2015-09-29 Steven G. Kargl <kargl@gcc.gnu.org>
[official-gcc.git] / gcc / cse.c
blobef375791d4893602a83358b3e403846ee4cb02e6
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 "backend.h"
24 #include "cfghooks.h"
25 #include "tree.h"
26 #include "rtl.h"
27 #include "df.h"
28 #include "tm_p.h"
29 #include "regs.h"
30 #include "cfgrtl.h"
31 #include "cfganal.h"
32 #include "cfgcleanup.h"
33 #include "flags.h"
34 #include "insn-config.h"
35 #include "recog.h"
36 #include "alias.h"
37 #include "expmed.h"
38 #include "dojump.h"
39 #include "explow.h"
40 #include "calls.h"
41 #include "emit-rtl.h"
42 #include "varasm.h"
43 #include "stmt.h"
44 #include "expr.h"
45 #include "diagnostic-core.h"
46 #include "toplev.h"
47 #include "except.h"
48 #include "target.h"
49 #include "params.h"
50 #include "rtlhooks-def.h"
51 #include "tree-pass.h"
52 #include "dbgcnt.h"
53 #include "rtl-iter.h"
55 #ifndef LOAD_EXTEND_OP
56 #define LOAD_EXTEND_OP(M) UNKNOWN
57 #endif
59 /* The basic idea of common subexpression elimination is to go
60 through the code, keeping a record of expressions that would
61 have the same value at the current scan point, and replacing
62 expressions encountered with the cheapest equivalent expression.
64 It is too complicated to keep track of the different possibilities
65 when control paths merge in this code; so, at each label, we forget all
66 that is known and start fresh. This can be described as processing each
67 extended basic block separately. We have a separate pass to perform
68 global CSE.
70 Note CSE can turn a conditional or computed jump into a nop or
71 an unconditional jump. When this occurs we arrange to run the jump
72 optimizer after CSE to delete the unreachable code.
74 We use two data structures to record the equivalent expressions:
75 a hash table for most expressions, and a vector of "quantity
76 numbers" to record equivalent (pseudo) registers.
78 The use of the special data structure for registers is desirable
79 because it is faster. It is possible because registers references
80 contain a fairly small number, the register number, taken from
81 a contiguously allocated series, and two register references are
82 identical if they have the same number. General expressions
83 do not have any such thing, so the only way to retrieve the
84 information recorded on an expression other than a register
85 is to keep it in a hash table.
87 Registers and "quantity numbers":
89 At the start of each basic block, all of the (hardware and pseudo)
90 registers used in the function are given distinct quantity
91 numbers to indicate their contents. During scan, when the code
92 copies one register into another, we copy the quantity number.
93 When a register is loaded in any other way, we allocate a new
94 quantity number to describe the value generated by this operation.
95 `REG_QTY (N)' records what quantity register N is currently thought
96 of as containing.
98 All real quantity numbers are greater than or equal to zero.
99 If register N has not been assigned a quantity, `REG_QTY (N)' will
100 equal -N - 1, which is always negative.
102 Quantity numbers below zero do not exist and none of the `qty_table'
103 entries should be referenced with a negative index.
105 We also maintain a bidirectional chain of registers for each
106 quantity number. The `qty_table` members `first_reg' and `last_reg',
107 and `reg_eqv_table' members `next' and `prev' hold these chains.
109 The first register in a chain is the one whose lifespan is least local.
110 Among equals, it is the one that was seen first.
111 We replace any equivalent register with that one.
113 If two registers have the same quantity number, it must be true that
114 REG expressions with qty_table `mode' must be in the hash table for both
115 registers and must be in the same class.
117 The converse is not true. Since hard registers may be referenced in
118 any mode, two REG expressions might be equivalent in the hash table
119 but not have the same quantity number if the quantity number of one
120 of the registers is not the same mode as those expressions.
122 Constants and quantity numbers
124 When a quantity has a known constant value, that value is stored
125 in the appropriate qty_table `const_rtx'. This is in addition to
126 putting the constant in the hash table as is usual for non-regs.
128 Whether a reg or a constant is preferred is determined by the configuration
129 macro CONST_COSTS and will often depend on the constant value. In any
130 event, expressions containing constants can be simplified, by fold_rtx.
132 When a quantity has a known nearly constant value (such as an address
133 of a stack slot), that value is stored in the appropriate qty_table
134 `const_rtx'.
136 Integer constants don't have a machine mode. However, cse
137 determines the intended machine mode from the destination
138 of the instruction that moves the constant. The machine mode
139 is recorded in the hash table along with the actual RTL
140 constant expression so that different modes are kept separate.
142 Other expressions:
144 To record known equivalences among expressions in general
145 we use a hash table called `table'. It has a fixed number of buckets
146 that contain chains of `struct table_elt' elements for expressions.
147 These chains connect the elements whose expressions have the same
148 hash codes.
150 Other chains through the same elements connect the elements which
151 currently have equivalent values.
153 Register references in an expression are canonicalized before hashing
154 the expression. This is done using `reg_qty' and qty_table `first_reg'.
155 The hash code of a register reference is computed using the quantity
156 number, not the register number.
158 When the value of an expression changes, it is necessary to remove from the
159 hash table not just that expression but all expressions whose values
160 could be different as a result.
162 1. If the value changing is in memory, except in special cases
163 ANYTHING referring to memory could be changed. That is because
164 nobody knows where a pointer does not point.
165 The function `invalidate_memory' removes what is necessary.
167 The special cases are when the address is constant or is
168 a constant plus a fixed register such as the frame pointer
169 or a static chain pointer. When such addresses are stored in,
170 we can tell exactly which other such addresses must be invalidated
171 due to overlap. `invalidate' does this.
172 All expressions that refer to non-constant
173 memory addresses are also invalidated. `invalidate_memory' does this.
175 2. If the value changing is a register, all expressions
176 containing references to that register, and only those,
177 must be removed.
179 Because searching the entire hash table for expressions that contain
180 a register is very slow, we try to figure out when it isn't necessary.
181 Precisely, this is necessary only when expressions have been
182 entered in the hash table using this register, and then the value has
183 changed, and then another expression wants to be added to refer to
184 the register's new value. This sequence of circumstances is rare
185 within any one basic block.
187 `REG_TICK' and `REG_IN_TABLE', accessors for members of
188 cse_reg_info, are used to detect this case. REG_TICK (i) is
189 incremented whenever a value is stored in register i.
190 REG_IN_TABLE (i) holds -1 if no references to register i have been
191 entered in the table; otherwise, it contains the value REG_TICK (i)
192 had when the references were entered. If we want to enter a
193 reference and REG_IN_TABLE (i) != REG_TICK (i), we must scan and
194 remove old references. Until we want to enter a new entry, the
195 mere fact that the two vectors don't match makes the entries be
196 ignored if anyone tries to match them.
198 Registers themselves are entered in the hash table as well as in
199 the equivalent-register chains. However, `REG_TICK' and
200 `REG_IN_TABLE' do not apply to expressions which are simple
201 register references. These expressions are removed from the table
202 immediately when they become invalid, and this can be done even if
203 we do not immediately search for all the expressions that refer to
204 the register.
206 A CLOBBER rtx in an instruction invalidates its operand for further
207 reuse. A CLOBBER or SET rtx whose operand is a MEM:BLK
208 invalidates everything that resides in memory.
210 Related expressions:
212 Constant expressions that differ only by an additive integer
213 are called related. When a constant expression is put in
214 the table, the related expression with no constant term
215 is also entered. These are made to point at each other
216 so that it is possible to find out if there exists any
217 register equivalent to an expression related to a given expression. */
219 /* Length of qty_table vector. We know in advance we will not need
220 a quantity number this big. */
222 static int max_qty;
224 /* Next quantity number to be allocated.
225 This is 1 + the largest number needed so far. */
227 static int next_qty;
229 /* Per-qty information tracking.
231 `first_reg' and `last_reg' track the head and tail of the
232 chain of registers which currently contain this quantity.
234 `mode' contains the machine mode of this quantity.
236 `const_rtx' holds the rtx of the constant value of this
237 quantity, if known. A summations of the frame/arg pointer
238 and a constant can also be entered here. When this holds
239 a known value, `const_insn' is the insn which stored the
240 constant value.
242 `comparison_{code,const,qty}' are used to track when a
243 comparison between a quantity and some constant or register has
244 been passed. In such a case, we know the results of the comparison
245 in case we see it again. These members record a comparison that
246 is known to be true. `comparison_code' holds the rtx code of such
247 a comparison, else it is set to UNKNOWN and the other two
248 comparison members are undefined. `comparison_const' holds
249 the constant being compared against, or zero if the comparison
250 is not against a constant. `comparison_qty' holds the quantity
251 being compared against when the result is known. If the comparison
252 is not with a register, `comparison_qty' is -1. */
254 struct qty_table_elem
256 rtx const_rtx;
257 rtx_insn *const_insn;
258 rtx comparison_const;
259 int comparison_qty;
260 unsigned int first_reg, last_reg;
261 /* The sizes of these fields should match the sizes of the
262 code and mode fields of struct rtx_def (see rtl.h). */
263 ENUM_BITFIELD(rtx_code) comparison_code : 16;
264 ENUM_BITFIELD(machine_mode) mode : 8;
267 /* The table of all qtys, indexed by qty number. */
268 static struct qty_table_elem *qty_table;
270 /* For machines that have a CC0, we do not record its value in the hash
271 table since its use is guaranteed to be the insn immediately following
272 its definition and any other insn is presumed to invalidate it.
274 Instead, we store below the current and last value assigned to CC0.
275 If it should happen to be a constant, it is stored in preference
276 to the actual assigned value. In case it is a constant, we store
277 the mode in which the constant should be interpreted. */
279 static rtx this_insn_cc0, prev_insn_cc0;
280 static machine_mode this_insn_cc0_mode, prev_insn_cc0_mode;
282 /* Insn being scanned. */
284 static rtx_insn *this_insn;
285 static bool optimize_this_for_speed_p;
287 /* Index by register number, gives the number of the next (or
288 previous) register in the chain of registers sharing the same
289 value.
291 Or -1 if this register is at the end of the chain.
293 If REG_QTY (N) == -N - 1, reg_eqv_table[N].next is undefined. */
295 /* Per-register equivalence chain. */
296 struct reg_eqv_elem
298 int next, prev;
301 /* The table of all register equivalence chains. */
302 static struct reg_eqv_elem *reg_eqv_table;
304 struct cse_reg_info
306 /* The timestamp at which this register is initialized. */
307 unsigned int timestamp;
309 /* The quantity number of the register's current contents. */
310 int reg_qty;
312 /* The number of times the register has been altered in the current
313 basic block. */
314 int reg_tick;
316 /* The REG_TICK value at which rtx's containing this register are
317 valid in the hash table. If this does not equal the current
318 reg_tick value, such expressions existing in the hash table are
319 invalid. */
320 int reg_in_table;
322 /* The SUBREG that was set when REG_TICK was last incremented. Set
323 to -1 if the last store was to the whole register, not a subreg. */
324 unsigned int subreg_ticked;
327 /* A table of cse_reg_info indexed by register numbers. */
328 static struct cse_reg_info *cse_reg_info_table;
330 /* The size of the above table. */
331 static unsigned int cse_reg_info_table_size;
333 /* The index of the first entry that has not been initialized. */
334 static unsigned int cse_reg_info_table_first_uninitialized;
336 /* The timestamp at the beginning of the current run of
337 cse_extended_basic_block. We increment this variable at the beginning of
338 the current run of cse_extended_basic_block. The timestamp field of a
339 cse_reg_info entry matches the value of this variable if and only
340 if the entry has been initialized during the current run of
341 cse_extended_basic_block. */
342 static unsigned int cse_reg_info_timestamp;
344 /* A HARD_REG_SET containing all the hard registers for which there is
345 currently a REG expression in the hash table. Note the difference
346 from the above variables, which indicate if the REG is mentioned in some
347 expression in the table. */
349 static HARD_REG_SET hard_regs_in_table;
351 /* True if CSE has altered the CFG. */
352 static bool cse_cfg_altered;
354 /* True if CSE has altered conditional jump insns in such a way
355 that jump optimization should be redone. */
356 static bool cse_jumps_altered;
358 /* True if we put a LABEL_REF into the hash table for an INSN
359 without a REG_LABEL_OPERAND, we have to rerun jump after CSE
360 to put in the note. */
361 static bool recorded_label_ref;
363 /* canon_hash stores 1 in do_not_record
364 if it notices a reference to CC0, PC, or some other volatile
365 subexpression. */
367 static int do_not_record;
369 /* canon_hash stores 1 in hash_arg_in_memory
370 if it notices a reference to memory within the expression being hashed. */
372 static int hash_arg_in_memory;
374 /* The hash table contains buckets which are chains of `struct table_elt's,
375 each recording one expression's information.
376 That expression is in the `exp' field.
378 The canon_exp field contains a canonical (from the point of view of
379 alias analysis) version of the `exp' field.
381 Those elements with the same hash code are chained in both directions
382 through the `next_same_hash' and `prev_same_hash' fields.
384 Each set of expressions with equivalent values
385 are on a two-way chain through the `next_same_value'
386 and `prev_same_value' fields, and all point with
387 the `first_same_value' field at the first element in
388 that chain. The chain is in order of increasing cost.
389 Each element's cost value is in its `cost' field.
391 The `in_memory' field is nonzero for elements that
392 involve any reference to memory. These elements are removed
393 whenever a write is done to an unidentified location in memory.
394 To be safe, we assume that a memory address is unidentified unless
395 the address is either a symbol constant or a constant plus
396 the frame pointer or argument pointer.
398 The `related_value' field is used to connect related expressions
399 (that differ by adding an integer).
400 The related expressions are chained in a circular fashion.
401 `related_value' is zero for expressions for which this
402 chain is not useful.
404 The `cost' field stores the cost of this element's expression.
405 The `regcost' field stores the value returned by approx_reg_cost for
406 this element's expression.
408 The `is_const' flag is set if the element is a constant (including
409 a fixed address).
411 The `flag' field is used as a temporary during some search routines.
413 The `mode' field is usually the same as GET_MODE (`exp'), but
414 if `exp' is a CONST_INT and has no machine mode then the `mode'
415 field is the mode it was being used as. Each constant is
416 recorded separately for each mode it is used with. */
418 struct table_elt
420 rtx exp;
421 rtx canon_exp;
422 struct table_elt *next_same_hash;
423 struct table_elt *prev_same_hash;
424 struct table_elt *next_same_value;
425 struct table_elt *prev_same_value;
426 struct table_elt *first_same_value;
427 struct table_elt *related_value;
428 int cost;
429 int regcost;
430 /* The size of this field should match the size
431 of the mode field of struct rtx_def (see rtl.h). */
432 ENUM_BITFIELD(machine_mode) mode : 8;
433 char in_memory;
434 char is_const;
435 char flag;
438 /* We don't want a lot of buckets, because we rarely have very many
439 things stored in the hash table, and a lot of buckets slows
440 down a lot of loops that happen frequently. */
441 #define HASH_SHIFT 5
442 #define HASH_SIZE (1 << HASH_SHIFT)
443 #define HASH_MASK (HASH_SIZE - 1)
445 /* Compute hash code of X in mode M. Special-case case where X is a pseudo
446 register (hard registers may require `do_not_record' to be set). */
448 #define HASH(X, M) \
449 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
450 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
451 : canon_hash (X, M)) & HASH_MASK)
453 /* Like HASH, but without side-effects. */
454 #define SAFE_HASH(X, M) \
455 ((REG_P (X) && REGNO (X) >= FIRST_PSEUDO_REGISTER \
456 ? (((unsigned) REG << 7) + (unsigned) REG_QTY (REGNO (X))) \
457 : safe_hash (X, M)) & HASH_MASK)
459 /* Determine whether register number N is considered a fixed register for the
460 purpose of approximating register costs.
461 It is desirable to replace other regs with fixed regs, to reduce need for
462 non-fixed hard regs.
463 A reg wins if it is either the frame pointer or designated as fixed. */
464 #define FIXED_REGNO_P(N) \
465 ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
466 || fixed_regs[N] || global_regs[N])
468 /* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
469 hard registers and pointers into the frame are the cheapest with a cost
470 of 0. Next come pseudos with a cost of one and other hard registers with
471 a cost of 2. Aside from these special cases, call `rtx_cost'. */
473 #define CHEAP_REGNO(N) \
474 (REGNO_PTR_FRAME_P (N) \
475 || (HARD_REGISTER_NUM_P (N) \
476 && FIXED_REGNO_P (N) && REGNO_REG_CLASS (N) != NO_REGS))
478 #define COST(X, MODE) \
479 (REG_P (X) ? 0 : notreg_cost (X, MODE, SET, 1))
480 #define COST_IN(X, MODE, OUTER, OPNO) \
481 (REG_P (X) ? 0 : notreg_cost (X, MODE, OUTER, OPNO))
483 /* Get the number of times this register has been updated in this
484 basic block. */
486 #define REG_TICK(N) (get_cse_reg_info (N)->reg_tick)
488 /* Get the point at which REG was recorded in the table. */
490 #define REG_IN_TABLE(N) (get_cse_reg_info (N)->reg_in_table)
492 /* Get the SUBREG set at the last increment to REG_TICK (-1 if not a
493 SUBREG). */
495 #define SUBREG_TICKED(N) (get_cse_reg_info (N)->subreg_ticked)
497 /* Get the quantity number for REG. */
499 #define REG_QTY(N) (get_cse_reg_info (N)->reg_qty)
501 /* Determine if the quantity number for register X represents a valid index
502 into the qty_table. */
504 #define REGNO_QTY_VALID_P(N) (REG_QTY (N) >= 0)
506 /* Compare table_elt X and Y and return true iff X is cheaper than Y. */
508 #define CHEAPER(X, Y) \
509 (preferable ((X)->cost, (X)->regcost, (Y)->cost, (Y)->regcost) < 0)
511 static struct table_elt *table[HASH_SIZE];
513 /* Chain of `struct table_elt's made so far for this function
514 but currently removed from the table. */
516 static struct table_elt *free_element_chain;
518 /* Set to the cost of a constant pool reference if one was found for a
519 symbolic constant. If this was found, it means we should try to
520 convert constants into constant pool entries if they don't fit in
521 the insn. */
523 static int constant_pool_entries_cost;
524 static int constant_pool_entries_regcost;
526 /* Trace a patch through the CFG. */
528 struct branch_path
530 /* The basic block for this path entry. */
531 basic_block bb;
534 /* This data describes a block that will be processed by
535 cse_extended_basic_block. */
537 struct cse_basic_block_data
539 /* Total number of SETs in block. */
540 int nsets;
541 /* Size of current branch path, if any. */
542 int path_size;
543 /* Current path, indicating which basic_blocks will be processed. */
544 struct branch_path *path;
548 /* Pointers to the live in/live out bitmaps for the boundaries of the
549 current EBB. */
550 static bitmap cse_ebb_live_in, cse_ebb_live_out;
552 /* A simple bitmap to track which basic blocks have been visited
553 already as part of an already processed extended basic block. */
554 static sbitmap cse_visited_basic_blocks;
556 static bool fixed_base_plus_p (rtx x);
557 static int notreg_cost (rtx, machine_mode, enum rtx_code, int);
558 static int preferable (int, int, int, int);
559 static void new_basic_block (void);
560 static void make_new_qty (unsigned int, machine_mode);
561 static void make_regs_eqv (unsigned int, unsigned int);
562 static void delete_reg_equiv (unsigned int);
563 static int mention_regs (rtx);
564 static int insert_regs (rtx, struct table_elt *, int);
565 static void remove_from_table (struct table_elt *, unsigned);
566 static void remove_pseudo_from_table (rtx, unsigned);
567 static struct table_elt *lookup (rtx, unsigned, machine_mode);
568 static struct table_elt *lookup_for_remove (rtx, unsigned, machine_mode);
569 static rtx lookup_as_function (rtx, enum rtx_code);
570 static struct table_elt *insert_with_costs (rtx, struct table_elt *, unsigned,
571 machine_mode, int, int);
572 static struct table_elt *insert (rtx, struct table_elt *, unsigned,
573 machine_mode);
574 static void merge_equiv_classes (struct table_elt *, struct table_elt *);
575 static void invalidate (rtx, machine_mode);
576 static void remove_invalid_refs (unsigned int);
577 static void remove_invalid_subreg_refs (unsigned int, unsigned int,
578 machine_mode);
579 static void rehash_using_reg (rtx);
580 static void invalidate_memory (void);
581 static void invalidate_for_call (void);
582 static rtx use_related_value (rtx, struct table_elt *);
584 static inline unsigned canon_hash (rtx, machine_mode);
585 static inline unsigned safe_hash (rtx, machine_mode);
586 static inline unsigned hash_rtx_string (const char *);
588 static rtx canon_reg (rtx, rtx_insn *);
589 static enum rtx_code find_comparison_args (enum rtx_code, rtx *, rtx *,
590 machine_mode *,
591 machine_mode *);
592 static rtx fold_rtx (rtx, rtx_insn *);
593 static rtx equiv_constant (rtx);
594 static void record_jump_equiv (rtx_insn *, bool);
595 static void record_jump_cond (enum rtx_code, machine_mode, rtx, rtx,
596 int);
597 static void cse_insn (rtx_insn *);
598 static void cse_prescan_path (struct cse_basic_block_data *);
599 static void invalidate_from_clobbers (rtx_insn *);
600 static void invalidate_from_sets_and_clobbers (rtx_insn *);
601 static rtx cse_process_notes (rtx, rtx, bool *);
602 static void cse_extended_basic_block (struct cse_basic_block_data *);
603 extern void dump_class (struct table_elt*);
604 static void get_cse_reg_info_1 (unsigned int regno);
605 static struct cse_reg_info * get_cse_reg_info (unsigned int regno);
607 static void flush_hash_table (void);
608 static bool insn_live_p (rtx_insn *, int *);
609 static bool set_live_p (rtx, rtx_insn *, int *);
610 static void cse_change_cc_mode_insn (rtx_insn *, rtx);
611 static void cse_change_cc_mode_insns (rtx_insn *, rtx_insn *, rtx);
612 static machine_mode cse_cc_succs (basic_block, basic_block, rtx, rtx,
613 bool);
616 #undef RTL_HOOKS_GEN_LOWPART
617 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_if_possible
619 static const struct rtl_hooks cse_rtl_hooks = RTL_HOOKS_INITIALIZER;
621 /* Nonzero if X has the form (PLUS frame-pointer integer). */
623 static bool
624 fixed_base_plus_p (rtx x)
626 switch (GET_CODE (x))
628 case REG:
629 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx)
630 return true;
631 if (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM])
632 return true;
633 return false;
635 case PLUS:
636 if (!CONST_INT_P (XEXP (x, 1)))
637 return false;
638 return fixed_base_plus_p (XEXP (x, 0));
640 default:
641 return false;
645 /* Dump the expressions in the equivalence class indicated by CLASSP.
646 This function is used only for debugging. */
647 DEBUG_FUNCTION void
648 dump_class (struct table_elt *classp)
650 struct table_elt *elt;
652 fprintf (stderr, "Equivalence chain for ");
653 print_rtl (stderr, classp->exp);
654 fprintf (stderr, ": \n");
656 for (elt = classp->first_same_value; elt; elt = elt->next_same_value)
658 print_rtl (stderr, elt->exp);
659 fprintf (stderr, "\n");
663 /* Return an estimate of the cost of the registers used in an rtx.
664 This is mostly the number of different REG expressions in the rtx;
665 however for some exceptions like fixed registers we use a cost of
666 0. If any other hard register reference occurs, return MAX_COST. */
668 static int
669 approx_reg_cost (const_rtx x)
671 int cost = 0;
672 subrtx_iterator::array_type array;
673 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
675 const_rtx x = *iter;
676 if (REG_P (x))
678 unsigned int regno = REGNO (x);
679 if (!CHEAP_REGNO (regno))
681 if (regno < FIRST_PSEUDO_REGISTER)
683 if (targetm.small_register_classes_for_mode_p (GET_MODE (x)))
684 return MAX_COST;
685 cost += 2;
687 else
688 cost += 1;
692 return cost;
695 /* Return a negative value if an rtx A, whose costs are given by COST_A
696 and REGCOST_A, is more desirable than an rtx B.
697 Return a positive value if A is less desirable, or 0 if the two are
698 equally good. */
699 static int
700 preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
702 /* First, get rid of cases involving expressions that are entirely
703 unwanted. */
704 if (cost_a != cost_b)
706 if (cost_a == MAX_COST)
707 return 1;
708 if (cost_b == MAX_COST)
709 return -1;
712 /* Avoid extending lifetimes of hardregs. */
713 if (regcost_a != regcost_b)
715 if (regcost_a == MAX_COST)
716 return 1;
717 if (regcost_b == MAX_COST)
718 return -1;
721 /* Normal operation costs take precedence. */
722 if (cost_a != cost_b)
723 return cost_a - cost_b;
724 /* Only if these are identical consider effects on register pressure. */
725 if (regcost_a != regcost_b)
726 return regcost_a - regcost_b;
727 return 0;
730 /* Internal function, to compute cost when X is not a register; called
731 from COST macro to keep it simple. */
733 static int
734 notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
736 return ((GET_CODE (x) == SUBREG
737 && REG_P (SUBREG_REG (x))
738 && GET_MODE_CLASS (mode) == MODE_INT
739 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_INT
740 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
741 && subreg_lowpart_p (x)
742 && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (SUBREG_REG (x))))
744 : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
748 /* Initialize CSE_REG_INFO_TABLE. */
750 static void
751 init_cse_reg_info (unsigned int nregs)
753 /* Do we need to grow the table? */
754 if (nregs > cse_reg_info_table_size)
756 unsigned int new_size;
758 if (cse_reg_info_table_size < 2048)
760 /* Compute a new size that is a power of 2 and no smaller
761 than the large of NREGS and 64. */
762 new_size = (cse_reg_info_table_size
763 ? cse_reg_info_table_size : 64);
765 while (new_size < nregs)
766 new_size *= 2;
768 else
770 /* If we need a big table, allocate just enough to hold
771 NREGS registers. */
772 new_size = nregs;
775 /* Reallocate the table with NEW_SIZE entries. */
776 free (cse_reg_info_table);
777 cse_reg_info_table = XNEWVEC (struct cse_reg_info, new_size);
778 cse_reg_info_table_size = new_size;
779 cse_reg_info_table_first_uninitialized = 0;
782 /* Do we have all of the first NREGS entries initialized? */
783 if (cse_reg_info_table_first_uninitialized < nregs)
785 unsigned int old_timestamp = cse_reg_info_timestamp - 1;
786 unsigned int i;
788 /* Put the old timestamp on newly allocated entries so that they
789 will all be considered out of date. We do not touch those
790 entries beyond the first NREGS entries to be nice to the
791 virtual memory. */
792 for (i = cse_reg_info_table_first_uninitialized; i < nregs; i++)
793 cse_reg_info_table[i].timestamp = old_timestamp;
795 cse_reg_info_table_first_uninitialized = nregs;
799 /* Given REGNO, initialize the cse_reg_info entry for REGNO. */
801 static void
802 get_cse_reg_info_1 (unsigned int regno)
804 /* Set TIMESTAMP field to CSE_REG_INFO_TIMESTAMP so that this
805 entry will be considered to have been initialized. */
806 cse_reg_info_table[regno].timestamp = cse_reg_info_timestamp;
808 /* Initialize the rest of the entry. */
809 cse_reg_info_table[regno].reg_tick = 1;
810 cse_reg_info_table[regno].reg_in_table = -1;
811 cse_reg_info_table[regno].subreg_ticked = -1;
812 cse_reg_info_table[regno].reg_qty = -regno - 1;
815 /* Find a cse_reg_info entry for REGNO. */
817 static inline struct cse_reg_info *
818 get_cse_reg_info (unsigned int regno)
820 struct cse_reg_info *p = &cse_reg_info_table[regno];
822 /* If this entry has not been initialized, go ahead and initialize
823 it. */
824 if (p->timestamp != cse_reg_info_timestamp)
825 get_cse_reg_info_1 (regno);
827 return p;
830 /* Clear the hash table and initialize each register with its own quantity,
831 for a new basic block. */
833 static void
834 new_basic_block (void)
836 int i;
838 next_qty = 0;
840 /* Invalidate cse_reg_info_table. */
841 cse_reg_info_timestamp++;
843 /* Clear out hash table state for this pass. */
844 CLEAR_HARD_REG_SET (hard_regs_in_table);
846 /* The per-quantity values used to be initialized here, but it is
847 much faster to initialize each as it is made in `make_new_qty'. */
849 for (i = 0; i < HASH_SIZE; i++)
851 struct table_elt *first;
853 first = table[i];
854 if (first != NULL)
856 struct table_elt *last = first;
858 table[i] = NULL;
860 while (last->next_same_hash != NULL)
861 last = last->next_same_hash;
863 /* Now relink this hash entire chain into
864 the free element list. */
866 last->next_same_hash = free_element_chain;
867 free_element_chain = first;
871 prev_insn_cc0 = 0;
874 /* Say that register REG contains a quantity in mode MODE not in any
875 register before and initialize that quantity. */
877 static void
878 make_new_qty (unsigned int reg, machine_mode mode)
880 int q;
881 struct qty_table_elem *ent;
882 struct reg_eqv_elem *eqv;
884 gcc_assert (next_qty < max_qty);
886 q = REG_QTY (reg) = next_qty++;
887 ent = &qty_table[q];
888 ent->first_reg = reg;
889 ent->last_reg = reg;
890 ent->mode = mode;
891 ent->const_rtx = ent->const_insn = NULL;
892 ent->comparison_code = UNKNOWN;
894 eqv = &reg_eqv_table[reg];
895 eqv->next = eqv->prev = -1;
898 /* Make reg NEW equivalent to reg OLD.
