG_REG (x), NULL, 0);
1179 mention_regs (x);
1180 return 1;
1181 }
1182 else
1183 return mention_regs (x);
1184 }
1185 \f
1186
1187 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1188 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1189 CST is equal to an anchor. */
1190
1191 static bool
1192 compute_const_anchors (rtx cst,
1193 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1194 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1195 {
1196 HOST_WIDE_INT n = INTVAL (cst);
1197
1198 *lower_base = n & ~(targetm.const_anchor - 1);
1199 if (*lower_base == n)
1200 return false;
1201
1202 *upper_base =
1203 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1204 *upper_offs = n - *upper_base;
1205 *lower_offs = n - *lower_base;
1206 return true;
1207 }
1208
1209 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1210
1211 static void
1212 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1213 machine_mode mode)
1214 {
1215 struct table_elt *elt;
1216 unsigned hash;
1217 rtx anchor_exp;
1218 rtx exp;
1219
1220 anchor_exp = GEN_INT (anchor);
1221 hash = HASH (anchor_exp, mode);
1222 elt = lookup (anchor_exp, hash, mode);
1223 if (!elt)
1224 elt = insert (anchor_exp, NULL, hash, mode);
1225
1226 exp = plus_constant (mode, reg, offs);
1227 /* REG has just been inserted and the hash codes recomputed. */
1228 mention_regs (exp);
1229 hash = HASH (exp, mode);
1230
1231 /* Use the cost of the register rather than the whole expression. When
1232 looking up constant anchors we will further offset the corresponding
1233 expression therefore it does not make sense to prefer REGs over
1234 reg-immediate additions. Prefer instead the oldest expression. Also
1235 don't prefer pseudos over hard regs so that we derive constants in
1236 argument registers from other argument registers rather than from the
1237 original pseudo that was used to synthesize the constant. */
1238 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1239 }
1240
1241 /* The constant CST is equivalent to the register REG. Create
1242 equivalences between the two anchors of CST and the corresponding
1243 register-offset expressions using REG. */
1244
1245 static void
1246 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1247 {
1248 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1249
1250 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1251 &upper_base, &upper_offs))
1252 return;
1253
1254 /* Ignore anchors of value 0. Constants accessible from zero are
1255 simple. */
1256 if (lower_base != 0)
1257 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1258
1259 if (upper_base != 0)
1260 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1261 }
1262
1263 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1264 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1265 valid expression. Return the cheapest and oldest of such expressions. In
1266 *OLD, return how old the resulting expression is compared to the other
1267 equivalent expressions. */
1268
1269 static rtx
1270 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1271 unsigned *old)
1272 {
1273 struct table_elt *elt;
1274 unsigned idx;
1275 struct table_elt *match_elt;
1276 rtx match;
1277
1278 /* Find the cheapest and *oldest* expression to maximize the chance of
1279 reusing the same pseudo. */
1280
1281 match_elt = NULL;
1282 match = NULL_RTX;
1283 for (elt = anchor_elt->first_same_value, idx = 0;
1284 elt;
1285 elt = elt->next_same_value, idx++)
1286 {
1287 if (match_elt && CHEAPER (match_elt, elt))
1288 return match;
1289
1290 if (REG_P (elt->exp)
1291 || (GET_CODE (elt->exp) == PLUS
1292 && REG_P (XEXP (elt->exp, 0))
1293 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1294 {
1295 rtx x;
1296
1297 /* Ignore expressions that are no longer valid. */
1298 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1299 continue;
1300
1301 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1302 if (REG_P (x)
1303 || (GET_CODE (x) == PLUS
1304 && IN_RANGE (INTVAL (XEXP (x, 1)),
1305 -targetm.const_anchor,
1306 targetm.const_anchor - 1)))
1307 {
1308 match = x;
1309 match_elt = elt;
1310 *old = idx;
1311 }
1312 }
1313 }
1314
1315 return match;
1316 }
1317
1318 /* Try to express the constant SRC_CONST using a register+offset expression
1319 derived from a constant anchor. Return it if successful or NULL_RTX,
1320 otherwise. */
1321
1322 static rtx
1323 try_const_anchors (rtx src_const, machine_mode mode)
1324 {
1325 struct table_elt *lower_elt, *upper_elt;
1326 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1327 rtx lower_anchor_rtx, upper_anchor_rtx;
1328 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1329 unsigned lower_old, upper_old;
1330
1331 /* CONST_INT is used for CC modes, but we should leave those alone. */
1332 if (GET_MODE_CLASS (mode) == MODE_CC)
1333 return NULL_RTX;
1334
1335 gcc_assert (SCALAR_INT_MODE_P (mode));
1336 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1337 &upper_base, &upper_offs))
1338 return NULL_RTX;
1339
1340 lower_anchor_rtx = GEN_INT (lower_base);
1341 upper_anchor_rtx = GEN_INT (upper_base);
1342 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1343 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1344
1345 if (lower_elt)
1346 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1347 if (upper_elt)
1348 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1349
1350 if (!lower_exp)
1351 return upper_exp;
1352 if (!upper_exp)
1353 return lower_exp;
1354
1355 /* Return the older expression. */
1356 return (upper_old > lower_old ? upper_exp : lower_exp);
1357 }
1358 \f
1359 /* Look in or update the hash table. */
1360
1361 /* Remove table element ELT from use in the table.
1362 HASH is its hash code, made using the HASH macro.
1363 It's an argument because often that is known in advance
1364 and we save much time not recomputing it. */
1365
1366 static void
1367 remove_from_table (struct table_elt *elt, unsigned int hash)
1368 {
1369 if (elt == 0)
1370 return;
1371
1372 /* Mark this element as removed. See cse_insn. */
1373 elt->first_same_value = 0;
1374
1375 /* Remove the table element from its equivalence class. */
1376
1377 {
1378 struct table_elt *prev = elt->prev_same_value;
1379 struct table_elt *next = elt->next_same_value;
1380
1381 if (next)
1382 next->prev_same_value = prev;
1383
1384 if (prev)
1385 prev->next_same_value = next;
1386 else
1387 {
1388 struct table_elt *newfirst = next;
1389 while (next)
1390 {
1391 next->first_same_value = newfirst;
1392 next = next->next_same_value;
1393 }
1394 }
1395 }
1396
1397 /* Remove the table element from its hash bucket. */
1398
1399 {
1400 struct table_elt *prev = elt->prev_same_hash;
1401 struct table_elt *next = elt->next_same_hash;
1402
1403 if (next)
1404 next->prev_same_hash = prev;
1405
1406 if (prev)
1407 prev->next_same_hash = next;
1408 else if (table[hash] == elt)
1409 table[hash] = next;
1410 else
1411 {
1412 /* This entry is not in the proper hash bucket. This can happen
1413 when two classes were merged by `merge_equiv_classes'. Search
1414 for the hash bucket that it heads. This happens only very
1415 rarely, so the cost is acceptable. */
1416 for (hash = 0; hash < HASH_SIZE; hash++)
1417 if (table[hash] == elt)
1418 table[hash] = next;
1419 }
1420 }
1421
1422 /* Remove the table element from its related-value circular chain. */
1423
1424 if (elt->related_value != 0 && elt->related_value != elt)
1425 {
1426 struct table_elt *p = elt->related_value;
1427
1428 while (p->related_value != elt)
1429 p = p->related_value;
1430 p->related_value = elt->related_value;
1431 if (p->related_value == p)
1432 p->related_value = 0;
1433 }
1434
1435 /* Now add it to the free element chain. */
1436 elt->next_same_hash = free_element_chain;
1437 free_element_chain = elt;
1438 }
1439
1440 /* Same as above, but X is a pseudo-register. */
1441
1442 static void
1443 remove_pseudo_from_table (rtx x, unsigned int hash)
1444 {
1445 struct table_elt *elt;
1446
1447 /* Because a pseudo-register can be referenced in more than one
1448 mode, we might have to remove more than one table entry. */
1449 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1450 remove_from_table (elt, hash);
1451 }
1452
1453 /* Look up X in the hash table and return its table element,
1454 or 0 if X is not in the table.
1455
1456 MODE is the machine-mode of X, or if X is an integer constant
1457 with VOIDmode then MODE is the mode with which X will be used.
1458
1459 Here we are satisfied to find an expression whose tree structure
1460 looks like X. */
1461
1462 static struct table_elt *
1463 lookup (rtx x, unsigned int hash, machine_mode mode)
1464 {
1465 struct table_elt *p;
1466
1467 for (p = table[hash]; p; p = p->next_same_hash)
1468 if (mode == p->mode && ((x == p->exp && REG_P (x))
1469 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1470 return p;
1471
1472 return 0;
1473 }
1474
1475 /* Like `lookup' but don't care whether the table element uses invalid regs.
1476 Also ignore discrepancies in the machine mode of a register. */
1477
1478 static struct table_elt *
1479 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1480 {
1481 struct table_elt *p;
1482
1483 if (REG_P (x))
1484 {
1485 unsigned int regno = REGNO (x);
1486
1487 /* Don't check the machine mode when comparing registers;
1488 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1489 for (p = table[hash]; p; p = p->next_same_hash)
1490 if (REG_P (p->exp)
1491 && REGNO (p->exp) == regno)
1492 return p;
1493 }
1494 else
1495 {
1496 for (p = table[hash]; p; p = p->next_same_hash)
1497 if (mode == p->mode
1498 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1499 return p;
1500 }
1501
1502 return 0;
1503 }
1504
1505 /* Look for an expression equivalent to X and with code CODE.
1506 If one is found, return that expression. */
1507
1508 static rtx
1509 lookup_as_function (rtx x, enum rtx_code code)
1510 {
1511 struct table_elt *p
1512 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1513
1514 if (p == 0)
1515 return 0;
1516
1517 for (p = p->first_same_value; p; p = p->next_same_value)
1518 if (GET_CODE (p->exp) == code
1519 /* Make sure this is a valid entry in the table. */
1520 && exp_equiv_p (p->exp, p->exp, 1, false))
1521 return p->exp;
1522
1523 return 0;
1524 }
1525
1526 /* Insert X in the hash table, assuming HASH is its hash code and
1527 CLASSP is an element of the class it should go in (or 0 if a new
1528 class should be made). COST is the code of X and reg_cost is the
1529 cost of registers in X. It is inserted at the proper position to
1530 keep the class in the order cheapest first.
1531
1532 MODE is the machine-mode of X, or if X is an integer constant
1533 with VOIDmode then MODE is the mode with which X will be used.
1534
1535 For elements of equal cheapness, the most recent one
1536 goes in front, except that the first element in the list
1537 remains first unless a cheaper element is added. The order of
1538 pseudo-registers does not matter, as canon_reg will be called to
1539 find the cheapest when a register is retrieved from the table.
1540
1541 The in_memory field in the hash table element is set to 0.
1542 The caller must set it nonzero if appropriate.
1543
1544 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1545 and if insert_regs returns a nonzero value
1546 you must then recompute its hash code before calling here.
1547
1548 If necessary, update table showing constant values of quantities. */
1549
1550 static struct table_elt *
1551 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1552 machine_mode mode, int cost, int reg_cost)
1553 {
1554 struct table_elt *elt;
1555
1556 /* If X is a register and we haven't made a quantity for it,
1557 something is wrong. */
1558 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1559
1560 /* If X is a hard register, show it is being put in the table. */
1561 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1562 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1563
1564 /* Put an element for X into the right hash bucket. */
1565
1566 elt = free_element_chain;
1567 if (elt)
1568 free_element_chain = elt->next_same_hash;
1569 else
1570 elt = XNEW (struct table_elt);
1571
1572 elt->exp = x;
1573 elt->canon_exp = NULL_RTX;
1574 elt->cost = cost;
1575 elt->regcost = reg_cost;
1576 elt->next_same_value = 0;
1577 elt->prev_same_value = 0;
1578 elt->next_same_hash = table[hash];
1579 elt->prev_same_hash = 0;
1580 elt->related_value = 0;
1581 elt->in_memory = 0;
1582 elt->mode = mode;
1583 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1584
1585 if (table[hash])
1586 table[hash]->prev_same_hash = elt;
1587 table[hash] = elt;
1588
1589 /* Put it into the proper value-class. */
1590 if (classp)
1591 {
1592 classp = classp->first_same_value;
1593 if (CHEAPER (elt, classp))
1594 /* Insert at the head of the class. */
1595 {
1596 struct table_elt *p;
1597 elt->next_same_value = classp;
1598 classp->prev_same_value = elt;
1599 elt->first_same_value = elt;
1600
1601 for (p = classp; p; p = p->next_same_value)
1602 p->first_same_value = elt;
1603 }
1604 else
1605 {
1606 /* Insert not at head of the class. */
1607 /* Put it after the last element cheaper than X. */
1608 struct table_elt *p, *next;
1609
1610 for (p = classp;
1611 (next = p->next_same_value) && CHEAPER (next, elt);
1612 p = next)
1613 ;
1614
1615 /* Put it after P and before NEXT. */
1616 elt->next_same_value = next;
1617 if (next)
1618 next->prev_same_value = elt;
1619
1620 elt->prev_same_value = p;
1621 p->next_same_value = elt;
1622 elt->first_same_value = classp;
1623 }
1624 }
1625 else
1626 elt->first_same_value = elt;
1627
1628 /* If this is a constant being set equivalent to a register or a register
1629 being set equivalent to a constant, note the constant equivalence.
1630
1631 If this is a constant, it cannot be equivalent to a different constant,
1632 and a constant is the only thing that can be cheaper than a register. So
1633 we know the register is the head of the class (before the constant was
1634 inserted).
1635
1636 If this is a register that is not already known equivalent to a
1637 constant, we must check the entire class.
1638
1639 If this is a register that is already known equivalent to an insn,
1640 update the qtys `const_insn' to show that `this_insn' is the latest
1641 insn making that quantity equivalent to the constant. */
1642
1643 if (elt->is_const && classp && REG_P (classp->exp)
1644 && !REG_P (x))
1645 {
1646 int exp_q = REG_QTY (REGNO (classp->exp));
1647 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1648
1649 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1650 exp_ent->const_insn = this_insn;
1651 }
1652
1653 else if (REG_P (x)
1654 && classp
1655 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1656 && ! elt->is_const)
1657 {
1658 struct table_elt *p;
1659
1660 for (p = classp; p != 0; p = p->next_same_value)
1661 {
1662 if (p->is_const && !REG_P (p->exp))
1663 {
1664 int x_q = REG_QTY (REGNO (x));
1665 struct qty_table_elem *x_ent = &qty_table[x_q];
1666
1667 x_ent->const_rtx
1668 = gen_lowpart (GET_MODE (x), p->exp);
1669 x_ent->const_insn = this_insn;
1670 break;
1671 }
1672 }
1673 }
1674
1675 else if (REG_P (x)
1676 && qty_table[REG_QTY (REGNO (x))].const_rtx
1677 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1678 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1679
1680 /* If this is a constant with symbolic value,
1681 and it has a term with an explicit integer value,
1682 link it up with related expressions. */
1683 if (GET_CODE (x) == CONST)
1684 {
1685 rtx subexp = get_related_value (x);
1686 unsigned subhash;
1687 struct table_elt *subelt, *subelt_prev;
1688
1689 if (subexp != 0)
1690 {
1691 /* Get the integer-free subexpression in the hash table. */
1692 subhash = SAFE_HASH (subexp, mode);
1693 subelt = lookup (subexp, subhash, mode);
1694 if (subelt == 0)
1695 subelt = insert (subexp, NULL, subhash, mode);
1696 /* Initialize SUBELT's circular chain if it has none. */
1697 if (subelt->related_value == 0)
1698 subelt->related_value = subelt;
1699 /* Find the element in the circular chain that precedes SUBELT. */
1700 subelt_prev = subelt;
1701 while (subelt_prev->related_value != subelt)
1702 subelt_prev = subelt_prev->related_value;
1703 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1704 This way the element that follows SUBELT is the oldest one. */
1705 elt->related_value = subelt_prev->related_value;
1706 subelt_prev->related_value = elt;
1707 }
1708 }
1709
1710 return elt;
1711 }
1712
1713 /* Wrap insert_with_costs by passing the default costs. */
1714
1715 static struct table_elt *
1716 insert (rtx x, struct table_elt *classp, unsigned int hash,
1717 machine_mode mode)
1718 {
1719 return insert_with_costs (x, classp, hash, mode,
1720 COST (x, mode), approx_reg_cost (x));
1721 }
1722
1723 \f
1724 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1725 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1726 the two classes equivalent.
1727
1728 CLASS1 will be the surviving class; CLASS2 should not be used after this
1729 call.
1730
1731 Any invalid entries in CLASS2 will not be copied. */
1732
1733 static void
1734 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1735 {
1736 struct table_elt *elt, *next, *new_elt;
1737
1738 /* Ensure we start with the head of the classes. */
1739 class1 = class1->first_same_value;
1740 class2 = class2->first_same_value;
1741
1742 /* If they were already equal, forget it. */
1743 if (class1 == class2)
1744 return;
1745
1746 for (elt = class2; elt; elt = next)
1747 {
1748 unsigned int hash;
1749 rtx exp = elt->exp;
1750 machine_mode mode = elt->mode;
1751
1752 next = elt->next_same_value;
1753
1754 /* Remove old entry, make a new one in CLASS1's class.
1755 Don't do this for invalid entries as we cannot find their
1756 hash code (it also isn't necessary). */
1757 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1758 {
1759 bool need_rehash = false;
1760
1761 hash_arg_in_memory = 0;
1762 hash = HASH (exp, mode);
1763
1764 if (REG_P (exp))
1765 {
1766 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1767 delete_reg_equiv (REGNO (exp));
1768 }
1769
1770 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1771 remove_pseudo_from_table (exp, hash);
1772 else
1773 remove_from_table (elt, hash);
1774
1775 if (insert_regs (exp, class1, 0) || need_rehash)
1776 {
1777 rehash_using_reg (exp);
1778 hash = HASH (exp, mode);
1779 }
1780 new_elt = insert (exp, class1, hash, mode);
1781 new_elt->in_memory = hash_arg_in_memory;
1782 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1783 new_elt->cost = MAX_COST;
1784 }
1785 }
1786 }
1787 \f
1788 /* Flush the entire hash table. */
1789
1790 static void
1791 flush_hash_table (void)
1792 {
1793 int i;
1794 struct table_elt *p;
1795
1796 for (i = 0; i < HASH_SIZE; i++)
1797 for (p = table[i]; p; p = table[i])
1798 {
1799 /* Note that invalidate can remove elements
1800 after P in the current hash chain. */
1801 if (REG_P (p->exp))
1802 invalidate (p->exp, VOIDmode);
1803 else
1804 remove_from_table (p, i);
1805 }
1806 }
1807 \f
1808 /* Check whether an anti dependence exists between X and EXP. MODE and
1809 ADDR are as for canon_anti_dependence. */
1810
1811 static bool
1812 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1813 {
1814 subrtx_iterator::array_type array;
1815 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1816 {
1817 const_rtx x = *iter;
1818 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1819 return true;
1820 }
1821 return false;
1822 }
1823 \f
1824 /* Remove from the hash table, or mark as invalid, all expressions whose
1825 values could be altered by storing in X. X is a register, a subreg, or
1826 a memory reference with nonvarying address (because, when a memory
1827 reference with a varying address is stored in, all memory references are
1828 removed by invalidate_memory so specific invalidation is superfluous).
1829 FULL_MODE, if not VOIDmode, indicates that this much should be
1830 invalidated instead of just the amount indicated by the mode of X. This
1831 is only used for bitfield stores into memory.
1832
1833 A nonvarying address may be just a register or just a symbol reference,
1834 or it may be either of those plus a numeric offset. */
1835
1836 static void
1837 invalidate (rtx x, machine_mode full_mode)
1838 {
1839 int i;
1840 struct table_elt *p;
1841 rtx addr;
1842
1843 switch (GET_CODE (x))
1844 {
1845 case REG:
1846 {
1847 /* If X is a register, dependencies on its contents are recorded
1848 through the qty number mechanism. Just change the qty number of
1849 the register, mark it as invalid for expressions that refer to it,
1850 and remove it itself. */
1851 unsigned int regno = REGNO (x);
1852 unsigned int hash = HASH (x, GET_MODE (x));
1853
1854 /* Remove REGNO from any quantity list it might be on and indicate
1855 that its value might have changed. If it is a pseudo, remove its
1856 entry from the hash table.
1857
1858 For a hard register, we do the first two actions above for any
1859 additional hard registers corresponding to X. Then, if any of these
1860 registers are in the table, we must remove any REG entries that
1861 overlap these registers. */
1862
1863 delete_reg_equiv (regno);
1864 REG_TICK (regno)++;
1865 SUBREG_TICKED (regno) = -1;
1866
1867 if (regno >= FIRST_PSEUDO_REGISTER)
1868 remove_pseudo_from_table (x, hash);
1869 else
1870 {
1871 HOST_WIDE_INT in_table
1872 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1873 unsigned int endregno = END_REGNO (x);
1874 unsigned int tregno, tendregno, rn;
1875 struct table_elt *p, *next;
1876
1877 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1878
1879 for (rn = regno + 1; rn < endregno; rn++)
1880 {
1881 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1882 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1883 delete_reg_equiv (rn);
1884 REG_TICK (rn)++;
1885 SUBREG_TICKED (rn) = -1;
1886 }
1887
1888 if (in_table)
1889 for (hash = 0; hash < HASH_SIZE; hash++)
1890 for (p = table[hash]; p; p = next)
1891 {
1892 next = p->next_same_hash;
1893
1894 if (!REG_P (p->exp)
1895 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1896 continue;
1897
1898 tregno = REGNO (p->exp);
1899 tendregno = END_REGNO (p->exp);
1900 if (tendregno > regno && tregno < endregno)
1901 remove_from_table (p, hash);
1902 }
1903 }
1904 }
1905 return;
1906
1907 case SUBREG:
1908 invalidate (SUBREG_REG (x), VOIDmode);
1909 return;
1910
1911 case PARALLEL:
1912 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1913 invalidate (XVECEXP (x, 0, i), VOIDmode);
1914 return;
1915
1916 case EXPR_LIST:
1917 /* This is part of a disjoint return value; extract the location in
1918 question ignoring the offset. */
1919 invalidate (XEXP (x, 0), VOIDmode);
1920 return;
1921
1922 case MEM:
1923 addr = canon_rtx (get_addr (XEXP (x, 0)));
1924 /* Calculate the canonical version of X here so that
1925 true_dependence doesn't generate new RTL for X on each call. */
1926 x = canon_rtx (x);
1927
1928 /* Remove all hash table elements that refer to overlapping pieces of
1929 memory. */
1930 if (full_mode == VOIDmode)
1931 full_mode = GET_MODE (x);
1932
1933 for (i = 0; i < HASH_SIZE; i++)
1934 {
1935 struct table_elt *next;
1936
1937 for (p = table[i]; p; p = next)
1938 {
1939 next = p->next_same_hash;
1940 if (p->in_memory)
1941 {
1942 /* Just canonicalize the expression once;
1943 otherwise each time we call invalidate
1944 true_dependence will canonicalize the
1945 expression again. */
1946 if (!p->canon_exp)
1947 p->canon_exp = canon_rtx (p->exp);
1948 if (check_dependence (p->canon_exp, x, full_mode, addr))
1949 remove_from_table (p, i);
1950 }
1951 }
1952 }
1953 return;
1954
1955 default:
1956 gcc_unreachable ();
1957 }
1958 }
1959
1960 /* Invalidate DEST. Used when DEST is not going to be added
1961 into the hash table for some reason, e.g. do_not_record
1962 flagged on it. */
1963
1964 static void
1965 invalidate_dest (rtx dest)
1966 {
1967 if (REG_P (dest)
1968 || GET_CODE (dest) == SUBREG
1969 || MEM_P (dest))
1970 invalidate (dest, VOIDmode);
1971 else if (GET_CODE (dest) == STRICT_LOW_PART
1972 || GET_CODE (dest) == ZERO_EXTRACT)
1973 invalidate (XEXP (dest, 0), GET_MODE (dest));
1974 }
1975 \f
1976 /* Remove all expressions that refer to register REGNO,
1977 since they are already invalid, and we are about to
1978 mark that register valid again and don't want the old
1979 expressions to reappear as valid. */
1980
1981 static void
1982 remove_invalid_refs (unsigned int regno)
1983 {
1984 unsigned int i;
1985 struct table_elt *p, *next;
1986
1987 for (i = 0; i < HASH_SIZE; i++)
1988 for (p = table[i]; p; p = next)
1989 {
1990 next = p->next_same_hash;
1991 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
1992 remove_from_table (p, i);
1993 }
1994 }
1995
1996 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
1997 and mode MODE. */
1998 static void
1999 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2000 machine_mode mode)
2001 {
2002 unsigned int i;
2003 struct table_elt *p, *next;
2004 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2005
2006 for (i = 0; i < HASH_SIZE; i++)
2007 for (p = table[i]; p; p = next)
2008 {
2009 rtx exp = p->exp;
2010 next = p->next_same_hash;
2011
2012 if (!REG_P (exp)
2013 && (GET_CODE (exp) != SUBREG
2014 || !REG_P (SUBREG_REG (exp))
2015 || REGNO (SUBREG_REG (exp)) != regno
2016 || (((SUBREG_BYTE (exp)
2017 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2018 && SUBREG_BYTE (exp) <= end))
2019 && refers_to_regno_p (regno, p->exp))
2020 remove_from_table (p, i);
2021 }
2022 }
2023 \f
2024 /* Recompute the hash codes of any valid entries in the hash table that
2025 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2026
2027 This is called when we make a jump equivalence. */
2028
2029 static void
2030 rehash_using_reg (rtx x)
2031 {
2032 unsigned int i;
2033 struct table_elt *p, *next;
2034 unsigned hash;
2035
2036 if (GET_CODE (x) == SUBREG)
2037 x = SUBREG_REG (x);
2038
2039 /* If X is not a register or if the register is known not to be in any
2040 valid entries in the table, we have no work to do. */
2041
2042 if (!REG_P (x)
2043 || REG_IN_TABLE (REGNO (x)) < 0
2044 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2045 return;
2046
2047 /* Scan all hash chains looking for valid entries that mention X.
