[ARM/AArch64][testsuite] Add vqdmull_n tests.
[official-gcc.git] / gcc / cse.c
blob4732d81a72a60273102b88aebfeb3422d3211a55
1 /* Common subexpression elimination for GNU compiler.
2 Copyright (C) 1987-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "rtl.h"
25 #include "tm_p.h"
26 #include "hard-reg-set.h"
27 #include "regs.h"
28 #include "predict.h"
29 #include "vec.h"
30 #include "hashtab.h"
31 #include "hash-set.h"
32 #include "machmode.h"
33 #include "input.h"
34 #include "function.h"
35 #include "dominance.h"
36 #include "cfg.h"
37 #include "cfgrtl.h"
38 #include "cfganal.h"
39 #include "cfgcleanup.h"
40 #include "basic-block.h"
41 #include "flags.h"
42 #include "insn-config.h"
43 #include "recog.h"
44 #include "symtab.h"
45 #include "statistics.h"
46 #include "double-int.h"
47 #include "real.h"
48 #include "fixed-value.h"
49 #include "alias.h"
50 #include "wide-int.h"
51 #include "inchash.h"
52 #include "tree.h"
53 #include "expmed.h"
54 #include "dojump.h"
55 #include "explow.h"
56 #include "calls.h"
57 #include "emit-rtl.h"
58 #include "varasm.h"
59 #include "stmt.h"
60 #include "expr.h"
61 #include "diagnostic-core.h"
62 #include "toplev.h"
63 #include "ggc.h"
64 #include "except.h"
65 #include "target.h"
66 #include "params.h"
67 #include "rtlhooks-def.h"
68 #include "tree-pass.h"
69 #include "df.h"
70 #include "dbgcnt.h"
71 #include "rtl-iter.h"
73 /* The basic idea of common subexpression elimination is to go
74 through the code, keeping a record of expressions that would
75 have the same value at the current scan point, and replacing
76 expressions encountered with the cheapest equivalent expression.
78 It is too complicated to keep track of the different possibilities
79 when control paths merge in this code; so, at each label, we forget all
80 that is known and start fresh. This can be described as processing each
81 extended basic block separately. We have a separate pass to perform
82 global CSE.
84 Note CSE can turn a conditional or computed jump into a nop or
85 an unconditional jump. When this occurs we arrange to run the jump
86 optimizer after CSE to delete the unreachable code.
88 We use two data structures to record the equivalent expressions:
89 a hash table for most expressions, and a vector of "quantity
90 numbers" to record equivalent (pseudo) registers.
92 The use of the special data structure for registers is desirable
93 because it is faster. It is possible because registers references
94 contain a fairly small number, the register number, taken from
95 a contiguously allocated series, and two register references are
96 identical if they have the same number. General expressions
97 do not have any such thing, so the only way to retrieve the
98 information recorded on an expression other than a register
99 is to keep it in a hash table.
101 Registers and "quantity numbers":
103 At the start of each basic block, all of the (hardware and pseudo)
104 registers used in the function are given distinct quantity
105 numbers to indicate their contents. During scan, when the code
106 copies one register into another, we copy the quantity number.
107 When a register is loaded in any other way, we allocate a new
108 quantity number to describe the value generated by this operation.
109 `REG_QTY (N)' records what quantity register N is currently thought
110 of as containing.
112 All real quantity numbers are greater than or equal to zero.
113 If register N has not been assigned a quantity, `REG_QTY (N)' will
114 equal -N - 1, which is always negative.
116 Quantity numbers below zero do not exist and none of the `qty_table'
117 entries should be referenced with a negative index.
119 We also maintain a bidirectional chain of registers for each
120 quantity number. The `qty_table` members `first_reg' and `last_reg',
121 and `reg_eqv_table' members `next' and `prev' hold these chains.
123 The first register in a chain is the one whose lifespan is least local.
124 Among equals, it is the one that was seen first.
125 We replace any equivalent register with that one.
127 If two registers have the same quantity number, it must be true that
128 REG expressions with qty_table `mode' must be in the hash table for both
129 registers and must be in the same class.
131 The converse is not true. Since hard registers may be referenced in
132 any mode, two REG expressions might be equivalent in the hash table
133 but not have the same quantity number if the quantity number of one
134 of the registers is not the same mode as those expressions.
136 Constants and quantity numbers
138 When a quantity has a known constant value, that value is stored
139 in the appropriate qty_table `const_rtx'. This is in addition to
140 putting the constant in the hash table as is usual for non-regs.
142 Whether a reg or a constant is preferred is determined by the configuration
143 macro CONST_COSTS and will often depend on the constant value. In any
144 event, expressions containing constants can be simplified, by fold_rtx.
146 When a quantity has a known nearly constant value (such as an address
147 of a stack slot), that value is stored in the appropriate qty_table
148 `const_rtx'.
150 Integer constants don't have a machine mode. However, cse
151 determines the intended machine mode from the destination
152 of the instruction that moves the constant. The machine mode
153 is recorded in the hash table along with the actual RTL
154 constant expression so that different modes are kept separate.
156 Other expressions:
158 To record known equivalences among expressions in general
159 we use a hash table called `table'. It has a fixed number of buckets
160 that contain chains of `struct table_elt' elements for expressions.
161 These chains connect the elements whose expressions have the same
162 hash codes.
164 Other chains through the same elements connect the elements which
165 currently have equivalent values.
167 Register references in an expression are canonicalized before hashing
168 the expression. This is done using `reg_qty' and qty_table `first_reg'.
169 The hash code of a register reference is computed using the quantity
170 number, not the register number.
172 When the value of an expression changes, it is necessary to remove from the
173 hash table not just that expression but all expressions whose values
174 could be different as a result.
176 1. If the value changing is in memory, except in special cases
177 ANYTHING referring to memory could be changed. That is because
178 nobody knows where a pointer does not point.
179 The function `invalidate_memory' removes what is necessary.
181 The special cases are when the address is constant or is
182 a constant plus a fixed register such as the frame pointer
183 or a static chain pointer. When such addresses are stored in,
184 we can tell exactly which other such addresses must be invalidated
185 due to overlap. `invalidate' does this.
186 All expressions that refer to non-constant
187 memory addresses are also invalidated. `invalidate_memory' does this.
189 2. If the value changing is a register, all expressions
190 containing references to that register, and only those,
191 must be removed.
193 Because searching the entire hash table for expressions that contain
194 a register is very slow, we try to figure out when it isn't necessary.
195 Precisely, this is necessary only when expressions have been
196 entered in the hash table using this register, and then the value has
197 changed, and then another expression wants to be added to refer to
198 the register's new value. This sequence of circumstances is rare
199 within any one basic block.
201 `REG_TICK' and `REG_IN_TABLE', accessors for members of
202 cse_reg_info, are used to detect this case. REG_TICK (i) is
203 incremented whenever a value is stored in register i.
204 REG_IN_TABLE (i) holds -1 if no references to register i have been
205 entered in the table; otherwise, it contains the value REG_TICK (i)
206 had when the references were entered. If we want to enter a
207 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
208 remove old references. Until we want to enter a new entry, the
209 mere fact that the two vectors don't match makes the entries be
210 ignored if anyone tries to match them.
212 Registers themselves are entered in the hash table as well as in
213 the equivalent-register chains. However, `REG_TICK' and
214 `REG_IN_TABLE' do not apply to expressions which are simple
215 register references. These expressions are removed from the table
216 immediately when they become invalid, and this can be done even if
217 we do not immediately search for all the expressions that refer to
218 the register.
220 A CLOBBER rtx in an instruction invalidates its operand for further
221 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
222 invalidates everything that resides in memory.
224 Related expressions:
226 Constant expressions that differ only by an additive integer
227 are called related. When a constant expression is put in
228 the table, the related expression with no constant term
229 is also entered. These are made to point at each other
230 so that it is possible to find out if there exists any
231 register equivalent to an expression related to a given expression. */
233 /* Length of qty_table vector. We know in advance we will not need
234 a quantity number this big. */
236 static int max_qty;
238 /* Next quantity number to be allocated.
239 This is 1 + the largest number needed so far. */
241 static int next_qty;
243 /* Per-qty information tracking.
245 `first_reg' and `last_reg' track the head and tail of the
246 chain of registers which currently contain this quantity.
248 `mode' contains the machine mode of this quantity.
250 `const_rtx' holds the rtx of the constant value of this
251 quantity, if known. A summations of the frame/arg pointer
252 and a constant can also be entered here. When this holds
253 a known value, `const_insn' is the insn which stored the
254 constant value.
256 `comparison_{code,const,qty}' are used to track when a
257 comparison between a quantity and some constant or register has
258 been passed. In such a case, we know the results of the comparison
259 in case we see it again. These members record a comparison that
260 is known to be true. `comparison_code' holds the rtx code of such
261 a comparison, else it is set to UNKNOWN and the other two
262 comparison members are undefined. `comparison_const' holds
263 the constant being compared against, or zero if the comparison
264 is not against a constant. `comparison_qty' holds the quantity
265 being compared against when the result is known. If the comparison
266 is not with a register, `comparison_qty' is -1. */
268 struct qty_table_elem
270 rtx const_rtx;
271 rtx_insn *const_insn;
272 rtx comparison_const;
273 int comparison_qty;
274 unsigned int first_reg, last_reg;
275 /* The sizes of these fields should match the sizes of the
276 code and mode fields of struct rtx_def (see rtl.h). */
277 ENUM_BITFIELD(rtx_code) comparison_code : 16;
278 ENUM_BITFIELD(machine_mode) mode : 8;
281 /* The table of all qtys, indexed by qty number. */
282 static struct qty_table_elem *qty_table;
284 #ifdef HAVE_cc0
285 /* For machines that have a CC0, we do not record its value in the hash
286 table since its use is guaranteed to be the insn immediately following
287 its definition and any other insn is presumed to invalidate it.
289 Instead, we store below the current and last value assigned to CC0.
290 If it should happen to be a constant, it is stored in preference
291 to the actual assigned value. In case it is a constant, we store
292 the mode in which the constant should be interpreted. */
294 static rtx this_insn_cc0, prev_insn_cc0;
295 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
296 #endif
298 /* Insn being scanned. */
300 static rtx_insn *this_insn;
301 static bool optimize_this_for_speed_p;
303 /* Index by register number, gives the number of the next (or
304 previous) register in the chain of registers sharing the same
305 value.
307 Or -1 if this register is at the end of the chain.
309 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
311 /* Per-register equivalence chain. */
312 struct reg_eqv_elem
314 int next, prev;
317 /* The table of all register equivalence chains. */
318 static struct reg_eqv_elem *reg_eqv_table;
320 struct cse_reg_info
322 /* The timestamp at which this register is initialized. */
323 unsigned int timestamp;
325 /* The quantity number of the register's current contents. */
326 int reg_qty;
328 /* The number of times the register has been altered in the current
329 basic block. */
330 int reg_tick;
332 /* The REG_TICK value at which rtx's containing this register are
333 valid in the hash table. If this does not equal the current
334 reg_tick value, such expressions existing in the hash table are
335 invalid. */
336 int reg_in_table;
338 /* The SUBREG that was set when REG_TICK was last incremented. Set
339 to -1 if the last store was to the whole register, not a subreg. */
340 unsigned int subreg_ticked;
343 /* A table of cse_reg_info indexed by register numbers. */
344 static struct cse_reg_info *cse_reg_info_table;
346 /* The size of the above table. */
347 static unsigned int cse_reg_info_table_size;
349 /* The index of the first entry that has not been initialized. */
350 static unsigned int cse_reg_info_table_first_uninitialized;
352 /* The timestamp at the beginning of the current run of
353 cse_extended_basic_block. We increment this variable at the beginning of
354 the current run of cse_extended_basic_block. The timestamp field of a
355 cse_reg_info entry matches the value of this variable if and only
356 if the entry has been initialized during the current run of
357 cse_extended_basic_block. */
358 static unsigned int cse_reg_info_timestamp;
360 /* A HARD_REG_SET containing all the hard registers for which there is
361 currently a REG expression in the hash table. Note the difference
362 from the above variables, which indicate if the REG is mentioned in some
363 expression in the table. */
365 static HARD_REG_SET hard_regs_in_table;
367 /* True if CSE has altered the CFG. */
368 static bool cse_cfg_altered;
370 /* True if CSE has altered conditional jump insns in such a way
371 that jump optimization should be redone. */
372 static bool cse_jumps_altered;
374 /* True if we put a LABEL_REF into the hash table for an INSN
375 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
376 to put in the note. */
377 static bool recorded_label_ref;
379 /* canon_hash stores 1 in do_not_record
380 if it notices a reference to CC0, PC, or some other volatile
381 subexpression. */
383 static int do_not_record;
385 /* canon_hash stores 1 in hash_arg_in_memory
386 if it notices a reference to memory within the expression being hashed. */
388 static int hash_arg_in_memory;
390 /* The hash table contains buckets which are chains of `struct table_elt's,
391 each recording one expression's information.
392 That expression is in the `exp' field.
394 The canon_exp field contains a canonical (from the point of view of
395 alias analysis) version of the `exp' field.
397 Those elements with the same hash code are chained in both directions
398 through the `next_same_hash' and `prev_same_hash' fields.
400 Each set of expressions with equivalent values
401 are on a two-way chain through the `next_same_value'
402 and `prev_same_value' fields, and all point with
403 the `first_same_value' field at the first element in
404 that chain. The chain is in order of increasing cost.
405 Each element's cost value is in its `cost' field.
407 The `in_memory' field is nonzero for elements that
408 involve any reference to memory. These elements are removed
409 whenever a write is done to an unidentified location in memory.
410 To be safe, we assume that a memory address is unidentified unless
411 the address is either a symbol constant or a constant plus
412 the frame pointer or argument pointer.
414 The `related_value' field is used to connect related expressions
415 (that differ by adding an integer).
416 The related expressions are chained in a circular fashion.
417 `related_value' is zero for expressions for which this
418 chain is not useful.
420 The `cost' field stores the cost of this element's expression.
421 The `regcost' field stores the value returned by approx_reg_cost for
422 this element's expression.
424 The `is_const' flag is set if the element is a constant (including
425 a fixed address).
427 The `flag' field is used as a temporary during some search routines.
429 The `mode' field is usually the same as GET_MODE (`exp'), but
430 if `exp' is a CONST_INT and has no machine mode then the `mode'
431 field is the mode it was being used as. Each constant is
432 recorded separately for each mode it is used with. */
434 struct table_elt
436 rtx exp;
437 rtx canon_exp;
438 struct table_elt *next_same_hash;
439 struct table_elt *prev_same_hash;
440 struct table_elt *next_same_value;
441 struct table_elt *prev_same_value;
442 struct table_elt *first_same_value;
443 struct table_elt *related_value;
444 int cost;
445 int regcost;
446 /* The size of this field should match the size
447 of the mode field of struct rtx_def (see rtl.h). */
448 ENUM_BITFIELD(machine_mode) mode : 8;
449 char in_memory;
450 char is_const;
451 char flag;
454 /* We don't want a lot of buckets, because we rarely have very many
455 things stored in the hash table, and a lot of buckets slows
456 down a lot of loops that happen frequently. */
457 #define HASH_SHIFT 5
458 #define HASH_SIZE (1 << HASH_SHIFT)
459 #define HASH_MASK (HASH_SIZE - 1)
461 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
462 register (hard registers may require `do_not_record' to be set). */
464 #define HASH(X, M) \
465 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
466 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
467 : canon_hash (X, M)) & HASH_MASK)
469 /* Like HASH, but without side-effects. */
470 #define SAFE_HASH(X, M) \
471 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
472 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
473 : safe_hash (X, M)) & HASH_MASK)
475 /* Determine whether register number N is considered a fixed register for the
476 purpose of approximating register costs.
477 It is desirable to replace other regs with fixed regs, to reduce need for
478 non-fixed hard regs.
479 A reg wins if it is either the frame pointer or designated as fixed. */
480 #define FIXED_REGNO_P(N) \
481 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
482 || fixed_regs[N] || global_regs[N])
484 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
485 hard registers and pointers into the frame are the cheapest with a cost
486 of 0. Next come pseudos with a cost of one and other hard registers with
487 a cost of 2. Aside from these special cases, call `rtx_cost'. */
489 #define CHEAP_REGNO(N) \
490 (REGNO_PTR_FRAME_P (N) \
491 || (HARD_REGISTER_NUM_P (N) \
492 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
494 #define COST(X) (REG_P (X) ? 0 : notreg_cost (X, SET, 1))
495 #define COST_IN(X, OUTER, OPNO) (REG_P (X) ? 0 : notreg_cost (X, OUTER, OPNO))
497 /* Get the number of times this register has been updated in this
498 basic block. */
500 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
502 /* Get the point at which REG was recorded in the table. */
504 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
506 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
507 SUBREG). */
509 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
511 /* Get the quantity number for REG. */
513 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
515 /* Determine if the quantity number for register X represents a valid index
516 into the qty_table. */
518 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
520 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
522 #define CHEAPER(X, Y) \
523 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
525 static struct table_elt *table[HASH_SIZE];
527 /* Chain of `struct table_elt's made so far for this function
528 but currently removed from the table. */
530 static struct table_elt *free_element_chain;
532 /* Set to the cost of a constant pool reference if one was found for a
533 symbolic constant. If this was found, it means we should try to
534 convert constants into constant pool entries if they don't fit in
535 the insn. */
537 static int constant_pool_entries_cost;
538 static int constant_pool_entries_regcost;
540 /* Trace a patch through the CFG. */
542 struct branch_path
544 /* The basic block for this path entry. */
545 basic_block bb;
548 /* This data describes a block that will be processed by
549 cse_extended_basic_block. */
551 struct cse_basic_block_data
553 /* Total number of SETs in block. */
554 int nsets;
555 /* Size of current branch path, if any. */
556 int path_size;
557 /* Current path, indicating which basic_blocks will be processed. */
558 struct branch_path *path;
562 /* Pointers to the live in/live out bitmaps for the boundaries of the
563 current EBB. */
564 static bitmap cse_ebb_live_in, cse_ebb_live_out;
566 /* A simple bitmap to track which basic blocks have been visited
567 already as part of an already processed extended basic block. */
568 static sbitmap cse_visited_basic_blocks;
570 static bool fixed_base_plus_p (rtx x);
571 static int notreg_cost (rtx, enum rtx_code, int);
572 static int preferable (int, int, int, int);
573 static void new_basic_block (void);
574 static void make_new_qty (unsigned int, machine_mode);
575 static void make_regs_eqv (unsigned int, unsigned int);
576 static void delete_reg_equiv (unsigned int);
577 static int mention_regs (rtx);
578 static int insert_regs (rtx, struct table_elt *, int);
579 static void remove_from_table (struct table_elt *, unsigned);
580 static void remove_pseudo_from_table (rtx, unsigned);
581 static struct table_elt *lookup (rtx, unsigned, machine_mode);
582 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
583 static rtx lookup_as_function (rtx, enum rtx_code);
584 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
585 machine_mode, int, int);
586 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
587 machine_mode);
588 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
589 static void invalidate (rtx, machine_mode);
590 static void remove_invalid_refs (unsigned int);
591 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
592 machine_mode);
593 static void rehash_using_reg (rtx);
594 static void invalidate_memory (void);
595 static void invalidate_for_call (void);
596 static rtx use_related_value (rtx, struct table_elt *);
598 static inline unsigned canon_hash (rtx, machine_mode);
599 static inline unsigned safe_hash (rtx, machine_mode);
600 static inline unsigned hash_rtx_string (const char *);
602 static rtx canon_reg (rtx, rtx_insn *);
603 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
604 machine_mode *,
605 machine_mode *);
606 static rtx fold_rtx (rtx, rtx_insn *);
607 static rtx equiv_constant (rtx);
608 static void record_jump_equiv (rtx_insn *, bool);
609 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
610 int);
611 static void cse_insn (rtx_insn *);
612 static void cse_prescan_path (struct cse_basic_block_data *);
613 static void invalidate_from_clobbers (rtx_insn *);
614 static void invalidate_from_sets_and_clobbers (rtx_insn *);
615 static rtx cse_process_notes (rtx, rtx, bool *);
616 static void cse_extended_basic_block (struct cse_basic_block_data *);
617 extern void dump_class (struct table_elt*);
618 static void get_cse_reg_info_1 (unsigned int regno);
619 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
621 static void flush_hash_table (void);
622 static bool insn_live_p (rtx_insn *, int *);
623 static bool set_live_p (rtx, rtx_insn *, int *);
624 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
625 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
626 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
627 bool);
630 #undef RTL_HOOKS_GEN_LOWPART
631 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
633 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
635 /* Nonzero if X has the form (PLUS frame-pointer integer). */
637 static bool
638 fixed_base_plus_p (rtx x)
640 switch (GET_CODE (x))
642 case REG:
643 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
644 return true;
645 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
646 return true;
647 return false;
649 case PLUS:
650 if (!CONST_INT_P (XEXP (x, 1)))
651 return false;
652 return fixed_base_plus_p (XEXP (x, 0));
654 default:
655 return false;
659 /* Dump the expressions in the equivalence class indicated by CLASSP.
660 This function is used only for debugging. */
661 DEBUG_FUNCTION void
662 dump_class (struct table_elt *classp)
664 struct table_elt *elt;
666 fprintf (stderr, "Equivalence chain for ");
667 print_rtl (stderr, classp->exp);
668 fprintf (stderr, ": \n");
670 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
672 print_rtl (stderr, elt->exp);
673 fprintf (stderr, "\n");
677 /* Return an estimate of the cost of the registers used in an rtx.
678 This is mostly the number of different REG expressions in the rtx;
679 however for some exceptions like fixed registers we use a cost of
680 0. If any other hard register reference occurs, return MAX_COST. */
682 static int
683 approx_reg_cost (const_rtx x)
685 int cost = 0;
686 subrtx_iterator::array_type array;
687 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
689 const_rtx x = *iter;
690 if (REG_P (x))
692 unsigned int regno = REGNO (x);
693 if (!CHEAP_REGNO (regno))
695 if (regno < FIRST_PSEUDO_REGISTER)
697 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
698 return MAX_COST;
699 cost += 2;
701 else
702 cost += 1;
706 return cost;
709 /* Return a negative value if an rtx A, whose costs are given by COST_A
710 and REGCOST_A, is more desirable than an rtx B.
711 Return a positive value if A is less desirable, or 0 if the two are
712 equally good. */
713 static int
714 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
716 /* First, get rid of cases involving expressions that are entirely
717 unwanted. */
718 if (cost_a != cost_b)
720 if (cost_a == MAX_COST)
721 return 1;
722 if (cost_b == MAX_COST)
723 return -1;
726 /* Avoid extending lifetimes of hardregs. */
727 if (regcost_a != regcost_b)
729 if (regcost_a == MAX_COST)
730 return 1;
731 if (regcost_b == MAX_COST)
732 return -1;
735 /* Normal operation costs take precedence. */
736 if (cost_a != cost_b)
737 return cost_a - cost_b;
738 /* Only if these are identical consider effects on register pressure. */
739 if (regcost_a != regcost_b)
740 return regcost_a - regcost_b;
741 return 0;
744 /* Internal function, to compute cost when X is not a register; called
745 from COST macro to keep it simple. */
747 static int
748 notreg_cost (rtx x, enum rtx_code outer, int opno)
750 return ((GET_CODE (x) == SUBREG
751 && REG_P (SUBREG_REG (x))
752 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT
753 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
754 && (GET_MODE_SIZE (GET_MODE (x))
755 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
756 && subreg_lowpart_p (x)
757 && TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (x),
758 GET_MODE (SUBREG_REG (x))))
760 : rtx_cost (x, outer, opno, optimize_this_for_speed_p) * 2);
764 /* Initialize CSE_REG_INFO_TABLE. */
766 static void
767 init_cse_reg_info (unsigned int nregs)
769 /* Do we need to grow the table? */
770 if (nregs > cse_reg_info_table_size)
772 unsigned int new_size;
774 if (cse_reg_info_table_size < 2048)
776 /* Compute a new size that is a power of 2 and no smaller
777 than the large of NREGS and 64. */
778 new_size = (cse_reg_info_table_size
779 ? cse_reg_info_table_size : 64);
781 while (new_size < nregs)
782 new_size *= 2;
784 else
786 /* If we need a big table, allocate just enough to hold
787 NREGS registers. */
788 new_size = nregs;
791 /* Reallocate the table with NEW_SIZE entries. */
792 free (cse_reg_info_table);
793 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
794 cse_reg_info_table_size = new_size;
795 cse_reg_info_table_first_uninitialized = 0;
798 /* Do we have all of the first NREGS entries initialized? */
799 if (cse_reg_info_table_first_uninitialized < nregs)
801 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
802 unsigned int i;
804 /* Put the old timestamp on newly allocated entries so that they
805 will all be considered out of date. We do not touch those
806 entries beyond the first NREGS entries to be nice to the
807 virtual memory. */
808 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
809 cse_reg_info_table[i].timestamp = old_timestamp;
811 cse_reg_info_table_first_uninitialized = nregs;
815 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
817 static void
818 get_cse_reg_info_1 (unsigned int regno)
820 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
821 entry will be considered to have been initialized. */
822 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
824 /* Initialize the rest of the entry. */
825 cse_reg_info_table[regno].reg_tick = 1;
826 cse_reg_info_table[regno].reg_in_table = -1;
827 cse_reg_info_table[regno].subreg_ticked = -1;
828 cse_reg_info_table[regno].reg_qty = -regno - 1;
831 /* Find a cse_reg_info entry for REGNO. */
833 static inline struct cse_reg_info *
834 get_cse_reg_info (unsigned int regno)
836 struct cse_reg_info *p = &cse_reg_info_table[regno];
838 /* If this entry has not been initialized, go ahead and initialize
839 it. */
840 if (p->timestamp != cse_reg_info_timestamp)
841 get_cse_reg_info_1 (regno);
843 return p;
846 /* Clear the hash table and initialize each register with its own quantity,
847 for a new basic block. */
849 static void
850 new_basic_block (void)
852 int i;
854 next_qty = 0;
856 /* Invalidate cse_reg_info_table. */
857 cse_reg_info_timestamp++;
859 /* Clear out hash table state for this pass. */
860 CLEAR_HARD_REG_SET (hard_regs_in_table);
862 /* The per-quantity values used to be initialized here, but it is
863 much faster to initialize each as it is made in `make_new_qty'. */
865 for (i = 0; i < HASH_SIZE; i++)
867 struct table_elt *first;
869 first = table[i];
870 if (first != NULL)
872 struct table_elt *last = first;
874 table[i] = NULL;
876 while (last->next_same_hash != NULL)
877 last = last->next_same_hash;
879 /* Now relink this hash entire chain into
880 the free element list. */
882 last->next_same_hash = free_element_chain;
883 free_element_chain = first;
887 #ifdef HAVE_cc0
888 prev_insn_cc0 = 0;
889 #endif
892 /* Say that register REG contains a quantity in mode MODE not in any
893 register before and initialize that quantity. */
895 static void
896 make_new_qty (unsigned int reg, machine_mode mode)
898 int q;
899 struct qty_table_elem *ent;
900 struct reg_eqv_elem *eqv;
902 gcc_assert (next_qty < max_qty);
904 q = REG_QTY (reg) = next_qty++;
905 ent = &qty_table[q];
906 ent->first_reg = reg;
907 ent->last_reg = reg;
908 ent->mode = mode;
909 ent->const_rtx = ent->const_insn = NULL;
910 ent->comparison_code = UNKNOWN;
912 eqv = &reg_eqv_table[reg];
913 eqv->next = eqv->prev = -1;
916 /* Make reg NEW equivalent to reg OLD.