899 OLD is not changing; NEW is. */
901 static void
902 make_regs_eqv (unsigned int new_reg, unsigned int old_reg)
904 unsigned int lastr, firstr;
905 int q = REG_QTY (old_reg);
906 struct qty_table_elem *ent;
908 ent = &qty_table[q];
910 /* Nothing should become eqv until it has a "non-invalid" qty number. */
911 gcc_assert (REGNO_QTY_VALID_P (old_reg));
913 REG_QTY (new_reg) = q;
914 firstr = ent->first_reg;
915 lastr = ent->last_reg;
917 /* Prefer fixed hard registers to anything. Prefer pseudo regs to other
918 hard regs. Among pseudos, if NEW will live longer than any other reg
919 of the same qty, and that is beyond the current basic block,
920 make it the new canonical replacement for this qty. */
921 if (! (firstr < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (firstr))
922 /* Certain fixed registers might be of the class NO_REGS. This means
923 that not only can they not be allocated by the compiler, but
924 they cannot be used in substitutions or canonicalizations
925 either. */
926 && (new_reg >= FIRST_PSEUDO_REGISTER || REGNO_REG_CLASS (new_reg) != NO_REGS)
927 && ((new_reg < FIRST_PSEUDO_REGISTER && FIXED_REGNO_P (new_reg))
928 || (new_reg >= FIRST_PSEUDO_REGISTER
929 && (firstr < FIRST_PSEUDO_REGISTER
930 || (bitmap_bit_p (cse_ebb_live_out, new_reg)
931 && !bitmap_bit_p (cse_ebb_live_out, firstr))
932 || (bitmap_bit_p (cse_ebb_live_in, new_reg)
933 && !bitmap_bit_p (cse_ebb_live_in, firstr))))))
935 reg_eqv_table[firstr].prev = new_reg;
936 reg_eqv_table[new_reg].next = firstr;
937 reg_eqv_table[new_reg].prev = -1;
938 ent->first_reg = new_reg;
940 else
942 /* If NEW is a hard reg (known to be non-fixed), insert at end.
943 Otherwise, insert before any non-fixed hard regs that are at the
944 end. Registers of class NO_REGS cannot be used as an
945 equivalent for anything. */
946 while (lastr < FIRST_PSEUDO_REGISTER && reg_eqv_table[lastr].prev >= 0
947 && (REGNO_REG_CLASS (lastr) == NO_REGS || ! FIXED_REGNO_P (lastr))
948 && new_reg >= FIRST_PSEUDO_REGISTER)
949 lastr = reg_eqv_table[lastr].prev;
950 reg_eqv_table[new_reg].next = reg_eqv_table[lastr].next;
951 if (reg_eqv_table[lastr].next >= 0)
952 reg_eqv_table[reg_eqv_table[lastr].next].prev = new_reg;
953 else
954 qty_table[q].last_reg = new_reg;
955 reg_eqv_table[lastr].next = new_reg;
956 reg_eqv_table[new_reg].prev = lastr;
960 /* Remove REG from its equivalence class. */
962 static void
963 delete_reg_equiv (unsigned int reg)
965 struct qty_table_elem *ent;
966 int q = REG_QTY (reg);
967 int p, n;
969 /* If invalid, do nothing. */
970 if (! REGNO_QTY_VALID_P (reg))
971 return;
973 ent = &qty_table[q];
975 p = reg_eqv_table[reg].prev;
976 n = reg_eqv_table[reg].next;
978 if (n != -1)
979 reg_eqv_table[n].prev = p;
980 else
981 ent->last_reg = p;
982 if (p != -1)
983 reg_eqv_table[p].next = n;
984 else
985 ent->first_reg = n;
987 REG_QTY (reg) = -reg - 1;
990 /* Remove any invalid expressions from the hash table
991 that refer to any of the registers contained in expression X.
993 Make sure that newly inserted references to those registers
994 as subexpressions will be considered valid.
996 mention_regs is not called when a register itself
997 is being stored in the table.
999 Return 1 if we have done something that may have changed the hash code
1000 of X. */
1002 static int
1003 mention_regs (rtx x)
1005 enum rtx_code code;
1006 int i, j;
1007 const char *fmt;
1008 int changed = 0;
1010 if (x == 0)
1011 return 0;
1013 code = GET_CODE (x);
1014 if (code == REG)
1016 unsigned int regno = REGNO (x);
1017 unsigned int endregno = END_REGNO (x);
1018 unsigned int i;
1020 for (i = regno; i < endregno; i++)
1022 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1023 remove_invalid_refs (i);
1025 REG_IN_TABLE (i) = REG_TICK (i);
1026 SUBREG_TICKED (i) = -1;
1029 return 0;
1032 /* If this is a SUBREG, we don't want to discard other SUBREGs of the same
1033 pseudo if they don't use overlapping words. We handle only pseudos
1034 here for simplicity. */
1035 if (code == SUBREG && REG_P (SUBREG_REG (x))
1036 && REGNO (SUBREG_REG (x)) >= FIRST_PSEUDO_REGISTER)
1038 unsigned int i = REGNO (SUBREG_REG (x));
1040 if (REG_IN_TABLE (i) >= 0 && REG_IN_TABLE (i) != REG_TICK (i))
1042 /* If REG_IN_TABLE (i) differs from REG_TICK (i) by one, and
1043 the last store to this register really stored into this
1044 subreg, then remove the memory of this subreg.
1045 Otherwise, remove any memory of the entire register and
1046 all its subregs from the table. */
1047 if (REG_TICK (i) - REG_IN_TABLE (i) > 1
1048 || SUBREG_TICKED (i) != REGNO (SUBREG_REG (x)))
1049 remove_invalid_refs (i);
1050 else
1051 remove_invalid_subreg_refs (i, SUBREG_BYTE (x), GET_MODE (x));
1054 REG_IN_TABLE (i) = REG_TICK (i);
1055 SUBREG_TICKED (i) = REGNO (SUBREG_REG (x));
1056 return 0;
1059 /* If X is a comparison or a COMPARE and either operand is a register
1060 that does not have a quantity, give it one. This is so that a later
1061 call to record_jump_equiv won't cause X to be assigned a different
1062 hash code and not found in the table after that call.
1064 It is not necessary to do this here, since rehash_using_reg can
1065 fix up the table later, but doing this here eliminates the need to
1066 call that expensive function in the most common case where the only
1067 use of the register is in the comparison. */
1069 if (code == COMPARE || COMPARISON_P (x))
1071 if (REG_P (XEXP (x, 0))
1072 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 0))))
1073 if (insert_regs (XEXP (x, 0), NULL, 0))
1075 rehash_using_reg (XEXP (x, 0));
1076 changed = 1;
1079 if (REG_P (XEXP (x, 1))
1080 && ! REGNO_QTY_VALID_P (REGNO (XEXP (x, 1))))
1081 if (insert_regs (XEXP (x, 1), NULL, 0))
1083 rehash_using_reg (XEXP (x, 1));
1084 changed = 1;
1088 fmt = GET_RTX_FORMAT (code);
1089 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1090 if (fmt[i] == 'e')
1091 changed |= mention_regs (XEXP (x, i));
1092 else if (fmt[i] == 'E')
1093 for (j = 0; j < XVECLEN (x, i); j++)
1094 changed |= mention_regs (XVECEXP (x, i, j));
1096 return changed;
1099 /* Update the register quantities for inserting X into the hash table
1100 with a value equivalent to CLASSP.
1101 (If the class does not contain a REG, it is irrelevant.)
1102 If MODIFIED is nonzero, X is a destination; it is being modified.
1103 Note that delete_reg_equiv should be called on a register
1104 before insert_regs is done on that register with MODIFIED != 0.
1106 Nonzero value means that elements of reg_qty have changed
1107 so X's hash code may be different. */
1109 static int
1110 insert_regs (rtx x, struct table_elt *classp, int modified)
1112 if (REG_P (x))
1114 unsigned int regno = REGNO (x);
1115 int qty_valid;
1117 /* If REGNO is in the equivalence table already but is of the
1118 wrong mode for that equivalence, don't do anything here. */
1120 qty_valid = REGNO_QTY_VALID_P (regno);
1121 if (qty_valid)
1123 struct qty_table_elem *ent = &qty_table[REG_QTY (regno)];
1125 if (ent->mode != GET_MODE (x))
1126 return 0;
1129 if (modified || ! qty_valid)
1131 if (classp)
1132 for (classp = classp->first_same_value;
1133 classp != 0;
1134 classp = classp->next_same_value)
1135 if (REG_P (classp->exp)
1136 && GET_MODE (classp->exp) == GET_MODE (x))
1138 unsigned c_regno = REGNO (classp->exp);
1140 gcc_assert (REGNO_QTY_VALID_P (c_regno));
1142 /* Suppose that 5 is hard reg and 100 and 101 are
1143 pseudos. Consider
1145 (set (reg:si 100) (reg:si 5))
1146 (set (reg:si 5) (reg:si 100))
1147 (set (reg:di 101) (reg:di 5))
1149 We would now set REG_QTY (101) = REG_QTY (5), but the
1150 entry for 5 is in SImode. When we use this later in
1151 copy propagation, we get the register in wrong mode. */
1152 if (qty_table[REG_QTY (c_regno)].mode != GET_MODE (x))
1153 continue;
1155 make_regs_eqv (regno, c_regno);
1156 return 1;
1159 /* Mention_regs for a SUBREG checks if REG_TICK is exactly one larger
1160 than REG_IN_TABLE to find out if there was only a single preceding
1161 invalidation - for the SUBREG - or another one, which would be
1162 for the full register. However, if we find here that REG_TICK
1163 indicates that the register is invalid, it means that it has
1164 been invalidated in a separate operation. The SUBREG might be used
1165 now (then this is a recursive call), or we might use the full REG
1166 now and a SUBREG of it later. So bump up REG_TICK so that
1167 mention_regs will do the right thing. */
1168 if (! modified
1169 && REG_IN_TABLE (regno) >= 0
1170 && REG_TICK (regno) == REG_IN_TABLE (regno) + 1)
1171 REG_TICK (regno)++;
1172 make_new_qty (regno, GET_MODE (x));
1173 return 1;
1176 return 0;
1179 /* If X is a SUBREG, we will likely be inserting the inner register in the
1180 table. If that register doesn't have an assigned quantity number at
1181 this point but does later, the insertion that we will be doing now will
1182 not be accessible because its hash code will have changed. So assign
1183 a quantity number now. */
1185 else if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
1186 && ! REGNO_QTY_VALID_P (REGNO (SUBREG_REG (x))))
1188 insert_regs (SUBREG_REG (x), NULL, 0);
1189 mention_regs (x);
1190 return 1;
1192 else
1193 return mention_regs (x);
1197 /* Compute upper and lower anchors for CST. Also compute the offset of CST
1198 from these anchors/bases such that *_BASE + *_OFFS = CST. Return false iff
1199 CST is equal to an anchor. */
1201 static bool
1202 compute_const_anchors (rtx cst,
1203 HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
1204 HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
1206 HOST_WIDE_INT n = INTVAL (cst);
1208 *lower_base = n & ~(targetm.const_anchor - 1);
1209 if (*lower_base == n)
1210 return false;
1212 *upper_base =
1213 (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
1214 *upper_offs = n - *upper_base;
1215 *lower_offs = n - *lower_base;
1216 return true;
1219 /* Insert the equivalence between ANCHOR and (REG + OFF) in mode MODE. */
1221 static void
1222 insert_const_anchor (HOST_WIDE_INT anchor, rtx reg, HOST_WIDE_INT offs,
1223 machine_mode mode)
1225 struct table_elt *elt;
1226 unsigned hash;
1227 rtx anchor_exp;
1228 rtx exp;
1230 anchor_exp = GEN_INT (anchor);
1231 hash = HASH (anchor_exp, mode);
1232 elt = lookup (anchor_exp, hash, mode);
1233 if (!elt)
1234 elt = insert (anchor_exp, NULL, hash, mode);
1236 exp = plus_constant (mode, reg, offs);
1237 /* REG has just been inserted and the hash codes recomputed. */
1238 mention_regs (exp);
1239 hash = HASH (exp, mode);
1241 /* Use the cost of the register rather than the whole expression. When
1242 looking up constant anchors we will further offset the corresponding
1243 expression therefore it does not make sense to prefer REGs over
1244 reg-immediate additions. Prefer instead the oldest expression. Also
1245 don't prefer pseudos over hard regs so that we derive constants in
1246 argument registers from other argument registers rather than from the
1247 original pseudo that was used to synthesize the constant. */
1248 insert_with_costs (exp, elt, hash, mode, COST (reg, mode), 1);
1251 /* The constant CST is equivalent to the register REG. Create
1252 equivalences between the two anchors of CST and the corresponding
1253 register-offset expressions using REG. */
1255 static void
1256 insert_const_anchors (rtx reg, rtx cst, machine_mode mode)
1258 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1260 if (!compute_const_anchors (cst, &lower_base, &lower_offs,
1261 &upper_base, &upper_offs))
1262 return;
1264 /* Ignore anchors of value 0. Constants accessible from zero are
1265 simple. */
1266 if (lower_base != 0)
1267 insert_const_anchor (lower_base, reg, -lower_offs, mode);
1269 if (upper_base != 0)
1270 insert_const_anchor (upper_base, reg, -upper_offs, mode);
1273 /* We need to express ANCHOR_ELT->exp + OFFS. Walk the equivalence list of
1274 ANCHOR_ELT and see if offsetting any of the entries by OFFS would create a
1275 valid expression. Return the cheapest and oldest of such expressions. In
1276 *OLD, return how old the resulting expression is compared to the other
1277 equivalent expressions. */
1279 static rtx
1280 find_reg_offset_for_const (struct table_elt *anchor_elt, HOST_WIDE_INT offs,
1281 unsigned *old)
1283 struct table_elt *elt;
1284 unsigned idx;
1285 struct table_elt *match_elt;
1286 rtx match;
1288 /* Find the cheapest and *oldest* expression to maximize the chance of
1289 reusing the same pseudo. */
1291 match_elt = NULL;
1292 match = NULL_RTX;
1293 for (elt = anchor_elt->first_same_value, idx = 0;
1294 elt;
1295 elt = elt->next_same_value, idx++)
1297 if (match_elt && CHEAPER (match_elt, elt))
1298 return match;
1300 if (REG_P (elt->exp)
1301 || (GET_CODE (elt->exp) == PLUS
1302 && REG_P (XEXP (elt->exp, 0))
1303 && GET_CODE (XEXP (elt->exp, 1)) == CONST_INT))
1305 rtx x;
1307 /* Ignore expressions that are no longer valid. */
1308 if (!REG_P (elt->exp) && !exp_equiv_p (elt->exp, elt->exp, 1, false))
1309 continue;
1311 x = plus_constant (GET_MODE (elt->exp), elt->exp, offs);
1312 if (REG_P (x)
1313 || (GET_CODE (x) == PLUS
1314 && IN_RANGE (INTVAL (XEXP (x, 1)),
1315 -targetm.const_anchor,
1316 targetm.const_anchor - 1)))
1318 match = x;
1319 match_elt = elt;
1320 *old = idx;
1325 return match;
1328 /* Try to express the constant SRC_CONST using a register+offset expression
1329 derived from a constant anchor. Return it if successful or NULL_RTX,
1330 otherwise. */
1332 static rtx
1333 try_const_anchors (rtx src_const, machine_mode mode)
1335 struct table_elt *lower_elt, *upper_elt;
1336 HOST_WIDE_INT lower_base, lower_offs, upper_base, upper_offs;
1337 rtx lower_anchor_rtx, upper_anchor_rtx;
1338 rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
1339 unsigned lower_old, upper_old;
1341 /* CONST_INT is used for CC modes, but we should leave those alone. */
1342 if (GET_MODE_CLASS (mode) == MODE_CC)
1343 return NULL_RTX;
1345 gcc_assert (SCALAR_INT_MODE_P (mode));
1346 if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
1347 &upper_base, &upper_offs))
1348 return NULL_RTX;
1350 lower_anchor_rtx = GEN_INT (lower_base);
1351 upper_anchor_rtx = GEN_INT (upper_base);
1352 lower_elt = lookup (lower_anchor_rtx, HASH (lower_anchor_rtx, mode), mode);
1353 upper_elt = lookup (upper_anchor_rtx, HASH (upper_anchor_rtx, mode), mode);
1355 if (lower_elt)
1356 lower_exp = find_reg_offset_for_const (lower_elt, lower_offs, &lower_old);
1357 if (upper_elt)
1358 upper_exp = find_reg_offset_for_const (upper_elt, upper_offs, &upper_old);
1360 if (!lower_exp)
1361 return upper_exp;
1362 if (!upper_exp)
1363 return lower_exp;
1365 /* Return the older expression. */
1366 return (upper_old > lower_old ? upper_exp : lower_exp);
1369 /* Look in or update the hash table. */
1371 /* Remove table element ELT from use in the table.
1372 HASH is its hash code, made using the HASH macro.
1373 It's an argument because often that is known in advance
1374 and we save much time not recomputing it. */
1376 static void
1377 remove_from_table (struct table_elt *elt, unsigned int hash)
1379 if (elt == 0)
1380 return;
1382 /* Mark this element as removed. See cse_insn. */
1383 elt->first_same_value = 0;
1385 /* Remove the table element from its equivalence class. */
1388 struct table_elt *prev = elt->prev_same_value;
1389 struct table_elt *next = elt->next_same_value;
1391 if (next)
1392 next->prev_same_value = prev;
1394 if (prev)
1395 prev->next_same_value = next;
1396 else
1398 struct table_elt *newfirst = next;
1399 while (next)
1401 next->first_same_value = newfirst;
1402 next = next->next_same_value;
1407 /* Remove the table element from its hash bucket. */
1410 struct table_elt *prev = elt->prev_same_hash;
1411 struct table_elt *next = elt->next_same_hash;
1413 if (next)
1414 next->prev_same_hash = prev;
1416 if (prev)
1417 prev->next_same_hash = next;
1418 else if (table[hash] == elt)
1419 table[hash] = next;
1420 else
1422 /* This entry is not in the proper hash bucket. This can happen
1423 when two classes were merged by `merge_equiv_classes'. Search
1424 for the hash bucket that it heads. This happens only very
1425 rarely, so the cost is acceptable. */
1426 for (hash = 0; hash < HASH_SIZE; hash++)
1427 if (table[hash] == elt)
1428 table[hash] = next;
1432 /* Remove the table element from its related-value circular chain. */
1434 if (elt->related_value != 0 && elt->related_value != elt)
1436 struct table_elt *p = elt->related_value;
1438 while (p->related_value != elt)
1439 p = p->related_value;
1440 p->related_value = elt->related_value;
1441 if (p->related_value == p)
1442 p->related_value = 0;
1445 /* Now add it to the free element chain. */
1446 elt->next_same_hash = free_element_chain;
1447 free_element_chain = elt;
1450 /* Same as above, but X is a pseudo-register. */
1452 static void
1453 remove_pseudo_from_table (rtx x, unsigned int hash)
1455 struct table_elt *elt;
1457 /* Because a pseudo-register can be referenced in more than one
1458 mode, we might have to remove more than one table entry. */
1459 while ((elt = lookup_for_remove (x, hash, VOIDmode)))
1460 remove_from_table (elt, hash);
1463 /* Look up X in the hash table and return its table element,
1464 or 0 if X is not in the table.
1466 MODE is the machine-mode of X, or if X is an integer constant
1467 with VOIDmode then MODE is the mode with which X will be used.
1469 Here we are satisfied to find an expression whose tree structure
1470 looks like X. */
1472 static struct table_elt *
1473 lookup (rtx x, unsigned int hash, machine_mode mode)
1475 struct table_elt *p;
1477 for (p = table[hash]; p; p = p->next_same_hash)
1478 if (mode == p->mode && ((x == p->exp && REG_P (x))
1479 || exp_equiv_p (x, p->exp, !REG_P (x), false)))
1480 return p;
1482 return 0;
1485 /* Like `lookup' but don't care whether the table element uses invalid regs.
1486 Also ignore discrepancies in the machine mode of a register. */
1488 static struct table_elt *
1489 lookup_for_remove (rtx x, unsigned int hash, machine_mode mode)
1491 struct table_elt *p;
1493 if (REG_P (x))
1495 unsigned int regno = REGNO (x);
1497 /* Don't check the machine mode when comparing registers;
1498 invalidating (REG:SI 0) also invalidates (REG:DF 0). */
1499 for (p = table[hash]; p; p = p->next_same_hash)
1500 if (REG_P (p->exp)
1501 && REGNO (p->exp) == regno)
1502 return p;
1504 else
1506 for (p = table[hash]; p; p = p->next_same_hash)
1507 if (mode == p->mode
1508 && (x == p->exp || exp_equiv_p (x, p->exp, 0, false)))
1509 return p;
1512 return 0;
1515 /* Look for an expression equivalent to X and with code CODE.
1516 If one is found, return that expression. */
1518 static rtx
1519 lookup_as_function (rtx x, enum rtx_code code)
1521 struct table_elt *p
1522 = lookup (x, SAFE_HASH (x, VOIDmode), GET_MODE (x));
1524 if (p == 0)
1525 return 0;
1527 for (p = p->first_same_value; p; p = p->next_same_value)
1528 if (GET_CODE (p->exp) == code
1529 /* Make sure this is a valid entry in the table. */
1530 && exp_equiv_p (p->exp, p->exp, 1, false))
1531 return p->exp;
1533 return 0;
1536 /* Insert X in the hash table, assuming HASH is its hash code and
1537 CLASSP is an element of the class it should go in (or 0 if a new
1538 class should be made). COST is the code of X and reg_cost is the
1539 cost of registers in X. It is inserted at the proper position to
1540 keep the class in the order cheapest first.
1542 MODE is the machine-mode of X, or if X is an integer constant
1543 with VOIDmode then MODE is the mode with which X will be used.
1545 For elements of equal cheapness, the most recent one
1546 goes in front, except that the first element in the list
1547 remains first unless a cheaper element is added. The order of
1548 pseudo-registers does not matter, as canon_reg will be called to
1549 find the cheapest when a register is retrieved from the table.
1551 The in_memory field in the hash table element is set to 0.
1552 The caller must set it nonzero if appropriate.
1554 You should call insert_regs (X, CLASSP, MODIFY) before calling here,
1555 and if insert_regs returns a nonzero value
1556 you must then recompute its hash code before calling here.
1558 If necessary, update table showing constant values of quantities. */
1560 static struct table_elt *
1561 insert_with_costs (rtx x, struct table_elt *classp, unsigned int hash,
1562 machine_mode mode, int cost, int reg_cost)
1564 struct table_elt *elt;
1566 /* If X is a register and we haven't made a quantity for it,
1567 something is wrong. */
1568 gcc_assert (!REG_P (x) || REGNO_QTY_VALID_P (REGNO (x)));
1570 /* If X is a hard register, show it is being put in the table. */
1571 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
1572 add_to_hard_reg_set (&hard_regs_in_table, GET_MODE (x), REGNO (x));
1574 /* Put an element for X into the right hash bucket. */
1576 elt = free_element_chain;
1577 if (elt)
1578 free_element_chain = elt->next_same_hash;
1579 else
1580 elt = XNEW (struct table_elt);
1582 elt->exp = x;
1583 elt->canon_exp = NULL_RTX;
1584 elt->cost = cost;
1585 elt->regcost = reg_cost;
1586 elt->next_same_value = 0;
1587 elt->prev_same_value = 0;
1588 elt->next_same_hash = table[hash];
1589 elt->prev_same_hash = 0;
1590 elt->related_value = 0;
1591 elt->in_memory = 0;
1592 elt->mode = mode;
1593 elt->is_const = (CONSTANT_P (x) || fixed_base_plus_p (x));
1595 if (table[hash])
1596 table[hash]->prev_same_hash = elt;
1597 table[hash] = elt;
1599 /* Put it into the proper value-class. */
1600 if (classp)
1602 classp = classp->first_same_value;
1603 if (CHEAPER (elt, classp))
1604 /* Insert at the head of the class. */
1606 struct table_elt *p;
1607 elt->next_same_value = classp;
1608 classp->prev_same_value = elt;
1609 elt->first_same_value = elt;
1611 for (p = classp; p; p = p->next_same_value)
1612 p->first_same_value = elt;
1614 else
1616 /* Insert not at head of the class. */
1617 /* Put it after the last element cheaper than X. */
1618 struct table_elt *p, *next;
1620 for (p = classp;
1621 (next = p->next_same_value) && CHEAPER (next, elt);
1622 p = next)
1625 /* Put it after P and before NEXT. */
1626 elt->next_same_value = next;
1627 if (next)
1628 next->prev_same_value = elt;
1630 elt->prev_same_value = p;
1631 p->next_same_value = elt;
1632 elt->first_same_value = classp;
1635 else
1636 elt->first_same_value = elt;
1638 /* If this is a constant being set equivalent to a register or a register
1639 being set equivalent to a constant, note the constant equivalence.
1641 If this is a constant, it cannot be equivalent to a different constant,
1642 and a constant is the only thing that can be cheaper than a register. So
1643 we know the register is the head of the class (before the constant was
1644 inserted).
1646 If this is a register that is not already known equivalent to a
1647 constant, we must check the entire class.
1649 If this is a register that is already known equivalent to an insn,
1650 update the qtys `const_insn' to show that `this_insn' is the latest
1651 insn making that quantity equivalent to the constant. */
1653 if (elt->is_const && classp && REG_P (classp->exp)
1654 && !REG_P (x))
1656 int exp_q = REG_QTY (REGNO (classp->exp));
1657 struct qty_table_elem *exp_ent = &qty_table[exp_q];
1659 exp_ent->const_rtx = gen_lowpart (exp_ent->mode, x);
1660 exp_ent->const_insn = this_insn;
1663 else if (REG_P (x)
1664 && classp
1665 && ! qty_table[REG_QTY (REGNO (x))].const_rtx
1666 && ! elt->is_const)
1668 struct table_elt *p;
1670 for (p = classp; p != 0; p = p->next_same_value)
1672 if (p->is_const && !REG_P (p->exp))
1674 int x_q = REG_QTY (REGNO (x));
1675 struct qty_table_elem *x_ent = &qty_table[x_q];
1677 x_ent->const_rtx
1678 = gen_lowpart (GET_MODE (x), p->exp);
1679 x_ent->const_insn = this_insn;
1680 break;
1685 else if (REG_P (x)
1686 && qty_table[REG_QTY (REGNO (x))].const_rtx
1687 && GET_MODE (x) == qty_table[REG_QTY (REGNO (x))].mode)
1688 qty_table[REG_QTY (REGNO (x))].const_insn = this_insn;
1690 /* If this is a constant with symbolic value,
1691 and it has a term with an explicit integer value,
1692 link it up with related expressions. */
1693 if (GET_CODE (x) == CONST)
1695 rtx subexp = get_related_value (x);
1696 unsigned subhash;
1697 struct table_elt *subelt, *subelt_prev;
1699 if (subexp != 0)
1701 /* Get the integer-free subexpression in the hash table. */
1702 subhash = SAFE_HASH (subexp, mode);
1703 subelt = lookup (subexp, subhash, mode);
1704 if (subelt == 0)
1705 subelt = insert (subexp, NULL, subhash, mode);
1706 /* Initialize SUBELT's circular chain if it has none. */
1707 if (subelt->related_value == 0)
1708 subelt->related_value = subelt;
1709 /* Find the element in the circular chain that precedes SUBELT. */
1710 subelt_prev = subelt;
1711 while (subelt_prev->related_value != subelt)
1712 subelt_prev = subelt_prev->related_value;
1713 /* Put new ELT into SUBELT's circular chain just before SUBELT.