2048 If we find one and it is in the wrong hash chain, move it. */
2049
2050 for (i = 0; i < HASH_SIZE; i++)
2051 for (p = table[i]; p; p = next)
2052 {
2053 next = p->next_same_hash;
2054 if (reg_mentioned_p (x, p->exp)
2055 && exp_equiv_p (p->exp, p->exp, 1, false)
2056 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2057 {
2058 if (p->next_same_hash)
2059 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2060
2061 if (p->prev_same_hash)
2062 p->prev_same_hash->next_same_hash = p->next_same_hash;
2063 else
2064 table[i] = p->next_same_hash;
2065
2066 p->next_same_hash = table[hash];
2067 p->prev_same_hash = 0;
2068 if (table[hash])
2069 table[hash]->prev_same_hash = p;
2070 table[hash] = p;
2071 }
2072 }
2073 }
2074 \f
2075 /* Remove from the hash table any expression that is a call-clobbered
2076 register. Also update their TICK values. */
2077
2078 static void
2079 invalidate_for_call (void)
2080 {
2081 unsigned int regno, endregno;
2082 unsigned int i;
2083 unsigned hash;
2084 struct table_elt *p, *next;
2085 int in_table = 0;
2086 hard_reg_set_iterator hrsi;
2087
2088 /* Go through all the hard registers. For each that is clobbered in
2089 a CALL_INSN, remove the register from quantity chains and update
2090 reg_tick if defined. Also see if any of these registers is currently
2091 in the table. */
2092 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2093 {
2094 delete_reg_equiv (regno);
2095 if (REG_TICK (regno) >= 0)
2096 {
2097 REG_TICK (regno)++;
2098 SUBREG_TICKED (regno) = -1;
2099 }
2100 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2101 }
2102
2103 /* In the case where we have no call-clobbered hard registers in the
2104 table, we are done. Otherwise, scan the table and remove any
2105 entry that overlaps a call-clobbered register. */
2106
2107 if (in_table)
2108 for (hash = 0; hash < HASH_SIZE; hash++)
2109 for (p = table[hash]; p; p = next)
2110 {
2111 next = p->next_same_hash;
2112
2113 if (!REG_P (p->exp)
2114 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2115 continue;
2116
2117 regno = REGNO (p->exp);
2118 endregno = END_REGNO (p->exp);
2119
2120 for (i = regno; i < endregno; i++)
2121 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2122 {
2123 remove_from_table (p, hash);
2124 break;
2125 }
2126 }
2127 }
2128 \f
2129 /* Given an expression X of type CONST,
2130 and ELT which is its table entry (or 0 if it
2131 is not in the hash table),
2132 return an alternate expression for X as a register plus integer.
2133 If none can be found, return 0. */
2134
2135 static rtx
2136 use_related_value (rtx x, struct table_elt *elt)
2137 {
2138 struct table_elt *relt = 0;
2139 struct table_elt *p, *q;
2140 HOST_WIDE_INT offset;
2141
2142 /* First, is there anything related known?
2143 If we have a table element, we can tell from that.
2144 Otherwise, must look it up. */
2145
2146 if (elt != 0 && elt->related_value != 0)
2147 relt = elt;
2148 else if (elt == 0 && GET_CODE (x) == CONST)
2149 {
2150 rtx subexp = get_related_value (x);
2151 if (subexp != 0)
2152 relt = lookup (subexp,
2153 SAFE_HASH (subexp, GET_MODE (subexp)),
2154 GET_MODE (subexp));
2155 }
2156
2157 if (relt == 0)
2158 return 0;
2159
2160 /* Search all related table entries for one that has an
2161 equivalent register. */
2162
2163 p = relt;
2164 while (1)
2165 {
2166 /* This loop is strange in that it is executed in two different cases.
2167 The first is when X is already in the table. Then it is searching
2168 the RELATED_VALUE list of X's class (RELT). The second case is when
2169 X is not in the table. Then RELT points to a class for the related
2170 value.
2171
2172 Ensure that, whatever case we are in, that we ignore classes that have
2173 the same value as X. */
2174
2175 if (rtx_equal_p (x, p->exp))
2176 q = 0;
2177 else
2178 for (q = p->first_same_value; q; q = q->next_same_value)
2179 if (REG_P (q->exp))
2180 break;
2181
2182 if (q)
2183 break;
2184
2185 p = p->related_value;
2186
2187 /* We went all the way around, so there is nothing to be found.
2188 Alternatively, perhaps RELT was in the table for some other reason
2189 and it has no related values recorded. */
2190 if (p == relt || p == 0)
2191 break;
2192 }
2193
2194 if (q == 0)
2195 return 0;
2196
2197 offset = (get_integer_term (x) - get_integer_term (p->exp));
2198 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2199 return plus_constant (q->mode, q->exp, offset);
2200 }
2201 \f
2202
2203 /* Hash a string. Just add its bytes up. */
2204 static inline unsigned
2205 hash_rtx_string (const char *ps)
2206 {
2207 unsigned hash = 0;
2208 const unsigned char *p = (const unsigned char *) ps;
2209
2210 if (p)
2211 while (*p)
2212 hash += *p++;
2213
2214 return hash;
2215 }
2216
2217 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2218 When the callback returns true, we continue with the new rtx. */
2219
2220 unsigned
2221 hash_rtx_cb (const_rtx x, machine_mode mode,
2222 int *do_not_record_p, int *hash_arg_in_memory_p,
2223 bool have_reg_qty, hash_rtx_callback_function cb)
2224 {
2225 int i, j;
2226 unsigned hash = 0;
2227 enum rtx_code code;
2228 const char *fmt;
2229 machine_mode newmode;
2230 rtx newx;
2231
2232 /* Used to turn recursion into iteration. We can't rely on GCC's
2233 tail-recursion elimination since we need to keep accumulating values
2234 in HASH. */
2235 repeat:
2236 if (x == 0)
2237 return hash;
2238
2239 /* Invoke the callback first. */
2240 if (cb != NULL
2241 && ((*cb) (x, mode, &newx, &newmode)))
2242 {
2243 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2244 hash_arg_in_memory_p, have_reg_qty, cb);
2245 return hash;
2246 }
2247
2248 code = GET_CODE (x);
2249 switch (code)
2250 {
2251 case REG:
2252 {
2253 unsigned int regno = REGNO (x);
2254
2255 if (do_not_record_p && !reload_completed)
2256 {
2257 /* On some machines, we can't record any non-fixed hard register,
2258 because extending its life will cause reload problems. We
2259 consider ap, fp, sp, gp to be fixed for this purpose.
2260
2261 We also consider CCmode registers to be fixed for this purpose;
2262 failure to do so leads to failure to simplify 0<100 type of
2263 conditionals.
2264
2265 On all machines, we can't record any global registers.
2266 Nor should we record any register that is in a small
2267 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2268 bool record;
2269
2270 if (regno >= FIRST_PSEUDO_REGISTER)
2271 record = true;
2272 else if (x == frame_pointer_rtx
2273 || x == hard_frame_pointer_rtx
2274 || x == arg_pointer_rtx
2275 || x == stack_pointer_rtx
2276 || x == pic_offset_table_rtx)
2277 record = true;
2278 else if (global_regs[regno])
2279 record = false;
2280 else if (fixed_regs[regno])
2281 record = true;
2282 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2283 record = true;
2284 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2285 record = false;
2286 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2287 record = false;
2288 else
2289 record = true;
2290
2291 if (!record)
2292 {
2293 *do_not_record_p = 1;
2294 return 0;
2295 }
2296 }
2297
2298 hash += ((unsigned int) REG << 7);
2299 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2300 return hash;
2301 }
2302
2303 /* We handle SUBREG of a REG specially because the underlying
2304 reg changes its hash value with every value change; we don't
2305 want to have to forget unrelated subregs when one subreg changes. */
2306 case SUBREG:
2307 {
2308 if (REG_P (SUBREG_REG (x)))
2309 {
2310 hash += (((unsigned int) SUBREG << 7)
2311 + REGNO (SUBREG_REG (x))
2312 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2313 return hash;
2314 }
2315 break;
2316 }
2317
2318 case CONST_INT:
2319 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2320 + (unsigned int) INTVAL (x));
2321 return hash;
2322
2323 case CONST_WIDE_INT:
2324 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2325 hash += CONST_WIDE_INT_ELT (x, i);
2326 return hash;
2327
2328 case CONST_DOUBLE:
2329 /* This is like the general case, except that it only counts
2330 the integers representing the constant. */
2331 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2332 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2333 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2334 + (unsigned int) CONST_DOUBLE_HIGH (x));
2335 else
2336 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2337 return hash;
2338
2339 case CONST_FIXED:
2340 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2341 hash += fixed_hash (CONST_FIXED_VALUE (x));
2342 return hash;
2343
2344 case CONST_VECTOR:
2345 {
2346 int units;
2347 rtx elt;
2348
2349 units = CONST_VECTOR_NUNITS (x);
2350
2351 for (i = 0; i < units; ++i)
2352 {
2353 elt = CONST_VECTOR_ELT (x, i);
2354 hash += hash_rtx_cb (elt, GET_MODE (elt),
2355 do_not_record_p, hash_arg_in_memory_p,
2356 have_reg_qty, cb);
2357 }
2358
2359 return hash;
2360 }
2361
2362 /* Assume there is only one rtx object for any given label. */
2363 case LABEL_REF:
2364 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2365 differences and differences between each stage's debugging dumps. */
2366 hash += (((unsigned int) LABEL_REF << 7)
2367 + CODE_LABEL_NUMBER (LABEL_REF_LABEL (x)));
2368 return hash;
2369
2370 case SYMBOL_REF:
2371 {
2372 /* Don't hash on the symbol's address to avoid bootstrap differences.
2373 Different hash values may cause expressions to be recorded in
2374 different orders and thus different registers to be used in the
2375 final assembler. This also avoids differences in the dump files
2376 between various stages. */
2377 unsigned int h = 0;
2378 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2379
2380 while (*p)
2381 h += (h << 7) + *p++; /* ??? revisit */
2382
2383 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2384 return hash;
2385 }
2386
2387 case MEM:
2388 /* We don't record if marked volatile or if BLKmode since we don't
2389 know the size of the move. */
2390 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2391 {
2392 *do_not_record_p = 1;
2393 return 0;
2394 }
2395 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2396 *hash_arg_in_memory_p = 1;
2397
2398 /* Now that we have already found this special case,
2399 might as well speed it up as much as possible. */
2400 hash += (unsigned) MEM;
2401 x = XEXP (x, 0);
2402 goto repeat;
2403
2404 case USE:
2405 /* A USE that mentions non-volatile memory needs special
2406 handling since the MEM may be BLKmode which normally
2407 prevents an entry from being made. Pure calls are
2408 marked by a USE which mentions BLKmode memory.
2409 See calls.c:emit_call_1. */
2410 if (MEM_P (XEXP (x, 0))
2411 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2412 {
2413 hash += (unsigned) USE;
2414 x = XEXP (x, 0);
2415
2416 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2417 *hash_arg_in_memory_p = 1;
2418
2419 /* Now that we have already found this special case,
2420 might as well speed it up as much as possible. */
2421 hash += (unsigned) MEM;
2422 x = XEXP (x, 0);
2423 goto repeat;
2424 }
2425 break;
2426
2427 case PRE_DEC:
2428 case PRE_INC:
2429 case POST_DEC:
2430 case POST_INC:
2431 case PRE_MODIFY:
2432 case POST_MODIFY:
2433 case PC:
2434 case CC0:
2435 case CALL:
2436 case UNSPEC_VOLATILE:
2437 if (do_not_record_p) {
2438 *do_not_record_p = 1;
2439 return 0;
2440 }
2441 else
2442 return hash;
2443 break;
2444
2445 case ASM_OPERANDS:
2446 if (do_not_record_p && MEM_VOLATILE_P (x))
2447 {
2448 *do_not_record_p = 1;
2449 return 0;
2450 }
2451 else
2452 {
2453 /* We don't want to take the filename and line into account. */
2454 hash += (unsigned) code + (unsigned) GET_MODE (x)
2455 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2456 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2457 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2458
2459 if (ASM_OPERANDS_INPUT_LENGTH (x))
2460 {
2461 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2462 {
2463 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2464 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2465 do_not_record_p, hash_arg_in_memory_p,
2466 have_reg_qty, cb)
2467 + hash_rtx_string
2468 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2469 }
2470
2471 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2472 x = ASM_OPERANDS_INPUT (x, 0);
2473 mode = GET_MODE (x);
2474 goto repeat;
2475 }
2476
2477 return hash;
2478 }
2479 break;
2480
2481 default:
2482 break;
2483 }
2484
2485 i = GET_RTX_LENGTH (code) - 1;
2486 hash += (unsigned) code + (unsigned) GET_MODE (x);
2487 fmt = GET_RTX_FORMAT (code);
2488 for (; i >= 0; i--)
2489 {
2490 switch (fmt[i])
2491 {
2492 case 'e':
2493 /* If we are about to do the last recursive call
2494 needed at this level, change it into iteration.
2495 This function is called enough to be worth it. */
2496 if (i == 0)
2497 {
2498 x = XEXP (x, i);
2499 goto repeat;
2500 }
2501
2502 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2503 hash_arg_in_memory_p,
2504 have_reg_qty, cb);
2505 break;
2506
2507 case 'E':
2508 for (j = 0; j < XVECLEN (x, i); j++)
2509 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2510 hash_arg_in_memory_p,
2511 have_reg_qty, cb);
2512 break;
2513
2514 case 's':
2515 hash += hash_rtx_string (XSTR (x, i));
2516 break;
2517
2518 case 'i':
2519 hash += (unsigned int) XINT (x, i);
2520 break;
2521
2522 case '0': case 't':
2523 /* Unused. */
2524 break;
2525
2526 default:
2527 gcc_unreachable ();
2528 }
2529 }
2530
2531 return hash;
2532 }
2533
2534 /* Hash an rtx. We are careful to make sure the value is never negative.
2535 Equivalent registers hash identically.
2536 MODE is used in hashing for CONST_INTs only;
2537 otherwise the mode of X is used.
2538
2539 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2540
2541 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2542 a MEM rtx which does not have the MEM_READONLY_P flag set.
2543
2544 Note that cse_insn knows that the hash code of a MEM expression
2545 is just (int) MEM plus the hash code of the address. */
2546
2547 unsigned
2548 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2549 int *hash_arg_in_memory_p, bool have_reg_qty)
2550 {
2551 return hash_rtx_cb (x, mode, do_not_record_p,
2552 hash_arg_in_memory_p, have_reg_qty, NULL);
2553 }
2554
2555 /* Hash an rtx X for cse via hash_rtx.
2556 Stores 1 in do_not_record if any subexpression is volatile.
2557 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2558 does not have the MEM_READONLY_P flag set. */
2559
2560 static inline unsigned
2561 canon_hash (rtx x, machine_mode mode)
2562 {
2563 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2564 }
2565
2566 /* Like canon_hash but with no side effects, i.e. do_not_record
2567 and hash_arg_in_memory are not changed. */
2568
2569 static inline unsigned
2570 safe_hash (rtx x, machine_mode mode)
2571 {
2572 int dummy_do_not_record;
2573 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2574 }
2575 \f
2576 /* Return 1 iff X and Y would canonicalize into the same thing,
2577 without actually constructing the canonicalization of either one.
2578 If VALIDATE is nonzero,
2579 we assume X is an expression being processed from the rtl
2580 and Y was found in the hash table. We check register refs
2581 in Y for being marked as valid.
2582
2583 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2584
2585 int
2586 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2587 {
2588 int i, j;
2589 enum rtx_code code;
2590 const char *fmt;
2591
2592 /* Note: it is incorrect to assume an expression is equivalent to itself
2593 if VALIDATE is nonzero. */
2594 if (x == y && !validate)
2595 return 1;
2596
2597 if (x == 0 || y == 0)
2598 return x == y;
2599
2600 code = GET_CODE (x);
2601 if (code != GET_CODE (y))
2602 return 0;
2603
2604 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2605 if (GET_MODE (x) != GET_MODE (y))
2606 return 0;
2607
2608 /* MEMs referring to different address space are not equivalent. */
2609 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2610 return 0;
2611
2612 switch (code)
2613 {
2614 case PC:
2615 case CC0:
2616 CASE_CONST_UNIQUE:
2617 return x == y;
2618
2619 case LABEL_REF:
2620 return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y);
2621
2622 case SYMBOL_REF:
2623 return XSTR (x, 0) == XSTR (y, 0);
2624
2625 case REG:
2626 if (for_gcse)
2627 return REGNO (x) == REGNO (y);
2628 else
2629 {
2630 unsigned int regno = REGNO (y);
2631 unsigned int i;
2632 unsigned int endregno = END_REGNO (y);
2633
2634 /* If the quantities are not the same, the expressions are not
2635 equivalent. If there are and we are not to validate, they
2636 are equivalent. Otherwise, ensure all regs are up-to-date. */
2637
2638 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2639 return 0;
2640
2641 if (! validate)
2642 return 1;
2643
2644 for (i = regno; i < endregno; i++)
2645 if (REG_IN_TABLE (i) != REG_TICK (i))
2646 return 0;
2647
2648 return 1;
2649 }
2650
2651 case MEM:
2652 if (for_gcse)
2653 {
2654 /* A volatile mem should not be considered equivalent to any
2655 other. */
2656 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2657 return 0;
2658
2659 /* Can't merge two expressions in different alias sets, since we
2660 can decide that the expression is transparent in a block when
2661 it isn't, due to it being set with the different alias set.
2662
2663 Also, can't merge two expressions with different MEM_ATTRS.
2664 They could e.g. be two different entities allocated into the
2665 same space on the stack (see e.g. PR25130). In that case, the
2666 MEM addresses can be the same, even though the two MEMs are
2667 absolutely not equivalent.
2668
2669 But because really all MEM attributes should be the same for
2670 equivalent MEMs, we just use the invariant that MEMs that have
2671 the same attributes share the same mem_attrs data structure. */
2672 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2673 return 0;
2674
2675 /* If we are handling exceptions, we cannot consider two expressions
2676 with different trapping status as equivalent, because simple_mem
2677 might accept one and reject the other. */
2678 if (cfun->can_throw_non_call_exceptions
2679 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2680 return 0;
2681 }
2682 break;
2683
2684 /* For commutative operations, check both orders. */
2685 case PLUS:
2686 case MULT:
2687 case AND:
2688 case IOR:
2689 case XOR:
2690 case NE:
2691 case EQ:
2692 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2693 validate, for_gcse)
2694 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2695 validate, for_gcse))
2696 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2697 validate, for_gcse)
2698 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2699 validate, for_gcse)));
2700
2701 case ASM_OPERANDS:
2702 /* We don't use the generic code below because we want to
2703 disregard filename and line numbers. */
2704
2705 /* A volatile asm isn't equivalent to any other. */
2706 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2707 return 0;
2708
2709 if (GET_MODE (x) != GET_MODE (y)
2710 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2711 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2712 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2713 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2714 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2715 return 0;
2716
2717 if (ASM_OPERANDS_INPUT_LENGTH (x))
2718 {
2719 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2720 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2721 ASM_OPERANDS_INPUT (y, i),
2722 validate, for_gcse)
2723 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2724 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2725 return 0;
2726 }
2727
2728 return 1;
2729
2730 default:
2731 break;
2732 }
2733
2734 /* Compare the elements. If any pair of corresponding elements
2735 fail to match, return 0 for the whole thing. */
2736
2737 fmt = GET_RTX_FORMAT (code);
2738 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2739 {
2740 switch (fmt[i])
2741 {
2742 case 'e':
2743 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2744 validate, for_gcse))
2745 return 0;
2746 break;
2747
2748 case 'E':
2749 if (XVECLEN (x, i) != XVECLEN (y, i))
2750 return 0;
2751 for (j = 0; j < XVECLEN (x, i); j++)
2752 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2753 validate, for_gcse))
2754 return 0;
2755 break;
2756
2757 case 's':
2758 if (strcmp (XSTR (x, i), XSTR (y, i)))
2759 return 0;
2760 break;
2761
2762 case 'i':
2763 if (XINT (x, i) != XINT (y, i))
2764 return 0;
2765 break;
2766
2767 case 'w':
2768 if (XWINT (x, i) != XWINT (y, i))
2769 return 0;
2770 break;
2771
2772 case '0':
2773 case 't':
2774 break;
2775
2776 default:
2777 gcc_unreachable ();
2778 }
2779 }
2780
2781 return 1;
2782 }
2783 \f
2784 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2785 the result if necessary. INSN is as for canon_reg. */
2786
2787 static void
2788 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2789 {
2790 if (*xloc)
2791 {
2792 rtx new_rtx = canon_reg (*xloc, insn);
2793
2794 /* If replacing pseudo with hard reg or vice versa, ensure the
2795 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2796 gcc_assert (insn && new_rtx);
2797 validate_change (insn, xloc, new_rtx, 1);
2798 }
2799 }
2800
2801 /* Canonicalize an expression:
2802 replace each register reference inside it
2803 with the "oldest" equivalent register.
2804
2805 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2806 after we make our substitution. The calls are made with IN_GROUP nonzero
2807 so apply_change_group must be called upon the outermost return from this
2808 function (unless INSN is zero). The result of apply_change_group can
2809 generally be discarded since the changes we are making are optional. */
2810
2811 static rtx
2812 canon_reg (rtx x, rtx_insn *insn)
2813 {
2814 int i;
2815 enum rtx_code code;
2816 const char *fmt;
2817
2818 if (x == 0)
2819 return x;
2820
2821 code = GET_CODE (x);
2822 switch (code)
2823 {
2824 case PC:
2825 case CC0:
2826 case CONST:
2827 CASE_CONST_ANY:
2828 case SYMBOL_REF:
2829 case LABEL_REF:
2830 case ADDR_VEC:
2831 case ADDR_DIFF_VEC:
2832 return x;
2833
2834 case REG:
2835 {
2836 int first;
2837 int q;
2838 struct qty_table_elem *ent;
2839
2840 /* Never replace a hard reg, because hard regs can appear
2841 in more than one machine mode, and we must preserve the mode
2842 of each occurrence. Also, some hard regs appear in
2843 MEMs that are shared and mustn't be altered. Don't try to
2844 replace any reg that maps to a reg of class NO_REGS. */
2845 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2846 || ! REGNO_QTY_VALID_P (REGNO (x)))
2847 return x;
2848
2849 q = REG_QTY (REGNO (x));
2850 ent = &qty_table[q];
2851 first = ent->first_reg;
2852 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2853 : REGNO_REG_CLASS (first) == NO_REGS ? x
2854 : gen_rtx_REG (ent->mode, first));
2855 }
2856
2857 default:
2858 break;
2859 }
2860
2861 fmt = GET_RTX_FORMAT (code);
2862 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2863 {
2864 int j;
2865
2866 if (fmt[i] == 'e')
2867 validate_canon_reg (&XEXP (x, i), insn);
2868 else if (fmt[i] == 'E')
2869 for (j = 0; j < XVECLEN (x, i); j++)
2870 validate_canon_reg (&XVECEXP (x, i, j), insn);
2871 }
2872
2873 return x;
2874 }
2875 \f
2876 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2877 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2878 what values are being compared.
2879
2880 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2881 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2882 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2883 compared to produce cc0.