917 OLD is not changing; NEW is. */
919 static void
920 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
922 unsigned int lastr, firstr;
923 int q = REG_QTY (old_reg);
924 struct qty_table_elem *ent;
926 ent = &qty_table[q];
928 /* Nothing should become eqv until it has a "non-invalid" qty number. */
929 gcc_assert (REGNO_QTY_VALID_P (old_reg));
931 REG_QTY (new_reg) = q;
932 firstr = ent->first_reg;
933 lastr = ent->last_reg;
935 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
936 hard regs. Among pseudos, if NEW will live longer than any other reg
937 of the same qty, and that is beyond the current basic block,
938 make it the new canonical replacement for this qty. */
939 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
940 /* Certain fixed registers might be of the class NO_REGS. This means
941 that not only can they not be allocated by the compiler, but
942 they cannot be used in substitutions or canonicalizations
943 either. */
944 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
945 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
946 || (new_reg >= FIRST_PSEUDO_REGISTER
947 && (firstr < FIRST_PSEUDO_REGISTER
948 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
949 && !bitmap_bit_p (cse_ebb_live_out, firstr))
950 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
951 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
953 reg_eqv_table[firstr].prev = new_reg;
954 reg_eqv_table[new_reg].next = firstr;
955 reg_eqv_table[new_reg].prev = -1;
956 ent->first_reg = new_reg;
958 else
960 /* If NEW is a hard reg (known to be non-fixed), insert at end.
961 Otherwise, insert before any non-fixed hard regs that are at the
962 end. Registers of class NO_REGS cannot be used as an
963 equivalent for anything. */
964 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
965 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
966 && new_reg >= FIRST_PSEUDO_REGISTER)
967 lastr = reg_eqv_table[lastr].prev;
968 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
969 if (reg_eqv_table[lastr].next >= 0)
970 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
971 else
972 qty_table[q].last_reg = new_reg;
973 reg_eqv_table[lastr].next = new_reg;
974 reg_eqv_table[new_reg].prev = lastr;
978 /* Remove REG from its equivalence class. */
980 static void
981 delete_reg_equiv (unsigned int reg)
983 struct qty_table_elem *ent;
984 int q = REG_QTY (reg);
985 int p, n;
987 /* If invalid, do nothing. */
988 if (! REGNO_QTY_VALID_P (reg))
989 return;
991 ent = &qty_table[q];
993 p = reg_eqv_table[reg].prev;
994 n = reg_eqv_table[reg].next;
996 if (n != -1)
997 reg_eqv_table[n].prev = p;
998 else
999 ent->last_reg = p;
1000 if (p != -1)
1001 reg_eqv_table[p].next = n;
1002 else
1003 ent->first_reg = n;
1005 REG_QTY (reg) = -reg - 1;
1008 /* Remove any invalid expressions from the hash table
1009 that refer to any of the registers contained in expression X.
1011 Make sure that newly inserted references to those registers
1012 as subexpressions will be considered valid.
1014 mention_regs is not called when a register itself
1015 is being stored in the table.
1017 Return 1 if we have done something that may have changed the hash code
1018 of X. */
1020 static int
1021 mention_regs (rtx x)
1023 enum rtx_code code;
1024 int i, j;
1025 const char *fmt;
1026 int changed = 0;
1028 if (x == 0)
1029 return 0;
1031 code = GET_CODE (x);
1032 if (code == REG)
1034 unsigned int regno = REGNO (x);
1035 unsigned int endregno = END_REGNO (x);
1036 unsigned int i;
1038 for (i = regno; i < endregno; i++)
1040 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1041 remove_invalid_refs (i);
1043 REG_IN_TABLE (i) = REG_TICK (i);
1044 SUBREG_TICKED (i) = -1;
1047 return 0;
1050 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1051 pseudo if they don't use overlapping words. We handle only pseudos
1052 here for simplicity. */
1053 if (code == SUBREG && REG_P (SUBREG_REG (x))
1054 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1056 unsigned int i = REGNO (SUBREG_REG (x));
1058 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1060 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1061 the last store to this register really stored into this
1062 subreg, then remove the memory of this subreg.
1063 Otherwise, remove any memory of the entire register and
1064 all its subregs from the table. */
1065 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1066 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1067 remove_invalid_refs (i);
1068 else
1069 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1072 REG_IN_TABLE (i) = REG_TICK (i);
1073 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1074 return 0;
1077 /* If X is a comparison or a COMPARE and either operand is a register
1078 that does not have a quantity, give it one. This is so that a later
1079 call to record_jump_equiv won't cause X to be assigned a different
1080 hash code and not found in the table after that call.
1082 It is not necessary to do this here, since rehash_using_reg can
1083 fix up the table later, but doing this here eliminates the need to
1084 call that expensive function in the most common case where the only
1085 use of the register is in the comparison. */
1087 if (code == COMPARE || COMPARISON_P (x))
1089 if (REG_P (XEXP (x, 0))
1090 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1091 if (insert_regs (XEXP (x, 0), NULL, 0))
1093 rehash_using_reg (XEXP (x, 0));
1094 changed = 1;
1097 if (REG_P (XEXP (x, 1))
1098 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1099 if (insert_regs (XEXP (x, 1), NULL, 0))
1101 rehash_using_reg (XEXP (x, 1));
1102 changed = 1;
1106 fmt = GET_RTX_FORMAT (code);
1107 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1108 if (fmt[i] == 'e')
1109 changed |= mention_regs (XEXP (x, i));
1110 else if (fmt[i] == 'E')
1111 for (j = 0; j < XVECLEN (x, i); j++)
1112 changed |= mention_regs (XVECEXP (x, i, j));
1114 return changed;
1117 /* Update the register quantities for inserting X into the hash table
1118 with a value equivalent to CLASSP.
1119 (If the class does not contain a REG, it is irrelevant.)
1120 If MODIFIED is nonzero, X is a destination; it is being modified.
1121 Note that delete_reg_equiv should be called on a register
1122 before insert_regs is done on that register with MODIFIED != 0.
1124 Nonzero value means that elements of reg_qty have changed
1125 so X's hash code may be different. */
1127 static int
1128 insert_regs (rtx x, struct table_elt *classp, int modified)
1130 if (REG_P (x))
1132 unsigned int regno = REGNO (x);
1133 int qty_valid;
1135 /* If REGNO is in the equivalence table already but is of the
1136 wrong mode for that equivalence, don't do anything here. */
1138 qty_valid = REGNO_QTY_VALID_P (regno);
1139 if (qty_valid)
1141 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1143 if (ent->mode != GET_MODE (x))
1144 return 0;
1147 if (modified || ! qty_valid)
1149 if (classp)
1150 for (classp = classp->first_same_value;
1151 classp != 0;
1152 classp = classp->next_same_value)
1153 if (REG_P (classp->exp)
1154 && GET_MODE (classp->exp) == GET_MODE (x))
1156 unsigned c_regno = REGNO (classp->exp);
1158 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1160 /* Suppose that 5 is hard reg and 100 and 101 are
1161 pseudos. Consider
1163 (set (reg:si 100) (reg:si 5))
1164 (set (reg:si 5) (reg:si 100))
1165 (set (reg:di 101) (reg:di 5))
1167 We would now set REG_QTY (101) = REG_QTY (5), but the
1168 entry for 5 is in SImode. When we use this later in
1169 copy propagation, we get the register in wrong mode. */
1170 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1171 continue;
1173 make_regs_eqv (regno, c_regno);
1174 return 1;
1177 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1178 than REG_IN_TABLE to find out if there was only a single preceding
1179 invalidation - for the SUBREG - or another one, which would be
1180 for the full register. However, if we find here that REG_TICK
1181 indicates that the register is invalid, it means that it has
1182 been invalidated in a separate operation. The SUBREG might be used
1183 now (then this is a recursive call), or we might use the full REG
1184 now and a SUBREG of it later. So bump up REG_TICK so that
1185 mention_regs will do the right thing. */
1186 if (! modified
1187 && REG_IN_TABLE (regno) >= 0
1188 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1189 REG_TICK (regno)++;
1190 make_new_qty (regno, GET_MODE (x));
1191 return 1;
1194 return 0;
1197 /* If X is a SUBREG, we will likely be inserting the inner register in the
1198 table. If that register doesn't have an assigned quantity number at
1199 this point but does later, the insertion that we will be doing now will
1200 not be accessible because its hash code will have changed. So assign
1201 a quantity number now. */
1203 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1204 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1206 insert_regs (SUBREG_REG (x), NULL, 0);
1207 mention_regs (x);
1208 return 1;
1210 else
1211 return mention_regs (x);
1215 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1216 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1217 CST is equal to an anchor. */
1219 static bool
1220 compute_const_anchors (rtx cst,
1221 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1222 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1224 HOST_WIDE_INT n = INTVAL (cst);
1226 *lower_base = n & ~(targetm.const_anchor - 1);
1227 if (*lower_base == n)
1228 return false;
1230 *upper_base =
1231 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1232 *upper_offs = n - *upper_base;
1233 *lower_offs = n - *lower_base;
1234 return true;
1237 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1239 static void
1240 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1241 machine_mode mode)
1243 struct table_elt *elt;
1244 unsigned hash;
1245 rtx anchor_exp;
1246 rtx exp;
1248 anchor_exp = GEN_INT (anchor);
1249 hash = HASH (anchor_exp, mode);
1250 elt = lookup (anchor_exp, hash, mode);
1251 if (!elt)
1252 elt = insert (anchor_exp, NULL, hash, mode);
1254 exp = plus_constant (mode, reg, offs);
1255 /* REG has just been inserted and the hash codes recomputed. */
1256 mention_regs (exp);
1257 hash = HASH (exp, mode);
1259 /* Use the cost of the register rather than the whole expression. When
1260 looking up constant anchors we will further offset the corresponding
1261 expression therefore it does not make sense to prefer REGs over
1262 reg-immediate additions. Prefer instead the oldest expression. Also
1263 don't prefer pseudos over hard regs so that we derive constants in
1264 argument registers from other argument registers rather than from the
1265 original pseudo that was used to synthesize the constant. */
1266 insert_with_costs (exp, elt, hash, mode, COST (reg), 1);
1269 /* The constant CST is equivalent to the register REG. Create
1270 equivalences between the two anchors of CST and the corresponding
1271 register-offset expressions using REG. */
1273 static void
1274 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1276 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1278 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1279 &upper_base, &upper_offs))
1280 return;
1282 /* Ignore anchors of value 0. Constants accessible from zero are
1283 simple. */
1284 if (lower_base != 0)
1285 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1287 if (upper_base != 0)
1288 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1291 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1292 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1293 valid expression. Return the cheapest and oldest of such expressions. In
1294 *OLD, return how old the resulting expression is compared to the other
1295 equivalent expressions. */
1297 static rtx
1298 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1299 unsigned *old)
1301 struct table_elt *elt;
1302 unsigned idx;
1303 struct table_elt *match_elt;
1304 rtx match;
1306 /* Find the cheapest and *oldest* expression to maximize the chance of
1307 reusing the same pseudo. */
1309 match_elt = NULL;
1310 match = NULL_RTX;
1311 for (elt = anchor_elt->first_same_value, idx = 0;
1312 elt;
1313 elt = elt->next_same_value, idx++)
1315 if (match_elt && CHEAPER (match_elt, elt))
1316 return match;
1318 if (REG_P (elt->exp)
1319 || (GET_CODE (elt->exp) == PLUS
1320 && REG_P (XEXP (elt->exp, 0))
1321 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1323 rtx x;
1325 /* Ignore expressions that are no longer valid. */
1326 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1327 continue;
1329 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1330 if (REG_P (x)
1331 || (GET_CODE (x) == PLUS
1332 && IN_RANGE (INTVAL (XEXP (x, 1)),
1333 -targetm.const_anchor,
1334 targetm.const_anchor - 1)))
1336 match = x;
1337 match_elt = elt;
1338 *old = idx;
1343 return match;
1346 /* Try to express the constant SRC_CONST using a register+offset expression
1347 derived from a constant anchor. Return it if successful or NULL_RTX,
1348 otherwise. */
1350 static rtx
1351 try_const_anchors (rtx src_const, machine_mode mode)
1353 struct table_elt *lower_elt, *upper_elt;
1354 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1355 rtx lower_anchor_rtx, upper_anchor_rtx;
1356 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1357 unsigned lower_old, upper_old;
1359 /* CONST_INT is used for CC modes, but we should leave those alone. */
1360 if (GET_MODE_CLASS (mode) == MODE_CC)
1361 return NULL_RTX;
1363 gcc_assert (SCALAR_INT_MODE_P (mode));
1364 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1365 &upper_base, &upper_offs))
1366 return NULL_RTX;
1368 lower_anchor_rtx = GEN_INT (lower_base);
1369 upper_anchor_rtx = GEN_INT (upper_base);
1370 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1371 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1373 if (lower_elt)
1374 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1375 if (upper_elt)
1376 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1378 if (!lower_exp)
1379 return upper_exp;
1380 if (!upper_exp)
1381 return lower_exp;
1383 /* Return the older expression. */
1384 return (upper_old > lower_old ? upper_exp : lower_exp);
1387 /* Look in or update the hash table. */
1389 /* Remove table element ELT from use in the table.
1390 HASH is its hash code, made using the HASH macro.
1391 It's an argument because often that is known in advance
1392 and we save much time not recomputing it. */
1394 static void
1395 remove_from_table (struct table_elt *elt, unsigned int hash)
1397 if (elt == 0)
1398 return;
1400 /* Mark this element as removed. See cse_insn. */
1401 elt->first_same_value = 0;
1403 /* Remove the table element from its equivalence class. */
1406 struct table_elt *prev = elt->prev_same_value;
1407 struct table_elt *next = elt->next_same_value;
1409 if (next)
1410 next->prev_same_value = prev;
1412 if (prev)
1413 prev->next_same_value = next;
1414 else
1416 struct table_elt *newfirst = next;
1417 while (next)
1419 next->first_same_value = newfirst;
1420 next = next->next_same_value;
1425 /* Remove the table element from its hash bucket. */
1428 struct table_elt *prev = elt->prev_same_hash;
1429 struct table_elt *next = elt->next_same_hash;
1431 if (next)
1432 next->prev_same_hash = prev;
1434 if (prev)
1435 prev->next_same_hash = next;
1436 else if (table[hash] == elt)
1437 table[hash] = next;
1438 else
1440 /* This entry is not in the proper hash bucket. This can happen
1441 when two classes were merged by `merge_equiv_classes'. Search
1442 for the hash bucket that it heads. This happens only very
1443 rarely, so the cost is acceptable. */
1444 for (hash = 0; hash < HASH_SIZE; hash++)
1445 if (table[hash] == elt)
1446 table[hash] = next;
1450 /* Remove the table element from its related-value circular chain. */
1452 if (elt->related_value != 0 && elt->related_value != elt)
1454 struct table_elt *p = elt->related_value;
1456 while (p->related_value != elt)
1457 p = p->related_value;
1458 p->related_value = elt->related_value;
1459 if (p->related_value == p)
1460 p->related_value = 0;
1463 /* Now add it to the free element chain. */
1464 elt->next_same_hash = free_element_chain;
1465 free_element_chain = elt;
1468 /* Same as above, but X is a pseudo-register. */
1470 static void
1471 remove_pseudo_from_table (rtx x, unsigned int hash)
1473 struct table_elt *elt;
1475 /* Because a pseudo-register can be referenced in more than one
1476 mode, we might have to remove more than one table entry. */
1477 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1478 remove_from_table (elt, hash);
1481 /* Look up X in the hash table and return its table element,
1482 or 0 if X is not in the table.
1484 MODE is the machine-mode of X, or if X is an integer constant
1485 with VOIDmode then MODE is the mode with which X will be used.
1487 Here we are satisfied to find an expression whose tree structure
1488 looks like X. */
1490 static struct table_elt *
1491 lookup (rtx x, unsigned int hash, machine_mode mode)
1493 struct table_elt *p;
1495 for (p = table[hash]; p; p = p->next_same_hash)
1496 if (mode == p->mode && ((x == p->exp && REG_P (x))
1497 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1498 return p;
1500 return 0;
1503 /* Like `lookup' but don't care whether the table element uses invalid regs.
1504 Also ignore discrepancies in the machine mode of a register. */
1506 static struct table_elt *
1507 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1509 struct table_elt *p;
1511 if (REG_P (x))
1513 unsigned int regno = REGNO (x);
1515 /* Don't check the machine mode when comparing registers;
1516 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1517 for (p = table[hash]; p; p = p->next_same_hash)
1518 if (REG_P (p->exp)
1519 && REGNO (p->exp) == regno)
1520 return p;
1522 else
1524 for (p = table[hash]; p; p = p->next_same_hash)
1525 if (mode == p->mode
1526 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1527 return p;
1530 return 0;
1533 /* Look for an expression equivalent to X and with code CODE.
1534 If one is found, return that expression. */
1536 static rtx
1537 lookup_as_function (rtx x, enum rtx_code code)
1539 struct table_elt *p
1540 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1542 if (p == 0)
1543 return 0;
1545 for (p = p->first_same_value; p; p = p->next_same_value)
1546 if (GET_CODE (p->exp) == code
1547 /* Make sure this is a valid entry in the table. */
1548 && exp_equiv_p (p->exp, p->exp, 1, false))
1549 return p->exp;
1551 return 0;
1554 /* Insert X in the hash table, assuming HASH is its hash code and
1555 CLASSP is an element of the class it should go in (or 0 if a new
1556 class should be made). COST is the code of X and reg_cost is the
1557 cost of registers in X. It is inserted at the proper position to
1558 keep the class in the order cheapest first.
1560 MODE is the machine-mode of X, or if X is an integer constant
1561 with VOIDmode then MODE is the mode with which X will be used.
1563 For elements of equal cheapness, the most recent one
1564 goes in front, except that the first element in the list
1565 remains first unless a cheaper element is added. The order of
1566 pseudo-registers does not matter, as canon_reg will be called to
1567 find the cheapest when a register is retrieved from the table.
1569 The in_memory field in the hash table element is set to 0.
1570 The caller must set it nonzero if appropriate.
1572 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1573 and if insert_regs returns a nonzero value
1574 you must then recompute its hash code before calling here.
1576 If necessary, update table showing constant values of quantities. */
1578 static struct table_elt *
1579 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1580 machine_mode mode, int cost, int reg_cost)
1582 struct table_elt *elt;
1584 /* If X is a register and we haven't made a quantity for it,
1585 something is wrong. */
1586 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1588 /* If X is a hard register, show it is being put in the table. */
1589 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1590 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1592 /* Put an element for X into the right hash bucket. */
1594 elt = free_element_chain;
1595 if (elt)
1596 free_element_chain = elt->next_same_hash;
1597 else
1598 elt = XNEW (struct table_elt);
1600 elt->exp = x;
1601 elt->canon_exp = NULL_RTX;
1602 elt->cost = cost;
1603 elt->regcost = reg_cost;
1604 elt->next_same_value = 0;
1605 elt->prev_same_value = 0;
1606 elt->next_same_hash = table[hash];
1607 elt->prev_same_hash = 0;
1608 elt->related_value = 0;
1609 elt->in_memory = 0;
1610 elt->mode = mode;
1611 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1613 if (table[hash])
1614 table[hash]->prev_same_hash = elt;
1615 table[hash] = elt;
1617 /* Put it into the proper value-class. */
1618 if (classp)
1620 classp = classp->first_same_value;
1621 if (CHEAPER (elt, classp))
1622 /* Insert at the head of the class. */
1624 struct table_elt *p;
1625 elt->next_same_value = classp;
1626 classp->prev_same_value = elt;
1627 elt->first_same_value = elt;
1629 for (p = classp; p; p = p->next_same_value)
1630 p->first_same_value = elt;
1632 else
1634 /* Insert not at head of the class. */
1635 /* Put it after the last element cheaper than X. */
1636 struct table_elt *p, *next;
1638 for (p = classp;
1639 (next = p->next_same_value) && CHEAPER (next, elt);
1640 p = next)
1643 /* Put it after P and before NEXT. */
1644 elt->next_same_value = next;
1645 if (next)
1646 next->prev_same_value = elt;
1648 elt->prev_same_value = p;
1649 p->next_same_value = elt;
1650 elt->first_same_value = classp;
1653 else
1654 elt->first_same_value = elt;
1656 /* If this is a constant being set equivalent to a register or a register
1657 being set equivalent to a constant, note the constant equivalence.
1659 If this is a constant, it cannot be equivalent to a different constant,
1660 and a constant is the only thing that can be cheaper than a register. So
1661 we know the register is the head of the class (before the constant was
1662 inserted).
1664 If this is a register that is not already known equivalent to a
1665 constant, we must check the entire class.
1667 If this is a register that is already known equivalent to an insn,
1668 update the qtys `const_insn' to show that `this_insn' is the latest
1669 insn making that quantity equivalent to the constant. */
1671 if (elt->is_const && classp && REG_P (classp->exp)
1672 && !REG_P (x))
1674 int exp_q = REG_QTY (REGNO (classp->exp));
1675 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1677 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1678 exp_ent->const_insn = this_insn;
1681 else if (REG_P (x)
1682 && classp
1683 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1684 && ! elt->is_const)
1686 struct table_elt *p;
1688 for (p = classp; p != 0; p = p->next_same_value)
1690 if (p->is_const && !REG_P (p->exp))
1692 int x_q = REG_QTY (REGNO (x));
1693 struct qty_table_elem *x_ent = &qty_table[x_q];
1695 x_ent->const_rtx
1696 = gen_lowpart (GET_MODE (x), p->exp);
1697 x_ent->const_insn = this_insn;
1698 break;
1703 else if (REG_P (x)
1704 && qty_table[REG_QTY (REGNO (x))].const_rtx
1705 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1706 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1708 /* If this is a constant with symbolic value,
1709 and it has a term with an explicit integer value,
1710 link it up with related expressions. */
1711 if (GET_CODE (x) == CONST)
1713 rtx subexp = get_related_value (x);
1714 unsigned subhash;
1715 struct table_elt *subelt, *subelt_prev;
1717 if (subexp != 0)
1719 /* Get the integer-free subexpression in the hash table. */
1720 subhash = SAFE_HASH (subexp, mode);
1721 subelt = lookup (subexp, subhash, mode);
1722 if (subelt == 0)
1723 subelt = insert (subexp, NULL, subhash, mode);
1724 /* Initialize SUBELT's circular chain if it has none. */
1725 if (subelt->related_value == 0)
1726 subelt->related_value = subelt;
1727 /* Find the element in the circular chain that precedes SUBELT. */
1728 subelt_prev = subelt;
1729 while (subelt_prev->related_value != subelt)
1730 subelt_prev = subelt_prev->related_value;
1731 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1732 This way the element that follows SUBELT is the oldest one. */
1733 elt->related_value = subelt_prev->related_value;
1734 subelt_prev->related_value = elt;
1738 return elt;
1741 /* Wrap insert_with_costs by passing the default costs. */
1743 static struct table_elt *
1744 insert (rtx x, struct table_elt *classp, unsigned int hash,
1745 machine_mode mode)
1747 return
1748 insert_with_costs (x, classp, hash, mode, COST (x), approx_reg_cost (x));
1752 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1753 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1754 the two classes equivalent.
1756 CLASS1 will be the surviving class; CLASS2 should not be used after this
1757 call.
1759 Any invalid entries in CLASS2 will not be copied. */
1761 static void
1762 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1764 struct table_elt *elt, *next, *new_elt;
1766 /* Ensure we start with the head of the classes. */
1767 class1 = class1->first_same_value;
1768 class2 = class2->first_same_value;
1770 /* If they were already equal, forget it. */
1771 if (class1 == class2)
1772 return;
1774 for (elt = class2; elt; elt = next)
1776 unsigned int hash;
1777 rtx exp = elt->exp;
1778 machine_mode mode = elt->mode;
1780 next = elt->next_same_value;
1782 /* Remove old entry, make a new one in CLASS1's class.
1783 Don't do this for invalid entries as we cannot find their
1784 hash code (it also isn't necessary). */
1785 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1787 bool need_rehash = false;
1789 hash_arg_in_memory = 0;
1790 hash = HASH (exp, mode);
1792 if (REG_P (exp))
1794 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1795 delete_reg_equiv (REGNO (exp));
1798 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1799 remove_pseudo_from_table (exp, hash);
1800 else
1801 remove_from_table (elt, hash);
1803 if (insert_regs (exp, class1, 0) || need_rehash)
1805 rehash_using_reg (exp);
1806 hash = HASH (exp, mode);
1808 new_elt = insert (exp, class1, hash, mode);
1809 new_elt->in_memory = hash_arg_in_memory;
1814 /* Flush the entire hash table. */
1816 static void
1817 flush_hash_table (void)
1819 int i;
1820 struct table_elt *p;
1822 for (i = 0; i < HASH_SIZE; i++)
1823 for (p = table[i]; p; p = table[i])
1825 /* Note that invalidate can remove elements
1826 after P in the current hash chain. */
1827 if (REG_P (p->exp))
1828 invalidate (p->exp, VOIDmode);
1829 else
1830 remove_from_table (p, i);
1834 /* Check whether an anti dependence exists between X and EXP. MODE and
1835 ADDR are as for canon_anti_dependence. */
1837 static bool
1838 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1840 subrtx_iterator::array_type array;
1841 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1843 const_rtx x = *iter;
1844 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1845 return true;
1847 return false;
1850 /* Remove from the hash table, or mark as invalid, all expressions whose
1851 values could be altered by storing in X. X is a register, a subreg, or
1852 a memory reference with nonvarying address (because, when a memory
1853 reference with a varying address is stored in, all memory references are
1854 removed by invalidate_memory so specific invalidation is superfluous).