1714 This way the element that follows SUBELT is the oldest one. */
1715 elt->related_value = subelt_prev->related_value;
1716 subelt_prev->related_value = elt;
1720 return elt;
1723 /* Wrap insert_with_costs by passing the default costs. */
1725 static struct table_elt *
1726 insert (rtx x, struct table_elt *classp, unsigned int hash,
1727 machine_mode mode)
1729 return insert_with_costs (x, classp, hash, mode,
1730 COST (x, mode), approx_reg_cost (x));
1734 /* Given two equivalence classes, CLASS1 and CLASS2, put all the entries from
1735 CLASS2 into CLASS1. This is done when we have reached an insn which makes
1736 the two classes equivalent.
1738 CLASS1 will be the surviving class; CLASS2 should not be used after this
1739 call.
1741 Any invalid entries in CLASS2 will not be copied. */
1743 static void
1744 merge_equiv_classes (struct table_elt *class1, struct table_elt *class2)
1746 struct table_elt *elt, *next, *new_elt;
1748 /* Ensure we start with the head of the classes. */
1749 class1 = class1->first_same_value;
1750 class2 = class2->first_same_value;
1752 /* If they were already equal, forget it. */
1753 if (class1 == class2)
1754 return;
1756 for (elt = class2; elt; elt = next)
1758 unsigned int hash;
1759 rtx exp = elt->exp;
1760 machine_mode mode = elt->mode;
1762 next = elt->next_same_value;
1764 /* Remove old entry, make a new one in CLASS1's class.
1765 Don't do this for invalid entries as we cannot find their
1766 hash code (it also isn't necessary). */
1767 if (REG_P (exp) || exp_equiv_p (exp, exp, 1, false))
1769 bool need_rehash = false;
1771 hash_arg_in_memory = 0;
1772 hash = HASH (exp, mode);
1774 if (REG_P (exp))
1776 need_rehash = REGNO_QTY_VALID_P (REGNO (exp));
1777 delete_reg_equiv (REGNO (exp));
1780 if (REG_P (exp) && REGNO (exp) >= FIRST_PSEUDO_REGISTER)
1781 remove_pseudo_from_table (exp, hash);
1782 else
1783 remove_from_table (elt, hash);
1785 if (insert_regs (exp, class1, 0) || need_rehash)
1787 rehash_using_reg (exp);
1788 hash = HASH (exp, mode);
1790 new_elt = insert (exp, class1, hash, mode);
1791 new_elt->in_memory = hash_arg_in_memory;
1792 if (GET_CODE (exp) == ASM_OPERANDS && elt->cost == MAX_COST)
1793 new_elt->cost = MAX_COST;
1798 /* Flush the entire hash table. */
1800 static void
1801 flush_hash_table (void)
1803 int i;
1804 struct table_elt *p;
1806 for (i = 0; i < HASH_SIZE; i++)
1807 for (p = table[i]; p; p = table[i])
1809 /* Note that invalidate can remove elements
1810 after P in the current hash chain. */
1811 if (REG_P (p->exp))
1812 invalidate (p->exp, VOIDmode);
1813 else
1814 remove_from_table (p, i);
1818 /* Check whether an anti dependence exists between X and EXP. MODE and
1819 ADDR are as for canon_anti_dependence. */
1821 static bool
1822 check_dependence (const_rtx x, rtx exp, machine_mode mode, rtx addr)
1824 subrtx_iterator::array_type array;
1825 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
1827 const_rtx x = *iter;
1828 if (MEM_P (x) && canon_anti_dependence (x, true, exp, mode, addr))
1829 return true;
1831 return false;
1834 /* Remove from the hash table, or mark as invalid, all expressions whose
1835 values could be altered by storing in X. X is a register, a subreg, or
1836 a memory reference with nonvarying address (because, when a memory
1837 reference with a varying address is stored in, all memory references are
1838 removed by invalidate_memory so specific invalidation is superfluous).
1839 FULL_MODE, if not VOIDmode, indicates that this much should be
1840 invalidated instead of just the amount indicated by the mode of X. This
1841 is only used for bitfield stores into memory.
1843 A nonvarying address may be just a register or just a symbol reference,
1844 or it may be either of those plus a numeric offset. */
1846 static void
1847 invalidate (rtx x, machine_mode full_mode)
1849 int i;
1850 struct table_elt *p;
1851 rtx addr;
1853 switch (GET_CODE (x))
1855 case REG:
1857 /* If X is a register, dependencies on its contents are recorded
1858 through the qty number mechanism. Just change the qty number of
1859 the register, mark it as invalid for expressions that refer to it,
1860 and remove it itself. */
1861 unsigned int regno = REGNO (x);
1862 unsigned int hash = HASH (x, GET_MODE (x));
1864 /* Remove REGNO from any quantity list it might be on and indicate
1865 that its value might have changed. If it is a pseudo, remove its
1866 entry from the hash table.
1868 For a hard register, we do the first two actions above for any
1869 additional hard registers corresponding to X. Then, if any of these
1870 registers are in the table, we must remove any REG entries that
1871 overlap these registers. */
1873 delete_reg_equiv (regno);
1874 REG_TICK (regno)++;
1875 SUBREG_TICKED (regno) = -1;
1877 if (regno >= FIRST_PSEUDO_REGISTER)
1878 remove_pseudo_from_table (x, hash);
1879 else
1881 HOST_WIDE_INT in_table
1882 = TEST_HARD_REG_BIT (hard_regs_in_table, regno);
1883 unsigned int endregno = END_REGNO (x);
1884 unsigned int tregno, tendregno, rn;
1885 struct table_elt *p, *next;
1887 CLEAR_HARD_REG_BIT (hard_regs_in_table, regno);
1889 for (rn = regno + 1; rn < endregno; rn++)
1891 in_table |= TEST_HARD_REG_BIT (hard_regs_in_table, rn);
1892 CLEAR_HARD_REG_BIT (hard_regs_in_table, rn);
1893 delete_reg_equiv (rn);
1894 REG_TICK (rn)++;
1895 SUBREG_TICKED (rn) = -1;
1898 if (in_table)
1899 for (hash = 0; hash < HASH_SIZE; hash++)
1900 for (p = table[hash]; p; p = next)
1902 next = p->next_same_hash;
1904 if (!REG_P (p->exp)
1905 || REGNO (p->exp) >= FIRST_PSEUDO_REGISTER)
1906 continue;
1908 tregno = REGNO (p->exp);
1909 tendregno = END_REGNO (p->exp);
1910 if (tendregno > regno && tregno < endregno)
1911 remove_from_table (p, hash);
1915 return;
1917 case SUBREG:
1918 invalidate (SUBREG_REG (x), VOIDmode);
1919 return;
1921 case PARALLEL:
1922 for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
1923 invalidate (XVECEXP (x, 0, i), VOIDmode);
1924 return;
1926 case EXPR_LIST:
1927 /* This is part of a disjoint return value; extract the location in
1928 question ignoring the offset. */
1929 invalidate (XEXP (x, 0), VOIDmode);
1930 return;
1932 case MEM:
1933 addr = canon_rtx (get_addr (XEXP (x, 0)));
1934 /* Calculate the canonical version of X here so that
1935 true_dependence doesn't generate new RTL for X on each call. */
1936 x = canon_rtx (x);
1938 /* Remove all hash table elements that refer to overlapping pieces of
1939 memory. */
1940 if (full_mode == VOIDmode)
1941 full_mode = GET_MODE (x);
1943 for (i = 0; i < HASH_SIZE; i++)
1945 struct table_elt *next;
1947 for (p = table[i]; p; p = next)
1949 next = p->next_same_hash;
1950 if (p->in_memory)
1952 /* Just canonicalize the expression once;
1953 otherwise each time we call invalidate
1954 true_dependence will canonicalize the
1955 expression again. */
1956 if (!p->canon_exp)
1957 p->canon_exp = canon_rtx (p->exp);
1958 if (check_dependence (p->canon_exp, x, full_mode, addr))
1959 remove_from_table (p, i);
1963 return;
1965 default:
1966 gcc_unreachable ();
1970 /* Invalidate DEST. Used when DEST is not going to be added
1971 into the hash table for some reason, e.g. do_not_record
1972 flagged on it. */
1974 static void
1975 invalidate_dest (rtx dest)
1977 if (REG_P (dest)
1978 || GET_CODE (dest) == SUBREG
1979 || MEM_P (dest))
1980 invalidate (dest, VOIDmode);
1981 else if (GET_CODE (dest) == STRICT_LOW_PART
1982 || GET_CODE (dest) == ZERO_EXTRACT)
1983 invalidate (XEXP (dest, 0), GET_MODE (dest));
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_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 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 /* The first operand of a SIGN/ZERO_EXTRACT has a different meaning
3136 than it would in other contexts. Basically its mode does not
3137 signify the size of the object read. That information is carried
3138 by size operand. If we happen to have a MEM of the appropriate
3139 mode in our tables with a constant value we could simplify the
3140 extraction incorrectly if we allowed substitution of that value
3141 for the MEM. */
3142 case ZERO_EXTRACT:
3143 case SIGN_EXTRACT:
3144 if ((new_rtx = equiv_constant (x)) != NULL_RTX)
3145 return new_rtx;
3146 return x;
3148 case CONST:
3149 CASE_CONST_ANY:
3150 case SYMBOL_REF:
3151 case LABEL_REF:
3152 case REG:
3153 case PC:
3154 /* No use simplifying an EXPR_LIST
3155 since they are used only for lists of args
3156 in a function call's REG_EQUAL note. */
3157 case EXPR_LIST:
3158 return x;
3160 case CC0:
3161 return prev_insn_cc0;
3163 case ASM_OPERANDS:
3164 if (insn)
3166 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
3167 validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
3168 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
3170 return x;
3172 case CALL:
3173 if (NO_FUNCTION_CSE && CONSTANT_P (XEXP (XEXP (x, 0), 0)))
3174 return x;
3175 break;
3177 /* Anything else goes through the loop below. */
3178 default:
3179 break;
3182 mode = GET_MODE (x);
3183 const_arg0 = 0;
3184 const_arg1 = 0;
3185 const_arg2 = 0;
3186 mode_arg0 = VOIDmode;
3188 /* Try folding our operands.
3189 Then see which ones have constant values known. */
3191 fmt = GET_RTX_FORMAT (code);
3192 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3193 if (fmt[i] == 'e')
3195 rtx folded_arg = XEXP (x, i), const_arg;
3196 machine_mode mode_arg = GET_MODE (folded_arg);
3198 switch (GET_CODE (folded_arg))
3200 case MEM:
3201 case REG:
3202 case SUBREG:
3203 const_arg = equiv_constant (folded_arg);
3204 break;
3206 case CONST:
3207 CASE_CONST_ANY:
3208 case SYMBOL_REF:
3209 case LABEL_REF:
3210 const_arg = folded_arg;
3211 break;
3213 case CC0:
3214 /* The cc0-user and cc0-setter may be in different blocks if
3215 the cc0-setter potentially traps. In that case PREV_INSN_CC0
3216 will have been cleared as we exited the block with the
3217 setter.
3219 While we could potentially track cc0 in this case, it just
3220 doesn't seem to be worth it given that cc0 targets are not
3221 terribly common or important these days and trapping math
3222 is rarely used. The combination of those two conditions
3223 necessary to trip this situation is exceedingly rare in the
3224 real world. */
3225 if (!prev_insn_cc0)
3227 const_arg = NULL_RTX;
3229 else
3231 folded_arg = prev_insn_cc0;
3232 mode_arg = prev_insn_cc0_mode;
3233 const_arg = equiv_constant (folded_arg);
3235 break;
3237 default:
3238 folded_arg = fold_rtx (folded_arg, insn);
3239 const_arg = equiv_constant (folded_arg);
3240 break;
3243 /* For the first three operands, see if the operand
3244 is constant or equivalent to a constant. */
3245 switch (i)
3247 case 0:
3248 folded_arg0 = folded_arg;
3249 const_arg0 = const_arg;
3250 mode_arg0 = mode_arg;
3251 break;
3252 case 1:
3253 folded_arg1 = folded_arg;
3254 const_arg1 = const_arg;
3255 break;
3256 case 2:
3257 const_arg2 = const_arg;
3258 break;
3261 /* Pick the least expensive of the argument and an equivalent constant
3262 argument. */
3263 if (const_arg != 0
3264 && const_arg != folded_arg
3265 && (COST_IN (const_arg, mode_arg, code, i)
3266 <= COST_IN (folded_arg, mode_arg, code, i))
3268 /* It's not safe to substitute the operand of a conversion
3269 operator with a constant, as the conversion's identity
3270 depends upon the mode of its operand. This optimization
3271 is handled by the call to simplify_unary_operation. */
3272 && (GET_RTX_CLASS (code) != RTX_UNARY
3273 || GET_MODE (const_arg) == mode_arg0
3274 || (code != ZERO_EXTEND
3275 && code != SIGN_EXTEND
3276 && code != TRUNCATE
3277 && code != FLOAT_TRUNCATE
3278 && code != FLOAT_EXTEND
3279 && code != FLOAT
3280 && code != FIX
3281 && code != UNSIGNED_FLOAT
3282 && code != UNSIGNED_FIX)))
3283 folded_arg = const_arg;
3285 if (folded_arg == XEXP (x, i))
3286 continue;
3288 if (insn == NULL_RTX && !changed)
3289 x = copy_rtx (x);
3290 changed = 1;
3291 validate_unshare_change (insn, &XEXP (x, i), folded_arg, 1);
3294 if (changed)
3296 /* Canonicalize X if necessary, and keep const_argN and folded_argN
3297 consistent with the order in X. */
3298 if (canonicalize_change_group (insn, x))
3300 std::swap (const_arg0, const_arg1);
3301 std::swap (folded_arg0, folded_arg1);
3304 apply_change_group ();
3307 /* If X is an arithmetic operation, see if we can simplify it. */
3309 switch (GET_RTX_CLASS (code))
3311 case RTX_UNARY:
3313 /* We can't simplify extension ops unless we know the
3314 original mode. */
3315 if ((code == ZERO_EXTEND || code == SIGN_EXTEND)
3316 && mode_arg0 == VOIDmode)
3317 break;
3319 new_rtx = simplify_unary_operation (code, mode,
3320 const_arg0 ? const_arg0 : folded_arg0,
3321 mode_arg0);
3323 break;
3325 case RTX_COMPARE:
3326 case RTX_COMM_COMPARE:
3327 /* See what items are actually being compared and set FOLDED_ARG[01]
3328 to those values and CODE to the actual comparison code. If any are
3329 constant, set CONST_ARG0 and CONST_ARG1 appropriately. We needn't
3330 do anything if both operands are already known to be constant. */
3332 /* ??? Vector mode comparisons are not supported yet. */
3333 if (VECTOR_MODE_P (mode))
3334 break;
3336 if (const_arg0 == 0 || const_arg1 == 0)
3338 struct table_elt *p0, *p1;
3339 rtx true_rtx, false_rtx;
3340 machine_mode mode_arg1;
3342 if (SCALAR_FLOAT_MODE_P (mode))
3344 #ifdef FLOAT_STORE_FLAG_VALUE
3345 true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
3346 (FLOAT_STORE_FLAG_VALUE (mode), mode));
3347 #else
3348 true_rtx = NULL_RTX;
3349 #endif
3350 false_rtx = CONST0_RTX (mode);
3352 else
3354 true_rtx = const_true_rtx;
3355 false_rtx = const0_rtx;
3358 code = find_comparison_args (code, &folded_arg0, &folded_arg1,
3359 &mode_arg0, &mode_arg1);
3361 /* If the mode is VOIDmode or a MODE_CC mode, we don't know
3362 what kinds of things are being compared, so we can't do
3363 anything with this comparison. */
3365 if (mode_arg0 == VOIDmode || GET_MODE_CLASS (mode_arg0) == MODE_CC)
3366 break;
3368 const_arg0 = equiv_constant (folded_arg0);
3369 const_arg1 = equiv_constant (folded_arg1);
3371 /* If we do not now have two constants being compared, see
3372 if we can nevertheless deduce some things about the
3373 comparison. */
3374 if (const_arg0 == 0 || const_arg1 == 0)
3376 if (const_arg1 != NULL)
3378 rtx cheapest_simplification;
3379 int cheapest_cost;
3380 rtx simp_result;
3381 struct table_elt *p;
3383 /* See if we can find an equivalent of folded_arg0
3384 that gets us a cheaper expression, possibly a
3385 constant through simplifications. */
3386 p = lookup (folded_arg0, SAFE_HASH (folded_arg0, mode_arg0),
3387 mode_arg0);
3389 if (p != NULL)
3391 cheapest_simplification = x;
3392 cheapest_cost = COST (x, mode);
3394 for (p = p->first_same_value; p != NULL; p = p->next_same_value)
3396 int cost;
3398 /* If the entry isn't valid, skip it. */
3399 if (! exp_equiv_p (p->exp, p->exp, 1, false))
3400 continue;
3402 /* Try to simplify using this equivalence. */
3403 simp_result
3404 = simplify_relational_operation (code, mode,
3405 mode_arg0,
3406 p->exp,
3407 const_arg1);
3409 if (simp_result == NULL)
3410 continue;
3412 cost = COST (simp_result, mode);
3413 if (cost < cheapest_cost)
3415 cheapest_cost = cost;
3416 cheapest_simplification = simp_result;
3420 /* If we have a cheaper expression now, use that
3421 and try folding it further, from the top. */
3422 if (cheapest_simplification != x)
3423 return fold_rtx (copy_rtx (cheapest_simplification),
3424 insn);
3428 /* See if the two operands are the same. */
3430 if ((REG_P (folded_arg0)
3431 && REG_P (folded_arg1)
3432 && (REG_QTY (REGNO (folded_arg0))
3433 == REG_QTY (REGNO (folded_arg1))))
3434 || ((p0 = lookup (folded_arg0,
3435 SAFE_HASH (folded_arg0, mode_arg0),
3436 mode_arg0))
3437 && (p1 = lookup (folded_arg1,
3438 SAFE_HASH (folded_arg1, mode_arg0),
3439 mode_arg0))
3440 && p0->first_same_value == p1->first_same_value))
3441 folded_arg1 = folded_arg0;
3443 /* If FOLDED_ARG0 is a register, see if the comparison we are
3444 doing now is either the same as we did before or the reverse
3445 (we only check the reverse if not floating-point). */
3446 else if (REG_P (folded_arg0))
3448 int qty = REG_QTY (REGNO (folded_arg0));
3450 if (REGNO_QTY_VALID_P (REGNO (folded_arg0)))
3452 struct qty_table_elem *ent = &qty_table[qty];
3454 if ((comparison_dominates_p (ent->comparison_code, code)
3455 || (! FLOAT_MODE_P (mode_arg0)
3456 && comparison_dominates_p (ent->comparison_code,
3457 reverse_condition (code))))
3458 && (rtx_equal_p (ent->comparison_const, folded_arg1)
3459 || (const_arg1
3460 && rtx_equal_p (ent->comparison_const,
3461 const_arg1))
3462 || (REG_P (folded_arg1)
3463 && (REG_QTY (REGNO (folded_arg1)) == ent->comparison_qty))))
3465 if (comparison_dominates_p (ent->comparison_code, code))
3467 if (true_rtx)
3468 return true_rtx;
3469 else
3470 break;
3472 else
3473 return false_rtx;
3480 /* If we are comparing against zero, see if the first operand is
3481 equivalent to an IOR with a constant. If so, we may be able to
3482 determine the result of this comparison. */
3483 if (const_arg1 == const0_rtx && !const_arg0)
3485 rtx y = lookup_as_function (folded_arg0, IOR);
3486 rtx inner_const;
3488 if (y != 0
3489 && (inner_const = equiv_constant (XEXP (y, 1))) != 0
3490 && CONST_INT_P (inner_const)
3491 && INTVAL (inner_const) != 0)
3492 folded_arg0 = gen_rtx_IOR (mode_arg0, XEXP (y, 0), inner_const);
3496 rtx op0 = const_arg0 ? const_arg0 : copy_rtx (folded_arg0);
3497 rtx op1 = const_arg1 ? const_arg1 : copy_rtx (folded_arg1);
3498 new_rtx = simplify_relational_operation (code, mode, mode_arg0,
3499 op0, op1);
3501 break;
3503 case RTX_BIN_ARITH:
3504 case RTX_COMM_ARITH:
3505 switch (code)
3507 case PLUS:
3508 /* If the second operand is a LABEL_REF, see if the first is a MINUS
3509 with that LABEL_REF as its second operand. If so, the result is
3510 the first operand of that MINUS. This handles switches with an
3511 ADDR_DIFF_VEC table. */
3512 if (const_arg1 && GET_CODE (const_arg1) == LABEL_REF)
3514 rtx y
3515 = GET_CODE (folded_arg0) == MINUS ? folded_arg0
3516 : lookup_as_function (folded_arg0, MINUS);
3518 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3519 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg1))
3520 return XEXP (y, 0);
3522 /* Now try for a CONST of a MINUS like the above. */
3523 if ((y = (GET_CODE (folded_arg0) == CONST ? folded_arg0
3524 : lookup_as_function (folded_arg0, CONST))) != 0
3525 && GET_CODE (XEXP (y, 0)) == MINUS
3526 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3527 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg1))
3528 return XEXP (XEXP (y, 0), 0);
3531 /* Likewise if the operands are in the other order. */
3532 if (const_arg0 && GET_CODE (const_arg0) == LABEL_REF)
3534 rtx y
3535 = GET_CODE (folded_arg1) == MINUS ? folded_arg1
3536 : lookup_as_function (folded_arg1, MINUS);
3538 if (y != 0 && GET_CODE (XEXP (y, 1)) == LABEL_REF
3539 && LABEL_REF_LABEL (XEXP (y, 1)) == LABEL_REF_LABEL (const_arg0))
3540 return XEXP (y, 0);
3542 /* Now try for a CONST of a MINUS like the above. */
3543 if ((y = (GET_CODE (folded_arg1) == CONST ? folded_arg1
3544 : lookup_as_function (folded_arg1, CONST))) != 0
3545 && GET_CODE (XEXP (y, 0)) == MINUS
3546 && GET_CODE (XEXP (XEXP (y, 0), 1)) == LABEL_REF
3547 && LABEL_REF_LABEL (XEXP (XEXP (y, 0), 1)) == LABEL_REF_LABEL (const_arg0))
3548 return XEXP (XEXP (y, 0), 0);
3551 /* If second operand is a register equivalent to a negative
3552 CONST_INT, see if we can find a register equivalent to the
3553 positive constant. Make a MINUS if so. Don't do this for
3554 a non-negative constant since we might then alternate between
3555 choosing positive and negative constants. Having the positive
3556 constant previously-used is the more common case. Be sure
3557 the resulting constant is non-negative; if const_arg1 were
3558 the smallest negative number this would overflow: depending
3559 on the mode, this would either just be the same value (and
3560 hence not save anything) or be incorrect. */
3561 if (const_arg1 != 0 && CONST_INT_P (const_arg1)
3562 && INTVAL (const_arg1) < 0
3563 /* This used to test
3565 -INTVAL (const_arg1) >= 0
3567 But The Sun V5.0 compilers mis-compiled that test. So
3568 instead we test for the problematic value in a more direct
3569 manner and hope the Sun compilers get it correct. */
3570 && INTVAL (const_arg1) !=
3571 ((HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
3572 && REG_P (folded_arg1))
3574 rtx new_const = GEN_INT (-INTVAL (const_arg1));
3575 struct table_elt *p
3576 = lookup (new_const, SAFE_HASH (new_const, mode), mode);
3578 if (p)
3579 for (p = p->first_same_value; p; p = p->next_same_value)
3580 if (REG_P (p->exp))
3581 return simplify_gen_binary (MINUS, mode, folded_arg0,
3582 canon_reg (p->exp, NULL));
3584 goto from_plus;
3586 case MINUS:
3587 /* If we have (MINUS Y C), see if Y is known to be (PLUS Z C2).
3588 If so, produce (PLUS Z C2-C). */
3589 if (const_arg1 != 0 && CONST_INT_P (const_arg1))
3591 rtx y = lookup_as_function (XEXP (x, 0), PLUS);
3592 if (y && CONST_INT_P (XEXP (y, 1)))
3593 return fold_rtx (plus_constant (mode, copy_rtx (y),
3594 -INTVAL (const_arg1)),
3595 NULL);
3598 /* Fall through. */
3600 from_plus:
3601 case SMIN: case SMAX: case UMIN: case UMAX:
3602 case IOR: case AND: case XOR:
3603 case MULT:
3604 case ASHIFT: case LSHIFTRT: case ASHIFTRT:
3605 /* If we have (<op> <reg> <const_int>) for an associative OP and REG
3606 is known to be of similar form, we may be able to replace the
3607 operation with a combined operation. This may eliminate the
3608 intermediate operation if every use is simplified in this way.
3609 Note that the similar optimization done by combine.c only works
3610 if the intermediate operation's result has only one reference. */
3612 if (REG_P (folded_arg0)
3613 && const_arg1 && CONST_INT_P (const_arg1))
3615 int is_shift
3616 = (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT);
3617 rtx y, inner_const, new_const;
3618 rtx canon_const_arg1 = const_arg1;
3619 enum rtx_code associate_code;
3621 if (is_shift
3622 && (INTVAL (const_arg1) >= GET_MODE_PRECISION (mode)
3623 || INTVAL (const_arg1) < 0))
3625 if (SHIFT_COUNT_TRUNCATED)
3626 canon_const_arg1 = GEN_INT (INTVAL (const_arg1)
3627 & (GET_MODE_BITSIZE (mode)
3628 - 1));
3629 else
3630 break;
3633 y = lookup_as_function (folded_arg0, code);
3634 if (y == 0)
3635 break;
3637 /* If we have compiled a statement like
3638 "if (x == (x & mask1))", and now are looking at
3639 "x & mask2", we will have a case where the first operand
3640 of Y is the same as our first operand. Unless we detect
3641 this case, an infinite loop will result. */
3642 if (XEXP (y, 0) == folded_arg0)
3643 break;
3645 inner_const = equiv_constant (fold_rtx (XEXP (y, 1), 0));
3646 if (!inner_const || !CONST_INT_P (inner_const))
3647 break;
3649 /* Don't associate these operations if they are a PLUS with the
3650 same constant and it is a power of two. These might be doable
3651 with a pre- or post-increment. Similarly for two subtracts of
3652 identical powers of two with post decrement. */
3654 if (code == PLUS && const_arg1 == inner_const
3655 && ((HAVE_PRE_INCREMENT
3656 && exact_log2 (INTVAL (const_arg1)) >= 0)
3657 || (HAVE_POST_INCREMENT
3658 && exact_log2 (INTVAL (const_arg1)) >= 0)
3659 || (HAVE_PRE_DECREMENT
3660 && exact_log2 (- INTVAL (const_arg1)) >= 0)
3661 || (HAVE_POST_DECREMENT
3662 && exact_log2 (- INTVAL (const_arg1)) >= 0)))
3663 break;
3665 /* ??? Vector mode shifts by scalar
3666 shift operand are not supported yet. */
3667 if (is_shift && VECTOR_MODE_P (mode))
3668 break;
3670 if (is_shift
3671 && (INTVAL (inner_const) >= GET_MODE_PRECISION (mode)
3672 || INTVAL (inner_const) < 0))
3674 if (SHIFT_COUNT_TRUNCATED)
3675 inner_const = GEN_INT (INTVAL (inner_const)
3676 & (GET_MODE_BITSIZE (mode) - 1));
3677 else
3678 break;
3681 /* Compute the code used to compose the constants. For example,
3682 A-C1-C2 is A-(C1 + C2), so if CODE == MINUS, we want PLUS. */
3684 associate_code = (is_shift || code == MINUS ? PLUS : code);
3686 new_const = simplify_binary_operation (associate_code, mode,
3687 canon_const_arg1,
3688 inner_const);
3690 if (new_const == 0)
3691 break;
3693 /* If we are associating shift operations, don't let this
3694 produce a shift of the size of the object or larger.