2884
2885 The return value is the comparison operator and is either the code of
2886 A or the code corresponding to the inverse of the comparison. */
2887
2888 static enum rtx_code
2889 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2890 machine_mode *pmode1, machine_mode *pmode2)
2891 {
2892 rtx arg1, arg2;
2893 hash_set<rtx> *visited = NULL;
2894 /* Set nonzero when we find something of interest. */
2895 rtx x = NULL;
2896
2897 arg1 = *parg1, arg2 = *parg2;
2898
2899 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2900
2901 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2902 {
2903 int reverse_code = 0;
2904 struct table_elt *p = 0;
2905
2906 /* Remember state from previous iteration. */
2907 if (x)
2908 {
2909 if (!visited)
2910 visited = new hash_set<rtx>;
2911 visited->add (x);
2912 x = 0;
2913 }
2914
2915 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2916 On machines with CC0, this is the only case that can occur, since
2917 fold_rtx will return the COMPARE or item being compared with zero
2918 when given CC0. */
2919
2920 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2921 x = arg1;
2922
2923 /* If ARG1 is a comparison operator and CODE is testing for
2924 STORE_FLAG_VALUE, get the inner arguments. */
2925
2926 else if (COMPARISON_P (arg1))
2927 {
2928 #ifdef FLOAT_STORE_FLAG_VALUE
2929 REAL_VALUE_TYPE fsfv;
2930 #endif
2931
2932 if (code == NE
2933 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2934 && code == LT && STORE_FLAG_VALUE == -1)
2935 #ifdef FLOAT_STORE_FLAG_VALUE
2936 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2937 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2938 REAL_VALUE_NEGATIVE (fsfv)))
2939 #endif
2940 )
2941 x = arg1;
2942 else if (code == EQ
2943 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2944 && code == GE && STORE_FLAG_VALUE == -1)
2945 #ifdef FLOAT_STORE_FLAG_VALUE
2946 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2947 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2948 REAL_VALUE_NEGATIVE (fsfv)))
2949 #endif
2950 )
2951 x = arg1, reverse_code = 1;
2952 }
2953
2954 /* ??? We could also check for
2955
2956 (ne (and (eq (...) (const_int 1))) (const_int 0))
2957
2958 and related forms, but let's wait until we see them occurring. */
2959
2960 if (x == 0)
2961 /* Look up ARG1 in the hash table and see if it has an equivalence
2962 that lets us see what is being compared. */
2963 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2964 if (p)
2965 {
2966 p = p->first_same_value;
2967
2968 /* If what we compare is already known to be constant, that is as
2969 good as it gets.
2970 We need to break the loop in this case, because otherwise we
2971 can have an infinite loop when looking at a reg that is known
2972 to be a constant which is the same as a comparison of a reg
2973 against zero which appears later in the insn stream, which in
2974 turn is constant and the same as the comparison of the first reg
2975 against zero... */
2976 if (p->is_const)
2977 break;
2978 }
2979
2980 for (; p; p = p->next_same_value)
2981 {
2982 machine_mode inner_mode = GET_MODE (p->exp);
2983 #ifdef FLOAT_STORE_FLAG_VALUE
2984 REAL_VALUE_TYPE fsfv;
2985 #endif
2986
2987 /* If the entry isn't valid, skip it. */
2988 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2989 continue;
2990
2991 /* If it's a comparison we've used before, skip it. */
2992 if (visited && visited->contains (p->exp))
2993 continue;
2994
2995 if (GET_CODE (p->exp) == COMPARE
2996 /* Another possibility is that this machine has a compare insn
2997 that includes the comparison code. In that case, ARG1 would
2998 be equivalent to a comparison operation that would set ARG1 to
2999 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3000 ORIG_CODE is the actual comparison being done; if it is an EQ,
3001 we must reverse ORIG_CODE. On machine with a negative value
3002 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3003 || ((code == NE
3004 || (code == LT
3005 && val_signbit_known_set_p (inner_mode,
3006 STORE_FLAG_VALUE))
3007 #ifdef FLOAT_STORE_FLAG_VALUE
3008 || (code == LT
3009 && SCALAR_FLOAT_MODE_P (inner_mode)
3010 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3011 REAL_VALUE_NEGATIVE (fsfv)))
3012 #endif
3013 )
3014 && COMPARISON_P (p->exp)))
3015 {
3016 x = p->exp;
3017 break;
3018 }
3019 else if ((code == EQ
3020 || (code == GE
3021 && val_signbit_known_set_p (inner_mode,
3022 STORE_FLAG_VALUE))
3023 #ifdef FLOAT_STORE_FLAG_VALUE
3024 || (code == GE
3025 && SCALAR_FLOAT_MODE_P (inner_mode)
3026 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3027 REAL_VALUE_NEGATIVE (fsfv)))
3028 #endif
3029 )
3030 && COMPARISON_P (p->exp))
3031 {
3032 reverse_code = 1;
3033 x = p->exp;
3034 break;
3035 }
3036
3037 /* If this non-trapping address, e.g. fp + constant, the
3038 equivalent is a better operand since it may let us predict
3039 the value of the comparison. */
3040 else if (!rtx_addr_can_trap_p (p->exp))
3041 {
3042 arg1 = p->exp;
3043 continue;
3044 }
3045 }
3046
3047 /* If we didn't find a useful equivalence for ARG1, we are done.
3048 Otherwise, set up for the next iteration. */
3049 if (x == 0)
3050 break;
3051
3052 /* If we need to reverse the comparison, make sure that is
3053 possible -- we can't necessarily infer the value of GE from LT
3054 with floating-point operands. */
3055 if (reverse_code)
3056 {
3057 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3058 if (reversed == UNKNOWN)
3059 break;
3060 else
3061 code = reversed;
3062 }
3063 else if (COMPARISON_P (x))
3064 code = GET_CODE (x);
3065 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3066 }
3067
3068 /* Return our results. Return the modes from before fold_rtx
3069 because fold_rtx might produce const_int, and then it's too late. */
3070 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3071 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3072
3073 if (visited)
3074 delete visited;
3075 return code;
3076 }
3077 \f
3078 /* If X is a nontrivial arithmetic operation on an argument for which
3079 a constant value can be determined, return the result of operating
3080 on that value, as a constant. Otherwise, return X, possibly with
3081 one or more operands changed to a forward-propagated constant.
3082
3083 If X is a register whose contents are known, we do NOT return
3084 those contents here; equiv_constant is called to perform that task.
3085 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3086
3087 INSN is the insn that we may be modifying. If it is 0, make a copy
3088 of X before modifying it. */
3089
3090 static rtx
3091 fold_rtx (rtx x, rtx_insn *insn)
3092 {
3093 enum rtx_code code;
3094 machine_mode mode;
3095 const char *fmt;
3096 int i;
3097 rtx new_rtx = 0;
3098 int changed = 0;
3099
3100 /* Operands of X. */
3101 /* Workaround -Wmaybe-uninitialized false positive during
3102 profiledbootstrap by initializing them. */
3103 rtx folded_arg0 = NULL_RTX;
3104 rtx folded_arg1 = NULL_RTX;
3105
3106 /* Constant equivalents of first three operands of X;
3107 0 when no such equivalent is known. */
3108 rtx const_arg0;
3109 rtx const_arg1;
3110 rtx const_arg2;
3111
3112 /* The mode of the first operand of X. We need this for sign and zero
3113 extends. */
3114 machine_mode mode_arg0;
3115
3116 if (x == 0)
3117 return x;
3118
3119 /* Try to perform some initial simplifications on X. */
3120 code = GET_CODE (x);
3121 switch (code)
3122 {
3123 case MEM:
3124 case SUBREG:
3125 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3126 than it would in other contexts. Basically its mode does not
3127 signify the size of the object read. That information is carried
3128 by size operand. If we happen to have a MEM of the appropriate
3129 mode in our tables with a constant value we could simplify the
3130 extraction incorrectly if we allowed substitution of that value
3131 for the MEM. */
3132 case ZERO_EXTRACT:
3133 case SIGN_EXTRACT:
3134 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3135 return new_rtx;
3136 return x;
3137
3138 case CONST:
3139 CASE_CONST_ANY:
3140 case SYMBOL_REF:
3141 case LABEL_REF:
3142 case REG:
3143 case PC:
3144 /* No use simplifying an EXPR_LIST
3145 since they are used only for lists of args
3146 in a function call's REG_EQUAL note. */
3147 case EXPR_LIST:
3148 return x;
3149
3150 case CC0:
3151 return prev_insn_cc0;
3152
3153 case ASM_OPERANDS:
3154 if (insn)
3155 {
3156 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3157 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3158 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3159 }
3160 return x;
3161
3162 case CALL:
3163 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3164 return x;
3165 break;
3166
3167 /* Anything else goes through the loop below. */
3168 default:
3169 break;
3170 }
3171
3172 mode = GET_MODE (x);
3173 const_arg0 = 0;
3174 const_arg1 = 0;
3175 const_arg2 = 0;
3176 mode_arg0 = VOIDmode;
3177
3178 /* Try folding our operands.
3179 Then see which ones have constant values known. */
3180
3181 fmt = GET_RTX_FORMAT (code);
3182 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3183 if (fmt[i] == 'e')
3184 {
3185 rtx folded_arg = XEXP (x, i), const_arg;
3186 machine_mode mode_arg = GET_MODE (folded_arg);
3187
3188 switch (GET_CODE (folded_arg))
3189 {
3190 case MEM:
3191 case REG:
3192 case SUBREG:
3193 const_arg = equiv_constant (folded_arg);
3194 break;
3195
3196 case CONST:
3197 CASE_CONST_ANY:
3198 case SYMBOL_REF:
3199 case LABEL_REF:
3200 const_arg = folded_arg;
3201 break;
3202
3203 case CC0:
3204 /* The cc0-user and cc0-setter may be in different blocks if
3205 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3206 will have been cleared as we exited the block with the
3207 setter.
3208
3209 While we could potentially track cc0 in this case, it just
3210 doesn't seem to be worth it given that cc0 targets are not
3211 terribly common or important these days and trapping math
3212 is rarely used. The combination of those two conditions
3213 necessary to trip this situation is exceedingly rare in the
3214 real world. */
3215 if (!prev_insn_cc0)
3216 {
3217 const_arg = NULL_RTX;
3218 }
3219 else
3220 {
3221 folded_arg = prev_insn_cc0;
3222 mode_arg = prev_insn_cc0_mode;
3223 const_arg = equiv_constant (folded_arg);
3224 }
3225 break;
3226
3227 default:
3228 folded_arg = fold_rtx (folded_arg, insn);
3229 const_arg = equiv_constant (folded_arg);
3230 break;
3231 }
3232
3233 /* For the first three operands, see if the operand
3234 is constant or equivalent to a constant. */
3235 switch (i)
3236 {
3237 case 0:
3238 folded_arg0 = folded_arg;
3239 const_arg0 = const_arg;
3240 mode_arg0 = mode_arg;
3241 break;
3242 case 1:
3243 folded_arg1 = folded_arg;
3244 const_arg1 = const_arg;
3245 break;
3246 case 2:
3247 const_arg2 = const_arg;
3248 break;
3249 }
3250
3251 /* Pick the least expensive of the argument and an equivalent constant
3252 argument. */
3253 if (const_arg != 0
3254 && const_arg != folded_arg
3255 && (COST_IN (const_arg, mode_arg, code, i)
3256 <= COST_IN (folded_arg, mode_arg, code, i))
3257
3258 /* It's not safe to substitute the operand of a conversion
3259 operator with a constant, as the conversion's identity
3260 depends upon the mode of its operand. This optimization
3261 is handled by the call to simplify_unary_operation. */
3262 && (GET_RTX_CLASS (code) != RTX_UNARY
3263 || GET_MODE (const_arg) == mode_arg0
3264 || (code != ZERO_EXTEND
3265 && code != SIGN_EXTEND
3266 && code != TRUNCATE
3267 && code != FLOAT_TRUNCATE
3268 && code != FLOAT_EXTEND
3269 && code != FLOAT
3270 && code != FIX
3271 && code != UNSIGNED_FLOAT
3272 && code != UNSIGNED_FIX)))
3273 folded_arg = const_arg;
3274
3275 if (folded_arg == XEXP (x, i))
3276 continue;
3277
3278 if (insn == NULL_RTX && !changed)
3279 x = copy_rtx (x);
3280 changed = 1;
3281 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3282 }
3283
3284 if (changed)
3285 {
3286 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3287 consistent with the order in X. */
3288 if (canonicalize_change_group (insn, x))
3289 {
3290 std::swap (const_arg0, const_arg1);
3291 std::swap (folded_arg0, folded_arg1);
3292 }
3293
3294 apply_change_group ();
3295 }
3296
3297 /* If X is an arithmetic operation, see if we can simplify it. */
3298
3299 switch (GET_RTX_CLASS (code))
3300 {
3301 case RTX_UNARY:
3302 {
3303 /* We can't simplify extension ops unless we know the
3304 original mode. */
3305 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3306 && mode_arg0 == VOIDmode)
3307 break;
3308
3309 new_rtx = simplify_unary_operation (code, mode,
3310 const_arg0 ? const_arg0 : folded_arg0,
3311 mode_arg0);
3312 }
3313 break;
3314
3315 case RTX_COMPARE:
3316 case RTX_COMM_COMPARE:
3317 /* See what items are actually being compared and set FOLDED_ARG[01]
3318 to those values and CODE to the actual comparison code. If any are
3319 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3320 do anything if both operands are already known to be constant. */
3321
3322 /* ??? Vector mode comparisons are not supported yet. */
3323 if (VECTOR_MODE_P (mode))
3324 break;
3325
3326 if (const_arg0 == 0 || const_arg1 == 0)
3327 {
3328 struct table_elt *p0, *p1;
3329 rtx true_rtx, false_rtx;
3330 machine_mode mode_arg1;
3331
3332 if (SCALAR_FLOAT_MODE_P (mode))
3333 {
3334 #ifdef FLOAT_STORE_FLAG_VALUE
3335 true_rtx = (const_double_from_real_value
3336 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3337 #else
3338 true_rtx = NULL_RTX;
3339 #endif
3340 false_rtx = CONST0_RTX (mode);
3341 }
3342 else
3343 {
3344 true_rtx = const_true_rtx;
3345 false_rtx = const0_rtx;
3346 }
3347
3348 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3349 &mode_arg0, &mode_arg1);
3350
3351 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3352 what kinds of things are being compared, so we can't do
3353 anything with this comparison. */
3354
3355 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3356 break;
3357
3358 const_arg0 = equiv_constant (folded_arg0);
3359 const_arg1 = equiv_constant (folded_arg1);
3360
3361 /* If we do not now have two constants being compared, see
3362 if we can nevertheless deduce some things about the
3363 comparison. */
3364 if (const_arg0 == 0 || const_arg1 == 0)
3365 {
3366 if (const_arg1 != NULL)
3367 {
3368 rtx cheapest_simplification;
3369 int cheapest_cost;
3370 rtx simp_result;
3371 struct table_elt *p;
3372
3373 /* See if we can find an equivalent of folded_arg0
3374 that gets us a cheaper expression, possibly a
3375 constant through simplifications. */
3376 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3377 mode_arg0);
3378
3379 if (p != NULL)
3380 {
3381 cheapest_simplification = x;
3382 cheapest_cost = COST (x, mode);
3383
3384 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3385 {
3386 int cost;
3387
3388 /* If the entry isn't valid, skip it. */
3389 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3390 continue;
3391
3392 /* Try to simplify using this equivalence. */
3393 simp_result
3394 = simplify_relational_operation (code, mode,
3395 mode_arg0,
3396 p->exp,
3397 const_arg1);
3398
3399 if (simp_result == NULL)
3400 continue;
3401
3402 cost = COST (simp_result, mode);
3403 if (cost < cheapest_cost)
3404 {
3405 cheapest_cost = cost;
3406 cheapest_simplification = simp_result;
3407 }
3408 }
3409
3410 /* If we have a cheaper expression now, use that
3411 and try folding it further, from the top. */
3412 if (cheapest_simplification != x)
3413 return fold_rtx (copy_rtx (cheapest_simplification),
3414 insn);
3415 }
3416 }
3417
3418 /* See if the two operands are the same. */
3419
3420 if ((REG_P (folded_arg0)
3421 && REG_P (folded_arg1)
3422 && (REG_QTY (REGNO (folded_arg0))
3423 == REG_QTY (REGNO (folded_arg1))))
3424 || ((p0 = lookup (folded_arg0,
3425 SAFE_HASH (folded_arg0, mode_arg0),
3426 mode_arg0))
3427 && (p1 = lookup (folded_arg1,
3428 SAFE_HASH (folded_arg1, mode_arg0),
3429 mode_arg0))
3430 && p0->first_same_value == p1->first_same_value))
3431 folded_arg1 = folded_arg0;
3432
3433 /* If FOLDED_ARG0 is a register, see if the comparison we are
3434 doing now is either the same as we did before or the reverse
3435 (we only check the reverse if not floating-point). */
3436 else if (REG_P (folded_arg0))
3437 {
3438 int qty = REG_QTY (REGNO (folded_arg0));
3439
3440 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3441 {
3442 struct qty_table_elem *ent = &qty_table[qty];
3443
3444 if ((comparison_dominates_p (ent->comparison_code, code)
3445 || (! FLOAT_MODE_P (mode_arg0)
3446 && comparison_dominates_p (ent->comparison_code,
3447 reverse_condition (code))))
3448 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3449 || (const_arg1
3450 && rtx_equal_p (ent->comparison_const,
3451 const_arg1))
3452 || (REG_P (folded_arg1)
3453 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3454 {
3455 if (comparison_dominates_p (ent->comparison_code, code))
3456 {
3457 if (true_rtx)
3458 return true_rtx;
3459 else
3460 break;
3461 }
3462 else
3463 return false_rtx;
3464 }
3465 }
3466 }
3467 }
3468 }
3469
3470 /* If we are comparing against zero, see if the first operand is
3471 equivalent to an IOR with a constant. If so, we may be able to
3472 determine the result of this comparison. */
3473 if (const_arg1 == const0_rtx && !const_arg0)
3474 {
3475 rtx y = lookup_as_function (folded_arg0, IOR);
3476 rtx inner_const;
3477
3478 if (y != 0
3479 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3480 && CONST_INT_P (inner_const)
3481 && INTVAL (inner_const) != 0)
3482 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3483 }
3484
3485 {
3486 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3487 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3488 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3489 op0, op1);
3490 }
3491 break;
3492
3493 case RTX_BIN_ARITH:
3494 case RTX_COMM_ARITH:
3495 switch (code)
3496 {
3497 case PLUS:
3498 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3499 with that LABEL_REF as its second operand. If so, the result is
3500 the first operand of that MINUS. This handles switches with an
3501 ADDR_DIFF_VEC table. */
3502 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3503 {
3504 rtx y
3505 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3506 : lookup_as_function (folded_arg0, MINUS);
3507
3508 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3509 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg1))
3510 return XEXP (y, 0);
3511
3512 /* Now try for a CONST of a MINUS like the above. */
3513 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3514 : lookup_as_function (folded_arg0, CONST))) != 0
3515 && GET_CODE (XEXP (y, 0)) == MINUS
3516 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3517 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg1))
3518 return XEXP (XEXP (y, 0), 0);
3519 }
3520
3521 /* Likewise if the operands are in the other order. */
3522 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3523 {
3524 rtx y
3525 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3526 : lookup_as_function (folded_arg1, MINUS);
3527
3528 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3529 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg0))
3530 return XEXP (y, 0);
3531
3532 /* Now try for a CONST of a MINUS like the above. */
3533 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3534 : lookup_as_function (folded_arg1, CONST))) != 0
3535 && GET_CODE (XEXP (y, 0)) == MINUS
3536 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3537 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg0))
3538 return XEXP (XEXP (y, 0), 0);
3539 }
3540
3541 /* If second operand is a register equivalent to a negative
3542 CONST_INT, see if we can find a register equivalent to the
3543 positive constant. Make a MINUS if so. Don't do this for
3544 a non-negative constant since we might then alternate between
3545 choosing positive and negative constants. Having the positive
3546 constant previously-used is the more common case. Be sure
3547 the resulting constant is non-negative; if const_arg1 were
3548 the smallest negative number this would overflow: depending
3549 on the mode, this would either just be the same value (and
3550 hence not save anything) or be incorrect. */
3551 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3552 && INTVAL (const_arg1) < 0
3553 /* This used to test
3554
3555 -INTVAL (const_arg1) >= 0
3556
3557 But The Sun V5.0 compilers mis-compiled that test. So
3558 instead we test for the problematic value in a more direct
3559 manner and hope the Sun compilers get it correct. */
3560 && INTVAL (const_arg1) !=
3561 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3562 && REG_P (folded_arg1))
3563 {
3564 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3565 struct table_elt *p
3566 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3567
3568 if (p)
3569 for (p = p->first_same_value; p; p = p->next_same_value)
3570 if (REG_P (p->exp))
3571 return simplify_gen_binary (MINUS, mode, folded_arg0,
3572 canon_reg (p->exp, NULL));
3573 }
3574 goto from_plus;
3575
3576 case MINUS:
3577 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3578 If so, produce (PLUS Z C2-C). */
3579 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3580 {
3581 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3582 if (y && CONST_INT_P (XEXP (y, 1)))
3583 return fold_rtx (plus_constant (mode, copy_rtx (y),
3584 -INTVAL (const_arg1)),
3585 NULL);
3586 }
3587
3588 /* Fall through. */
3589
3590 from_plus:
3591 case SMIN: case SMAX: case UMIN: case UMAX:
3592 case IOR: case AND: case XOR:
3593 case MULT:
3594 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3595 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3596 is known to be of similar form, we may be able to replace the
3597 operation with a combined operation. This may eliminate the
3598 intermediate operation if every use is simplified in this way.
3599 Note that the similar optimization done by combine.c only works
3600 if the intermediate operation's result has only one reference. */
3601
3602 if (REG_P (folded_arg0)
3603 && const_arg1 && CONST_INT_P (const_arg1))
3604 {
3605 int is_shift
3606 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3607 rtx y, inner_const, new_const;
3608 rtx canon_const_arg1 = const_arg1;
3609 enum rtx_code associate_code;
3610
3611 if (is_shift
3612 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3613 || INTVAL (const_arg1) < 0))
3614 {
3615 if (SHIFT_COUNT_TRUNCATED)
3616 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3617 & (GET_MODE_BITSIZE (mode)
3618 - 1));
3619 else
3620 break;
3621 }
3622
3623 y = lookup_as_function (folded_arg0, code);
3624 if (y == 0)
3625 break;
3626
3627 /* If we have compiled a statement like
3628 "if (x == (x & mask1))", and now are looking at
3629 "x & mask2", we will have a case where the first operand
3630 of Y is the same as our first operand. Unless we detect
3631 this case, an infinite loop will result. */
3632 if (XEXP (y, 0) == folded_arg0)
3633 break;
3634
3635 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3636 if (!inner_const || !CONST_INT_P (inner_const))
3637 break;
3638
3639 /* Don't associate these operations if they are a PLUS with the
3640 same constant and it is a power of two. These might be doable
3641 with a pre- or post-increment. Similarly for two subtracts of
3642 identical powers of two with post decrement. */
3643
3644 if (code == PLUS && const_arg1 == inner_const
3645 && ((HAVE_PRE_INCREMENT
3646 && exact_log2 (INTVAL (const_arg1)) >= 0)
3647 || (HAVE_POST_INCREMENT
3648 && exact_log2 (INTVAL (const_arg1)) >= 0)
3649 || (HAVE_PRE_DECREMENT
3650 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3651 || (HAVE_POST_DECREMENT
3652 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3653 break;
3654
3655 /* ??? Vector mode shifts by scalar
3656 shift operand are not supported yet. */
3657 if (is_shift && VECTOR_MODE_P (mode))
3658 break;
3659
3660 if (is_shift
3661 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3662 || INTVAL (inner_const) < 0))
3663 {
3664 if (SHIFT_COUNT_TRUNCATED)
3665 inner_const = GEN_INT (INTVAL (inner_const)
3666 & (GET_MODE_BITSIZE (mode) - 1));
3667 else
3668 break;
3669 }
3670
3671 /* Compute the code used to compose the constants. For example,
3672 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3673
3674 associate_code = (is_shift || code == MINUS ? PLUS : code);
3675
3676 new_const = simplify_binary_operation (associate_code, mode,
3677 canon_const_arg1,
3678 inner_const);
3679
3680 if (new_const == 0)
3681 break;
3682
3683 /* If we are associating shift operations, don't let this
3684 produce a shift of the size of the object or larger.
3685 This could occur when we follow a sign-extend by a right
3686 shift on a machine that does a sign-extend as a pair
3687 of shifts. */
3688
3689 if (is_shift
3690 && CONST_INT_P (new_const)
3691 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3692 {
3693 /* As an exception, we can turn an ASHIFTRT of this
3694 form into a shift of the number of bits - 1. */
3695 if (code == ASHIFTRT)
3696 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3697 else if (!side_effects_p (XEXP (y, 0)))
3698 return CONST0_RTX (mode);
3699 else
3700 break;
3701 }
3702
3703 y = copy_rtx (XEXP (y, 0));
3704
3705 /* If Y contains our first operand (the most common way this
3706 can happen is if Y is a MEM), we would do into an infinite
3707 loop if we tried to fold it. So don't in that case. */
3708
3709 if (! reg_mentioned_p (folded_arg0, y))
3710 y = fold_rtx (y, insn);
3711
3712 return simplify_gen_binary (code, mode, y, new_const);
3713 }
3714 break;
3715
3716 case DIV: case UDIV:
3717 /* ??? The associative optimization performed immediately above is
3718 also possible for DIV and UDIV using associate_code of MULT.