1855 FULL_MODE, if not VOIDmode, indicates that this much should be
1856 invalidated instead of just the amount indicated by the mode of X. This
1857 is only used for bitfield stores into memory.
1859 A nonvarying address may be just a register or just a symbol reference,
1860 or it may be either of those plus a numeric offset. */
1862 static void
1863 invalidate (rtx x, machine_mode full_mode)
1865 int i;
1866 struct table_elt *p;
1867 rtx addr;
1869 switch (GET_CODE (x))
1871 case REG:
1873 /* If X is a register, dependencies on its contents are recorded
1874 through the qty number mechanism. Just change the qty number of
1875 the register, mark it as invalid for expressions that refer to it,
1876 and remove it itself. */
1877 unsigned int regno = REGNO (x);
1878 unsigned int hash = HASH (x, GET_MODE (x));
1880 /* Remove REGNO from any quantity list it might be on and indicate
1881 that its value might have changed. If it is a pseudo, remove its
1882 entry from the hash table.
1884 For a hard register, we do the first two actions above for any
1885 additional hard registers corresponding to X. Then, if any of these
1886 registers are in the table, we must remove any REG entries that
1887 overlap these registers. */
1889 delete_reg_equiv (regno);
1890 REG_TICK (regno)++;
1891 SUBREG_TICKED (regno) = -1;
1893 if (regno >= FIRST_PSEUDO_REGISTER)
1894 remove_pseudo_from_table (x, hash);
1895 else
1897 HOST_WIDE_INT in_table
1898 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1899 unsigned int endregno = END_HARD_REGNO (x);
1900 unsigned int tregno, tendregno, rn;
1901 struct table_elt *p, *next;
1903 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1905 for (rn = regno + 1; rn < endregno; rn++)
1907 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1908 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1909 delete_reg_equiv (rn);
1910 REG_TICK (rn)++;
1911 SUBREG_TICKED (rn) = -1;
1914 if (in_table)
1915 for (hash = 0; hash < HASH_SIZE; hash++)
1916 for (p = table[hash]; p; p = next)
1918 next = p->next_same_hash;
1920 if (!REG_P (p->exp)
1921 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1922 continue;
1924 tregno = REGNO (p->exp);
1925 tendregno = END_HARD_REGNO (p->exp);
1926 if (tendregno > regno && tregno < endregno)
1927 remove_from_table (p, hash);
1931 return;
1933 case SUBREG:
1934 invalidate (SUBREG_REG (x), VOIDmode);
1935 return;
1937 case PARALLEL:
1938 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1939 invalidate (XVECEXP (x, 0, i), VOIDmode);
1940 return;
1942 case EXPR_LIST:
1943 /* This is part of a disjoint return value; extract the location in
1944 question ignoring the offset. */
1945 invalidate (XEXP (x, 0), VOIDmode);
1946 return;
1948 case MEM:
1949 addr = canon_rtx (get_addr (XEXP (x, 0)));
1950 /* Calculate the canonical version of X here so that
1951 true_dependence doesn't generate new RTL for X on each call. */
1952 x = canon_rtx (x);
1954 /* Remove all hash table elements that refer to overlapping pieces of
1955 memory. */
1956 if (full_mode == VOIDmode)
1957 full_mode = GET_MODE (x);
1959 for (i = 0; i < HASH_SIZE; i++)
1961 struct table_elt *next;
1963 for (p = table[i]; p; p = next)
1965 next = p->next_same_hash;
1966 if (p->in_memory)
1968 /* Just canonicalize the expression once;
1969 otherwise each time we call invalidate
1970 true_dependence will canonicalize the
1971 expression again. */
1972 if (!p->canon_exp)
1973 p->canon_exp = canon_rtx (p->exp);
1974 if (check_dependence (p->canon_exp, x, full_mode, addr))
1975 remove_from_table (p, i);
1979 return;
1981 default:
1982 gcc_unreachable ();
1986 /* Remove all expressions that refer to register REGNO,
1987 since they are already invalid, and we are about to
1988 mark that register valid again and don't want the old
1989 expressions to reappear as valid. */
1991 static void
1992 remove_invalid_refs (unsigned int regno)
1994 unsigned int i;
1995 struct table_elt *p, *next;
1997 for (i = 0; i < HASH_SIZE; i++)
1998 for (p = table[i]; p; p = next)
2000 next = p->next_same_hash;
2001 if (!REG_P (p->exp) && refers_to_regno_p (regno, p->exp))
2002 remove_from_table (p, i);
2006 /* Likewise for a subreg with subreg_reg REGNO, subreg_byte OFFSET,
2007 and mode MODE. */
2008 static void
2009 remove_invalid_subreg_refs (unsigned int regno, unsigned int offset,
2010 machine_mode mode)
2012 unsigned int i;
2013 struct table_elt *p, *next;
2014 unsigned int end = offset + (GET_MODE_SIZE (mode) - 1);
2016 for (i = 0; i < HASH_SIZE; i++)
2017 for (p = table[i]; p; p = next)
2019 rtx exp = p->exp;
2020 next = p->next_same_hash;
2022 if (!REG_P (exp)
2023 && (GET_CODE (exp) != SUBREG
2024 || !REG_P (SUBREG_REG (exp))
2025 || REGNO (SUBREG_REG (exp)) != regno
2026 || (((SUBREG_BYTE (exp)
2027 + (GET_MODE_SIZE (GET_MODE (exp)) - 1)) >= offset)
2028 && SUBREG_BYTE (exp) <= end))
2029 && refers_to_regno_p (regno, p->exp))
2030 remove_from_table (p, i);
2034 /* Recompute the hash codes of any valid entries in the hash table that
2035 reference X, if X is a register, or SUBREG_REG (X) if X is a SUBREG.
2037 This is called when we make a jump equivalence. */
2039 static void
2040 rehash_using_reg (rtx x)
2042 unsigned int i;
2043 struct table_elt *p, *next;
2044 unsigned hash;
2046 if (GET_CODE (x) == SUBREG)
2047 x = SUBREG_REG (x);
2049 /* If X is not a register or if the register is known not to be in any
2050 valid entries in the table, we have no work to do. */
2052 if (!REG_P (x)
2053 || REG_IN_TABLE (REGNO (x)) < 0
2054 || REG_IN_TABLE (REGNO (x)) != REG_TICK (REGNO (x)))
2055 return;
2057 /* Scan all hash chains looking for valid entries that mention X.
2058 If we find one and it is in the wrong hash chain, move it. */
2060 for (i = 0; i < HASH_SIZE; i++)
2061 for (p = table[i]; p; p = next)
2063 next = p->next_same_hash;
2064 if (reg_mentioned_p (x, p->exp)
2065 && exp_equiv_p (p->exp, p->exp, 1, false)
2066 && i != (hash = SAFE_HASH (p->exp, p->mode)))
2068 if (p->next_same_hash)
2069 p->next_same_hash->prev_same_hash = p->prev_same_hash;
2071 if (p->prev_same_hash)
2072 p->prev_same_hash->next_same_hash = p->next_same_hash;
2073 else
2074 table[i] = p->next_same_hash;
2076 p->next_same_hash = table[hash];
2077 p->prev_same_hash = 0;
2078 if (table[hash])
2079 table[hash]->prev_same_hash = p;
2080 table[hash] = p;
2085 /* Remove from the hash table any expression that is a call-clobbered
2086 register. Also update their TICK values. */
2088 static void
2089 invalidate_for_call (void)
2091 unsigned int regno, endregno;
2092 unsigned int i;
2093 unsigned hash;
2094 struct table_elt *p, *next;
2095 int in_table = 0;
2096 hard_reg_set_iterator hrsi;
2098 /* Go through all the hard registers. For each that is clobbered in
2099 a CALL_INSN, remove the register from quantity chains and update
2100 reg_tick if defined. Also see if any of these registers is currently
2101 in the table. */
2102 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi)
2104 delete_reg_equiv (regno);
2105 if (REG_TICK (regno) >= 0)
2107 REG_TICK (regno)++;
2108 SUBREG_TICKED (regno) = -1;
2110 in_table |= (TEST_HARD_REG_BIT (hard_regs_in_table, regno) != 0);
2113 /* In the case where we have no call-clobbered hard registers in the
2114 table, we are done. Otherwise, scan the table and remove any
2115 entry that overlaps a call-clobbered register. */
2117 if (in_table)
2118 for (hash = 0; hash < HASH_SIZE; hash++)
2119 for (p = table[hash]; p; p = next)
2121 next = p->next_same_hash;
2123 if (!REG_P (p->exp)
2124 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
2125 continue;
2127 regno = REGNO (p->exp);
2128 endregno = END_HARD_REGNO (p->exp);
2130 for (i = regno; i < endregno; i++)
2131 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
2133 remove_from_table (p, hash);
2134 break;
2139 /* Given an expression X of type CONST,
2140 and ELT which is its table entry (or 0 if it
2141 is not in the hash table),
2142 return an alternate expression for X as a register plus integer.
2143 If none can be found, return 0. */
2145 static rtx
2146 use_related_value (rtx x, struct table_elt *elt)
2148 struct table_elt *relt = 0;
2149 struct table_elt *p, *q;
2150 HOST_WIDE_INT offset;
2152 /* First, is there anything related known?
2153 If we have a table element, we can tell from that.
2154 Otherwise, must look it up. */
2156 if (elt != 0 && elt->related_value != 0)
2157 relt = elt;
2158 else if (elt == 0 && GET_CODE (x) == CONST)
2160 rtx subexp = get_related_value (x);
2161 if (subexp != 0)
2162 relt = lookup (subexp,
2163 SAFE_HASH (subexp, GET_MODE (subexp)),
2164 GET_MODE (subexp));
2167 if (relt == 0)
2168 return 0;
2170 /* Search all related table entries for one that has an
2171 equivalent register. */
2173 p = relt;
2174 while (1)
2176 /* This loop is strange in that it is executed in two different cases.
2177 The first is when X is already in the table. Then it is searching
2178 the RELATED_VALUE list of X's class (RELT). The second case is when
2179 X is not in the table. Then RELT points to a class for the related
2180 value.
2182 Ensure that, whatever case we are in, that we ignore classes that have
2183 the same value as X. */
2185 if (rtx_equal_p (x, p->exp))
2186 q = 0;
2187 else
2188 for (q = p->first_same_value; q; q = q->next_same_value)
2189 if (REG_P (q->exp))
2190 break;
2192 if (q)
2193 break;
2195 p = p->related_value;
2197 /* We went all the way around, so there is nothing to be found.
2198 Alternatively, perhaps RELT was in the table for some other reason
2199 and it has no related values recorded. */
2200 if (p == relt || p == 0)
2201 break;
2204 if (q == 0)
2205 return 0;
2207 offset = (get_integer_term (x) - get_integer_term (p->exp));
2208 /* Note: OFFSET may be 0 if P->xexp and X are related by commutativity. */
2209 return plus_constant (q->mode, q->exp, offset);
2213 /* Hash a string. Just add its bytes up. */
2214 static inline unsigned
2215 hash_rtx_string (const char *ps)
2217 unsigned hash = 0;
2218 const unsigned char *p = (const unsigned char *) ps;
2220 if (p)
2221 while (*p)
2222 hash += *p++;
2224 return hash;
2227 /* Same as hash_rtx, but call CB on each rtx if it is not NULL.
2228 When the callback returns true, we continue with the new rtx. */
2230 unsigned
2231 hash_rtx_cb (const_rtx x, machine_mode mode,
2232 int *do_not_record_p, int *hash_arg_in_memory_p,
2233 bool have_reg_qty, hash_rtx_callback_function cb)
2235 int i, j;
2236 unsigned hash = 0;
2237 enum rtx_code code;
2238 const char *fmt;
2239 machine_mode newmode;
2240 rtx newx;
2242 /* Used to turn recursion into iteration. We can't rely on GCC's
2243 tail-recursion elimination since we need to keep accumulating values
2244 in HASH. */
2245 repeat:
2246 if (x == 0)
2247 return hash;
2249 /* Invoke the callback first. */
2250 if (cb != NULL
2251 && ((*cb) (x, mode, &newx, &newmode)))
2253 hash += hash_rtx_cb (newx, newmode, do_not_record_p,
2254 hash_arg_in_memory_p, have_reg_qty, cb);
2255 return hash;
2258 code = GET_CODE (x);
2259 switch (code)
2261 case REG:
2263 unsigned int regno = REGNO (x);
2265 if (do_not_record_p && !reload_completed)
2267 /* On some machines, we can't record any non-fixed hard register,
2268 because extending its life will cause reload problems. We
2269 consider ap, fp, sp, gp to be fixed for this purpose.
2271 We also consider CCmode registers to be fixed for this purpose;
2272 failure to do so leads to failure to simplify 0<100 type of
2273 conditionals.
2275 On all machines, we can't record any global registers.
2276 Nor should we record any register that is in a small
2277 class, as defined by TARGET_CLASS_LIKELY_SPILLED_P. */
2278 bool record;
2280 if (regno >= FIRST_PSEUDO_REGISTER)
2281 record = true;
2282 else if (x == frame_pointer_rtx
2283 || x == hard_frame_pointer_rtx
2284 || x == arg_pointer_rtx
2285 || x == stack_pointer_rtx
2286 || x == pic_offset_table_rtx)
2287 record = true;
2288 else if (global_regs[regno])
2289 record = false;
2290 else if (fixed_regs[regno])
2291 record = true;
2292 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
2293 record = true;
2294 else if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
2295 record = false;
2296 else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno)))
2297 record = false;
2298 else
2299 record = true;
2301 if (!record)
2303 *do_not_record_p = 1;
2304 return 0;
2308 hash += ((unsigned int) REG << 7);
2309 hash += (have_reg_qty ? (unsigned) REG_QTY (regno) : regno);
2310 return hash;
2313 /* We handle SUBREG of a REG specially because the underlying
2314 reg changes its hash value with every value change; we don't
2315 want to have to forget unrelated subregs when one subreg changes. */
2316 case SUBREG:
2318 if (REG_P (SUBREG_REG (x)))
2320 hash += (((unsigned int) SUBREG << 7)
2321 + REGNO (SUBREG_REG (x))
2322 + (SUBREG_BYTE (x) / UNITS_PER_WORD));
2323 return hash;
2325 break;
2328 case CONST_INT:
2329 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
2330 + (unsigned int) INTVAL (x));
2331 return hash;
2333 case CONST_WIDE_INT:
2334 for (i = 0; i < CONST_WIDE_INT_NUNITS (x); i++)
2335 hash += CONST_WIDE_INT_ELT (x, i);
2336 return hash;
2338 case CONST_DOUBLE:
2339 /* This is like the general case, except that it only counts
2340 the integers representing the constant. */
2341 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2342 if (TARGET_SUPPORTS_WIDE_INT == 0 && GET_MODE (x) == VOIDmode)
2343 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
2344 + (unsigned int) CONST_DOUBLE_HIGH (x));
2345 else
2346 hash += real_hash (CONST_DOUBLE_REAL_VALUE (x));
2347 return hash;
2349 case CONST_FIXED:
2350 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
2351 hash += fixed_hash (CONST_FIXED_VALUE (x));
2352 return hash;
2354 case CONST_VECTOR:
2356 int units;
2357 rtx elt;
2359 units = CONST_VECTOR_NUNITS (x);
2361 for (i = 0; i < units; ++i)
2363 elt = CONST_VECTOR_ELT (x, i);
2364 hash += hash_rtx_cb (elt, GET_MODE (elt),
2365 do_not_record_p, hash_arg_in_memory_p,
2366 have_reg_qty, cb);
2369 return hash;
2372 /* Assume there is only one rtx object for any given label. */
2373 case LABEL_REF:
2374 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
2375 differences and differences between each stage's debugging dumps. */
2376 hash += (((unsigned int) LABEL_REF << 7)
2377 + CODE_LABEL_NUMBER (LABEL_REF_LABEL (x)));
2378 return hash;
2380 case SYMBOL_REF:
2382 /* Don't hash on the symbol's address to avoid bootstrap differences.
2383 Different hash values may cause expressions to be recorded in
2384 different orders and thus different registers to be used in the
2385 final assembler. This also avoids differences in the dump files
2386 between various stages. */
2387 unsigned int h = 0;
2388 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
2390 while (*p)
2391 h += (h << 7) + *p++; /* ??? revisit */
2393 hash += ((unsigned int) SYMBOL_REF << 7) + h;
2394 return hash;
2397 case MEM:
2398 /* We don't record if marked volatile or if BLKmode since we don't
2399 know the size of the move. */
2400 if (do_not_record_p && (MEM_VOLATILE_P (x) || GET_MODE (x) == BLKmode))
2402 *do_not_record_p = 1;
2403 return 0;
2405 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2406 *hash_arg_in_memory_p = 1;
2408 /* Now that we have already found this special case,
2409 might as well speed it up as much as possible. */
2410 hash += (unsigned) MEM;
2411 x = XEXP (x, 0);
2412 goto repeat;
2414 case USE:
2415 /* A USE that mentions non-volatile memory needs special
2416 handling since the MEM may be BLKmode which normally
2417 prevents an entry from being made. Pure calls are
2418 marked by a USE which mentions BLKmode memory.
2419 See calls.c:emit_call_1. */
2420 if (MEM_P (XEXP (x, 0))
2421 && ! MEM_VOLATILE_P (XEXP (x, 0)))
2423 hash += (unsigned) USE;
2424 x = XEXP (x, 0);
2426 if (hash_arg_in_memory_p && !MEM_READONLY_P (x))
2427 *hash_arg_in_memory_p = 1;
2429 /* Now that we have already found this special case,
2430 might as well speed it up as much as possible. */
2431 hash += (unsigned) MEM;
2432 x = XEXP (x, 0);
2433 goto repeat;
2435 break;
2437 case PRE_DEC:
2438 case PRE_INC:
2439 case POST_DEC:
2440 case POST_INC:
2441 case PRE_MODIFY:
2442 case POST_MODIFY:
2443 case PC:
2444 case CC0:
2445 case CALL:
2446 case UNSPEC_VOLATILE:
2447 if (do_not_record_p) {
2448 *do_not_record_p = 1;
2449 return 0;
2451 else
2452 return hash;
2453 break;
2455 case ASM_OPERANDS:
2456 if (do_not_record_p && MEM_VOLATILE_P (x))
2458 *do_not_record_p = 1;
2459 return 0;
2461 else
2463 /* We don't want to take the filename and line into account. */
2464 hash += (unsigned) code + (unsigned) GET_MODE (x)
2465 + hash_rtx_string (ASM_OPERANDS_TEMPLATE (x))
2466 + hash_rtx_string (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
2467 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
2469 if (ASM_OPERANDS_INPUT_LENGTH (x))
2471 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
2473 hash += (hash_rtx_cb (ASM_OPERANDS_INPUT (x, i),
2474 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
2475 do_not_record_p, hash_arg_in_memory_p,
2476 have_reg_qty, cb)
2477 + hash_rtx_string
2478 (ASM_OPERANDS_INPUT_CONSTRAINT (x, i)));
2481 hash += hash_rtx_string (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
2482 x = ASM_OPERANDS_INPUT (x, 0);
2483 mode = GET_MODE (x);
2484 goto repeat;
2487 return hash;
2489 break;
2491 default:
2492 break;
2495 i = GET_RTX_LENGTH (code) - 1;
2496 hash += (unsigned) code + (unsigned) GET_MODE (x);
2497 fmt = GET_RTX_FORMAT (code);
2498 for (; i >= 0; i--)
2500 switch (fmt[i])
2502 case 'e':
2503 /* If we are about to do the last recursive call
2504 needed at this level, change it into iteration.
2505 This function is called enough to be worth it. */
2506 if (i == 0)
2508 x = XEXP (x, i);
2509 goto repeat;
2512 hash += hash_rtx_cb (XEXP (x, i), VOIDmode, do_not_record_p,
2513 hash_arg_in_memory_p,
2514 have_reg_qty, cb);
2515 break;
2517 case 'E':
2518 for (j = 0; j < XVECLEN (x, i); j++)
2519 hash += hash_rtx_cb (XVECEXP (x, i, j), VOIDmode, do_not_record_p,
2520 hash_arg_in_memory_p,
2521 have_reg_qty, cb);
2522 break;
2524 case 's':
2525 hash += hash_rtx_string (XSTR (x, i));
2526 break;
2528 case 'i':
2529 hash += (unsigned int) XINT (x, i);
2530 break;
2532 case '0': case 't':
2533 /* Unused. */
2534 break;
2536 default:
2537 gcc_unreachable ();
2541 return hash;
2544 /* Hash an rtx. We are careful to make sure the value is never negative.
2545 Equivalent registers hash identically.
2546 MODE is used in hashing for CONST_INTs only;
2547 otherwise the mode of X is used.
2549 Store 1 in DO_NOT_RECORD_P if any subexpression is volatile.
2551 If HASH_ARG_IN_MEMORY_P is not NULL, store 1 in it if X contains
2552 a MEM rtx which does not have the MEM_READONLY_P flag set.
2554 Note that cse_insn knows that the hash code of a MEM expression
2555 is just (int) MEM plus the hash code of the address. */
2557 unsigned
2558 hash_rtx (const_rtx x, machine_mode mode, int *do_not_record_p,
2559 int *hash_arg_in_memory_p, bool have_reg_qty)
2561 return hash_rtx_cb (x, mode, do_not_record_p,
2562 hash_arg_in_memory_p, have_reg_qty, NULL);
2565 /* Hash an rtx X for cse via hash_rtx.
2566 Stores 1 in do_not_record if any subexpression is volatile.
2567 Stores 1 in hash_arg_in_memory if X contains a mem rtx which
2568 does not have the MEM_READONLY_P flag set. */
2570 static inline unsigned
2571 canon_hash (rtx x, machine_mode mode)
2573 return hash_rtx (x, mode, &do_not_record, &hash_arg_in_memory, true);
2576 /* Like canon_hash but with no side effects, i.e. do_not_record
2577 and hash_arg_in_memory are not changed. */
2579 static inline unsigned
2580 safe_hash (rtx x, machine_mode mode)
2582 int dummy_do_not_record;
2583 return hash_rtx (x, mode, &dummy_do_not_record, NULL, true);
2586 /* Return 1 iff X and Y would canonicalize into the same thing,
2587 without actually constructing the canonicalization of either one.
2588 If VALIDATE is nonzero,
2589 we assume X is an expression being processed from the rtl
2590 and Y was found in the hash table. We check register refs
2591 in Y for being marked as valid.
2593 If FOR_GCSE is true, we compare X and Y for equivalence for GCSE. */
2596 exp_equiv_p (const_rtx x, const_rtx y, int validate, bool for_gcse)
2598 int i, j;
2599 enum rtx_code code;
2600 const char *fmt;
2602 /* Note: it is incorrect to assume an expression is equivalent to itself
2603 if VALIDATE is nonzero. */
2604 if (x == y && !validate)
2605 return 1;
2607 if (x == 0 || y == 0)
2608 return x == y;
2610 code = GET_CODE (x);
2611 if (code != GET_CODE (y))
2612 return 0;
2614 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
2615 if (GET_MODE (x) != GET_MODE (y))
2616 return 0;
2618 /* MEMs referring to different address space are not equivalent. */
2619 if (code == MEM && MEM_ADDR_SPACE (x) != MEM_ADDR_SPACE (y))
2620 return 0;
2622 switch (code)
2624 case PC:
2625 case CC0:
2626 CASE_CONST_UNIQUE:
2627 return x == y;
2629 case LABEL_REF:
2630 return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y);
2632 case SYMBOL_REF:
2633 return XSTR (x, 0) == XSTR (y, 0);
2635 case REG:
2636 if (for_gcse)
2637 return REGNO (x) == REGNO (y);
2638 else
2640 unsigned int regno = REGNO (y);
2641 unsigned int i;
2642 unsigned int endregno = END_REGNO (y);
2644 /* If the quantities are not the same, the expressions are not
2645 equivalent. If there are and we are not to validate, they
2646 are equivalent. Otherwise, ensure all regs are up-to-date. */
2648 if (REG_QTY (REGNO (x)) != REG_QTY (regno))
2649 return 0;
2651 if (! validate)
2652 return 1;
2654 for (i = regno; i < endregno; i++)
2655 if (REG_IN_TABLE (i) != REG_TICK (i))
2656 return 0;
2658 return 1;
2661 case MEM:
2662 if (for_gcse)
2664 /* A volatile mem should not be considered equivalent to any
2665 other. */
2666 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2667 return 0;
2669 /* Can't merge two expressions in different alias sets, since we
2670 can decide that the expression is transparent in a block when
2671 it isn't, due to it being set with the different alias set.
2673 Also, can't merge two expressions with different MEM_ATTRS.
2674 They could e.g. be two different entities allocated into the
2675 same space on the stack (see e.g. PR25130). In that case, the
2676 MEM addresses can be the same, even though the two MEMs are
2677 absolutely not equivalent.