3695 This could occur when we follow a sign-extend by a right
3696 shift on a machine that does a sign-extend as a pair
3697 of shifts. */
3699 if (is_shift
3700 && CONST_INT_P (new_const)
3701 && INTVAL (new_const) >= GET_MODE_PRECISION (mode))
3703 /* As an exception, we can turn an ASHIFTRT of this
3704 form into a shift of the number of bits - 1. */
3705 if (code == ASHIFTRT)
3706 new_const = GEN_INT (GET_MODE_BITSIZE (mode) - 1);
3707 else if (!side_effects_p (XEXP (y, 0)))
3708 return CONST0_RTX (mode);
3709 else
3710 break;
3713 y = copy_rtx (XEXP (y, 0));
3715 /* If Y contains our first operand (the most common way this
3716 can happen is if Y is a MEM), we would do into an infinite
3717 loop if we tried to fold it. So don't in that case. */
3719 if (! reg_mentioned_p (folded_arg0, y))
3720 y = fold_rtx (y, insn);
3722 return simplify_gen_binary (code, mode, y, new_const);
3724 break;
3726 case DIV: case UDIV:
3727 /* ??? The associative optimization performed immediately above is
3728 also possible for DIV and UDIV using associate_code of MULT.
3729 However, we would need extra code to verify that the
3730 multiplication does not overflow, that is, there is no overflow
3731 in the calculation of new_const. */
3732 break;
3734 default:
3735 break;
3738 new_rtx = simplify_binary_operation (code, mode,
3739 const_arg0 ? const_arg0 : folded_arg0,
3740 const_arg1 ? const_arg1 : folded_arg1);
3741 break;
3743 case RTX_OBJ:
3744 /* (lo_sum (high X) X) is simply X. */
3745 if (code == LO_SUM && const_arg0 != 0
3746 && GET_CODE (const_arg0) == HIGH
3747 && rtx_equal_p (XEXP (const_arg0, 0), const_arg1))
3748 return const_arg1;
3749 break;
3751 case RTX_TERNARY:
3752 case RTX_BITFIELD_OPS:
3753 new_rtx = simplify_ternary_operation (code, mode, mode_arg0,
3754 const_arg0 ? const_arg0 : folded_arg0,
3755 const_arg1 ? const_arg1 : folded_arg1,
3756 const_arg2 ? const_arg2 : XEXP (x, 2));
3757 break;
3759 default:
3760 break;
3763 return new_rtx ? new_rtx : x;
3766 /* Return a constant value currently equivalent to X.
3767 Return 0 if we don't know one. */
3769 static rtx
3770 equiv_constant (rtx x)
3772 if (REG_P (x)
3773 && REGNO_QTY_VALID_P (REGNO (x)))
3775 int x_q = REG_QTY (REGNO (x));
3776 struct qty_table_elem *x_ent = &qty_table[x_q];
3778 if (x_ent->const_rtx)
3779 x = gen_lowpart (GET_MODE (x), x_ent->const_rtx);
3782 if (x == 0 || CONSTANT_P (x))
3783 return x;
3785 if (GET_CODE (x) == SUBREG)
3787 machine_mode mode = GET_MODE (x);
3788 machine_mode imode = GET_MODE (SUBREG_REG (x));
3789 rtx new_rtx;
3791 /* See if we previously assigned a constant value to this SUBREG. */
3792 if ((new_rtx = lookup_as_function (x, CONST_INT)) != 0
3793 || (new_rtx = lookup_as_function (x, CONST_WIDE_INT)) != 0
3794 || (new_rtx = lookup_as_function (x, CONST_DOUBLE)) != 0
3795 || (new_rtx = lookup_as_function (x, CONST_FIXED)) != 0)
3796 return new_rtx;
3798 /* If we didn't and if doing so makes sense, see if we previously
3799 assigned a constant value to the enclosing word mode SUBREG. */
3800 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (word_mode)
3801 && GET_MODE_SIZE (word_mode) < GET_MODE_SIZE (imode))
3803 int byte = SUBREG_BYTE (x) - subreg_lowpart_offset (mode, word_mode);
3804 if (byte >= 0 && (byte % UNITS_PER_WORD) == 0)
3806 rtx y = gen_rtx_SUBREG (word_mode, SUBREG_REG (x), byte);
3807 new_rtx = lookup_as_function (y, CONST_INT);
3808 if (new_rtx)
3809 return gen_lowpart (mode, new_rtx);
3813 /* Otherwise see if we already have a constant for the inner REG,
3814 and if that is enough to calculate an equivalent constant for
3815 the subreg. Note that the upper bits of paradoxical subregs
3816 are undefined, so they cannot be said to equal anything. */
3817 if (REG_P (SUBREG_REG (x))
3818 && GET_MODE_SIZE (mode) <= GET_MODE_SIZE (imode)
3819 && (new_rtx = equiv_constant (SUBREG_REG (x))) != 0)
3820 return simplify_subreg (mode, new_rtx, imode, SUBREG_BYTE (x));
3822 return 0;
3825 /* If X is a MEM, see if it is a constant-pool reference, or look it up in
3826 the hash table in case its value was seen before. */
3828 if (MEM_P (x))
3830 struct table_elt *elt;
3832 x = avoid_constant_pool_reference (x);
3833 if (CONSTANT_P (x))
3834 return x;
3836 elt = lookup (x, SAFE_HASH (x, GET_MODE (x)), GET_MODE (x));
3837 if (elt == 0)
3838 return 0;
3840 for (elt = elt->first_same_value; elt; elt = elt->next_same_value)
3841 if (elt->is_const && CONSTANT_P (elt->exp))
3842 return elt->exp;
3845 return 0;
3848 /* Given INSN, a jump insn, TAKEN indicates if we are following the
3849 "taken" branch.
3851 In certain cases, this can cause us to add an equivalence. For example,
3852 if we are following the taken case of
3853 if (i == 2)
3854 we can add the fact that `i' and '2' are now equivalent.
3856 In any case, we can record that this comparison was passed. If the same
3857 comparison is seen later, we will know its value. */
3859 static void
3860 record_jump_equiv (rtx_insn *insn, bool taken)
3862 int cond_known_true;
3863 rtx op0, op1;
3864 rtx set;
3865 machine_mode mode, mode0, mode1;
3866 int reversed_nonequality = 0;
3867 enum rtx_code code;
3869 /* Ensure this is the right kind of insn. */
3870 gcc_assert (any_condjump_p (insn));
3872 set = pc_set (insn);
3874 /* See if this jump condition is known true or false. */
3875 if (taken)
3876 cond_known_true = (XEXP (SET_SRC (set), 2) == pc_rtx);
3877 else
3878 cond_known_true = (XEXP (SET_SRC (set), 1) == pc_rtx);
3880 /* Get the type of comparison being done and the operands being compared.
3881 If we had to reverse a non-equality condition, record that fact so we
3882 know that it isn't valid for floating-point. */
3883 code = GET_CODE (XEXP (SET_SRC (set), 0));
3884 op0 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 0), insn);
3885 op1 = fold_rtx (XEXP (XEXP (SET_SRC (set), 0), 1), insn);
3887 code = find_comparison_args (code, &op0, &op1, &mode0, &mode1);
3888 if (! cond_known_true)
3890 code = reversed_comparison_code_parts (code, op0, op1, insn);
3892 /* Don't remember if we can't find the inverse. */
3893 if (code == UNKNOWN)
3894 return;
3897 /* The mode is the mode of the non-constant. */
3898 mode = mode0;
3899 if (mode1 != VOIDmode)
3900 mode = mode1;
3902 record_jump_cond (code, mode, op0, op1, reversed_nonequality);
3905 /* Yet another form of subreg creation. In this case, we want something in
3906 MODE, and we should assume OP has MODE iff it is naturally modeless. */
3908 static rtx
3909 record_jump_cond_subreg (machine_mode mode, rtx op)
3911 machine_mode op_mode = GET_MODE (op);
3912 if (op_mode == mode || op_mode == VOIDmode)
3913 return op;
3914 return lowpart_subreg (mode, op, op_mode);
3917 /* We know that comparison CODE applied to OP0 and OP1 in MODE is true.
3918 REVERSED_NONEQUALITY is nonzero if CODE had to be swapped.
3919 Make any useful entries we can with that information. Called from
3920 above function and called recursively. */
3922 static void
3923 record_jump_cond (enum rtx_code code, machine_mode mode, rtx op0,
3924 rtx op1, int reversed_nonequality)
3926 unsigned op0_hash, op1_hash;
3927 int op0_in_memory, op1_in_memory;
3928 struct table_elt *op0_elt, *op1_elt;
3930 /* If OP0 and OP1 are known equal, and either is a paradoxical SUBREG,
3931 we know that they are also equal in the smaller mode (this is also
3932 true for all smaller modes whether or not there is a SUBREG, but
3933 is not worth testing for with no SUBREG). */
3935 /* Note that GET_MODE (op0) may not equal MODE. */
3936 if (code == EQ && paradoxical_subreg_p (op0))
3938 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3939 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3940 if (tem)
3941 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3942 reversed_nonequality);
3945 if (code == EQ && paradoxical_subreg_p (op1))
3947 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3948 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3949 if (tem)
3950 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3951 reversed_nonequality);
3954 /* Similarly, if this is an NE comparison, and either is a SUBREG
3955 making a smaller mode, we know the whole thing is also NE. */
3957 /* Note that GET_MODE (op0) may not equal MODE;
3958 if we test MODE instead, we can get an infinite recursion
3959 alternating between two modes each wider than MODE. */
3961 if (code == NE && GET_CODE (op0) == SUBREG
3962 && subreg_lowpart_p (op0)
3963 && (GET_MODE_SIZE (GET_MODE (op0))
3964 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
3966 machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
3967 rtx tem = record_jump_cond_subreg (inner_mode, op1);
3968 if (tem)
3969 record_jump_cond (code, mode, SUBREG_REG (op0), tem,
3970 reversed_nonequality);
3973 if (code == NE && GET_CODE (op1) == SUBREG
3974 && subreg_lowpart_p (op1)
3975 && (GET_MODE_SIZE (GET_MODE (op1))
3976 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1)))))
3978 machine_mode inner_mode = GET_MODE (SUBREG_REG (op1));
3979 rtx tem = record_jump_cond_subreg (inner_mode, op0);
3980 if (tem)
3981 record_jump_cond (code, mode, SUBREG_REG (op1), tem,
3982 reversed_nonequality);
3985 /* Hash both operands. */
3987 do_not_record = 0;
3988 hash_arg_in_memory = 0;
3989 op0_hash = HASH (op0, mode);
3990 op0_in_memory = hash_arg_in_memory;
3992 if (do_not_record)
3993 return;
3995 do_not_record = 0;
3996 hash_arg_in_memory = 0;
3997 op1_hash = HASH (op1, mode);
3998 op1_in_memory = hash_arg_in_memory;
4000 if (do_not_record)
4001 return;
4003 /* Look up both operands. */
4004 op0_elt = lookup (op0, op0_hash, mode);
4005 op1_elt = lookup (op1, op1_hash, mode);
4007 /* If both operands are already equivalent or if they are not in the
4008 table but are identical, do nothing. */
4009 if ((op0_elt != 0 && op1_elt != 0
4010 && op0_elt->first_same_value == op1_elt->first_same_value)
4011 || op0 == op1 || rtx_equal_p (op0, op1))
4012 return;
4014 /* If we aren't setting two things equal all we can do is save this
4015 comparison. Similarly if this is floating-point. In the latter
4016 case, OP1 might be zero and both -0.0 and 0.0 are equal to it.
4017 If we record the equality, we might inadvertently delete code
4018 whose intent was to change -0 to +0. */
4020 if (code != EQ || FLOAT_MODE_P (GET_MODE (op0)))
4022 struct qty_table_elem *ent;
4023 int qty;
4025 /* If we reversed a floating-point comparison, if OP0 is not a
4026 register, or if OP1 is neither a register or constant, we can't
4027 do anything. */
4029 if (!REG_P (op1))
4030 op1 = equiv_constant (op1);
4032 if ((reversed_nonequality && FLOAT_MODE_P (mode))
4033 || !REG_P (op0) || op1 == 0)
4034 return;
4036 /* Put OP0 in the hash table if it isn't already. This gives it a
4037 new quantity number. */
4038 if (op0_elt == 0)
4040 if (insert_regs (op0, NULL, 0))
4042 rehash_using_reg (op0);
4043 op0_hash = HASH (op0, mode);
4045 /* If OP0 is contained in OP1, this changes its hash code
4046 as well. Faster to rehash than to check, except
4047 for the simple case of a constant. */
4048 if (! CONSTANT_P (op1))
4049 op1_hash = HASH (op1,mode);
4052 op0_elt = insert (op0, NULL, op0_hash, mode);
4053 op0_elt->in_memory = op0_in_memory;
4056 qty = REG_QTY (REGNO (op0));
4057 ent = &qty_table[qty];
4059 ent->comparison_code = code;
4060 if (REG_P (op1))
4062 /* Look it up again--in case op0 and op1 are the same. */
4063 op1_elt = lookup (op1, op1_hash, mode);
4065 /* Put OP1 in the hash table so it gets a new quantity number. */
4066 if (op1_elt == 0)
4068 if (insert_regs (op1, NULL, 0))
4070 rehash_using_reg (op1);
4071 op1_hash = HASH (op1, mode);
4074 op1_elt = insert (op1, NULL, op1_hash, mode);
4075 op1_elt->in_memory = op1_in_memory;
4078 ent->comparison_const = NULL_RTX;
4079 ent->comparison_qty = REG_QTY (REGNO (op1));
4081 else
4083 ent->comparison_const = op1;
4084 ent->comparison_qty = -1;
4087 return;
4090 /* If either side is still missing an equivalence, make it now,
4091 then merge the equivalences. */
4093 if (op0_elt == 0)
4095 if (insert_regs (op0, NULL, 0))
4097 rehash_using_reg (op0);
4098 op0_hash = HASH (op0, mode);
4101 op0_elt = insert (op0, NULL, op0_hash, mode);
4102 op0_elt->in_memory = op0_in_memory;
4105 if (op1_elt == 0)
4107 if (insert_regs (op1, NULL, 0))
4109 rehash_using_reg (op1);
4110 op1_hash = HASH (op1, mode);
4113 op1_elt = insert (op1, NULL, op1_hash, mode);
4114 op1_elt->in_memory = op1_in_memory;
4117 merge_equiv_classes (op0_elt, op1_elt);
4120 /* CSE processing for one instruction.
4122 Most "true" common subexpressions are mostly optimized away in GIMPLE,
4123 but the few that "leak through" are cleaned up by cse_insn, and complex
4124 addressing modes are often formed here.
4126 The main function is cse_insn, and between here and that function
4127 a couple of helper functions is defined to keep the size of cse_insn
4128 within reasonable proportions.
4130 Data is shared between the main and helper functions via STRUCT SET,
4131 that contains all data related for every set in the instruction that
4132 is being processed.
4134 Note that cse_main processes all sets in the instruction. Most
4135 passes in GCC only process simple SET insns or single_set insns, but
4136 CSE processes insns with multiple sets as well. */
4138 /* Data on one SET contained in the instruction. */
4140 struct set
4142 /* The SET rtx itself. */
4143 rtx rtl;
4144 /* The SET_SRC of the rtx (the original value, if it is changing). */
4145 rtx src;
4146 /* The hash-table element for the SET_SRC of the SET. */
4147 struct table_elt *src_elt;
4148 /* Hash value for the SET_SRC. */
4149 unsigned src_hash;
4150 /* Hash value for the SET_DEST. */
4151 unsigned dest_hash;
4152 /* The SET_DEST, with SUBREG, etc., stripped. */
4153 rtx inner_dest;
4154 /* Nonzero if the SET_SRC is in memory. */
4155 char src_in_memory;
4156 /* Nonzero if the SET_SRC contains something
4157 whose value cannot be predicted and understood. */
4158 char src_volatile;
4159 /* Original machine mode, in case it becomes a CONST_INT.
4160 The size of this field should match the size of the mode
4161 field of struct rtx_def (see rtl.h). */
4162 ENUM_BITFIELD(machine_mode) mode : 8;
4163 /* A constant equivalent for SET_SRC, if any. */
4164 rtx src_const;
4165 /* Hash value of constant equivalent for SET_SRC. */
4166 unsigned src_const_hash;
4167 /* Table entry for constant equivalent for SET_SRC, if any. */
4168 struct table_elt *src_const_elt;
4169 /* Table entry for the destination address. */
4170 struct table_elt *dest_addr_elt;
4173 /* Special handling for (set REG0 REG1) where REG0 is the
4174 "cheapest", cheaper than REG1. After cse, REG1 will probably not
4175 be used in the sequel, so (if easily done) change this insn to
4176 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
4177 that computed their value. Then REG1 will become a dead store
4178 and won't cloud the situation for later optimizations.
4180 Do not make this change if REG1 is a hard register, because it will
4181 then be used in the sequel and we may be changing a two-operand insn
4182 into a three-operand insn.
4184 This is the last transformation that cse_insn will try to do. */
4186 static void
4187 try_back_substitute_reg (rtx set, rtx_insn *insn)
4189 rtx dest = SET_DEST (set);
4190 rtx src = SET_SRC (set);
4192 if (REG_P (dest)
4193 && REG_P (src) && ! HARD_REGISTER_P (src)
4194 && REGNO_QTY_VALID_P (REGNO (src)))
4196 int src_q = REG_QTY (REGNO (src));
4197 struct qty_table_elem *src_ent = &qty_table[src_q];
4199 if (src_ent->first_reg == REGNO (dest))
4201 /* Scan for the previous nonnote insn, but stop at a basic
4202 block boundary. */
4203 rtx_insn *prev = insn;
4204 rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (insn));
4207 prev = PREV_INSN (prev);
4209 while (prev != bb_head && (NOTE_P (prev) || DEBUG_INSN_P (prev)));
4211 /* Do not swap the registers around if the previous instruction
4212 attaches a REG_EQUIV note to REG1.
4214 ??? It's not entirely clear whether we can transfer a REG_EQUIV
4215 from the pseudo that originally shadowed an incoming argument
4216 to another register. Some uses of REG_EQUIV might rely on it
4217 being attached to REG1 rather than REG2.
4219 This section previously turned the REG_EQUIV into a REG_EQUAL
4220 note. We cannot do that because REG_EQUIV may provide an
4221 uninitialized stack slot when REG_PARM_STACK_SPACE is used. */
4222 if (NONJUMP_INSN_P (prev)
4223 && GET_CODE (PATTERN (prev)) == SET
4224 && SET_DEST (PATTERN (prev)) == src
4225 && ! find_reg_note (prev, REG_EQUIV, NULL_RTX))
4227 rtx note;
4229 validate_change (prev, &SET_DEST (PATTERN (prev)), dest, 1);
4230 validate_change (insn, &SET_DEST (set), src, 1);
4231 validate_change (insn, &SET_SRC (set), dest, 1);
4232 apply_change_group ();
4234 /* If INSN has a REG_EQUAL note, and this note mentions
4235 REG0, then we must delete it, because the value in
4236 REG0 has changed. If the note's value is REG1, we must
4237 also delete it because that is now this insn's dest. */
4238 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
4239 if (note != 0
4240 && (reg_mentioned_p (dest, XEXP (note, 0))
4241 || rtx_equal_p (src, XEXP (note, 0))))
4242 remove_note (insn, note);
4248 /* Record all the SETs in this instruction into SETS_PTR,
4249 and return the number of recorded sets. */
4250 static int
4251 find_sets_in_insn (rtx_insn *insn, struct set **psets)
4253 struct set *sets = *psets;
4254 int n_sets = 0;
4255 rtx x = PATTERN (insn);
4257 if (GET_CODE (x) == SET)
4259 /* Ignore SETs that are unconditional jumps.
4260 They never need cse processing, so this does not hurt.
4261 The reason is not efficiency but rather
4262 so that we can test at the end for instructions
4263 that have been simplified to unconditional jumps
4264 and not be misled by unchanged instructions
4265 that were unconditional jumps to begin with. */
4266 if (SET_DEST (x) == pc_rtx
4267 && GET_CODE (SET_SRC (x)) == LABEL_REF)
4269 /* Don't count call-insns, (set (reg 0) (call ...)), as a set.
4270 The hard function value register is used only once, to copy to
4271 someplace else, so it isn't worth cse'ing. */
4272 else if (GET_CODE (SET_SRC (x)) == CALL)
4274 else
4275 sets[n_sets++].rtl = x;
4277 else if (GET_CODE (x) == PARALLEL)
4279 int i, lim = XVECLEN (x, 0);
4281 /* Go over the expressions of the PARALLEL in forward order, to
4282 put them in the same order in the SETS array. */
4283 for (i = 0; i < lim; i++)
4285 rtx y = XVECEXP (x, 0, i);
4286 if (GET_CODE (y) == SET)
4288 /* As above, we ignore unconditional jumps and call-insns and
4289 ignore the result of apply_change_group. */
4290 if (SET_DEST (y) == pc_rtx
4291 && GET_CODE (SET_SRC (y)) == LABEL_REF)
4293 else if (GET_CODE (SET_SRC (y)) == CALL)
4295 else
4296 sets[n_sets++].rtl = y;
4301 return n_sets;
4304 /* Where possible, substitute every register reference in the N_SETS
4305 number of SETS in INSN with the canonical register.
4307 Register canonicalization propagatest the earliest register (i.e.
4308 one that is set before INSN) with the same value. This is a very
4309 useful, simple form of CSE, to clean up warts from expanding GIMPLE
4310 to RTL. For instance, a CONST for an address is usually expanded
4311 multiple times to loads into different registers, thus creating many
4312 subexpressions of the form:
4314 (set (reg1) (some_const))
4315 (set (mem (... reg1 ...) (thing)))
4316 (set (reg2) (some_const))
4317 (set (mem (... reg2 ...) (thing)))
4319 After canonicalizing, the code takes the following form:
4321 (set (reg1) (some_const))
4322 (set (mem (... reg1 ...) (thing)))
4323 (set (reg2) (some_const))
4324 (set (mem (... reg1 ...) (thing)))
4326 The set to reg2 is now trivially dead, and the memory reference (or
4327 address, or whatever) may be a candidate for further CSEing.
4329 In this function, the result of apply_change_group can be ignored;
4330 see canon_reg. */
4332 static void
4333 canonicalize_insn (rtx_insn *insn, struct set **psets, int n_sets)
4335 struct set *sets = *psets;
4336 rtx tem;
4337 rtx x = PATTERN (insn);
4338 int i;
4340 if (CALL_P (insn))
4342 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
4343 if (GET_CODE (XEXP (tem, 0)) != SET)
4344 XEXP (tem, 0) = canon_reg (XEXP (tem, 0), insn);
4347 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
4349 canon_reg (SET_SRC (x), insn);
4350 apply_change_group ();
4351 fold_rtx (SET_SRC (x), insn);
4353 else if (GET_CODE (x) == CLOBBER)
4355 /* If we clobber memory, canon the address.
4356 This does nothing when a register is clobbered
4357 because we have already invalidated the reg. */
4358 if (MEM_P (XEXP (x, 0)))
4359 canon_reg (XEXP (x, 0), insn);
4361 else if (GET_CODE (x) == USE
4362 && ! (REG_P (XEXP (x, 0))
4363 && REGNO (XEXP (x, 0)) < FIRST_PSEUDO_REGISTER))
4364 /* Canonicalize a USE of a pseudo register or memory location. */
4365 canon_reg (x, insn);
4366 else if (GET_CODE (x) == ASM_OPERANDS)
4368 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
4370 rtx input = ASM_OPERANDS_INPUT (x, i);
4371 if (!(REG_P (input) && REGNO (input) < FIRST_PSEUDO_REGISTER))
4373 input = canon_reg (input, insn);
4374 validate_change (insn, &ASM_OPERANDS_INPUT (x, i), input, 1);
4378 else if (GET_CODE (x) == CALL)
4380 canon_reg (x, insn);
4381 apply_change_group ();
4382 fold_rtx (x, insn);
4384 else if (DEBUG_INSN_P (insn))
4385 canon_reg (PATTERN (insn), insn);
4386 else if (GET_CODE (x) == PARALLEL)
4388 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
4390 rtx y = XVECEXP (x, 0, i);
4391 if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
4393 canon_reg (SET_SRC (y), insn);
4394 apply_change_group ();
4395 fold_rtx (SET_SRC (y), insn);
4397 else if (GET_CODE (y) == CLOBBER)
4399 if (MEM_P (XEXP (y, 0)))
4400 canon_reg (XEXP (y, 0), insn);
4402 else if (GET_CODE (y) == USE
4403 && ! (REG_P (XEXP (y, 0))
4404 && REGNO (XEXP (y, 0)) < FIRST_PSEUDO_REGISTER))
4405 canon_reg (y, insn);
4406 else if (GET_CODE (y) == CALL)
4408 canon_reg (y, insn);
4409 apply_change_group ();
4410 fold_rtx (y, insn);
4415 if (n_sets == 1 && REG_NOTES (insn) != 0
4416 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4418 /* We potentially will process this insn many times. Therefore,
4419 drop the REG_EQUAL note if it is equal to the SET_SRC of the
4420 unique set in INSN.
4422 Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
4423 because cse_insn handles those specially. */
4424 if (GET_CODE (SET_DEST (sets[0].rtl)) != STRICT_LOW_PART
4425 && rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl)))
4426 remove_note (insn, tem);
4427 else
4429 canon_reg (XEXP (tem, 0), insn);
4430 apply_change_group ();
4431 XEXP (tem, 0) = fold_rtx (XEXP (tem, 0), insn);
4432 df_notes_rescan (insn);
4436 /* Canonicalize sources and addresses of destinations.
4437 We do this in a separate pass to avoid problems when a MATCH_DUP is
4438 present in the insn pattern. In that case, we want to ensure that
4439 we don't break the duplicate nature of the pattern. So we will replace
4440 both operands at the same time. Otherwise, we would fail to find an
4441 equivalent substitution in the loop calling validate_change below.
4443 We used to suppress canonicalization of DEST if it appears in SRC,
4444 but we don't do this any more. */
4446 for (i = 0; i < n_sets; i++)
4448 rtx dest = SET_DEST (sets[i].rtl);
4449 rtx src = SET_SRC (sets[i].rtl);
4450 rtx new_rtx = canon_reg (src, insn);
4452 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
4454 if (GET_CODE (dest) == ZERO_EXTRACT)
4456 validate_change (insn, &XEXP (dest, 1),
4457 canon_reg (XEXP (dest, 1), insn), 1);
4458 validate_change (insn, &XEXP (dest, 2),
4459 canon_reg (XEXP (dest, 2), insn), 1);
4462 while (GET_CODE (dest) == SUBREG
4463 || GET_CODE (dest) == ZERO_EXTRACT
4464 || GET_CODE (dest) == STRICT_LOW_PART)
4465 dest = XEXP (dest, 0);
4467 if (MEM_P (dest))
4468 canon_reg (dest, insn);
4471 /* Now that we have done all the replacements, we can apply the change
4472 group and see if they all work. Note that this will cause some
4473 canonicalizations that would have worked individually not to be applied
4474 because some other canonicalization didn't work, but this should not
4475 occur often.
4477 The result of apply_change_group can be ignored; see canon_reg. */
4479 apply_change_group ();
4482 /* Main function of CSE.
4483 First simplify sources and addresses of all assignments
4484 in the instruction, using previously-computed equivalents values.