3719 However, we would need extra code to verify that the
3720 multiplication does not overflow, that is, there is no overflow
3721 in the calculation of new_const. */
3722 break;
3723
3724 default:
3725 break;
3726 }
3727
3728 new_rtx = simplify_binary_operation (code, mode,
3729 const_arg0 ? const_arg0 : folded_arg0,
3730 const_arg1 ? const_arg1 : folded_arg1);
3731 break;
3732
3733 case RTX_OBJ:
3734 /* (lo_sum (high X) X) is simply X. */
3735 if (code == LO_SUM && const_arg0 != 0
3736 && GET_CODE (const_arg0) == HIGH
3737 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3738 return const_arg1;
3739 break;
3740
3741 case RTX_TERNARY:
3742 case RTX_BITFIELD_OPS:
3743 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3744 const_arg0 ? const_arg0 : folded_arg0,
3745 const_arg1 ? const_arg1 : folded_arg1,
3746 const_arg2 ? const_arg2 : XEXP (x, 2));
3747 break;
3748
3749 default:
3750 break;
3751 }
3752
3753 return new_rtx ? new_rtx : x;
3754 }
3755 \f
3756 /* Return a constant value currently equivalent to X.
3757 Return 0 if we don't know one. */
3758
3759 static rtx
3760 equiv_constant (rtx x)
3761 {
3762 if (REG_P (x)
3763 && REGNO_QTY_VALID_P (REGNO (x)))
3764 {
3765 int x_q = REG_QTY (REGNO (x));
3766 struct qty_table_elem *x_ent = &qty_table[x_q];
3767
3768 if (x_ent->const_rtx)
3769 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3770 }
3771
3772 if (x == 0 || CONSTANT_P (x))
3773 return x;
3774
3775 if (GET_CODE (x) == SUBREG)
3776 {
3777 machine_mode mode = GET_MODE (x);
3778 machine_mode imode = GET_MODE (SUBREG_REG (x));
3779 rtx new_rtx;
3780
3781 /* See if we previously assigned a constant value to this SUBREG. */
3782 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3783 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3784 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3785 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3786 return new_rtx;
3787
3788 /* If we didn't and if doing so makes sense, see if we previously
3789 assigned a constant value to the enclosing word mode SUBREG. */
3790 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3791 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3792 {
3793 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3794 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3795 {
3796 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3797 new_rtx = lookup_as_function (y, CONST_INT);
3798 if (new_rtx)
3799 return gen_lowpart (mode, new_rtx);
3800 }
3801 }
3802
3803 /* Otherwise see if we already have a constant for the inner REG,
3804 and if that is enough to calculate an equivalent constant for
3805 the subreg. Note that the upper bits of paradoxical subregs
3806 are undefined, so they cannot be said to equal anything. */
3807 if (REG_P (SUBREG_REG (x))
3808 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3809 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3810 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3811
3812 return 0;
3813 }
3814
3815 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3816 the hash table in case its value was seen before. */
3817
3818 if (MEM_P (x))
3819 {
3820 struct table_elt *elt;
3821
3822 x = avoid_constant_pool_reference (x);
3823 if (CONSTANT_P (x))
3824 return x;
3825
3826 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3827 if (elt == 0)
3828 return 0;
3829
3830 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3831 if (elt->is_const && CONSTANT_P (elt->exp))
3832 return elt->exp;
3833 }
3834
3835 return 0;
3836 }
3837 \f
3838 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3839 "taken" branch.
3840
3841 In certain cases, this can cause us to add an equivalence. For example,
3842 if we are following the taken case of
3843 if (i == 2)
3844 we can add the fact that `i' and '2' are now equivalent.
3845
3846 In any case, we can record that this comparison was passed. If the same
3847 comparison is seen later, we will know its value. */
3848
3849 static void
3850 record_jump_equiv (rtx_insn *insn, bool taken)
3851 {
3852 int cond_known_true;
3853 rtx op0, op1;
3854 rtx set;
3855 machine_mode mode, mode0, mode1;
3856 int reversed_nonequality = 0;
3857 enum rtx_code code;
3858
3859 /* Ensure this is the right kind of insn. */
3860 gcc_assert (any_condjump_p (insn));
3861
3862 set = pc_set (insn);
3863
3864 /* See if this jump condition is known true or false. */
3865 if (taken)
3866 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3867 else
3868 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3869
3870 /* Get the type of comparison being done and the operands being compared.
3871 If we had to reverse a non-equality condition, record that fact so we
3872 know that it isn't valid for floating-point. */
3873 code = GET_CODE (XEXP (SET_SRC (set), 0));
3874 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3875 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3876
3877 /* On a cc0 target the cc0-setter and cc0-user may end up in different
3878 blocks. When that happens the tracking of the cc0-setter via
3879 PREV_INSN_CC0 is spoiled. That means that fold_rtx may return
3880 NULL_RTX. In those cases, there's nothing to record. */
3881 if (op0 == NULL_RTX || op1 == NULL_RTX)
3882 return;
3883
3884 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3885 if (! cond_known_true)
3886 {
3887 code = reversed_comparison_code_parts (code, op0, op1, insn);
3888
3889 /* Don't remember if we can't find the inverse. */
3890 if (code == UNKNOWN)
3891 return;
3892 }
3893
3894 /* The mode is the mode of the non-constant. */
3895 mode = mode0;
3896 if (mode1 != VOIDmode)
3897 mode = mode1;
3898
3899 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3900 }
3901
3902 /* Yet another form of subreg creation. In this case, we want something in
3903 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3904
3905 static rtx
3906 record_jump_cond_subreg (machine_mode mode, rtx op)
3907 {
3908 machine_mode op_mode = GET_MODE (op);
3909 if (op_mode == mode || op_mode == VOIDmode)
3910 return op;
3911 return lowpart_subreg (mode, op, op_mode);
3912 }
3913
3914 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3915 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3916 Make any useful entries we can with that information. Called from
3917 above function and called recursively. */
3918
3919 static void
3920 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3921 rtx op1, int reversed_nonequality)
3922 {
3923 unsigned op0_hash, op1_hash;
3924 int op0_in_memory, op1_in_memory;
3925 struct table_elt *op0_elt, *op1_elt;
3926
3927 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3928 we know that they are also equal in the smaller mode (this is also
3929 true for all smaller modes whether or not there is a SUBREG, but
3930 is not worth testing for with no SUBREG). */
3931
3932 /* Note that GET_MODE (op0) may not equal MODE. */
3933 if (code == EQ && paradoxical_subreg_p (op0))
3934 {
3935 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3936 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3937 if (tem)
3938 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3939 reversed_nonequality);
3940 }
3941
3942 if (code == EQ && paradoxical_subreg_p (op1))
3943 {
3944 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3945 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3946 if (tem)
3947 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3948 reversed_nonequality);
3949 }
3950
3951 /* Similarly, if this is an NE comparison, and either is a SUBREG
3952 making a smaller mode, we know the whole thing is also NE. */
3953
3954 /* Note that GET_MODE (op0) may not equal MODE;
3955 if we test MODE instead, we can get an infinite recursion
3956 alternating between two modes each wider than MODE. */
3957
3958 if (code == NE && GET_CODE (op0) == SUBREG
3959 && subreg_lowpart_p (op0)
3960 && (GET_MODE_SIZE (GET_MODE (op0))
3961 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3962 {
3963 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3964 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3965 if (tem)
3966 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3967 reversed_nonequality);
3968 }
3969
3970 if (code == NE && GET_CODE (op1) == SUBREG
3971 && subreg_lowpart_p (op1)
3972 && (GET_MODE_SIZE (GET_MODE (op1))
3973 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3974 {
3975 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3976 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3977 if (tem)
3978 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3979 reversed_nonequality);
3980 }
3981
3982 /* Hash both operands. */
3983
3984 do_not_record = 0;
3985 hash_arg_in_memory = 0;
3986 op0_hash = HASH (op0, mode);
3987 op0_in_memory = hash_arg_in_memory;
3988
3989 if (do_not_record)
3990 return;
3991
3992 do_not_record = 0;
3993 hash_arg_in_memory = 0;
3994 op1_hash = HASH (op1, mode);
3995 op1_in_memory = hash_arg_in_memory;
3996
3997 if (do_not_record)
3998 return;
3999
4000 /* Look up both operands. */
4001 op0_elt = lookup (op0, op0_hash, mode);
4002 op1_elt = lookup (op1, op1_hash, mode);
4003
4004 /* If both operands are already equivalent or if they are not in the
4005 table but are identical, do nothing. */
4006 if ((op0_elt != 0 && op1_elt != 0
4007 && op0_elt->first_same_value == op1_elt->first_same_value)
4008 || op0 == op1 || rtx_equal_p (op0, op1))
4009 return;
4010
4011 /* If we aren't setting two things equal all we can do is save this
4012 comparison. Similarly if this is floating-point. In the latter
4013 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4014 If we record the equality, we might inadvertently delete code
4015 whose intent was to change -0 to +0. */
4016
4017 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4018 {
4019 struct qty_table_elem *ent;
4020 int qty;
4021
4022 /* If we reversed a floating-point comparison, if OP0 is not a
4023 register, or if OP1 is neither a register or constant, we can't
4024 do anything. */
4025
4026 if (!REG_P (op1))
4027 op1 = equiv_constant (op1);
4028
4029 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4030 || !REG_P (op0) || op1 == 0)
4031 return;
4032
4033 /* Put OP0 in the hash table if it isn't already. This gives it a
4034 new quantity number. */
4035 if (op0_elt == 0)
4036 {
4037 if (insert_regs (op0, NULL, 0))
4038 {
4039 rehash_using_reg (op0);
4040 op0_hash = HASH (op0, mode);
4041
4042 /* If OP0 is contained in OP1, this changes its hash code
4043 as well. Faster to rehash than to check, except
4044 for the simple case of a constant. */
4045 if (! CONSTANT_P (op1))
4046 op1_hash = HASH (op1,mode);
4047 }
4048
4049 op0_elt = insert (op0, NULL, op0_hash, mode);
4050 op0_elt->in_memory = op0_in_memory;
4051 }
4052
4053 qty = REG_QTY (REGNO (op0));
4054 ent = &qty_table[qty];
4055
4056 ent->comparison_code = code;
4057 if (REG_P (op1))
4058 {
4059 /* Look it up again--in case op0 and op1 are the same. */
4060 op1_elt = lookup (op1, op1_hash, mode);
4061
4062 /* Put OP1 in the hash table so it gets a new quantity number. */
4063 if (op1_elt == 0)
4064 {
4065 if (insert_regs (op1, NULL, 0))
4066 {
4067 rehash_using_reg (op1);
4068 op1_hash = HASH (op1, mode);
4069 }
4070
4071 op1_elt = insert (op1, NULL, op1_hash, mode);
4072 op1_elt->in_memory = op1_in_memory;
4073 }
4074
4075 ent->comparison_const = NULL_RTX;
4076 ent->comparison_qty = REG_QTY (REGNO (op1));
4077 }
4078 else
4079 {
4080 ent->comparison_const = op1;
4081 ent->comparison_qty = -1;
4082 }
4083
4084 return;
4085 }
4086
4087 /* If either side is still missing an equivalence, make it now,
4088 then merge the equivalences. */
4089
4090 if (op0_elt == 0)
4091 {
4092 if (insert_regs (op0, NULL, 0))
4093 {
4094 rehash_using_reg (op0);
4095 op0_hash = HASH (op0, mode);
4096 }
4097
4098 op0_elt = insert (op0, NULL, op0_hash, mode);
4099 op0_elt->in_memory = op0_in_memory;
4100 }
4101
4102 if (op1_elt == 0)
4103 {
4104 if (insert_regs (op1, NULL, 0))
4105 {
4106 rehash_using_reg (op1);
4107 op1_hash = HASH (op1, mode);
4108 }
4109
4110 op1_elt = insert (op1, NULL, op1_hash, mode);
4111 op1_elt->in_memory = op1_in_memory;
4112 }
4113
4114 merge_equiv_classes (op0_elt, op1_elt);
4115 }
4116 \f
4117 /* CSE processing for one instruction.
4118
4119 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4120 but the few that "leak through" are cleaned up by cse_insn, and complex
4121 addressing modes are often formed here.
4122
4123 The main function is cse_insn, and between here and that function
4124 a couple of helper functions is defined to keep the size of cse_insn
4125 within reasonable proportions.
4126
4127 Data is shared between the main and helper functions via STRUCT SET,
4128 that contains all data related for every set in the instruction that
4129 is being processed.
4130
4131 Note that cse_main processes all sets in the instruction. Most
4132 passes in GCC only process simple SET insns or single_set insns, but
4133 CSE processes insns with multiple sets as well. */
4134
4135 /* Data on one SET contained in the instruction. */
4136
4137 struct set
4138 {
4139 /* The SET rtx itself. */
4140 rtx rtl;
4141 /* The SET_SRC of the rtx (the original value, if it is changing). */
4142 rtx src;
4143 /* The hash-table element for the SET_SRC of the SET. */
4144 struct table_elt *src_elt;
4145 /* Hash value for the SET_SRC. */
4146 unsigned src_hash;
4147 /* Hash value for the SET_DEST. */
4148 unsigned dest_hash;
4149 /* The SET_DEST, with SUBREG, etc., stripped. */
4150 rtx inner_dest;
4151 /* Nonzero if the SET_SRC is in memory. */
4152 char src_in_memory;
4153 /* Nonzero if the SET_SRC contains something
4154 whose value cannot be predicted and understood. */
4155 char src_volatile;
4156 /* Original machine mode, in case it becomes a CONST_INT.
4157 The size of this field should match the size of the mode
4158 field of struct rtx_def (see rtl.h). */
4159 ENUM_BITFIELD(machine_mode) mode : 8;
4160 /* A constant equivalent for SET_SRC, if any. */
4161 rtx src_const;
4162 /* Hash value of constant equivalent for SET_SRC. */
4163 unsigned src_const_hash;
4164 /* Table entry for constant equivalent for SET_SRC, if any. */
4165 struct table_elt *src_const_elt;
4166 /* Table entry for the destination address. */
4167 struct table_elt *dest_addr_elt;
4168 };
4169 \f
4170 /* Special handling for (set REG0 REG1) where REG0 is the
4171 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4172 be used in the sequel, so (if easily done) change this insn to
4173 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4174 that computed their value. Then REG1 will become a dead store
4175 and won't cloud the situation for later optimizations.
4176
4177 Do not make this change if REG1 is a hard register, because it will
4178 then be used in the sequel and we may be changing a two-operand insn
4179 into a three-operand insn.
4180
4181 This is the last transformation that cse_insn will try to do. */
4182
4183 static void
4184 try_back_substitute_reg (rtx set, rtx_insn *insn)
4185 {
4186 rtx dest = SET_DEST (set);
4187 rtx src = SET_SRC (set);
4188
4189 if (REG_P (dest)
4190 && REG_P (src) && ! HARD_REGISTER_P (src)
4191 && REGNO_QTY_VALID_P (REGNO (src)))
4192 {
4193 int src_q = REG_QTY (REGNO (src));
4194 struct qty_table_elem *src_ent = &qty_table[src_q];
4195
4196 if (src_ent->first_reg == REGNO (dest))
4197 {
4198 /* Scan for the previous nonnote insn, but stop at a basic
4199 block boundary. */
4200 rtx_insn *prev = insn;
4201 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4202 do
4203 {
4204 prev = PREV_INSN (prev);
4205 }
4206 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4207
4208 /* Do not swap the registers around if the previous instruction
4209 attaches a REG_EQUIV note to REG1.
4210
4211 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4212 from the pseudo that originally shadowed an incoming argument
4213 to another register. Some uses of REG_EQUIV might rely on it
4214 being attached to REG1 rather than REG2.
4215
4216 This section previously turned the REG_EQUIV into a REG_EQUAL
4217 note. We cannot do that because REG_EQUIV may provide an
4218 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4219 if (NONJUMP_INSN_P (prev)
4220 && GET_CODE (PATTERN (prev)) == SET
4221 && SET_DEST (PATTERN (prev)) == src
4222 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4223 {
4224 rtx note;
4225
4226 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4227 validate_change (insn, &SET_DEST (set), src, 1);
4228 validate_change (insn, &SET_SRC (set), dest, 1);
4229 apply_change_group ();
4230
4231 /* If INSN has a REG_EQUAL note, and this note mentions
4232 REG0, then we must delete it, because the value in
4233 REG0 has changed. If the note's value is REG1, we must
4234 also delete it because that is now this insn's dest. */
4235 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4236 if (note != 0
4237 && (reg_mentioned_p (dest, XEXP (note, 0))
4238 || rtx_equal_p (src, XEXP (note, 0))))
4239 remove_note (insn, note);
4240 }
4241 }
4242 }
4243 }
4244 \f
4245 /* Record all the SETs in this instruction into SETS_PTR,
4246 and return the number of recorded sets. */
4247 static int
4248 find_sets_in_insn (rtx_insn *insn, struct set **psets)
4249 {
4250 struct set *sets = *psets;
4251 int n_sets = 0;
4252 rtx x = PATTERN (insn);
4253
4254 if (GET_CODE (x) == SET)
4255 {
4256 /* Ignore SETs that are unconditional jumps.
4257 They never need cse processing, so this does not hurt.
4258 The reason is not efficiency but rather
4259 so that we can test at the end for instructions
4260 that have been simplified to unconditional jumps
4261 and not be misled by unchanged instructions
4262 that were unconditional jumps to begin with. */
4263 if (SET_DEST (x) == pc_rtx
4264 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4265 ;
4266 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4267 The hard function value register is used only once, to copy to
4268 someplace else, so it isn't worth cse'ing. */
4269 else if (GET_CODE (SET_SRC (x)) == CALL)
4270 ;
4271 else
4272 sets[n_sets++].rtl = x;
4273 }
4274 else if (GET_CODE (x) == PARALLEL)
4275 {
4276 int i, lim = XVECLEN (x, 0);
4277
4278 /* Go over the expressions of the PARALLEL in forward order, to
4279 put them in the same order in the SETS array. */
4280 for (i = 0; i < lim; i++)
4281 {
4282 rtx y = XVECEXP (x, 0, i);
4283 if (GET_CODE (y) == SET)
4284 {
4285 /* As above, we ignore unconditional jumps and call-insns and
4286 ignore the result of apply_change_group. */
4287 if (SET_DEST (y) == pc_rtx
4288 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4289 ;
4290 else if (GET_CODE (SET_SRC (y)) == CALL)
4291 ;
4292 else
4293 sets[n_sets++].rtl = y;
4294 }
4295 }
4296 }
4297
4298 return n_sets;
4299 }
4300 \f
4301 /* Where possible, substitute every register reference in the N_SETS
4302 number of SETS in INSN with the canonical register.
4303
4304 Register canonicalization propagatest the earliest register (i.e.
4305 one that is set before INSN) with the same value. This is a very
4306 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4307 to RTL. For instance, a CONST for an address is usually expanded
4308 multiple times to loads into different registers, thus creating many
4309 subexpressions of the form:
4310
4311 (set (reg1) (some_const))
4312 (set (mem (... reg1 ...) (thing)))
4313 (set (reg2) (some_const))
4314 (set (mem (... reg2 ...) (thing)))
4315
4316 After canonicalizing, the code takes the following form:
4317
4318 (set (reg1) (some_const))
4319 (set (mem (... reg1 ...) (thing)))
4320 (set (reg2) (some_const))
4321 (set (mem (... reg1 ...) (thing)))
4322
4323 The set to reg2 is now trivially dead, and the memory reference (or
4324 address, or whatever) may be a candidate for further CSEing.
4325
4326 In this function, the result of apply_change_group can be ignored;
4327 see canon_reg. */
4328
4329 static void
4330 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4331 {
4332 struct set *sets = *psets;
4333 rtx tem;
4334 rtx x = PATTERN (insn);
4335 int i;
4336
4337 if (CALL_P (insn))
4338 {
4339 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4340 if (GET_CODE (XEXP (tem, 0)) != SET)
4341 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4342 }
4343
4344 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4345 {
4346 canon_reg (SET_SRC (x), insn);
4347 apply_change_group ();
4348 fold_rtx (SET_SRC (x), insn);
4349 }
4350 else if (GET_CODE (x) == CLOBBER)
4351 {
4352 /* If we clobber memory, canon the address.
4353 This does nothing when a register is clobbered
4354 because we have already invalidated the reg. */
4355 if (MEM_P (XEXP (x, 0)))
4356 canon_reg (XEXP (x, 0), insn);
4357 }
4358 else if (GET_CODE (x) == USE
4359 && ! (REG_P (XEXP (x, 0))
4360 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4361 /* Canonicalize a USE of a pseudo register or memory location. */
4362 canon_reg (x, insn);
4363 else if (GET_CODE (x) == ASM_OPERANDS)
4364 {
4365 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4366 {
4367 rtx input = ASM_OPERANDS_INPUT (x, i);
4368 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4369 {
4370 input = canon_reg (input, insn);
4371 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4372 }
4373 }
4374 }
4375 else if (GET_CODE (x) == CALL)
4376 {
4377 canon_reg (x, insn);
4378 apply_change_group ();
4379 fold_rtx (x, insn);
4380 }
4381 else if (DEBUG_INSN_P (insn))
4382 canon_reg (PATTERN (insn), insn);
4383 else if (GET_CODE (x) == PARALLEL)
4384 {
4385 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4386 {
4387 rtx y = XVECEXP (x, 0, i);
4388 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4389 {
4390 canon_reg (SET_SRC (y), insn);
4391 apply_change_group ();
4392 fold_rtx (SET_SRC (y), insn);
4393 }
4394 else if (GET_CODE (y) == CLOBBER)
4395 {
4396 if (MEM_P (XEXP (y, 0)))
4397 canon_reg (XEXP (y, 0), insn);
4398 }
4399 else if (GET_CODE (y) == USE
4400 && ! (REG_P (XEXP (y, 0))
4401 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4402 canon_reg (y, insn);
4403 else if (GET_CODE (y) == CALL)
4404 {
4405 canon_reg (y, insn);
4406 apply_change_group ();
4407 fold_rtx (y, insn);
4408 }
4409 }
4410 }
4411
4412 if (n_sets == 1 && REG_NOTES (insn) != 0
4413 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4414 {
4415 /* We potentially will process this insn many times. Therefore,
4416 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4417 unique set in INSN.
4418
4419 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4420 because cse_insn handles those specially. */
4421 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4422 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4423 remove_note (insn, tem);
4424 else
4425 {
4426 canon_reg (XEXP (tem, 0), insn);
4427 apply_change_group ();
4428 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4429 df_notes_rescan (insn);
4430 }
4431 }
4432
4433 /* Canonicalize sources and addresses of destinations.
4434 We do this in a separate pass to avoid problems when a MATCH_DUP is
4435 present in the insn pattern. In that case, we want to ensure that
4436 we don't break the duplicate nature of the pattern. So we will replace
4437 both operands at the same time. Otherwise, we would fail to find an
4438 equivalent substitution in the loop calling validate_change below.
4439
4440 We used to suppress canonicalization of DEST if it appears in SRC,
4441 but we don't do this any more. */
4442
4443 for (i = 0; i < n_sets; i++)
4444 {
4445 rtx dest = SET_DEST (sets[i].rtl);
4446 rtx src = SET_SRC (sets[i].rtl);
4447 rtx new_rtx = canon_reg (src, insn);
4448
4449 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4450
4451 if (GET_CODE (dest) == ZERO_EXTRACT)
4452 {
4453 validate_change (insn, &XEXP (dest, 1),
4454 canon_reg (XEXP (dest, 1), insn), 1);
4455 validate_change (insn, &XEXP (dest, 2),
4456 canon_reg (XEXP (dest, 2), insn), 1);
4457 }
4458
4459 while (GET_CODE (dest) == SUBREG
4460 || GET_CODE (dest) == ZERO_EXTRACT
4461 || GET_CODE (dest) == STRICT_LOW_PART)
4462 dest = XEXP (dest, 0);
4463
4464 if (MEM_P (dest))
4465 canon_reg (dest, insn);
4466 }
4467
4468 /* Now that we have done all the replacements, we can apply the change
4469 group and see if they all work. Note that this will cause some
4470 canonicalizations that would have worked individually not to be applied
4471 because some other canonicalization didn't work, but this should not
4472 occur often.
4473
4474 The result of apply_change_group can be ignored; see canon_reg. */
4475
4476 apply_change_group ();
4477 }
4478 \f
4479 /* Main function of CSE.
4480 First simplify sources and addresses of all assignments
4481 in the instruction, using previously-computed equivalents values.