2679 But because really all MEM attributes should be the same for
2680 equivalent MEMs, we just use the invariant that MEMs that have
2681 the same attributes share the same mem_attrs data structure. */
2682 if (!mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
2683 return 0;
2685 /* If we are handling exceptions, we cannot consider two expressions
2686 with different trapping status as equivalent, because simple_mem
2687 might accept one and reject the other. */
2688 if (cfun->can_throw_non_call_exceptions
2689 && (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y)))
2690 return 0;
2692 break;
2694 /* For commutative operations, check both orders. */
2695 case PLUS:
2696 case MULT:
2697 case AND:
2698 case IOR:
2699 case XOR:
2700 case NE:
2701 case EQ:
2702 return ((exp_equiv_p (XEXP (x, 0), XEXP (y, 0),
2703 validate, for_gcse)
2704 && exp_equiv_p (XEXP (x, 1), XEXP (y, 1),
2705 validate, for_gcse))
2706 || (exp_equiv_p (XEXP (x, 0), XEXP (y, 1),
2707 validate, for_gcse)
2708 && exp_equiv_p (XEXP (x, 1), XEXP (y, 0),
2709 validate, for_gcse)));
2711 case ASM_OPERANDS:
2712 /* We don't use the generic code below because we want to
2713 disregard filename and line numbers. */
2715 /* A volatile asm isn't equivalent to any other. */
2716 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
2717 return 0;
2719 if (GET_MODE (x) != GET_MODE (y)
2720 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
2721 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
2722 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
2723 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
2724 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
2725 return 0;
2727 if (ASM_OPERANDS_INPUT_LENGTH (x))
2729 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
2730 if (! exp_equiv_p (ASM_OPERANDS_INPUT (x, i),
2731 ASM_OPERANDS_INPUT (y, i),
2732 validate, for_gcse)
2733 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
2734 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
2735 return 0;
2738 return 1;
2740 default:
2741 break;
2744 /* Compare the elements. If any pair of corresponding elements
2745 fail to match, return 0 for the whole thing. */
2747 fmt = GET_RTX_FORMAT (code);
2748 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2750 switch (fmt[i])
2752 case 'e':
2753 if (! exp_equiv_p (XEXP (x, i), XEXP (y, i),
2754 validate, for_gcse))
2755 return 0;
2756 break;
2758 case 'E':
2759 if (XVECLEN (x, i) != XVECLEN (y, i))
2760 return 0;
2761 for (j = 0; j < XVECLEN (x, i); j++)
2762 if (! exp_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2763 validate, for_gcse))
2764 return 0;
2765 break;
2767 case 's':
2768 if (strcmp (XSTR (x, i), XSTR (y, i)))
2769 return 0;
2770 break;
2772 case 'i':
2773 if (XINT (x, i) != XINT (y, i))
2774 return 0;
2775 break;
2777 case 'w':
2778 if (XWINT (x, i) != XWINT (y, i))
2779 return 0;
2780 break;
2782 case '0':
2783 case 't':
2784 break;
2786 default:
2787 gcc_unreachable ();
2791 return 1;
2794 /* Subroutine of canon_reg. Pass *XLOC through canon_reg, and validate
2795 the result if necessary. INSN is as for canon_reg. */
2797 static void
2798 validate_canon_reg (rtx *xloc, rtx_insn *insn)
2800 if (*xloc)
2802 rtx new_rtx = canon_reg (*xloc, insn);
2804 /* If replacing pseudo with hard reg or vice versa, ensure the
2805 insn remains valid. Likewise if the insn has MATCH_DUPs. */
2806 gcc_assert (insn && new_rtx);
2807 validate_change (insn, xloc, new_rtx, 1);
2811 /* Canonicalize an expression:
2812 replace each register reference inside it
2813 with the "oldest" equivalent register.
2815 If INSN is nonzero validate_change is used to ensure that INSN remains valid
2816 after we make our substitution. The calls are made with IN_GROUP nonzero
2817 so apply_change_group must be called upon the outermost return from this
2818 function (unless INSN is zero). The result of apply_change_group can
2819 generally be discarded since the changes we are making are optional. */
2821 static rtx
2822 canon_reg (rtx x, rtx_insn *insn)
2824 int i;
2825 enum rtx_code code;
2826 const char *fmt;
2828 if (x == 0)
2829 return x;
2831 code = GET_CODE (x);
2832 switch (code)
2834 case PC:
2835 case CC0:
2836 case CONST:
2837 CASE_CONST_ANY:
2838 case SYMBOL_REF:
2839 case LABEL_REF:
2840 case ADDR_VEC:
2841 case ADDR_DIFF_VEC:
2842 return x;
2844 case REG:
2846 int first;
2847 int q;
2848 struct qty_table_elem *ent;
2850 /* Never replace a hard reg, because hard regs can appear
2851 in more than one machine mode, and we must preserve the mode
2852 of each occurrence. Also, some hard regs appear in
2853 MEMs that are shared and mustn't be altered. Don't try to
2854 replace any reg that maps to a reg of class NO_REGS. */
2855 if (REGNO (x) < FIRST_PSEUDO_REGISTER
2856 || ! REGNO_QTY_VALID_P (REGNO (x)))
2857 return x;
2859 q = REG_QTY (REGNO (x));
2860 ent = &qty_table[q];
2861 first = ent->first_reg;
2862 return (first >= FIRST_PSEUDO_REGISTER ? regno_reg_rtx[first]
2863 : REGNO_REG_CLASS (first) == NO_REGS ? x
2864 : gen_rtx_REG (ent->mode, first));
2867 default:
2868 break;
2871 fmt = GET_RTX_FORMAT (code);
2872 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2874 int j;
2876 if (fmt[i] == 'e')
2877 validate_canon_reg (&XEXP (x, i), insn);
2878 else if (fmt[i] == 'E')
2879 for (j = 0; j < XVECLEN (x, i); j++)
2880 validate_canon_reg (&XVECEXP (x, i, j), insn);
2883 return x;
2886 /* Given an operation (CODE, *PARG1, *PARG2), where code is a comparison
2887 operation (EQ, NE, GT, etc.), follow it back through the hash table and
2888 what values are being compared.
2890 *PARG1 and *PARG2 are updated to contain the rtx representing the values
2891 actually being compared. For example, if *PARG1 was (cc0) and *PARG2
2892 was (const_int 0), *PARG1 and *PARG2 will be set to the objects that were
2893 compared to produce cc0.
2895 The return value is the comparison operator and is either the code of
2896 A or the code corresponding to the inverse of the comparison. */
2898 static enum rtx_code
2899 find_comparison_args (enum rtx_code code, rtx *parg1, rtx *parg2,
2900 machine_mode *pmode1, machine_mode *pmode2)
2902 rtx arg1, arg2;
2903 hash_set<rtx> *visited = NULL;
2904 /* Set nonzero when we find something of interest. */
2905 rtx x = NULL;
2907 arg1 = *parg1, arg2 = *parg2;
2909 /* If ARG2 is const0_rtx, see what ARG1 is equivalent to. */
2911 while (arg2 == CONST0_RTX (GET_MODE (arg1)))
2913 int reverse_code = 0;
2914 struct table_elt *p = 0;
2916 /* Remember state from previous iteration. */
2917 if (x)
2919 if (!visited)
2920 visited = new hash_set<rtx>;
2921 visited->add (x);
2922 x = 0;
2925 /* If arg1 is a COMPARE, extract the comparison arguments from it.
2926 On machines with CC0, this is the only case that can occur, since
2927 fold_rtx will return the COMPARE or item being compared with zero
2928 when given CC0. */
2930 if (GET_CODE (arg1) == COMPARE && arg2 == const0_rtx)
2931 x = arg1;
2933 /* If ARG1 is a comparison operator and CODE is testing for
2934 STORE_FLAG_VALUE, get the inner arguments. */
2936 else if (COMPARISON_P (arg1))
2938 #ifdef FLOAT_STORE_FLAG_VALUE
2939 REAL_VALUE_TYPE fsfv;
2940 #endif
2942 if (code == NE
2943 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2944 && code == LT && 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
2951 x = arg1;
2952 else if (code == EQ
2953 || (GET_MODE_CLASS (GET_MODE (arg1)) == MODE_INT
2954 && code == GE && STORE_FLAG_VALUE == -1)
2955 #ifdef FLOAT_STORE_FLAG_VALUE
2956 || (SCALAR_FLOAT_MODE_P (GET_MODE (arg1))
2957 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
2958 REAL_VALUE_NEGATIVE (fsfv)))
2959 #endif
2961 x = arg1, reverse_code = 1;
2964 /* ??? We could also check for
2966 (ne (and (eq (...) (const_int 1))) (const_int 0))
2968 and related forms, but let's wait until we see them occurring. */
2970 if (x == 0)
2971 /* Look up ARG1 in the hash table and see if it has an equivalence
2972 that lets us see what is being compared. */
2973 p = lookup (arg1, SAFE_HASH (arg1, GET_MODE (arg1)), GET_MODE (arg1));
2974 if (p)
2976 p = p->first_same_value;
2978 /* If what we compare is already known to be constant, that is as
2979 good as it gets.
2980 We need to break the loop in this case, because otherwise we
2981 can have an infinite loop when looking at a reg that is known
2982 to be a constant which is the same as a comparison of a reg
2983 against zero which appears later in the insn stream, which in
2984 turn is constant and the same as the comparison of the first reg
2985 against zero... */
2986 if (p->is_const)
2987 break;
2990 for (; p; p = p->next_same_value)
2992 machine_mode inner_mode = GET_MODE (p->exp);
2993 #ifdef FLOAT_STORE_FLAG_VALUE
2994 REAL_VALUE_TYPE fsfv;
2995 #endif
2997 /* If the entry isn't valid, skip it. */
2998 if (! exp_equiv_p (p->exp, p->exp, 1, false))
2999 continue;
3001 /* If it's a comparison we've used before, skip it. */
3002 if (visited && visited->contains (p->exp))
3003 continue;
3005 if (GET_CODE (p->exp) == COMPARE
3006 /* Another possibility is that this machine has a compare insn
3007 that includes the comparison code. In that case, ARG1 would
3008 be equivalent to a comparison operation that would set ARG1 to
3009 either STORE_FLAG_VALUE or zero. If this is an NE operation,
3010 ORIG_CODE is the actual comparison being done; if it is an EQ,
3011 we must reverse ORIG_CODE. On machine with a negative value
3012 for STORE_FLAG_VALUE, also look at LT and GE operations. */
3013 || ((code == NE
3014 || (code == LT
3015 && val_signbit_known_set_p (inner_mode,
3016 STORE_FLAG_VALUE))
3017 #ifdef FLOAT_STORE_FLAG_VALUE
3018 || (code == LT
3019 && SCALAR_FLOAT_MODE_P (inner_mode)
3020 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3021 REAL_VALUE_NEGATIVE (fsfv)))
3022 #endif
3024 && COMPARISON_P (p->exp)))
3026 x = p->exp;
3027 break;
3029 else if ((code == EQ
3030 || (code == GE
3031 && val_signbit_known_set_p (inner_mode,
3032 STORE_FLAG_VALUE))
3033 #ifdef FLOAT_STORE_FLAG_VALUE
3034 || (code == GE
3035 && SCALAR_FLOAT_MODE_P (inner_mode)
3036 && (fsfv = FLOAT_STORE_FLAG_VALUE (GET_MODE (arg1)),
3037 REAL_VALUE_NEGATIVE (fsfv)))
3038 #endif
3040 && COMPARISON_P (p->exp))
3042 reverse_code = 1;
3043 x = p->exp;
3044 break;
3047 /* If this non-trapping address, e.g. fp + constant, the
3048 equivalent is a better operand since it may let us predict
3049 the value of the comparison. */
3050 else if (!rtx_addr_can_trap_p (p->exp))
3052 arg1 = p->exp;
3053 continue;
3057 /* If we didn't find a useful equivalence for ARG1, we are done.
3058 Otherwise, set up for the next iteration. */
3059 if (x == 0)
3060 break;
3062 /* If we need to reverse the comparison, make sure that that is
3063 possible -- we can't necessarily infer the value of GE from LT
3064 with floating-point operands. */
3065 if (reverse_code)
3067 enum rtx_code reversed = reversed_comparison_code (x, NULL_RTX);
3068 if (reversed == UNKNOWN)
3069 break;
3070 else
3071 code = reversed;
3073 else if (COMPARISON_P (x))
3074 code = GET_CODE (x);
3075 arg1 = XEXP (x, 0), arg2 = XEXP (x, 1);
3078 /* Return our results. Return the modes from before fold_rtx
3079 because fold_rtx might produce const_int, and then it's too late. */
3080 *pmode1 = GET_MODE (arg1), *pmode2 = GET_MODE (arg2);
3081 *parg1 = fold_rtx (arg1, 0), *parg2 = fold_rtx (arg2, 0);
3083 if (visited)
3084 delete visited;
3085 return code;
3088 /* If X is a nontrivial arithmetic operation on an argument for which
3089 a constant value can be determined, return the result of operating
3090 on that value, as a constant. Otherwise, return X, possibly with
3091 one or more operands changed to a forward-propagated constant.
3093 If X is a register whose contents are known, we do NOT return
3094 those contents here; equiv_constant is called to perform that task.
3095 For SUBREGs and MEMs, we do that both here and in equiv_constant.
3097 INSN is the insn that we may be modifying. If it is 0, make a copy
3098 of X before modifying it. */
3100 static rtx
3101 fold_rtx (rtx x, rtx_insn *insn)
3103 enum rtx_code code;
3104 machine_mode mode;
3105 const char *fmt;
3106 int i;
3107 rtx new_rtx = 0;
3108 int changed = 0;
3110 /* Operands of X. */
3111 /* Workaround -Wmaybe-uninitialized false positive during
3112 profiledbootstrap by initializing them. */
3113 rtx folded_arg0 = NULL_RTX;
3114 rtx folded_arg1 = NULL_RTX;
3116 /* Constant equivalents of first three operands of X;
3117 0 when no such equivalent is known. */
3118 rtx const_arg0;
3119 rtx const_arg1;
3120 rtx const_arg2;
3122 /* The mode of the first operand of X. We need this for sign and zero
3123 extends. */
3124 machine_mode mode_arg0;
3126 if (x == 0)
3127 return x;
3129 /* Try to perform some initial simplifications on X. */
3130 code = GET_CODE (x);
3131 switch (code)
3133 case MEM:
3134 case SUBREG:
3135 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3136 return new_rtx;
3137 return x;
3139 case CONST:
3140 CASE_CONST_ANY:
3141 case SYMBOL_REF:
3142 case LABEL_REF:
3143 case REG:
3144 case PC:
3145 /* No use simplifying an EXPR_LIST
3146 since they are used only for lists of args
3147 in a function call's REG_EQUAL note. */
3148 case EXPR_LIST:
3149 return x;
3151 #ifdef HAVE_cc0
3152 case CC0:
3153 return prev_insn_cc0;
3154 #endif
3156 case ASM_OPERANDS:
3157 if (insn)
3159 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3160 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3161 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3163 return x;
3165 #ifdef NO_FUNCTION_CSE
3166 case CALL:
3167 if (CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3168 return x;
3169 break;
3170 #endif
3172 /* Anything else goes through the loop below. */
3173 default:
3174 break;
3177 mode = GET_MODE (x);
3178 const_arg0 = 0;
3179 const_arg1 = 0;
3180 const_arg2 = 0;
3181 mode_arg0 = VOIDmode;
3183 /* Try folding our operands.
3184 Then see which ones have constant values known. */
3186 fmt = GET_RTX_FORMAT (code);
3187 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3188 if (fmt[i] == 'e')
3190 rtx folded_arg = XEXP (x, i), const_arg;
3191 machine_mode mode_arg = GET_MODE (folded_arg);
3193 switch (GET_CODE (folded_arg))
3195 case MEM:
3196 case REG:
3197 case SUBREG:
3198 const_arg = equiv_constant (folded_arg);
3199 break;
3201 case CONST:
3202 CASE_CONST_ANY:
3203 case SYMBOL_REF:
3204 case LABEL_REF:
3205 const_arg = folded_arg;
3206 break;
3208 #ifdef HAVE_cc0
3209 case CC0:
3210 /* The cc0-user and cc0-setter may be in different blocks if
3211 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3212 will have been cleared as we exited the block with the
3213 setter.
3215 While we could potentially track cc0 in this case, it just
3216 doesn't seem to be worth it given that cc0 targets are not
3217 terribly common or important these days and trapping math
3218 is rarely used. The combination of those two conditions
3219 necessary to trip this situation is exceedingly rare in the
3220 real world. */
3221 if (!prev_insn_cc0)
3223 const_arg = NULL_RTX;
3225 else
3227 folded_arg = prev_insn_cc0;
3228 mode_arg = prev_insn_cc0_mode;
3229 const_arg = equiv_constant (folded_arg);
3231 break;
3232 #endif
3234 default:
3235 folded_arg = fold_rtx (folded_arg, insn);
3236 const_arg = equiv_constant (folded_arg);
3237 break;
3240 /* For the first three operands, see if the operand
3241 is constant or equivalent to a constant. */
3242 switch (i)
3244 case 0:
3245 folded_arg0 = folded_arg;
3246 const_arg0 = const_arg;
3247 mode_arg0 = mode_arg;
3248 break;
3249 case 1:
3250 folded_arg1 = folded_arg;
3251 const_arg1 = const_arg;
3252 break;
3253 case 2:
3254 const_arg2 = const_arg;
3255 break;
3258 /* Pick the least expensive of the argument and an equivalent constant
3259 argument. */
3260 if (const_arg != 0
3261 && const_arg != folded_arg
3262 && COST_IN (const_arg, code, i) <= COST_IN (folded_arg, code, i)
3264 /* It's not safe to substitute the operand of a conversion
3265 operator with a constant, as the conversion's identity
3266 depends upon the mode of its operand. This optimization
3267 is handled by the call to simplify_unary_operation. */
3268 && (GET_RTX_CLASS (code) != RTX_UNARY
3269 || GET_MODE (const_arg) == mode_arg0
3270 || (code != ZERO_EXTEND
3271 && code != SIGN_EXTEND
3272 && code != TRUNCATE
3273 && code != FLOAT_TRUNCATE
3274 && code != FLOAT_EXTEND
3275 && code != FLOAT
3276 && code != FIX
3277 && code != UNSIGNED_FLOAT
3278 && code != UNSIGNED_FIX)))
3279 folded_arg = const_arg;
3281 if (folded_arg == XEXP (x, i))
3282 continue;
3284 if (insn == NULL_RTX && !changed)
3285 x = copy_rtx (x);
3286 changed = 1;
3287 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3290 if (changed)
3292 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3293 consistent with the order in X. */
3294 if (canonicalize_change_group (insn, x))
3296 rtx tem;
3297 tem = const_arg0, const_arg0 = const_arg1, const_arg1 = tem;
3298 tem = folded_arg0, folded_arg0 = folded_arg1, folded_arg1 = tem;
3301 apply_change_group ();
3304 /* If X is an arithmetic operation, see if we can simplify it. */
3306 switch (GET_RTX_CLASS (code))
3308 case RTX_UNARY:
3310 /* We can't simplify extension ops unless we know the
3311 original mode. */
3312 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3313 && mode_arg0 == VOIDmode)
3314 break;
3316 new_rtx = simplify_unary_operation (code, mode,
3317 const_arg0 ? const_arg0 : folded_arg0,
3318 mode_arg0);
3320 break;
3322 case RTX_COMPARE:
3323 case RTX_COMM_COMPARE:
3324 /* See what items are actually being compared and set FOLDED_ARG[01]
3325 to those values and CODE to the actual comparison code. If any are
3326 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3327 do anything if both operands are already known to be constant. */
3329 /* ??? Vector mode comparisons are not supported yet. */
3330 if (VECTOR_MODE_P (mode))
3331 break;
3333 if (const_arg0 == 0 || const_arg1 == 0)
3335 struct table_elt *p0, *p1;
3336 rtx true_rtx, false_rtx;
3337 machine_mode mode_arg1;
3339 if (SCALAR_FLOAT_MODE_P (mode))
3341 #ifdef FLOAT_STORE_FLAG_VALUE
3342 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3343 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3344 #else
3345 true_rtx = NULL_RTX;
3346 #endif
3347 false_rtx = CONST0_RTX (mode);
3349 else
3351 true_rtx = const_true_rtx;
3352 false_rtx = const0_rtx;
3355 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3356 &mode_arg0, &mode_arg1);
3358 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3359 what kinds of things are being compared, so we can't do
3360 anything with this comparison. */
3362 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3363 break;
3365 const_arg0 = equiv_constant (folded_arg0);
3366 const_arg1 = equiv_constant (folded_arg1);
3368 /* If we do not now have two constants being compared, see
3369 if we can nevertheless deduce some things about the
3370 comparison. */
3371 if (const_arg0 == 0 || const_arg1 == 0)
3373 if (const_arg1 != NULL)
3375 rtx cheapest_simplification;
3376 int cheapest_cost;
3377 rtx simp_result;
3378 struct table_elt *p;
3380 /* See if we can find an equivalent of folded_arg0
3381 that gets us a cheaper expression, possibly a
3382 constant through simplifications. */
3383 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3384 mode_arg0);
3386 if (p != NULL)
3388 cheapest_simplification = x;
3389 cheapest_cost = COST (x);
3391 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3393 int cost;
3395 /* If the entry isn't valid, skip it. */
3396 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3397 continue;
3399 /* Try to simplify using this equivalence. */
3400 simp_result
3401 = simplify_relational_operation (code, mode,
3402 mode_arg0,
3403 p->exp,
3404 const_arg1);
3406 if (simp_result == NULL)
3407 continue;
3409 cost = COST (simp_result);
3410 if (cost < cheapest_cost)
3412 cheapest_cost = cost;
3413 cheapest_simplification = simp_result;
3417 /* If we have a cheaper expression now, use that
3418 and try folding it further, from the top. */
3419 if (cheapest_simplification != x)
3420 return fold_rtx (copy_rtx (cheapest_simplification),
3421 insn);
3425 /* See if the two operands are the same. */
3427 if ((REG_P (folded_arg0)
3428 && REG_P (folded_arg1)
3429 && (REG_QTY (REGNO (folded_arg0))
3430 == REG_QTY (REGNO (folded_arg1))))
3431 || ((p0 = lookup (folded_arg0,
3432 SAFE_HASH (folded_arg0, mode_arg0),
3433 mode_arg0))
3434 && (p1 = lookup (folded_arg1,
3435 SAFE_HASH (folded_arg1, mode_arg0),
3436 mode_arg0))
3437 && p0->first_same_value == p1->first_same_value))
3438 folded_arg1 = folded_arg0;
3440 /* If FOLDED_ARG0 is a register, see if the comparison we are
3441 doing now is either the same as we did before or the reverse
3442 (we only check the reverse if not floating-point). */
3443 else if (REG_P (folded_arg0))
3445 int qty = REG_QTY (REGNO (folded_arg0));
3447 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3449 struct qty_table_elem *ent = &qty_table[qty];
3451 if ((comparison_dominates_p (ent->comparison_code, code)
3452 || (! FLOAT_MODE_P (mode_arg0)
3453 && comparison_dominates_p (ent->comparison_code,
3454 reverse_condition (code))))
3455 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3456 || (const_arg1
3457 && rtx_equal_p (ent->comparison_const,
3458 const_arg1))
3459 || (REG_P (folded_arg1)
3460 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3462 if (comparison_dominates_p (ent->comparison_code, code))
3464 if (true_rtx)
3465 return true_rtx;
3466 else
3467 break;
3469 else
3470 return false_rtx;
3477 /* If we are comparing against zero, see if the first operand is
3478 equivalent to an IOR with a constant. If so, we may be able to
3479 determine the result of this comparison. */
3480 if (const_arg1 == const0_rtx && !const_arg0)
3482 rtx y = lookup_as_function (folded_arg0, IOR);
3483 rtx inner_const;
3485 if (y != 0
3486 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3487 && CONST_INT_P (inner_const)
3488 && INTVAL (inner_const) != 0)
3489 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3493 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3494 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3495 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3496 op0, op1);
3498 break;
3500 case RTX_BIN_ARITH:
3501 case RTX_COMM_ARITH:
3502 switch (code)
3504 case PLUS:
3505 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3506 with that LABEL_REF as its second operand. If so, the result is
3507 the first operand of that MINUS. This handles switches with an
3508 ADDR_DIFF_VEC table. */
3509 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3511 rtx y
3512 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3513 : lookup_as_function (folded_arg0, MINUS);
3515 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3516 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg1))
3517 return XEXP (y, 0);
3519 /* Now try for a CONST of a MINUS like the above. */
3520 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3521 : lookup_as_function (folded_arg0, CONST))) != 0
3522 && GET_CODE (XEXP (y, 0)) == MINUS
3523 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3524 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg1))
3525 return XEXP (XEXP (y, 0), 0);
3528 /* Likewise if the operands are in the other order. */
3529 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3531 rtx y
3532 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3533 : lookup_as_function (folded_arg1, MINUS);
3535 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3536 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg0))
3537 return XEXP (y, 0);
3539 /* Now try for a CONST of a MINUS like the above. */
3540 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3541 : lookup_as_function (folded_arg1, CONST))) != 0
3542 && GET_CODE (XEXP (y, 0)) == MINUS
3543 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3544 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg0))
3545 return XEXP (XEXP (y, 0), 0);
3548 /* If second operand is a register equivalent to a negative
3549 CONST_INT, see if we can find a register equivalent to the
3550 positive constant. Make a MINUS if so. Don't do this for
3551 a non-negative constant since we might then alternate between
3552 choosing positive and negative constants. Having the positive
3553 constant previously-used is the more common case. Be sure
3554 the resulting constant is non-negative; if const_arg1 were
3555 the smallest negative number this would overflow: depending
3556 on the mode, this would either just be the same value (and
3557 hence not save anything) or be incorrect. */
3558 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3559 && INTVAL (const_arg1) < 0
3560 /* This used to test
3562 -INTVAL (const_arg1) >= 0
3564 But The Sun V5.0 compilers mis-compiled that test. So
3565 instead we test for the problematic value in a more direct
3566 manner and hope the Sun compilers get it correct. */
3567 && INTVAL (const_arg1) !=
3568 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3569 && REG_P (folded_arg1))
3571 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3572 struct table_elt *p
3573 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3575 if (p)
3576 for (p = p->first_same_value; p; p = p->next_same_value)
3577 if (REG_P (p->exp))
3578 return simplify_gen_binary (MINUS, mode, folded_arg0,
3579 canon_reg (p->exp, NULL));
3581 goto from_plus;
3583 case MINUS:
3584 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3585 If so, produce (PLUS Z C2-C). */
3586 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3588 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3589 if (y && CONST_INT_P (XEXP (y, 1)))
3590 return fold_rtx (plus_constant (mode, copy_rtx (y),
3591 -INTVAL (const_arg1)),
3592 NULL);
3595 /* Fall through. */
3597 from_plus:
3598 case SMIN: case SMAX: case UMIN: case UMAX:
3599 case IOR: case AND: case XOR:
3600 case MULT:
3601 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3602 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3603 is known to be of similar form, we may be able to replace the
3604 operation with a combined operation. This may eliminate the
3605 intermediate operation if every use is simplified in this way.