4485 Then install the new sources and destinations in the table
4486 of available values. */
4488 static void
4489 cse_insn (rtx_insn *insn)
4491 rtx x = PATTERN (insn);
4492 int i;
4493 rtx tem;
4494 int n_sets = 0;
4496 rtx src_eqv = 0;
4497 struct table_elt *src_eqv_elt = 0;
4498 int src_eqv_volatile = 0;
4499 int src_eqv_in_memory = 0;
4500 unsigned src_eqv_hash = 0;
4502 struct set *sets = (struct set *) 0;
4504 if (GET_CODE (x) == SET)
4505 sets = XALLOCA (struct set);
4506 else if (GET_CODE (x) == PARALLEL)
4507 sets = XALLOCAVEC (struct set, XVECLEN (x, 0));
4509 this_insn = insn;
4510 /* Records what this insn does to set CC0. */
4511 this_insn_cc0 = 0;
4512 this_insn_cc0_mode = VOIDmode;
4514 /* Find all regs explicitly clobbered in this insn,
4515 to ensure they are not replaced with any other regs
4516 elsewhere in this insn. */
4517 invalidate_from_sets_and_clobbers (insn);
4519 /* Record all the SETs in this instruction. */
4520 n_sets = find_sets_in_insn (insn, &sets);
4522 /* Substitute the canonical register where possible. */
4523 canonicalize_insn (insn, &sets, n_sets);
4525 /* If this insn has a REG_EQUAL note, store the equivalent value in SRC_EQV,
4526 if different, or if the DEST is a STRICT_LOW_PART/ZERO_EXTRACT. The
4527 latter condition is necessary because SRC_EQV is handled specially for
4528 this case, and if it isn't set, then there will be no equivalence
4529 for the destination. */
4530 if (n_sets == 1 && REG_NOTES (insn) != 0
4531 && (tem = find_reg_note (insn, REG_EQUAL, NULL_RTX)) != 0)
4534 if (GET_CODE (SET_DEST (sets[0].rtl)) != ZERO_EXTRACT
4535 && (! rtx_equal_p (XEXP (tem, 0), SET_SRC (sets[0].rtl))
4536 || GET_CODE (SET_DEST (sets[0].rtl)) == STRICT_LOW_PART))
4537 src_eqv = copy_rtx (XEXP (tem, 0));
4538 /* If DEST is of the form ZERO_EXTACT, as in:
4539 (set (zero_extract:SI (reg:SI 119)
4540 (const_int 16 [0x10])
4541 (const_int 16 [0x10]))
4542 (const_int 51154 [0xc7d2]))
4543 REG_EQUAL note will specify the value of register (reg:SI 119) at this
4544 point. Note that this is different from SRC_EQV. We can however
4545 calculate SRC_EQV with the position and width of ZERO_EXTRACT. */
4546 else if (GET_CODE (SET_DEST (sets[0].rtl)) == ZERO_EXTRACT
4547 && CONST_INT_P (XEXP (tem, 0))
4548 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 1))
4549 && CONST_INT_P (XEXP (SET_DEST (sets[0].rtl), 2)))
4551 rtx dest_reg = XEXP (SET_DEST (sets[0].rtl), 0);
4552 rtx width = XEXP (SET_DEST (sets[0].rtl), 1);
4553 rtx pos = XEXP (SET_DEST (sets[0].rtl), 2);
4554 HOST_WIDE_INT val = INTVAL (XEXP (tem, 0));
4555 HOST_WIDE_INT mask;
4556 unsigned int shift;
4557 if (BITS_BIG_ENDIAN)
4558 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
4559 - INTVAL (pos) - INTVAL (width);
4560 else
4561 shift = INTVAL (pos);
4562 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
4563 mask = ~(HOST_WIDE_INT) 0;
4564 else
4565 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
4566 val = (val >> shift) & mask;
4567 src_eqv = GEN_INT (val);
4571 /* Set sets[i].src_elt to the class each source belongs to.
4572 Detect assignments from or to volatile things
4573 and set set[i] to zero so they will be ignored
4574 in the rest of this function.
4576 Nothing in this loop changes the hash table or the register chains. */
4578 for (i = 0; i < n_sets; i++)
4580 bool repeat = false;
4581 rtx src, dest;
4582 rtx src_folded;
4583 struct table_elt *elt = 0, *p;
4584 machine_mode mode;
4585 rtx src_eqv_here;
4586 rtx src_const = 0;
4587 rtx src_related = 0;
4588 bool src_related_is_const_anchor = false;
4589 struct table_elt *src_const_elt = 0;
4590 int src_cost = MAX_COST;
4591 int src_eqv_cost = MAX_COST;
4592 int src_folded_cost = MAX_COST;
4593 int src_related_cost = MAX_COST;
4594 int src_elt_cost = MAX_COST;
4595 int src_regcost = MAX_COST;
4596 int src_eqv_regcost = MAX_COST;
4597 int src_folded_regcost = MAX_COST;
4598 int src_related_regcost = MAX_COST;
4599 int src_elt_regcost = MAX_COST;
4600 /* Set nonzero if we need to call force_const_mem on with the
4601 contents of src_folded before using it. */
4602 int src_folded_force_flag = 0;
4604 dest = SET_DEST (sets[i].rtl);
4605 src = SET_SRC (sets[i].rtl);
4607 /* If SRC is a constant that has no machine mode,
4608 hash it with the destination's machine mode.
4609 This way we can keep different modes separate. */
4611 mode = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
4612 sets[i].mode = mode;
4614 if (src_eqv)
4616 machine_mode eqvmode = mode;
4617 if (GET_CODE (dest) == STRICT_LOW_PART)
4618 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
4619 do_not_record = 0;
4620 hash_arg_in_memory = 0;
4621 src_eqv_hash = HASH (src_eqv, eqvmode);
4623 /* Find the equivalence class for the equivalent expression. */
4625 if (!do_not_record)
4626 src_eqv_elt = lookup (src_eqv, src_eqv_hash, eqvmode);
4628 src_eqv_volatile = do_not_record;
4629 src_eqv_in_memory = hash_arg_in_memory;
4632 /* If this is a STRICT_LOW_PART assignment, src_eqv corresponds to the
4633 value of the INNER register, not the destination. So it is not
4634 a valid substitution for the source. But save it for later. */
4635 if (GET_CODE (dest) == STRICT_LOW_PART)
4636 src_eqv_here = 0;
4637 else
4638 src_eqv_here = src_eqv;
4640 /* Simplify and foldable subexpressions in SRC. Then get the fully-
4641 simplified result, which may not necessarily be valid. */
4642 src_folded = fold_rtx (src, insn);
4644 #if 0
4645 /* ??? This caused bad code to be generated for the m68k port with -O2.
4646 Suppose src is (CONST_INT -1), and that after truncation src_folded
4647 is (CONST_INT 3). Suppose src_folded is then used for src_const.
4648 At the end we will add src and src_const to the same equivalence
4649 class. We now have 3 and -1 on the same equivalence class. This
4650 causes later instructions to be mis-optimized. */
4651 /* If storing a constant in a bitfield, pre-truncate the constant
4652 so we will be able to record it later. */
4653 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
4655 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
4657 if (CONST_INT_P (src)
4658 && CONST_INT_P (width)
4659 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
4660 && (INTVAL (src) & ((HOST_WIDE_INT) (-1) << INTVAL (width))))
4661 src_folded
4662 = GEN_INT (INTVAL (src) & (((HOST_WIDE_INT) 1
4663 << INTVAL (width)) - 1));
4665 #endif
4667 /* Compute SRC's hash code, and also notice if it
4668 should not be recorded at all. In that case,
4669 prevent any further processing of this assignment. */
4670 do_not_record = 0;
4671 hash_arg_in_memory = 0;
4673 sets[i].src = src;
4674 sets[i].src_hash = HASH (src, mode);
4675 sets[i].src_volatile = do_not_record;
4676 sets[i].src_in_memory = hash_arg_in_memory;
4678 /* If SRC is a MEM, there is a REG_EQUIV note for SRC, and DEST is
4679 a pseudo, do not record SRC. Using SRC as a replacement for
4680 anything else will be incorrect in that situation. Note that
4681 this usually occurs only for stack slots, in which case all the
4682 RTL would be referring to SRC, so we don't lose any optimization
4683 opportunities by not having SRC in the hash table. */
4685 if (MEM_P (src)
4686 && find_reg_note (insn, REG_EQUIV, NULL_RTX) != 0
4687 && REG_P (dest)
4688 && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
4689 sets[i].src_volatile = 1;
4691 else if (GET_CODE (src) == ASM_OPERANDS
4692 && GET_CODE (x) == PARALLEL)
4694 /* Do not record result of a non-volatile inline asm with
4695 more than one result. */
4696 if (n_sets > 1)
4697 sets[i].src_volatile = 1;
4699 int j, lim = XVECLEN (x, 0);
4700 for (j = 0; j < lim; j++)
4702 rtx y = XVECEXP (x, 0, j);
4703 /* And do not record result of a non-volatile inline asm
4704 with "memory" clobber. */
4705 if (GET_CODE (y) == CLOBBER && MEM_P (XEXP (y, 0)))
4707 sets[i].src_volatile = 1;
4708 break;
4713 #if 0
4714 /* It is no longer clear why we used to do this, but it doesn't
4715 appear to still be needed. So let's try without it since this
4716 code hurts cse'ing widened ops. */
4717 /* If source is a paradoxical subreg (such as QI treated as an SI),
4718 treat it as volatile. It may do the work of an SI in one context
4719 where the extra bits are not being used, but cannot replace an SI
4720 in general. */
4721 if (paradoxical_subreg_p (src))
4722 sets[i].src_volatile = 1;
4723 #endif
4725 /* Locate all possible equivalent forms for SRC. Try to replace
4726 SRC in the insn with each cheaper equivalent.
4728 We have the following types of equivalents: SRC itself, a folded
4729 version, a value given in a REG_EQUAL note, or a value related
4730 to a constant.
4732 Each of these equivalents may be part of an additional class
4733 of equivalents (if more than one is in the table, they must be in
4734 the same class; we check for this).
4736 If the source is volatile, we don't do any table lookups.
4738 We note any constant equivalent for possible later use in a
4739 REG_NOTE. */
4741 if (!sets[i].src_volatile)
4742 elt = lookup (src, sets[i].src_hash, mode);
4744 sets[i].src_elt = elt;
4746 if (elt && src_eqv_here && src_eqv_elt)
4748 if (elt->first_same_value != src_eqv_elt->first_same_value)
4750 /* The REG_EQUAL is indicating that two formerly distinct
4751 classes are now equivalent. So merge them. */
4752 merge_equiv_classes (elt, src_eqv_elt);
4753 src_eqv_hash = HASH (src_eqv, elt->mode);
4754 src_eqv_elt = lookup (src_eqv, src_eqv_hash, elt->mode);
4757 src_eqv_here = 0;
4760 else if (src_eqv_elt)
4761 elt = src_eqv_elt;
4763 /* Try to find a constant somewhere and record it in `src_const'.
4764 Record its table element, if any, in `src_const_elt'. Look in
4765 any known equivalences first. (If the constant is not in the
4766 table, also set `sets[i].src_const_hash'). */
4767 if (elt)
4768 for (p = elt->first_same_value; p; p = p->next_same_value)
4769 if (p->is_const)
4771 src_const = p->exp;
4772 src_const_elt = elt;
4773 break;
4776 if (src_const == 0
4777 && (CONSTANT_P (src_folded)
4778 /* Consider (minus (label_ref L1) (label_ref L2)) as
4779 "constant" here so we will record it. This allows us
4780 to fold switch statements when an ADDR_DIFF_VEC is used. */
4781 || (GET_CODE (src_folded) == MINUS
4782 && GET_CODE (XEXP (src_folded, 0)) == LABEL_REF
4783 && GET_CODE (XEXP (src_folded, 1)) == LABEL_REF)))
4784 src_const = src_folded, src_const_elt = elt;
4785 else if (src_const == 0 && src_eqv_here && CONSTANT_P (src_eqv_here))
4786 src_const = src_eqv_here, src_const_elt = src_eqv_elt;
4788 /* If we don't know if the constant is in the table, get its
4789 hash code and look it up. */
4790 if (src_const && src_const_elt == 0)
4792 sets[i].src_const_hash = HASH (src_const, mode);
4793 src_const_elt = lookup (src_const, sets[i].src_const_hash, mode);
4796 sets[i].src_const = src_const;
4797 sets[i].src_const_elt = src_const_elt;
4799 /* If the constant and our source are both in the table, mark them as
4800 equivalent. Otherwise, if a constant is in the table but the source
4801 isn't, set ELT to it. */
4802 if (src_const_elt && elt
4803 && src_const_elt->first_same_value != elt->first_same_value)
4804 merge_equiv_classes (elt, src_const_elt);
4805 else if (src_const_elt && elt == 0)
4806 elt = src_const_elt;
4808 /* See if there is a register linearly related to a constant
4809 equivalent of SRC. */
4810 if (src_const
4811 && (GET_CODE (src_const) == CONST
4812 || (src_const_elt && src_const_elt->related_value != 0)))
4814 src_related = use_related_value (src_const, src_const_elt);
4815 if (src_related)
4817 struct table_elt *src_related_elt
4818 = lookup (src_related, HASH (src_related, mode), mode);
4819 if (src_related_elt && elt)
4821 if (elt->first_same_value
4822 != src_related_elt->first_same_value)
4823 /* This can occur when we previously saw a CONST
4824 involving a SYMBOL_REF and then see the SYMBOL_REF
4825 twice. Merge the involved classes. */
4826 merge_equiv_classes (elt, src_related_elt);
4828 src_related = 0;
4829 src_related_elt = 0;
4831 else if (src_related_elt && elt == 0)
4832 elt = src_related_elt;
4836 /* See if we have a CONST_INT that is already in a register in a
4837 wider mode. */
4839 if (src_const && src_related == 0 && CONST_INT_P (src_const)
4840 && GET_MODE_CLASS (mode) == MODE_INT
4841 && GET_MODE_PRECISION (mode) < BITS_PER_WORD)
4843 machine_mode wider_mode;
4845 for (wider_mode = GET_MODE_WIDER_MODE (mode);
4846 wider_mode != VOIDmode
4847 && GET_MODE_PRECISION (wider_mode) <= BITS_PER_WORD
4848 && src_related == 0;
4849 wider_mode = GET_MODE_WIDER_MODE (wider_mode))
4851 struct table_elt *const_elt
4852 = lookup (src_const, HASH (src_const, wider_mode), wider_mode);
4854 if (const_elt == 0)
4855 continue;
4857 for (const_elt = const_elt->first_same_value;
4858 const_elt; const_elt = const_elt->next_same_value)
4859 if (REG_P (const_elt->exp))
4861 src_related = gen_lowpart (mode, const_elt->exp);
4862 break;
4867 /* Another possibility is that we have an AND with a constant in
4868 a mode narrower than a word. If so, it might have been generated
4869 as part of an "if" which would narrow the AND. If we already
4870 have done the AND in a wider mode, we can use a SUBREG of that
4871 value. */
4873 if (flag_expensive_optimizations && ! src_related
4874 && GET_CODE (src) == AND && CONST_INT_P (XEXP (src, 1))
4875 && GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4877 machine_mode tmode;
4878 rtx new_and = gen_rtx_AND (VOIDmode, NULL_RTX, XEXP (src, 1));
4880 for (tmode = GET_MODE_WIDER_MODE (mode);
4881 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4882 tmode = GET_MODE_WIDER_MODE (tmode))
4884 rtx inner = gen_lowpart (tmode, XEXP (src, 0));
4885 struct table_elt *larger_elt;
4887 if (inner)
4889 PUT_MODE (new_and, tmode);
4890 XEXP (new_and, 0) = inner;
4891 larger_elt = lookup (new_and, HASH (new_and, tmode), tmode);
4892 if (larger_elt == 0)
4893 continue;
4895 for (larger_elt = larger_elt->first_same_value;
4896 larger_elt; larger_elt = larger_elt->next_same_value)
4897 if (REG_P (larger_elt->exp))
4899 src_related
4900 = gen_lowpart (mode, larger_elt->exp);
4901 break;
4904 if (src_related)
4905 break;
4910 /* See if a MEM has already been loaded with a widening operation;
4911 if it has, we can use a subreg of that. Many CISC machines
4912 also have such operations, but this is only likely to be
4913 beneficial on these machines. */
4915 if (flag_expensive_optimizations && src_related == 0
4916 && (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
4917 && GET_MODE_CLASS (mode) == MODE_INT
4918 && MEM_P (src) && ! do_not_record
4919 && LOAD_EXTEND_OP (mode) != UNKNOWN)
4921 struct rtx_def memory_extend_buf;
4922 rtx memory_extend_rtx = &memory_extend_buf;
4923 machine_mode tmode;
4925 /* Set what we are trying to extend and the operation it might
4926 have been extended with. */
4927 memset (memory_extend_rtx, 0, sizeof (*memory_extend_rtx));
4928 PUT_CODE (memory_extend_rtx, LOAD_EXTEND_OP (mode));
4929 XEXP (memory_extend_rtx, 0) = src;
4931 for (tmode = GET_MODE_WIDER_MODE (mode);
4932 GET_MODE_SIZE (tmode) <= UNITS_PER_WORD;
4933 tmode = GET_MODE_WIDER_MODE (tmode))
4935 struct table_elt *larger_elt;
4937 PUT_MODE (memory_extend_rtx, tmode);
4938 larger_elt = lookup (memory_extend_rtx,
4939 HASH (memory_extend_rtx, tmode), tmode);
4940 if (larger_elt == 0)
4941 continue;
4943 for (larger_elt = larger_elt->first_same_value;
4944 larger_elt; larger_elt = larger_elt->next_same_value)
4945 if (REG_P (larger_elt->exp))
4947 src_related = gen_lowpart (mode, larger_elt->exp);
4948 break;
4951 if (src_related)
4952 break;
4956 /* Try to express the constant using a register+offset expression
4957 derived from a constant anchor. */
4959 if (targetm.const_anchor
4960 && !src_related
4961 && src_const
4962 && GET_CODE (src_const) == CONST_INT)
4964 src_related = try_const_anchors (src_const, mode);
4965 src_related_is_const_anchor = src_related != NULL_RTX;
4969 if (src == src_folded)
4970 src_folded = 0;
4972 /* At this point, ELT, if nonzero, points to a class of expressions
4973 equivalent to the source of this SET and SRC, SRC_EQV, SRC_FOLDED,
4974 and SRC_RELATED, if nonzero, each contain additional equivalent
4975 expressions. Prune these latter expressions by deleting expressions
4976 already in the equivalence class.
4978 Check for an equivalent identical to the destination. If found,
4979 this is the preferred equivalent since it will likely lead to
4980 elimination of the insn. Indicate this by placing it in
4981 `src_related'. */
4983 if (elt)
4984 elt = elt->first_same_value;
4985 for (p = elt; p; p = p->next_same_value)
4987 enum rtx_code code = GET_CODE (p->exp);
4989 /* If the expression is not valid, ignore it. Then we do not
4990 have to check for validity below. In most cases, we can use
4991 `rtx_equal_p', since canonicalization has already been done. */
4992 if (code != REG && ! exp_equiv_p (p->exp, p->exp, 1, false))
4993 continue;
4995 /* Also skip paradoxical subregs, unless that's what we're
4996 looking for. */
4997 if (paradoxical_subreg_p (p->exp)
4998 && ! (src != 0
4999 && GET_CODE (src) == SUBREG
5000 && GET_MODE (src) == GET_MODE (p->exp)
5001 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5002 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (p->exp))))))
5003 continue;
5005 if (src && GET_CODE (src) == code && rtx_equal_p (src, p->exp))
5006 src = 0;
5007 else if (src_folded && GET_CODE (src_folded) == code
5008 && rtx_equal_p (src_folded, p->exp))
5009 src_folded = 0;
5010 else if (src_eqv_here && GET_CODE (src_eqv_here) == code
5011 && rtx_equal_p (src_eqv_here, p->exp))
5012 src_eqv_here = 0;
5013 else if (src_related && GET_CODE (src_related) == code
5014 && rtx_equal_p (src_related, p->exp))
5015 src_related = 0;
5017 /* This is the same as the destination of the insns, we want
5018 to prefer it. Copy it to src_related. The code below will
5019 then give it a negative cost. */
5020 if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
5021 src_related = dest;
5024 /* Find the cheapest valid equivalent, trying all the available
5025 possibilities. Prefer items not in the hash table to ones
5026 that are when they are equal cost. Note that we can never
5027 worsen an insn as the current contents will also succeed.
5028 If we find an equivalent identical to the destination, use it as best,
5029 since this insn will probably be eliminated in that case. */
5030 if (src)
5032 if (rtx_equal_p (src, dest))
5033 src_cost = src_regcost = -1;
5034 else
5036 src_cost = COST (src, mode);
5037 src_regcost = approx_reg_cost (src);
5041 if (src_eqv_here)
5043 if (rtx_equal_p (src_eqv_here, dest))
5044 src_eqv_cost = src_eqv_regcost = -1;
5045 else
5047 src_eqv_cost = COST (src_eqv_here, mode);
5048 src_eqv_regcost = approx_reg_cost (src_eqv_here);
5052 if (src_folded)
5054 if (rtx_equal_p (src_folded, dest))
5055 src_folded_cost = src_folded_regcost = -1;
5056 else
5058 src_folded_cost = COST (src_folded, mode);
5059 src_folded_regcost = approx_reg_cost (src_folded);
5063 if (src_related)
5065 if (rtx_equal_p (src_related, dest))
5066 src_related_cost = src_related_regcost = -1;
5067 else
5069 src_related_cost = COST (src_related, mode);
5070 src_related_regcost = approx_reg_cost (src_related);
5072 /* If a const-anchor is used to synthesize a constant that
5073 normally requires multiple instructions then slightly prefer
5074 it over the original sequence. These instructions are likely
5075 to become redundant now. We can't compare against the cost
5076 of src_eqv_here because, on MIPS for example, multi-insn
5077 constants have zero cost; they are assumed to be hoisted from
5078 loops. */
5079 if (src_related_is_const_anchor
5080 && src_related_cost == src_cost
5081 && src_eqv_here)
5082 src_related_cost--;
5086 /* If this was an indirect jump insn, a known label will really be
5087 cheaper even though it looks more expensive. */
5088 if (dest == pc_rtx && src_const && GET_CODE (src_const) == LABEL_REF)
5089 src_folded = src_const, src_folded_cost = src_folded_regcost = -1;
5091 /* Terminate loop when replacement made. This must terminate since
5092 the current contents will be tested and will always be valid. */
5093 while (1)
5095 rtx trial;
5097 /* Skip invalid entries. */
5098 while (elt && !REG_P (elt->exp)
5099 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5100 elt = elt->next_same_value;
5102 /* A paradoxical subreg would be bad here: it'll be the right
5103 size, but later may be adjusted so that the upper bits aren't
5104 what we want. So reject it. */
5105 if (elt != 0
5106 && paradoxical_subreg_p (elt->exp)
5107 /* It is okay, though, if the rtx we're trying to match
5108 will ignore any of the bits we can't predict. */
5109 && ! (src != 0
5110 && GET_CODE (src) == SUBREG
5111 && GET_MODE (src) == GET_MODE (elt->exp)
5112 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5113 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (elt->exp))))))
5115 elt = elt->next_same_value;
5116 continue;
5119 if (elt)
5121 src_elt_cost = elt->cost;
5122 src_elt_regcost = elt->regcost;
5125 /* Find cheapest and skip it for the next time. For items
5126 of equal cost, use this order:
5127 src_folded, src, src_eqv, src_related and hash table entry. */
5128 if (src_folded
5129 && preferable (src_folded_cost, src_folded_regcost,
5130 src_cost, src_regcost) <= 0
5131 && preferable (src_folded_cost, src_folded_regcost,
5132 src_eqv_cost, src_eqv_regcost) <= 0
5133 && preferable (src_folded_cost, src_folded_regcost,
5134 src_related_cost, src_related_regcost) <= 0
5135 && preferable (src_folded_cost, src_folded_regcost,
5136 src_elt_cost, src_elt_regcost) <= 0)
5138 trial = src_folded, src_folded_cost = MAX_COST;
5139 if (src_folded_force_flag)
5141 rtx forced = force_const_mem (mode, trial);
5142 if (forced)
5143 trial = forced;
5146 else if (src
5147 && preferable (src_cost, src_regcost,
5148 src_eqv_cost, src_eqv_regcost) <= 0
5149 && preferable (src_cost, src_regcost,
5150 src_related_cost, src_related_regcost) <= 0
5151 && preferable (src_cost, src_regcost,
5152 src_elt_cost, src_elt_regcost) <= 0)
5153 trial = src, src_cost = MAX_COST;
5154 else if (src_eqv_here
5155 && preferable (src_eqv_cost, src_eqv_regcost,
5156 src_related_cost, src_related_regcost) <= 0
5157 && preferable (src_eqv_cost, src_eqv_regcost,
5158 src_elt_cost, src_elt_regcost) <= 0)
5159 trial = src_eqv_here, src_eqv_cost = MAX_COST;
5160 else if (src_related
5161 && preferable (src_related_cost, src_related_regcost,
5162 src_elt_cost, src_elt_regcost) <= 0)
5163 trial = src_related, src_related_cost = MAX_COST;
5164 else
5166 trial = elt->exp;
5167 elt = elt->next_same_value;
5168 src_elt_cost = MAX_COST;
5171 /* Avoid creation of overlapping memory moves. */
5172 if (MEM_P (trial) && MEM_P (SET_DEST (sets[i].rtl)))
5174 rtx src, dest;
5176 /* BLKmode moves are not handled by cse anyway. */
5177 if (GET_MODE (trial) == BLKmode)
5178 break;
5180 src = canon_rtx (trial);
5181 dest = canon_rtx (SET_DEST (sets[i].rtl));
5183 if (!MEM_P (src) || !MEM_P (dest)
5184 || !nonoverlapping_memrefs_p (src, dest, false))
5185 break;
5188 /* Try to optimize
5189 (set (reg:M N) (const_int A))
5190 (set (reg:M2 O) (const_int B))
5191 (set (zero_extract:M2 (reg:M N) (const_int C) (const_int D))
5192 (reg:M2 O)). */
5193 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT
5194 && CONST_INT_P (trial)
5195 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 1))
5196 && CONST_INT_P (XEXP (SET_DEST (sets[i].rtl), 2))
5197 && REG_P (XEXP (SET_DEST (sets[i].rtl), 0))
5198 && (GET_MODE_PRECISION (GET_MODE (SET_DEST (sets[i].rtl)))
5199 >= INTVAL (XEXP (SET_DEST (sets[i].rtl), 1)))
5200 && ((unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 1))
5201 + (unsigned) INTVAL (XEXP (SET_DEST (sets[i].rtl), 2))
5202 <= HOST_BITS_PER_WIDE_INT))
5204 rtx dest_reg = XEXP (SET_DEST (sets[i].rtl), 0);
5205 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5206 rtx pos = XEXP (SET_DEST (sets[i].rtl), 2);
5207 unsigned int dest_hash = HASH (dest_reg, GET_MODE (dest_reg));
5208 struct table_elt *dest_elt
5209 = lookup (dest_reg, dest_hash, GET_MODE (dest_reg));
5210 rtx dest_cst = NULL;
5212 if (dest_elt)
5213 for (p = dest_elt->first_same_value; p; p = p->next_same_value)
5214 if (p->is_const && CONST_INT_P (p->exp))
5216 dest_cst = p->exp;
5217 break;
5219 if (dest_cst)
5221 HOST_WIDE_INT val = INTVAL (dest_cst);
5222 HOST_WIDE_INT mask;
5223 unsigned int shift;
5224 if (BITS_BIG_ENDIAN)
5225 shift = GET_MODE_PRECISION (GET_MODE (dest_reg))
5226 - INTVAL (pos) - INTVAL (width);
5227 else
5228 shift = INTVAL (pos);
5229 if (INTVAL (width) == HOST_BITS_PER_WIDE_INT)
5230 mask = ~(HOST_WIDE_INT) 0;
5231 else
5232 mask = ((HOST_WIDE_INT) 1 << INTVAL (width)) - 1;
5233 val &= ~(mask << shift);
5234 val |= (INTVAL (trial) & mask) << shift;
5235 val = trunc_int_for_mode (val, GET_MODE (dest_reg));
5236 validate_unshare_change (insn, &SET_DEST (sets[i].rtl),
5237 dest_reg, 1);
5238 validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5239 GEN_INT (val), 1);
5240 if (apply_change_group ())
5242 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5243 if (note)
5245 remove_note (insn, note);
5246 df_notes_rescan (insn);
5248 src_eqv = NULL_RTX;
5249 src_eqv_elt = NULL;
5250 src_eqv_volatile = 0;
5251 src_eqv_in_memory = 0;
5252 src_eqv_hash = 0;
5253 repeat = true;
5254 break;
5259 /* We don't normally have an insn matching (set (pc) (pc)), so
5260 check for this separately here. We will delete such an
5261 insn below.