4482 Then install the new sources and destinations in the table
4483 of available values. */
4484
4485 static void
4486 cse_insn (rtx_insn *insn)
4487 {
4488 rtx x = PATTERN (insn);
4489 int i;
4490 rtx tem;
4491 int n_sets = 0;
4492
4493 rtx src_eqv = 0;
4494 struct table_elt *src_eqv_elt = 0;
4495 int src_eqv_volatile = 0;
4496 int src_eqv_in_memory = 0;
4497 unsigned src_eqv_hash = 0;
4498
4499 struct set *sets = (struct set *) 0;
4500
4501 if (GET_CODE (x) == SET)
4502 sets = XALLOCA (struct set);
4503 else if (GET_CODE (x) == PARALLEL)
4504 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4505
4506 this_insn = insn;
4507 /* Records what this insn does to set CC0. */
4508 this_insn_cc0 = 0;
4509 this_insn_cc0_mode = VOIDmode;
4510
4511 /* Find all regs explicitly clobbered in this insn,
4512 to ensure they are not replaced with any other regs
4513 elsewhere in this insn. */
4514 invalidate_from_sets_and_clobbers (insn);
4515
4516 /* Record all the SETs in this instruction. */
4517 n_sets = find_sets_in_insn (insn, &sets);
4518
4519 /* Substitute the canonical register where possible. */
4520 canonicalize_insn (insn, &sets, n_sets);
4521
4522 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4523 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4524 latter condition is necessary because SRC_EQV is handled specially for
4525 this case, and if it isn't set, then there will be no equivalence
4526 for the destination. */
4527 if (n_sets == 1 && REG_NOTES (insn) != 0
4528 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4529 {
4530
4531 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4532 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4533 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4534 src_eqv = copy_rtx (XEXP (tem, 0));
4535 /* If DEST is of the form ZERO_EXTACT, as in:
4536 (set (zero_extract:SI (reg:SI 119)
4537 (const_int 16 [0x10])
4538 (const_int 16 [0x10]))
4539 (const_int 51154 [0xc7d2]))
4540 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4541 point. Note that this is different from SRC_EQV. We can however
4542 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4543 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4544 && CONST_INT_P (XEXP (tem, 0))
4545 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4546 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4547 {
4548 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4549 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4550 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4551 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4552 HOST_WIDE_INT mask;
4553 unsigned int shift;
4554 if (BITS_BIG_ENDIAN)
4555 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
4556 - INTVAL (pos) - INTVAL (width);
4557 else
4558 shift = INTVAL (pos);
4559 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4560 mask = ~(HOST_WIDE_INT) 0;
4561 else
4562 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
4563 val = (val >> shift) & mask;
4564 src_eqv = GEN_INT (val);
4565 }
4566 }
4567
4568 /* Set sets[i].src_elt to the class each source belongs to.
4569 Detect assignments from or to volatile things
4570 and set set[i] to zero so they will be ignored
4571 in the rest of this function.
4572
4573 Nothing in this loop changes the hash table or the register chains. */
4574
4575 for (i = 0; i < n_sets; i++)
4576 {
4577 bool repeat = false;
4578 rtx src, dest;
4579 rtx src_folded;
4580 struct table_elt *elt = 0, *p;
4581 machine_mode mode;
4582 rtx src_eqv_here;
4583 rtx src_const = 0;
4584 rtx src_related = 0;
4585 bool src_related_is_const_anchor = false;
4586 struct table_elt *src_const_elt = 0;
4587 int src_cost = MAX_COST;
4588 int src_eqv_cost = MAX_COST;
4589 int src_folded_cost = MAX_COST;
4590 int src_related_cost = MAX_COST;
4591 int src_elt_cost = MAX_COST;
4592 int src_regcost = MAX_COST;
4593 int src_eqv_regcost = MAX_COST;
4594 int src_folded_regcost = MAX_COST;
4595 int src_related_regcost = MAX_COST;
4596 int src_elt_regcost = MAX_COST;
4597 /* Set nonzero if we need to call force_const_mem on with the
4598 contents of src_folded before using it. */
4599 int src_folded_force_flag = 0;
4600
4601 dest = SET_DEST (sets[i].rtl);
4602 src = SET_SRC (sets[i].rtl);
4603
4604 /* If SRC is a constant that has no machine mode,
4605 hash it with the destination's machine mode.
4606 This way we can keep different modes separate. */
4607
4608 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4609 sets[i].mode = mode;
4610
4611 if (src_eqv)
4612 {
4613 machine_mode eqvmode = mode;
4614 if (GET_CODE (dest) == STRICT_LOW_PART)
4615 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4616 do_not_record = 0;
4617 hash_arg_in_memory = 0;
4618 src_eqv_hash = HASH (src_eqv, eqvmode);
4619
4620 /* Find the equivalence class for the equivalent expression. */
4621
4622 if (!do_not_record)
4623 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4624
4625 src_eqv_volatile = do_not_record;
4626 src_eqv_in_memory = hash_arg_in_memory;
4627 }
4628
4629 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4630 value of the INNER register, not the destination. So it is not
4631 a valid substitution for the source. But save it for later. */
4632 if (GET_CODE (dest) == STRICT_LOW_PART)
4633 src_eqv_here = 0;
4634 else
4635 src_eqv_here = src_eqv;
4636
4637 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4638 simplified result, which may not necessarily be valid. */
4639 src_folded = fold_rtx (src, insn);
4640
4641 #if 0
4642 /* ??? This caused bad code to be generated for the m68k port with -O2.
4643 Suppose src is (CONST_INT -1), and that after truncation src_folded
4644 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4645 At the end we will add src and src_const to the same equivalence
4646 class. We now have 3 and -1 on the same equivalence class. This
4647 causes later instructions to be mis-optimized. */
4648 /* If storing a constant in a bitfield, pre-truncate the constant
4649 so we will be able to record it later. */
4650 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4651 {
4652 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4653
4654 if (CONST_INT_P (src)
4655 && CONST_INT_P (width)
4656 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4657 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4658 src_folded
4659 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4660 << INTVAL (width)) - 1));
4661 }
4662 #endif
4663
4664 /* Compute SRC's hash code, and also notice if it
4665 should not be recorded at all. In that case,
4666 prevent any further processing of this assignment. */
4667 do_not_record = 0;
4668 hash_arg_in_memory = 0;
4669
4670 sets[i].src = src;
4671 sets[i].src_hash = HASH (src, mode);
4672 sets[i].src_volatile = do_not_record;
4673 sets[i].src_in_memory = hash_arg_in_memory;
4674
4675 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4676 a pseudo, do not record SRC. Using SRC as a replacement for
4677 anything else will be incorrect in that situation. Note that
4678 this usually occurs only for stack slots, in which case all the
4679 RTL would be referring to SRC, so we don't lose any optimization
4680 opportunities by not having SRC in the hash table. */
4681
4682 if (MEM_P (src)
4683 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4684 && REG_P (dest)
4685 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4686 sets[i].src_volatile = 1;
4687
4688 else if (GET_CODE (src) == ASM_OPERANDS
4689 && GET_CODE (x) == PARALLEL)
4690 {
4691 /* Do not record result of a non-volatile inline asm with
4692 more than one result. */
4693 if (n_sets > 1)
4694 sets[i].src_volatile = 1;
4695
4696 int j, lim = XVECLEN (x, 0);
4697 for (j = 0; j < lim; j++)
4698 {
4699 rtx y = XVECEXP (x, 0, j);
4700 /* And do not record result of a non-volatile inline asm
4701 with "memory" clobber. */
4702 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4703 {
4704 sets[i].src_volatile = 1;
4705 break;
4706 }
4707 }
4708 }
4709
4710 #if 0
4711 /* It is no longer clear why we used to do this, but it doesn't
4712 appear to still be needed. So let's try without it since this
4713 code hurts cse'ing widened ops. */
4714 /* If source is a paradoxical subreg (such as QI treated as an SI),
4715 treat it as volatile. It may do the work of an SI in one context
4716 where the extra bits are not being used, but cannot replace an SI
4717 in general. */
4718 if (paradoxical_subreg_p (src))
4719 sets[i].src_volatile = 1;
4720 #endif
4721
4722 /* Locate all possible equivalent forms for SRC. Try to replace
4723 SRC in the insn with each cheaper equivalent.
4724
4725 We have the following types of equivalents: SRC itself, a folded
4726 version, a value given in a REG_EQUAL note, or a value related
4727 to a constant.
4728
4729 Each of these equivalents may be part of an additional class
4730 of equivalents (if more than one is in the table, they must be in
4731 the same class; we check for this).
4732
4733 If the source is volatile, we don't do any table lookups.
4734
4735 We note any constant equivalent for possible later use in a
4736 REG_NOTE. */
4737
4738 if (!sets[i].src_volatile)
4739 elt = lookup (src, sets[i].src_hash, mode);
4740
4741 sets[i].src_elt = elt;
4742
4743 if (elt && src_eqv_here && src_eqv_elt)
4744 {
4745 if (elt->first_same_value != src_eqv_elt->first_same_value)
4746 {
4747 /* The REG_EQUAL is indicating that two formerly distinct
4748 classes are now equivalent. So merge them. */
4749 merge_equiv_classes (elt, src_eqv_elt);
4750 src_eqv_hash = HASH (src_eqv, elt->mode);
4751 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4752 }
4753
4754 src_eqv_here = 0;
4755 }
4756
4757 else if (src_eqv_elt)
4758 elt = src_eqv_elt;
4759
4760 /* Try to find a constant somewhere and record it in `src_const'.
4761 Record its table element, if any, in `src_const_elt'. Look in
4762 any known equivalences first. (If the constant is not in the
4763 table, also set `sets[i].src_const_hash'). */
4764 if (elt)
4765 for (p = elt->first_same_value; p; p = p->next_same_value)
4766 if (p->is_const)
4767 {
4768 src_const = p->exp;
4769 src_const_elt = elt;
4770 break;
4771 }
4772
4773 if (src_const == 0
4774 && (CONSTANT_P (src_folded)
4775 /* Consider (minus (label_ref L1) (label_ref L2)) as
4776 "constant" here so we will record it. This allows us
4777 to fold switch statements when an ADDR_DIFF_VEC is used. */
4778 || (GET_CODE (src_folded) == MINUS
4779 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4780 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4781 src_const = src_folded, src_const_elt = elt;
4782 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4783 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4784
4785 /* If we don't know if the constant is in the table, get its
4786 hash code and look it up. */
4787 if (src_const && src_const_elt == 0)
4788 {
4789 sets[i].src_const_hash = HASH (src_const, mode);
4790 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4791 }
4792
4793 sets[i].src_const = src_const;
4794 sets[i].src_const_elt = src_const_elt;
4795
4796 /* If the constant and our source are both in the table, mark them as
4797 equivalent. Otherwise, if a constant is in the table but the source
4798 isn't, set ELT to it. */
4799 if (src_const_elt && elt
4800 && src_const_elt->first_same_value != elt->first_same_value)
4801 merge_equiv_classes (elt, src_const_elt);
4802 else if (src_const_elt && elt == 0)
4803 elt = src_const_elt;
4804
4805 /* See if there is a register linearly related to a constant
4806 equivalent of SRC. */
4807 if (src_const
4808 && (GET_CODE (src_const) == CONST
4809 || (src_const_elt && src_const_elt->related_value != 0)))
4810 {
4811 src_related = use_related_value (src_const, src_const_elt);
4812 if (src_related)
4813 {
4814 struct table_elt *src_related_elt
4815 = lookup (src_related, HASH (src_related, mode), mode);
4816 if (src_related_elt && elt)
4817 {
4818 if (elt->first_same_value
4819 != src_related_elt->first_same_value)
4820 /* This can occur when we previously saw a CONST
4821 involving a SYMBOL_REF and then see the SYMBOL_REF
4822 twice. Merge the involved classes. */
4823 merge_equiv_classes (elt, src_related_elt);
4824
4825 src_related = 0;
4826 src_related_elt = 0;
4827 }
4828 else if (src_related_elt && elt == 0)
4829 elt = src_related_elt;
4830 }
4831 }
4832
4833 /* See if we have a CONST_INT that is already in a register in a
4834 wider mode. */
4835
4836 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4837 && GET_MODE_CLASS (mode) == MODE_INT
4838 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4839 {
4840 machine_mode wider_mode;
4841
4842 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4843 wider_mode != VOIDmode
4844 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4845 && src_related == 0;
4846 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4847 {
4848 struct table_elt *const_elt
4849 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4850
4851 if (const_elt == 0)
4852 continue;
4853
4854 for (const_elt = const_elt->first_same_value;
4855 const_elt; const_elt = const_elt->next_same_value)
4856 if (REG_P (const_elt->exp))
4857 {
4858 src_related = gen_lowpart (mode, const_elt->exp);
4859 break;
4860 }
4861 }
4862 }
4863
4864 /* Another possibility is that we have an AND with a constant in
4865 a mode narrower than a word. If so, it might have been generated
4866 as part of an "if" which would narrow the AND. If we already
4867 have done the AND in a wider mode, we can use a SUBREG of that
4868 value. */
4869
4870 if (flag_expensive_optimizations && ! src_related
4871 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4872 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4873 {
4874 machine_mode tmode;
4875 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4876
4877 for (tmode = GET_MODE_WIDER_MODE (mode);
4878 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4879 tmode = GET_MODE_WIDER_MODE (tmode))
4880 {
4881 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4882 struct table_elt *larger_elt;
4883
4884 if (inner)
4885 {
4886 PUT_MODE (new_and, tmode);
4887 XEXP (new_and, 0) = inner;
4888 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4889 if (larger_elt == 0)
4890 continue;
4891
4892 for (larger_elt = larger_elt->first_same_value;
4893 larger_elt; larger_elt = larger_elt->next_same_value)
4894 if (REG_P (larger_elt->exp))
4895 {
4896 src_related
4897 = gen_lowpart (mode, larger_elt->exp);
4898 break;
4899 }
4900
4901 if (src_related)
4902 break;
4903 }
4904 }
4905 }
4906
4907 /* See if a MEM has already been loaded with a widening operation;
4908 if it has, we can use a subreg of that. Many CISC machines
4909 also have such operations, but this is only likely to be
4910 beneficial on these machines. */
4911
4912 if (flag_expensive_optimizations && src_related == 0
4913 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4914 && GET_MODE_CLASS (mode) == MODE_INT
4915 && MEM_P (src) && ! do_not_record
4916 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4917 {
4918 struct rtx_def memory_extend_buf;
4919 rtx memory_extend_rtx = &memory_extend_buf;
4920 machine_mode tmode;
4921
4922 /* Set what we are trying to extend and the operation it might
4923 have been extended with. */
4924 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4925 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4926 XEXP (memory_extend_rtx, 0) = src;
4927
4928 for (tmode = GET_MODE_WIDER_MODE (mode);
4929 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4930 tmode = GET_MODE_WIDER_MODE (tmode))
4931 {
4932 struct table_elt *larger_elt;
4933
4934 PUT_MODE (memory_extend_rtx, tmode);
4935 larger_elt = lookup (memory_extend_rtx,
4936 HASH (memory_extend_rtx, tmode), tmode);
4937 if (larger_elt == 0)
4938 continue;
4939
4940 for (larger_elt = larger_elt->first_same_value;
4941 larger_elt; larger_elt = larger_elt->next_same_value)
4942 if (REG_P (larger_elt->exp))
4943 {
4944 src_related = gen_lowpart (mode, larger_elt->exp);
4945 break;
4946 }
4947
4948 if (src_related)
4949 break;
4950 }
4951 }
4952
4953 /* Try to express the constant using a register+offset expression
4954 derived from a constant anchor. */
4955
4956 if (targetm.const_anchor
4957 && !src_related
4958 && src_const
4959 && GET_CODE (src_const) == CONST_INT)
4960 {
4961 src_related = try_const_anchors (src_const, mode);
4962 src_related_is_const_anchor = src_related != NULL_RTX;
4963 }
4964
4965
4966 if (src == src_folded)
4967 src_folded = 0;
4968
4969 /* At this point, ELT, if nonzero, points to a class of expressions
4970 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4971 and SRC_RELATED, if nonzero, each contain additional equivalent
4972 expressions. Prune these latter expressions by deleting expressions
4973 already in the equivalence class.
4974
4975 Check for an equivalent identical to the destination. If found,
4976 this is the preferred equivalent since it will likely lead to
4977 elimination of the insn. Indicate this by placing it in
4978 `src_related'. */
4979
4980 if (elt)
4981 elt = elt->first_same_value;
4982 for (p = elt; p; p = p->next_same_value)
4983 {
4984 enum rtx_code code = GET_CODE (p->exp);
4985
4986 /* If the expression is not valid, ignore it. Then we do not
4987 have to check for validity below. In most cases, we can use
4988 `rtx_equal_p', since canonicalization has already been done. */
4989 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4990 continue;
4991
4992 /* Also skip paradoxical subregs, unless that's what we're
4993 looking for. */
4994 if (paradoxical_subreg_p (p->exp)
4995 && ! (src != 0
4996 && GET_CODE (src) == SUBREG
4997 && GET_MODE (src) == GET_MODE (p->exp)
4998 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4999 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5000 continue;
5001
5002 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5003 src = 0;
5004 else if (src_folded && GET_CODE (src_folded) == code
5005 && rtx_equal_p (src_folded, p->exp))
5006 src_folded = 0;
5007 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5008 && rtx_equal_p (src_eqv_here, p->exp))
5009 src_eqv_here = 0;
5010 else if (src_related && GET_CODE (src_related) == code
5011 && rtx_equal_p (src_related, p->exp))
5012 src_related = 0;
5013
5014 /* This is the same as the destination of the insns, we want
5015 to prefer it. Copy it to src_related. The code below will
5016 then give it a negative cost. */
5017 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5018 src_related = dest;
5019 }
5020
5021 /* Find the cheapest valid equivalent, trying all the available
5022 possibilities. Prefer items not in the hash table to ones
5023 that are when they are equal cost. Note that we can never
5024 worsen an insn as the current contents will also succeed.
5025 If we find an equivalent identical to the destination, use it as best,
5026 since this insn will probably be eliminated in that case. */
5027 if (src)
5028 {
5029 if (rtx_equal_p (src, dest))
5030 src_cost = src_regcost = -1;
5031 else
5032 {
5033 src_cost = COST (src, mode);
5034 src_regcost = approx_reg_cost (src);
5035 }
5036 }
5037
5038 if (src_eqv_here)
5039 {
5040 if (rtx_equal_p (src_eqv_here, dest))
5041 src_eqv_cost = src_eqv_regcost = -1;
5042 else
5043 {
5044 src_eqv_cost = COST (src_eqv_here, mode);
5045 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5046 }
5047 }
5048
5049 if (src_folded)
5050 {
5051 if (rtx_equal_p (src_folded, dest))
5052 src_folded_cost = src_folded_regcost = -1;
5053 else
5054 {
5055 src_folded_cost = COST (src_folded, mode);
5056 src_folded_regcost = approx_reg_cost (src_folded);
5057 }
5058 }
5059
5060 if (src_related)
5061 {
5062 if (rtx_equal_p (src_related, dest))
5063 src_related_cost = src_related_regcost = -1;
5064 else
5065 {
5066 src_related_cost = COST (src_related, mode);
5067 src_related_regcost = approx_reg_cost (src_related);
5068
5069 /* If a const-anchor is used to synthesize a constant that
5070 normally requires multiple instructions then slightly prefer
5071 it over the original sequence. These instructions are likely
5072 to become redundant now. We can't compare against the cost
5073 of src_eqv_here because, on MIPS for example, multi-insn
5074 constants have zero cost; they are assumed to be hoisted from
5075 loops. */
5076 if (src_related_is_const_anchor
5077 && src_related_cost == src_cost
5078 && src_eqv_here)
5079 src_related_cost--;
5080 }
5081 }
5082
5083 /* If this was an indirect jump insn, a known label will really be
5084 cheaper even though it looks more expensive. */
5085 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5086 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5087
5088 /* Terminate loop when replacement made. This must terminate since
5089 the current contents will be tested and will always be valid. */
5090 while (1)
5091 {
5092 rtx trial;
5093
5094 /* Skip invalid entries. */
5095 while (elt && !REG_P (elt->exp)
5096 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5097 elt = elt->next_same_value;
5098
5099 /* A paradoxical subreg would be bad here: it'll be the right
5100 size, but later may be adjusted so that the upper bits aren't
5101 what we want. So reject it. */
5102 if (elt != 0
5103 && paradoxical_subreg_p (elt->exp)
5104 /* It is okay, though, if the rtx we're trying to match
5105 will ignore any of the bits we can't predict. */
5106 && ! (src != 0
5107 && GET_CODE (src) == SUBREG
5108 && GET_MODE (src) == GET_MODE (elt->exp)
5109 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5110 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5111 {
5112 elt = elt->next_same_value;
5113 continue;
5114 }
5115
5116 if (elt)
5117 {
5118 src_elt_cost = elt->cost;
5119 src_elt_regcost = elt->regcost;
5120 }
5121
5122 /* Find cheapest and skip it for the next time. For items
5123 of equal cost, use this order:
5124 src_folded, src, src_eqv, src_related and hash table entry. */
5125 if (src_folded
5126 && preferable (src_folded_cost, src_folded_regcost,
5127 src_cost, src_regcost) <= 0
5128 && preferable (src_folded_cost, src_folded_regcost,
5129 src_eqv_cost, src_eqv_regcost) <= 0
5130 && preferable (src_folded_cost, src_folded_regcost,
5131 src_related_cost, src_related_regcost) <= 0
5132 && preferable (src_folded_cost, src_folded_regcost,
5133 src_elt_cost, src_elt_regcost) <= 0)
5134 {
5135 trial = src_folded, src_folded_cost = MAX_COST;
5136 if (src_folded_force_flag)
5137 {
5138 rtx forced = force_const_mem (mode, trial);
5139 if (forced)
5140 trial = forced;
5141 }
5142 }
5143 else if (src
5144 && preferable (src_cost, src_regcost,
5145 src_eqv_cost, src_eqv_regcost) <= 0
5146 && preferable (src_cost, src_regcost,
5147 src_related_cost, src_related_regcost) <= 0
5148 && preferable (src_cost, src_regcost,
5149 src_elt_cost, src_elt_regcost) <= 0)
5150 trial = src, src_cost = MAX_COST;
5151 else if (src_eqv_here
5152 && preferable (src_eqv_cost, src_eqv_regcost,
5153 src_related_cost, src_related_regcost) <= 0
5154 && preferable (src_eqv_cost, src_eqv_regcost,
5155 src_elt_cost, src_elt_regcost) <= 0)
5156 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5157 else if (src_related
5158 && preferable (src_related_cost, src_related_regcost,
5159 src_elt_cost, src_elt_regcost) <= 0)
5160 trial = src_related, src_related_cost = MAX_COST;
5161 else
5162 {
5163 trial = elt->exp;
5164 elt = elt->next_same_value;
5165 src_elt_cost = MAX_COST;
5166 }
5167
5168 /* Avoid creation of overlapping memory moves. */
5169 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5170 {
5171 rtx src, dest;
5172
5173 /* BLKmode moves are not handled by cse anyway. */
5174 if (GET_MODE (trial) == BLKmode)
5175 break;
5176
5177 src = canon_rtx (trial);
5178 dest = canon_rtx (SET_DEST (sets[i].rtl));
5179
5180 if (!MEM_P (src) || !MEM_P (dest)
5181 || !nonoverlapping_memrefs_p (src, dest, false))
5182 break;
5183 }
5184
5185 /* Try to optimize
5186 (set (reg:M N) (const_int A))
5187 (set (reg:M2 O) (const_int B))
5188 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5189 (reg:M2 O)). */
5190 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5191 && CONST_INT_P (trial)
5192 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5193 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5194 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5195 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5196 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5197 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5198 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5199 <= HOST_BITS_PER_WIDE_INT))
5200 {
5201 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5202 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5203 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5204 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5205 struct table_elt *dest_elt
5206 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5207 rtx dest_cst = NULL;
5208
5209 if (dest_elt)
5210 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5211 if (p->is_const && CONST_INT_P (p->exp))
5212 {
5213 dest_cst = p->exp;
5214 break;
5215 }
5216 if (dest_cst)
5217 {
5218 HOST_WIDE_INT val = INTVAL (dest_cst);
5219 HOST_WIDE_INT mask;
5220 unsigned int shift;
5221 if (BITS_BIG_ENDIAN)
5222 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5223 - INTVAL (pos) - INTVAL (width);
5224 else
5225 shift = INTVAL (pos);
5226 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5227 mask = ~(HOST_WIDE_INT) 0;
5228 else
5229 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5230 val &= ~(mask << shift);
5231 val |= (INTVAL (trial) & mask) << shift;
5232 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5233 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5234 dest_reg, 1);
5235 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5236 GEN_INT (val), 1);
5237 if (apply_change_group ())
5238 {
5239 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5240 if (note)
5241 {
5242 remove_note (insn, note);
5243 df_notes_rescan (insn);
5244 }
5245 src_eqv = NULL_RTX;
5246 src_eqv_elt = NULL;
5247 src_eqv_volatile = 0;
5248 src_eqv_in_memory = 0;
5249 src_eqv_hash = 0;
5250 repeat = true;
5251 break;
5252 }
5253 }
5254 }
5255
5256 /* We don't normally have an insn matching (set (pc) (pc)), so
5257 check for this separately here. We will delete such an
5258 insn below.