3606 Note that the similar optimization done by combine.c only works
3607 if the intermediate operation's result has only one reference. */
3609 if (REG_P (folded_arg0)
3610 && const_arg1 && CONST_INT_P (const_arg1))
3612 int is_shift
3613 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3614 rtx y, inner_const, new_const;
3615 rtx canon_const_arg1 = const_arg1;
3616 enum rtx_code associate_code;
3618 if (is_shift
3619 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3620 || INTVAL (const_arg1) < 0))
3622 if (SHIFT_COUNT_TRUNCATED)
3623 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3624 & (GET_MODE_BITSIZE (mode)
3625 - 1));
3626 else
3627 break;
3630 y = lookup_as_function (folded_arg0, code);
3631 if (y == 0)
3632 break;
3634 /* If we have compiled a statement like
3635 "if (x == (x & mask1))", and now are looking at
3636 "x & mask2", we will have a case where the first operand
3637 of Y is the same as our first operand. Unless we detect
3638 this case, an infinite loop will result. */
3639 if (XEXP (y, 0) == folded_arg0)
3640 break;
3642 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3643 if (!inner_const || !CONST_INT_P (inner_const))
3644 break;
3646 /* Don't associate these operations if they are a PLUS with the
3647 same constant and it is a power of two. These might be doable
3648 with a pre- or post-increment. Similarly for two subtracts of
3649 identical powers of two with post decrement. */
3651 if (code == PLUS && const_arg1 == inner_const
3652 && ((HAVE_PRE_INCREMENT
3653 && exact_log2 (INTVAL (const_arg1)) >= 0)
3654 || (HAVE_POST_INCREMENT
3655 && exact_log2 (INTVAL (const_arg1)) >= 0)
3656 || (HAVE_PRE_DECREMENT
3657 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3658 || (HAVE_POST_DECREMENT
3659 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3660 break;
3662 /* ??? Vector mode shifts by scalar
3663 shift operand are not supported yet. */
3664 if (is_shift && VECTOR_MODE_P (mode))
3665 break;
3667 if (is_shift
3668 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3669 || INTVAL (inner_const) < 0))
3671 if (SHIFT_COUNT_TRUNCATED)
3672 inner_const = GEN_INT (INTVAL (inner_const)
3673 & (GET_MODE_BITSIZE (mode) - 1));
3674 else
3675 break;
3678 /* Compute the code used to compose the constants. For example,
3679 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3681 associate_code = (is_shift || code == MINUS ? PLUS : code);
3683 new_const = simplify_binary_operation (associate_code, mode,
3684 canon_const_arg1,
3685 inner_const);
3687 if (new_const == 0)
3688 break;
3690 /* If we are associating shift operations, don't let this
3691 produce a shift of the size of the object or larger.
3692 This could occur when we follow a sign-extend by a right
3693 shift on a machine that does a sign-extend as a pair
3694 of shifts. */
3696 if (is_shift
3697 && CONST_INT_P (new_const)
3698 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3700 /* As an exception, we can turn an ASHIFTRT of this
3701 form into a shift of the number of bits - 1. */
3702 if (code == ASHIFTRT)
3703 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3704 else if (!side_effects_p (XEXP (y, 0)))
3705 return CONST0_RTX (mode);
3706 else
3707 break;
3710 y = copy_rtx (XEXP (y, 0));
3712 /* If Y contains our first operand (the most common way this
3713 can happen is if Y is a MEM), we would do into an infinite
3714 loop if we tried to fold it. So don't in that case. */
3716 if (! reg_mentioned_p (folded_arg0, y))
3717 y = fold_rtx (y, insn);
3719 return simplify_gen_binary (code, mode, y, new_const);
3721 break;
3723 case DIV: case UDIV:
3724 /* ??? The associative optimization performed immediately above is
3725 also possible for DIV and UDIV using associate_code of MULT.
3726 However, we would need extra code to verify that the
3727 multiplication does not overflow, that is, there is no overflow
3728 in the calculation of new_const. */
3729 break;
3731 default:
3732 break;
3735 new_rtx = simplify_binary_operation (code, mode,
3736 const_arg0 ? const_arg0 : folded_arg0,
3737 const_arg1 ? const_arg1 : folded_arg1);
3738 break;
3740 case RTX_OBJ:
3741 /* (lo_sum (high X) X) is simply X. */
3742 if (code == LO_SUM && const_arg0 != 0
3743 && GET_CODE (const_arg0) == HIGH
3744 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3745 return const_arg1;
3746 break;
3748 case RTX_TERNARY:
3749 case RTX_BITFIELD_OPS:
3750 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3751 const_arg0 ? const_arg0 : folded_arg0,
3752 const_arg1 ? const_arg1 : folded_arg1,
3753 const_arg2 ? const_arg2 : XEXP (x, 2));
3754 break;
3756 default:
3757 break;
3760 return new_rtx ? new_rtx : x;
3763 /* Return a constant value currently equivalent to X.
3764 Return 0 if we don't know one. */
3766 static rtx
3767 equiv_constant (rtx x)
3769 if (REG_P (x)
3770 && REGNO_QTY_VALID_P (REGNO (x)))
3772 int x_q = REG_QTY (REGNO (x));
3773 struct qty_table_elem *x_ent = &qty_table[x_q];
3775 if (x_ent->const_rtx)
3776 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3779 if (x == 0 || CONSTANT_P (x))
3780 return x;
3782 if (GET_CODE (x) == SUBREG)
3784 machine_mode mode = GET_MODE (x);
3785 machine_mode imode = GET_MODE (SUBREG_REG (x));
3786 rtx new_rtx;
3788 /* See if we previously assigned a constant value to this SUBREG. */
3789 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3790 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3791 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3792 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3793 return new_rtx;
3795 /* If we didn't and if doing so makes sense, see if we previously
3796 assigned a constant value to the enclosing word mode SUBREG. */
3797 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3798 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3800 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3801 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3803 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3804 new_rtx = lookup_as_function (y, CONST_INT);
3805 if (new_rtx)
3806 return gen_lowpart (mode, new_rtx);
3810 /* Otherwise see if we already have a constant for the inner REG,
3811 and if that is enough to calculate an equivalent constant for
3812 the subreg. Note that the upper bits of paradoxical subregs
3813 are undefined, so they cannot be said to equal anything. */
3814 if (REG_P (SUBREG_REG (x))
3815 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3816 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3817 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3819 return 0;
3822 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3823 the hash table in case its value was seen before. */
3825 if (MEM_P (x))
3827 struct table_elt *elt;
3829 x = avoid_constant_pool_reference (x);
3830 if (CONSTANT_P (x))
3831 return x;
3833 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3834 if (elt == 0)
3835 return 0;
3837 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3838 if (elt->is_const && CONSTANT_P (elt->exp))
3839 return elt->exp;
3842 return 0;
3845 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3846 "taken" branch.
3848 In certain cases, this can cause us to add an equivalence. For example,
3849 if we are following the taken case of
3850 if (i == 2)
3851 we can add the fact that `i' and '2' are now equivalent.
3853 In any case, we can record that this comparison was passed. If the same
3854 comparison is seen later, we will know its value. */
3856 static void
3857 record_jump_equiv (rtx_insn *insn, bool taken)
3859 int cond_known_true;
3860 rtx op0, op1;
3861 rtx set;
3862 machine_mode mode, mode0, mode1;
3863 int reversed_nonequality = 0;
3864 enum rtx_code code;
3866 /* Ensure this is the right kind of insn. */
3867 gcc_assert (any_condjump_p (insn));
3869 set = pc_set (insn);
3871 /* See if this jump condition is known true or false. */
3872 if (taken)
3873 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3874 else
3875 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3877 /* Get the type of comparison being done and the operands being compared.
3878 If we had to reverse a non-equality condition, record that fact so we
3879 know that it isn't valid for floating-point. */
3880 code = GET_CODE (XEXP (SET_SRC (set), 0));
3881 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3882 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3884 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3885 if (! cond_known_true)
3887 code = reversed_comparison_code_parts (code, op0, op1, insn);
3889 /* Don't remember if we can't find the inverse. */
3890 if (code == UNKNOWN)
3891 return;
3894 /* The mode is the mode of the non-constant. */
3895 mode = mode0;
3896 if (mode1 != VOIDmode)
3897 mode = mode1;
3899 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
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. */
3905 static rtx
3906 record_jump_cond_subreg (machine_mode mode, rtx op)
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);
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. */
3919 static void
3920 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3921 rtx op1, int reversed_nonequality)
3923 unsigned op0_hash, op1_hash;
3924 int op0_in_memory, op1_in_memory;
3925 struct table_elt *op0_elt, *op1_elt;
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). */
3932 /* Note that GET_MODE (op0) may not equal MODE. */
3933 if (code == EQ && paradoxical_subreg_p (op0))
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);
3942 if (code == EQ && paradoxical_subreg_p (op1))
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);
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. */
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. */
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)))))
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);
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)))))
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);
3982 /* Hash both operands. */
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;
3989 if (do_not_record)
3990 return;
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;
3997 if (do_not_record)
3998 return;
4000 /* Look up both operands. */
4001 op0_elt = lookup (op0, op0_hash, mode);
4002 op1_elt = lookup (op1, op1_hash, mode);
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;
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. */
4017 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4019 struct qty_table_elem *ent;
4020 int qty;
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. */
4026 if (!REG_P (op1))
4027 op1 = equiv_constant (op1);
4029 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4030 || !REG_P (op0) || op1 == 0)
4031 return;
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)
4037 if (insert_regs (op0, NULL, 0))
4039 rehash_using_reg (op0);
4040 op0_hash = HASH (op0, mode);
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);
4049 op0_elt = insert (op0, NULL, op0_hash, mode);
4050 op0_elt->in_memory = op0_in_memory;
4053 qty = REG_QTY (REGNO (op0));
4054 ent = &qty_table[qty];
4056 ent->comparison_code = code;
4057 if (REG_P (op1))
4059 /* Look it up again--in case op0 and op1 are the same. */
4060 op1_elt = lookup (op1, op1_hash, mode);
4062 /* Put OP1 in the hash table so it gets a new quantity number. */
4063 if (op1_elt == 0)
4065 if (insert_regs (op1, NULL, 0))
4067 rehash_using_reg (op1);
4068 op1_hash = HASH (op1, mode);
4071 op1_elt = insert (op1, NULL, op1_hash, mode);
4072 op1_elt->in_memory = op1_in_memory;
4075 ent->comparison_const = NULL_RTX;
4076 ent->comparison_qty = REG_QTY (REGNO (op1));
4078 else
4080 ent->comparison_const = op1;
4081 ent->comparison_qty = -1;
4084 return;
4087 /* If either side is still missing an equivalence, make it now,
4088 then merge the equivalences. */
4090 if (op0_elt == 0)
4092 if (insert_regs (op0, NULL, 0))
4094 rehash_using_reg (op0);
4095 op0_hash = HASH (op0, mode);
4098 op0_elt = insert (op0, NULL, op0_hash, mode);
4099 op0_elt->in_memory = op0_in_memory;
4102 if (op1_elt == 0)
4104 if (insert_regs (op1, NULL, 0))
4106 rehash_using_reg (op1);
4107 op1_hash = HASH (op1, mode);
4110 op1_elt = insert (op1, NULL, op1_hash, mode);
4111 op1_elt->in_memory = op1_in_memory;
4114 merge_equiv_classes (op0_elt, op1_elt);
4117 /* CSE processing for one instruction.
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.
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.
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.
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. */
4135 /* Data on one SET contained in the instruction. */
4137 struct set
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;
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.
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.
4181 This is the last transformation that cse_insn will try to do. */
4183 static void
4184 try_back_substitute_reg (rtx set, rtx_insn *insn)
4186 rtx dest = SET_DEST (set);
4187 rtx src = SET_SRC (set);
4189 if (REG_P (dest)
4190 && REG_P (src) && ! HARD_REGISTER_P (src)
4191 && REGNO_QTY_VALID_P (REGNO (src)))
4193 int src_q = REG_QTY (REGNO (src));
4194 struct qty_table_elem *src_ent = &qty_table[src_q];
4196 if (src_ent->first_reg == REGNO (dest))
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));
4204 prev = PREV_INSN (prev);
4206 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4208 /* Do not swap the registers around if the previous instruction
4209 attaches a REG_EQUIV note to REG1.
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.
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))
4224 rtx note;
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 ();
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);
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)
4250 struct set *sets = *psets;
4251 int n_sets = 0;
4252 rtx x = PATTERN (insn);
4254 if (GET_CODE (x) == SET)
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)
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)
4271 else
4272 sets[n_sets++].rtl = x;
4274 else if (GET_CODE (x) == PARALLEL)
4276 int i, lim = XVECLEN (x, 0);
4278 /* Go over the epressions 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++)
4282 rtx y = XVECEXP (x, 0, i);
4283 if (GET_CODE (y) == SET)
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)
4290 else if (GET_CODE (SET_SRC (y)) == CALL)
4292 else
4293 sets[n_sets++].rtl = y;
4298 return n_sets;
4301 /* Where possible, substitute every register reference in the N_SETS
4302 number of SETS in INSN with the the canonical register.
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:
4311 (set (reg1) (some_const))
4312 (set (mem (... reg1 ...) (thing)))
4313 (set (reg2) (some_const))
4314 (set (mem (... reg2 ...) (thing)))
4316 After canonicalizing, the code takes the following form:
4318 (set (reg1) (some_const))
4319 (set (mem (... reg1 ...) (thing)))
4320 (set (reg2) (some_const))
4321 (set (mem (... reg1 ...) (thing)))
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.
4326 In this function, the result of apply_change_group can be ignored;
4327 see canon_reg. */
4329 static void
4330 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4332 struct set *sets = *psets;
4333 rtx tem;
4334 rtx x = PATTERN (insn);
4335 int i;
4337 if (CALL_P (insn))
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);
4344 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4346 canon_reg (SET_SRC (x), insn);
4347 apply_change_group ();
4348 fold_rtx (SET_SRC (x), insn);
4350 else if (GET_CODE (x) == CLOBBER)
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);
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)
4365 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4367 rtx input = ASM_OPERANDS_INPUT (x, i);
4368 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4370 input = canon_reg (input, insn);
4371 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4375 else if (GET_CODE (x) == CALL)
4377 canon_reg (x, insn);
4378 apply_change_group ();
4379 fold_rtx (x, insn);
4381 else if (DEBUG_INSN_P (insn))
4382 canon_reg (PATTERN (insn), insn);
4383 else if (GET_CODE (x) == PARALLEL)
4385 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4387 rtx y = XVECEXP (x, 0, i);
4388 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4390 canon_reg (SET_SRC (y), insn);
4391 apply_change_group ();
4392 fold_rtx (SET_SRC (y), insn);
4394 else if (GET_CODE (y) == CLOBBER)
4396 if (MEM_P (XEXP (y, 0)))
4397 canon_reg (XEXP (y, 0), insn);
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)
4405 canon_reg (y, insn);
4406 apply_change_group ();
4407 fold_rtx (y, insn);
4412 if (n_sets == 1 && REG_NOTES (insn) != 0
4413 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
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.
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
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);
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.
4440 We used to suppress canonicalization of DEST if it appears in SRC,
4441 but we don't do this any more. */
4443 for (i = 0; i < n_sets; i++)
4445 rtx dest = SET_DEST (sets[i].rtl);
4446 rtx src = SET_SRC (sets[i].rtl);
4447 rtx new_rtx = canon_reg (src, insn);
4449 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4451 if (GET_CODE (dest) == ZERO_EXTRACT)
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);
4459 while (GET_CODE (dest) == SUBREG
4460 || GET_CODE (dest) == ZERO_EXTRACT
4461 || GET_CODE (dest) == STRICT_LOW_PART)
4462 dest = XEXP (dest, 0);
4464 if (MEM_P (dest))
4465 canon_reg (dest, insn);
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.
4474 The result of apply_change_group can be ignored; see canon_reg. */
4476 apply_change_group ();
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. */
4485 static void
4486 cse_insn (rtx_insn *insn)
4488 rtx x = PATTERN (insn);
4489 int i;
4490 rtx tem;
4491 int n_sets = 0;
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;
4499 struct set *sets = (struct set *) 0;
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));
4506 this_insn = insn;
4507 #ifdef HAVE_cc0
4508 /* Records what this insn does to set CC0. */
4509 this_insn_cc0 = 0;
4510 this_insn_cc0_mode = VOIDmode;
4511 #endif
4513 /* Find all regs explicitly clobbered in this insn,
4514 to ensure they are not replaced with any other regs
4515 elsewhere in this insn. */
4516 invalidate_from_sets_and_clobbers (insn);
4518 /* Record all the SETs in this instruction. */
4519 n_sets = find_sets_in_insn (insn, &sets);
4521 /* Substitute the canonical register where possible. */
4522 canonicalize_insn (insn, &sets, n_sets);
4524 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4525 if different, or if the DEST is a STRICT_LOW_PART. The latter condition
4526 is necessary because SRC_EQV is handled specially for this case, and if
4527 it isn't set, then there will be no equivalence for the destination. */
4528 if (n_sets == 1 && REG_NOTES (insn) != 0
4529 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0
4530 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4531 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4532 src_eqv = copy_rtx (XEXP (tem, 0));
4534 /* Set sets[i].src_elt to the class each source belongs to.
4535 Detect assignments from or to volatile things
4536 and set set[i] to zero so they will be ignored
4537 in the rest of this function.
4539 Nothing in this loop changes the hash table or the register chains. */
4541 for (i = 0; i < n_sets; i++)
4543 bool repeat = false;
4544 rtx src, dest;
4545 rtx src_folded;
4546 struct table_elt *elt = 0, *p;
4547 machine_mode mode;
4548 rtx src_eqv_here;
4549 rtx src_const = 0;
4550 rtx src_related = 0;
4551 bool src_related_is_const_anchor = false;
4552 struct table_elt *src_const_elt = 0;
4553 int src_cost = MAX_COST;
4554 int src_eqv_cost = MAX_COST;
4555 int src_folded_cost = MAX_COST;
4556 int src_related_cost = MAX_COST;
4557 int src_elt_cost = MAX_COST;
4558 int src_regcost = MAX_COST;
4559 int src_eqv_regcost = MAX_COST;
4560 int src_folded_regcost = MAX_COST;
4561 int src_related_regcost = MAX_COST;
4562 int src_elt_regcost = MAX_COST;
4563 /* Set nonzero if we need to call force_const_mem on with the
4564 contents of src_folded before using it. */
4565 int src_folded_force_flag = 0;
4567 dest = SET_DEST (sets[i].rtl);
4568 src = SET_SRC (sets[i].rtl);
4570 /* If SRC is a constant that has no machine mode,
4571 hash it with the destination's machine mode.
4572 This way we can keep different modes separate. */
4574 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4575 sets[i].mode = mode;
4577 if (src_eqv)
4579 machine_mode eqvmode = mode;
4580 if (GET_CODE (dest) == STRICT_LOW_PART)
4581 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4582 do_not_record = 0;
4583 hash_arg_in_memory = 0;
4584 src_eqv_hash = HASH (src_eqv, eqvmode);
4586 /* Find the equivalence class for the equivalent expression. */
4588 if (!do_not_record)
4589 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4591 src_eqv_volatile = do_not_record;
4592 src_eqv_in_memory = hash_arg_in_memory;
4595 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4596 value of the INNER register, not the destination. So it is not
4597 a valid substitution for the source. But save it for later. */
4598 if (GET_CODE (dest) == STRICT_LOW_PART)
4599 src_eqv_here = 0;
4600 else
4601 src_eqv_here = src_eqv;
4603 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4604 simplified result, which may not necessarily be valid. */
4605 src_folded = fold_rtx (src, insn);
4607 #if 0
4608 /* ??? This caused bad code to be generated for the m68k port with -O2.
4609 Suppose src is (CONST_INT -1), and that after truncation src_folded
4610 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4611 At the end we will add src and src_const to the same equivalence
4612 class. We now have 3 and -1 on the same equivalence class. This
4613 causes later instructions to be mis-optimized. */
4614 /* If storing a constant in a bitfield, pre-truncate the constant
4615 so we will be able to record it later. */
4616 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4618 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4620 if (CONST_INT_P (src)
4621 && CONST_INT_P (width)
4622 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4623 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4624 src_folded
4625 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4626 << INTVAL (width)) - 1));
4628 #endif
4630 /* Compute SRC's hash code, and also notice if it
4631 should not be recorded at all. In that case,
4632 prevent any further processing of this assignment. */
4633 do_not_record = 0;
4634 hash_arg_in_memory = 0;
4636 sets[i].src = src;
4637 sets[i].src_hash = HASH (src, mode);
4638 sets[i].src_volatile = do_not_record;
4639 sets[i].src_in_memory = hash_arg_in_memory;
4641 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4642 a pseudo, do not record SRC. Using SRC as a replacement for
4643 anything else will be incorrect in that situation. Note that
4644 this usually occurs only for stack slots, in which case all the
4645 RTL would be referring to SRC, so we don't lose any optimization
4646 opportunities by not having SRC in the hash table. */
4648 if (MEM_P (src)
4649 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4650 && REG_P (dest)
4651 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4652 sets[i].src_volatile = 1;
4654 /* Also do not record result of a non-volatile inline asm with
4655 more than one result or with clobbers, we do not want CSE to
4656 break the inline asm apart. */
4657 else if (GET_CODE (src) == ASM_OPERANDS
4658 && GET_CODE (x) == PARALLEL)
4659 sets[i].src_volatile = 1;
4661 #if 0
4662 /* It is no longer clear why we used to do this, but it doesn't
4663 appear to still be needed. So let's try without it since this
4664 code hurts cse'ing widened ops. */
4665 /* If source is a paradoxical subreg (such as QI treated as an SI),
4666 treat it as volatile. It may do the work of an SI in one context
4667 where the extra bits are not being used, but cannot replace an SI
4668 in general. */
4669 if (paradoxical_subreg_p (src))
4670 sets[i].src_volatile = 1;
4671 #endif
4673 /* Locate all possible equivalent forms for SRC. Try to replace
4674 SRC in the insn with each cheaper equivalent.
4676 We have the following types of equivalents: SRC itself, a folded
4677 version, a value given in a REG_EQUAL note, or a value related
4678 to a constant.
4680 Each of these equivalents may be part of an additional class
4681 of equivalents (if more than one is in the table, they must be in
4682 the same class; we check for this).
4684 If the source is volatile, we don't do any table lookups.
4686 We note any constant equivalent for possible later use in a
4687 REG_NOTE. */
4689 if (!sets[i].src_volatile)
4690 elt = lookup (src, sets[i].src_hash, mode);
4692 sets[i].src_elt = elt;
4694 if (elt && src_eqv_here && src_eqv_elt)
4696 if (elt->first_same_value != src_eqv_elt->first_same_value)
4698 /* The REG_EQUAL is indicating that two formerly distinct
4699 classes are now equivalent. So merge them. */
4700 merge_equiv_classes (elt, src_eqv_elt);
4701 src_eqv_hash = HASH (src_eqv, elt->mode);
4702 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4705 src_eqv_here = 0;
4708 else if (src_eqv_elt)
4709 elt = src_eqv_elt;
4711 /* Try to find a constant somewhere and record it in `src_const'.
4712 Record its table element, if any, in `src_const_elt'. Look in
4713 any known equivalences first. (If the constant is not in the
4714 table, also set `sets[i].src_const_hash'). */
4715 if (elt)
4716 for (p = elt->first_same_value; p; p = p->next_same_value)
4717 if (p->is_const)
4719 src_const = p->exp;
4720 src_const_elt = elt;
4721 break;
4724 if (src_const == 0
4725 && (CONSTANT_P (src_folded)
4726 /* Consider (minus (label_ref L1) (label_ref L2)) as
4727 "constant" here so we will record it. This allows us
4728 to fold switch statements when an ADDR_DIFF_VEC is used. */
4729 || (GET_CODE (src_folded) == MINUS
4730 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4731 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4732 src_const = src_folded, src_const_elt = elt;
4733 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4734 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4736 /* If we don't know if the constant is in the table, get its
4737 hash code and look it up. */
4738 if (src_const && src_const_elt == 0)
4740 sets[i].src_const_hash = HASH (src_const, mode);
4741 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4744 sets[i].src_const = src_const;
4745 sets[i].src_const_elt = src_const_elt;
4747 /* If the constant and our source are both in the table, mark them as
4748 equivalent. Otherwise, if a constant is in the table but the source
4749 isn't, set ELT to it. */
4750 if (src_const_elt && elt
4751 && src_const_elt->first_same_value != elt->first_same_value)
4752 merge_equiv_classes (elt, src_const_elt);
4753 else if (src_const_elt && elt == 0)
4754 elt = src_const_elt;
4756 /* See if there is a register linearly related to a constant
4757 equivalent of SRC. */
4758 if (src_const
4759 && (GET_CODE (src_const) == CONST
4760 || (src_const_elt && src_const_elt->related_value != 0)))
4762 src_related = use_related_value (src_const, src_const_elt);
4763 if (src_related)
4765 struct table_elt *src_related_elt
4766 = lookup (src_related, HASH (src_related, mode), mode);
4767 if (src_related_elt && elt)
4769 if (elt->first_same_value
4770 != src_related_elt->first_same_value)
4771 /* This can occur when we previously saw a CONST
4772 involving a SYMBOL_REF and then see the SYMBOL_REF
4773 twice. Merge the involved classes. */
4774 merge_equiv_classes (elt, src_related_elt);
4776 src_related = 0;
4777 src_related_elt = 0;
4779 else if (src_related_elt && elt == 0)
4780 elt = src_related_elt;
4784 /* See if we have a CONST_INT that is already in a register in a
4785 wider mode. */
4787 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4788 && GET_MODE_CLASS (mode) == MODE_INT
4789 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4791 machine_mode wider_mode;
4793 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4794 wider_mode != VOIDmode
4795 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4796 && src_related == 0;
4797 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4799 struct table_elt *const_elt
4800 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4802 if (const_elt == 0)
4803 continue;
4805 for (const_elt = const_elt->first_same_value;
4806 const_elt; const_elt = const_elt->next_same_value)
4807 if (REG_P (const_elt->exp))
4809 src_related = gen_lowpart (mode, const_elt->exp);
4810 break;
4815 /* Another possibility is that we have an AND with a constant in
4816 a mode narrower than a word. If so, it might have been generated
4817 as part of an "if" which would narrow the AND. If we already
4818 have done the AND in a wider mode, we can use a SUBREG of that
4819 value. */
4821 if (flag_expensive_optimizations && ! src_related
4822 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4823 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4825 machine_mode tmode;
4826 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4828 for (tmode = GET_MODE_WIDER_MODE (mode);
4829 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4830 tmode = GET_MODE_WIDER_MODE (tmode))
4832 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4833 struct table_elt *larger_elt;
4835 if (inner)
4837 PUT_MODE (new_and, tmode);
4838 XEXP (new_and, 0) = inner;
4839 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4840 if (larger_elt == 0)
4841 continue;
4843 for (larger_elt = larger_elt->first_same_value;
4844 larger_elt; larger_elt = larger_elt->next_same_value)
4845 if (REG_P (larger_elt->exp))
4847 src_related
4848 = gen_lowpart (mode, larger_elt->exp);
4849 break;
4852 if (src_related)
4853 break;
4858 #ifdef LOAD_EXTEND_OP
4859 /* See if a MEM has already been loaded with a widening operation;
4860 if it has, we can use a subreg of that. Many CISC machines
4861 also have such operations, but this is only likely to be
4862 beneficial on these machines. */
4864 if (flag_expensive_optimizations && src_related == 0
4865 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4866 && GET_MODE_CLASS (mode) == MODE_INT
4867 && MEM_P (src) && ! do_not_record
4868 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4870 struct rtx_def memory_extend_buf;
4871 rtx memory_extend_rtx = &memory_extend_buf;
4872 machine_mode tmode;
4874 /* Set what we are trying to extend and the operation it might
4875 have been extended with. */
4876 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4877 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4878 XEXP (memory_extend_rtx, 0) = src;
4880 for (tmode = GET_MODE_WIDER_MODE (mode);
4881 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4882 tmode = GET_MODE_WIDER_MODE (tmode))
4884 struct table_elt *larger_elt;
4886 PUT_MODE (memory_extend_rtx, tmode);
4887 larger_elt = lookup (memory_extend_rtx,
4888 HASH (memory_extend_rtx, tmode), tmode);
4889 if (larger_elt == 0)
4890 continue;
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))
4896 src_related = gen_lowpart (mode, larger_elt->exp);
4897 break;
4900 if (src_related)
4901 break;
4904 #endif /* LOAD_EXTEND_OP */
4906 /* Try to express the constant using a register+offset expression
4907 derived from a constant anchor. */
4909 if (targetm.const_anchor
4910 && !src_related
4911 && src_const
4912 && GET_CODE (src_const) == CONST_INT)
4914 src_related = try_const_anchors (src_const, mode);
4915 src_related_is_const_anchor = src_related != NULL_RTX;
4919 if (src == src_folded)
4920 src_folded = 0;
4922 /* At this point, ELT, if nonzero, points to a class of expressions
4923 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4924 and SRC_RELATED, if nonzero, each contain additional equivalent
4925 expressions. Prune these latter expressions by deleting expressions
4926 already in the equivalence class.