5263 For other cases such as a table jump or conditional jump
5264 where we know the ultimate target, go ahead and replace the
5265 operand. While that may not make a valid insn, we will
5266 reemit the jump below (and also insert any necessary
5267 barriers). */
5268 if (n_sets == 1 && dest == pc_rtx
5269 && (trial == pc_rtx
5270 || (GET_CODE (trial) == LABEL_REF
5271 && ! condjump_p (insn))))
5273 /* Don't substitute non-local labels, this confuses CFG. */
5274 if (GET_CODE (trial) == LABEL_REF
5275 && LABEL_REF_NONLOCAL_P (trial))
5276 continue;
5278 SET_SRC (sets[i].rtl) = trial;
5279 cse_jumps_altered = true;
5280 break;
5283 /* Reject certain invalid forms of CONST that we create. */
5284 else if (CONSTANT_P (trial)
5285 && GET_CODE (trial) == CONST
5286 /* Reject cases that will cause decode_rtx_const to
5287 die. On the alpha when simplifying a switch, we
5288 get (const (truncate (minus (label_ref)
5289 (label_ref)))). */
5290 && (GET_CODE (XEXP (trial, 0)) == TRUNCATE
5291 /* Likewise on IA-64, except without the
5292 truncate. */
5293 || (GET_CODE (XEXP (trial, 0)) == MINUS
5294 && GET_CODE (XEXP (XEXP (trial, 0), 0)) == LABEL_REF
5295 && GET_CODE (XEXP (XEXP (trial, 0), 1)) == LABEL_REF)))
5296 /* Do nothing for this case. */
5299 /* Look for a substitution that makes a valid insn. */
5300 else if (validate_unshare_change (insn, &SET_SRC (sets[i].rtl),
5301 trial, 0))
5303 rtx new_rtx = canon_reg (SET_SRC (sets[i].rtl), insn);
5305 /* The result of apply_change_group can be ignored; see
5306 canon_reg. */
5308 validate_change (insn, &SET_SRC (sets[i].rtl), new_rtx, 1);
5309 apply_change_group ();
5311 break;
5314 /* If we previously found constant pool entries for
5315 constants and this is a constant, try making a
5316 pool entry. Put it in src_folded unless we already have done
5317 this since that is where it likely came from. */
5319 else if (constant_pool_entries_cost
5320 && CONSTANT_P (trial)
5321 && (src_folded == 0
5322 || (!MEM_P (src_folded)
5323 && ! src_folded_force_flag))
5324 && GET_MODE_CLASS (mode) != MODE_CC
5325 && mode != VOIDmode)
5327 src_folded_force_flag = 1;
5328 src_folded = trial;
5329 src_folded_cost = constant_pool_entries_cost;
5330 src_folded_regcost = constant_pool_entries_regcost;
5334 /* If we changed the insn too much, handle this set from scratch. */
5335 if (repeat)
5337 i--;
5338 continue;
5341 src = SET_SRC (sets[i].rtl);
5343 /* In general, it is good to have a SET with SET_SRC == SET_DEST.
5344 However, there is an important exception: If both are registers
5345 that are not the head of their equivalence class, replace SET_SRC
5346 with the head of the class. If we do not do this, we will have
5347 both registers live over a portion of the basic block. This way,
5348 their lifetimes will likely abut instead of overlapping. */
5349 if (REG_P (dest)
5350 && REGNO_QTY_VALID_P (REGNO (dest)))
5352 int dest_q = REG_QTY (REGNO (dest));
5353 struct qty_table_elem *dest_ent = &qty_table[dest_q];
5355 if (dest_ent->mode == GET_MODE (dest)
5356 && dest_ent->first_reg != REGNO (dest)
5357 && REG_P (src) && REGNO (src) == REGNO (dest)
5358 /* Don't do this if the original insn had a hard reg as
5359 SET_SRC or SET_DEST. */
5360 && (!REG_P (sets[i].src)
5361 || REGNO (sets[i].src) >= FIRST_PSEUDO_REGISTER)
5362 && (!REG_P (dest) || REGNO (dest) >= FIRST_PSEUDO_REGISTER))
5363 /* We can't call canon_reg here because it won't do anything if
5364 SRC is a hard register. */
5366 int src_q = REG_QTY (REGNO (src));
5367 struct qty_table_elem *src_ent = &qty_table[src_q];
5368 int first = src_ent->first_reg;
5369 rtx new_src
5370 = (first >= FIRST_PSEUDO_REGISTER
5371 ? regno_reg_rtx[first] : gen_rtx_REG (GET_MODE (src), first));
5373 /* We must use validate-change even for this, because this
5374 might be a special no-op instruction, suitable only to
5375 tag notes onto. */
5376 if (validate_change (insn, &SET_SRC (sets[i].rtl), new_src, 0))
5378 src = new_src;
5379 /* If we had a constant that is cheaper than what we are now
5380 setting SRC to, use that constant. We ignored it when we
5381 thought we could make this into a no-op. */
5382 if (src_const && COST (src_const, mode) < COST (src, mode)
5383 && validate_change (insn, &SET_SRC (sets[i].rtl),
5384 src_const, 0))
5385 src = src_const;
5390 /* If we made a change, recompute SRC values. */
5391 if (src != sets[i].src)
5393 do_not_record = 0;
5394 hash_arg_in_memory = 0;
5395 sets[i].src = src;
5396 sets[i].src_hash = HASH (src, mode);
5397 sets[i].src_volatile = do_not_record;
5398 sets[i].src_in_memory = hash_arg_in_memory;
5399 sets[i].src_elt = lookup (src, sets[i].src_hash, mode);
5402 /* If this is a single SET, we are setting a register, and we have an
5403 equivalent constant, we want to add a REG_EQUAL note if the constant
5404 is different from the source. We don't want to do it for a constant
5405 pseudo since verifying that this pseudo hasn't been eliminated is a
5406 pain; moreover such a note won't help anything.
5408 Avoid a REG_EQUAL note for (CONST (MINUS (LABEL_REF) (LABEL_REF)))
5409 which can be created for a reference to a compile time computable
5410 entry in a jump table. */
5411 if (n_sets == 1
5412 && REG_P (dest)
5413 && src_const
5414 && !REG_P (src_const)
5415 && !(GET_CODE (src_const) == SUBREG
5416 && REG_P (SUBREG_REG (src_const)))
5417 && !(GET_CODE (src_const) == CONST
5418 && GET_CODE (XEXP (src_const, 0)) == MINUS
5419 && GET_CODE (XEXP (XEXP (src_const, 0), 0)) == LABEL_REF
5420 && GET_CODE (XEXP (XEXP (src_const, 0), 1)) == LABEL_REF)
5421 && !rtx_equal_p (src, src_const))
5423 /* Make sure that the rtx is not shared. */
5424 src_const = copy_rtx (src_const);
5426 /* Record the actual constant value in a REG_EQUAL note,
5427 making a new one if one does not already exist. */
5428 set_unique_reg_note (insn, REG_EQUAL, src_const);
5429 df_notes_rescan (insn);
5432 /* Now deal with the destination. */
5433 do_not_record = 0;
5435 /* Look within any ZERO_EXTRACT to the MEM or REG within it. */
5436 while (GET_CODE (dest) == SUBREG
5437 || GET_CODE (dest) == ZERO_EXTRACT
5438 || GET_CODE (dest) == STRICT_LOW_PART)
5439 dest = XEXP (dest, 0);
5441 sets[i].inner_dest = dest;
5443 if (MEM_P (dest))
5445 #ifdef PUSH_ROUNDING
5446 /* Stack pushes invalidate the stack pointer. */
5447 rtx addr = XEXP (dest, 0);
5448 if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC
5449 && XEXP (addr, 0) == stack_pointer_rtx)
5450 invalidate (stack_pointer_rtx, VOIDmode);
5451 #endif
5452 dest = fold_rtx (dest, insn);
5455 /* Compute the hash code of the destination now,
5456 before the effects of this instruction are recorded,
5457 since the register values used in the address computation
5458 are those before this instruction. */
5459 sets[i].dest_hash = HASH (dest, mode);
5461 /* Don't enter a bit-field in the hash table
5462 because the value in it after the store
5463 may not equal what was stored, due to truncation. */
5465 if (GET_CODE (SET_DEST (sets[i].rtl)) == ZERO_EXTRACT)
5467 rtx width = XEXP (SET_DEST (sets[i].rtl), 1);
5469 if (src_const != 0 && CONST_INT_P (src_const)
5470 && CONST_INT_P (width)
5471 && INTVAL (width) < HOST_BITS_PER_WIDE_INT
5472 && ! (INTVAL (src_const)
5473 & (HOST_WIDE_INT_M1U << INTVAL (width))))
5474 /* Exception: if the value is constant,
5475 and it won't be truncated, record it. */
5477 else
5479 /* This is chosen so that the destination will be invalidated
5480 but no new value will be recorded.
5481 We must invalidate because sometimes constant
5482 values can be recorded for bitfields. */
5483 sets[i].src_elt = 0;
5484 sets[i].src_volatile = 1;
5485 src_eqv = 0;
5486 src_eqv_elt = 0;
5490 /* If only one set in a JUMP_INSN and it is now a no-op, we can delete
5491 the insn. */
5492 else if (n_sets == 1 && dest == pc_rtx && src == pc_rtx)
5494 /* One less use of the label this insn used to jump to. */
5495 delete_insn_and_edges (insn);
5496 cse_jumps_altered = true;
5497 /* No more processing for this set. */
5498 sets[i].rtl = 0;
5501 /* If this SET is now setting PC to a label, we know it used to
5502 be a conditional or computed branch. */
5503 else if (dest == pc_rtx && GET_CODE (src) == LABEL_REF
5504 && !LABEL_REF_NONLOCAL_P (src))
5506 /* We reemit the jump in as many cases as possible just in
5507 case the form of an unconditional jump is significantly
5508 different than a computed jump or conditional jump.
5510 If this insn has multiple sets, then reemitting the
5511 jump is nontrivial. So instead we just force rerecognition
5512 and hope for the best. */
5513 if (n_sets == 1)
5515 rtx_jump_insn *new_rtx;
5516 rtx note;
5518 rtx_insn *seq = targetm.gen_jump (XEXP (src, 0));
5519 new_rtx = emit_jump_insn_before (seq, insn);
5520 JUMP_LABEL (new_rtx) = XEXP (src, 0);
5521 LABEL_NUSES (XEXP (src, 0))++;
5523 /* Make sure to copy over REG_NON_LOCAL_GOTO. */
5524 note = find_reg_note (insn, REG_NON_LOCAL_GOTO, 0);
5525 if (note)
5527 XEXP (note, 1) = NULL_RTX;
5528 REG_NOTES (new_rtx) = note;
5531 delete_insn_and_edges (insn);
5532 insn = new_rtx;
5534 else
5535 INSN_CODE (insn) = -1;
5537 /* Do not bother deleting any unreachable code, let jump do it. */
5538 cse_jumps_altered = true;
5539 sets[i].rtl = 0;
5542 /* If destination is volatile, invalidate it and then do no further
5543 processing for this assignment. */
5545 else if (do_not_record)
5547 invalidate_dest (dest);
5548 sets[i].rtl = 0;
5551 if (sets[i].rtl != 0 && dest != SET_DEST (sets[i].rtl))
5553 do_not_record = 0;
5554 sets[i].dest_hash = HASH (SET_DEST (sets[i].rtl), mode);
5555 if (do_not_record)
5557 invalidate_dest (SET_DEST (sets[i].rtl));
5558 sets[i].rtl = 0;
5562 /* If setting CC0, record what it was set to, or a constant, if it
5563 is equivalent to a constant. If it is being set to a floating-point
5564 value, make a COMPARE with the appropriate constant of 0. If we
5565 don't do this, later code can interpret this as a test against
5566 const0_rtx, which can cause problems if we try to put it into an
5567 insn as a floating-point operand. */
5568 if (dest == cc0_rtx)
5570 this_insn_cc0 = src_const && mode != VOIDmode ? src_const : src;
5571 this_insn_cc0_mode = mode;
5572 if (FLOAT_MODE_P (mode))
5573 this_insn_cc0 = gen_rtx_COMPARE (VOIDmode, this_insn_cc0,
5574 CONST0_RTX (mode));
5578 /* Now enter all non-volatile source expressions in the hash table
5579 if they are not already present.
5580 Record their equivalence classes in src_elt.
5581 This way we can insert the corresponding destinations into
5582 the same classes even if the actual sources are no longer in them
5583 (having been invalidated). */
5585 if (src_eqv && src_eqv_elt == 0 && sets[0].rtl != 0 && ! src_eqv_volatile
5586 && ! rtx_equal_p (src_eqv, SET_DEST (sets[0].rtl)))
5588 struct table_elt *elt;
5589 struct table_elt *classp = sets[0].src_elt;
5590 rtx dest = SET_DEST (sets[0].rtl);
5591 machine_mode eqvmode = GET_MODE (dest);
5593 if (GET_CODE (dest) == STRICT_LOW_PART)
5595 eqvmode = GET_MODE (SUBREG_REG (XEXP (dest, 0)));
5596 classp = 0;
5598 if (insert_regs (src_eqv, classp, 0))
5600 rehash_using_reg (src_eqv);
5601 src_eqv_hash = HASH (src_eqv, eqvmode);
5603 elt = insert (src_eqv, classp, src_eqv_hash, eqvmode);
5604 elt->in_memory = src_eqv_in_memory;
5605 src_eqv_elt = elt;
5607 /* Check to see if src_eqv_elt is the same as a set source which
5608 does not yet have an elt, and if so set the elt of the set source
5609 to src_eqv_elt. */
5610 for (i = 0; i < n_sets; i++)
5611 if (sets[i].rtl && sets[i].src_elt == 0
5612 && rtx_equal_p (SET_SRC (sets[i].rtl), src_eqv))
5613 sets[i].src_elt = src_eqv_elt;
5616 for (i = 0; i < n_sets; i++)
5617 if (sets[i].rtl && ! sets[i].src_volatile
5618 && ! rtx_equal_p (SET_SRC (sets[i].rtl), SET_DEST (sets[i].rtl)))
5620 if (GET_CODE (SET_DEST (sets[i].rtl)) == STRICT_LOW_PART)
5622 /* REG_EQUAL in setting a STRICT_LOW_PART
5623 gives an equivalent for the entire destination register,
5624 not just for the subreg being stored in now.
5625 This is a more interesting equivalence, so we arrange later
5626 to treat the entire reg as the destination. */
5627 sets[i].src_elt = src_eqv_elt;
5628 sets[i].src_hash = src_eqv_hash;
5630 else
5632 /* Insert source and constant equivalent into hash table, if not
5633 already present. */
5634 struct table_elt *classp = src_eqv_elt;
5635 rtx src = sets[i].src;
5636 rtx dest = SET_DEST (sets[i].rtl);
5637 machine_mode mode
5638 = GET_MODE (src) == VOIDmode ? GET_MODE (dest) : GET_MODE (src);
5640 /* It's possible that we have a source value known to be
5641 constant but don't have a REG_EQUAL note on the insn.
5642 Lack of a note will mean src_eqv_elt will be NULL. This
5643 can happen where we've generated a SUBREG to access a
5644 CONST_INT that is already in a register in a wider mode.
5645 Ensure that the source expression is put in the proper
5646 constant class. */
5647 if (!classp)
5648 classp = sets[i].src_const_elt;
5650 if (sets[i].src_elt == 0)
5652 struct table_elt *elt;
5654 /* Note that these insert_regs calls cannot remove
5655 any of the src_elt's, because they would have failed to
5656 match if not still valid. */
5657 if (insert_regs (src, classp, 0))
5659 rehash_using_reg (src);
5660 sets[i].src_hash = HASH (src, mode);
5662 elt = insert (src, classp, sets[i].src_hash, mode);
5663 elt->in_memory = sets[i].src_in_memory;
5664 /* If inline asm has any clobbers, ensure we only reuse
5665 existing inline asms and never try to put the ASM_OPERANDS
5666 into an insn that isn't inline asm. */
5667 if (GET_CODE (src) == ASM_OPERANDS
5668 && GET_CODE (x) == PARALLEL)
5669 elt->cost = MAX_COST;
5670 sets[i].src_elt = classp = elt;
5672 if (sets[i].src_const && sets[i].src_const_elt == 0
5673 && src != sets[i].src_const
5674 && ! rtx_equal_p (sets[i].src_const, src))
5675 sets[i].src_elt = insert (sets[i].src_const, classp,
5676 sets[i].src_const_hash, mode);
5679 else if (sets[i].src_elt == 0)
5680 /* If we did not insert the source into the hash table (e.g., it was
5681 volatile), note the equivalence class for the REG_EQUAL value, if any,
5682 so that the destination goes into that class. */
5683 sets[i].src_elt = src_eqv_elt;
5685 /* Record destination addresses in the hash table. This allows us to
5686 check if they are invalidated by other sets. */
5687 for (i = 0; i < n_sets; i++)
5689 if (sets[i].rtl)
5691 rtx x = sets[i].inner_dest;
5692 struct table_elt *elt;
5693 machine_mode mode;
5694 unsigned hash;
5696 if (MEM_P (x))
5698 x = XEXP (x, 0);
5699 mode = GET_MODE (x);
5700 hash = HASH (x, mode);
5701 elt = lookup (x, hash, mode);
5702 if (!elt)
5704 if (insert_regs (x, NULL, 0))
5706 rtx dest = SET_DEST (sets[i].rtl);
5708 rehash_using_reg (x);
5709 hash = HASH (x, mode);
5710 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5712 elt = insert (x, NULL, hash, mode);
5715 sets[i].dest_addr_elt = elt;
5717 else
5718 sets[i].dest_addr_elt = NULL;
5722 invalidate_from_clobbers (insn);
5724 /* Some registers are invalidated by subroutine calls. Memory is
5725 invalidated by non-constant calls. */
5727 if (CALL_P (insn))
5729 if (!(RTL_CONST_OR_PURE_CALL_P (insn)))
5730 invalidate_memory ();
5731 invalidate_for_call ();
5734 /* Now invalidate everything set by this instruction.
5735 If a SUBREG or other funny destination is being set,
5736 sets[i].rtl is still nonzero, so here we invalidate the reg
5737 a part of which is being set. */
5739 for (i = 0; i < n_sets; i++)
5740 if (sets[i].rtl)
5742 /* We can't use the inner dest, because the mode associated with
5743 a ZERO_EXTRACT is significant. */
5744 rtx dest = SET_DEST (sets[i].rtl);
5746 /* Needed for registers to remove the register from its
5747 previous quantity's chain.
5748 Needed for memory if this is a nonvarying address, unless
5749 we have just done an invalidate_memory that covers even those. */
5750 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5751 invalidate (dest, VOIDmode);
5752 else if (MEM_P (dest))
5753 invalidate (dest, VOIDmode);
5754 else if (GET_CODE (dest) == STRICT_LOW_PART
5755 || GET_CODE (dest) == ZERO_EXTRACT)
5756 invalidate (XEXP (dest, 0), GET_MODE (dest));
5759 /* Don't cse over a call to setjmp; on some machines (eg VAX)
5760 the regs restored by the longjmp come from a later time
5761 than the setjmp. */
5762 if (CALL_P (insn) && find_reg_note (insn, REG_SETJMP, NULL))
5764 flush_hash_table ();
5765 goto done;
5768 /* Make sure registers mentioned in destinations
5769 are safe for use in an expression to be inserted.
5770 This removes from the hash table
5771 any invalid entry that refers to one of these registers.
5773 We don't care about the return value from mention_regs because
5774 we are going to hash the SET_DEST values unconditionally. */
5776 for (i = 0; i < n_sets; i++)
5778 if (sets[i].rtl)
5780 rtx x = SET_DEST (sets[i].rtl);
5782 if (!REG_P (x))
5783 mention_regs (x);
5784 else
5786 /* We used to rely on all references to a register becoming
5787 inaccessible when a register changes to a new quantity,
5788 since that changes the hash code. However, that is not
5789 safe, since after HASH_SIZE new quantities we get a
5790 hash 'collision' of a register with its own invalid
5791 entries. And since SUBREGs have been changed not to
5792 change their hash code with the hash code of the register,
5793 it wouldn't work any longer at all. So we have to check
5794 for any invalid references lying around now.
5795 This code is similar to the REG case in mention_regs,
5796 but it knows that reg_tick has been incremented, and
5797 it leaves reg_in_table as -1 . */
5798 unsigned int regno = REGNO (x);
5799 unsigned int endregno = END_REGNO (x);
5800 unsigned int i;
5802 for (i = regno; i < endregno; i++)
5804 if (REG_IN_TABLE (i) >= 0)
5806 remove_invalid_refs (i);
5807 REG_IN_TABLE (i) = -1;
5814 /* We may have just removed some of the src_elt's from the hash table.
5815 So replace each one with the current head of the same class.
5816 Also check if destination addresses have been removed. */
5818 for (i = 0; i < n_sets; i++)
5819 if (sets[i].rtl)
5821 if (sets[i].dest_addr_elt
5822 && sets[i].dest_addr_elt->first_same_value == 0)
5824 /* The elt was removed, which means this destination is not
5825 valid after this instruction. */
5826 sets[i].rtl = NULL_RTX;
5828 else if (sets[i].src_elt && sets[i].src_elt->first_same_value == 0)
5829 /* If elt was removed, find current head of same class,
5830 or 0 if nothing remains of that class. */
5832 struct table_elt *elt = sets[i].src_elt;
5834 while (elt && elt->prev_same_value)
5835 elt = elt->prev_same_value;
5837 while (elt && elt->first_same_value == 0)
5838 elt = elt->next_same_value;
5839 sets[i].src_elt = elt ? elt->first_same_value : 0;
5843 /* Now insert the destinations into their equivalence classes. */
5845 for (i = 0; i < n_sets; i++)
5846 if (sets[i].rtl)
5848 rtx dest = SET_DEST (sets[i].rtl);
5849 struct table_elt *elt;
5851 /* Don't record value if we are not supposed to risk allocating
5852 floating-point values in registers that might be wider than
5853 memory. */
5854 if ((flag_float_store
5855 && MEM_P (dest)
5856 && FLOAT_MODE_P (GET_MODE (dest)))
5857 /* Don't record BLKmode values, because we don't know the
5858 size of it, and can't be sure that other BLKmode values
5859 have the same or smaller size. */
5860 || GET_MODE (dest) == BLKmode
5861 /* If we didn't put a REG_EQUAL value or a source into the hash
5862 table, there is no point is recording DEST. */
5863 || sets[i].src_elt == 0
5864 /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
5865 or SIGN_EXTEND, don't record DEST since it can cause
5866 some tracking to be wrong.
5868 ??? Think about this more later. */
5869 || (paradoxical_subreg_p (dest)
5870 && (GET_CODE (sets[i].src) == SIGN_EXTEND
5871 || GET_CODE (sets[i].src) == ZERO_EXTEND)))
5872 continue;
5874 /* STRICT_LOW_PART isn't part of the value BEING set,
5875 and neither is the SUBREG inside it.
5876 Note that in this case SETS[I].SRC_ELT is really SRC_EQV_ELT. */
5877 if (GET_CODE (dest) == STRICT_LOW_PART)
5878 dest = SUBREG_REG (XEXP (dest, 0));
5880 if (REG_P (dest) || GET_CODE (dest) == SUBREG)
5881 /* Registers must also be inserted into chains for quantities. */
5882 if (insert_regs (dest, sets[i].src_elt, 1))
5884 /* If `insert_regs' changes something, the hash code must be
5885 recalculated. */
5886 rehash_using_reg (dest);
5887 sets[i].dest_hash = HASH (dest, GET_MODE (dest));
5890 elt = insert (dest, sets[i].src_elt,
5891 sets[i].dest_hash, GET_MODE (dest));
5893 /* If this is a constant, insert the constant anchors with the
5894 equivalent register-offset expressions using register DEST. */
5895 if (targetm.const_anchor
5896 && REG_P (dest)
5897 && SCALAR_INT_MODE_P (GET_MODE (dest))
5898 && GET_CODE (sets[i].src_elt->exp) == CONST_INT)
5899 insert_const_anchors (dest, sets[i].src_elt->exp, GET_MODE (dest));
5901 elt->in_memory = (MEM_P (sets[i].inner_dest)
5902 && !MEM_READONLY_P (sets[i].inner_dest));
5904 /* If we have (set (subreg:m1 (reg:m2 foo) 0) (bar:m1)), M1 is no
5905 narrower than M2, and both M1 and M2 are the same number of words,
5906 we are also doing (set (reg:m2 foo) (subreg:m2 (bar:m1) 0)) so
5907 make that equivalence as well.
5909 However, BAR may have equivalences for which gen_lowpart
5910 will produce a simpler value than gen_lowpart applied to
5911 BAR (e.g., if BAR was ZERO_EXTENDed from M2), so we will scan all
5912 BAR's equivalences. If we don't get a simplified form, make
5913 the SUBREG. It will not be used in an equivalence, but will
5914 cause two similar assignments to be detected.
5916 Note the loop below will find SUBREG_REG (DEST) since we have
5917 already entered SRC and DEST of the SET in the table. */
5919 if (GET_CODE (dest) == SUBREG
5920 && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))) - 1)
5921 / UNITS_PER_WORD)
5922 == (GET_MODE_SIZE (GET_MODE (dest)) - 1) / UNITS_PER_WORD)
5923 && (GET_MODE_SIZE (GET_MODE (dest))
5924 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
5925 && sets[i].src_elt != 0)
5927 machine_mode new_mode = GET_MODE (SUBREG_REG (dest));
5928 struct table_elt *elt, *classp = 0;
5930 for (elt = sets[i].src_elt->first_same_value; elt;
5931 elt = elt->next_same_value)
5933 rtx new_src = 0;
5934 unsigned src_hash;
5935 struct table_elt *src_elt;
5936 int byte = 0;
5938 /* Ignore invalid entries. */
5939 if (!REG_P (elt->exp)
5940 && ! exp_equiv_p (elt->exp, elt->exp, 1, false))
5941 continue;
5943 /* We may have already been playing subreg games. If the
5944 mode is already correct for the destination, use it. */
5945 if (GET_MODE (elt->exp) == new_mode)
5946 new_src = elt->exp;
5947 else
5949 /* Calculate big endian correction for the SUBREG_BYTE.
5950 We have already checked that M1 (GET_MODE (dest))
5951 is not narrower than M2 (new_mode). */
5952 if (BYTES_BIG_ENDIAN)
5953 byte = (GET_MODE_SIZE (GET_MODE (dest))
5954 - GET_MODE_SIZE (new_mode));
5956 new_src = simplify_gen_subreg (new_mode, elt->exp,
5957 GET_MODE (dest), byte);
5960 /* The call to simplify_gen_subreg fails if the value
5961 is VOIDmode, yet we can't do any simplification, e.g.
5962 for EXPR_LISTs denoting function call results.
5963 It is invalid to construct a SUBREG with a VOIDmode
5964 SUBREG_REG, hence a zero new_src means we can't do
5965 this substitution. */
5966 if (! new_src)
5967 continue;
5969 src_hash = HASH (new_src, new_mode);
5970 src_elt = lookup (new_src, src_hash, new_mode);
5972 /* Put the new source in the hash table is if isn't
5973 already. */
5974 if (src_elt == 0)
5976 if (insert_regs (new_src, classp, 0))
5978 rehash_using_reg (new_src);
5979 src_hash = HASH (new_src, new_mode);
5981 src_elt = insert (new_src, classp, src_hash, new_mode);
5982 src_elt->in_memory = elt->in_memory;
5983 if (GET_CODE (new_src) == ASM_OPERANDS
5984 && elt->cost == MAX_COST)
5985 src_elt->cost = MAX_COST;
5987 else if (classp && classp != src_elt->first_same_value)
5988 /* Show that two things that we've seen before are
5989 actually the same. */
5990 merge_equiv_classes (src_elt, classp);
5992 classp = src_elt->first_same_value;
5993 /* Ignore invalid entries. */
5994 while (classp
5995 && !REG_P (classp->exp)
5996 && ! exp_equiv_p (classp->exp, classp->exp, 1, false))
5997 classp = classp->next_same_value;
6002 /* Special handling for (set REG0 REG1) where REG0 is the
6003 "cheapest", cheaper than REG1. After cse, REG1 will probably not
6004 be used in the sequel, so (if easily done) change this insn to
6005 (set REG1 REG0) and replace REG1 with REG0 in the previous insn
6006 that computed their value. Then REG1 will become a dead store
6007 and won't cloud the situation for later optimizations.