5259
5260 For other cases such as a table jump or conditional jump
5261 where we know the ultimate target, go ahead and replace the
5262 operand. While that may not make a valid insn, we will
5263 reemit the jump below (and also insert any necessary
5264 barriers). */
5265 if (n_sets == 1 && dest == pc_rtx
5266 && (trial == pc_rtx
5267 || (GET_CODE (trial) == LABEL_REF
5268 && ! condjump_p (insn))))
5269 {
5270 /* Don't substitute non-local labels, this confuses CFG. */
5271 if (GET_CODE (trial) == LABEL_REF
5272 && LABEL_REF_NONLOCAL_P (trial))
5273 continue;
5274
5275 SET_SRC (sets[i].rtl) = trial;
5276 cse_jumps_altered = true;
5277 break;
5278 }
5279
5280 /* Reject certain invalid forms of CONST that we create. */
5281 else if (CONSTANT_P (trial)
5282 && GET_CODE (trial) == CONST
5283 /* Reject cases that will cause decode_rtx_const to
5284 die. On the alpha when simplifying a switch, we
5285 get (const (truncate (minus (label_ref)
5286 (label_ref)))). */
5287 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5288 /* Likewise on IA-64, except without the
5289 truncate. */
5290 || (GET_CODE (XEXP (trial, 0)) == MINUS
5291 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5292 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5293 /* Do nothing for this case. */
5294 ;
5295
5296 /* Look for a substitution that makes a valid insn. */
5297 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5298 trial, 0))
5299 {
5300 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5301
5302 /* The result of apply_change_group can be ignored; see
5303 canon_reg. */
5304
5305 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5306 apply_change_group ();
5307
5308 break;
5309 }
5310
5311 /* If we previously found constant pool entries for
5312 constants and this is a constant, try making a
5313 pool entry. Put it in src_folded unless we already have done
5314 this since that is where it likely came from. */
5315
5316 else if (constant_pool_entries_cost
5317 && CONSTANT_P (trial)
5318 && (src_folded == 0
5319 || (!MEM_P (src_folded)
5320 && ! src_folded_force_flag))
5321 && GET_MODE_CLASS (mode) != MODE_CC
5322 && mode != VOIDmode)
5323 {
5324 src_folded_force_flag = 1;
5325 src_folded = trial;
5326 src_folded_cost = constant_pool_entries_cost;
5327 src_folded_regcost = constant_pool_entries_regcost;
5328 }
5329 }
5330
5331 /* If we changed the insn too much, handle this set from scratch. */
5332 if (repeat)
5333 {
5334 i--;
5335 continue;
5336 }
5337
5338 src = SET_SRC (sets[i].rtl);
5339
5340 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5341 However, there is an important exception: If both are registers
5342 that are not the head of their equivalence class, replace SET_SRC
5343 with the head of the class. If we do not do this, we will have
5344 both registers live over a portion of the basic block. This way,
5345 their lifetimes will likely abut instead of overlapping. */
5346 if (REG_P (dest)
5347 && REGNO_QTY_VALID_P (REGNO (dest)))
5348 {
5349 int dest_q = REG_QTY (REGNO (dest));
5350 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5351
5352 if (dest_ent->mode == GET_MODE (dest)
5353 && dest_ent->first_reg != REGNO (dest)
5354 && REG_P (src) && REGNO (src) == REGNO (dest)
5355 /* Don't do this if the original insn had a hard reg as
5356 SET_SRC or SET_DEST. */
5357 && (!REG_P (sets[i].src)
5358 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5359 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5360 /* We can't call canon_reg here because it won't do anything if
5361 SRC is a hard register. */
5362 {
5363 int src_q = REG_QTY (REGNO (src));
5364 struct qty_table_elem *src_ent = &qty_table[src_q];
5365 int first = src_ent->first_reg;
5366 rtx new_src
5367 = (first >= FIRST_PSEUDO_REGISTER
5368 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5369
5370 /* We must use validate-change even for this, because this
5371 might be a special no-op instruction, suitable only to
5372 tag notes onto. */
5373 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5374 {
5375 src = new_src;
5376 /* If we had a constant that is cheaper than what we are now
5377 setting SRC to, use that constant. We ignored it when we
5378 thought we could make this into a no-op. */
5379 if (src_const && COST (src_const, mode) < COST (src, mode)
5380 && validate_change (insn, &SET_SRC (sets[i].rtl),
5381 src_const, 0))
5382 src = src_const;
5383 }
5384 }
5385 }
5386
5387 /* If we made a change, recompute SRC values. */
5388 if (src != sets[i].src)
5389 {
5390 do_not_record = 0;
5391 hash_arg_in_memory = 0;
5392 sets[i].src = src;
5393 sets[i].src_hash = HASH (src, mode);
5394 sets[i].src_volatile = do_not_record;
5395 sets[i].src_in_memory = hash_arg_in_memory;
5396 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5397 }
5398
5399 /* If this is a single SET, we are setting a register, and we have an
5400 equivalent constant, we want to add a REG_EQUAL note if the constant
5401 is different from the source. We don't want to do it for a constant
5402 pseudo since verifying that this pseudo hasn't been eliminated is a
5403 pain; moreover such a note won't help anything.
5404
5405 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5406 which can be created for a reference to a compile time computable
5407 entry in a jump table. */
5408 if (n_sets == 1
5409 && REG_P (dest)
5410 && src_const
5411 && !REG_P (src_const)
5412 && !(GET_CODE (src_const) == SUBREG
5413 && REG_P (SUBREG_REG (src_const)))
5414 && !(GET_CODE (src_const) == CONST
5415 && GET_CODE (XEXP (src_const, 0)) == MINUS
5416 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5417 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5418 && !rtx_equal_p (src, src_const))
5419 {
5420 /* Make sure that the rtx is not shared. */
5421 src_const = copy_rtx (src_const);
5422
5423 /* Record the actual constant value in a REG_EQUAL note,
5424 making a new one if one does not already exist. */
5425 set_unique_reg_note (insn, REG_EQUAL, src_const);
5426 df_notes_rescan (insn);
5427 }
5428
5429 /* Now deal with the destination. */
5430 do_not_record = 0;
5431
5432 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5433 while (GET_CODE (dest) == SUBREG
5434 || GET_CODE (dest) == ZERO_EXTRACT
5435 || GET_CODE (dest) == STRICT_LOW_PART)
5436 dest = XEXP (dest, 0);
5437
5438 sets[i].inner_dest = dest;
5439
5440 if (MEM_P (dest))
5441 {
5442 #ifdef PUSH_ROUNDING
5443 /* Stack pushes invalidate the stack pointer. */
5444 rtx addr = XEXP (dest, 0);
5445 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5446 && XEXP (addr, 0) == stack_pointer_rtx)
5447 invalidate (stack_pointer_rtx, VOIDmode);
5448 #endif
5449 dest = fold_rtx (dest, insn);
5450 }
5451
5452 /* Compute the hash code of the destination now,
5453 before the effects of this instruction are recorded,
5454 since the register values used in the address computation
5455 are those before this instruction. */
5456 sets[i].dest_hash = HASH (dest, mode);
5457
5458 /* Don't enter a bit-field in the hash table
5459 because the value in it after the store
5460 may not equal what was stored, due to truncation. */
5461
5462 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5463 {
5464 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5465
5466 if (src_const != 0 && CONST_INT_P (src_const)
5467 && CONST_INT_P (width)
5468 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5469 && ! (INTVAL (src_const)
5470 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5471 /* Exception: if the value is constant,
5472 and it won't be truncated, record it. */
5473 ;
5474 else
5475 {
5476 /* This is chosen so that the destination will be invalidated
5477 but no new value will be recorded.
5478 We must invalidate because sometimes constant
5479 values can be recorded for bitfields. */
5480 sets[i].src_elt = 0;
5481 sets[i].src_volatile = 1;
5482 src_eqv = 0;
5483 src_eqv_elt = 0;
5484 }
5485 }
5486
5487 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5488 the insn. */
5489 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5490 {
5491 /* One less use of the label this insn used to jump to. */
5492 delete_insn_and_edges (insn);
5493 cse_jumps_altered = true;
5494 /* No more processing for this set. */
5495 sets[i].rtl = 0;
5496 }
5497
5498 /* If this SET is now setting PC to a label, we know it used to
5499 be a conditional or computed branch. */
5500 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5501 && !LABEL_REF_NONLOCAL_P (src))
5502 {
5503 /* We reemit the jump in as many cases as possible just in
5504 case the form of an unconditional jump is significantly
5505 different than a computed jump or conditional jump.
5506
5507 If this insn has multiple sets, then reemitting the
5508 jump is nontrivial. So instead we just force rerecognition
5509 and hope for the best. */
5510 if (n_sets == 1)
5511 {
5512 rtx_jump_insn *new_rtx;
5513 rtx note;
5514
5515 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5516 new_rtx = emit_jump_insn_before (seq, insn);
5517 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5518 LABEL_NUSES (XEXP (src, 0))++;
5519
5520 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5521 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5522 if (note)
5523 {
5524 XEXP (note, 1) = NULL_RTX;
5525 REG_NOTES (new_rtx) = note;
5526 }
5527
5528 delete_insn_and_edges (insn);
5529 insn = new_rtx;
5530 }
5531 else
5532 INSN_CODE (insn) = -1;
5533
5534 /* Do not bother deleting any unreachable code, let jump do it. */
5535 cse_jumps_altered = true;
5536 sets[i].rtl = 0;
5537 }
5538
5539 /* If destination is volatile, invalidate it and then do no further
5540 processing for this assignment. */
5541
5542 else if (do_not_record)
5543 {
5544 invalidate_dest (dest);
5545 sets[i].rtl = 0;
5546 }
5547
5548 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5549 {
5550 do_not_record = 0;
5551 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5552 if (do_not_record)
5553 {
5554 invalidate_dest (SET_DEST (sets[i].rtl));
5555 sets[i].rtl = 0;
5556 }
5557 }
5558
5559 /* If setting CC0, record what it was set to, or a constant, if it
5560 is equivalent to a constant. If it is being set to a floating-point
5561 value, make a COMPARE with the appropriate constant of 0. If we
5562 don't do this, later code can interpret this as a test against
5563 const0_rtx, which can cause problems if we try to put it into an
5564 insn as a floating-point operand. */
5565 if (dest == cc0_rtx)
5566 {
5567 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5568 this_insn_cc0_mode = mode;
5569 if (FLOAT_MODE_P (mode))
5570 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5571 CONST0_RTX (mode));
5572 }
5573 }
5574
5575 /* Now enter all non-volatile source expressions in the hash table
5576 if they are not already present.
5577 Record their equivalence classes in src_elt.
5578 This way we can insert the corresponding destinations into
5579 the same classes even if the actual sources are no longer in them
5580 (having been invalidated). */
5581
5582 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5583 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5584 {
5585 struct table_elt *elt;
5586 struct table_elt *classp = sets[0].src_elt;
5587 rtx dest = SET_DEST (sets[0].rtl);
5588 machine_mode eqvmode = GET_MODE (dest);
5589
5590 if (GET_CODE (dest) == STRICT_LOW_PART)
5591 {
5592 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5593 classp = 0;
5594 }
5595 if (insert_regs (src_eqv, classp, 0))
5596 {
5597 rehash_using_reg (src_eqv);
5598 src_eqv_hash = HASH (src_eqv, eqvmode);
5599 }
5600 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5601 elt->in_memory = src_eqv_in_memory;
5602 src_eqv_elt = elt;
5603
5604 /* Check to see if src_eqv_elt is the same as a set source which
5605 does not yet have an elt, and if so set the elt of the set source
5606 to src_eqv_elt. */
5607 for (i = 0; i < n_sets; i++)
5608 if (sets[i].rtl && sets[i].src_elt == 0
5609 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5610 sets[i].src_elt = src_eqv_elt;
5611 }
5612
5613 for (i = 0; i < n_sets; i++)
5614 if (sets[i].rtl && ! sets[i].src_volatile
5615 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5616 {
5617 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5618 {
5619 /* REG_EQUAL in setting a STRICT_LOW_PART
5620 gives an equivalent for the entire destination register,
5621 not just for the subreg being stored in now.
5622 This is a more interesting equivalence, so we arrange later
5623 to treat the entire reg as the destination. */
5624 sets[i].src_elt = src_eqv_elt;
5625 sets[i].src_hash = src_eqv_hash;
5626 }
5627 else
5628 {
5629 /* Insert source and constant equivalent into hash table, if not
5630 already present. */
5631 struct table_elt *classp = src_eqv_elt;
5632 rtx src = sets[i].src;
5633 rtx dest = SET_DEST (sets[i].rtl);
5634 machine_mode mode
5635 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5636
5637 /* It's possible that we have a source value known to be
5638 constant but don't have a REG_EQUAL note on the insn.
5639 Lack of a note will mean src_eqv_elt will be NULL. This
5640 can happen where we've generated a SUBREG to access a
5641 CONST_INT that is already in a register in a wider mode.
5642 Ensure that the source expression is put in the proper
5643 constant class. */
5644 if (!classp)
5645 classp = sets[i].src_const_elt;
5646
5647 if (sets[i].src_elt == 0)
5648 {
5649 struct table_elt *elt;
5650
5651 /* Note that these insert_regs calls cannot remove
5652 any of the src_elt's, because they would have failed to
5653 match if not still valid. */
5654 if (insert_regs (src, classp, 0))
5655 {
5656 rehash_using_reg (src);
5657 sets[i].src_hash = HASH (src, mode);
5658 }
5659 elt = insert (src, classp, sets[i].src_hash, mode);
5660 elt->in_memory = sets[i].src_in_memory;
5661 /* If inline asm has any clobbers, ensure we only reuse
5662 existing inline asms and never try to put the ASM_OPERANDS
5663 into an insn that isn't inline asm. */
5664 if (GET_CODE (src) == ASM_OPERANDS
5665 && GET_CODE (x) == PARALLEL)
5666 elt->cost = MAX_COST;
5667 sets[i].src_elt = classp = elt;
5668 }
5669 if (sets[i].src_const && sets[i].src_const_elt == 0
5670 && src != sets[i].src_const
5671 && ! rtx_equal_p (sets[i].src_const, src))
5672 sets[i].src_elt = insert (sets[i].src_const, classp,
5673 sets[i].src_const_hash, mode);
5674 }
5675 }
5676 else if (sets[i].src_elt == 0)
5677 /* If we did not insert the source into the hash table (e.g., it was
5678 volatile), note the equivalence class for the REG_EQUAL value, if any,
5679 so that the destination goes into that class. */
5680 sets[i].src_elt = src_eqv_elt;
5681
5682 /* Record destination addresses in the hash table. This allows us to
5683 check if they are invalidated by other sets. */
5684 for (i = 0; i < n_sets; i++)
5685 {
5686 if (sets[i].rtl)
5687 {
5688 rtx x = sets[i].inner_dest;
5689 struct table_elt *elt;
5690 machine_mode mode;
5691 unsigned hash;
5692
5693 if (MEM_P (x))
5694 {
5695 x = XEXP (x, 0);
5696 mode = GET_MODE (x);
5697 hash = HASH (x, mode);
5698 elt = lookup (x, hash, mode);
5699 if (!elt)
5700 {
5701 if (insert_regs (x, NULL, 0))
5702 {
5703 rtx dest = SET_DEST (sets[i].rtl);
5704
5705 rehash_using_reg (x);
5706 hash = HASH (x, mode);
5707 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5708 }
5709 elt = insert (x, NULL, hash, mode);
5710 }
5711
5712 sets[i].dest_addr_elt = elt;
5713 }
5714 else
5715 sets[i].dest_addr_elt = NULL;
5716 }
5717 }
5718
5719 invalidate_from_clobbers (insn);
5720
5721 /* Some registers are invalidated by subroutine calls. Memory is
5722 invalidated by non-constant calls. */
5723
5724 if (CALL_P (insn))
5725 {
5726 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5727 invalidate_memory ();
5728 invalidate_for_call ();
5729 }
5730
5731 /* Now invalidate everything set by this instruction.
5732 If a SUBREG or other funny destination is being set,
5733 sets[i].rtl is still nonzero, so here we invalidate the reg
5734 a part of which is being set. */
5735
5736 for (i = 0; i < n_sets; i++)
5737 if (sets[i].rtl)
5738 {
5739 /* We can't use the inner dest, because the mode associated with
5740 a ZERO_EXTRACT is significant. */
5741 rtx dest = SET_DEST (sets[i].rtl);
5742
5743 /* Needed for registers to remove the register from its
5744 previous quantity's chain.
5745 Needed for memory if this is a nonvarying address, unless
5746 we have just done an invalidate_memory that covers even those. */
5747 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5748 invalidate (dest, VOIDmode);
5749 else if (MEM_P (dest))
5750 invalidate (dest, VOIDmode);
5751 else if (GET_CODE (dest) == STRICT_LOW_PART
5752 || GET_CODE (dest) == ZERO_EXTRACT)
5753 invalidate (XEXP (dest, 0), GET_MODE (dest));
5754 }
5755
5756 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5757 the regs restored by the longjmp come from a later time
5758 than the setjmp. */
5759 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5760 {
5761 flush_hash_table ();
5762 goto done;
5763 }
5764
5765 /* Make sure registers mentioned in destinations
5766 are safe for use in an expression to be inserted.
5767 This removes from the hash table
5768 any invalid entry that refers to one of these registers.
5769
5770 We don't care about the return value from mention_regs because
5771 we are going to hash the SET_DEST values unconditionally. */
5772
5773 for (i = 0; i < n_sets; i++)
5774 {
5775 if (sets[i].rtl)
5776 {
5777 rtx x = SET_DEST (sets[i].rtl);
5778
5779 if (!REG_P (x))
5780 mention_regs (x);
5781 else
5782 {
5783 /* We used to rely on all references to a register becoming
5784 inaccessible when a register changes to a new quantity,
5785 since that changes the hash code. However, that is not
5786 safe, since after HASH_SIZE new quantities we get a
5787 hash 'collision' of a register with its own invalid
5788 entries. And since SUBREGs have been changed not to
5789 change their hash code with the hash code of the register,
5790 it wouldn't work any longer at all. So we have to check
5791 for any invalid references lying around now.
5792 This code is similar to the REG case in mention_regs,
5793 but it knows that reg_tick has been incremented, and
5794 it leaves reg_in_table as -1 . */
5795 unsigned int regno = REGNO (x);
5796 unsigned int endregno = END_REGNO (x);
5797 unsigned int i;
5798
5799 for (i = regno; i < endregno; i++)
5800 {
5801 if (REG_IN_TABLE (i) >= 0)
5802 {
5803 remove_invalid_refs (i);
5804 REG_IN_TABLE (i) = -1;
5805 }
5806 }
5807 }
5808 }
5809 }
5810
5811 /* We may have just removed some of the src_elt's from the hash table.
5812 So replace each one with the current head of the same class.
5813 Also check if destination addresses have been removed. */
5814
5815 for (i = 0; i < n_sets; i++)
5816 if (sets[i].rtl)
5817 {
5818 if (sets[i].dest_addr_elt
5819 && sets[i].dest_addr_elt->first_same_value == 0)
5820 {
5821 /* The elt was removed, which means this destination is not
5822 valid after this instruction. */
5823 sets[i].rtl = NULL_RTX;
5824 }
5825 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5826 /* If elt was removed, find current head of same class,
5827 or 0 if nothing remains of that class. */
5828 {
5829 struct table_elt *elt = sets[i].src_elt;
5830
5831 while (elt && elt->prev_same_value)
5832 elt = elt->prev_same_value;
5833
5834 while (elt && elt->first_same_value == 0)
5835 elt = elt->next_same_value;
5836 sets[i].src_elt = elt ? elt->first_same_value : 0;
5837 }
5838 }
5839
5840 /* Now insert the destinations into their equivalence classes. */
5841
5842 for (i = 0; i < n_sets; i++)
5843 if (sets[i].rtl)
5844 {
5845 rtx dest = SET_DEST (sets[i].rtl);
5846 struct table_elt *elt;
5847
5848 /* Don't record value if we are not supposed to risk allocating
5849 floating-point values in registers that might be wider than
5850 memory. */
5851 if ((flag_float_store
5852 && MEM_P (dest)
5853 && FLOAT_MODE_P (GET_MODE (dest)))
5854 /* Don't record BLKmode values, because we don't know the
5855 size of it, and can't be sure that other BLKmode values
5856 have the same or smaller size. */
5857 || GET_MODE (dest) == BLKmode
5858 /* If we didn't put a REG_EQUAL value or a source into the hash
5859 table, there is no point is recording DEST. */
5860 || sets[i].src_elt == 0
5861 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5862 or SIGN_EXTEND, don't record DEST since it can cause
5863 some tracking to be wrong.
5864
5865 ??? Think about this more later. */
5866 || (paradoxical_subreg_p (dest)
5867 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5868 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5869 continue;
5870
5871 /* STRICT_LOW_PART isn't part of the value BEING set,
5872 and neither is the SUBREG inside it.
5873 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5874 if (GET_CODE (dest) == STRICT_LOW_PART)
5875 dest = SUBREG_REG (XEXP (dest, 0));
5876
5877 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5878 /* Registers must also be inserted into chains for quantities. */
5879 if (insert_regs (dest, sets[i].src_elt, 1))
5880 {
5881 /* If `insert_regs' changes something, the hash code must be
5882 recalculated. */
5883 rehash_using_reg (dest);
5884 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5885 }
5886
5887 elt = insert (dest, sets[i].src_elt,
5888 sets[i].dest_hash, GET_MODE (dest));
5889
5890 /* If this is a constant, insert the constant anchors with the
5891 equivalent register-offset expressions using register DEST. */
5892 if (targetm.const_anchor
5893 && REG_P (dest)
5894 && SCALAR_INT_MODE_P (GET_MODE (dest))
5895 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5896 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5897
5898 elt->in_memory = (MEM_P (sets[i].inner_dest)
5899 && !MEM_READONLY_P (sets[i].inner_dest));
5900
5901 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5902 narrower than M2, and both M1 and M2 are the same number of words,
5903 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5904 make that equivalence as well.
5905
5906 However, BAR may have equivalences for which gen_lowpart
5907 will produce a simpler value than gen_lowpart applied to
5908 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5909 BAR's equivalences. If we don't get a simplified form, make
5910 the SUBREG. It will not be used in an equivalence, but will
5911 cause two similar assignments to be detected.
5912
5913 Note the loop below will find SUBREG_REG (DEST) since we have
5914 already entered SRC and DEST of the SET in the table. */
5915
5916 if (GET_CODE (dest) == SUBREG
5917 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5918 / UNITS_PER_WORD)
5919 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5920 && (GET_MODE_SIZE (GET_MODE (dest))
5921 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5922 && sets[i].src_elt != 0)
5923 {
5924 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5925 struct table_elt *elt, *classp = 0;
5926
5927 for (elt = sets[i].src_elt->first_same_value; elt;
5928 elt = elt->next_same_value)
5929 {
5930 rtx new_src = 0;
5931 unsigned src_hash;
5932 struct table_elt *src_elt;
5933 int byte = 0;
5934
5935 /* Ignore invalid entries. */
5936 if (!REG_P (elt->exp)
5937 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5938 continue;
5939
5940 /* We may have already been playing subreg games. If the
5941 mode is already correct for the destination, use it. */
5942 if (GET_MODE (elt->exp) == new_mode)
5943 new_src = elt->exp;
5944 else
5945 {
5946 /* Calculate big endian correction for the SUBREG_BYTE.
5947 We have already checked that M1 (GET_MODE (dest))
5948 is not narrower than M2 (new_mode). */
5949 if (BYTES_BIG_ENDIAN)
5950 byte = (GET_MODE_SIZE (GET_MODE (dest))
5951 - GET_MODE_SIZE (new_mode));
5952
5953 new_src = simplify_gen_subreg (new_mode, elt->exp,
5954 GET_MODE (dest), byte);
5955 }
5956
5957 /* The call to simplify_gen_subreg fails if the value
5958 is VOIDmode, yet we can't do any simplification, e.g.
5959 for EXPR_LISTs denoting function call results.
5960 It is invalid to construct a SUBREG with a VOIDmode
5961 SUBREG_REG, hence a zero new_src means we can't do
5962 this substitution. */
5963 if (! new_src)
5964 continue;
5965
5966 src_hash = HASH (new_src, new_mode);
5967 src_elt = lookup (new_src, src_hash, new_mode);
5968
5969 /* Put the new source in the hash table is if isn't
5970 already. */
5971 if (src_elt == 0)
5972 {
5973 if (insert_regs (new_src, classp, 0))
5974 {
5975 rehash_using_reg (new_src);
5976 src_hash = HASH (new_src, new_mode);
5977 }
5978 src_elt = insert (new_src, classp, src_hash, new_mode);
5979 src_elt->in_memory = elt->in_memory;
5980 if (GET_CODE (new_src) == ASM_OPERANDS
5981 && elt->cost == MAX_COST)
5982 src_elt->cost = MAX_COST;
5983 }
5984 else if (classp && classp != src_elt->first_same_value)
5985 /* Show that two things that we've seen before are
5986 actually the same. */
5987 merge_equiv_classes (src_elt, classp);
5988
5989 classp = src_elt->first_same_value;
5990 /* Ignore invalid entries. */
5991 while (classp
5992 && !REG_P (classp->exp)
5993 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5994 classp = classp->next_same_value;
5995 }
5996 }
5997 }
5998
5999 /* Special handling for (set REG0 REG1) where REG0 is the
6000 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6001 be used in the sequel, so (if easily done) change this insn to
6002 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6003 that computed their value. Then REG1 will become a dead store
6004 and won't cloud the situation for later optimizations.