4928 Check for an equivalent identical to the destination. If found,
4929 this is the preferred equivalent since it will likely lead to
4930 elimination of the insn. Indicate this by placing it in
4931 `src_related'. */
4933 if (elt)
4934 elt = elt->first_same_value;
4935 for (p = elt; p; p = p->next_same_value)
4937 enum rtx_code code = GET_CODE (p->exp);
4939 /* If the expression is not valid, ignore it. Then we do not
4940 have to check for validity below. In most cases, we can use
4941 `rtx_equal_p', since canonicalization has already been done. */
4942 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4943 continue;
4945 /* Also skip paradoxical subregs, unless that's what we're
4946 looking for. */
4947 if (paradoxical_subreg_p (p->exp)
4948 && ! (src != 0
4949 && GET_CODE (src) == SUBREG
4950 && GET_MODE (src) == GET_MODE (p->exp)
4951 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
4952 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
4953 continue;
4955 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
4956 src = 0;
4957 else if (src_folded && GET_CODE (src_folded) == code
4958 && rtx_equal_p (src_folded, p->exp))
4959 src_folded = 0;
4960 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
4961 && rtx_equal_p (src_eqv_here, p->exp))
4962 src_eqv_here = 0;
4963 else if (src_related && GET_CODE (src_related) == code
4964 && rtx_equal_p (src_related, p->exp))
4965 src_related = 0;
4967 /* This is the same as the destination of the insns, we want
4968 to prefer it. Copy it to src_related. The code below will
4969 then give it a negative cost. */
4970 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
4971 src_related = dest;
4974 /* Find the cheapest valid equivalent, trying all the available
4975 possibilities. Prefer items not in the hash table to ones
4976 that are when they are equal cost. Note that we can never
4977 worsen an insn as the current contents will also succeed.
4978 If we find an equivalent identical to the destination, use it as best,
4979 since this insn will probably be eliminated in that case. */
4980 if (src)
4982 if (rtx_equal_p (src, dest))
4983 src_cost = src_regcost = -1;
4984 else
4986 src_cost = COST (src);
4987 src_regcost = approx_reg_cost (src);
4991 if (src_eqv_here)
4993 if (rtx_equal_p (src_eqv_here, dest))
4994 src_eqv_cost = src_eqv_regcost = -1;
4995 else
4997 src_eqv_cost = COST (src_eqv_here);
4998 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5002 if (src_folded)
5004 if (rtx_equal_p (src_folded, dest))
5005 src_folded_cost = src_folded_regcost = -1;
5006 else
5008 src_folded_cost = COST (src_folded);
5009 src_folded_regcost = approx_reg_cost (src_folded);
5013 if (src_related)
5015 if (rtx_equal_p (src_related, dest))
5016 src_related_cost = src_related_regcost = -1;
5017 else
5019 src_related_cost = COST (src_related);
5020 src_related_regcost = approx_reg_cost (src_related);
5022 /* If a const-anchor is used to synthesize a constant that
5023 normally requires multiple instructions then slightly prefer
5024 it over the original sequence. These instructions are likely
5025 to become redundant now. We can't compare against the cost
5026 of src_eqv_here because, on MIPS for example, multi-insn
5027 constants have zero cost; they are assumed to be hoisted from
5028 loops. */
5029 if (src_related_is_const_anchor
5030 && src_related_cost == src_cost
5031 && src_eqv_here)
5032 src_related_cost--;
5036 /* If this was an indirect jump insn, a known label will really be
5037 cheaper even though it looks more expensive. */
5038 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5039 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5041 /* Terminate loop when replacement made. This must terminate since
5042 the current contents will be tested and will always be valid. */
5043 while (1)
5045 rtx trial;
5047 /* Skip invalid entries. */
5048 while (elt && !REG_P (elt->exp)
5049 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5050 elt = elt->next_same_value;
5052 /* A paradoxical subreg would be bad here: it'll be the right
5053 size, but later may be adjusted so that the upper bits aren't
5054 what we want. So reject it. */
5055 if (elt != 0
5056 && paradoxical_subreg_p (elt->exp)
5057 /* It is okay, though, if the rtx we're trying to match
5058 will ignore any of the bits we can't predict. */
5059 && ! (src != 0
5060 && GET_CODE (src) == SUBREG
5061 && GET_MODE (src) == GET_MODE (elt->exp)
5062 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5063 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5065 elt = elt->next_same_value;
5066 continue;
5069 if (elt)
5071 src_elt_cost = elt->cost;
5072 src_elt_regcost = elt->regcost;
5075 /* Find cheapest and skip it for the next time. For items
5076 of equal cost, use this order:
5077 src_folded, src, src_eqv, src_related and hash table entry. */
5078 if (src_folded
5079 && preferable (src_folded_cost, src_folded_regcost,
5080 src_cost, src_regcost) <= 0
5081 && preferable (src_folded_cost, src_folded_regcost,
5082 src_eqv_cost, src_eqv_regcost) <= 0
5083 && preferable (src_folded_cost, src_folded_regcost,
5084 src_related_cost, src_related_regcost) <= 0
5085 && preferable (src_folded_cost, src_folded_regcost,
5086 src_elt_cost, src_elt_regcost) <= 0)
5088 trial = src_folded, src_folded_cost = MAX_COST;
5089 if (src_folded_force_flag)
5091 rtx forced = force_const_mem (mode, trial);
5092 if (forced)
5093 trial = forced;
5096 else if (src
5097 && preferable (src_cost, src_regcost,
5098 src_eqv_cost, src_eqv_regcost) <= 0
5099 && preferable (src_cost, src_regcost,
5100 src_related_cost, src_related_regcost) <= 0
5101 && preferable (src_cost, src_regcost,
5102 src_elt_cost, src_elt_regcost) <= 0)
5103 trial = src, src_cost = MAX_COST;
5104 else if (src_eqv_here
5105 && preferable (src_eqv_cost, src_eqv_regcost,
5106 src_related_cost, src_related_regcost) <= 0
5107 && preferable (src_eqv_cost, src_eqv_regcost,
5108 src_elt_cost, src_elt_regcost) <= 0)
5109 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5110 else if (src_related
5111 && preferable (src_related_cost, src_related_regcost,
5112 src_elt_cost, src_elt_regcost) <= 0)
5113 trial = src_related, src_related_cost = MAX_COST;
5114 else
5116 trial = elt->exp;
5117 elt = elt->next_same_value;
5118 src_elt_cost = MAX_COST;
5121 /* Avoid creation of overlapping memory moves. */
5122 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5124 rtx src, dest;
5126 /* BLKmode moves are not handled by cse anyway. */
5127 if (GET_MODE (trial) == BLKmode)
5128 break;
5130 src = canon_rtx (trial);
5131 dest = canon_rtx (SET_DEST (sets[i].rtl));
5133 if (!MEM_P (src) || !MEM_P (dest)
5134 || !nonoverlapping_memrefs_p (src, dest, false))
5135 break;
5138 /* Try to optimize
5139 (set (reg:M N) (const_int A))
5140 (set (reg:M2 O) (const_int B))
5141 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5142 (reg:M2 O)). */
5143 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5144 && CONST_INT_P (trial)
5145 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5146 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5147 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5148 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5149 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5150 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5151 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5152 <= HOST_BITS_PER_WIDE_INT))
5154 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5155 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5156 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5157 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5158 struct table_elt *dest_elt
5159 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5160 rtx dest_cst = NULL;
5162 if (dest_elt)
5163 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5164 if (p->is_const && CONST_INT_P (p->exp))
5166 dest_cst = p->exp;
5167 break;
5169 if (dest_cst)
5171 HOST_WIDE_INT val = INTVAL (dest_cst);
5172 HOST_WIDE_INT mask;
5173 unsigned int shift;
5174 if (BITS_BIG_ENDIAN)
5175 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5176 - INTVAL (pos) - INTVAL (width);
5177 else
5178 shift = INTVAL (pos);
5179 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5180 mask = ~(HOST_WIDE_INT) 0;
5181 else
5182 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5183 val &= ~(mask << shift);
5184 val |= (INTVAL (trial) & mask) << shift;
5185 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5186 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5187 dest_reg, 1);
5188 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5189 GEN_INT (val), 1);
5190 if (apply_change_group ())
5192 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5193 if (note)
5195 remove_note (insn, note);
5196 df_notes_rescan (insn);
5198 src_eqv = NULL_RTX;
5199 src_eqv_elt = NULL;
5200 src_eqv_volatile = 0;
5201 src_eqv_in_memory = 0;
5202 src_eqv_hash = 0;
5203 repeat = true;
5204 break;
5209 /* We don't normally have an insn matching (set (pc) (pc)), so
5210 check for this separately here. We will delete such an
5211 insn below.
5213 For other cases such as a table jump or conditional jump
5214 where we know the ultimate target, go ahead and replace the
5215 operand. While that may not make a valid insn, we will
5216 reemit the jump below (and also insert any necessary
5217 barriers). */
5218 if (n_sets == 1 && dest == pc_rtx
5219 && (trial == pc_rtx
5220 || (GET_CODE (trial) == LABEL_REF
5221 && ! condjump_p (insn))))
5223 /* Don't substitute non-local labels, this confuses CFG. */
5224 if (GET_CODE (trial) == LABEL_REF
5225 && LABEL_REF_NONLOCAL_P (trial))
5226 continue;
5228 SET_SRC (sets[i].rtl) = trial;
5229 cse_jumps_altered = true;
5230 break;
5233 /* Reject certain invalid forms of CONST that we create. */
5234 else if (CONSTANT_P (trial)
5235 && GET_CODE (trial) == CONST
5236 /* Reject cases that will cause decode_rtx_const to
5237 die. On the alpha when simplifying a switch, we
5238 get (const (truncate (minus (label_ref)
5239 (label_ref)))). */
5240 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5241 /* Likewise on IA-64, except without the
5242 truncate. */
5243 || (GET_CODE (XEXP (trial, 0)) == MINUS
5244 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5245 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5246 /* Do nothing for this case. */
5249 /* Look for a substitution that makes a valid insn. */
5250 else if (validate_unshare_change
5251 (insn, &SET_SRC (sets[i].rtl), trial, 0))
5253 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5255 /* The result of apply_change_group can be ignored; see
5256 canon_reg. */
5258 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5259 apply_change_group ();
5261 break;
5264 /* If we previously found constant pool entries for
5265 constants and this is a constant, try making a
5266 pool entry. Put it in src_folded unless we already have done
5267 this since that is where it likely came from. */
5269 else if (constant_pool_entries_cost
5270 && CONSTANT_P (trial)
5271 && (src_folded == 0
5272 || (!MEM_P (src_folded)
5273 && ! src_folded_force_flag))
5274 && GET_MODE_CLASS (mode) != MODE_CC
5275 && mode != VOIDmode)
5277 src_folded_force_flag = 1;
5278 src_folded = trial;
5279 src_folded_cost = constant_pool_entries_cost;
5280 src_folded_regcost = constant_pool_entries_regcost;
5284 /* If we changed the insn too much, handle this set from scratch. */
5285 if (repeat)
5287 i--;
5288 continue;
5291 src = SET_SRC (sets[i].rtl);
5293 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5294 However, there is an important exception: If both are registers
5295 that are not the head of their equivalence class, replace SET_SRC
5296 with the head of the class. If we do not do this, we will have
5297 both registers live over a portion of the basic block. This way,
5298 their lifetimes will likely abut instead of overlapping. */
5299 if (REG_P (dest)
5300 && REGNO_QTY_VALID_P (REGNO (dest)))
5302 int dest_q = REG_QTY (REGNO (dest));
5303 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5305 if (dest_ent->mode == GET_MODE (dest)
5306 && dest_ent->first_reg != REGNO (dest)
5307 && REG_P (src) && REGNO (src) == REGNO (dest)
5308 /* Don't do this if the original insn had a hard reg as
5309 SET_SRC or SET_DEST. */
5310 && (!REG_P (sets[i].src)
5311 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5312 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5313 /* We can't call canon_reg here because it won't do anything if
5314 SRC is a hard register. */
5316 int src_q = REG_QTY (REGNO (src));
5317 struct qty_table_elem *src_ent = &qty_table[src_q];
5318 int first = src_ent->first_reg;
5319 rtx new_src
5320 = (first >= FIRST_PSEUDO_REGISTER
5321 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5323 /* We must use validate-change even for this, because this
5324 might be a special no-op instruction, suitable only to
5325 tag notes onto. */
5326 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5328 src = new_src;
5329 /* If we had a constant that is cheaper than what we are now
5330 setting SRC to, use that constant. We ignored it when we
5331 thought we could make this into a no-op. */
5332 if (src_const && COST (src_const) < COST (src)
5333 && validate_change (insn, &SET_SRC (sets[i].rtl),
5334 src_const, 0))
5335 src = src_const;
5340 /* If we made a change, recompute SRC values. */
5341 if (src != sets[i].src)
5343 do_not_record = 0;
5344 hash_arg_in_memory = 0;
5345 sets[i].src = src;
5346 sets[i].src_hash = HASH (src, mode);
5347 sets[i].src_volatile = do_not_record;
5348 sets[i].src_in_memory = hash_arg_in_memory;
5349 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5352 /* If this is a single SET, we are setting a register, and we have an
5353 equivalent constant, we want to add a REG_EQUAL note if the constant
5354 is different from the source. We don't want to do it for a constant
5355 pseudo since verifying that this pseudo hasn't been eliminated is a
5356 pain; moreover such a note won't help anything.
5358 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5359 which can be created for a reference to a compile time computable
5360 entry in a jump table. */
5361 if (n_sets == 1
5362 && REG_P (dest)
5363 && src_const
5364 && !REG_P (src_const)
5365 && !(GET_CODE (src_const) == SUBREG
5366 && REG_P (SUBREG_REG (src_const)))
5367 && !(GET_CODE (src_const) == CONST
5368 && GET_CODE (XEXP (src_const, 0)) == MINUS
5369 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5370 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5371 && !rtx_equal_p (src, src_const))
5373 /* Make sure that the rtx is not shared. */
5374 src_const = copy_rtx (src_const);
5376 /* Record the actual constant value in a REG_EQUAL note,
5377 making a new one if one does not already exist. */
5378 set_unique_reg_note (insn, REG_EQUAL, src_const);
5379 df_notes_rescan (insn);
5382 /* Now deal with the destination. */
5383 do_not_record = 0;
5385 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5386 while (GET_CODE (dest) == SUBREG
5387 || GET_CODE (dest) == ZERO_EXTRACT
5388 || GET_CODE (dest) == STRICT_LOW_PART)
5389 dest = XEXP (dest, 0);
5391 sets[i].inner_dest = dest;
5393 if (MEM_P (dest))
5395 #ifdef PUSH_ROUNDING
5396 /* Stack pushes invalidate the stack pointer. */
5397 rtx addr = XEXP (dest, 0);
5398 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5399 && XEXP (addr, 0) == stack_pointer_rtx)
5400 invalidate (stack_pointer_rtx, VOIDmode);
5401 #endif
5402 dest = fold_rtx (dest, insn);
5405 /* Compute the hash code of the destination now,
5406 before the effects of this instruction are recorded,
5407 since the register values used in the address computation
5408 are those before this instruction. */
5409 sets[i].dest_hash = HASH (dest, mode);
5411 /* Don't enter a bit-field in the hash table
5412 because the value in it after the store
5413 may not equal what was stored, due to truncation. */
5415 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5417 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5419 if (src_const != 0 && CONST_INT_P (src_const)
5420 && CONST_INT_P (width)
5421 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5422 && ! (INTVAL (src_const)
5423 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5424 /* Exception: if the value is constant,
5425 and it won't be truncated, record it. */
5427 else
5429 /* This is chosen so that the destination will be invalidated
5430 but no new value will be recorded.
5431 We must invalidate because sometimes constant
5432 values can be recorded for bitfields. */
5433 sets[i].src_elt = 0;
5434 sets[i].src_volatile = 1;
5435 src_eqv = 0;
5436 src_eqv_elt = 0;
5440 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5441 the insn. */
5442 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5444 /* One less use of the label this insn used to jump to. */
5445 delete_insn_and_edges (insn);
5446 cse_jumps_altered = true;
5447 /* No more processing for this set. */
5448 sets[i].rtl = 0;
5451 /* If this SET is now setting PC to a label, we know it used to
5452 be a conditional or computed branch. */
5453 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5454 && !LABEL_REF_NONLOCAL_P (src))
5456 /* We reemit the jump in as many cases as possible just in
5457 case the form of an unconditional jump is significantly
5458 different than a computed jump or conditional jump.
5460 If this insn has multiple sets, then reemitting the
5461 jump is nontrivial. So instead we just force rerecognition
5462 and hope for the best. */
5463 if (n_sets == 1)
5465 rtx_insn *new_rtx;
5466 rtx note;
5468 new_rtx = emit_jump_insn_before (gen_jump (XEXP (src, 0)), insn);
5469 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5470 LABEL_NUSES (XEXP (src, 0))++;
5472 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5473 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5474 if (note)
5476 XEXP (note, 1) = NULL_RTX;
5477 REG_NOTES (new_rtx) = note;
5480 delete_insn_and_edges (insn);
5481 insn = new_rtx;
5483 else
5484 INSN_CODE (insn) = -1;
5486 /* Do not bother deleting any unreachable code, let jump do it. */
5487 cse_jumps_altered = true;
5488 sets[i].rtl = 0;
5491 /* If destination is volatile, invalidate it and then do no further
5492 processing for this assignment. */
5494 else if (do_not_record)
5496 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5497 invalidate (dest, VOIDmode);
5498 else if (MEM_P (dest))
5499 invalidate (dest, VOIDmode);
5500 else if (GET_CODE (dest) == STRICT_LOW_PART
5501 || GET_CODE (dest) == ZERO_EXTRACT)
5502 invalidate (XEXP (dest, 0), GET_MODE (dest));
5503 sets[i].rtl = 0;
5506 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5507 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5509 #ifdef HAVE_cc0
5510 /* If setting CC0, record what it was set to, or a constant, if it
5511 is equivalent to a constant. If it is being set to a floating-point
5512 value, make a COMPARE with the appropriate constant of 0. If we
5513 don't do this, later code can interpret this as a test against
5514 const0_rtx, which can cause problems if we try to put it into an
5515 insn as a floating-point operand. */
5516 if (dest == cc0_rtx)
5518 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5519 this_insn_cc0_mode = mode;
5520 if (FLOAT_MODE_P (mode))
5521 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5522 CONST0_RTX (mode));
5524 #endif
5527 /* Now enter all non-volatile source expressions in the hash table
5528 if they are not already present.
5529 Record their equivalence classes in src_elt.
5530 This way we can insert the corresponding destinations into
5531 the same classes even if the actual sources are no longer in them
5532 (having been invalidated). */
5534 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5535 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5537 struct table_elt *elt;
5538 struct table_elt *classp = sets[0].src_elt;
5539 rtx dest = SET_DEST (sets[0].rtl);
5540 machine_mode eqvmode = GET_MODE (dest);
5542 if (GET_CODE (dest) == STRICT_LOW_PART)
5544 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5545 classp = 0;
5547 if (insert_regs (src_eqv, classp, 0))
5549 rehash_using_reg (src_eqv);
5550 src_eqv_hash = HASH (src_eqv, eqvmode);
5552 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5553 elt->in_memory = src_eqv_in_memory;
5554 src_eqv_elt = elt;
5556 /* Check to see if src_eqv_elt is the same as a set source which
5557 does not yet have an elt, and if so set the elt of the set source
5558 to src_eqv_elt. */
5559 for (i = 0; i < n_sets; i++)
5560 if (sets[i].rtl && sets[i].src_elt == 0
5561 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5562 sets[i].src_elt = src_eqv_elt;
5565 for (i = 0; i < n_sets; i++)
5566 if (sets[i].rtl && ! sets[i].src_volatile
5567 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5569 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5571 /* REG_EQUAL in setting a STRICT_LOW_PART
5572 gives an equivalent for the entire destination register,
5573 not just for the subreg being stored in now.
5574 This is a more interesting equivalence, so we arrange later
5575 to treat the entire reg as the destination. */
5576 sets[i].src_elt = src_eqv_elt;
5577 sets[i].src_hash = src_eqv_hash;
5579 else
5581 /* Insert source and constant equivalent into hash table, if not
5582 already present. */
5583 struct table_elt *classp = src_eqv_elt;
5584 rtx src = sets[i].src;
5585 rtx dest = SET_DEST (sets[i].rtl);
5586 machine_mode mode
5587 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5589 /* It's possible that we have a source value known to be
5590 constant but don't have a REG_EQUAL note on the insn.
5591 Lack of a note will mean src_eqv_elt will be NULL. This
5592 can happen where we've generated a SUBREG to access a
5593 CONST_INT that is already in a register in a wider mode.
5594 Ensure that the source expression is put in the proper
5595 constant class. */
5596 if (!classp)
5597 classp = sets[i].src_const_elt;
5599 if (sets[i].src_elt == 0)
5601 struct table_elt *elt;
5603 /* Note that these insert_regs calls cannot remove
5604 any of the src_elt's, because they would have failed to
5605 match if not still valid. */
5606 if (insert_regs (src, classp, 0))
5608 rehash_using_reg (src);
5609 sets[i].src_hash = HASH (src, mode);
5611 elt = insert (src, classp, sets[i].src_hash, mode);
5612 elt->in_memory = sets[i].src_in_memory;
5613 sets[i].src_elt = classp = elt;
5615 if (sets[i].src_const && sets[i].src_const_elt == 0
5616 && src != sets[i].src_const
5617 && ! rtx_equal_p (sets[i].src_const, src))
5618 sets[i].src_elt = insert (sets[i].src_const, classp,
5619 sets[i].src_const_hash, mode);
5622 else if (sets[i].src_elt == 0)
5623 /* If we did not insert the source into the hash table (e.g., it was
5624 volatile), note the equivalence class for the REG_EQUAL value, if any,
5625 so that the destination goes into that class. */
5626 sets[i].src_elt = src_eqv_elt;
5628 /* Record destination addresses in the hash table. This allows us to
5629 check if they are invalidated by other sets. */
5630 for (i = 0; i < n_sets; i++)
5632 if (sets[i].rtl)
5634 rtx x = sets[i].inner_dest;
5635 struct table_elt *elt;
5636 machine_mode mode;
5637 unsigned hash;
5639 if (MEM_P (x))
5641 x = XEXP (x, 0);
5642 mode = GET_MODE (x);
5643 hash = HASH (x, mode);
5644 elt = lookup (x, hash, mode);
5645 if (!elt)
5647 if (insert_regs (x, NULL, 0))
5649 rtx dest = SET_DEST (sets[i].rtl);
5651 rehash_using_reg (x);
5652 hash = HASH (x, mode);
5653 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5655 elt = insert (x, NULL, hash, mode);
5658 sets[i].dest_addr_elt = elt;
5660 else
5661 sets[i].dest_addr_elt = NULL;
5665 invalidate_from_clobbers (insn);
5667 /* Some registers are invalidated by subroutine calls. Memory is
5668 invalidated by non-constant calls. */
5670 if (CALL_P (insn))
5672 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5673 invalidate_memory ();
5674 invalidate_for_call ();
5677 /* Now invalidate everything set by this instruction.
5678 If a SUBREG or other funny destination is being set,
5679 sets[i].rtl is still nonzero, so here we invalidate the reg
5680 a part of which is being set. */
5682 for (i = 0; i < n_sets; i++)
5683 if (sets[i].rtl)
5685 /* We can't use the inner dest, because the mode associated with
5686 a ZERO_EXTRACT is significant. */
5687 rtx dest = SET_DEST (sets[i].rtl);
5689 /* Needed for registers to remove the register from its
5690 previous quantity's chain.
5691 Needed for memory if this is a nonvarying address, unless
5692 we have just done an invalidate_memory that covers even those. */
5693 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5694 invalidate (dest, VOIDmode);
5695 else if (MEM_P (dest))
5696 invalidate (dest, VOIDmode);
5697 else if (GET_CODE (dest) == STRICT_LOW_PART
5698 || GET_CODE (dest) == ZERO_EXTRACT)
5699 invalidate (XEXP (dest, 0), GET_MODE (dest));
5702 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5703 the regs restored by the longjmp come from a later time
5704 than the setjmp. */
5705 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5707 flush_hash_table ();
5708 goto done;
5711 /* Make sure registers mentioned in destinations
5712 are safe for use in an expression to be inserted.
5713 This removes from the hash table
5714 any invalid entry that refers to one of these registers.