6009 Do not make this change if REG1 is a hard register, because it will
6010 then be used in the sequel and we may be changing a two-operand insn
6011 into a three-operand insn.
6013 Also do not do this if we are operating on a copy of INSN. */
6015 if (n_sets == 1 && sets[0].rtl)
6016 try_back_substitute_reg (sets[0].rtl, insn);
6018 done:;
6021 /* Remove from the hash table all expressions that reference memory. */
6023 static void
6024 invalidate_memory (void)
6026 int i;
6027 struct table_elt *p, *next;
6029 for (i = 0; i < HASH_SIZE; i++)
6030 for (p = table[i]; p; p = next)
6032 next = p->next_same_hash;
6033 if (p->in_memory)
6034 remove_from_table (p, i);
6038 /* Perform invalidation on the basis of everything about INSN,
6039 except for invalidating the actual places that are SET in it.
6040 This includes the places CLOBBERed, and anything that might
6041 alias with something that is SET or CLOBBERed. */
6043 static void
6044 invalidate_from_clobbers (rtx_insn *insn)
6046 rtx x = PATTERN (insn);
6048 if (GET_CODE (x) == CLOBBER)
6050 rtx ref = XEXP (x, 0);
6051 if (ref)
6053 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6054 || MEM_P (ref))
6055 invalidate (ref, VOIDmode);
6056 else if (GET_CODE (ref) == STRICT_LOW_PART
6057 || GET_CODE (ref) == ZERO_EXTRACT)
6058 invalidate (XEXP (ref, 0), GET_MODE (ref));
6061 else if (GET_CODE (x) == PARALLEL)
6063 int i;
6064 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6066 rtx y = XVECEXP (x, 0, i);
6067 if (GET_CODE (y) == CLOBBER)
6069 rtx ref = XEXP (y, 0);
6070 if (REG_P (ref) || GET_CODE (ref) == SUBREG
6071 || MEM_P (ref))
6072 invalidate (ref, VOIDmode);
6073 else if (GET_CODE (ref) == STRICT_LOW_PART
6074 || GET_CODE (ref) == ZERO_EXTRACT)
6075 invalidate (XEXP (ref, 0), GET_MODE (ref));
6081 /* Perform invalidation on the basis of everything about INSN.
6082 This includes the places CLOBBERed, and anything that might
6083 alias with something that is SET or CLOBBERed. */
6085 static void
6086 invalidate_from_sets_and_clobbers (rtx_insn *insn)
6088 rtx tem;
6089 rtx x = PATTERN (insn);
6091 if (CALL_P (insn))
6093 for (tem = CALL_INSN_FUNCTION_USAGE (insn); tem; tem = XEXP (tem, 1))
6094 if (GET_CODE (XEXP (tem, 0)) == CLOBBER)
6095 invalidate (SET_DEST (XEXP (tem, 0)), VOIDmode);
6098 /* Ensure we invalidate the destination register of a CALL insn.
6099 This is necessary for machines where this register is a fixed_reg,
6100 because no other code would invalidate it. */
6101 if (GET_CODE (x) == SET && GET_CODE (SET_SRC (x)) == CALL)
6102 invalidate (SET_DEST (x), VOIDmode);
6104 else if (GET_CODE (x) == PARALLEL)
6106 int i;
6108 for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
6110 rtx y = XVECEXP (x, 0, i);
6111 if (GET_CODE (y) == CLOBBER)
6113 rtx clobbered = XEXP (y, 0);
6115 if (REG_P (clobbered)
6116 || GET_CODE (clobbered) == SUBREG)
6117 invalidate (clobbered, VOIDmode);
6118 else if (GET_CODE (clobbered) == STRICT_LOW_PART
6119 || GET_CODE (clobbered) == ZERO_EXTRACT)
6120 invalidate (XEXP (clobbered, 0), GET_MODE (clobbered));
6122 else if (GET_CODE (y) == SET && GET_CODE (SET_SRC (y)) == CALL)
6123 invalidate (SET_DEST (y), VOIDmode);
6128 /* Process X, part of the REG_NOTES of an insn. Look at any REG_EQUAL notes
6129 and replace any registers in them with either an equivalent constant
6130 or the canonical form of the register. If we are inside an address,
6131 only do this if the address remains valid.
6133 OBJECT is 0 except when within a MEM in which case it is the MEM.
6135 Return the replacement for X. */
6137 static rtx
6138 cse_process_notes_1 (rtx x, rtx object, bool *changed)
6140 enum rtx_code code = GET_CODE (x);
6141 const char *fmt = GET_RTX_FORMAT (code);
6142 int i;
6144 switch (code)
6146 case CONST:
6147 case SYMBOL_REF:
6148 case LABEL_REF:
6149 CASE_CONST_ANY:
6150 case PC:
6151 case CC0:
6152 case LO_SUM:
6153 return x;
6155 case MEM:
6156 validate_change (x, &XEXP (x, 0),
6157 cse_process_notes (XEXP (x, 0), x, changed), 0);
6158 return x;
6160 case EXPR_LIST:
6161 if (REG_NOTE_KIND (x) == REG_EQUAL)
6162 XEXP (x, 0) = cse_process_notes (XEXP (x, 0), NULL_RTX, changed);
6163 /* Fall through. */
6165 case INSN_LIST:
6166 case INT_LIST:
6167 if (XEXP (x, 1))
6168 XEXP (x, 1) = cse_process_notes (XEXP (x, 1), NULL_RTX, changed);
6169 return x;
6171 case SIGN_EXTEND:
6172 case ZERO_EXTEND:
6173 case SUBREG:
6175 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6176 /* We don't substitute VOIDmode constants into these rtx,
6177 since they would impede folding. */
6178 if (GET_MODE (new_rtx) != VOIDmode)
6179 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6180 return x;
6183 case UNSIGNED_FLOAT:
6185 rtx new_rtx = cse_process_notes (XEXP (x, 0), object, changed);
6186 /* We don't substitute negative VOIDmode constants into these rtx,
6187 since they would impede folding. */
6188 if (GET_MODE (new_rtx) != VOIDmode
6189 || (CONST_INT_P (new_rtx) && INTVAL (new_rtx) >= 0)
6190 || (CONST_DOUBLE_P (new_rtx) && CONST_DOUBLE_HIGH (new_rtx) >= 0))
6191 validate_change (object, &XEXP (x, 0), new_rtx, 0);
6192 return x;
6195 case REG:
6196 i = REG_QTY (REGNO (x));
6198 /* Return a constant or a constant register. */
6199 if (REGNO_QTY_VALID_P (REGNO (x)))
6201 struct qty_table_elem *ent = &qty_table[i];
6203 if (ent->const_rtx != NULL_RTX
6204 && (CONSTANT_P (ent->const_rtx)
6205 || REG_P (ent->const_rtx)))
6207 rtx new_rtx = gen_lowpart (GET_MODE (x), ent->const_rtx);
6208 if (new_rtx)
6209 return copy_rtx (new_rtx);
6213 /* Otherwise, canonicalize this register. */
6214 return canon_reg (x, NULL);
6216 default:
6217 break;
6220 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6221 if (fmt[i] == 'e')
6222 validate_change (object, &XEXP (x, i),
6223 cse_process_notes (XEXP (x, i), object, changed), 0);
6225 return x;
6228 static rtx
6229 cse_process_notes (rtx x, rtx object, bool *changed)
6231 rtx new_rtx = cse_process_notes_1 (x, object, changed);
6232 if (new_rtx != x)
6233 *changed = true;
6234 return new_rtx;
6238 /* Find a path in the CFG, starting with FIRST_BB to perform CSE on.
6240 DATA is a pointer to a struct cse_basic_block_data, that is used to
6241 describe the path.
6242 It is filled with a queue of basic blocks, starting with FIRST_BB
6243 and following a trace through the CFG.
6245 If all paths starting at FIRST_BB have been followed, or no new path
6246 starting at FIRST_BB can be constructed, this function returns FALSE.
6247 Otherwise, DATA->path is filled and the function returns TRUE indicating
6248 that a path to follow was found.
6250 If FOLLOW_JUMPS is false, the maximum path length is 1 and the only
6251 block in the path will be FIRST_BB. */
6253 static bool
6254 cse_find_path (basic_block first_bb, struct cse_basic_block_data *data,
6255 int follow_jumps)
6257 basic_block bb;
6258 edge e;
6259 int path_size;
6261 bitmap_set_bit (cse_visited_basic_blocks, first_bb->index);
6263 /* See if there is a previous path. */
6264 path_size = data->path_size;
6266 /* There is a previous path. Make sure it started with FIRST_BB. */
6267 if (path_size)
6268 gcc_assert (data->path[0].bb == first_bb);
6270 /* There was only one basic block in the last path. Clear the path and
6271 return, so that paths starting at another basic block can be tried. */
6272 if (path_size == 1)
6274 path_size = 0;
6275 goto done;
6278 /* If the path was empty from the beginning, construct a new path. */
6279 if (path_size == 0)
6280 data->path[path_size++].bb = first_bb;
6281 else
6283 /* Otherwise, path_size must be equal to or greater than 2, because
6284 a previous path exists that is at least two basic blocks long.
6286 Update the previous branch path, if any. If the last branch was
6287 previously along the branch edge, take the fallthrough edge now. */
6288 while (path_size >= 2)
6290 basic_block last_bb_in_path, previous_bb_in_path;
6291 edge e;
6293 --path_size;
6294 last_bb_in_path = data->path[path_size].bb;
6295 previous_bb_in_path = data->path[path_size - 1].bb;
6297 /* If we previously followed a path along the branch edge, try
6298 the fallthru edge now. */
6299 if (EDGE_COUNT (previous_bb_in_path->succs) == 2
6300 && any_condjump_p (BB_END (previous_bb_in_path))
6301 && (e = find_edge (previous_bb_in_path, last_bb_in_path))
6302 && e == BRANCH_EDGE (previous_bb_in_path))
6304 bb = FALLTHRU_EDGE (previous_bb_in_path)->dest;
6305 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
6306 && single_pred_p (bb)
6307 /* We used to assert here that we would only see blocks
6308 that we have not visited yet. But we may end up
6309 visiting basic blocks twice if the CFG has changed
6310 in this run of cse_main, because when the CFG changes
6311 the topological sort of the CFG also changes. A basic
6312 blocks that previously had more than two predecessors
6313 may now have a single predecessor, and become part of
6314 a path that starts at another basic block.
6316 We still want to visit each basic block only once, so
6317 halt the path here if we have already visited BB. */
6318 && !bitmap_bit_p (cse_visited_basic_blocks, bb->index))
6320 bitmap_set_bit (cse_visited_basic_blocks, bb->index);
6321 data->path[path_size++].bb = bb;
6322 break;
6326 data->path[path_size].bb = NULL;
6329 /* If only one block remains in the path, bail. */
6330 if (path_size == 1)
6332 path_size = 0;
6333 goto done;
6337 /* Extend the path if possible. */
6338 if (follow_jumps)
6340 bb = data->path[path_size - 1].bb;
6341 while (bb && path_size < PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH))
6343 if (single_succ_p (bb))
6344 e = single_succ_edge (bb);
6345 else if (EDGE_COUNT (bb->succs) == 2
6346 && any_condjump_p (BB_END (bb)))
6348 /* First try to follow the branch. If that doesn't lead
6349 to a useful path, follow the fallthru edge. */
6350 e = BRANCH_EDGE (bb);
6351 if (!single_pred_p (e->dest))
6352 e = FALLTHRU_EDGE (bb);
6354 else
6355 e = NULL;
6357 if (e
6358 && !((e->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label)
6359 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
6360 && single_pred_p (e->dest)
6361 /* Avoid visiting basic blocks twice. The large comment
6362 above explains why this can happen. */
6363 && !bitmap_bit_p (cse_visited_basic_blocks, e->dest->index))
6365 basic_block bb2 = e->dest;
6366 bitmap_set_bit (cse_visited_basic_blocks, bb2->index);
6367 data->path[path_size++].bb = bb2;
6368 bb = bb2;
6370 else
6371 bb = NULL;
6375 done:
6376 data->path_size = path_size;
6377 return path_size != 0;
6380 /* Dump the path in DATA to file F. NSETS is the number of sets
6381 in the path. */
6383 static void
6384 cse_dump_path (struct cse_basic_block_data *data, int nsets, FILE *f)
6386 int path_entry;
6388 fprintf (f, ";; Following path with %d sets: ", nsets);
6389 for (path_entry = 0; path_entry < data->path_size; path_entry++)
6390 fprintf (f, "%d ", (data->path[path_entry].bb)->index);
6391 fputc ('\n', dump_file);
6392 fflush (f);
6396 /* Return true if BB has exception handling successor edges. */
6398 static bool
6399 have_eh_succ_edges (basic_block bb)
6401 edge e;
6402 edge_iterator ei;
6404 FOR_EACH_EDGE (e, ei, bb->succs)
6405 if (e->flags & EDGE_EH)
6406 return true;
6408 return false;
6412 /* Scan to the end of the path described by DATA. Return an estimate of
6413 the total number of SETs of all insns in the path. */
6415 static void
6416 cse_prescan_path (struct cse_basic_block_data *data)
6418 int nsets = 0;
6419 int path_size = data->path_size;
6420 int path_entry;
6422 /* Scan to end of each basic block in the path. */
6423 for (path_entry = 0; path_entry < path_size; path_entry++)
6425 basic_block bb;
6426 rtx_insn *insn;
6428 bb = data->path[path_entry].bb;
6430 FOR_BB_INSNS (bb, insn)
6432 if (!INSN_P (insn))
6433 continue;
6435 /* A PARALLEL can have lots of SETs in it,
6436 especially if it is really an ASM_OPERANDS. */
6437 if (GET_CODE (PATTERN (insn)) == PARALLEL)
6438 nsets += XVECLEN (PATTERN (insn), 0);
6439 else
6440 nsets += 1;
6444 data->nsets = nsets;
6447 /* Return true if the pattern of INSN uses a LABEL_REF for which
6448 there isn't a REG_LABEL_OPERAND note. */
6450 static bool
6451 check_for_label_ref (rtx_insn *insn)
6453 /* If this insn uses a LABEL_REF and there isn't a REG_LABEL_OPERAND
6454 note for it, we must rerun jump since it needs to place the note. If
6455 this is a LABEL_REF for a CODE_LABEL that isn't in the insn chain,
6456 don't do this since no REG_LABEL_OPERAND will be added. */
6457 subrtx_iterator::array_type array;
6458 FOR_EACH_SUBRTX (iter, array, PATTERN (insn), ALL)
6460 const_rtx x = *iter;
6461 if (GET_CODE (x) == LABEL_REF
6462 && !LABEL_REF_NONLOCAL_P (x)
6463 && (!JUMP_P (insn)
6464 || !label_is_jump_target_p (LABEL_REF_LABEL (x), insn))
6465 && LABEL_P (LABEL_REF_LABEL (x))
6466 && INSN_UID (LABEL_REF_LABEL (x)) != 0
6467 && !find_reg_note (insn, REG_LABEL_OPERAND, LABEL_REF_LABEL (x)))
6468 return true;
6470 return false;
6473 /* Process a single extended basic block described by EBB_DATA. */
6475 static void
6476 cse_extended_basic_block (struct cse_basic_block_data *ebb_data)
6478 int path_size = ebb_data->path_size;
6479 int path_entry;
6480 int num_insns = 0;
6482 /* Allocate the space needed by qty_table. */
6483 qty_table = XNEWVEC (struct qty_table_elem, max_qty);
6485 new_basic_block ();
6486 cse_ebb_live_in = df_get_live_in (ebb_data->path[0].bb);
6487 cse_ebb_live_out = df_get_live_out (ebb_data->path[path_size - 1].bb);
6488 for (path_entry = 0; path_entry < path_size; path_entry++)
6490 basic_block bb;
6491 rtx_insn *insn;
6493 bb = ebb_data->path[path_entry].bb;
6495 /* Invalidate recorded information for eh regs if there is an EH
6496 edge pointing to that bb. */
6497 if (bb_has_eh_pred (bb))
6499 df_ref def;
6501 FOR_EACH_ARTIFICIAL_DEF (def, bb->index)
6502 if (DF_REF_FLAGS (def) & DF_REF_AT_TOP)
6503 invalidate (DF_REF_REG (def), GET_MODE (DF_REF_REG (def)));
6506 optimize_this_for_speed_p = optimize_bb_for_speed_p (bb);
6507 FOR_BB_INSNS (bb, insn)
6509 /* If we have processed 1,000 insns, flush the hash table to
6510 avoid extreme quadratic behavior. We must not include NOTEs
6511 in the count since there may be more of them when generating
6512 debugging information. If we clear the table at different
6513 times, code generated with -g -O might be different than code
6514 generated with -O but not -g.
6516 FIXME: This is a real kludge and needs to be done some other
6517 way. */
6518 if (NONDEBUG_INSN_P (insn)
6519 && num_insns++ > PARAM_VALUE (PARAM_MAX_CSE_INSNS))
6521 flush_hash_table ();
6522 num_insns = 0;
6525 if (INSN_P (insn))
6527 /* Process notes first so we have all notes in canonical forms
6528 when looking for duplicate operations. */
6529 if (REG_NOTES (insn))
6531 bool changed = false;
6532 REG_NOTES (insn) = cse_process_notes (REG_NOTES (insn),
6533 NULL_RTX, &changed);
6534 if (changed)
6535 df_notes_rescan (insn);
6538 cse_insn (insn);
6540 /* If we haven't already found an insn where we added a LABEL_REF,
6541 check this one. */
6542 if (INSN_P (insn) && !recorded_label_ref
6543 && check_for_label_ref (insn))
6544 recorded_label_ref = true;
6546 if (HAVE_cc0 && NONDEBUG_INSN_P (insn))
6548 /* If the previous insn sets CC0 and this insn no
6549 longer references CC0, delete the previous insn.
6550 Here we use fact that nothing expects CC0 to be
6551 valid over an insn, which is true until the final
6552 pass. */
6553 rtx_insn *prev_insn;
6554 rtx tem;
6556 prev_insn = prev_nonnote_nondebug_insn (insn);
6557 if (prev_insn && NONJUMP_INSN_P (prev_insn)
6558 && (tem = single_set (prev_insn)) != NULL_RTX
6559 && SET_DEST (tem) == cc0_rtx
6560 && ! reg_mentioned_p (cc0_rtx, PATTERN (insn)))
6561 delete_insn (prev_insn);
6563 /* If this insn is not the last insn in the basic
6564 block, it will be PREV_INSN(insn) in the next
6565 iteration. If we recorded any CC0-related
6566 information for this insn, remember it. */
6567 if (insn != BB_END (bb))
6569 prev_insn_cc0 = this_insn_cc0;
6570 prev_insn_cc0_mode = this_insn_cc0_mode;
6576 /* With non-call exceptions, we are not always able to update
6577 the CFG properly inside cse_insn. So clean up possibly
6578 redundant EH edges here. */
6579 if (cfun->can_throw_non_call_exceptions && have_eh_succ_edges (bb))
6580 cse_cfg_altered |= purge_dead_edges (bb);
6582 /* If we changed a conditional jump, we may have terminated
6583 the path we are following. Check that by verifying that
6584 the edge we would take still exists. If the edge does
6585 not exist anymore, purge the remainder of the path.
6586 Note that this will cause us to return to the caller. */
6587 if (path_entry < path_size - 1)
6589 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6590 if (!find_edge (bb, next_bb))
6594 path_size--;
6596 /* If we truncate the path, we must also reset the
6597 visited bit on the remaining blocks in the path,
6598 or we will never visit them at all. */
6599 bitmap_clear_bit (cse_visited_basic_blocks,
6600 ebb_data->path[path_size].bb->index);
6601 ebb_data->path[path_size].bb = NULL;
6603 while (path_size - 1 != path_entry);
6604 ebb_data->path_size = path_size;
6608 /* If this is a conditional jump insn, record any known
6609 equivalences due to the condition being tested. */
6610 insn = BB_END (bb);
6611 if (path_entry < path_size - 1
6612 && JUMP_P (insn)
6613 && single_set (insn)
6614 && any_condjump_p (insn))
6616 basic_block next_bb = ebb_data->path[path_entry + 1].bb;
6617 bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
6618 record_jump_equiv (insn, taken);
6621 /* Clear the CC0-tracking related insns, they can't provide
6622 useful information across basic block boundaries. */
6623 prev_insn_cc0 = 0;
6626 gcc_assert (next_qty <= max_qty);
6628 free (qty_table);
6632 /* Perform cse on the instructions of a function.
6633 F is the first instruction.
6634 NREGS is one plus the highest pseudo-reg number used in the instruction.
6636 Return 2 if jump optimizations should be redone due to simplifications
6637 in conditional jump instructions.
6638 Return 1 if the CFG should be cleaned up because it has been modified.
6639 Return 0 otherwise. */
6641 static int
6642 cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
6644 struct cse_basic_block_data ebb_data;
6645 basic_block bb;
6646 int *rc_order = XNEWVEC (int, last_basic_block_for_fn (cfun));
6647 int i, n_blocks;
6649 df_set_flags (DF_LR_RUN_DCE);
6650 df_note_add_problem ();
6651 df_analyze ();
6652 df_set_flags (DF_DEFER_INSN_RESCAN);
6654 reg_scan (get_insns (), max_reg_num ());
6655 init_cse_reg_info (nregs);
6657 ebb_data.path = XNEWVEC (struct branch_path,
6658 PARAM_VALUE (PARAM_MAX_CSE_PATH_LENGTH));
6660 cse_cfg_altered = false;
6661 cse_jumps_altered = false;
6662 recorded_label_ref = false;
6663 constant_pool_entries_cost = 0;
6664 constant_pool_entries_regcost = 0;
6665 ebb_data.path_size = 0;
6666 ebb_data.nsets = 0;
6667 rtl_hooks = cse_rtl_hooks;
6669 init_recog ();
6670 init_alias_analysis ();
6672 reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
6674 /* Set up the table of already visited basic blocks. */
6675 cse_visited_basic_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun));
6676 bitmap_clear (cse_visited_basic_blocks);
6678 /* Loop over basic blocks in reverse completion order (RPO),
6679 excluding the ENTRY and EXIT blocks. */
6680 n_blocks = pre_and_rev_post_order_compute (NULL, rc_order, false);
6681 i = 0;
6682 while (i < n_blocks)
6684 /* Find the first block in the RPO queue that we have not yet
6685 processed before. */
6688 bb = BASIC_BLOCK_FOR_FN (cfun, rc_order[i++]);
6690 while (bitmap_bit_p (cse_visited_basic_blocks, bb->index)
6691 && i < n_blocks);
6693 /* Find all paths starting with BB, and process them. */
6694 while (cse_find_path (bb, &ebb_data, flag_cse_follow_jumps))
6696 /* Pre-scan the path. */
6697 cse_prescan_path (&ebb_data);
6699 /* If this basic block has no sets, skip it. */
6700 if (ebb_data.nsets == 0)
6701 continue;
6703 /* Get a reasonable estimate for the maximum number of qty's
6704 needed for this path. For this, we take the number of sets
6705 and multiply that by MAX_RECOG_OPERANDS. */
6706 max_qty = ebb_data.nsets * MAX_RECOG_OPERANDS;
6708 /* Dump the path we're about to process. */
6709 if (dump_file)
6710 cse_dump_path (&ebb_data, ebb_data.nsets, dump_file);
6712 cse_extended_basic_block (&ebb_data);
6716 /* Clean up. */
6717 end_alias_analysis ();
6718 free (reg_eqv_table);
6719 free (ebb_data.path);
6720 sbitmap_free (cse_visited_basic_blocks);
6721 free (rc_order);
6722 rtl_hooks = general_rtl_hooks;
6724 if (cse_jumps_altered || recorded_label_ref)
6725 return 2;
6726 else if (cse_cfg_altered)
6727 return 1;
6728 else
6729 return 0;
6732 /* Count the number of times registers are used (not set) in X.
6733 COUNTS is an array in which we accumulate the count, INCR is how much
6734 we count each register usage.
6736 Don't count a usage of DEST, which is the SET_DEST of a SET which
6737 contains X in its SET_SRC. This is because such a SET does not
6738 modify the liveness of DEST.
6739 DEST is set to pc_rtx for a trapping insn, or for an insn with side effects.
6740 We must then count uses of a SET_DEST regardless, because the insn can't be
6741 deleted here. */
6743 static void
6744 count_reg_usage (rtx x, int *counts, rtx dest, int incr)
6746 enum rtx_code code;
6747 rtx note;
6748 const char *fmt;
6749 int i, j;
6751 if (x == 0)
6752 return;
6754 switch (code = GET_CODE (x))
6756 case REG:
6757 if (x != dest)
6758 counts[REGNO (x)] += incr;
6759 return;
6761 case PC:
6762 case CC0:
6763 case CONST:
6764 CASE_CONST_ANY:
6765 case SYMBOL_REF:
6766 case LABEL_REF:
6767 return;
6769 case CLOBBER:
6770 /* If we are clobbering a MEM, mark any registers inside the address
6771 as being used. */
6772 if (MEM_P (XEXP (x, 0)))
6773 count_reg_usage (XEXP (XEXP (x, 0), 0), counts, NULL_RTX, incr);
6774 return;
6776 case SET:
6777 /* Unless we are setting a REG, count everything in SET_DEST. */
6778 if (!REG_P (SET_DEST (x)))
6779 count_reg_usage (SET_DEST (x), counts, NULL_RTX, incr);
6780 count_reg_usage (SET_SRC (x), counts,
6781 dest ? dest : SET_DEST (x),
6782 incr);
6783 return;
6785 case DEBUG_INSN:
6786 return;
6788 case CALL_INSN:
6789 case INSN:
6790 case JUMP_INSN:
6791 /* We expect dest to be NULL_RTX here. If the insn may throw,
6792 or if it cannot be deleted due to side-effects, mark this fact
6793 by setting DEST to pc_rtx. */
6794 if ((!cfun->can_delete_dead_exceptions && !insn_nothrow_p (x))
6795 || side_effects_p (PATTERN (x)))
6796 dest = pc_rtx;
6797 if (code == CALL_INSN)
6798 count_reg_usage (CALL_INSN_FUNCTION_USAGE (x), counts, dest, incr);
6799 count_reg_usage (PATTERN (x), counts, dest, incr);
6801 /* Things used in a REG_EQUAL note aren't dead since loop may try to
6802 use them. */
6804 note = find_reg_equal_equiv_note (x);
6805 if (note)
6807 rtx eqv = XEXP (note, 0);
6809 if (GET_CODE (eqv) == EXPR_LIST)
6810 /* This REG_EQUAL note describes the result of a function call.