6005
6006 Do not make this change if REG1 is a hard register, because it will
6007 then be used in the sequel and we may be changing a two-operand insn
6008 into a three-operand insn.
6009
6010 Also do not do this if we are operating on a copy of INSN. */
6011
6012 if (n_sets == 1 && sets[0].rtl)
6013 try_back_substitute_reg (sets[0].rtl, insn);
6014
6015 done:;
6016 }
6017 \f
6018 /* Remove from the hash table all expressions that reference memory. */
6019
6020 static void
6021 invalidate_memory (void)
6022 {
6023 int i;
6024 struct table_elt *p, *next;
6025
6026 for (i = 0; i < HASH_SIZE; i++)
6027 for (p = table[i]; p; p = next)
6028 {
6029 next = p->next_same_hash;
6030 if (p->in_memory)
6031 remove_from_table (p, i);
6032 }
6033 }
6034
6035 /* Perform invalidation on the basis of everything about INSN,
6036 except for invalidating the actual places that are SET in it.
6037 This includes the places CLOBBERed, and anything that might
6038 alias with something that is SET or CLOBBERed. */
6039
6040 static void
6041 invalidate_from_clobbers (rtx_insn *insn)
6042 {
6043 rtx x = PATTERN (insn);
6044
6045 if (GET_CODE (x) == CLOBBER)
6046 {
6047 rtx ref = XEXP (x, 0);
6048 if (ref)
6049 {
6050 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6051 || MEM_P (ref))
6052 invalidate (ref, VOIDmode);
6053 else if (GET_CODE (ref) == STRICT_LOW_PART
6054 || GET_CODE (ref) == ZERO_EXTRACT)
6055 invalidate (XEXP (ref, 0), GET_MODE (ref));
6056 }
6057 }
6058 else if (GET_CODE (x) == PARALLEL)
6059 {
6060 int i;
6061 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6062 {
6063 rtx y = XVECEXP (x, 0, i);
6064 if (GET_CODE (y) == CLOBBER)
6065 {
6066 rtx ref = XEXP (y, 0);
6067 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6068 || MEM_P (ref))
6069 invalidate (ref, VOIDmode);
6070 else if (GET_CODE (ref) == STRICT_LOW_PART
6071 || GET_CODE (ref) == ZERO_EXTRACT)
6072 invalidate (XEXP (ref, 0), GET_MODE (ref));
6073 }
6074 }
6075 }
6076 }
6077 \f
6078 /* Perform invalidation on the basis of everything about INSN.
6079 This includes the places CLOBBERed, and anything that might
6080 alias with something that is SET or CLOBBERed. */
6081
6082 static void
6083 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6084 {
6085 rtx tem;
6086 rtx x = PATTERN (insn);
6087
6088 if (CALL_P (insn))
6089 {
6090 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6091 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6092 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6093 }
6094
6095 /* Ensure we invalidate the destination register of a CALL insn.
6096 This is necessary for machines where this register is a fixed_reg,
6097 because no other code would invalidate it. */
6098 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6099 invalidate (SET_DEST (x), VOIDmode);
6100
6101 else if (GET_CODE (x) == PARALLEL)
6102 {
6103 int i;
6104
6105 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6106 {
6107 rtx y = XVECEXP (x, 0, i);
6108 if (GET_CODE (y) == CLOBBER)
6109 {
6110 rtx clobbered = XEXP (y, 0);
6111
6112 if (REG_P (clobbered)
6113 || GET_CODE (clobbered) == SUBREG)
6114 invalidate (clobbered, VOIDmode);
6115 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6116 || GET_CODE (clobbered) == ZERO_EXTRACT)
6117 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6118 }
6119 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6120 invalidate (SET_DEST (y), VOIDmode);
6121 }
6122 }
6123 }
6124 \f
6125 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6126 and replace any registers in them with either an equivalent constant
6127 or the canonical form of the register. If we are inside an address,
6128 only do this if the address remains valid.
6129
6130 OBJECT is 0 except when within a MEM in which case it is the MEM.
6131
6132 Return the replacement for X. */
6133
6134 static rtx
6135 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6136 {
6137 enum rtx_code code = GET_CODE (x);
6138 const char *fmt = GET_RTX_FORMAT (code);
6139 int i;
6140
6141 switch (code)
6142 {
6143 case CONST:
6144 case SYMBOL_REF:
6145 case LABEL_REF:
6146 CASE_CONST_ANY:
6147 case PC:
6148 case CC0:
6149 case LO_SUM:
6150 return x;
6151
6152 case MEM:
6153 validate_change (x, &XEXP (x, 0),
6154 cse_process_notes (XEXP (x, 0), x, changed), 0);
6155 return x;
6156
6157 case EXPR_LIST:
6158 if (REG_NOTE_KIND (x) == REG_EQUAL)
6159 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6160 /* Fall through. */
6161
6162 case INSN_LIST:
6163 case INT_LIST:
6164 if (XEXP (x, 1))
6165 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6166 return x;
6167
6168 case SIGN_EXTEND:
6169 case ZERO_EXTEND:
6170 case SUBREG:
6171 {
6172 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6173 /* We don't substitute VOIDmode constants into these rtx,
6174 since they would impede folding. */
6175 if (GET_MODE (new_rtx) != VOIDmode)
6176 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6177 return x;
6178 }
6179
6180 case UNSIGNED_FLOAT:
6181 {
6182 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6183 /* We don't substitute negative VOIDmode constants into these rtx,
6184 since they would impede folding. */
6185 if (GET_MODE (new_rtx) != VOIDmode
6186 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6187 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6188 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6189 return x;
6190 }
6191
6192 case REG:
6193 i = REG_QTY (REGNO (x));
6194
6195 /* Return a constant or a constant register. */
6196 if (REGNO_QTY_VALID_P (REGNO (x)))
6197 {
6198 struct qty_table_elem *ent = &qty_table[i];
6199
6200 if (ent->const_rtx != NULL_RTX
6201 && (CONSTANT_P (ent->const_rtx)
6202 || REG_P (ent->const_rtx)))
6203 {
6204 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6205 if (new_rtx)
6206 return copy_rtx (new_rtx);
6207 }
6208 }
6209
6210 /* Otherwise, canonicalize this register. */
6211 return canon_reg (x, NULL);
6212
6213 default:
6214 break;
6215 }
6216
6217 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6218 if (fmt[i] == 'e')
6219 validate_change (object, &XEXP (x, i),
6220 cse_process_notes (XEXP (x, i), object, changed), 0);
6221
6222 return x;
6223 }
6224
6225 static rtx
6226 cse_process_notes (rtx x, rtx object, bool *changed)
6227 {
6228 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6229 if (new_rtx != x)
6230 *changed = true;
6231 return new_rtx;
6232 }
6233
6234 \f
6235 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6236
6237 DATA is a pointer to a struct cse_basic_block_data, that is used to
6238 describe the path.
6239 It is filled with a queue of basic blocks, starting with FIRST_BB
6240 and following a trace through the CFG.
6241
6242 If all paths starting at FIRST_BB have been followed, or no new path
6243 starting at FIRST_BB can be constructed, this function returns FALSE.
6244 Otherwise, DATA->path is filled and the function returns TRUE indicating
6245 that a path to follow was found.
6246
6247 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6248 block in the path will be FIRST_BB. */
6249
6250 static bool
6251 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6252 int follow_jumps)
6253 {
6254 basic_block bb;
6255 edge e;
6256 int path_size;
6257
6258 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6259
6260 /* See if there is a previous path. */
6261 path_size = data->path_size;
6262
6263 /* There is a previous path. Make sure it started with FIRST_BB. */
6264 if (path_size)
6265 gcc_assert (data->path[0].bb == first_bb);
6266
6267 /* There was only one basic block in the last path. Clear the path and
6268 return, so that paths starting at another basic block can be tried. */
6269 if (path_size == 1)
6270 {
6271 path_size = 0;
6272 goto done;
6273 }
6274
6275 /* If the path was empty from the beginning, construct a new path. */
6276 if (path_size == 0)
6277 data->path[path_size++].bb = first_bb;
6278 else
6279 {
6280 /* Otherwise, path_size must be equal to or greater than 2, because
6281 a previous path exists that is at least two basic blocks long.
6282
6283 Update the previous branch path, if any. If the last branch was
6284 previously along the branch edge, take the fallthrough edge now. */
6285 while (path_size >= 2)
6286 {
6287 basic_block last_bb_in_path, previous_bb_in_path;
6288 edge e;
6289
6290 --path_size;
6291 last_bb_in_path = data->path[path_size].bb;
6292 previous_bb_in_path = data->path[path_size - 1].bb;
6293
6294 /* If we previously followed a path along the branch edge, try
6295 the fallthru edge now. */
6296 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6297 && any_condjump_p (BB_END (previous_bb_in_path))
6298 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6299 && e == BRANCH_EDGE (previous_bb_in_path))
6300 {
6301 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6302 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6303 && single_pred_p (bb)
6304 /* We used to assert here that we would only see blocks
6305 that we have not visited yet. But we may end up
6306 visiting basic blocks twice if the CFG has changed
6307 in this run of cse_main, because when the CFG changes
6308 the topological sort of the CFG also changes. A basic
6309 blocks that previously had more than two predecessors
6310 may now have a single predecessor, and become part of
6311 a path that starts at another basic block.
6312
6313 We still want to visit each basic block only once, so
6314 halt the path here if we have already visited BB. */
6315 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6316 {
6317 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6318 data->path[path_size++].bb = bb;
6319 break;
6320 }
6321 }
6322
6323 data->path[path_size].bb = NULL;
6324 }
6325
6326 /* If only one block remains in the path, bail. */
6327 if (path_size == 1)
6328 {
6329 path_size = 0;
6330 goto done;
6331 }
6332 }
6333
6334 /* Extend the path if possible. */
6335 if (follow_jumps)
6336 {
6337 bb = data->path[path_size - 1].bb;
6338 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6339 {
6340 if (single_succ_p (bb))
6341 e = single_succ_edge (bb);
6342 else if (EDGE_COUNT (bb->succs) == 2
6343 && any_condjump_p (BB_END (bb)))
6344 {
6345 /* First try to follow the branch. If that doesn't lead
6346 to a useful path, follow the fallthru edge. */
6347 e = BRANCH_EDGE (bb);
6348 if (!single_pred_p (e->dest))
6349 e = FALLTHRU_EDGE (bb);
6350 }
6351 else
6352 e = NULL;
6353
6354 if (e
6355 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6356 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6357 && single_pred_p (e->dest)
6358 /* Avoid visiting basic blocks twice. The large comment
6359 above explains why this can happen. */
6360 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6361 {
6362 basic_block bb2 = e->dest;
6363 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6364 data->path[path_size++].bb = bb2;
6365 bb = bb2;
6366 }
6367 else
6368 bb = NULL;
6369 }
6370 }
6371
6372 done:
6373 data->path_size = path_size;
6374 return path_size != 0;
6375 }
6376 \f
6377 /* Dump the path in DATA to file F. NSETS is the number of sets
6378 in the path. */
6379
6380 static void
6381 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6382 {
6383 int path_entry;
6384
6385 fprintf (f, ";; Following path with %d sets: ", nsets);
6386 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6387 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6388 fputc ('\n', dump_file);
6389 fflush (f);
6390 }
6391
6392 \f
6393 /* Return true if BB has exception handling successor edges. */
6394
6395 static bool
6396 have_eh_succ_edges (basic_block bb)
6397 {
6398 edge e;
6399 edge_iterator ei;
6400
6401 FOR_EACH_EDGE (e, ei, bb->succs)
6402 if (e->flags & EDGE_EH)
6403 return true;
6404
6405 return false;
6406 }
6407
6408 \f
6409 /* Scan to the end of the path described by DATA. Return an estimate of
6410 the total number of SETs of all insns in the path. */
6411
6412 static void
6413 cse_prescan_path (struct cse_basic_block_data *data)
6414 {
6415 int nsets = 0;
6416 int path_size = data->path_size;
6417 int path_entry;
6418
6419 /* Scan to end of each basic block in the path. */
6420 for (path_entry = 0; path_entry < path_size; path_entry++)
6421 {
6422 basic_block bb;
6423 rtx_insn *insn;
6424
6425 bb = data->path[path_entry].bb;
6426
6427 FOR_BB_INSNS (bb, insn)
6428 {
6429 if (!INSN_P (insn))
6430 continue;
6431
6432 /* A PARALLEL can have lots of SETs in it,
6433 especially if it is really an ASM_OPERANDS. */
6434 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6435 nsets += XVECLEN (PATTERN (insn), 0);
6436 else
6437 nsets += 1;
6438 }
6439 }
6440
6441 data->nsets = nsets;
6442 }
6443 \f
6444 /* Return true if the pattern of INSN uses a LABEL_REF for which
6445 there isn't a REG_LABEL_OPERAND note. */
6446
6447 static bool
6448 check_for_label_ref (rtx_insn *insn)
6449 {
6450 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6451 note for it, we must rerun jump since it needs to place the note. If
6452 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6453 don't do this since no REG_LABEL_OPERAND will be added. */
6454 subrtx_iterator::array_type array;
6455 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6456 {
6457 const_rtx x = *iter;
6458 if (GET_CODE (x) == LABEL_REF
6459 && !LABEL_REF_NONLOCAL_P (x)
6460 && (!JUMP_P (insn)
6461 || !label_is_jump_target_p (LABEL_REF_LABEL (x), insn))
6462 && LABEL_P (LABEL_REF_LABEL (x))
6463 && INSN_UID (LABEL_REF_LABEL (x)) != 0
6464 && !find_reg_note (insn, REG_LABEL_OPERAND, LABEL_REF_LABEL (x)))
6465 return true;
6466 }
6467 return false;
6468 }
6469
6470 /* Process a single extended basic block described by EBB_DATA. */
6471
6472 static void
6473 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6474 {
6475 int path_size = ebb_data->path_size;
6476 int path_entry;
6477 int num_insns = 0;
6478
6479 /* Allocate the space needed by qty_table. */
6480 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6481
6482 new_basic_block ();
6483 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6484 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6485 for (path_entry = 0; path_entry < path_size; path_entry++)
6486 {
6487 basic_block bb;
6488 rtx_insn *insn;
6489
6490 bb = ebb_data->path[path_entry].bb;
6491
6492 /* Invalidate recorded information for eh regs if there is an EH
6493 edge pointing to that bb. */
6494 if (bb_has_eh_pred (bb))
6495 {
6496 df_ref def;
6497
6498 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6499 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6500 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6501 }
6502
6503 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6504 FOR_BB_INSNS (bb, insn)
6505 {
6506 /* If we have processed 1,000 insns, flush the hash table to
6507 avoid extreme quadratic behavior. We must not include NOTEs
6508 in the count since there may be more of them when generating
6509 debugging information. If we clear the table at different
6510 times, code generated with -g -O might be different than code
6511 generated with -O but not -g.
6512
6513 FIXME: This is a real kludge and needs to be done some other
6514 way. */
6515 if (NONDEBUG_INSN_P (insn)
6516 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6517 {
6518 flush_hash_table ();
6519 num_insns = 0;
6520 }
6521
6522 if (INSN_P (insn))
6523 {
6524 /* Process notes first so we have all notes in canonical forms
6525 when looking for duplicate operations. */
6526 if (REG_NOTES (insn))
6527 {
6528 bool changed = false;
6529 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6530 NULL_RTX, &changed);
6531 if (changed)
6532 df_notes_rescan (insn);
6533 }
6534
6535 cse_insn (insn);
6536
6537 /* If we haven't already found an insn where we added a LABEL_REF,
6538 check this one. */
6539 if (INSN_P (insn) && !recorded_label_ref
6540 && check_for_label_ref (insn))
6541 recorded_label_ref = true;
6542
6543 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6544 {
6545 /* If the previous insn sets CC0 and this insn no
6546 longer references CC0, delete the previous insn.
6547 Here we use fact that nothing expects CC0 to be
6548 valid over an insn, which is true until the final
6549 pass. */
6550 rtx_insn *prev_insn;
6551 rtx tem;
6552
6553 prev_insn = prev_nonnote_nondebug_insn (insn);
6554 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6555 && (tem = single_set (prev_insn)) != NULL_RTX
6556 && SET_DEST (tem) == cc0_rtx
6557 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6558 delete_insn (prev_insn);
6559
6560 /* If this insn is not the last insn in the basic
6561 block, it will be PREV_INSN(insn) in the next
6562 iteration. If we recorded any CC0-related
6563 information for this insn, remember it. */
6564 if (insn != BB_END (bb))
6565 {
6566 prev_insn_cc0 = this_insn_cc0;
6567 prev_insn_cc0_mode = this_insn_cc0_mode;
6568 }
6569 }
6570 }
6571 }
6572
6573 /* With non-call exceptions, we are not always able to update
6574 the CFG properly inside cse_insn. So clean up possibly
6575 redundant EH edges here. */
6576 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6577 cse_cfg_altered |= purge_dead_edges (bb);
6578
6579 /* If we changed a conditional jump, we may have terminated
6580 the path we are following. Check that by verifying that
6581 the edge we would take still exists. If the edge does
6582 not exist anymore, purge the remainder of the path.
6583 Note that this will cause us to return to the caller. */
6584 if (path_entry < path_size - 1)
6585 {
6586 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6587 if (!find_edge (bb, next_bb))
6588 {
6589 do
6590 {
6591 path_size--;
6592
6593 /* If we truncate the path, we must also reset the
6594 visited bit on the remaining blocks in the path,
6595 or we will never visit them at all. */
6596 bitmap_clear_bit (cse_visited_basic_blocks,
6597 ebb_data->path[path_size].bb->index);
6598 ebb_data->path[path_size].bb = NULL;
6599 }
6600 while (path_size - 1 != path_entry);
6601 ebb_data->path_size = path_size;
6602 }
6603 }
6604
6605 /* If this is a conditional jump insn, record any known
6606 equivalences due to the condition being tested. */
6607 insn = BB_END (bb);
6608 if (path_entry < path_size - 1
6609 && JUMP_P (insn)
6610 && single_set (insn)
6611 && any_condjump_p (insn))
6612 {
6613 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6614 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6615 record_jump_equiv (insn, taken);
6616 }
6617
6618 /* Clear the CC0-tracking related insns, they can't provide
6619 useful information across basic block boundaries. */
6620 prev_insn_cc0 = 0;
6621 }
6622
6623 gcc_assert (next_qty <= max_qty);
6624
6625 free (qty_table);
6626 }
6627
6628 \f
6629 /* Perform cse on the instructions of a function.
6630 F is the first instruction.
6631 NREGS is one plus the highest pseudo-reg number used in the instruction.
6632
6633 Return 2 if jump optimizations should be redone due to simplifications
6634 in conditional jump instructions.
6635 Return 1 if the CFG should be cleaned up because it has been modified.
6636 Return 0 otherwise. */
6637
6638 static int
6639 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6640 {
6641 struct cse_basic_block_data ebb_data;
6642 basic_block bb;
6643 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6644 int i, n_blocks;
6645
6646 df_set_flags (DF_LR_RUN_DCE);
6647 df_note_add_problem ();
6648 df_analyze ();
6649 df_set_flags (DF_DEFER_INSN_RESCAN);
6650
6651 reg_scan (get_insns (), max_reg_num ());
6652 init_cse_reg_info (nregs);
6653
6654 ebb_data.path = XNEWVEC (struct branch_path,
6655 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6656
6657 cse_cfg_altered = false;
6658 cse_jumps_altered = false;
6659 recorded_label_ref = false;
6660 constant_pool_entries_cost = 0;
6661 constant_pool_entries_regcost = 0;
6662 ebb_data.path_size = 0;
6663 ebb_data.nsets = 0;
6664 rtl_hooks = cse_rtl_hooks;
6665
6666 init_recog ();
6667 init_alias_analysis ();
6668
6669 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6670
6671 /* Set up the table of already visited basic blocks. */
6672 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6673 bitmap_clear (cse_visited_basic_blocks);
6674
6675 /* Loop over basic blocks in reverse completion order (RPO),
6676 excluding the ENTRY and EXIT blocks. */
6677 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6678 i = 0;
6679 while (i < n_blocks)
6680 {
6681 /* Find the first block in the RPO queue that we have not yet
6682 processed before. */
6683 do
6684 {
6685 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6686 }
6687 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6688 && i < n_blocks);
6689
6690 /* Find all paths starting with BB, and process them. */
6691 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6692 {
6693 /* Pre-scan the path. */
6694 cse_prescan_path (&ebb_data);
6695
6696 /* If this basic block has no sets, skip it. */
6697 if (ebb_data.nsets == 0)
6698 continue;
6699
6700 /* Get a reasonable estimate for the maximum number of qty's
6701 needed for this path. For this, we take the number of sets
6702 and multiply that by MAX_RECOG_OPERANDS. */
6703 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6704
6705 /* Dump the path we're about to process. */
6706 if (dump_file)
6707 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6708
6709 cse_extended_basic_block (&ebb_data);
6710 }
6711 }
6712
6713 /* Clean up. */
6714 end_alias_analysis ();
6715 free (reg_eqv_table);
6716 free (ebb_data.path);
6717 sbitmap_free (cse_visited_basic_blocks);
6718 free (rc_order);
6719 rtl_hooks = general_rtl_hooks;
6720
6721 if (cse_jumps_altered || recorded_label_ref)
6722 return 2;
6723 else if (cse_cfg_altered)
6724 return 1;
6725 else
6726 return 0;
6727 }
6728 \f
6729 /* Count the number of times registers are used (not set) in X.
6730 COUNTS is an array in which we accumulate the count, INCR is how much
6731 we count each register usage.
6732
6733 Don't count a usage of DEST, which is the SET_DEST of a SET which
6734 contains X in its SET_SRC. This is because such a SET does not
6735 modify the liveness of DEST.
6736 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6737 We must then count uses of a SET_DEST regardless, because the insn can't be
6738 deleted here. */
6739
6740 static void
6741 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6742 {
6743 enum rtx_code code;
6744 rtx note;
6745 const char *fmt;
6746 int i, j;
6747
6748 if (x == 0)
6749 return;
6750
6751 switch (code = GET_CODE (x))
6752 {
6753 case REG:
6754 if (x != dest)
6755 counts[REGNO (x)] += incr;
6756 return;
6757
6758 case PC:
6759 case CC0:
6760 case CONST:
6761 CASE_CONST_ANY:
6762 case SYMBOL_REF:
6763 case LABEL_REF:
6764 return;
6765
6766 case CLOBBER:
6767 /* If we are clobbering a MEM, mark any registers inside the address
6768 as being used. */
6769 if (MEM_P (XEXP (x, 0)))
6770 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6771 return;
6772
6773 case SET:
6774 /* Unless we are setting a REG, count everything in SET_DEST. */
6775 if (!REG_P (SET_DEST (x)))
6776 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6777 count_reg_usage (SET_SRC (x), counts,
6778 dest ? dest : SET_DEST (x),
6779 incr);
6780 return;
6781
6782 case DEBUG_INSN:
6783 return;
6784
6785 case CALL_INSN:
6786 case INSN:
6787 case JUMP_INSN:
6788 /* We expect dest to be NULL_RTX here. If the insn may throw,
6789 or if it cannot be deleted due to side-effects, mark this fact
6790 by setting DEST to pc_rtx. */
6791 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6792 || side_effects_p (PATTERN (x)))
6793 dest = pc_rtx;
6794 if (code == CALL_INSN)
6795 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6796 count_reg_usage (PATTERN (x), counts, dest, incr);
6797
6798 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6799 use them. */
6800
6801 note = find_reg_equal_equiv_note (x);
6802 if (note)
6803 {
6804 rtx eqv = XEXP (note, 0);
6805
6806 if (GET_CODE (eqv) == EXPR_LIST)
6807 /* This REG_EQUAL note describes the result of a function call.