5716 We don't care about the return value from mention_regs because
5717 we are going to hash the SET_DEST values unconditionally. */
5719 for (i = 0; i < n_sets; i++)
5721 if (sets[i].rtl)
5723 rtx x = SET_DEST (sets[i].rtl);
5725 if (!REG_P (x))
5726 mention_regs (x);
5727 else
5729 /* We used to rely on all references to a register becoming
5730 inaccessible when a register changes to a new quantity,
5731 since that changes the hash code. However, that is not
5732 safe, since after HASH_SIZE new quantities we get a
5733 hash 'collision' of a register with its own invalid
5734 entries. And since SUBREGs have been changed not to
5735 change their hash code with the hash code of the register,
5736 it wouldn't work any longer at all. So we have to check
5737 for any invalid references lying around now.
5738 This code is similar to the REG case in mention_regs,
5739 but it knows that reg_tick has been incremented, and
5740 it leaves reg_in_table as -1 . */
5741 unsigned int regno = REGNO (x);
5742 unsigned int endregno = END_REGNO (x);
5743 unsigned int i;
5745 for (i = regno; i < endregno; i++)
5747 if (REG_IN_TABLE (i) >= 0)
5749 remove_invalid_refs (i);
5750 REG_IN_TABLE (i) = -1;
5757 /* We may have just removed some of the src_elt's from the hash table.
5758 So replace each one with the current head of the same class.
5759 Also check if destination addresses have been removed. */
5761 for (i = 0; i < n_sets; i++)
5762 if (sets[i].rtl)
5764 if (sets[i].dest_addr_elt
5765 && sets[i].dest_addr_elt->first_same_value == 0)
5767 /* The elt was removed, which means this destination is not
5768 valid after this instruction. */
5769 sets[i].rtl = NULL_RTX;
5771 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5772 /* If elt was removed, find current head of same class,
5773 or 0 if nothing remains of that class. */
5775 struct table_elt *elt = sets[i].src_elt;
5777 while (elt && elt->prev_same_value)
5778 elt = elt->prev_same_value;
5780 while (elt && elt->first_same_value == 0)
5781 elt = elt->next_same_value;
5782 sets[i].src_elt = elt ? elt->first_same_value : 0;
5786 /* Now insert the destinations into their equivalence classes. */
5788 for (i = 0; i < n_sets; i++)
5789 if (sets[i].rtl)
5791 rtx dest = SET_DEST (sets[i].rtl);
5792 struct table_elt *elt;
5794 /* Don't record value if we are not supposed to risk allocating
5795 floating-point values in registers that might be wider than
5796 memory. */
5797 if ((flag_float_store
5798 && MEM_P (dest)
5799 && FLOAT_MODE_P (GET_MODE (dest)))
5800 /* Don't record BLKmode values, because we don't know the
5801 size of it, and can't be sure that other BLKmode values
5802 have the same or smaller size. */
5803 || GET_MODE (dest) == BLKmode
5804 /* If we didn't put a REG_EQUAL value or a source into the hash
5805 table, there is no point is recording DEST. */
5806 || sets[i].src_elt == 0
5807 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5808 or SIGN_EXTEND, don't record DEST since it can cause
5809 some tracking to be wrong.
5811 ??? Think about this more later. */
5812 || (paradoxical_subreg_p (dest)
5813 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5814 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5815 continue;
5817 /* STRICT_LOW_PART isn't part of the value BEING set,
5818 and neither is the SUBREG inside it.
5819 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5820 if (GET_CODE (dest) == STRICT_LOW_PART)
5821 dest = SUBREG_REG (XEXP (dest, 0));
5823 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5824 /* Registers must also be inserted into chains for quantities. */
5825 if (insert_regs (dest, sets[i].src_elt, 1))
5827 /* If `insert_regs' changes something, the hash code must be
5828 recalculated. */
5829 rehash_using_reg (dest);
5830 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5833 elt = insert (dest, sets[i].src_elt,
5834 sets[i].dest_hash, GET_MODE (dest));
5836 /* If this is a constant, insert the constant anchors with the
5837 equivalent register-offset expressions using register DEST. */
5838 if (targetm.const_anchor
5839 && REG_P (dest)
5840 && SCALAR_INT_MODE_P (GET_MODE (dest))
5841 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5842 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5844 elt->in_memory = (MEM_P (sets[i].inner_dest)
5845 && !MEM_READONLY_P (sets[i].inner_dest));
5847 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5848 narrower than M2, and both M1 and M2 are the same number of words,
5849 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5850 make that equivalence as well.
5852 However, BAR may have equivalences for which gen_lowpart
5853 will produce a simpler value than gen_lowpart applied to
5854 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5855 BAR's equivalences. If we don't get a simplified form, make
5856 the SUBREG. It will not be used in an equivalence, but will
5857 cause two similar assignments to be detected.
5859 Note the loop below will find SUBREG_REG (DEST) since we have
5860 already entered SRC and DEST of the SET in the table. */
5862 if (GET_CODE (dest) == SUBREG
5863 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5864 / UNITS_PER_WORD)
5865 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5866 && (GET_MODE_SIZE (GET_MODE (dest))
5867 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5868 && sets[i].src_elt != 0)
5870 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5871 struct table_elt *elt, *classp = 0;
5873 for (elt = sets[i].src_elt->first_same_value; elt;
5874 elt = elt->next_same_value)
5876 rtx new_src = 0;
5877 unsigned src_hash;
5878 struct table_elt *src_elt;
5879 int byte = 0;
5881 /* Ignore invalid entries. */
5882 if (!REG_P (elt->exp)
5883 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5884 continue;
5886 /* We may have already been playing subreg games. If the
5887 mode is already correct for the destination, use it. */
5888 if (GET_MODE (elt->exp) == new_mode)
5889 new_src = elt->exp;
5890 else
5892 /* Calculate big endian correction for the SUBREG_BYTE.
5893 We have already checked that M1 (GET_MODE (dest))
5894 is not narrower than M2 (new_mode). */
5895 if (BYTES_BIG_ENDIAN)
5896 byte = (GET_MODE_SIZE (GET_MODE (dest))
5897 - GET_MODE_SIZE (new_mode));
5899 new_src = simplify_gen_subreg (new_mode, elt->exp,
5900 GET_MODE (dest), byte);
5903 /* The call to simplify_gen_subreg fails if the value
5904 is VOIDmode, yet we can't do any simplification, e.g.
5905 for EXPR_LISTs denoting function call results.
5906 It is invalid to construct a SUBREG with a VOIDmode
5907 SUBREG_REG, hence a zero new_src means we can't do
5908 this substitution. */
5909 if (! new_src)
5910 continue;
5912 src_hash = HASH (new_src, new_mode);
5913 src_elt = lookup (new_src, src_hash, new_mode);
5915 /* Put the new source in the hash table is if isn't
5916 already. */
5917 if (src_elt == 0)
5919 if (insert_regs (new_src, classp, 0))
5921 rehash_using_reg (new_src);
5922 src_hash = HASH (new_src, new_mode);
5924 src_elt = insert (new_src, classp, src_hash, new_mode);
5925 src_elt->in_memory = elt->in_memory;
5927 else if (classp && classp != src_elt->first_same_value)
5928 /* Show that two things that we've seen before are
5929 actually the same. */
5930 merge_equiv_classes (src_elt, classp);
5932 classp = src_elt->first_same_value;
5933 /* Ignore invalid entries. */
5934 while (classp
5935 && !REG_P (classp->exp)
5936 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5937 classp = classp->next_same_value;
5942 /* Special handling for (set REG0 REG1) where REG0 is the
5943 "cheapest", cheaper than REG1. After cse, REG1 will probably not
5944 be used in the sequel, so (if easily done) change this insn to
5945 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
5946 that computed their value. Then REG1 will become a dead store
5947 and won't cloud the situation for later optimizations.
5949 Do not make this change if REG1 is a hard register, because it will
5950 then be used in the sequel and we may be changing a two-operand insn
5951 into a three-operand insn.
5953 Also do not do this if we are operating on a copy of INSN. */
5955 if (n_sets == 1 && sets[0].rtl)
5956 try_back_substitute_reg (sets[0].rtl, insn);
5958 done:;
5961 /* Remove from the hash table all expressions that reference memory. */
5963 static void
5964 invalidate_memory (void)
5966 int i;
5967 struct table_elt *p, *next;
5969 for (i = 0; i < HASH_SIZE; i++)
5970 for (p = table[i]; p; p = next)
5972 next = p->next_same_hash;
5973 if (p->in_memory)
5974 remove_from_table (p, i);
5978 /* Perform invalidation on the basis of everything about INSN,
5979 except for invalidating the actual places that are SET in it.
5980 This includes the places CLOBBERed, and anything that might
5981 alias with something that is SET or CLOBBERed. */
5983 static void
5984 invalidate_from_clobbers (rtx_insn *insn)
5986 rtx x = PATTERN (insn);
5988 if (GET_CODE (x) == CLOBBER)
5990 rtx ref = XEXP (x, 0);
5991 if (ref)
5993 if (REG_P (ref) || GET_CODE (ref) == SUBREG
5994 || MEM_P (ref))
5995 invalidate (ref, VOIDmode);
5996 else if (GET_CODE (ref) == STRICT_LOW_PART
5997 || GET_CODE (ref) == ZERO_EXTRACT)
5998 invalidate (XEXP (ref, 0), GET_MODE (ref));
6001 else if (GET_CODE (x) == PARALLEL)
6003 int i;
6004 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6006 rtx y = XVECEXP (x, 0, i);
6007 if (GET_CODE (y) == CLOBBER)
6009 rtx ref = XEXP (y, 0);
6010 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6011 || MEM_P (ref))
6012 invalidate (ref, VOIDmode);
6013 else if (GET_CODE (ref) == STRICT_LOW_PART
6014 || GET_CODE (ref) == ZERO_EXTRACT)
6015 invalidate (XEXP (ref, 0), GET_MODE (ref));
6021 /* Perform invalidation on the basis of everything about INSN.
6022 This includes the places CLOBBERed, and anything that might
6023 alias with something that is SET or CLOBBERed. */
6025 static void
6026 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6028 rtx tem;
6029 rtx x = PATTERN (insn);
6031 if (CALL_P (insn))
6033 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6034 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6035 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6038 /* Ensure we invalidate the destination register of a CALL insn.
6039 This is necessary for machines where this register is a fixed_reg,
6040 because no other code would invalidate it. */
6041 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6042 invalidate (SET_DEST (x), VOIDmode);
6044 else if (GET_CODE (x) == PARALLEL)
6046 int i;
6048 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6050 rtx y = XVECEXP (x, 0, i);
6051 if (GET_CODE (y) == CLOBBER)
6053 rtx clobbered = XEXP (y, 0);
6055 if (REG_P (clobbered)
6056 || GET_CODE (clobbered) == SUBREG)
6057 invalidate (clobbered, VOIDmode);
6058 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6059 || GET_CODE (clobbered) == ZERO_EXTRACT)
6060 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6062 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6063 invalidate (SET_DEST (y), VOIDmode);
6068 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6069 and replace any registers in them with either an equivalent constant
6070 or the canonical form of the register. If we are inside an address,
6071 only do this if the address remains valid.
6073 OBJECT is 0 except when within a MEM in which case it is the MEM.
6075 Return the replacement for X. */
6077 static rtx
6078 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6080 enum rtx_code code = GET_CODE (x);
6081 const char *fmt = GET_RTX_FORMAT (code);
6082 int i;
6084 switch (code)
6086 case CONST:
6087 case SYMBOL_REF:
6088 case LABEL_REF:
6089 CASE_CONST_ANY:
6090 case PC:
6091 case CC0:
6092 case LO_SUM:
6093 return x;
6095 case MEM:
6096 validate_change (x, &XEXP (x, 0),
6097 cse_process_notes (XEXP (x, 0), x, changed), 0);
6098 return x;
6100 case EXPR_LIST:
6101 if (REG_NOTE_KIND (x) == REG_EQUAL)
6102 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6103 /* Fall through. */
6105 case INSN_LIST:
6106 case INT_LIST:
6107 if (XEXP (x, 1))
6108 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6109 return x;
6111 case SIGN_EXTEND:
6112 case ZERO_EXTEND:
6113 case SUBREG:
6115 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6116 /* We don't substitute VOIDmode constants into these rtx,
6117 since they would impede folding. */
6118 if (GET_MODE (new_rtx) != VOIDmode)
6119 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6120 return x;
6123 case UNSIGNED_FLOAT:
6125 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6126 /* We don't substitute negative VOIDmode constants into these rtx,
6127 since they would impede folding. */
6128 if (GET_MODE (new_rtx) != VOIDmode
6129 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6130 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6131 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6132 return x;
6135 case REG:
6136 i = REG_QTY (REGNO (x));
6138 /* Return a constant or a constant register. */
6139 if (REGNO_QTY_VALID_P (REGNO (x)))
6141 struct qty_table_elem *ent = &qty_table[i];
6143 if (ent->const_rtx != NULL_RTX
6144 && (CONSTANT_P (ent->const_rtx)
6145 || REG_P (ent->const_rtx)))
6147 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6148 if (new_rtx)
6149 return copy_rtx (new_rtx);
6153 /* Otherwise, canonicalize this register. */
6154 return canon_reg (x, NULL);
6156 default:
6157 break;
6160 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6161 if (fmt[i] == 'e')
6162 validate_change (object, &XEXP (x, i),
6163 cse_process_notes (XEXP (x, i), object, changed), 0);
6165 return x;
6168 static rtx
6169 cse_process_notes (rtx x, rtx object, bool *changed)
6171 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6172 if (new_rtx != x)
6173 *changed = true;
6174 return new_rtx;
6178 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6180 DATA is a pointer to a struct cse_basic_block_data, that is used to
6181 describe the path.
6182 It is filled with a queue of basic blocks, starting with FIRST_BB
6183 and following a trace through the CFG.
6185 If all paths starting at FIRST_BB have been followed, or no new path
6186 starting at FIRST_BB can be constructed, this function returns FALSE.
6187 Otherwise, DATA->path is filled and the function returns TRUE indicating
6188 that a path to follow was found.
6190 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6191 block in the path will be FIRST_BB. */
6193 static bool
6194 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6195 int follow_jumps)
6197 basic_block bb;
6198 edge e;
6199 int path_size;
6201 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6203 /* See if there is a previous path. */
6204 path_size = data->path_size;
6206 /* There is a previous path. Make sure it started with FIRST_BB. */
6207 if (path_size)
6208 gcc_assert (data->path[0].bb == first_bb);
6210 /* There was only one basic block in the last path. Clear the path and
6211 return, so that paths starting at another basic block can be tried. */
6212 if (path_size == 1)
6214 path_size = 0;
6215 goto done;
6218 /* If the path was empty from the beginning, construct a new path. */
6219 if (path_size == 0)
6220 data->path[path_size++].bb = first_bb;
6221 else
6223 /* Otherwise, path_size must be equal to or greater than 2, because
6224 a previous path exists that is at least two basic blocks long.
6226 Update the previous branch path, if any. If the last branch was
6227 previously along the branch edge, take the fallthrough edge now. */
6228 while (path_size >= 2)
6230 basic_block last_bb_in_path, previous_bb_in_path;
6231 edge e;
6233 --path_size;
6234 last_bb_in_path = data->path[path_size].bb;
6235 previous_bb_in_path = data->path[path_size - 1].bb;
6237 /* If we previously followed a path along the branch edge, try
6238 the fallthru edge now. */
6239 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6240 && any_condjump_p (BB_END (previous_bb_in_path))
6241 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6242 && e == BRANCH_EDGE (previous_bb_in_path))
6244 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6245 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6246 && single_pred_p (bb)
6247 /* We used to assert here that we would only see blocks
6248 that we have not visited yet. But we may end up
6249 visiting basic blocks twice if the CFG has changed
6250 in this run of cse_main, because when the CFG changes
6251 the topological sort of the CFG also changes. A basic
6252 blocks that previously had more than two predecessors
6253 may now have a single predecessor, and become part of
6254 a path that starts at another basic block.
6256 We still want to visit each basic block only once, so
6257 halt the path here if we have already visited BB. */
6258 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6260 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6261 data->path[path_size++].bb = bb;
6262 break;
6266 data->path[path_size].bb = NULL;
6269 /* If only one block remains in the path, bail. */
6270 if (path_size == 1)
6272 path_size = 0;
6273 goto done;
6277 /* Extend the path if possible. */
6278 if (follow_jumps)
6280 bb = data->path[path_size - 1].bb;
6281 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6283 if (single_succ_p (bb))
6284 e = single_succ_edge (bb);
6285 else if (EDGE_COUNT (bb->succs) == 2
6286 && any_condjump_p (BB_END (bb)))
6288 /* First try to follow the branch. If that doesn't lead
6289 to a useful path, follow the fallthru edge. */
6290 e = BRANCH_EDGE (bb);
6291 if (!single_pred_p (e->dest))
6292 e = FALLTHRU_EDGE (bb);
6294 else
6295 e = NULL;
6297 if (e
6298 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6299 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6300 && single_pred_p (e->dest)
6301 /* Avoid visiting basic blocks twice. The large comment
6302 above explains why this can happen. */
6303 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6305 basic_block bb2 = e->dest;
6306 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6307 data->path[path_size++].bb = bb2;
6308 bb = bb2;
6310 else
6311 bb = NULL;
6315 done:
6316 data->path_size = path_size;
6317 return path_size != 0;
6320 /* Dump the path in DATA to file F. NSETS is the number of sets
6321 in the path. */
6323 static void
6324 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6326 int path_entry;
6328 fprintf (f, ";; Following path with %d sets: ", nsets);
6329 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6330 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6331 fputc ('\n', dump_file);
6332 fflush (f);
6336 /* Return true if BB has exception handling successor edges. */
6338 static bool
6339 have_eh_succ_edges (basic_block bb)
6341 edge e;
6342 edge_iterator ei;
6344 FOR_EACH_EDGE (e, ei, bb->succs)
6345 if (e->flags & EDGE_EH)
6346 return true;
6348 return false;
6352 /* Scan to the end of the path described by DATA. Return an estimate of
6353 the total number of SETs of all insns in the path. */
6355 static void
6356 cse_prescan_path (struct cse_basic_block_data *data)
6358 int nsets = 0;
6359 int path_size = data->path_size;
6360 int path_entry;
6362 /* Scan to end of each basic block in the path. */
6363 for (path_entry = 0; path_entry < path_size; path_entry++)
6365 basic_block bb;
6366 rtx_insn *insn;
6368 bb = data->path[path_entry].bb;
6370 FOR_BB_INSNS (bb, insn)
6372 if (!INSN_P (insn))
6373 continue;
6375 /* A PARALLEL can have lots of SETs in it,
6376 especially if it is really an ASM_OPERANDS. */
6377 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6378 nsets += XVECLEN (PATTERN (insn), 0);
6379 else
6380 nsets += 1;
6384 data->nsets = nsets;
6387 /* Return true if the pattern of INSN uses a LABEL_REF for which
6388 there isn't a REG_LABEL_OPERAND note. */
6390 static bool
6391 check_for_label_ref (rtx_insn *insn)
6393 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6394 note for it, we must rerun jump since it needs to place the note. If
6395 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6396 don't do this since no REG_LABEL_OPERAND will be added. */
6397 subrtx_iterator::array_type array;
6398 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6400 const_rtx x = *iter;
6401 if (GET_CODE (x) == LABEL_REF
6402 && !LABEL_REF_NONLOCAL_P (x)
6403 && (!JUMP_P (insn)
6404 || !label_is_jump_target_p (LABEL_REF_LABEL (x), insn))
6405 && LABEL_P (LABEL_REF_LABEL (x))
6406 && INSN_UID (LABEL_REF_LABEL (x)) != 0
6407 && !find_reg_note (insn, REG_LABEL_OPERAND, LABEL_REF_LABEL (x)))
6408 return true;
6410 return false;
6413 /* Process a single extended basic block described by EBB_DATA. */
6415 static void
6416 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6418 int path_size = ebb_data->path_size;
6419 int path_entry;
6420 int num_insns = 0;
6422 /* Allocate the space needed by qty_table. */
6423 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6425 new_basic_block ();
6426 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6427 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6428 for (path_entry = 0; path_entry < path_size; path_entry++)
6430 basic_block bb;
6431 rtx_insn *insn;
6433 bb = ebb_data->path[path_entry].bb;
6435 /* Invalidate recorded information for eh regs if there is an EH
6436 edge pointing to that bb. */
6437 if (bb_has_eh_pred (bb))
6439 df_ref def;
6441 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6442 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6443 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6446 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6447 FOR_BB_INSNS (bb, insn)
6449 /* If we have processed 1,000 insns, flush the hash table to
6450 avoid extreme quadratic behavior. We must not include NOTEs
6451 in the count since there may be more of them when generating
6452 debugging information. If we clear the table at different
6453 times, code generated with -g -O might be different than code
6454 generated with -O but not -g.
6456 FIXME: This is a real kludge and needs to be done some other
6457 way. */
6458 if (NONDEBUG_INSN_P (insn)
6459 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6461 flush_hash_table ();
6462 num_insns = 0;
6465 if (INSN_P (insn))
6467 /* Process notes first so we have all notes in canonical forms
6468 when looking for duplicate operations. */
6469 if (REG_NOTES (insn))
6471 bool changed = false;
6472 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6473 NULL_RTX, &changed);
6474 if (changed)
6475 df_notes_rescan (insn);
6478 cse_insn (insn);
6480 /* If we haven't already found an insn where we added a LABEL_REF,
6481 check this one. */
6482 if (INSN_P (insn) && !recorded_label_ref
6483 && check_for_label_ref (insn))
6484 recorded_label_ref = true;
6486 #ifdef HAVE_cc0
6487 if (NONDEBUG_INSN_P (insn))
6489 /* If the previous insn sets CC0 and this insn no
6490 longer references CC0, delete the previous insn.
6491 Here we use fact that nothing expects CC0 to be
6492 valid over an insn, which is true until the final
6493 pass. */
6494 rtx_insn *prev_insn;
6495 rtx tem;
6497 prev_insn = prev_nonnote_nondebug_insn (insn);
6498 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6499 && (tem = single_set (prev_insn)) != NULL_RTX
6500 && SET_DEST (tem) == cc0_rtx
6501 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6502 delete_insn (prev_insn);
6504 /* If this insn is not the last insn in the basic
6505 block, it will be PREV_INSN(insn) in the next
6506 iteration. If we recorded any CC0-related
6507 information for this insn, remember it. */
6508 if (insn != BB_END (bb))
6510 prev_insn_cc0 = this_insn_cc0;
6511 prev_insn_cc0_mode = this_insn_cc0_mode;
6514 #endif
6518 /* With non-call exceptions, we are not always able to update
6519 the CFG properly inside cse_insn. So clean up possibly
6520 redundant EH edges here. */
6521 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6522 cse_cfg_altered |= purge_dead_edges (bb);
6524 /* If we changed a conditional jump, we may have terminated
6525 the path we are following. Check that by verifying that
6526 the edge we would take still exists. If the edge does
6527 not exist anymore, purge the remainder of the path.
6528 Note that this will cause us to return to the caller. */
6529 if (path_entry < path_size - 1)
6531 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6532 if (!find_edge (bb, next_bb))
6536 path_size--;
6538 /* If we truncate the path, we must also reset the
6539 visited bit on the remaining blocks in the path,
6540 or we will never visit them at all. */
6541 bitmap_clear_bit (cse_visited_basic_blocks,
6542 ebb_data->path[path_size].bb->index);
6543 ebb_data->path[path_size].bb = NULL;
6545 while (path_size - 1 != path_entry);
6546 ebb_data->path_size = path_size;
6550 /* If this is a conditional jump insn, record any known
6551 equivalences due to the condition being tested. */
6552 insn = BB_END (bb);
6553 if (path_entry < path_size - 1
6554 && JUMP_P (insn)
6555 && single_set (insn)
6556 && any_condjump_p (insn))
6558 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6559 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6560 record_jump_equiv (insn, taken);
6563 #ifdef HAVE_cc0
6564 /* Clear the CC0-tracking related insns, they can't provide
6565 useful information across basic block boundaries. */
6566 prev_insn_cc0 = 0;
6567 #endif
6570 gcc_assert (next_qty <= max_qty);
6572 free (qty_table);
6576 /* Perform cse on the instructions of a function.
6577 F is the first instruction.
6578 NREGS is one plus the highest pseudo-reg number used in the instruction.
6580 Return 2 if jump optimizations should be redone due to simplifications
6581 in conditional jump instructions.
6582 Return 1 if the CFG should be cleaned up because it has been modified.
6583 Return 0 otherwise. */
6585 static int
6586 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6588 struct cse_basic_block_data ebb_data;
6589 basic_block bb;
6590 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6591 int i, n_blocks;
6593 df_set_flags (DF_LR_RUN_DCE);
6594 df_note_add_problem ();
6595 df_analyze ();
6596 df_set_flags (DF_DEFER_INSN_RESCAN);
6598 reg_scan (get_insns (), max_reg_num ());
6599 init_cse_reg_info (nregs);
6601 ebb_data.path = XNEWVEC (struct branch_path,
6602 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6604 cse_cfg_altered = false;
6605 cse_jumps_altered = false;
6606 recorded_label_ref = false;
6607 constant_pool_entries_cost = 0;
6608 constant_pool_entries_regcost = 0;
6609 ebb_data.path_size = 0;
6610 ebb_data.nsets = 0;
6611 rtl_hooks = cse_rtl_hooks;
6613 init_recog ();
6614 init_alias_analysis ();
6616 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6618 /* Set up the table of already visited basic blocks. */
6619 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6620 bitmap_clear (cse_visited_basic_blocks);
6622 /* Loop over basic blocks in reverse completion order (RPO),
6623 excluding the ENTRY and EXIT blocks. */
6624 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6625 i = 0;
6626 while (i < n_blocks)
6628 /* Find the first block in the RPO queue that we have not yet
6629 processed before. */
6632 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6634 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6635 && i < n_blocks);
6637 /* Find all paths starting with BB, and process them. */
6638 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6640 /* Pre-scan the path. */
6641 cse_prescan_path (&ebb_data);
6643 /* If this basic block has no sets, skip it. */
6644 if (ebb_data.nsets == 0)
6645 continue;
6647 /* Get a reasonable estimate for the maximum number of qty's
6648 needed for this path. For this, we take the number of sets
6649 and multiply that by MAX_RECOG_OPERANDS. */
6650 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6652 /* Dump the path we're about to process. */
6653 if (dump_file)
6654 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6656 cse_extended_basic_block (&ebb_data);
6660 /* Clean up. */
6661 end_alias_analysis ();
6662 free (reg_eqv_table);
6663 free (ebb_data.path);
6664 sbitmap_free (cse_visited_basic_blocks);
6665 free (rc_order);
6666 rtl_hooks = general_rtl_hooks;
6668 if (cse_jumps_altered || recorded_label_ref)
6669 return 2;
6670 else if (cse_cfg_altered)
6671 return 1;
6672 else
6673 return 0;
6676 /* Count the number of times registers are used (not set) in X.