6811 Process all the arguments. */
6814 count_reg_usage (XEXP (eqv, 0), counts, dest, incr);
6815 eqv = XEXP (eqv, 1);
6817 while (eqv && GET_CODE (eqv) == EXPR_LIST);
6818 else
6819 count_reg_usage (eqv, counts, dest, incr);
6821 return;
6823 case EXPR_LIST:
6824 if (REG_NOTE_KIND (x) == REG_EQUAL
6825 || (REG_NOTE_KIND (x) != REG_NONNEG && GET_CODE (XEXP (x,0)) == USE)
6826 /* FUNCTION_USAGE expression lists may include (CLOBBER (mem /u)),
6827 involving registers in the address. */
6828 || GET_CODE (XEXP (x, 0)) == CLOBBER)
6829 count_reg_usage (XEXP (x, 0), counts, NULL_RTX, incr);
6831 count_reg_usage (XEXP (x, 1), counts, NULL_RTX, incr);
6832 return;
6834 case ASM_OPERANDS:
6835 /* Iterate over just the inputs, not the constraints as well. */
6836 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
6837 count_reg_usage (ASM_OPERANDS_INPUT (x, i), counts, dest, incr);
6838 return;
6840 case INSN_LIST:
6841 case INT_LIST:
6842 gcc_unreachable ();
6844 default:
6845 break;
6848 fmt = GET_RTX_FORMAT (code);
6849 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6851 if (fmt[i] == 'e')
6852 count_reg_usage (XEXP (x, i), counts, dest, incr);
6853 else if (fmt[i] == 'E')
6854 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6855 count_reg_usage (XVECEXP (x, i, j), counts, dest, incr);
6859 /* Return true if X is a dead register. */
6861 static inline int
6862 is_dead_reg (const_rtx x, int *counts)
6864 return (REG_P (x)
6865 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6866 && counts[REGNO (x)] == 0);
6869 /* Return true if set is live. */
6870 static bool
6871 set_live_p (rtx set, rtx_insn *insn ATTRIBUTE_UNUSED, /* Only used with HAVE_cc0. */
6872 int *counts)
6874 rtx_insn *tem;
6876 if (set_noop_p (set))
6879 else if (GET_CODE (SET_DEST (set)) == CC0
6880 && !side_effects_p (SET_SRC (set))
6881 && ((tem = next_nonnote_nondebug_insn (insn)) == NULL_RTX
6882 || !INSN_P (tem)
6883 || !reg_referenced_p (cc0_rtx, PATTERN (tem))))
6884 return false;
6885 else if (!is_dead_reg (SET_DEST (set), counts)
6886 || side_effects_p (SET_SRC (set)))
6887 return true;
6888 return false;
6891 /* Return true if insn is live. */
6893 static bool
6894 insn_live_p (rtx_insn *insn, int *counts)
6896 int i;
6897 if (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn))
6898 return true;
6899 else if (GET_CODE (PATTERN (insn)) == SET)
6900 return set_live_p (PATTERN (insn), insn, counts);
6901 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
6903 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
6905 rtx elt = XVECEXP (PATTERN (insn), 0, i);
6907 if (GET_CODE (elt) == SET)
6909 if (set_live_p (elt, insn, counts))
6910 return true;
6912 else if (GET_CODE (elt) != CLOBBER && GET_CODE (elt) != USE)
6913 return true;
6915 return false;
6917 else if (DEBUG_INSN_P (insn))
6919 rtx_insn *next;
6921 for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
6922 if (NOTE_P (next))
6923 continue;
6924 else if (!DEBUG_INSN_P (next))
6925 return true;
6926 else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL (next))
6927 return false;
6929 return true;
6931 else
6932 return true;
6935 /* Count the number of stores into pseudo. Callback for note_stores. */
6937 static void
6938 count_stores (rtx x, const_rtx set ATTRIBUTE_UNUSED, void *data)
6940 int *counts = (int *) data;
6941 if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER)
6942 counts[REGNO (x)]++;
6945 /* Return if DEBUG_INSN pattern PAT needs to be reset because some dead
6946 pseudo doesn't have a replacement. COUNTS[X] is zero if register X
6947 is dead and REPLACEMENTS[X] is null if it has no replacemenet.
6948 Set *SEEN_REPL to true if we see a dead register that does have
6949 a replacement. */
6951 static bool
6952 is_dead_debug_insn (const_rtx pat, int *counts, rtx *replacements,
6953 bool *seen_repl)
6955 subrtx_iterator::array_type array;
6956 FOR_EACH_SUBRTX (iter, array, pat, NONCONST)
6958 const_rtx x = *iter;
6959 if (is_dead_reg (x, counts))
6961 if (replacements && replacements[REGNO (x)] != NULL_RTX)
6962 *seen_repl = true;
6963 else
6964 return true;
6967 return false;
6970 /* Replace a dead pseudo in a DEBUG_INSN with replacement DEBUG_EXPR.
6971 Callback for simplify_replace_fn_rtx. */
6973 static rtx
6974 replace_dead_reg (rtx x, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
6976 rtx *replacements = (rtx *) data;
6978 if (REG_P (x)
6979 && REGNO (x) >= FIRST_PSEUDO_REGISTER
6980 && replacements[REGNO (x)] != NULL_RTX)
6982 if (GET_MODE (x) == GET_MODE (replacements[REGNO (x)]))
6983 return replacements[REGNO (x)];
6984 return lowpart_subreg (GET_MODE (x), replacements[REGNO (x)],
6985 GET_MODE (replacements[REGNO (x)]));
6987 return NULL_RTX;
6990 /* Scan all the insns and delete any that are dead; i.e., they store a register
6991 that is never used or they copy a register to itself.
6993 This is used to remove insns made obviously dead by cse, loop or other
6994 optimizations. It improves the heuristics in loop since it won't try to
6995 move dead invariants out of loops or make givs for dead quantities. The
6996 remaining passes of the compilation are also sped up. */
6999 delete_trivially_dead_insns (rtx_insn *insns, int nreg)
7001 int *counts;
7002 rtx_insn *insn, *prev;
7003 rtx *replacements = NULL;
7004 int ndead = 0;
7006 timevar_push (TV_DELETE_TRIVIALLY_DEAD);
7007 /* First count the number of times each register is used. */
7008 if (MAY_HAVE_DEBUG_INSNS)
7010 counts = XCNEWVEC (int, nreg * 3);
7011 for (insn = insns; insn; insn = NEXT_INSN (insn))
7012 if (DEBUG_INSN_P (insn))
7013 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7014 NULL_RTX, 1);
7015 else if (INSN_P (insn))
7017 count_reg_usage (insn, counts, NULL_RTX, 1);
7018 note_stores (PATTERN (insn), count_stores, counts + nreg * 2);
7020 /* If there can be debug insns, COUNTS are 3 consecutive arrays.
7021 First one counts how many times each pseudo is used outside
7022 of debug insns, second counts how many times each pseudo is
7023 used in debug insns and third counts how many times a pseudo
7024 is stored. */
7026 else
7028 counts = XCNEWVEC (int, nreg);
7029 for (insn = insns; insn; insn = NEXT_INSN (insn))
7030 if (INSN_P (insn))
7031 count_reg_usage (insn, counts, NULL_RTX, 1);
7032 /* If no debug insns can be present, COUNTS is just an array
7033 which counts how many times each pseudo is used. */
7035 /* Pseudo PIC register should be considered as used due to possible
7036 new usages generated. */
7037 if (!reload_completed
7038 && pic_offset_table_rtx
7039 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
7040 counts[REGNO (pic_offset_table_rtx)]++;
7041 /* Go from the last insn to the first and delete insns that only set unused
7042 registers or copy a register to itself. As we delete an insn, remove
7043 usage counts for registers it uses.
7045 The first jump optimization pass may leave a real insn as the last
7046 insn in the function. We must not skip that insn or we may end
7047 up deleting code that is not really dead.
7049 If some otherwise unused register is only used in DEBUG_INSNs,
7050 try to create a DEBUG_EXPR temporary and emit a DEBUG_INSN before
7051 the setter. Then go through DEBUG_INSNs and if a DEBUG_EXPR
7052 has been created for the unused register, replace it with
7053 the DEBUG_EXPR, otherwise reset the DEBUG_INSN. */
7054 for (insn = get_last_insn (); insn; insn = prev)
7056 int live_insn = 0;
7058 prev = PREV_INSN (insn);
7059 if (!INSN_P (insn))
7060 continue;
7062 live_insn = insn_live_p (insn, counts);
7064 /* If this is a dead insn, delete it and show registers in it aren't
7065 being used. */
7067 if (! live_insn && dbg_cnt (delete_trivial_dead))
7069 if (DEBUG_INSN_P (insn))
7070 count_reg_usage (INSN_VAR_LOCATION_LOC (insn), counts + nreg,
7071 NULL_RTX, -1);
7072 else
7074 rtx set;
7075 if (MAY_HAVE_DEBUG_INSNS
7076 && (set = single_set (insn)) != NULL_RTX
7077 && is_dead_reg (SET_DEST (set), counts)
7078 /* Used at least once in some DEBUG_INSN. */
7079 && counts[REGNO (SET_DEST (set)) + nreg] > 0
7080 /* And set exactly once. */
7081 && counts[REGNO (SET_DEST (set)) + nreg * 2] == 1
7082 && !side_effects_p (SET_SRC (set))
7083 && asm_noperands (PATTERN (insn)) < 0)
7085 rtx dval, bind_var_loc;
7086 rtx_insn *bind;
7088 /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */
7089 dval = make_debug_expr_from_rtl (SET_DEST (set));
7091 /* Emit a debug bind insn before the insn in which
7092 reg dies. */
7093 bind_var_loc =
7094 gen_rtx_VAR_LOCATION (GET_MODE (SET_DEST (set)),
7095 DEBUG_EXPR_TREE_DECL (dval),
7096 SET_SRC (set),
7097 VAR_INIT_STATUS_INITIALIZED);
7098 count_reg_usage (bind_var_loc, counts + nreg, NULL_RTX, 1);
7100 bind = emit_debug_insn_before (bind_var_loc, insn);
7101 df_insn_rescan (bind);
7103 if (replacements == NULL)
7104 replacements = XCNEWVEC (rtx, nreg);
7105 replacements[REGNO (SET_DEST (set))] = dval;
7108 count_reg_usage (insn, counts, NULL_RTX, -1);
7109 ndead++;
7111 delete_insn_and_edges (insn);
7115 if (MAY_HAVE_DEBUG_INSNS)
7117 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn))
7118 if (DEBUG_INSN_P (insn))
7120 /* If this debug insn references a dead register that wasn't replaced
7121 with an DEBUG_EXPR, reset the DEBUG_INSN. */
7122 bool seen_repl = false;
7123 if (is_dead_debug_insn (INSN_VAR_LOCATION_LOC (insn),
7124 counts, replacements, &seen_repl))
7126 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
7127 df_insn_rescan (insn);
7129 else if (seen_repl)
7131 INSN_VAR_LOCATION_LOC (insn)
7132 = simplify_replace_fn_rtx (INSN_VAR_LOCATION_LOC (insn),
7133 NULL_RTX, replace_dead_reg,
7134 replacements);
7135 df_insn_rescan (insn);
7138 free (replacements);
7141 if (dump_file && ndead)
7142 fprintf (dump_file, "Deleted %i trivially dead insns\n",
7143 ndead);
7144 /* Clean up. */
7145 free (counts);
7146 timevar_pop (TV_DELETE_TRIVIALLY_DEAD);
7147 return ndead;
7150 /* If LOC contains references to NEWREG in a different mode, change them
7151 to use NEWREG instead. */
7153 static void
7154 cse_change_cc_mode (subrtx_ptr_iterator::array_type &array,
7155 rtx *loc, rtx_insn *insn, rtx newreg)
7157 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
7159 rtx *loc = *iter;
7160 rtx x = *loc;
7161 if (x
7162 && REG_P (x)
7163 && REGNO (x) == REGNO (newreg)
7164 && GET_MODE (x) != GET_MODE (newreg))
7166 validate_change (insn, loc, newreg, 1);
7167 iter.skip_subrtxes ();
7172 /* Change the mode of any reference to the register REGNO (NEWREG) to
7173 GET_MODE (NEWREG) in INSN. */
7175 static void
7176 cse_change_cc_mode_insn (rtx_insn *insn, rtx newreg)
7178 int success;
7180 if (!INSN_P (insn))
7181 return;
7183 subrtx_ptr_iterator::array_type array;
7184 cse_change_cc_mode (array, &PATTERN (insn), insn, newreg);
7185 cse_change_cc_mode (array, &REG_NOTES (insn), insn, newreg);
7187 /* If the following assertion was triggered, there is most probably
7188 something wrong with the cc_modes_compatible back end function.
7189 CC modes only can be considered compatible if the insn - with the mode
7190 replaced by any of the compatible modes - can still be recognized. */
7191 success = apply_change_group ();
7192 gcc_assert (success);
7195 /* Change the mode of any reference to the register REGNO (NEWREG) to
7196 GET_MODE (NEWREG), starting at START. Stop before END. Stop at
7197 any instruction which modifies NEWREG. */
7199 static void
7200 cse_change_cc_mode_insns (rtx_insn *start, rtx_insn *end, rtx newreg)
7202 rtx_insn *insn;
7204 for (insn = start; insn != end; insn = NEXT_INSN (insn))
7206 if (! INSN_P (insn))
7207 continue;
7209 if (reg_set_p (newreg, insn))
7210 return;
7212 cse_change_cc_mode_insn (insn, newreg);
7216 /* BB is a basic block which finishes with CC_REG as a condition code
7217 register which is set to CC_SRC. Look through the successors of BB
7218 to find blocks which have a single predecessor (i.e., this one),
7219 and look through those blocks for an assignment to CC_REG which is
7220 equivalent to CC_SRC. CAN_CHANGE_MODE indicates whether we are
7221 permitted to change the mode of CC_SRC to a compatible mode. This
7222 returns VOIDmode if no equivalent assignments were found.
7223 Otherwise it returns the mode which CC_SRC should wind up with.
7224 ORIG_BB should be the same as BB in the outermost cse_cc_succs call,
7225 but is passed unmodified down to recursive calls in order to prevent
7226 endless recursion.
7228 The main complexity in this function is handling the mode issues.
7229 We may have more than one duplicate which we can eliminate, and we
7230 try to find a mode which will work for multiple duplicates. */
7232 static machine_mode
7233 cse_cc_succs (basic_block bb, basic_block orig_bb, rtx cc_reg, rtx cc_src,
7234 bool can_change_mode)
7236 bool found_equiv;
7237 machine_mode mode;
7238 unsigned int insn_count;
7239 edge e;
7240 rtx_insn *insns[2];
7241 machine_mode modes[2];
7242 rtx_insn *last_insns[2];
7243 unsigned int i;
7244 rtx newreg;
7245 edge_iterator ei;
7247 /* We expect to have two successors. Look at both before picking
7248 the final mode for the comparison. If we have more successors
7249 (i.e., some sort of table jump, although that seems unlikely),
7250 then we require all beyond the first two to use the same
7251 mode. */
7253 found_equiv = false;
7254 mode = GET_MODE (cc_src);
7255 insn_count = 0;
7256 FOR_EACH_EDGE (e, ei, bb->succs)
7258 rtx_insn *insn;
7259 rtx_insn *end;
7261 if (e->flags & EDGE_COMPLEX)
7262 continue;
7264 if (EDGE_COUNT (e->dest->preds) != 1
7265 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
7266 /* Avoid endless recursion on unreachable blocks. */
7267 || e->dest == orig_bb)
7268 continue;
7270 end = NEXT_INSN (BB_END (e->dest));
7271 for (insn = BB_HEAD (e->dest); insn != end; insn = NEXT_INSN (insn))
7273 rtx set;
7275 if (! INSN_P (insn))
7276 continue;
7278 /* If CC_SRC is modified, we have to stop looking for
7279 something which uses it. */
7280 if (modified_in_p (cc_src, insn))
7281 break;
7283 /* Check whether INSN sets CC_REG to CC_SRC. */
7284 set = single_set (insn);
7285 if (set
7286 && REG_P (SET_DEST (set))
7287 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7289 bool found;
7290 machine_mode set_mode;
7291 machine_mode comp_mode;
7293 found = false;
7294 set_mode = GET_MODE (SET_SRC (set));
7295 comp_mode = set_mode;
7296 if (rtx_equal_p (cc_src, SET_SRC (set)))
7297 found = true;
7298 else if (GET_CODE (cc_src) == COMPARE
7299 && GET_CODE (SET_SRC (set)) == COMPARE
7300 && mode != set_mode
7301 && rtx_equal_p (XEXP (cc_src, 0),
7302 XEXP (SET_SRC (set), 0))
7303 && rtx_equal_p (XEXP (cc_src, 1),
7304 XEXP (SET_SRC (set), 1)))
7307 comp_mode = targetm.cc_modes_compatible (mode, set_mode);
7308 if (comp_mode != VOIDmode
7309 && (can_change_mode || comp_mode == mode))
7310 found = true;
7313 if (found)
7315 found_equiv = true;
7316 if (insn_count < ARRAY_SIZE (insns))
7318 insns[insn_count] = insn;
7319 modes[insn_count] = set_mode;
7320 last_insns[insn_count] = end;
7321 ++insn_count;
7323 if (mode != comp_mode)
7325 gcc_assert (can_change_mode);
7326 mode = comp_mode;
7328 /* The modified insn will be re-recognized later. */
7329 PUT_MODE (cc_src, mode);
7332 else
7334 if (set_mode != mode)
7336 /* We found a matching expression in the
7337 wrong mode, but we don't have room to
7338 store it in the array. Punt. This case
7339 should be rare. */
7340 break;
7342 /* INSN sets CC_REG to a value equal to CC_SRC
7343 with the right mode. We can simply delete
7344 it. */
7345 delete_insn (insn);
7348 /* We found an instruction to delete. Keep looking,
7349 in the hopes of finding a three-way jump. */
7350 continue;
7353 /* We found an instruction which sets the condition
7354 code, so don't look any farther. */
7355 break;
7358 /* If INSN sets CC_REG in some other way, don't look any
7359 farther. */
7360 if (reg_set_p (cc_reg, insn))
7361 break;
7364 /* If we fell off the bottom of the block, we can keep looking
7365 through successors. We pass CAN_CHANGE_MODE as false because
7366 we aren't prepared to handle compatibility between the
7367 further blocks and this block. */
7368 if (insn == end)
7370 machine_mode submode;
7372 submode = cse_cc_succs (e->dest, orig_bb, cc_reg, cc_src, false);
7373 if (submode != VOIDmode)
7375 gcc_assert (submode == mode);
7376 found_equiv = true;
7377 can_change_mode = false;
7382 if (! found_equiv)
7383 return VOIDmode;
7385 /* Now INSN_COUNT is the number of instructions we found which set
7386 CC_REG to a value equivalent to CC_SRC. The instructions are in
7387 INSNS. The modes used by those instructions are in MODES. */
7389 newreg = NULL_RTX;
7390 for (i = 0; i < insn_count; ++i)
7392 if (modes[i] != mode)
7394 /* We need to change the mode of CC_REG in INSNS[i] and
7395 subsequent instructions. */
7396 if (! newreg)
7398 if (GET_MODE (cc_reg) == mode)
7399 newreg = cc_reg;
7400 else
7401 newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7403 cse_change_cc_mode_insns (NEXT_INSN (insns[i]), last_insns[i],
7404 newreg);
7407 delete_insn_and_edges (insns[i]);
7410 return mode;
7413 /* If we have a fixed condition code register (or two), walk through
7414 the instructions and try to eliminate duplicate assignments. */
7416 static void
7417 cse_condition_code_reg (void)
7419 unsigned int cc_regno_1;
7420 unsigned int cc_regno_2;
7421 rtx cc_reg_1;
7422 rtx cc_reg_2;
7423 basic_block bb;
7425 if (! targetm.fixed_condition_code_regs (&cc_regno_1, &cc_regno_2))
7426 return;
7428 cc_reg_1 = gen_rtx_REG (CCmode, cc_regno_1);
7429 if (cc_regno_2 != INVALID_REGNUM)
7430 cc_reg_2 = gen_rtx_REG (CCmode, cc_regno_2);
7431 else
7432 cc_reg_2 = NULL_RTX;
7434 FOR_EACH_BB_FN (bb, cfun)
7436 rtx_insn *last_insn;
7437 rtx cc_reg;
7438 rtx_insn *insn;
7439 rtx_insn *cc_src_insn;
7440 rtx cc_src;
7441 machine_mode mode;
7442 machine_mode orig_mode;
7444 /* Look for blocks which end with a conditional jump based on a
7445 condition code register. Then look for the instruction which
7446 sets the condition code register. Then look through the
7447 successor blocks for instructions which set the condition
7448 code register to the same value. There are other possible
7449 uses of the condition code register, but these are by far the
7450 most common and the ones which we are most likely to be able
7451 to optimize. */
7453 last_insn = BB_END (bb);
7454 if (!JUMP_P (last_insn))
7455 continue;
7457 if (reg_referenced_p (cc_reg_1, PATTERN (last_insn)))
7458 cc_reg = cc_reg_1;
7459 else if (cc_reg_2 && reg_referenced_p (cc_reg_2, PATTERN (last_insn)))
7460 cc_reg = cc_reg_2;
7461 else
7462 continue;
7464 cc_src_insn = NULL;
7465 cc_src = NULL_RTX;
7466 for (insn = PREV_INSN (last_insn);
7467 insn && insn != PREV_INSN (BB_HEAD (bb));
7468 insn = PREV_INSN (insn))
7470 rtx set;
7472 if (! INSN_P (insn))
7473 continue;
7474 set = single_set (insn);
7475 if (set
7476 && REG_P (SET_DEST (set))
7477 && REGNO (SET_DEST (set)) == REGNO (cc_reg))
7479 cc_src_insn = insn;
7480 cc_src = SET_SRC (set);
7481 break;
7483 else if (reg_set_p (cc_reg, insn))
7484 break;
7487 if (! cc_src_insn)
7488 continue;
7490 if (modified_between_p (cc_src, cc_src_insn, NEXT_INSN (last_insn)))
7491 continue;
7493 /* Now CC_REG is a condition code register used for a
7494 conditional jump at the end of the block, and CC_SRC, in
7495 CC_SRC_INSN, is the value to which that condition code
7496 register is set, and CC_SRC is still meaningful at the end of
7497 the basic block. */
7499 orig_mode = GET_MODE (cc_src);
7500 mode = cse_cc_succs (bb, bb, cc_reg, cc_src, true);
7501 if (mode != VOIDmode)
7503 gcc_assert (mode == GET_MODE (cc_src));
7504 if (mode != orig_mode)
7506 rtx newreg = gen_rtx_REG (mode, REGNO (cc_reg));
7508 cse_change_cc_mode_insn (cc_src_insn, newreg);
7510 /* Do the same in the following insns that use the
7511 current value of CC_REG within BB. */
7512 cse_change_cc_mode_insns (NEXT_INSN (cc_src_insn),
7513 NEXT_INSN (last_insn),
7514 newreg);
7521 /* Perform common subexpression elimination. Nonzero value from
7522 `cse_main' means that jumps were simplified and some code may now
7523 be unreachable, so do jump optimization again. */
7524 static unsigned int
7525 rest_of_handle_cse (void)
7527 int tem;
7529 if (dump_file)
7530 dump_flow_info (dump_file, dump_flags);
7532 tem = cse_main (get_insns (), max_reg_num ());
7534 /* If we are not running more CSE passes, then we are no longer
7535 expecting CSE to be run. But always rerun it in a cheap mode. */
7536 cse_not_expected = !flag_rerun_cse_after_loop && !flag_gcse;
7538 if (tem == 2)
7540 timevar_push (TV_JUMP);
7541 rebuild_jump_labels (get_insns ());
7542 cleanup_cfg (CLEANUP_CFG_CHANGED);
7543 timevar_pop (TV_JUMP);
7545 else if (tem == 1 || optimize > 1)
7546 cleanup_cfg (0);
7548 return 0;
7551 namespace {
7553 const pass_data pass_data_cse =
7555 RTL_PASS, /* type */
7556 "cse1", /* name */
7557 OPTGROUP_NONE, /* optinfo_flags */
7558 TV_CSE, /* tv_id */
7559 0, /* properties_required */
7560 0, /* properties_provided */
7561 0, /* properties_destroyed */
7562 0, /* todo_flags_start */
7563 TODO_df_finish, /* todo_flags_finish */
7566 class pass_cse : public rtl_opt_pass
7568 public:
7569 pass_cse (gcc::context *ctxt)
7570 : rtl_opt_pass (pass_data_cse, ctxt)
7573 /* opt_pass methods: */
7574 virtual bool gate (function *) { return optimize > 0; }
7575 virtual unsigned int execute (function *) { return rest_of_handle_cse (); }
7577 }; // class pass_cse
7579 } // anon namespace
7581 rtl_opt_pass *
7582 make_pass_cse (gcc::context *ctxt)
7584 return new pass_cse (ctxt);
7588 /* Run second CSE pass after loop optimizations. */
7589 static unsigned int
7590 rest_of_handle_cse2 (void)
7592 int tem;
7594 if (dump_file)
7595 dump_flow_info (dump_file, dump_flags);
7597 tem = cse_main (get_insns (), max_reg_num ());
7599 /* Run a pass to eliminate duplicated assignments to condition code
7600 registers. We have to run this after bypass_jumps, because it
7601 makes it harder for that pass to determine whether a jump can be
7602 bypassed safely. */
7603 cse_condition_code_reg ();
7605 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7607 if (tem == 2)
7609 timevar_push (TV_JUMP);
7610 rebuild_jump_labels (get_insns ());
7611 cleanup_cfg (CLEANUP_CFG_CHANGED);
7612 timevar_pop (TV_JUMP);
7614 else if (tem == 1)
7615 cleanup_cfg (0);
7617 cse_not_expected = 1;
7618 return 0;
7622 namespace {
7624 const pass_data pass_data_cse2 =
7626 RTL_PASS, /* type */
7627 "cse2", /* name */
7628 OPTGROUP_NONE, /* optinfo_flags */
7629 TV_CSE2, /* tv_id */
7630 0, /* properties_required */
7631 0, /* properties_provided */
7632 0, /* properties_destroyed */
7633 0, /* todo_flags_start */
7634 TODO_df_finish, /* todo_flags_finish */
7637 class pass_cse2 : public rtl_opt_pass
7639 public:
7640 pass_cse2 (gcc::context *ctxt)
7641 : rtl_opt_pass (pass_data_cse2, ctxt)
7644 /* opt_pass methods: */
7645 virtual bool gate (function *)
7647 return optimize > 0 && flag_rerun_cse_after_loop;
7650 virtual unsigned int execute (function *) { return rest_of_handle_cse2 (); }
7652 }; // class pass_cse2
7654 } // anon namespace
7656 rtl_opt_pass *
7657 make_pass_cse2 (gcc::context *ctxt)
7659 return new pass_cse2 (ctxt);
7662 /* Run second CSE pass after loop optimizations. */
7663 static unsigned int
7664 rest_of_handle_cse_after_global_opts (void)
7666 int save_cfj;
7667 int tem;
7669 /* We only want to do local CSE, so don't follow jumps. */
7670 save_cfj = flag_cse_follow_jumps;
7671 flag_cse_follow_jumps = 0;
7673 rebuild_jump_labels (get_insns ());
7674 tem = cse_main (get_insns (), max_reg_num ());
7675 purge_all_dead_edges ();
7676 delete_trivially_dead_insns (get_insns (), max_reg_num ());
7678 cse_not_expected = !flag_rerun_cse_after_loop;
7680 /* If cse altered any jumps, rerun jump opts to clean things up. */
7681 if (tem == 2)
7683 timevar_push (TV_JUMP);
7684 rebuild_jump_labels (get_insns ());
7685 cleanup_cfg (CLEANUP_CFG_CHANGED);
7686 timevar_pop (TV_JUMP);
7688 else if (tem == 1)
7689 cleanup_cfg (0);
7691 flag_cse_follow_jumps = save_cfj;
7692 return 0;
7695 namespace {
7697 const pass_data pass_data_cse_after_global_opts =
7699 RTL_PASS, /* type */
7700 "cse_local", /* name */
7701 OPTGROUP_NONE, /* optinfo_flags */
7702 TV_CSE, /* tv_id */
7703 0, /* properties_required */
7704 0, /* properties_provided */
7705 0, /* properties_destroyed */
7706 0, /* todo_flags_start */
7707 TODO_df_finish, /* todo_flags_finish */
7710 class pass_cse_after_global_opts : public rtl_opt_pass
7712 public:
7713 pass_cse_after_global_opts (gcc::context *ctxt)
7714 : rtl_opt_pass (pass_data_cse_after_global_opts, ctxt)
7717 /* opt_pass methods: */
7718 virtual bool gate (function *)
7720 return optimize > 0 && flag_rerun_cse_after_global_opts;
7723 virtual unsigned int execute (function *)
7725 return rest_of_handle_cse_after_global_opts ();
7728 }; // class pass_cse_after_global_opts
7730 } // anon namespace
7732 rtl_opt_pass *
7733 make_pass_cse_after_global_opts (gcc::context *ctxt)
7735 return new pass_cse_after_global_opts (ctxt);