6808 Process all the arguments. */
6809 do
6810 {
6811 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6812 eqv = XEXP (eqv, 1);
6813 }
6814 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6815 else
6816 count_reg_usage (eqv, counts, dest, incr);
6817 }
6818 return;
6819
6820 case EXPR_LIST:
6821 if (REG_NOTE_KIND (x) == REG_EQUAL
6822 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6823 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6824 involving registers in the address. */
6825 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6826 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6827
6828 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6829 return;
6830
6831 case ASM_OPERANDS:
6832 /* Iterate over just the inputs, not the constraints as well. */
6833 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6834 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6835 return;
6836
6837 case INSN_LIST:
6838 case INT_LIST:
6839 gcc_unreachable ();
6840
6841 default:
6842 break;
6843 }
6844
6845 fmt = GET_RTX_FORMAT (code);
6846 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6847 {
6848 if (fmt[i] == 'e')
6849 count_reg_usage (XEXP (x, i), counts, dest, incr);
6850 else if (fmt[i] == 'E')
6851 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6852 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6853 }
6854 }
6855 \f
6856 /* Return true if X is a dead register. */
6857
6858 static inline int
6859 is_dead_reg (const_rtx x, int *counts)
6860 {
6861 return (REG_P (x)
6862 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6863 && counts[REGNO (x)] == 0);
6864 }
6865
6866 /* Return true if set is live. */
6867 static bool
6868 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6869 int *counts)
6870 {
6871 rtx_insn *tem;
6872
6873 if (set_noop_p (set))
6874 ;
6875
6876 else if (GET_CODE (SET_DEST (set)) == CC0
6877 && !side_effects_p (SET_SRC (set))
6878 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6879 || !INSN_P (tem)
6880 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6881 return false;
6882 else if (!is_dead_reg (SET_DEST (set), counts)
6883 || side_effects_p (SET_SRC (set)))
6884 return true;
6885 return false;
6886 }
6887
6888 /* Return true if insn is live. */
6889
6890 static bool
6891 insn_live_p (rtx_insn *insn, int *counts)
6892 {
6893 int i;
6894 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6895 return true;
6896 else if (GET_CODE (PATTERN (insn)) == SET)
6897 return set_live_p (PATTERN (insn), insn, counts);
6898 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6899 {
6900 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6901 {
6902 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6903
6904 if (GET_CODE (elt) == SET)
6905 {
6906 if (set_live_p (elt, insn, counts))
6907 return true;
6908 }
6909 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6910 return true;
6911 }
6912 return false;
6913 }
6914 else if (DEBUG_INSN_P (insn))
6915 {
6916 rtx_insn *next;
6917
6918 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6919 if (NOTE_P (next))
6920 continue;
6921 else if (!DEBUG_INSN_P (next))
6922 return true;
6923 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6924 return false;
6925
6926 return true;
6927 }
6928 else
6929 return true;
6930 }
6931
6932 /* Count the number of stores into pseudo. Callback for note_stores. */
6933
6934 static void
6935 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6936 {
6937 int *counts = (int *) data;
6938 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6939 counts[REGNO (x)]++;
6940 }
6941
6942 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6943 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6944 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6945 Set *SEEN_REPL to true if we see a dead register that does have
6946 a replacement. */
6947
6948 static bool
6949 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6950 bool *seen_repl)
6951 {
6952 subrtx_iterator::array_type array;
6953 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6954 {
6955 const_rtx x = *iter;
6956 if (is_dead_reg (x, counts))
6957 {
6958 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6959 *seen_repl = true;
6960 else
6961 return true;
6962 }
6963 }
6964 return false;
6965 }
6966
6967 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6968 Callback for simplify_replace_fn_rtx. */
6969
6970 static rtx
6971 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6972 {
6973 rtx *replacements = (rtx *) data;
6974
6975 if (REG_P (x)
6976 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6977 && replacements[REGNO (x)] != NULL_RTX)
6978 {
6979 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6980 return replacements[REGNO (x)];
6981 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6982 GET_MODE (replacements[REGNO (x)]));
6983 }
6984 return NULL_RTX;
6985 }
6986
6987 /* Scan all the insns and delete any that are dead; i.e., they store a register
6988 that is never used or they copy a register to itself.
6989
6990 This is used to remove insns made obviously dead by cse, loop or other
6991 optimizations. It improves the heuristics in loop since it won't try to
6992 move dead invariants out of loops or make givs for dead quantities. The
6993 remaining passes of the compilation are also sped up. */
6994
6995 int
6996 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
6997 {
6998 int *counts;
6999 rtx_insn *insn, *prev;
7000 rtx *replacements = NULL;
7001 int ndead = 0;
7002
7003 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7004 /* First count the number of times each register is used. */
7005 if (MAY_HAVE_DEBUG_INSNS)
7006 {
7007 counts = XCNEWVEC (int, nreg * 3);
7008 for (insn = insns; insn; insn = NEXT_INSN (insn))
7009 if (DEBUG_INSN_P (insn))
7010 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7011 NULL_RTX, 1);
7012 else if (INSN_P (insn))
7013 {
7014 count_reg_usage (insn, counts, NULL_RTX, 1);
7015 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7016 }
7017 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7018 First one counts how many times each pseudo is used outside
7019 of debug insns, second counts how many times each pseudo is
7020 used in debug insns and third counts how many times a pseudo
7021 is stored. */
7022 }
7023 else
7024 {
7025 counts = XCNEWVEC (int, nreg);
7026 for (insn = insns; insn; insn = NEXT_INSN (insn))
7027 if (INSN_P (insn))
7028 count_reg_usage (insn, counts, NULL_RTX, 1);
7029 /* If no debug insns can be present, COUNTS is just an array
7030 which counts how many times each pseudo is used. */
7031 }
7032 /* Pseudo PIC register should be considered as used due to possible
7033 new usages generated. */
7034 if (!reload_completed
7035 && pic_offset_table_rtx
7036 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7037 counts[REGNO (pic_offset_table_rtx)]++;
7038 /* Go from the last insn to the first and delete insns that only set unused
7039 registers or copy a register to itself. As we delete an insn, remove
7040 usage counts for registers it uses.
7041
7042 The first jump optimization pass may leave a real insn as the last
7043 insn in the function. We must not skip that insn or we may end
7044 up deleting code that is not really dead.
7045
7046 If some otherwise unused register is only used in DEBUG_INSNs,
7047 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7048 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7049 has been created for the unused register, replace it with
7050 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7051 for (insn = get_last_insn (); insn; insn = prev)
7052 {
7053 int live_insn = 0;
7054
7055 prev = PREV_INSN (insn);
7056 if (!INSN_P (insn))
7057 continue;
7058
7059 live_insn = insn_live_p (insn, counts);
7060
7061 /* If this is a dead insn, delete it and show registers in it aren't
7062 being used. */
7063
7064 if (! live_insn && dbg_cnt (delete_trivial_dead))
7065 {
7066 if (DEBUG_INSN_P (insn))
7067 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7068 NULL_RTX, -1);
7069 else
7070 {
7071 rtx set;
7072 if (MAY_HAVE_DEBUG_INSNS
7073 && (set = single_set (insn)) != NULL_RTX
7074 && is_dead_reg (SET_DEST (set), counts)
7075 /* Used at least once in some DEBUG_INSN. */
7076 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7077 /* And set exactly once. */
7078 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7079 && !side_effects_p (SET_SRC (set))
7080 && asm_noperands (PATTERN (insn)) < 0)
7081 {
7082 rtx dval, bind_var_loc;
7083 rtx_insn *bind;
7084
7085 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7086 dval = make_debug_expr_from_rtl (SET_DEST (set));
7087
7088 /* Emit a debug bind insn before the insn in which
7089 reg dies. */
7090 bind_var_loc =
7091 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7092 DEBUG_EXPR_TREE_DECL (dval),
7093 SET_SRC (set),
7094 VAR_INIT_STATUS_INITIALIZED);
7095 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7096
7097 bind = emit_debug_insn_before (bind_var_loc, insn);
7098 df_insn_rescan (bind);
7099
7100 if (replacements == NULL)
7101 replacements = XCNEWVEC (rtx, nreg);
7102 replacements[REGNO (SET_DEST (set))] = dval;
7103 }
7104
7105 count_reg_usage (insn, counts, NULL_RTX, -1);
7106 ndead++;
7107 }
7108 delete_insn_and_edges (insn);
7109 }
7110 }
7111
7112 if (MAY_HAVE_DEBUG_INSNS)
7113 {
7114 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7115 if (DEBUG_INSN_P (insn))
7116 {
7117 /* If this debug insn references a dead register that wasn't replaced
7118 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7119 bool seen_repl = false;
7120 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7121 counts, replacements, &seen_repl))
7122 {
7123 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7124 df_insn_rescan (insn);
7125 }
7126 else if (seen_repl)
7127 {
7128 INSN_VAR_LOCATION_LOC (insn)
7129 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7130 NULL_RTX, replace_dead_reg,
7131 replacements);
7132 df_insn_rescan (insn);
7133 }
7134 }
7135 free (replacements);
7136 }
7137
7138 if (dump_file && ndead)
7139 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7140 ndead);
7141 /* Clean up. */
7142 free (counts);
7143 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7144 return ndead;
7145 }
7146
7147 /* If LOC contains references to NEWREG in a different mode, change them
7148 to use NEWREG instead. */
7149
7150 static void
7151 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7152 rtx *loc, rtx_insn *insn, rtx newreg)
7153 {
7154 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7155 {
7156 rtx *loc = *iter;
7157 rtx x = *loc;
7158 if (x
7159 && REG_P (x)
7160 && REGNO (x) == REGNO (newreg)
7161 && GET_MODE (x) != GET_MODE (newreg))
7162 {
7163 validate_change (insn, loc, newreg, 1);
7164 iter.skip_subrtxes ();
7165 }
7166 }
7167 }
7168
7169 /* Change the mode of any reference to the register REGNO (NEWREG) to
7170 GET_MODE (NEWREG) in INSN. */
7171
7172 static void
7173 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7174 {
7175 int success;
7176
7177 if (!INSN_P (insn))
7178 return;
7179
7180 subrtx_ptr_iterator::array_type array;
7181 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7182 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7183
7184 /* If the following assertion was triggered, there is most probably
7185 something wrong with the cc_modes_compatible back end function.
7186 CC modes only can be considered compatible if the insn - with the mode
7187 replaced by any of the compatible modes - can still be recognized. */
7188 success = apply_change_group ();
7189 gcc_assert (success);
7190 }
7191
7192 /* Change the mode of any reference to the register REGNO (NEWREG) to
7193 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7194 any instruction which modifies NEWREG. */
7195
7196 static void
7197 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7198 {
7199 rtx_insn *insn;
7200
7201 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7202 {
7203 if (! INSN_P (insn))
7204 continue;
7205
7206 if (reg_set_p (newreg, insn))
7207 return;
7208
7209 cse_change_cc_mode_insn (insn, newreg);
7210 }
7211 }
7212
7213 /* BB is a basic block which finishes with CC_REG as a condition code
7214 register which is set to CC_SRC. Look through the successors of BB
7215 to find blocks which have a single predecessor (i.e., this one),
7216 and look through those blocks for an assignment to CC_REG which is
7217 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7218 permitted to change the mode of CC_SRC to a compatible mode. This
7219 returns VOIDmode if no equivalent assignments were found.
7220 Otherwise it returns the mode which CC_SRC should wind up with.
7221 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7222 but is passed unmodified down to recursive calls in order to prevent
7223 endless recursion.
7224
7225 The main complexity in this function is handling the mode issues.
7226 We may have more than one duplicate which we can eliminate, and we
7227 try to find a mode which will work for multiple duplicates. */
7228
7229 static machine_mode
7230 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7231 bool can_change_mode)
7232 {
7233 bool found_equiv;
7234 machine_mode mode;
7235 unsigned int insn_count;
7236 edge e;
7237 rtx_insn *insns[2];
7238 machine_mode modes[2];
7239 rtx_insn *last_insns[2];
7240 unsigned int i;
7241 rtx newreg;
7242 edge_iterator ei;
7243
7244 /* We expect to have two successors. Look at both before picking
7245 the final mode for the comparison. If we have more successors
7246 (i.e., some sort of table jump, although that seems unlikely),
7247 then we require all beyond the first two to use the same
7248 mode. */
7249
7250 found_equiv = false;
7251 mode = GET_MODE (cc_src);
7252 insn_count = 0;
7253 FOR_EACH_EDGE (e, ei, bb->succs)
7254 {
7255 rtx_insn *insn;
7256 rtx_insn *end;
7257
7258 if (e->flags & EDGE_COMPLEX)
7259 continue;
7260
7261 if (EDGE_COUNT (e->dest->preds) != 1
7262 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7263 /* Avoid endless recursion on unreachable blocks. */
7264 || e->dest == orig_bb)
7265 continue;
7266
7267 end = NEXT_INSN (BB_END (e->dest));
7268 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7269 {
7270 rtx set;
7271
7272 if (! INSN_P (insn))
7273 continue;
7274
7275 /* If CC_SRC is modified, we have to stop looking for
7276 something which uses it. */
7277 if (modified_in_p (cc_src, insn))
7278 break;
7279
7280 /* Check whether INSN sets CC_REG to CC_SRC. */
7281 set = single_set (insn);
7282 if (set
7283 && REG_P (SET_DEST (set))
7284 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7285 {
7286 bool found;
7287 machine_mode set_mode;
7288 machine_mode comp_mode;
7289
7290 found = false;
7291 set_mode = GET_MODE (SET_SRC (set));
7292 comp_mode = set_mode;
7293 if (rtx_equal_p (cc_src, SET_SRC (set)))
7294 found = true;
7295 else if (GET_CODE (cc_src) == COMPARE
7296 && GET_CODE (SET_SRC (set)) == COMPARE
7297 && mode != set_mode
7298 && rtx_equal_p (XEXP (cc_src, 0),
7299 XEXP (SET_SRC (set), 0))
7300 && rtx_equal_p (XEXP (cc_src, 1),
7301 XEXP (SET_SRC (set), 1)))
7302
7303 {
7304 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7305 if (comp_mode != VOIDmode
7306 && (can_change_mode || comp_mode == mode))
7307 found = true;
7308 }
7309
7310 if (found)
7311 {
7312 found_equiv = true;
7313 if (insn_count < ARRAY_SIZE (insns))
7314 {
7315 insns[insn_count] = insn;
7316 modes[insn_count] = set_mode;
7317 last_insns[insn_count] = end;
7318 ++insn_count;
7319
7320 if (mode != comp_mode)
7321 {
7322 gcc_assert (can_change_mode);
7323 mode = comp_mode;
7324
7325 /* The modified insn will be re-recognized later. */
7326 PUT_MODE (cc_src, mode);
7327 }
7328 }
7329 else
7330 {
7331 if (set_mode != mode)
7332 {
7333 /* We found a matching expression in the
7334 wrong mode, but we don't have room to
7335 store it in the array. Punt. This case
7336 should be rare. */
7337 break;
7338 }
7339 /* INSN sets CC_REG to a value equal to CC_SRC
7340 with the right mode. We can simply delete
7341 it. */
7342 delete_insn (insn);
7343 }
7344
7345 /* We found an instruction to delete. Keep looking,
7346 in the hopes of finding a three-way jump. */
7347 continue;
7348 }
7349
7350 /* We found an instruction which sets the condition
7351 code, so don't look any farther. */
7352 break;
7353 }
7354
7355 /* If INSN sets CC_REG in some other way, don't look any
7356 farther. */
7357 if (reg_set_p (cc_reg, insn))
7358 break;
7359 }
7360
7361 /* If we fell off the bottom of the block, we can keep looking
7362 through successors. We pass CAN_CHANGE_MODE as false because
7363 we aren't prepared to handle compatibility between the
7364 further blocks and this block. */
7365 if (insn == end)
7366 {
7367 machine_mode submode;
7368
7369 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7370 if (submode != VOIDmode)
7371 {
7372 gcc_assert (submode == mode);
7373 found_equiv = true;
7374 can_change_mode = false;
7375 }
7376 }
7377 }
7378
7379 if (! found_equiv)
7380 return VOIDmode;
7381
7382 /* Now INSN_COUNT is the number of instructions we found which set
7383 CC_REG to a value equivalent to CC_SRC. The instructions are in
7384 INSNS. The modes used by those instructions are in MODES. */
7385
7386 newreg = NULL_RTX;
7387 for (i = 0; i < insn_count; ++i)
7388 {
7389 if (modes[i] != mode)
7390 {
7391 /* We need to change the mode of CC_REG in INSNS[i] and
7392 subsequent instructions. */
7393 if (! newreg)
7394 {
7395 if (GET_MODE (cc_reg) == mode)
7396 newreg = cc_reg;
7397 else
7398 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7399 }
7400 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7401 newreg);
7402 }
7403
7404 delete_insn_and_edges (insns[i]);
7405 }
7406
7407 return mode;
7408 }
7409
7410 /* If we have a fixed condition code register (or two), walk through
7411 the instructions and try to eliminate duplicate assignments. */
7412
7413 static void
7414 cse_condition_code_reg (void)
7415 {
7416 unsigned int cc_regno_1;
7417 unsigned int cc_regno_2;
7418 rtx cc_reg_1;
7419 rtx cc_reg_2;
7420 basic_block bb;
7421
7422 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7423 return;
7424
7425 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7426 if (cc_regno_2 != INVALID_REGNUM)
7427 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7428 else
7429 cc_reg_2 = NULL_RTX;
7430
7431 FOR_EACH_BB_FN (bb, cfun)
7432 {
7433 rtx_insn *last_insn;
7434 rtx cc_reg;
7435 rtx_insn *insn;
7436 rtx_insn *cc_src_insn;
7437 rtx cc_src;
7438 machine_mode mode;
7439 machine_mode orig_mode;
7440
7441 /* Look for blocks which end with a conditional jump based on a
7442 condition code register. Then look for the instruction which
7443 sets the condition code register. Then look through the
7444 successor blocks for instructions which set the condition
7445 code register to the same value. There are other possible
7446 uses of the condition code register, but these are by far the
7447 most common and the ones which we are most likely to be able
7448 to optimize. */
7449
7450 last_insn = BB_END (bb);
7451 if (!JUMP_P (last_insn))
7452 continue;
7453
7454 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7455 cc_reg = cc_reg_1;
7456 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7457 cc_reg = cc_reg_2;
7458 else
7459 continue;
7460
7461 cc_src_insn = NULL;
7462 cc_src = NULL_RTX;
7463 for (insn = PREV_INSN (last_insn);
7464 insn && insn != PREV_INSN (BB_HEAD (bb));
7465 insn = PREV_INSN (insn))
7466 {
7467 rtx set;
7468
7469 if (! INSN_P (insn))
7470 continue;
7471 set = single_set (insn);
7472 if (set
7473 && REG_P (SET_DEST (set))
7474 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7475 {
7476 cc_src_insn = insn;
7477 cc_src = SET_SRC (set);
7478 break;
7479 }
7480 else if (reg_set_p (cc_reg, insn))
7481 break;
7482 }
7483
7484 if (! cc_src_insn)
7485 continue;
7486
7487 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7488 continue;
7489
7490 /* Now CC_REG is a condition code register used for a
7491 conditional jump at the end of the block, and CC_SRC, in
7492 CC_SRC_INSN, is the value to which that condition code
7493 register is set, and CC_SRC is still meaningful at the end of
7494 the basic block. */
7495
7496 orig_mode = GET_MODE (cc_src);
7497 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7498 if (mode != VOIDmode)
7499 {
7500 gcc_assert (mode == GET_MODE (cc_src));
7501 if (mode != orig_mode)
7502 {
7503 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7504
7505 cse_change_cc_mode_insn (cc_src_insn, newreg);
7506
7507 /* Do the same in the following insns that use the
7508 current value of CC_REG within BB. */
7509 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7510 NEXT_INSN (last_insn),
7511 newreg);
7512 }
7513 }
7514 }
7515 }
7516 \f
7517
7518 /* Perform common subexpression elimination. Nonzero value from
7519 `cse_main' means that jumps were simplified and some code may now
7520 be unreachable, so do jump optimization again. */
7521 static unsigned int
7522 rest_of_handle_cse (void)
7523 {
7524 int tem;
7525
7526 if (dump_file)
7527 dump_flow_info (dump_file, dump_flags);
7528
7529 tem = cse_main (get_insns (), max_reg_num ());
7530
7531 /* If we are not running more CSE passes, then we are no longer
7532 expecting CSE to be run. But always rerun it in a cheap mode. */
7533 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7534
7535 if (tem == 2)
7536 {
7537 timevar_push (TV_JUMP);
7538 rebuild_jump_labels (get_insns ());
7539 cleanup_cfg (CLEANUP_CFG_CHANGED);
7540 timevar_pop (TV_JUMP);
7541 }
7542 else if (tem == 1 || optimize > 1)
7543 cleanup_cfg (0);
7544
7545 return 0;
7546 }
7547
7548 namespace {
7549
7550 const pass_data pass_data_cse =
7551 {
7552 RTL_PASS, /* type */
7553 "cse1", /* name */
7554 OPTGROUP_NONE, /* optinfo_flags */
7555 TV_CSE, /* tv_id */
7556 0, /* properties_required */
7557 0, /* properties_provided */
7558 0, /* properties_destroyed */
7559 0, /* todo_flags_start */
7560 TODO_df_finish, /* todo_flags_finish */
7561 };
7562
7563 class pass_cse : public rtl_opt_pass
7564 {
7565 public:
7566 pass_cse (gcc::context *ctxt)
7567 : rtl_opt_pass (pass_data_cse, ctxt)
7568 {}
7569
7570 /* opt_pass methods: */
7571 virtual bool gate (function *) { return optimize > 0; }
7572 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7573
7574 }; // class pass_cse
7575
7576 } // anon namespace
7577
7578 rtl_opt_pass *
7579 make_pass_cse (gcc::context *ctxt)
7580 {
7581 return new pass_cse (ctxt);
7582 }
7583
7584
7585 /* Run second CSE pass after loop optimizations. */
7586 static unsigned int
7587 rest_of_handle_cse2 (void)
7588 {
7589 int tem;
7590
7591 if (dump_file)
7592 dump_flow_info (dump_file, dump_flags);
7593
7594 tem = cse_main (get_insns (), max_reg_num ());
7595
7596 /* Run a pass to eliminate duplicated assignments to condition code
7597 registers. We have to run this after bypass_jumps, because it
7598 makes it harder for that pass to determine whether a jump can be
7599 bypassed safely. */
7600 cse_condition_code_reg ();
7601
7602 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7603
7604 if (tem == 2)
7605 {
7606 timevar_push (TV_JUMP);
7607 rebuild_jump_labels (get_insns ());
7608 cleanup_cfg (CLEANUP_CFG_CHANGED);
7609 timevar_pop (TV_JUMP);
7610 }
7611 else if (tem == 1)
7612 cleanup_cfg (0);
7613
7614 cse_not_expected = 1;
7615 return 0;
7616 }
7617
7618
7619 namespace {
7620
7621 const pass_data pass_data_cse2 =
7622 {
7623 RTL_PASS, /* type */
7624 "cse2", /* name */
7625 OPTGROUP_NONE, /* optinfo_flags */
7626 TV_CSE2, /* tv_id */
7627 0, /* properties_required */
7628 0, /* properties_provided */
7629 0, /* properties_destroyed */
7630 0, /* todo_flags_start */
7631 TODO_df_finish, /* todo_flags_finish */
7632 };
7633
7634 class pass_cse2 : public rtl_opt_pass
7635 {
7636 public:
7637 pass_cse2 (gcc::context *ctxt)
7638 : rtl_opt_pass (pass_data_cse2, ctxt)
7639 {}
7640
7641 /* opt_pass methods: */
7642 virtual bool gate (function *)
7643 {
7644 return optimize > 0 && flag_rerun_cse_after_loop;
7645 }
7646
7647 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7648
7649 }; // class pass_cse2
7650
7651 } // anon namespace
7652
7653 rtl_opt_pass *
7654 make_pass_cse2 (gcc::context *ctxt)
7655 {
7656 return new pass_cse2 (ctxt);
7657 }
7658
7659 /* Run second CSE pass after loop optimizations. */
7660 static unsigned int
7661 rest_of_handle_cse_after_global_opts (void)
7662 {
7663 int save_cfj;
7664 int tem;
7665
7666 /* We only want to do local CSE, so don't follow jumps. */
7667 save_cfj = flag_cse_follow_jumps;
7668 flag_cse_follow_jumps = 0;
7669
7670 rebuild_jump_labels (get_insns ());
7671 tem = cse_main (get_insns (), max_reg_num ());
7672 purge_all_dead_edges ();
7673 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7674
7675 cse_not_expected = !flag_rerun_cse_after_loop;
7676
7677 /* If cse altered any jumps, rerun jump opts to clean things up. */
7678 if (tem == 2)
7679 {
7680 timevar_push (TV_JUMP);
7681 rebuild_jump_labels (get_insns ());
7682 cleanup_cfg (CLEANUP_CFG_CHANGED);
7683 timevar_pop (TV_JUMP);
7684 }
7685 else if (tem == 1)
7686 cleanup_cfg (0);
7687
7688 flag_cse_follow_jumps = save_cfj;
7689 return 0;
7690 }
7691
7692 namespace {
7693
7694 const pass_data pass_data_cse_after_global_opts =
7695 {
7696 RTL_PASS, /* type */
7697 "cse_local", /* name */
7698 OPTGROUP_NONE, /* optinfo_flags */
7699 TV_CSE, /* tv_id */
7700 0, /* properties_required */
7701 0, /* properties_provided */
7702 0, /* properties_destroyed */
7703 0, /* todo_flags_start */
7704 TODO_df_finish, /* todo_flags_finish */
7705 };
7706
7707 class pass_cse_after_global_opts : public rtl_opt_pass
7708 {
7709 public:
7710 pass_cse_after_global_opts (gcc::context *ctxt)
7711 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7712 {}
7713
7714 /* opt_pass methods: */
7715 virtual bool gate (function *)
7716 {
7717 return optimize > 0 && flag_rerun_cse_after_global_opts;
7718 }
7719
7720 virtual unsigned int execute (function *)
7721 {
7722 return rest_of_handle_cse_after_global_opts ();
7723 }
7724
7725 }; // class pass_cse_after_global_opts
7726
7727 } // anon namespace
7728
7729 rtl_opt_pass *
7730 make_pass_cse_after_global_opts (gcc::context *ctxt)
7731 {
7732 return new pass_cse_after_global_opts (ctxt);
7733 }