6677 COUNTS is an array in which we accumulate the count, INCR is how much
6678 we count each register usage.
6680 Don't count a usage of DEST, which is the SET_DEST of a SET which
6681 contains X in its SET_SRC. This is because such a SET does not
6682 modify the liveness of DEST.
6683 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6684 We must then count uses of a SET_DEST regardless, because the insn can't be
6685 deleted here. */
6687 static void
6688 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6690 enum rtx_code code;
6691 rtx note;
6692 const char *fmt;
6693 int i, j;
6695 if (x == 0)
6696 return;
6698 switch (code = GET_CODE (x))
6700 case REG:
6701 if (x != dest)
6702 counts[REGNO (x)] += incr;
6703 return;
6705 case PC:
6706 case CC0:
6707 case CONST:
6708 CASE_CONST_ANY:
6709 case SYMBOL_REF:
6710 case LABEL_REF:
6711 return;
6713 case CLOBBER:
6714 /* If we are clobbering a MEM, mark any registers inside the address
6715 as being used. */
6716 if (MEM_P (XEXP (x, 0)))
6717 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6718 return;
6720 case SET:
6721 /* Unless we are setting a REG, count everything in SET_DEST. */
6722 if (!REG_P (SET_DEST (x)))
6723 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6724 count_reg_usage (SET_SRC (x), counts,
6725 dest ? dest : SET_DEST (x),
6726 incr);
6727 return;
6729 case DEBUG_INSN:
6730 return;
6732 case CALL_INSN:
6733 case INSN:
6734 case JUMP_INSN:
6735 /* We expect dest to be NULL_RTX here. If the insn may throw,
6736 or if it cannot be deleted due to side-effects, mark this fact
6737 by setting DEST to pc_rtx. */
6738 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6739 || side_effects_p (PATTERN (x)))
6740 dest = pc_rtx;
6741 if (code == CALL_INSN)
6742 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6743 count_reg_usage (PATTERN (x), counts, dest, incr);
6745 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6746 use them. */
6748 note = find_reg_equal_equiv_note (x);
6749 if (note)
6751 rtx eqv = XEXP (note, 0);
6753 if (GET_CODE (eqv) == EXPR_LIST)
6754 /* This REG_EQUAL note describes the result of a function call.
6755 Process all the arguments. */
6758 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6759 eqv = XEXP (eqv, 1);
6761 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6762 else
6763 count_reg_usage (eqv, counts, dest, incr);
6765 return;
6767 case EXPR_LIST:
6768 if (REG_NOTE_KIND (x) == REG_EQUAL
6769 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6770 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6771 involving registers in the address. */
6772 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6773 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6775 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6776 return;
6778 case ASM_OPERANDS:
6779 /* Iterate over just the inputs, not the constraints as well. */
6780 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6781 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6782 return;
6784 case INSN_LIST:
6785 case INT_LIST:
6786 gcc_unreachable ();
6788 default:
6789 break;
6792 fmt = GET_RTX_FORMAT (code);
6793 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6795 if (fmt[i] == 'e')
6796 count_reg_usage (XEXP (x, i), counts, dest, incr);
6797 else if (fmt[i] == 'E')
6798 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6799 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6803 /* Return true if X is a dead register. */
6805 static inline int
6806 is_dead_reg (const_rtx x, int *counts)
6808 return (REG_P (x)
6809 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6810 && counts[REGNO (x)] == 0);
6813 /* Return true if set is live. */
6814 static bool
6815 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6816 int *counts)
6818 #ifdef HAVE_cc0
6819 rtx tem;
6820 #endif
6822 if (set_noop_p (set))
6825 #ifdef HAVE_cc0
6826 else if (GET_CODE (SET_DEST (set)) == CC0
6827 && !side_effects_p (SET_SRC (set))
6828 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6829 || !INSN_P (tem)
6830 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6831 return false;
6832 #endif
6833 else if (!is_dead_reg (SET_DEST (set), counts)
6834 || side_effects_p (SET_SRC (set)))
6835 return true;
6836 return false;
6839 /* Return true if insn is live. */
6841 static bool
6842 insn_live_p (rtx_insn *insn, int *counts)
6844 int i;
6845 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6846 return true;
6847 else if (GET_CODE (PATTERN (insn)) == SET)
6848 return set_live_p (PATTERN (insn), insn, counts);
6849 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6851 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6853 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6855 if (GET_CODE (elt) == SET)
6857 if (set_live_p (elt, insn, counts))
6858 return true;
6860 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6861 return true;
6863 return false;
6865 else if (DEBUG_INSN_P (insn))
6867 rtx_insn *next;
6869 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6870 if (NOTE_P (next))
6871 continue;
6872 else if (!DEBUG_INSN_P (next))
6873 return true;
6874 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6875 return false;
6877 return true;
6879 else
6880 return true;
6883 /* Count the number of stores into pseudo. Callback for note_stores. */
6885 static void
6886 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6888 int *counts = (int *) data;
6889 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6890 counts[REGNO (x)]++;
6893 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6894 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6895 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6896 Set *SEEN_REPL to true if we see a dead register that does have
6897 a replacement. */
6899 static bool
6900 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6901 bool *seen_repl)
6903 subrtx_iterator::array_type array;
6904 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6906 const_rtx x = *iter;
6907 if (is_dead_reg (x, counts))
6909 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6910 *seen_repl = true;
6911 else
6912 return true;
6915 return false;
6918 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6919 Callback for simplify_replace_fn_rtx. */
6921 static rtx
6922 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6924 rtx *replacements = (rtx *) data;
6926 if (REG_P (x)
6927 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6928 && replacements[REGNO (x)] != NULL_RTX)
6930 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6931 return replacements[REGNO (x)];
6932 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6933 GET_MODE (replacements[REGNO (x)]));
6935 return NULL_RTX;
6938 /* Scan all the insns and delete any that are dead; i.e., they store a register
6939 that is never used or they copy a register to itself.
6941 This is used to remove insns made obviously dead by cse, loop or other
6942 optimizations. It improves the heuristics in loop since it won't try to
6943 move dead invariants out of loops or make givs for dead quantities. The
6944 remaining passes of the compilation are also sped up. */
6947 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
6949 int *counts;
6950 rtx_insn *insn, *prev;
6951 rtx *replacements = NULL;
6952 int ndead = 0;
6954 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
6955 /* First count the number of times each register is used. */
6956 if (MAY_HAVE_DEBUG_INSNS)
6958 counts = XCNEWVEC (int, nreg * 3);
6959 for (insn = insns; insn; insn = NEXT_INSN (insn))
6960 if (DEBUG_INSN_P (insn))
6961 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
6962 NULL_RTX, 1);
6963 else if (INSN_P (insn))
6965 count_reg_usage (insn, counts, NULL_RTX, 1);
6966 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
6968 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
6969 First one counts how many times each pseudo is used outside
6970 of debug insns, second counts how many times each pseudo is
6971 used in debug insns and third counts how many times a pseudo
6972 is stored. */
6974 else
6976 counts = XCNEWVEC (int, nreg);
6977 for (insn = insns; insn; insn = NEXT_INSN (insn))
6978 if (INSN_P (insn))
6979 count_reg_usage (insn, counts, NULL_RTX, 1);
6980 /* If no debug insns can be present, COUNTS is just an array
6981 which counts how many times each pseudo is used. */
6983 /* Pseudo PIC register should be considered as used due to possible
6984 new usages generated. */
6985 if (!reload_completed
6986 && pic_offset_table_rtx
6987 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
6988 counts[REGNO (pic_offset_table_rtx)]++;
6989 /* Go from the last insn to the first and delete insns that only set unused
6990 registers or copy a register to itself. As we delete an insn, remove
6991 usage counts for registers it uses.
6993 The first jump optimization pass may leave a real insn as the last
6994 insn in the function. We must not skip that insn or we may end
6995 up deleting code that is not really dead.
6997 If some otherwise unused register is only used in DEBUG_INSNs,
6998 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
6999 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7000 has been created for the unused register, replace it with
7001 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7002 for (insn = get_last_insn (); insn; insn = prev)
7004 int live_insn = 0;
7006 prev = PREV_INSN (insn);
7007 if (!INSN_P (insn))
7008 continue;
7010 live_insn = insn_live_p (insn, counts);
7012 /* If this is a dead insn, delete it and show registers in it aren't
7013 being used. */
7015 if (! live_insn && dbg_cnt (delete_trivial_dead))
7017 if (DEBUG_INSN_P (insn))
7018 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7019 NULL_RTX, -1);
7020 else
7022 rtx set;
7023 if (MAY_HAVE_DEBUG_INSNS
7024 && (set = single_set (insn)) != NULL_RTX
7025 && is_dead_reg (SET_DEST (set), counts)
7026 /* Used at least once in some DEBUG_INSN. */
7027 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7028 /* And set exactly once. */
7029 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7030 && !side_effects_p (SET_SRC (set))
7031 && asm_noperands (PATTERN (insn)) < 0)
7033 rtx dval, bind_var_loc;
7034 rtx_insn *bind;
7036 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7037 dval = make_debug_expr_from_rtl (SET_DEST (set));
7039 /* Emit a debug bind insn before the insn in which
7040 reg dies. */
7041 bind_var_loc =
7042 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7043 DEBUG_EXPR_TREE_DECL (dval),
7044 SET_SRC (set),
7045 VAR_INIT_STATUS_INITIALIZED);
7046 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7048 bind = emit_debug_insn_before (bind_var_loc, insn);
7049 df_insn_rescan (bind);
7051 if (replacements == NULL)
7052 replacements = XCNEWVEC (rtx, nreg);
7053 replacements[REGNO (SET_DEST (set))] = dval;
7056 count_reg_usage (insn, counts, NULL_RTX, -1);
7057 ndead++;
7059 delete_insn_and_edges (insn);
7063 if (MAY_HAVE_DEBUG_INSNS)
7065 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7066 if (DEBUG_INSN_P (insn))
7068 /* If this debug insn references a dead register that wasn't replaced
7069 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7070 bool seen_repl = false;
7071 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7072 counts, replacements, &seen_repl))
7074 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7075 df_insn_rescan (insn);
7077 else if (seen_repl)
7079 INSN_VAR_LOCATION_LOC (insn)
7080 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7081 NULL_RTX, replace_dead_reg,
7082 replacements);
7083 df_insn_rescan (insn);
7086 free (replacements);
7089 if (dump_file && ndead)
7090 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7091 ndead);
7092 /* Clean up. */
7093 free (counts);
7094 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7095 return ndead;
7098 /* If LOC contains references to NEWREG in a different mode, change them
7099 to use NEWREG instead. */
7101 static void
7102 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7103 rtx *loc, rtx insn, rtx newreg)
7105 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7107 rtx *loc = *iter;
7108 rtx x = *loc;
7109 if (x
7110 && REG_P (x)
7111 && REGNO (x) == REGNO (newreg)
7112 && GET_MODE (x) != GET_MODE (newreg))
7114 validate_change (insn, loc, newreg, 1);
7115 iter.skip_subrtxes ();
7120 /* Change the mode of any reference to the register REGNO (NEWREG) to
7121 GET_MODE (NEWREG) in INSN. */
7123 static void
7124 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7126 int success;
7128 if (!INSN_P (insn))
7129 return;
7131 subrtx_ptr_iterator::array_type array;
7132 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7133 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7135 /* If the following assertion was triggered, there is most probably
7136 something wrong with the cc_modes_compatible back end function.
7137 CC modes only can be considered compatible if the insn - with the mode
7138 replaced by any of the compatible modes - can still be recognized. */
7139 success = apply_change_group ();
7140 gcc_assert (success);
7143 /* Change the mode of any reference to the register REGNO (NEWREG) to
7144 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7145 any instruction which modifies NEWREG. */
7147 static void
7148 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7150 rtx_insn *insn;
7152 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7154 if (! INSN_P (insn))
7155 continue;
7157 if (reg_set_p (newreg, insn))
7158 return;
7160 cse_change_cc_mode_insn (insn, newreg);
7164 /* BB is a basic block which finishes with CC_REG as a condition code
7165 register which is set to CC_SRC. Look through the successors of BB
7166 to find blocks which have a single predecessor (i.e., this one),
7167 and look through those blocks for an assignment to CC_REG which is
7168 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7169 permitted to change the mode of CC_SRC to a compatible mode. This
7170 returns VOIDmode if no equivalent assignments were found.
7171 Otherwise it returns the mode which CC_SRC should wind up with.
7172 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7173 but is passed unmodified down to recursive calls in order to prevent
7174 endless recursion.
7176 The main complexity in this function is handling the mode issues.
7177 We may have more than one duplicate which we can eliminate, and we
7178 try to find a mode which will work for multiple duplicates. */
7180 static machine_mode
7181 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7182 bool can_change_mode)
7184 bool found_equiv;
7185 machine_mode mode;
7186 unsigned int insn_count;
7187 edge e;
7188 rtx_insn *insns[2];
7189 machine_mode modes[2];
7190 rtx_insn *last_insns[2];
7191 unsigned int i;
7192 rtx newreg;
7193 edge_iterator ei;
7195 /* We expect to have two successors. Look at both before picking
7196 the final mode for the comparison. If we have more successors
7197 (i.e., some sort of table jump, although that seems unlikely),
7198 then we require all beyond the first two to use the same
7199 mode. */
7201 found_equiv = false;
7202 mode = GET_MODE (cc_src);
7203 insn_count = 0;
7204 FOR_EACH_EDGE (e, ei, bb->succs)
7206 rtx_insn *insn;
7207 rtx_insn *end;
7209 if (e->flags & EDGE_COMPLEX)
7210 continue;
7212 if (EDGE_COUNT (e->dest->preds) != 1
7213 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7214 /* Avoid endless recursion on unreachable blocks. */
7215 || e->dest == orig_bb)
7216 continue;
7218 end = NEXT_INSN (BB_END (e->dest));
7219 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7221 rtx set;
7223 if (! INSN_P (insn))
7224 continue;
7226 /* If CC_SRC is modified, we have to stop looking for
7227 something which uses it. */
7228 if (modified_in_p (cc_src, insn))
7229 break;
7231 /* Check whether INSN sets CC_REG to CC_SRC. */
7232 set = single_set (insn);
7233 if (set
7234 && REG_P (SET_DEST (set))
7235 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7237 bool found;
7238 machine_mode set_mode;
7239 machine_mode comp_mode;
7241 found = false;
7242 set_mode = GET_MODE (SET_SRC (set));
7243 comp_mode = set_mode;
7244 if (rtx_equal_p (cc_src, SET_SRC (set)))
7245 found = true;
7246 else if (GET_CODE (cc_src) == COMPARE
7247 && GET_CODE (SET_SRC (set)) == COMPARE
7248 && mode != set_mode
7249 && rtx_equal_p (XEXP (cc_src, 0),
7250 XEXP (SET_SRC (set), 0))
7251 && rtx_equal_p (XEXP (cc_src, 1),
7252 XEXP (SET_SRC (set), 1)))
7255 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7256 if (comp_mode != VOIDmode
7257 && (can_change_mode || comp_mode == mode))
7258 found = true;
7261 if (found)
7263 found_equiv = true;
7264 if (insn_count < ARRAY_SIZE (insns))
7266 insns[insn_count] = insn;
7267 modes[insn_count] = set_mode;
7268 last_insns[insn_count] = end;
7269 ++insn_count;
7271 if (mode != comp_mode)
7273 gcc_assert (can_change_mode);
7274 mode = comp_mode;
7276 /* The modified insn will be re-recognized later. */
7277 PUT_MODE (cc_src, mode);
7280 else
7282 if (set_mode != mode)
7284 /* We found a matching expression in the
7285 wrong mode, but we don't have room to
7286 store it in the array. Punt. This case
7287 should be rare. */
7288 break;
7290 /* INSN sets CC_REG to a value equal to CC_SRC
7291 with the right mode. We can simply delete
7292 it. */
7293 delete_insn (insn);
7296 /* We found an instruction to delete. Keep looking,
7297 in the hopes of finding a three-way jump. */
7298 continue;
7301 /* We found an instruction which sets the condition
7302 code, so don't look any farther. */
7303 break;
7306 /* If INSN sets CC_REG in some other way, don't look any
7307 farther. */
7308 if (reg_set_p (cc_reg, insn))
7309 break;
7312 /* If we fell off the bottom of the block, we can keep looking
7313 through successors. We pass CAN_CHANGE_MODE as false because
7314 we aren't prepared to handle compatibility between the
7315 further blocks and this block. */
7316 if (insn == end)
7318 machine_mode submode;
7320 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7321 if (submode != VOIDmode)
7323 gcc_assert (submode == mode);
7324 found_equiv = true;
7325 can_change_mode = false;
7330 if (! found_equiv)
7331 return VOIDmode;
7333 /* Now INSN_COUNT is the number of instructions we found which set
7334 CC_REG to a value equivalent to CC_SRC. The instructions are in
7335 INSNS. The modes used by those instructions are in MODES. */
7337 newreg = NULL_RTX;
7338 for (i = 0; i < insn_count; ++i)
7340 if (modes[i] != mode)
7342 /* We need to change the mode of CC_REG in INSNS[i] and
7343 subsequent instructions. */
7344 if (! newreg)
7346 if (GET_MODE (cc_reg) == mode)
7347 newreg = cc_reg;
7348 else
7349 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7351 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7352 newreg);
7355 delete_insn_and_edges (insns[i]);
7358 return mode;
7361 /* If we have a fixed condition code register (or two), walk through
7362 the instructions and try to eliminate duplicate assignments. */
7364 static void
7365 cse_condition_code_reg (void)
7367 unsigned int cc_regno_1;
7368 unsigned int cc_regno_2;
7369 rtx cc_reg_1;
7370 rtx cc_reg_2;
7371 basic_block bb;
7373 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7374 return;
7376 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7377 if (cc_regno_2 != INVALID_REGNUM)
7378 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7379 else
7380 cc_reg_2 = NULL_RTX;
7382 FOR_EACH_BB_FN (bb, cfun)
7384 rtx_insn *last_insn;
7385 rtx cc_reg;
7386 rtx_insn *insn;
7387 rtx_insn *cc_src_insn;
7388 rtx cc_src;
7389 machine_mode mode;
7390 machine_mode orig_mode;
7392 /* Look for blocks which end with a conditional jump based on a
7393 condition code register. Then look for the instruction which
7394 sets the condition code register. Then look through the
7395 successor blocks for instructions which set the condition
7396 code register to the same value. There are other possible
7397 uses of the condition code register, but these are by far the
7398 most common and the ones which we are most likely to be able
7399 to optimize. */
7401 last_insn = BB_END (bb);
7402 if (!JUMP_P (last_insn))
7403 continue;
7405 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7406 cc_reg = cc_reg_1;
7407 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7408 cc_reg = cc_reg_2;
7409 else
7410 continue;
7412 cc_src_insn = NULL;
7413 cc_src = NULL_RTX;
7414 for (insn = PREV_INSN (last_insn);
7415 insn && insn != PREV_INSN (BB_HEAD (bb));
7416 insn = PREV_INSN (insn))
7418 rtx set;
7420 if (! INSN_P (insn))
7421 continue;
7422 set = single_set (insn);
7423 if (set
7424 && REG_P (SET_DEST (set))
7425 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7427 cc_src_insn = insn;
7428 cc_src = SET_SRC (set);
7429 break;
7431 else if (reg_set_p (cc_reg, insn))
7432 break;
7435 if (! cc_src_insn)
7436 continue;
7438 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7439 continue;
7441 /* Now CC_REG is a condition code register used for a
7442 conditional jump at the end of the block, and CC_SRC, in
7443 CC_SRC_INSN, is the value to which that condition code
7444 register is set, and CC_SRC is still meaningful at the end of
7445 the basic block. */
7447 orig_mode = GET_MODE (cc_src);
7448 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7449 if (mode != VOIDmode)
7451 gcc_assert (mode == GET_MODE (cc_src));
7452 if (mode != orig_mode)
7454 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7456 cse_change_cc_mode_insn (cc_src_insn, newreg);
7458 /* Do the same in the following insns that use the
7459 current value of CC_REG within BB. */
7460 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7461 NEXT_INSN (last_insn),
7462 newreg);
7469 /* Perform common subexpression elimination. Nonzero value from
7470 `cse_main' means that jumps were simplified and some code may now
7471 be unreachable, so do jump optimization again. */
7472 static unsigned int
7473 rest_of_handle_cse (void)
7475 int tem;
7477 if (dump_file)
7478 dump_flow_info (dump_file, dump_flags);
7480 tem = cse_main (get_insns (), max_reg_num ());
7482 /* If we are not running more CSE passes, then we are no longer
7483 expecting CSE to be run. But always rerun it in a cheap mode. */
7484 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7486 if (tem == 2)
7488 timevar_push (TV_JUMP);
7489 rebuild_jump_labels (get_insns ());
7490 cleanup_cfg (CLEANUP_CFG_CHANGED);
7491 timevar_pop (TV_JUMP);
7493 else if (tem == 1 || optimize > 1)
7494 cleanup_cfg (0);
7496 return 0;
7499 namespace {
7501 const pass_data pass_data_cse =
7503 RTL_PASS, /* type */
7504 "cse1", /* name */
7505 OPTGROUP_NONE, /* optinfo_flags */
7506 TV_CSE, /* tv_id */
7507 0, /* properties_required */
7508 0, /* properties_provided */
7509 0, /* properties_destroyed */
7510 0, /* todo_flags_start */
7511 TODO_df_finish, /* todo_flags_finish */
7514 class pass_cse : public rtl_opt_pass
7516 public:
7517 pass_cse (gcc::context *ctxt)
7518 : rtl_opt_pass (pass_data_cse, ctxt)
7521 /* opt_pass methods: */
7522 virtual bool gate (function *) { return optimize > 0; }
7523 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7525 }; // class pass_cse
7527 } // anon namespace
7529 rtl_opt_pass *
7530 make_pass_cse (gcc::context *ctxt)
7532 return new pass_cse (ctxt);
7536 /* Run second CSE pass after loop optimizations. */
7537 static unsigned int
7538 rest_of_handle_cse2 (void)
7540 int tem;
7542 if (dump_file)
7543 dump_flow_info (dump_file, dump_flags);
7545 tem = cse_main (get_insns (), max_reg_num ());
7547 /* Run a pass to eliminate duplicated assignments to condition code
7548 registers. We have to run this after bypass_jumps, because it
7549 makes it harder for that pass to determine whether a jump can be
7550 bypassed safely. */
7551 cse_condition_code_reg ();
7553 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7555 if (tem == 2)
7557 timevar_push (TV_JUMP);
7558 rebuild_jump_labels (get_insns ());
7559 cleanup_cfg (CLEANUP_CFG_CHANGED);
7560 timevar_pop (TV_JUMP);
7562 else if (tem == 1)
7563 cleanup_cfg (0);
7565 cse_not_expected = 1;
7566 return 0;
7570 namespace {
7572 const pass_data pass_data_cse2 =
7574 RTL_PASS, /* type */
7575 "cse2", /* name */
7576 OPTGROUP_NONE, /* optinfo_flags */
7577 TV_CSE2, /* tv_id */
7578 0, /* properties_required */
7579 0, /* properties_provided */
7580 0, /* properties_destroyed */
7581 0, /* todo_flags_start */
7582 TODO_df_finish, /* todo_flags_finish */
7585 class pass_cse2 : public rtl_opt_pass
7587 public:
7588 pass_cse2 (gcc::context *ctxt)
7589 : rtl_opt_pass (pass_data_cse2, ctxt)
7592 /* opt_pass methods: */
7593 virtual bool gate (function *)
7595 return optimize > 0 && flag_rerun_cse_after_loop;
7598 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7600 }; // class pass_cse2
7602 } // anon namespace
7604 rtl_opt_pass *
7605 make_pass_cse2 (gcc::context *ctxt)
7607 return new pass_cse2 (ctxt);
7610 /* Run second CSE pass after loop optimizations. */
7611 static unsigned int
7612 rest_of_handle_cse_after_global_opts (void)
7614 int save_cfj;
7615 int tem;
7617 /* We only want to do local CSE, so don't follow jumps. */
7618 save_cfj = flag_cse_follow_jumps;
7619 flag_cse_follow_jumps = 0;
7621 rebuild_jump_labels (get_insns ());
7622 tem = cse_main (get_insns (), max_reg_num ());
7623 purge_all_dead_edges ();
7624 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7626 cse_not_expected = !flag_rerun_cse_after_loop;
7628 /* If cse altered any jumps, rerun jump opts to clean things up. */
7629 if (tem == 2)
7631 timevar_push (TV_JUMP);
7632 rebuild_jump_labels (get_insns ());
7633 cleanup_cfg (CLEANUP_CFG_CHANGED);
7634 timevar_pop (TV_JUMP);
7636 else if (tem == 1)
7637 cleanup_cfg (0);
7639 flag_cse_follow_jumps = save_cfj;
7640 return 0;
7643 namespace {
7645 const pass_data pass_data_cse_after_global_opts =
7647 RTL_PASS, /* type */
7648 "cse_local", /* name */
7649 OPTGROUP_NONE, /* optinfo_flags */
7650 TV_CSE, /* tv_id */
7651 0, /* properties_required */
7652 0, /* properties_provided */
7653 0, /* properties_destroyed */
7654 0, /* todo_flags_start */
7655 TODO_df_finish, /* todo_flags_finish */
7658 class pass_cse_after_global_opts : public rtl_opt_pass
7660 public:
7661 pass_cse_after_global_opts (gcc::context *ctxt)
7662 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7665 /* opt_pass methods: */
7666 virtual bool gate (function *)
7668 return optimize > 0 && flag_rerun_cse_after_global_opts;
7671 virtual unsigned int execute (function *)
7673 return rest_of_handle_cse_after_global_opts ();
7676 }; // class pass_cse_after_global_opts
7678 } // anon namespace
7680 rtl_opt_pass *
7681 make_pass_cse_after_global_opts (gcc::context *ctxt)
7683 return new pass_cse_after_global_opts (ctxt);